f64aad61372763188bbbe252864685c862b766c3
[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, device or free
1020          * space trees, make sure that we do not finish block group creation of
1021          * pending block 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, device and free space trees,
1026          * therefore we could deadlock with ourselves since we are holding a
1027          * lock on an extent buffer that btrfs_create_pending_block_groups() may
1028          * try to COW later.
1029          */
1030         if (root == fs_info->extent_root ||
1031             root == fs_info->chunk_root ||
1032             root == fs_info->dev_root ||
1033             root == fs_info->free_space_root)
1034                 trans->can_flush_pending_bgs = false;
1035
1036         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1037                         root->root_key.objectid, &disk_key, level,
1038                         search_start, empty_size);
1039         trans->can_flush_pending_bgs = true;
1040         if (IS_ERR(cow))
1041                 return PTR_ERR(cow);
1042
1043         /* cow is set to blocking by btrfs_init_new_buffer */
1044
1045         copy_extent_buffer_full(cow, buf);
1046         btrfs_set_header_bytenr(cow, cow->start);
1047         btrfs_set_header_generation(cow, trans->transid);
1048         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1049         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1050                                      BTRFS_HEADER_FLAG_RELOC);
1051         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1052                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1053         else
1054                 btrfs_set_header_owner(cow, root->root_key.objectid);
1055
1056         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
1057
1058         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1059         if (ret) {
1060                 btrfs_abort_transaction(trans, ret);
1061                 return ret;
1062         }
1063
1064         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1065                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1066                 if (ret) {
1067                         btrfs_abort_transaction(trans, ret);
1068                         return ret;
1069                 }
1070         }
1071
1072         if (buf == root->node) {
1073                 WARN_ON(parent && parent != buf);
1074                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1075                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1076                         parent_start = buf->start;
1077
1078                 extent_buffer_get(cow);
1079                 ret = tree_mod_log_insert_root(root->node, cow, 1);
1080                 BUG_ON(ret < 0);
1081                 rcu_assign_pointer(root->node, cow);
1082
1083                 btrfs_free_tree_block(trans, root, buf, parent_start,
1084                                       last_ref);
1085                 free_extent_buffer(buf);
1086                 add_root_to_dirty_list(root);
1087         } else {
1088                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1089                 tree_mod_log_insert_key(parent, parent_slot,
1090                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1091                 btrfs_set_node_blockptr(parent, parent_slot,
1092                                         cow->start);
1093                 btrfs_set_node_ptr_generation(parent, parent_slot,
1094                                               trans->transid);
1095                 btrfs_mark_buffer_dirty(parent);
1096                 if (last_ref) {
1097                         ret = tree_mod_log_free_eb(buf);
1098                         if (ret) {
1099                                 btrfs_abort_transaction(trans, ret);
1100                                 return ret;
1101                         }
1102                 }
1103                 btrfs_free_tree_block(trans, root, buf, parent_start,
1104                                       last_ref);
1105         }
1106         if (unlock_orig)
1107                 btrfs_tree_unlock(buf);
1108         free_extent_buffer_stale(buf);
1109         btrfs_mark_buffer_dirty(cow);
1110         *cow_ret = cow;
1111         return 0;
1112 }
1113
1114 /*
1115  * returns the logical address of the oldest predecessor of the given root.
1116  * entries older than time_seq are ignored.
1117  */
1118 static struct tree_mod_elem *__tree_mod_log_oldest_root(
1119                 struct extent_buffer *eb_root, u64 time_seq)
1120 {
1121         struct tree_mod_elem *tm;
1122         struct tree_mod_elem *found = NULL;
1123         u64 root_logical = eb_root->start;
1124         int looped = 0;
1125
1126         if (!time_seq)
1127                 return NULL;
1128
1129         /*
1130          * the very last operation that's logged for a root is the
1131          * replacement operation (if it is replaced at all). this has
1132          * the logical address of the *new* root, making it the very
1133          * first operation that's logged for this root.
1134          */
1135         while (1) {
1136                 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1137                                                 time_seq);
1138                 if (!looped && !tm)
1139                         return NULL;
1140                 /*
1141                  * if there are no tree operation for the oldest root, we simply
1142                  * return it. this should only happen if that (old) root is at
1143                  * level 0.
1144                  */
1145                 if (!tm)
1146                         break;
1147
1148                 /*
1149                  * if there's an operation that's not a root replacement, we
1150                  * found the oldest version of our root. normally, we'll find a
1151                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1152                  */
1153                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1154                         break;
1155
1156                 found = tm;
1157                 root_logical = tm->old_root.logical;
1158                 looped = 1;
1159         }
1160
1161         /* if there's no old root to return, return what we found instead */
1162         if (!found)
1163                 found = tm;
1164
1165         return found;
1166 }
1167
1168 /*
1169  * tm is a pointer to the first operation to rewind within eb. then, all
1170  * previous operations will be rewound (until we reach something older than
1171  * time_seq).
1172  */
1173 static void
1174 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1175                       u64 time_seq, struct tree_mod_elem *first_tm)
1176 {
1177         u32 n;
1178         struct rb_node *next;
1179         struct tree_mod_elem *tm = first_tm;
1180         unsigned long o_dst;
1181         unsigned long o_src;
1182         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1183
1184         n = btrfs_header_nritems(eb);
1185         read_lock(&fs_info->tree_mod_log_lock);
1186         while (tm && tm->seq >= time_seq) {
1187                 /*
1188                  * all the operations are recorded with the operator used for
1189                  * the modification. as we're going backwards, we do the
1190                  * opposite of each operation here.
1191                  */
1192                 switch (tm->op) {
1193                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1194                         BUG_ON(tm->slot < n);
1195                         /* Fallthrough */
1196                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1197                 case MOD_LOG_KEY_REMOVE:
1198                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1199                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1200                         btrfs_set_node_ptr_generation(eb, tm->slot,
1201                                                       tm->generation);
1202                         n++;
1203                         break;
1204                 case MOD_LOG_KEY_REPLACE:
1205                         BUG_ON(tm->slot >= n);
1206                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1207                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1208                         btrfs_set_node_ptr_generation(eb, tm->slot,
1209                                                       tm->generation);
1210                         break;
1211                 case MOD_LOG_KEY_ADD:
1212                         /* if a move operation is needed it's in the log */
1213                         n--;
1214                         break;
1215                 case MOD_LOG_MOVE_KEYS:
1216                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1217                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1218                         memmove_extent_buffer(eb, o_dst, o_src,
1219                                               tm->move.nr_items * p_size);
1220                         break;
1221                 case MOD_LOG_ROOT_REPLACE:
1222                         /*
1223                          * this operation is special. for roots, this must be
1224                          * handled explicitly before rewinding.
1225                          * for non-roots, this operation may exist if the node
1226                          * was a root: root A -> child B; then A gets empty and
1227                          * B is promoted to the new root. in the mod log, we'll
1228                          * have a root-replace operation for B, a tree block
1229                          * that is no root. we simply ignore that operation.
1230                          */
1231                         break;
1232                 }
1233                 next = rb_next(&tm->node);
1234                 if (!next)
1235                         break;
1236                 tm = rb_entry(next, struct tree_mod_elem, node);
1237                 if (tm->logical != first_tm->logical)
1238                         break;
1239         }
1240         read_unlock(&fs_info->tree_mod_log_lock);
1241         btrfs_set_header_nritems(eb, n);
1242 }
1243
1244 /*
1245  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1246  * is returned. If rewind operations happen, a fresh buffer is returned. The
1247  * returned buffer is always read-locked. If the returned buffer is not the
1248  * input buffer, the lock on the input buffer is released and the input buffer
1249  * is freed (its refcount is decremented).
1250  */
1251 static struct extent_buffer *
1252 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1253                     struct extent_buffer *eb, u64 time_seq)
1254 {
1255         struct extent_buffer *eb_rewin;
1256         struct tree_mod_elem *tm;
1257
1258         if (!time_seq)
1259                 return eb;
1260
1261         if (btrfs_header_level(eb) == 0)
1262                 return eb;
1263
1264         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1265         if (!tm)
1266                 return eb;
1267
1268         btrfs_set_path_blocking(path);
1269         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1270
1271         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1272                 BUG_ON(tm->slot != 0);
1273                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1274                 if (!eb_rewin) {
1275                         btrfs_tree_read_unlock_blocking(eb);
1276                         free_extent_buffer(eb);
1277                         return NULL;
1278                 }
1279                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1280                 btrfs_set_header_backref_rev(eb_rewin,
1281                                              btrfs_header_backref_rev(eb));
1282                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1283                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1284         } else {
1285                 eb_rewin = btrfs_clone_extent_buffer(eb);
1286                 if (!eb_rewin) {
1287                         btrfs_tree_read_unlock_blocking(eb);
1288                         free_extent_buffer(eb);
1289                         return NULL;
1290                 }
1291         }
1292
1293         btrfs_tree_read_unlock_blocking(eb);
1294         free_extent_buffer(eb);
1295
1296         btrfs_tree_read_lock(eb_rewin);
1297         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1298         WARN_ON(btrfs_header_nritems(eb_rewin) >
1299                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1300
1301         return eb_rewin;
1302 }
1303
1304 /*
1305  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1306  * value. If there are no changes, the current root->root_node is returned. If
1307  * anything changed in between, there's a fresh buffer allocated on which the
1308  * rewind operations are done. In any case, the returned buffer is read locked.
1309  * Returns NULL on error (with no locks held).
1310  */
1311 static inline struct extent_buffer *
1312 get_old_root(struct btrfs_root *root, u64 time_seq)
1313 {
1314         struct btrfs_fs_info *fs_info = root->fs_info;
1315         struct tree_mod_elem *tm;
1316         struct extent_buffer *eb = NULL;
1317         struct extent_buffer *eb_root;
1318         struct extent_buffer *old;
1319         struct tree_mod_root *old_root = NULL;
1320         u64 old_generation = 0;
1321         u64 logical;
1322         int level;
1323
1324         eb_root = btrfs_read_lock_root_node(root);
1325         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1326         if (!tm)
1327                 return eb_root;
1328
1329         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1330                 old_root = &tm->old_root;
1331                 old_generation = tm->generation;
1332                 logical = old_root->logical;
1333                 level = old_root->level;
1334         } else {
1335                 logical = eb_root->start;
1336                 level = btrfs_header_level(eb_root);
1337         }
1338
1339         tm = tree_mod_log_search(fs_info, logical, time_seq);
1340         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1341                 btrfs_tree_read_unlock(eb_root);
1342                 free_extent_buffer(eb_root);
1343                 old = read_tree_block(fs_info, logical, 0, level, NULL);
1344                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1345                         if (!IS_ERR(old))
1346                                 free_extent_buffer(old);
1347                         btrfs_warn(fs_info,
1348                                    "failed to read tree block %llu from get_old_root",
1349                                    logical);
1350                 } else {
1351                         eb = btrfs_clone_extent_buffer(old);
1352                         free_extent_buffer(old);
1353                 }
1354         } else if (old_root) {
1355                 btrfs_tree_read_unlock(eb_root);
1356                 free_extent_buffer(eb_root);
1357                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1358         } else {
1359                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1360                 eb = btrfs_clone_extent_buffer(eb_root);
1361                 btrfs_tree_read_unlock_blocking(eb_root);
1362                 free_extent_buffer(eb_root);
1363         }
1364
1365         if (!eb)
1366                 return NULL;
1367         btrfs_tree_read_lock(eb);
1368         if (old_root) {
1369                 btrfs_set_header_bytenr(eb, eb->start);
1370                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1371                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1372                 btrfs_set_header_level(eb, old_root->level);
1373                 btrfs_set_header_generation(eb, old_generation);
1374         }
1375         if (tm)
1376                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1377         else
1378                 WARN_ON(btrfs_header_level(eb) != 0);
1379         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1380
1381         return eb;
1382 }
1383
1384 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1385 {
1386         struct tree_mod_elem *tm;
1387         int level;
1388         struct extent_buffer *eb_root = btrfs_root_node(root);
1389
1390         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1391         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1392                 level = tm->old_root.level;
1393         } else {
1394                 level = btrfs_header_level(eb_root);
1395         }
1396         free_extent_buffer(eb_root);
1397
1398         return level;
1399 }
1400
1401 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1402                                    struct btrfs_root *root,
1403                                    struct extent_buffer *buf)
1404 {
1405         if (btrfs_is_testing(root->fs_info))
1406                 return 0;
1407
1408         /* Ensure we can see the FORCE_COW bit */
1409         smp_mb__before_atomic();
1410
1411         /*
1412          * We do not need to cow a block if
1413          * 1) this block is not created or changed in this transaction;
1414          * 2) this block does not belong to TREE_RELOC tree;
1415          * 3) the root is not forced COW.
1416          *
1417          * What is forced COW:
1418          *    when we create snapshot during committing the transaction,
1419          *    after we've finished copying src root, we must COW the shared
1420          *    block to ensure the metadata consistency.
1421          */
1422         if (btrfs_header_generation(buf) == trans->transid &&
1423             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1424             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1425               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1426             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1427                 return 0;
1428         return 1;
1429 }
1430
1431 /*
1432  * cows a single block, see __btrfs_cow_block for the real work.
1433  * This version of it has extra checks so that a block isn't COWed more than
1434  * once per transaction, as long as it hasn't been written yet
1435  */
1436 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1437                     struct btrfs_root *root, struct extent_buffer *buf,
1438                     struct extent_buffer *parent, int parent_slot,
1439                     struct extent_buffer **cow_ret)
1440 {
1441         struct btrfs_fs_info *fs_info = root->fs_info;
1442         u64 search_start;
1443         int ret;
1444
1445         if (test_bit(BTRFS_ROOT_DELETING, &root->state))
1446                 btrfs_err(fs_info,
1447                         "COW'ing blocks on a fs root that's being dropped");
1448
1449         if (trans->transaction != fs_info->running_transaction)
1450                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1451                        trans->transid,
1452                        fs_info->running_transaction->transid);
1453
1454         if (trans->transid != fs_info->generation)
1455                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1456                        trans->transid, fs_info->generation);
1457
1458         if (!should_cow_block(trans, root, buf)) {
1459                 trans->dirty = true;
1460                 *cow_ret = buf;
1461                 return 0;
1462         }
1463
1464         search_start = buf->start & ~((u64)SZ_1G - 1);
1465
1466         if (parent)
1467                 btrfs_set_lock_blocking(parent);
1468         btrfs_set_lock_blocking(buf);
1469
1470         ret = __btrfs_cow_block(trans, root, buf, parent,
1471                                  parent_slot, cow_ret, search_start, 0);
1472
1473         trace_btrfs_cow_block(root, buf, *cow_ret);
1474
1475         return ret;
1476 }
1477
1478 /*
1479  * helper function for defrag to decide if two blocks pointed to by a
1480  * node are actually close by
1481  */
1482 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1483 {
1484         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1485                 return 1;
1486         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1487                 return 1;
1488         return 0;
1489 }
1490
1491 /*
1492  * compare two keys in a memcmp fashion
1493  */
1494 static int comp_keys(const struct btrfs_disk_key *disk,
1495                      const struct btrfs_key *k2)
1496 {
1497         struct btrfs_key k1;
1498
1499         btrfs_disk_key_to_cpu(&k1, disk);
1500
1501         return btrfs_comp_cpu_keys(&k1, k2);
1502 }
1503
1504 /*
1505  * same as comp_keys only with two btrfs_key's
1506  */
1507 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1508 {
1509         if (k1->objectid > k2->objectid)
1510                 return 1;
1511         if (k1->objectid < k2->objectid)
1512                 return -1;
1513         if (k1->type > k2->type)
1514                 return 1;
1515         if (k1->type < k2->type)
1516                 return -1;
1517         if (k1->offset > k2->offset)
1518                 return 1;
1519         if (k1->offset < k2->offset)
1520                 return -1;
1521         return 0;
1522 }
1523
1524 /*
1525  * this is used by the defrag code to go through all the
1526  * leaves pointed to by a node and reallocate them so that
1527  * disk order is close to key order
1528  */
1529 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1530                        struct btrfs_root *root, struct extent_buffer *parent,
1531                        int start_slot, u64 *last_ret,
1532                        struct btrfs_key *progress)
1533 {
1534         struct btrfs_fs_info *fs_info = root->fs_info;
1535         struct extent_buffer *cur;
1536         u64 blocknr;
1537         u64 gen;
1538         u64 search_start = *last_ret;
1539         u64 last_block = 0;
1540         u64 other;
1541         u32 parent_nritems;
1542         int end_slot;
1543         int i;
1544         int err = 0;
1545         int parent_level;
1546         int uptodate;
1547         u32 blocksize;
1548         int progress_passed = 0;
1549         struct btrfs_disk_key disk_key;
1550
1551         parent_level = btrfs_header_level(parent);
1552
1553         WARN_ON(trans->transaction != fs_info->running_transaction);
1554         WARN_ON(trans->transid != fs_info->generation);
1555
1556         parent_nritems = btrfs_header_nritems(parent);
1557         blocksize = fs_info->nodesize;
1558         end_slot = parent_nritems - 1;
1559
1560         if (parent_nritems <= 1)
1561                 return 0;
1562
1563         btrfs_set_lock_blocking(parent);
1564
1565         for (i = start_slot; i <= end_slot; i++) {
1566                 struct btrfs_key first_key;
1567                 int close = 1;
1568
1569                 btrfs_node_key(parent, &disk_key, i);
1570                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1571                         continue;
1572
1573                 progress_passed = 1;
1574                 blocknr = btrfs_node_blockptr(parent, i);
1575                 gen = btrfs_node_ptr_generation(parent, i);
1576                 btrfs_node_key_to_cpu(parent, &first_key, i);
1577                 if (last_block == 0)
1578                         last_block = blocknr;
1579
1580                 if (i > 0) {
1581                         other = btrfs_node_blockptr(parent, i - 1);
1582                         close = close_blocks(blocknr, other, blocksize);
1583                 }
1584                 if (!close && i < end_slot) {
1585                         other = btrfs_node_blockptr(parent, i + 1);
1586                         close = close_blocks(blocknr, other, blocksize);
1587                 }
1588                 if (close) {
1589                         last_block = blocknr;
1590                         continue;
1591                 }
1592
1593                 cur = find_extent_buffer(fs_info, blocknr);
1594                 if (cur)
1595                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1596                 else
1597                         uptodate = 0;
1598                 if (!cur || !uptodate) {
1599                         if (!cur) {
1600                                 cur = read_tree_block(fs_info, blocknr, gen,
1601                                                       parent_level - 1,
1602                                                       &first_key);
1603                                 if (IS_ERR(cur)) {
1604                                         return PTR_ERR(cur);
1605                                 } else if (!extent_buffer_uptodate(cur)) {
1606                                         free_extent_buffer(cur);
1607                                         return -EIO;
1608                                 }
1609                         } else if (!uptodate) {
1610                                 err = btrfs_read_buffer(cur, gen,
1611                                                 parent_level - 1,&first_key);
1612                                 if (err) {
1613                                         free_extent_buffer(cur);
1614                                         return err;
1615                                 }
1616                         }
1617                 }
1618                 if (search_start == 0)
1619                         search_start = last_block;
1620
1621                 btrfs_tree_lock(cur);
1622                 btrfs_set_lock_blocking(cur);
1623                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1624                                         &cur, search_start,
1625                                         min(16 * blocksize,
1626                                             (end_slot - i) * blocksize));
1627                 if (err) {
1628                         btrfs_tree_unlock(cur);
1629                         free_extent_buffer(cur);
1630                         break;
1631                 }
1632                 search_start = cur->start;
1633                 last_block = cur->start;
1634                 *last_ret = search_start;
1635                 btrfs_tree_unlock(cur);
1636                 free_extent_buffer(cur);
1637         }
1638         return err;
1639 }
1640
1641 /*
1642  * search for key in the extent_buffer.  The items start at offset p,
1643  * and they are item_size apart.  There are 'max' items in p.
1644  *
1645  * the slot in the array is returned via slot, and it points to
1646  * the place where you would insert key if it is not found in
1647  * the array.
1648  *
1649  * slot may point to max if the key is bigger than all of the keys
1650  */
1651 static noinline int generic_bin_search(struct extent_buffer *eb,
1652                                        unsigned long p, int item_size,
1653                                        const struct btrfs_key *key,
1654                                        int max, int *slot)
1655 {
1656         int low = 0;
1657         int high = max;
1658         int mid;
1659         int ret;
1660         struct btrfs_disk_key *tmp = NULL;
1661         struct btrfs_disk_key unaligned;
1662         unsigned long offset;
1663         char *kaddr = NULL;
1664         unsigned long map_start = 0;
1665         unsigned long map_len = 0;
1666         int err;
1667
1668         if (low > high) {
1669                 btrfs_err(eb->fs_info,
1670                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1671                           __func__, low, high, eb->start,
1672                           btrfs_header_owner(eb), btrfs_header_level(eb));
1673                 return -EINVAL;
1674         }
1675
1676         while (low < high) {
1677                 mid = (low + high) / 2;
1678                 offset = p + mid * item_size;
1679
1680                 if (!kaddr || offset < map_start ||
1681                     (offset + sizeof(struct btrfs_disk_key)) >
1682                     map_start + map_len) {
1683
1684                         err = map_private_extent_buffer(eb, offset,
1685                                                 sizeof(struct btrfs_disk_key),
1686                                                 &kaddr, &map_start, &map_len);
1687
1688                         if (!err) {
1689                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1690                                                         map_start);
1691                         } else if (err == 1) {
1692                                 read_extent_buffer(eb, &unaligned,
1693                                                    offset, sizeof(unaligned));
1694                                 tmp = &unaligned;
1695                         } else {
1696                                 return err;
1697                         }
1698
1699                 } else {
1700                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1701                                                         map_start);
1702                 }
1703                 ret = comp_keys(tmp, key);
1704
1705                 if (ret < 0)
1706                         low = mid + 1;
1707                 else if (ret > 0)
1708                         high = mid;
1709                 else {
1710                         *slot = mid;
1711                         return 0;
1712                 }
1713         }
1714         *slot = low;
1715         return 1;
1716 }
1717
1718 /*
1719  * simple bin_search frontend that does the right thing for
1720  * leaves vs nodes
1721  */
1722 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1723                      int level, int *slot)
1724 {
1725         if (level == 0)
1726                 return generic_bin_search(eb,
1727                                           offsetof(struct btrfs_leaf, items),
1728                                           sizeof(struct btrfs_item),
1729                                           key, btrfs_header_nritems(eb),
1730                                           slot);
1731         else
1732                 return generic_bin_search(eb,
1733                                           offsetof(struct btrfs_node, ptrs),
1734                                           sizeof(struct btrfs_key_ptr),
1735                                           key, btrfs_header_nritems(eb),
1736                                           slot);
1737 }
1738
1739 static void root_add_used(struct btrfs_root *root, u32 size)
1740 {
1741         spin_lock(&root->accounting_lock);
1742         btrfs_set_root_used(&root->root_item,
1743                             btrfs_root_used(&root->root_item) + size);
1744         spin_unlock(&root->accounting_lock);
1745 }
1746
1747 static void root_sub_used(struct btrfs_root *root, u32 size)
1748 {
1749         spin_lock(&root->accounting_lock);
1750         btrfs_set_root_used(&root->root_item,
1751                             btrfs_root_used(&root->root_item) - size);
1752         spin_unlock(&root->accounting_lock);
1753 }
1754
1755 /* given a node and slot number, this reads the blocks it points to.  The
1756  * extent buffer is returned with a reference taken (but unlocked).
1757  */
1758 static noinline struct extent_buffer *
1759 read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1760                int slot)
1761 {
1762         int level = btrfs_header_level(parent);
1763         struct extent_buffer *eb;
1764         struct btrfs_key first_key;
1765
1766         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1767                 return ERR_PTR(-ENOENT);
1768
1769         BUG_ON(level == 0);
1770
1771         btrfs_node_key_to_cpu(parent, &first_key, slot);
1772         eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1773                              btrfs_node_ptr_generation(parent, slot),
1774                              level - 1, &first_key);
1775         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1776                 free_extent_buffer(eb);
1777                 eb = ERR_PTR(-EIO);
1778         }
1779
1780         return eb;
1781 }
1782
1783 /*
1784  * node level balancing, used to make sure nodes are in proper order for
1785  * item deletion.  We balance from the top down, so we have to make sure
1786  * that a deletion won't leave an node completely empty later on.
1787  */
1788 static noinline int balance_level(struct btrfs_trans_handle *trans,
1789                          struct btrfs_root *root,
1790                          struct btrfs_path *path, int level)
1791 {
1792         struct btrfs_fs_info *fs_info = root->fs_info;
1793         struct extent_buffer *right = NULL;
1794         struct extent_buffer *mid;
1795         struct extent_buffer *left = NULL;
1796         struct extent_buffer *parent = NULL;
1797         int ret = 0;
1798         int wret;
1799         int pslot;
1800         int orig_slot = path->slots[level];
1801         u64 orig_ptr;
1802
1803         ASSERT(level > 0);
1804
1805         mid = path->nodes[level];
1806
1807         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1808                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1809         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1810
1811         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1812
1813         if (level < BTRFS_MAX_LEVEL - 1) {
1814                 parent = path->nodes[level + 1];
1815                 pslot = path->slots[level + 1];
1816         }
1817
1818         /*
1819          * deal with the case where there is only one pointer in the root
1820          * by promoting the node below to a root
1821          */
1822         if (!parent) {
1823                 struct extent_buffer *child;
1824
1825                 if (btrfs_header_nritems(mid) != 1)
1826                         return 0;
1827
1828                 /* promote the child to a root */
1829                 child = read_node_slot(fs_info, mid, 0);
1830                 if (IS_ERR(child)) {
1831                         ret = PTR_ERR(child);
1832                         btrfs_handle_fs_error(fs_info, ret, NULL);
1833                         goto enospc;
1834                 }
1835
1836                 btrfs_tree_lock(child);
1837                 btrfs_set_lock_blocking(child);
1838                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1839                 if (ret) {
1840                         btrfs_tree_unlock(child);
1841                         free_extent_buffer(child);
1842                         goto enospc;
1843                 }
1844
1845                 ret = tree_mod_log_insert_root(root->node, child, 1);
1846                 BUG_ON(ret < 0);
1847                 rcu_assign_pointer(root->node, child);
1848
1849                 add_root_to_dirty_list(root);
1850                 btrfs_tree_unlock(child);
1851
1852                 path->locks[level] = 0;
1853                 path->nodes[level] = NULL;
1854                 clean_tree_block(fs_info, mid);
1855                 btrfs_tree_unlock(mid);
1856                 /* once for the path */
1857                 free_extent_buffer(mid);
1858
1859                 root_sub_used(root, mid->len);
1860                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1861                 /* once for the root ptr */
1862                 free_extent_buffer_stale(mid);
1863                 return 0;
1864         }
1865         if (btrfs_header_nritems(mid) >
1866             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1867                 return 0;
1868
1869         left = read_node_slot(fs_info, parent, pslot - 1);
1870         if (IS_ERR(left))
1871                 left = NULL;
1872
1873         if (left) {
1874                 btrfs_tree_lock(left);
1875                 btrfs_set_lock_blocking(left);
1876                 wret = btrfs_cow_block(trans, root, left,
1877                                        parent, pslot - 1, &left);
1878                 if (wret) {
1879                         ret = wret;
1880                         goto enospc;
1881                 }
1882         }
1883
1884         right = read_node_slot(fs_info, parent, pslot + 1);
1885         if (IS_ERR(right))
1886                 right = NULL;
1887
1888         if (right) {
1889                 btrfs_tree_lock(right);
1890                 btrfs_set_lock_blocking(right);
1891                 wret = btrfs_cow_block(trans, root, right,
1892                                        parent, pslot + 1, &right);
1893                 if (wret) {
1894                         ret = wret;
1895                         goto enospc;
1896                 }
1897         }
1898
1899         /* first, try to make some room in the middle buffer */
1900         if (left) {
1901                 orig_slot += btrfs_header_nritems(left);
1902                 wret = push_node_left(trans, fs_info, left, mid, 1);
1903                 if (wret < 0)
1904                         ret = wret;
1905         }
1906
1907         /*
1908          * then try to empty the right most buffer into the middle
1909          */
1910         if (right) {
1911                 wret = push_node_left(trans, fs_info, mid, right, 1);
1912                 if (wret < 0 && wret != -ENOSPC)
1913                         ret = wret;
1914                 if (btrfs_header_nritems(right) == 0) {
1915                         clean_tree_block(fs_info, right);
1916                         btrfs_tree_unlock(right);
1917                         del_ptr(root, path, level + 1, pslot + 1);
1918                         root_sub_used(root, right->len);
1919                         btrfs_free_tree_block(trans, root, right, 0, 1);
1920                         free_extent_buffer_stale(right);
1921                         right = NULL;
1922                 } else {
1923                         struct btrfs_disk_key right_key;
1924                         btrfs_node_key(right, &right_key, 0);
1925                         ret = tree_mod_log_insert_key(parent, pslot + 1,
1926                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1927                         BUG_ON(ret < 0);
1928                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1929                         btrfs_mark_buffer_dirty(parent);
1930                 }
1931         }
1932         if (btrfs_header_nritems(mid) == 1) {
1933                 /*
1934                  * we're not allowed to leave a node with one item in the
1935                  * tree during a delete.  A deletion from lower in the tree
1936                  * could try to delete the only pointer in this node.
1937                  * So, pull some keys from the left.
1938                  * There has to be a left pointer at this point because
1939                  * otherwise we would have pulled some pointers from the
1940                  * right
1941                  */
1942                 if (!left) {
1943                         ret = -EROFS;
1944                         btrfs_handle_fs_error(fs_info, ret, NULL);
1945                         goto enospc;
1946                 }
1947                 wret = balance_node_right(trans, fs_info, mid, left);
1948                 if (wret < 0) {
1949                         ret = wret;
1950                         goto enospc;
1951                 }
1952                 if (wret == 1) {
1953                         wret = push_node_left(trans, fs_info, left, mid, 1);
1954                         if (wret < 0)
1955                                 ret = wret;
1956                 }
1957                 BUG_ON(wret == 1);
1958         }
1959         if (btrfs_header_nritems(mid) == 0) {
1960                 clean_tree_block(fs_info, mid);
1961                 btrfs_tree_unlock(mid);
1962                 del_ptr(root, path, level + 1, pslot);
1963                 root_sub_used(root, mid->len);
1964                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1965                 free_extent_buffer_stale(mid);
1966                 mid = NULL;
1967         } else {
1968                 /* update the parent key to reflect our changes */
1969                 struct btrfs_disk_key mid_key;
1970                 btrfs_node_key(mid, &mid_key, 0);
1971                 ret = tree_mod_log_insert_key(parent, pslot,
1972                                 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1973                 BUG_ON(ret < 0);
1974                 btrfs_set_node_key(parent, &mid_key, pslot);
1975                 btrfs_mark_buffer_dirty(parent);
1976         }
1977
1978         /* update the path */
1979         if (left) {
1980                 if (btrfs_header_nritems(left) > orig_slot) {
1981                         extent_buffer_get(left);
1982                         /* left was locked after cow */
1983                         path->nodes[level] = left;
1984                         path->slots[level + 1] -= 1;
1985                         path->slots[level] = orig_slot;
1986                         if (mid) {
1987                                 btrfs_tree_unlock(mid);
1988                                 free_extent_buffer(mid);
1989                         }
1990                 } else {
1991                         orig_slot -= btrfs_header_nritems(left);
1992                         path->slots[level] = orig_slot;
1993                 }
1994         }
1995         /* double check we haven't messed things up */
1996         if (orig_ptr !=
1997             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1998                 BUG();
1999 enospc:
2000         if (right) {
2001                 btrfs_tree_unlock(right);
2002                 free_extent_buffer(right);
2003         }
2004         if (left) {
2005                 if (path->nodes[level] != left)
2006                         btrfs_tree_unlock(left);
2007                 free_extent_buffer(left);
2008         }
2009         return ret;
2010 }
2011
2012 /* Node balancing for insertion.  Here we only split or push nodes around
2013  * when they are completely full.  This is also done top down, so we
2014  * have to be pessimistic.
2015  */
2016 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2017                                           struct btrfs_root *root,
2018                                           struct btrfs_path *path, int level)
2019 {
2020         struct btrfs_fs_info *fs_info = root->fs_info;
2021         struct extent_buffer *right = NULL;
2022         struct extent_buffer *mid;
2023         struct extent_buffer *left = NULL;
2024         struct extent_buffer *parent = NULL;
2025         int ret = 0;
2026         int wret;
2027         int pslot;
2028         int orig_slot = path->slots[level];
2029
2030         if (level == 0)
2031                 return 1;
2032
2033         mid = path->nodes[level];
2034         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2035
2036         if (level < BTRFS_MAX_LEVEL - 1) {
2037                 parent = path->nodes[level + 1];
2038                 pslot = path->slots[level + 1];
2039         }
2040
2041         if (!parent)
2042                 return 1;
2043
2044         left = read_node_slot(fs_info, parent, pslot - 1);
2045         if (IS_ERR(left))
2046                 left = NULL;
2047
2048         /* first, try to make some room in the middle buffer */
2049         if (left) {
2050                 u32 left_nr;
2051
2052                 btrfs_tree_lock(left);
2053                 btrfs_set_lock_blocking(left);
2054
2055                 left_nr = btrfs_header_nritems(left);
2056                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2057                         wret = 1;
2058                 } else {
2059                         ret = btrfs_cow_block(trans, root, left, parent,
2060                                               pslot - 1, &left);
2061                         if (ret)
2062                                 wret = 1;
2063                         else {
2064                                 wret = push_node_left(trans, fs_info,
2065                                                       left, mid, 0);
2066                         }
2067                 }
2068                 if (wret < 0)
2069                         ret = wret;
2070                 if (wret == 0) {
2071                         struct btrfs_disk_key disk_key;
2072                         orig_slot += left_nr;
2073                         btrfs_node_key(mid, &disk_key, 0);
2074                         ret = tree_mod_log_insert_key(parent, pslot,
2075                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2076                         BUG_ON(ret < 0);
2077                         btrfs_set_node_key(parent, &disk_key, pslot);
2078                         btrfs_mark_buffer_dirty(parent);
2079                         if (btrfs_header_nritems(left) > orig_slot) {
2080                                 path->nodes[level] = left;
2081                                 path->slots[level + 1] -= 1;
2082                                 path->slots[level] = orig_slot;
2083                                 btrfs_tree_unlock(mid);
2084                                 free_extent_buffer(mid);
2085                         } else {
2086                                 orig_slot -=
2087                                         btrfs_header_nritems(left);
2088                                 path->slots[level] = orig_slot;
2089                                 btrfs_tree_unlock(left);
2090                                 free_extent_buffer(left);
2091                         }
2092                         return 0;
2093                 }
2094                 btrfs_tree_unlock(left);
2095                 free_extent_buffer(left);
2096         }
2097         right = read_node_slot(fs_info, parent, pslot + 1);
2098         if (IS_ERR(right))
2099                 right = NULL;
2100
2101         /*
2102          * then try to empty the right most buffer into the middle
2103          */
2104         if (right) {
2105                 u32 right_nr;
2106
2107                 btrfs_tree_lock(right);
2108                 btrfs_set_lock_blocking(right);
2109
2110                 right_nr = btrfs_header_nritems(right);
2111                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2112                         wret = 1;
2113                 } else {
2114                         ret = btrfs_cow_block(trans, root, right,
2115                                               parent, pslot + 1,
2116                                               &right);
2117                         if (ret)
2118                                 wret = 1;
2119                         else {
2120                                 wret = balance_node_right(trans, fs_info,
2121                                                           right, mid);
2122                         }
2123                 }
2124                 if (wret < 0)
2125                         ret = wret;
2126                 if (wret == 0) {
2127                         struct btrfs_disk_key disk_key;
2128
2129                         btrfs_node_key(right, &disk_key, 0);
2130                         ret = tree_mod_log_insert_key(parent, pslot + 1,
2131                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2132                         BUG_ON(ret < 0);
2133                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2134                         btrfs_mark_buffer_dirty(parent);
2135
2136                         if (btrfs_header_nritems(mid) <= orig_slot) {
2137                                 path->nodes[level] = right;
2138                                 path->slots[level + 1] += 1;
2139                                 path->slots[level] = orig_slot -
2140                                         btrfs_header_nritems(mid);
2141                                 btrfs_tree_unlock(mid);
2142                                 free_extent_buffer(mid);
2143                         } else {
2144                                 btrfs_tree_unlock(right);
2145                                 free_extent_buffer(right);
2146                         }
2147                         return 0;
2148                 }
2149                 btrfs_tree_unlock(right);
2150                 free_extent_buffer(right);
2151         }
2152         return 1;
2153 }
2154
2155 /*
2156  * readahead one full node of leaves, finding things that are close
2157  * to the block in 'slot', and triggering ra on them.
2158  */
2159 static void reada_for_search(struct btrfs_fs_info *fs_info,
2160                              struct btrfs_path *path,
2161                              int level, int slot, u64 objectid)
2162 {
2163         struct extent_buffer *node;
2164         struct btrfs_disk_key disk_key;
2165         u32 nritems;
2166         u64 search;
2167         u64 target;
2168         u64 nread = 0;
2169         struct extent_buffer *eb;
2170         u32 nr;
2171         u32 blocksize;
2172         u32 nscan = 0;
2173
2174         if (level != 1)
2175                 return;
2176
2177         if (!path->nodes[level])
2178                 return;
2179
2180         node = path->nodes[level];
2181
2182         search = btrfs_node_blockptr(node, slot);
2183         blocksize = fs_info->nodesize;
2184         eb = find_extent_buffer(fs_info, search);
2185         if (eb) {
2186                 free_extent_buffer(eb);
2187                 return;
2188         }
2189
2190         target = search;
2191
2192         nritems = btrfs_header_nritems(node);
2193         nr = slot;
2194
2195         while (1) {
2196                 if (path->reada == READA_BACK) {
2197                         if (nr == 0)
2198                                 break;
2199                         nr--;
2200                 } else if (path->reada == READA_FORWARD) {
2201                         nr++;
2202                         if (nr >= nritems)
2203                                 break;
2204                 }
2205                 if (path->reada == READA_BACK && objectid) {
2206                         btrfs_node_key(node, &disk_key, nr);
2207                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2208                                 break;
2209                 }
2210                 search = btrfs_node_blockptr(node, nr);
2211                 if ((search <= target && target - search <= 65536) ||
2212                     (search > target && search - target <= 65536)) {
2213                         readahead_tree_block(fs_info, search);
2214                         nread += blocksize;
2215                 }
2216                 nscan++;
2217                 if ((nread > 65536 || nscan > 32))
2218                         break;
2219         }
2220 }
2221
2222 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2223                                        struct btrfs_path *path, int level)
2224 {
2225         int slot;
2226         int nritems;
2227         struct extent_buffer *parent;
2228         struct extent_buffer *eb;
2229         u64 gen;
2230         u64 block1 = 0;
2231         u64 block2 = 0;
2232
2233         parent = path->nodes[level + 1];
2234         if (!parent)
2235                 return;
2236
2237         nritems = btrfs_header_nritems(parent);
2238         slot = path->slots[level + 1];
2239
2240         if (slot > 0) {
2241                 block1 = btrfs_node_blockptr(parent, slot - 1);
2242                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2243                 eb = find_extent_buffer(fs_info, block1);
2244                 /*
2245                  * if we get -eagain from btrfs_buffer_uptodate, we
2246                  * don't want to return eagain here.  That will loop
2247                  * forever
2248                  */
2249                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2250                         block1 = 0;
2251                 free_extent_buffer(eb);
2252         }
2253         if (slot + 1 < nritems) {
2254                 block2 = btrfs_node_blockptr(parent, slot + 1);
2255                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2256                 eb = find_extent_buffer(fs_info, block2);
2257                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2258                         block2 = 0;
2259                 free_extent_buffer(eb);
2260         }
2261
2262         if (block1)
2263                 readahead_tree_block(fs_info, block1);
2264         if (block2)
2265                 readahead_tree_block(fs_info, block2);
2266 }
2267
2268
2269 /*
2270  * when we walk down the tree, it is usually safe to unlock the higher layers
2271  * in the tree.  The exceptions are when our path goes through slot 0, because
2272  * operations on the tree might require changing key pointers higher up in the
2273  * tree.
2274  *
2275  * callers might also have set path->keep_locks, which tells this code to keep
2276  * the lock if the path points to the last slot in the block.  This is part of
2277  * walking through the tree, and selecting the next slot in the higher block.
2278  *
2279  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2280  * if lowest_unlock is 1, level 0 won't be unlocked
2281  */
2282 static noinline void unlock_up(struct btrfs_path *path, int level,
2283                                int lowest_unlock, int min_write_lock_level,
2284                                int *write_lock_level)
2285 {
2286         int i;
2287         int skip_level = level;
2288         int no_skips = 0;
2289         struct extent_buffer *t;
2290
2291         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2292                 if (!path->nodes[i])
2293                         break;
2294                 if (!path->locks[i])
2295                         break;
2296                 if (!no_skips && path->slots[i] == 0) {
2297                         skip_level = i + 1;
2298                         continue;
2299                 }
2300                 if (!no_skips && path->keep_locks) {
2301                         u32 nritems;
2302                         t = path->nodes[i];
2303                         nritems = btrfs_header_nritems(t);
2304                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2305                                 skip_level = i + 1;
2306                                 continue;
2307                         }
2308                 }
2309                 if (skip_level < i && i >= lowest_unlock)
2310                         no_skips = 1;
2311
2312                 t = path->nodes[i];
2313                 if (i >= lowest_unlock && i > skip_level) {
2314                         btrfs_tree_unlock_rw(t, path->locks[i]);
2315                         path->locks[i] = 0;
2316                         if (write_lock_level &&
2317                             i > min_write_lock_level &&
2318                             i <= *write_lock_level) {
2319                                 *write_lock_level = i - 1;
2320                         }
2321                 }
2322         }
2323 }
2324
2325 /*
2326  * This releases any locks held in the path starting at level and
2327  * going all the way up to the root.
2328  *
2329  * btrfs_search_slot will keep the lock held on higher nodes in a few
2330  * corner cases, such as COW of the block at slot zero in the node.  This
2331  * ignores those rules, and it should only be called when there are no
2332  * more updates to be done higher up in the tree.
2333  */
2334 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2335 {
2336         int i;
2337
2338         if (path->keep_locks)
2339                 return;
2340
2341         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2342                 if (!path->nodes[i])
2343                         continue;
2344                 if (!path->locks[i])
2345                         continue;
2346                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2347                 path->locks[i] = 0;
2348         }
2349 }
2350
2351 /*
2352  * helper function for btrfs_search_slot.  The goal is to find a block
2353  * in cache without setting the path to blocking.  If we find the block
2354  * we return zero and the path is unchanged.
2355  *
2356  * If we can't find the block, we set the path blocking and do some
2357  * reada.  -EAGAIN is returned and the search must be repeated.
2358  */
2359 static int
2360 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2361                       struct extent_buffer **eb_ret, int level, int slot,
2362                       const struct btrfs_key *key)
2363 {
2364         struct btrfs_fs_info *fs_info = root->fs_info;
2365         u64 blocknr;
2366         u64 gen;
2367         struct extent_buffer *b = *eb_ret;
2368         struct extent_buffer *tmp;
2369         struct btrfs_key first_key;
2370         int ret;
2371         int parent_level;
2372
2373         blocknr = btrfs_node_blockptr(b, slot);
2374         gen = btrfs_node_ptr_generation(b, slot);
2375         parent_level = btrfs_header_level(b);
2376         btrfs_node_key_to_cpu(b, &first_key, slot);
2377
2378         tmp = find_extent_buffer(fs_info, blocknr);
2379         if (tmp) {
2380                 /* first we do an atomic uptodate check */
2381                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2382                         *eb_ret = tmp;
2383                         return 0;
2384                 }
2385
2386                 /* the pages were up to date, but we failed
2387                  * the generation number check.  Do a full
2388                  * read for the generation number that is correct.
2389                  * We must do this without dropping locks so
2390                  * we can trust our generation number
2391                  */
2392                 btrfs_set_path_blocking(p);
2393
2394                 /* now we're allowed to do a blocking uptodate check */
2395                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2396                 if (!ret) {
2397                         *eb_ret = tmp;
2398                         return 0;
2399                 }
2400                 free_extent_buffer(tmp);
2401                 btrfs_release_path(p);
2402                 return -EIO;
2403         }
2404
2405         /*
2406          * reduce lock contention at high levels
2407          * of the btree by dropping locks before
2408          * we read.  Don't release the lock on the current
2409          * level because we need to walk this node to figure
2410          * out which blocks to read.
2411          */
2412         btrfs_unlock_up_safe(p, level + 1);
2413         btrfs_set_path_blocking(p);
2414
2415         if (p->reada != READA_NONE)
2416                 reada_for_search(fs_info, p, level, slot, key->objectid);
2417
2418         ret = -EAGAIN;
2419         tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2420                               &first_key);
2421         if (!IS_ERR(tmp)) {
2422                 /*
2423                  * If the read above didn't mark this buffer up to date,
2424                  * it will never end up being up to date.  Set ret to EIO now
2425                  * and give up so that our caller doesn't loop forever
2426                  * on our EAGAINs.
2427                  */
2428                 if (!extent_buffer_uptodate(tmp))
2429                         ret = -EIO;
2430                 free_extent_buffer(tmp);
2431         } else {
2432                 ret = PTR_ERR(tmp);
2433         }
2434
2435         btrfs_release_path(p);
2436         return ret;
2437 }
2438
2439 /*
2440  * helper function for btrfs_search_slot.  This does all of the checks
2441  * for node-level blocks and does any balancing required based on
2442  * the ins_len.
2443  *
2444  * If no extra work was required, zero is returned.  If we had to
2445  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2446  * start over
2447  */
2448 static int
2449 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2450                        struct btrfs_root *root, struct btrfs_path *p,
2451                        struct extent_buffer *b, int level, int ins_len,
2452                        int *write_lock_level)
2453 {
2454         struct btrfs_fs_info *fs_info = root->fs_info;
2455         int ret;
2456
2457         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2458             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2459                 int sret;
2460
2461                 if (*write_lock_level < level + 1) {
2462                         *write_lock_level = level + 1;
2463                         btrfs_release_path(p);
2464                         goto again;
2465                 }
2466
2467                 btrfs_set_path_blocking(p);
2468                 reada_for_balance(fs_info, p, level);
2469                 sret = split_node(trans, root, p, level);
2470
2471                 BUG_ON(sret > 0);
2472                 if (sret) {
2473                         ret = sret;
2474                         goto done;
2475                 }
2476                 b = p->nodes[level];
2477         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2478                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2479                 int sret;
2480
2481                 if (*write_lock_level < level + 1) {
2482                         *write_lock_level = level + 1;
2483                         btrfs_release_path(p);
2484                         goto again;
2485                 }
2486
2487                 btrfs_set_path_blocking(p);
2488                 reada_for_balance(fs_info, p, level);
2489                 sret = balance_level(trans, root, p, level);
2490
2491                 if (sret) {
2492                         ret = sret;
2493                         goto done;
2494                 }
2495                 b = p->nodes[level];
2496                 if (!b) {
2497                         btrfs_release_path(p);
2498                         goto again;
2499                 }
2500                 BUG_ON(btrfs_header_nritems(b) == 1);
2501         }
2502         return 0;
2503
2504 again:
2505         ret = -EAGAIN;
2506 done:
2507         return ret;
2508 }
2509
2510 static void key_search_validate(struct extent_buffer *b,
2511                                 const struct btrfs_key *key,
2512                                 int level)
2513 {
2514 #ifdef CONFIG_BTRFS_ASSERT
2515         struct btrfs_disk_key disk_key;
2516
2517         btrfs_cpu_key_to_disk(&disk_key, key);
2518
2519         if (level == 0)
2520                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2521                     offsetof(struct btrfs_leaf, items[0].key),
2522                     sizeof(disk_key)));
2523         else
2524                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2525                     offsetof(struct btrfs_node, ptrs[0].key),
2526                     sizeof(disk_key)));
2527 #endif
2528 }
2529
2530 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2531                       int level, int *prev_cmp, int *slot)
2532 {
2533         if (*prev_cmp != 0) {
2534                 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2535                 return *prev_cmp;
2536         }
2537
2538         key_search_validate(b, key, level);
2539         *slot = 0;
2540
2541         return 0;
2542 }
2543
2544 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2545                 u64 iobjectid, u64 ioff, u8 key_type,
2546                 struct btrfs_key *found_key)
2547 {
2548         int ret;
2549         struct btrfs_key key;
2550         struct extent_buffer *eb;
2551
2552         ASSERT(path);
2553         ASSERT(found_key);
2554
2555         key.type = key_type;
2556         key.objectid = iobjectid;
2557         key.offset = ioff;
2558
2559         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2560         if (ret < 0)
2561                 return ret;
2562
2563         eb = path->nodes[0];
2564         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2565                 ret = btrfs_next_leaf(fs_root, path);
2566                 if (ret)
2567                         return ret;
2568                 eb = path->nodes[0];
2569         }
2570
2571         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2572         if (found_key->type != key.type ||
2573                         found_key->objectid != key.objectid)
2574                 return 1;
2575
2576         return 0;
2577 }
2578
2579 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2580                                                         struct btrfs_path *p,
2581                                                         int write_lock_level)
2582 {
2583         struct btrfs_fs_info *fs_info = root->fs_info;
2584         struct extent_buffer *b;
2585         int root_lock;
2586         int level = 0;
2587
2588         /* We try very hard to do read locks on the root */
2589         root_lock = BTRFS_READ_LOCK;
2590
2591         if (p->search_commit_root) {
2592                 /*
2593                  * The commit roots are read only so we always do read locks,
2594                  * and we always must hold the commit_root_sem when doing
2595                  * searches on them, the only exception is send where we don't
2596                  * want to block transaction commits for a long time, so
2597                  * we need to clone the commit root in order to avoid races
2598                  * with transaction commits that create a snapshot of one of
2599                  * the roots used by a send operation.
2600                  */
2601                 if (p->need_commit_sem) {
2602                         down_read(&fs_info->commit_root_sem);
2603                         b = btrfs_clone_extent_buffer(root->commit_root);
2604                         up_read(&fs_info->commit_root_sem);
2605                         if (!b)
2606                                 return ERR_PTR(-ENOMEM);
2607
2608                 } else {
2609                         b = root->commit_root;
2610                         extent_buffer_get(b);
2611                 }
2612                 level = btrfs_header_level(b);
2613                 /*
2614                  * Ensure that all callers have set skip_locking when
2615                  * p->search_commit_root = 1.
2616                  */
2617                 ASSERT(p->skip_locking == 1);
2618
2619                 goto out;
2620         }
2621
2622         if (p->skip_locking) {
2623                 b = btrfs_root_node(root);
2624                 level = btrfs_header_level(b);
2625                 goto out;
2626         }
2627
2628         /*
2629          * If the level is set to maximum, we can skip trying to get the read
2630          * lock.
2631          */
2632         if (write_lock_level < BTRFS_MAX_LEVEL) {
2633                 /*
2634                  * We don't know the level of the root node until we actually
2635                  * have it read locked
2636                  */
2637                 b = btrfs_read_lock_root_node(root);
2638                 level = btrfs_header_level(b);
2639                 if (level > write_lock_level)
2640                         goto out;
2641
2642                 /* Whoops, must trade for write lock */
2643                 btrfs_tree_read_unlock(b);
2644                 free_extent_buffer(b);
2645         }
2646
2647         b = btrfs_lock_root_node(root);
2648         root_lock = BTRFS_WRITE_LOCK;
2649
2650         /* The level might have changed, check again */
2651         level = btrfs_header_level(b);
2652
2653 out:
2654         p->nodes[level] = b;
2655         if (!p->skip_locking)
2656                 p->locks[level] = root_lock;
2657         /*
2658          * Callers are responsible for dropping b's references.
2659          */
2660         return b;
2661 }
2662
2663
2664 /*
2665  * btrfs_search_slot - look for a key in a tree and perform necessary
2666  * modifications to preserve tree invariants.
2667  *
2668  * @trans:      Handle of transaction, used when modifying the tree
2669  * @p:          Holds all btree nodes along the search path
2670  * @root:       The root node of the tree
2671  * @key:        The key we are looking for
2672  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2673  *              deletions it's -1. 0 for plain searches
2674  * @cow:        boolean should CoW operations be performed. Must always be 1
2675  *              when modifying the tree.
2676  *
2677  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2678  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2679  *
2680  * If @key is found, 0 is returned and you can find the item in the leaf level
2681  * of the path (level 0)
2682  *
2683  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2684  * points to the slot where it should be inserted
2685  *
2686  * If an error is encountered while searching the tree a negative error number
2687  * is returned
2688  */
2689 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2690                       const struct btrfs_key *key, struct btrfs_path *p,
2691                       int ins_len, int cow)
2692 {
2693         struct btrfs_fs_info *fs_info = root->fs_info;
2694         struct extent_buffer *b;
2695         int slot;
2696         int ret;
2697         int err;
2698         int level;
2699         int lowest_unlock = 1;
2700         /* everything at write_lock_level or lower must be write locked */
2701         int write_lock_level = 0;
2702         u8 lowest_level = 0;
2703         int min_write_lock_level;
2704         int prev_cmp;
2705
2706         lowest_level = p->lowest_level;
2707         WARN_ON(lowest_level && ins_len > 0);
2708         WARN_ON(p->nodes[0] != NULL);
2709         BUG_ON(!cow && ins_len);
2710
2711         if (ins_len < 0) {
2712                 lowest_unlock = 2;
2713
2714                 /* when we are removing items, we might have to go up to level
2715                  * two as we update tree pointers  Make sure we keep write
2716                  * for those levels as well
2717                  */
2718                 write_lock_level = 2;
2719         } else if (ins_len > 0) {
2720                 /*
2721                  * for inserting items, make sure we have a write lock on
2722                  * level 1 so we can update keys
2723                  */
2724                 write_lock_level = 1;
2725         }
2726
2727         if (!cow)
2728                 write_lock_level = -1;
2729
2730         if (cow && (p->keep_locks || p->lowest_level))
2731                 write_lock_level = BTRFS_MAX_LEVEL;
2732
2733         min_write_lock_level = write_lock_level;
2734
2735 again:
2736         prev_cmp = -1;
2737         b = btrfs_search_slot_get_root(root, p, write_lock_level);
2738         if (IS_ERR(b)) {
2739                 ret = PTR_ERR(b);
2740                 goto done;
2741         }
2742
2743         while (b) {
2744                 level = btrfs_header_level(b);
2745
2746                 /*
2747                  * setup the path here so we can release it under lock
2748                  * contention with the cow code
2749                  */
2750                 if (cow) {
2751                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2752
2753                         /*
2754                          * if we don't really need to cow this block
2755                          * then we don't want to set the path blocking,
2756                          * so we test it here
2757                          */
2758                         if (!should_cow_block(trans, root, b)) {
2759                                 trans->dirty = true;
2760                                 goto cow_done;
2761                         }
2762
2763                         /*
2764                          * must have write locks on this node and the
2765                          * parent
2766                          */
2767                         if (level > write_lock_level ||
2768                             (level + 1 > write_lock_level &&
2769                             level + 1 < BTRFS_MAX_LEVEL &&
2770                             p->nodes[level + 1])) {
2771                                 write_lock_level = level + 1;
2772                                 btrfs_release_path(p);
2773                                 goto again;
2774                         }
2775
2776                         btrfs_set_path_blocking(p);
2777                         if (last_level)
2778                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2779                                                       &b);
2780                         else
2781                                 err = btrfs_cow_block(trans, root, b,
2782                                                       p->nodes[level + 1],
2783                                                       p->slots[level + 1], &b);
2784                         if (err) {
2785                                 ret = err;
2786                                 goto done;
2787                         }
2788                 }
2789 cow_done:
2790                 p->nodes[level] = b;
2791                 /*
2792                  * Leave path with blocking locks to avoid massive
2793                  * lock context switch, this is made on purpose.
2794                  */
2795
2796                 /*
2797                  * we have a lock on b and as long as we aren't changing
2798                  * the tree, there is no way to for the items in b to change.
2799                  * It is safe to drop the lock on our parent before we
2800                  * go through the expensive btree search on b.
2801                  *
2802                  * If we're inserting or deleting (ins_len != 0), then we might
2803                  * be changing slot zero, which may require changing the parent.
2804                  * So, we can't drop the lock until after we know which slot
2805                  * we're operating on.
2806                  */
2807                 if (!ins_len && !p->keep_locks) {
2808                         int u = level + 1;
2809
2810                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2811                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2812                                 p->locks[u] = 0;
2813                         }
2814                 }
2815
2816                 ret = key_search(b, key, level, &prev_cmp, &slot);
2817                 if (ret < 0)
2818                         goto done;
2819
2820                 if (level != 0) {
2821                         int dec = 0;
2822                         if (ret && slot > 0) {
2823                                 dec = 1;
2824                                 slot -= 1;
2825                         }
2826                         p->slots[level] = slot;
2827                         err = setup_nodes_for_search(trans, root, p, b, level,
2828                                              ins_len, &write_lock_level);
2829                         if (err == -EAGAIN)
2830                                 goto again;
2831                         if (err) {
2832                                 ret = err;
2833                                 goto done;
2834                         }
2835                         b = p->nodes[level];
2836                         slot = p->slots[level];
2837
2838                         /*
2839                          * slot 0 is special, if we change the key
2840                          * we have to update the parent pointer
2841                          * which means we must have a write lock
2842                          * on the parent
2843                          */
2844                         if (slot == 0 && ins_len &&
2845                             write_lock_level < level + 1) {
2846                                 write_lock_level = level + 1;
2847                                 btrfs_release_path(p);
2848                                 goto again;
2849                         }
2850
2851                         unlock_up(p, level, lowest_unlock,
2852                                   min_write_lock_level, &write_lock_level);
2853
2854                         if (level == lowest_level) {
2855                                 if (dec)
2856                                         p->slots[level]++;
2857                                 goto done;
2858                         }
2859
2860                         err = read_block_for_search(root, p, &b, level,
2861                                                     slot, key);
2862                         if (err == -EAGAIN)
2863                                 goto again;
2864                         if (err) {
2865                                 ret = err;
2866                                 goto done;
2867                         }
2868
2869                         if (!p->skip_locking) {
2870                                 level = btrfs_header_level(b);
2871                                 if (level <= write_lock_level) {
2872                                         err = btrfs_try_tree_write_lock(b);
2873                                         if (!err) {
2874                                                 btrfs_set_path_blocking(p);
2875                                                 btrfs_tree_lock(b);
2876                                         }
2877                                         p->locks[level] = BTRFS_WRITE_LOCK;
2878                                 } else {
2879                                         err = btrfs_tree_read_lock_atomic(b);
2880                                         if (!err) {
2881                                                 btrfs_set_path_blocking(p);
2882                                                 btrfs_tree_read_lock(b);
2883                                         }
2884                                         p->locks[level] = BTRFS_READ_LOCK;
2885                                 }
2886                                 p->nodes[level] = b;
2887                         }
2888                 } else {
2889                         p->slots[level] = slot;
2890                         if (ins_len > 0 &&
2891                             btrfs_leaf_free_space(fs_info, b) < ins_len) {
2892                                 if (write_lock_level < 1) {
2893                                         write_lock_level = 1;
2894                                         btrfs_release_path(p);
2895                                         goto again;
2896                                 }
2897
2898                                 btrfs_set_path_blocking(p);
2899                                 err = split_leaf(trans, root, key,
2900                                                  p, ins_len, ret == 0);
2901
2902                                 BUG_ON(err > 0);
2903                                 if (err) {
2904                                         ret = err;
2905                                         goto done;
2906                                 }
2907                         }
2908                         if (!p->search_for_split)
2909                                 unlock_up(p, level, lowest_unlock,
2910                                           min_write_lock_level, NULL);
2911                         goto done;
2912                 }
2913         }
2914         ret = 1;
2915 done:
2916         /*
2917          * we don't really know what they plan on doing with the path
2918          * from here on, so for now just mark it as blocking
2919          */
2920         if (!p->leave_spinning)
2921                 btrfs_set_path_blocking(p);
2922         if (ret < 0 && !p->skip_release_on_error)
2923                 btrfs_release_path(p);
2924         return ret;
2925 }
2926
2927 /*
2928  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2929  * current state of the tree together with the operations recorded in the tree
2930  * modification log to search for the key in a previous version of this tree, as
2931  * denoted by the time_seq parameter.
2932  *
2933  * Naturally, there is no support for insert, delete or cow operations.
2934  *
2935  * The resulting path and return value will be set up as if we called
2936  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2937  */
2938 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2939                           struct btrfs_path *p, u64 time_seq)
2940 {
2941         struct btrfs_fs_info *fs_info = root->fs_info;
2942         struct extent_buffer *b;
2943         int slot;
2944         int ret;
2945         int err;
2946         int level;
2947         int lowest_unlock = 1;
2948         u8 lowest_level = 0;
2949         int prev_cmp = -1;
2950
2951         lowest_level = p->lowest_level;
2952         WARN_ON(p->nodes[0] != NULL);
2953
2954         if (p->search_commit_root) {
2955                 BUG_ON(time_seq);
2956                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2957         }
2958
2959 again:
2960         b = get_old_root(root, time_seq);
2961         if (!b) {
2962                 ret = -EIO;
2963                 goto done;
2964         }
2965         level = btrfs_header_level(b);
2966         p->locks[level] = BTRFS_READ_LOCK;
2967
2968         while (b) {
2969                 level = btrfs_header_level(b);
2970                 p->nodes[level] = b;
2971
2972                 /*
2973                  * we have a lock on b and as long as we aren't changing
2974                  * the tree, there is no way to for the items in b to change.
2975                  * It is safe to drop the lock on our parent before we
2976                  * go through the expensive btree search on b.
2977                  */
2978                 btrfs_unlock_up_safe(p, level + 1);
2979
2980                 /*
2981                  * Since we can unwind ebs we want to do a real search every
2982                  * time.
2983                  */
2984                 prev_cmp = -1;
2985                 ret = key_search(b, key, level, &prev_cmp, &slot);
2986
2987                 if (level != 0) {
2988                         int dec = 0;
2989                         if (ret && slot > 0) {
2990                                 dec = 1;
2991                                 slot -= 1;
2992                         }
2993                         p->slots[level] = slot;
2994                         unlock_up(p, level, lowest_unlock, 0, NULL);
2995
2996                         if (level == lowest_level) {
2997                                 if (dec)
2998                                         p->slots[level]++;
2999                                 goto done;
3000                         }
3001
3002                         err = read_block_for_search(root, p, &b, level,
3003                                                     slot, key);
3004                         if (err == -EAGAIN)
3005                                 goto again;
3006                         if (err) {
3007                                 ret = err;
3008                                 goto done;
3009                         }
3010
3011                         level = btrfs_header_level(b);
3012                         err = btrfs_tree_read_lock_atomic(b);
3013                         if (!err) {
3014                                 btrfs_set_path_blocking(p);
3015                                 btrfs_tree_read_lock(b);
3016                         }
3017                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3018                         if (!b) {
3019                                 ret = -ENOMEM;
3020                                 goto done;
3021                         }
3022                         p->locks[level] = BTRFS_READ_LOCK;
3023                         p->nodes[level] = b;
3024                 } else {
3025                         p->slots[level] = slot;
3026                         unlock_up(p, level, lowest_unlock, 0, NULL);
3027                         goto done;
3028                 }
3029         }
3030         ret = 1;
3031 done:
3032         if (!p->leave_spinning)
3033                 btrfs_set_path_blocking(p);
3034         if (ret < 0)
3035                 btrfs_release_path(p);
3036
3037         return ret;
3038 }
3039
3040 /*
3041  * helper to use instead of search slot if no exact match is needed but
3042  * instead the next or previous item should be returned.
3043  * When find_higher is true, the next higher item is returned, the next lower
3044  * otherwise.
3045  * When return_any and find_higher are both true, and no higher item is found,
3046  * return the next lower instead.
3047  * When return_any is true and find_higher is false, and no lower item is found,
3048  * return the next higher instead.
3049  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3050  * < 0 on error
3051  */
3052 int btrfs_search_slot_for_read(struct btrfs_root *root,
3053                                const struct btrfs_key *key,
3054                                struct btrfs_path *p, int find_higher,
3055                                int return_any)
3056 {
3057         int ret;
3058         struct extent_buffer *leaf;
3059
3060 again:
3061         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3062         if (ret <= 0)
3063                 return ret;
3064         /*
3065          * a return value of 1 means the path is at the position where the
3066          * item should be inserted. Normally this is the next bigger item,
3067          * but in case the previous item is the last in a leaf, path points
3068          * to the first free slot in the previous leaf, i.e. at an invalid
3069          * item.
3070          */
3071         leaf = p->nodes[0];
3072
3073         if (find_higher) {
3074                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3075                         ret = btrfs_next_leaf(root, p);
3076                         if (ret <= 0)
3077                                 return ret;
3078                         if (!return_any)
3079                                 return 1;
3080                         /*
3081                          * no higher item found, return the next
3082                          * lower instead
3083                          */
3084                         return_any = 0;
3085                         find_higher = 0;
3086                         btrfs_release_path(p);
3087                         goto again;
3088                 }
3089         } else {
3090                 if (p->slots[0] == 0) {
3091                         ret = btrfs_prev_leaf(root, p);
3092                         if (ret < 0)
3093                                 return ret;
3094                         if (!ret) {
3095                                 leaf = p->nodes[0];
3096                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3097                                         p->slots[0]--;
3098                                 return 0;
3099                         }
3100                         if (!return_any)
3101                                 return 1;
3102                         /*
3103                          * no lower item found, return the next
3104                          * higher instead
3105                          */
3106                         return_any = 0;
3107                         find_higher = 1;
3108                         btrfs_release_path(p);
3109                         goto again;
3110                 } else {
3111                         --p->slots[0];
3112                 }
3113         }
3114         return 0;
3115 }
3116
3117 /*
3118  * adjust the pointers going up the tree, starting at level
3119  * making sure the right key of each node is points to 'key'.
3120  * This is used after shifting pointers to the left, so it stops
3121  * fixing up pointers when a given leaf/node is not in slot 0 of the
3122  * higher levels
3123  *
3124  */
3125 static void fixup_low_keys(struct btrfs_path *path,
3126                            struct btrfs_disk_key *key, int level)
3127 {
3128         int i;
3129         struct extent_buffer *t;
3130         int ret;
3131
3132         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3133                 int tslot = path->slots[i];
3134
3135                 if (!path->nodes[i])
3136                         break;
3137                 t = path->nodes[i];
3138                 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3139                                 GFP_ATOMIC);
3140                 BUG_ON(ret < 0);
3141                 btrfs_set_node_key(t, key, tslot);
3142                 btrfs_mark_buffer_dirty(path->nodes[i]);
3143                 if (tslot != 0)
3144                         break;
3145         }
3146 }
3147
3148 /*
3149  * update item key.
3150  *
3151  * This function isn't completely safe. It's the caller's responsibility
3152  * that the new key won't break the order
3153  */
3154 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3155                              struct btrfs_path *path,
3156                              const struct btrfs_key *new_key)
3157 {
3158         struct btrfs_disk_key disk_key;
3159         struct extent_buffer *eb;
3160         int slot;
3161
3162         eb = path->nodes[0];
3163         slot = path->slots[0];
3164         if (slot > 0) {
3165                 btrfs_item_key(eb, &disk_key, slot - 1);
3166                 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
3167         }
3168         if (slot < btrfs_header_nritems(eb) - 1) {
3169                 btrfs_item_key(eb, &disk_key, slot + 1);
3170                 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
3171         }
3172
3173         btrfs_cpu_key_to_disk(&disk_key, new_key);
3174         btrfs_set_item_key(eb, &disk_key, slot);
3175         btrfs_mark_buffer_dirty(eb);
3176         if (slot == 0)
3177                 fixup_low_keys(path, &disk_key, 1);
3178 }
3179
3180 /*
3181  * try to push data from one node into the next node left in the
3182  * tree.
3183  *
3184  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3185  * error, and > 0 if there was no room in the left hand block.
3186  */
3187 static int push_node_left(struct btrfs_trans_handle *trans,
3188                           struct btrfs_fs_info *fs_info,
3189                           struct extent_buffer *dst,
3190                           struct extent_buffer *src, int empty)
3191 {
3192         int push_items = 0;
3193         int src_nritems;
3194         int dst_nritems;
3195         int ret = 0;
3196
3197         src_nritems = btrfs_header_nritems(src);
3198         dst_nritems = btrfs_header_nritems(dst);
3199         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3200         WARN_ON(btrfs_header_generation(src) != trans->transid);
3201         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3202
3203         if (!empty && src_nritems <= 8)
3204                 return 1;
3205
3206         if (push_items <= 0)
3207                 return 1;
3208
3209         if (empty) {
3210                 push_items = min(src_nritems, push_items);
3211                 if (push_items < src_nritems) {
3212                         /* leave at least 8 pointers in the node if
3213                          * we aren't going to empty it
3214                          */
3215                         if (src_nritems - push_items < 8) {
3216                                 if (push_items <= 8)
3217                                         return 1;
3218                                 push_items -= 8;
3219                         }
3220                 }
3221         } else
3222                 push_items = min(src_nritems - 8, push_items);
3223
3224         ret = tree_mod_log_eb_copy(fs_info, dst, src, dst_nritems, 0,
3225                                    push_items);
3226         if (ret) {
3227                 btrfs_abort_transaction(trans, ret);
3228                 return ret;
3229         }
3230         copy_extent_buffer(dst, src,
3231                            btrfs_node_key_ptr_offset(dst_nritems),
3232                            btrfs_node_key_ptr_offset(0),
3233                            push_items * sizeof(struct btrfs_key_ptr));
3234
3235         if (push_items < src_nritems) {
3236                 /*
3237                  * Don't call tree_mod_log_insert_move here, key removal was
3238                  * already fully logged by tree_mod_log_eb_copy above.
3239                  */
3240                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3241                                       btrfs_node_key_ptr_offset(push_items),
3242                                       (src_nritems - push_items) *
3243                                       sizeof(struct btrfs_key_ptr));
3244         }
3245         btrfs_set_header_nritems(src, src_nritems - push_items);
3246         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3247         btrfs_mark_buffer_dirty(src);
3248         btrfs_mark_buffer_dirty(dst);
3249
3250         return ret;
3251 }
3252
3253 /*
3254  * try to push data from one node into the next node right in the
3255  * tree.
3256  *
3257  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3258  * error, and > 0 if there was no room in the right hand block.
3259  *
3260  * this will  only push up to 1/2 the contents of the left node over
3261  */
3262 static int balance_node_right(struct btrfs_trans_handle *trans,
3263                               struct btrfs_fs_info *fs_info,
3264                               struct extent_buffer *dst,
3265                               struct extent_buffer *src)
3266 {
3267         int push_items = 0;
3268         int max_push;
3269         int src_nritems;
3270         int dst_nritems;
3271         int ret = 0;
3272
3273         WARN_ON(btrfs_header_generation(src) != trans->transid);
3274         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3275
3276         src_nritems = btrfs_header_nritems(src);
3277         dst_nritems = btrfs_header_nritems(dst);
3278         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3279         if (push_items <= 0)
3280                 return 1;
3281
3282         if (src_nritems < 4)
3283                 return 1;
3284
3285         max_push = src_nritems / 2 + 1;
3286         /* don't try to empty the node */
3287         if (max_push >= src_nritems)
3288                 return 1;
3289
3290         if (max_push < push_items)
3291                 push_items = max_push;
3292
3293         ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3294         BUG_ON(ret < 0);
3295         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3296                                       btrfs_node_key_ptr_offset(0),
3297                                       (dst_nritems) *
3298                                       sizeof(struct btrfs_key_ptr));
3299
3300         ret = tree_mod_log_eb_copy(fs_info, dst, src, 0,
3301                                    src_nritems - push_items, push_items);
3302         if (ret) {
3303                 btrfs_abort_transaction(trans, ret);
3304                 return ret;
3305         }
3306         copy_extent_buffer(dst, src,
3307                            btrfs_node_key_ptr_offset(0),
3308                            btrfs_node_key_ptr_offset(src_nritems - push_items),
3309                            push_items * sizeof(struct btrfs_key_ptr));
3310
3311         btrfs_set_header_nritems(src, src_nritems - push_items);
3312         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3313
3314         btrfs_mark_buffer_dirty(src);
3315         btrfs_mark_buffer_dirty(dst);
3316
3317         return ret;
3318 }
3319
3320 /*
3321  * helper function to insert a new root level in the tree.
3322  * A new node is allocated, and a single item is inserted to
3323  * point to the existing root
3324  *
3325  * returns zero on success or < 0 on failure.
3326  */
3327 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3328                            struct btrfs_root *root,
3329                            struct btrfs_path *path, int level)
3330 {
3331         struct btrfs_fs_info *fs_info = root->fs_info;
3332         u64 lower_gen;
3333         struct extent_buffer *lower;
3334         struct extent_buffer *c;
3335         struct extent_buffer *old;
3336         struct btrfs_disk_key lower_key;
3337         int ret;
3338
3339         BUG_ON(path->nodes[level]);
3340         BUG_ON(path->nodes[level-1] != root->node);
3341
3342         lower = path->nodes[level-1];
3343         if (level == 1)
3344                 btrfs_item_key(lower, &lower_key, 0);
3345         else
3346                 btrfs_node_key(lower, &lower_key, 0);
3347
3348         c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3349                                    &lower_key, level, root->node->start, 0);
3350         if (IS_ERR(c))
3351                 return PTR_ERR(c);
3352
3353         root_add_used(root, fs_info->nodesize);
3354
3355         btrfs_set_header_nritems(c, 1);
3356         btrfs_set_node_key(c, &lower_key, 0);
3357         btrfs_set_node_blockptr(c, 0, lower->start);
3358         lower_gen = btrfs_header_generation(lower);
3359         WARN_ON(lower_gen != trans->transid);
3360
3361         btrfs_set_node_ptr_generation(c, 0, lower_gen);
3362
3363         btrfs_mark_buffer_dirty(c);
3364
3365         old = root->node;
3366         ret = tree_mod_log_insert_root(root->node, c, 0);
3367         BUG_ON(ret < 0);
3368         rcu_assign_pointer(root->node, c);
3369
3370         /* the super has an extra ref to root->node */
3371         free_extent_buffer(old);
3372
3373         add_root_to_dirty_list(root);
3374         extent_buffer_get(c);
3375         path->nodes[level] = c;
3376         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3377         path->slots[level] = 0;
3378         return 0;
3379 }
3380
3381 /*
3382  * worker function to insert a single pointer in a node.
3383  * the node should have enough room for the pointer already
3384  *
3385  * slot and level indicate where you want the key to go, and
3386  * blocknr is the block the key points to.
3387  */
3388 static void insert_ptr(struct btrfs_trans_handle *trans,
3389                        struct btrfs_fs_info *fs_info, struct btrfs_path *path,
3390                        struct btrfs_disk_key *key, u64 bytenr,
3391                        int slot, int level)
3392 {
3393         struct extent_buffer *lower;
3394         int nritems;
3395         int ret;
3396
3397         BUG_ON(!path->nodes[level]);
3398         btrfs_assert_tree_locked(path->nodes[level]);
3399         lower = path->nodes[level];
3400         nritems = btrfs_header_nritems(lower);
3401         BUG_ON(slot > nritems);
3402         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(fs_info));
3403         if (slot != nritems) {
3404                 if (level) {
3405                         ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3406                                         nritems - slot);
3407                         BUG_ON(ret < 0);
3408                 }
3409                 memmove_extent_buffer(lower,
3410                               btrfs_node_key_ptr_offset(slot + 1),
3411                               btrfs_node_key_ptr_offset(slot),
3412                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
3413         }
3414         if (level) {
3415                 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3416                                 GFP_NOFS);
3417                 BUG_ON(ret < 0);
3418         }
3419         btrfs_set_node_key(lower, key, slot);
3420         btrfs_set_node_blockptr(lower, slot, bytenr);
3421         WARN_ON(trans->transid == 0);
3422         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3423         btrfs_set_header_nritems(lower, nritems + 1);
3424         btrfs_mark_buffer_dirty(lower);
3425 }
3426
3427 /*
3428  * split the node at the specified level in path in two.
3429  * The path is corrected to point to the appropriate node after the split
3430  *
3431  * Before splitting this tries to make some room in the node by pushing
3432  * left and right, if either one works, it returns right away.
3433  *
3434  * returns 0 on success and < 0 on failure
3435  */
3436 static noinline int split_node(struct btrfs_trans_handle *trans,
3437                                struct btrfs_root *root,
3438                                struct btrfs_path *path, int level)
3439 {
3440         struct btrfs_fs_info *fs_info = root->fs_info;
3441         struct extent_buffer *c;
3442         struct extent_buffer *split;
3443         struct btrfs_disk_key disk_key;
3444         int mid;
3445         int ret;
3446         u32 c_nritems;
3447
3448         c = path->nodes[level];
3449         WARN_ON(btrfs_header_generation(c) != trans->transid);
3450         if (c == root->node) {
3451                 /*
3452                  * trying to split the root, lets make a new one
3453                  *
3454                  * tree mod log: We don't log_removal old root in
3455                  * insert_new_root, because that root buffer will be kept as a
3456                  * normal node. We are going to log removal of half of the
3457                  * elements below with tree_mod_log_eb_copy. We're holding a
3458                  * tree lock on the buffer, which is why we cannot race with
3459                  * other tree_mod_log users.
3460                  */
3461                 ret = insert_new_root(trans, root, path, level + 1);
3462                 if (ret)
3463                         return ret;
3464         } else {
3465                 ret = push_nodes_for_insert(trans, root, path, level);
3466                 c = path->nodes[level];
3467                 if (!ret && btrfs_header_nritems(c) <
3468                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3469                         return 0;
3470                 if (ret < 0)
3471                         return ret;
3472         }
3473
3474         c_nritems = btrfs_header_nritems(c);
3475         mid = (c_nritems + 1) / 2;
3476         btrfs_node_key(c, &disk_key, mid);
3477
3478         split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3479                         &disk_key, level, c->start, 0);
3480         if (IS_ERR(split))
3481                 return PTR_ERR(split);
3482
3483         root_add_used(root, fs_info->nodesize);
3484         ASSERT(btrfs_header_level(c) == level);
3485
3486         ret = tree_mod_log_eb_copy(fs_info, split, c, 0, mid, c_nritems - mid);
3487         if (ret) {
3488                 btrfs_abort_transaction(trans, ret);
3489                 return ret;
3490         }
3491         copy_extent_buffer(split, c,
3492                            btrfs_node_key_ptr_offset(0),
3493                            btrfs_node_key_ptr_offset(mid),
3494                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3495         btrfs_set_header_nritems(split, c_nritems - mid);
3496         btrfs_set_header_nritems(c, mid);
3497         ret = 0;
3498
3499         btrfs_mark_buffer_dirty(c);
3500         btrfs_mark_buffer_dirty(split);
3501
3502         insert_ptr(trans, fs_info, path, &disk_key, split->start,
3503                    path->slots[level + 1] + 1, level + 1);
3504
3505         if (path->slots[level] >= mid) {
3506                 path->slots[level] -= mid;
3507                 btrfs_tree_unlock(c);
3508                 free_extent_buffer(c);
3509                 path->nodes[level] = split;
3510                 path->slots[level + 1] += 1;
3511         } else {
3512                 btrfs_tree_unlock(split);
3513                 free_extent_buffer(split);
3514         }
3515         return ret;
3516 }
3517
3518 /*
3519  * how many bytes are required to store the items in a leaf.  start
3520  * and nr indicate which items in the leaf to check.  This totals up the
3521  * space used both by the item structs and the item data
3522  */
3523 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3524 {
3525         struct btrfs_item *start_item;
3526         struct btrfs_item *end_item;
3527         struct btrfs_map_token token;
3528         int data_len;
3529         int nritems = btrfs_header_nritems(l);
3530         int end = min(nritems, start + nr) - 1;
3531
3532         if (!nr)
3533                 return 0;
3534         btrfs_init_map_token(&token);
3535         start_item = btrfs_item_nr(start);
3536         end_item = btrfs_item_nr(end);
3537         data_len = btrfs_token_item_offset(l, start_item, &token) +
3538                 btrfs_token_item_size(l, start_item, &token);
3539         data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3540         data_len += sizeof(struct btrfs_item) * nr;
3541         WARN_ON(data_len < 0);
3542         return data_len;
3543 }
3544
3545 /*
3546  * The space between the end of the leaf items and
3547  * the start of the leaf data.  IOW, how much room
3548  * the leaf has left for both items and data
3549  */
3550 noinline int btrfs_leaf_free_space(struct btrfs_fs_info *fs_info,
3551                                    struct extent_buffer *leaf)
3552 {
3553         int nritems = btrfs_header_nritems(leaf);
3554         int ret;
3555
3556         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3557         if (ret < 0) {
3558                 btrfs_crit(fs_info,
3559                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3560                            ret,
3561                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3562                            leaf_space_used(leaf, 0, nritems), nritems);
3563         }
3564         return ret;
3565 }
3566
3567 /*
3568  * min slot controls the lowest index we're willing to push to the
3569  * right.  We'll push up to and including min_slot, but no lower
3570  */
3571 static noinline int __push_leaf_right(struct btrfs_fs_info *fs_info,
3572                                       struct btrfs_path *path,
3573                                       int data_size, int empty,
3574                                       struct extent_buffer *right,
3575                                       int free_space, u32 left_nritems,
3576                                       u32 min_slot)
3577 {
3578         struct extent_buffer *left = path->nodes[0];
3579         struct extent_buffer *upper = path->nodes[1];
3580         struct btrfs_map_token token;
3581         struct btrfs_disk_key disk_key;
3582         int slot;
3583         u32 i;
3584         int push_space = 0;
3585         int push_items = 0;
3586         struct btrfs_item *item;
3587         u32 nr;
3588         u32 right_nritems;
3589         u32 data_end;
3590         u32 this_item_size;
3591
3592         btrfs_init_map_token(&token);
3593
3594         if (empty)
3595                 nr = 0;
3596         else
3597                 nr = max_t(u32, 1, min_slot);
3598
3599         if (path->slots[0] >= left_nritems)
3600                 push_space += data_size;
3601
3602         slot = path->slots[1];
3603         i = left_nritems - 1;
3604         while (i >= nr) {
3605                 item = btrfs_item_nr(i);
3606
3607                 if (!empty && push_items > 0) {
3608                         if (path->slots[0] > i)
3609                                 break;
3610                         if (path->slots[0] == i) {
3611                                 int space = btrfs_leaf_free_space(fs_info, left);
3612                                 if (space + push_space * 2 > free_space)
3613                                         break;
3614                         }
3615                 }
3616
3617                 if (path->slots[0] == i)
3618                         push_space += data_size;
3619
3620                 this_item_size = btrfs_item_size(left, item);
3621                 if (this_item_size + sizeof(*item) + push_space > free_space)
3622                         break;
3623
3624                 push_items++;
3625                 push_space += this_item_size + sizeof(*item);
3626                 if (i == 0)
3627                         break;
3628                 i--;
3629         }
3630
3631         if (push_items == 0)
3632                 goto out_unlock;
3633
3634         WARN_ON(!empty && push_items == left_nritems);
3635