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