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