Loading...
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
18static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
19 *root, struct btrfs_path *path, int level);
20static 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);
23static int push_node_left(struct btrfs_trans_handle *trans,
24 struct extent_buffer *dst,
25 struct extent_buffer *src, int empty);
26static int balance_node_right(struct btrfs_trans_handle *trans,
27 struct extent_buffer *dst_buf,
28 struct extent_buffer *src_buf);
29static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
30 int level, int slot);
31
32static const struct btrfs_csums {
33 u16 size;
34 const char *name;
35} btrfs_csums[] = {
36 [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
37};
38
39int btrfs_super_csum_size(const struct btrfs_super_block *s)
40{
41 u16 t = btrfs_super_csum_type(s);
42 /*
43 * csum type is validated at mount time
44 */
45 return btrfs_csums[t].size;
46}
47
48const char *btrfs_super_csum_name(u16 csum_type)
49{
50 /* csum type is validated at mount time */
51 return btrfs_csums[csum_type].name;
52}
53
54struct btrfs_path *btrfs_alloc_path(void)
55{
56 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
57}
58
59/*
60 * set all locked nodes in the path to blocking locks. This should
61 * be done before scheduling
62 */
63noinline void btrfs_set_path_blocking(struct btrfs_path *p)
64{
65 int i;
66 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
67 if (!p->nodes[i] || !p->locks[i])
68 continue;
69 /*
70 * If we currently have a spinning reader or writer lock this
71 * will bump the count of blocking holders and drop the
72 * spinlock.
73 */
74 if (p->locks[i] == BTRFS_READ_LOCK) {
75 btrfs_set_lock_blocking_read(p->nodes[i]);
76 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
77 } else if (p->locks[i] == BTRFS_WRITE_LOCK) {
78 btrfs_set_lock_blocking_write(p->nodes[i]);
79 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
80 }
81 }
82}
83
84/* this also releases the path */
85void btrfs_free_path(struct btrfs_path *p)
86{
87 if (!p)
88 return;
89 btrfs_release_path(p);
90 kmem_cache_free(btrfs_path_cachep, p);
91}
92
93/*
94 * path release drops references on the extent buffers in the path
95 * and it drops any locks held by this path
96 *
97 * It is safe to call this on paths that no locks or extent buffers held.
98 */
99noinline void btrfs_release_path(struct btrfs_path *p)
100{
101 int i;
102
103 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
104 p->slots[i] = 0;
105 if (!p->nodes[i])
106 continue;
107 if (p->locks[i]) {
108 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
109 p->locks[i] = 0;
110 }
111 free_extent_buffer(p->nodes[i]);
112 p->nodes[i] = NULL;
113 }
114}
115
116/*
117 * safely gets a reference on the root node of a tree. A lock
118 * is not taken, so a concurrent writer may put a different node
119 * at the root of the tree. See btrfs_lock_root_node for the
120 * looping required.
121 *
122 * The extent buffer returned by this has a reference taken, so
123 * it won't disappear. It may stop being the root of the tree
124 * at any time because there are no locks held.
125 */
126struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
127{
128 struct extent_buffer *eb;
129
130 while (1) {
131 rcu_read_lock();
132 eb = rcu_dereference(root->node);
133
134 /*
135 * RCU really hurts here, we could free up the root node because
136 * it was COWed but we may not get the new root node yet so do
137 * the inc_not_zero dance and if it doesn't work then
138 * synchronize_rcu and try again.
139 */
140 if (atomic_inc_not_zero(&eb->refs)) {
141 rcu_read_unlock();
142 break;
143 }
144 rcu_read_unlock();
145 synchronize_rcu();
146 }
147 return eb;
148}
149
150/* loop around taking references on and locking the root node of the
151 * tree until you end up with a lock on the root. A locked buffer
152 * is returned, with a reference held.
153 */
154struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
155{
156 struct extent_buffer *eb;
157
158 while (1) {
159 eb = btrfs_root_node(root);
160 btrfs_tree_lock(eb);
161 if (eb == root->node)
162 break;
163 btrfs_tree_unlock(eb);
164 free_extent_buffer(eb);
165 }
166 return eb;
167}
168
169/* loop around taking references on and locking the root node of the
170 * tree until you end up with a lock on the root. A locked buffer
171 * is returned, with a reference held.
172 */
173struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
174{
175 struct extent_buffer *eb;
176
177 while (1) {
178 eb = btrfs_root_node(root);
179 btrfs_tree_read_lock(eb);
180 if (eb == root->node)
181 break;
182 btrfs_tree_read_unlock(eb);
183 free_extent_buffer(eb);
184 }
185 return eb;
186}
187
188/* cowonly root (everything not a reference counted cow subvolume), just get
189 * put onto a simple dirty list. transaction.c walks this to make sure they
190 * get properly updated on disk.
191 */
192static void add_root_to_dirty_list(struct btrfs_root *root)
193{
194 struct btrfs_fs_info *fs_info = root->fs_info;
195
196 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
197 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
198 return;
199
200 spin_lock(&fs_info->trans_lock);
201 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
202 /* Want the extent tree to be the last on the list */
203 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
204 list_move_tail(&root->dirty_list,
205 &fs_info->dirty_cowonly_roots);
206 else
207 list_move(&root->dirty_list,
208 &fs_info->dirty_cowonly_roots);
209 }
210 spin_unlock(&fs_info->trans_lock);
211}
212
213/*
214 * used by snapshot creation to make a copy of a root for a tree with
215 * a given objectid. The buffer with the new root node is returned in
216 * cow_ret, and this func returns zero on success or a negative error code.
217 */
218int btrfs_copy_root(struct btrfs_trans_handle *trans,
219 struct btrfs_root *root,
220 struct extent_buffer *buf,
221 struct extent_buffer **cow_ret, u64 new_root_objectid)
222{
223 struct btrfs_fs_info *fs_info = root->fs_info;
224 struct extent_buffer *cow;
225 int ret = 0;
226 int level;
227 struct btrfs_disk_key disk_key;
228
229 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
230 trans->transid != fs_info->running_transaction->transid);
231 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
232 trans->transid != root->last_trans);
233
234 level = btrfs_header_level(buf);
235 if (level == 0)
236 btrfs_item_key(buf, &disk_key, 0);
237 else
238 btrfs_node_key(buf, &disk_key, 0);
239
240 cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
241 &disk_key, level, buf->start, 0);
242 if (IS_ERR(cow))
243 return PTR_ERR(cow);
244
245 copy_extent_buffer_full(cow, buf);
246 btrfs_set_header_bytenr(cow, cow->start);
247 btrfs_set_header_generation(cow, trans->transid);
248 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
249 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
250 BTRFS_HEADER_FLAG_RELOC);
251 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
252 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
253 else
254 btrfs_set_header_owner(cow, new_root_objectid);
255
256 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
257
258 WARN_ON(btrfs_header_generation(buf) > trans->transid);
259 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
260 ret = btrfs_inc_ref(trans, root, cow, 1);
261 else
262 ret = btrfs_inc_ref(trans, root, cow, 0);
263
264 if (ret)
265 return ret;
266
267 btrfs_mark_buffer_dirty(cow);
268 *cow_ret = cow;
269 return 0;
270}
271
272enum mod_log_op {
273 MOD_LOG_KEY_REPLACE,
274 MOD_LOG_KEY_ADD,
275 MOD_LOG_KEY_REMOVE,
276 MOD_LOG_KEY_REMOVE_WHILE_FREEING,
277 MOD_LOG_KEY_REMOVE_WHILE_MOVING,
278 MOD_LOG_MOVE_KEYS,
279 MOD_LOG_ROOT_REPLACE,
280};
281
282struct tree_mod_root {
283 u64 logical;
284 u8 level;
285};
286
287struct tree_mod_elem {
288 struct rb_node node;
289 u64 logical;
290 u64 seq;
291 enum mod_log_op op;
292
293 /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
294 int slot;
295
296 /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
297 u64 generation;
298
299 /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
300 struct btrfs_disk_key key;
301 u64 blockptr;
302
303 /* this is used for op == MOD_LOG_MOVE_KEYS */
304 struct {
305 int dst_slot;
306 int nr_items;
307 } move;
308
309 /* this is used for op == MOD_LOG_ROOT_REPLACE */
310 struct tree_mod_root old_root;
311};
312
313/*
314 * Pull a new tree mod seq number for our operation.
315 */
316static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
317{
318 return atomic64_inc_return(&fs_info->tree_mod_seq);
319}
320
321/*
322 * This adds a new blocker to the tree mod log's blocker list if the @elem
323 * passed does not already have a sequence number set. So when a caller expects
324 * to record tree modifications, it should ensure to set elem->seq to zero
325 * before calling btrfs_get_tree_mod_seq.
326 * Returns a fresh, unused tree log modification sequence number, even if no new
327 * blocker was added.
328 */
329u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
330 struct seq_list *elem)
331{
332 write_lock(&fs_info->tree_mod_log_lock);
333 spin_lock(&fs_info->tree_mod_seq_lock);
334 if (!elem->seq) {
335 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
336 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
337 }
338 spin_unlock(&fs_info->tree_mod_seq_lock);
339 write_unlock(&fs_info->tree_mod_log_lock);
340
341 return elem->seq;
342}
343
344void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
345 struct seq_list *elem)
346{
347 struct rb_root *tm_root;
348 struct rb_node *node;
349 struct rb_node *next;
350 struct seq_list *cur_elem;
351 struct tree_mod_elem *tm;
352 u64 min_seq = (u64)-1;
353 u64 seq_putting = elem->seq;
354
355 if (!seq_putting)
356 return;
357
358 spin_lock(&fs_info->tree_mod_seq_lock);
359 list_del(&elem->list);
360 elem->seq = 0;
361
362 list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
363 if (cur_elem->seq < min_seq) {
364 if (seq_putting > cur_elem->seq) {
365 /*
366 * blocker with lower sequence number exists, we
367 * cannot remove anything from the log
368 */
369 spin_unlock(&fs_info->tree_mod_seq_lock);
370 return;
371 }
372 min_seq = cur_elem->seq;
373 }
374 }
375 spin_unlock(&fs_info->tree_mod_seq_lock);
376
377 /*
378 * anything that's lower than the lowest existing (read: blocked)
379 * sequence number can be removed from the tree.
380 */
381 write_lock(&fs_info->tree_mod_log_lock);
382 tm_root = &fs_info->tree_mod_log;
383 for (node = rb_first(tm_root); node; node = next) {
384 next = rb_next(node);
385 tm = rb_entry(node, struct tree_mod_elem, node);
386 if (tm->seq > min_seq)
387 continue;
388 rb_erase(node, tm_root);
389 kfree(tm);
390 }
391 write_unlock(&fs_info->tree_mod_log_lock);
392}
393
394/*
395 * key order of the log:
396 * node/leaf start address -> sequence
397 *
398 * The 'start address' is the logical address of the *new* root node
399 * for root replace operations, or the logical address of the affected
400 * block for all other operations.
401 */
402static noinline int
403__tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
404{
405 struct rb_root *tm_root;
406 struct rb_node **new;
407 struct rb_node *parent = NULL;
408 struct tree_mod_elem *cur;
409
410 lockdep_assert_held_write(&fs_info->tree_mod_log_lock);
411
412 tm->seq = btrfs_inc_tree_mod_seq(fs_info);
413
414 tm_root = &fs_info->tree_mod_log;
415 new = &tm_root->rb_node;
416 while (*new) {
417 cur = rb_entry(*new, struct tree_mod_elem, node);
418 parent = *new;
419 if (cur->logical < tm->logical)
420 new = &((*new)->rb_left);
421 else if (cur->logical > tm->logical)
422 new = &((*new)->rb_right);
423 else if (cur->seq < tm->seq)
424 new = &((*new)->rb_left);
425 else if (cur->seq > tm->seq)
426 new = &((*new)->rb_right);
427 else
428 return -EEXIST;
429 }
430
431 rb_link_node(&tm->node, parent, new);
432 rb_insert_color(&tm->node, tm_root);
433 return 0;
434}
435
436/*
437 * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
438 * returns zero with the tree_mod_log_lock acquired. The caller must hold
439 * this until all tree mod log insertions are recorded in the rb tree and then
440 * write unlock fs_info::tree_mod_log_lock.
441 */
442static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
443 struct extent_buffer *eb) {
444 smp_mb();
445 if (list_empty(&(fs_info)->tree_mod_seq_list))
446 return 1;
447 if (eb && btrfs_header_level(eb) == 0)
448 return 1;
449
450 write_lock(&fs_info->tree_mod_log_lock);
451 if (list_empty(&(fs_info)->tree_mod_seq_list)) {
452 write_unlock(&fs_info->tree_mod_log_lock);
453 return 1;
454 }
455
456 return 0;
457}
458
459/* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
460static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
461 struct extent_buffer *eb)
462{
463 smp_mb();
464 if (list_empty(&(fs_info)->tree_mod_seq_list))
465 return 0;
466 if (eb && btrfs_header_level(eb) == 0)
467 return 0;
468
469 return 1;
470}
471
472static struct tree_mod_elem *
473alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
474 enum mod_log_op op, gfp_t flags)
475{
476 struct tree_mod_elem *tm;
477
478 tm = kzalloc(sizeof(*tm), flags);
479 if (!tm)
480 return NULL;
481
482 tm->logical = eb->start;
483 if (op != MOD_LOG_KEY_ADD) {
484 btrfs_node_key(eb, &tm->key, slot);
485 tm->blockptr = btrfs_node_blockptr(eb, slot);
486 }
487 tm->op = op;
488 tm->slot = slot;
489 tm->generation = btrfs_node_ptr_generation(eb, slot);
490 RB_CLEAR_NODE(&tm->node);
491
492 return tm;
493}
494
495static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
496 enum mod_log_op op, gfp_t flags)
497{
498 struct tree_mod_elem *tm;
499 int ret;
500
501 if (!tree_mod_need_log(eb->fs_info, eb))
502 return 0;
503
504 tm = alloc_tree_mod_elem(eb, slot, op, flags);
505 if (!tm)
506 return -ENOMEM;
507
508 if (tree_mod_dont_log(eb->fs_info, eb)) {
509 kfree(tm);
510 return 0;
511 }
512
513 ret = __tree_mod_log_insert(eb->fs_info, tm);
514 write_unlock(&eb->fs_info->tree_mod_log_lock);
515 if (ret)
516 kfree(tm);
517
518 return ret;
519}
520
521static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
522 int dst_slot, int src_slot, int nr_items)
523{
524 struct tree_mod_elem *tm = NULL;
525 struct tree_mod_elem **tm_list = NULL;
526 int ret = 0;
527 int i;
528 int locked = 0;
529
530 if (!tree_mod_need_log(eb->fs_info, eb))
531 return 0;
532
533 tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
534 if (!tm_list)
535 return -ENOMEM;
536
537 tm = kzalloc(sizeof(*tm), GFP_NOFS);
538 if (!tm) {
539 ret = -ENOMEM;
540 goto free_tms;
541 }
542
543 tm->logical = eb->start;
544 tm->slot = src_slot;
545 tm->move.dst_slot = dst_slot;
546 tm->move.nr_items = nr_items;
547 tm->op = MOD_LOG_MOVE_KEYS;
548
549 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
550 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
551 MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
552 if (!tm_list[i]) {
553 ret = -ENOMEM;
554 goto free_tms;
555 }
556 }
557
558 if (tree_mod_dont_log(eb->fs_info, eb))
559 goto free_tms;
560 locked = 1;
561
562 /*
563 * When we override something during the move, we log these removals.
564 * This can only happen when we move towards the beginning of the
565 * buffer, i.e. dst_slot < src_slot.
566 */
567 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
568 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
569 if (ret)
570 goto free_tms;
571 }
572
573 ret = __tree_mod_log_insert(eb->fs_info, tm);
574 if (ret)
575 goto free_tms;
576 write_unlock(&eb->fs_info->tree_mod_log_lock);
577 kfree(tm_list);
578
579 return 0;
580free_tms:
581 for (i = 0; i < nr_items; i++) {
582 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
583 rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
584 kfree(tm_list[i]);
585 }
586 if (locked)
587 write_unlock(&eb->fs_info->tree_mod_log_lock);
588 kfree(tm_list);
589 kfree(tm);
590
591 return ret;
592}
593
594static inline int
595__tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
596 struct tree_mod_elem **tm_list,
597 int nritems)
598{
599 int i, j;
600 int ret;
601
602 for (i = nritems - 1; i >= 0; i--) {
603 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
604 if (ret) {
605 for (j = nritems - 1; j > i; j--)
606 rb_erase(&tm_list[j]->node,
607 &fs_info->tree_mod_log);
608 return ret;
609 }
610 }
611
612 return 0;
613}
614
615static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
616 struct extent_buffer *new_root, int log_removal)
617{
618 struct btrfs_fs_info *fs_info = old_root->fs_info;
619 struct tree_mod_elem *tm = NULL;
620 struct tree_mod_elem **tm_list = NULL;
621 int nritems = 0;
622 int ret = 0;
623 int i;
624
625 if (!tree_mod_need_log(fs_info, NULL))
626 return 0;
627
628 if (log_removal && btrfs_header_level(old_root) > 0) {
629 nritems = btrfs_header_nritems(old_root);
630 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
631 GFP_NOFS);
632 if (!tm_list) {
633 ret = -ENOMEM;
634 goto free_tms;
635 }
636 for (i = 0; i < nritems; i++) {
637 tm_list[i] = alloc_tree_mod_elem(old_root, i,
638 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
639 if (!tm_list[i]) {
640 ret = -ENOMEM;
641 goto free_tms;
642 }
643 }
644 }
645
646 tm = kzalloc(sizeof(*tm), GFP_NOFS);
647 if (!tm) {
648 ret = -ENOMEM;
649 goto free_tms;
650 }
651
652 tm->logical = new_root->start;
653 tm->old_root.logical = old_root->start;
654 tm->old_root.level = btrfs_header_level(old_root);
655 tm->generation = btrfs_header_generation(old_root);
656 tm->op = MOD_LOG_ROOT_REPLACE;
657
658 if (tree_mod_dont_log(fs_info, NULL))
659 goto free_tms;
660
661 if (tm_list)
662 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
663 if (!ret)
664 ret = __tree_mod_log_insert(fs_info, tm);
665
666 write_unlock(&fs_info->tree_mod_log_lock);
667 if (ret)
668 goto free_tms;
669 kfree(tm_list);
670
671 return ret;
672
673free_tms:
674 if (tm_list) {
675 for (i = 0; i < nritems; i++)
676 kfree(tm_list[i]);
677 kfree(tm_list);
678 }
679 kfree(tm);
680
681 return ret;
682}
683
684static struct tree_mod_elem *
685__tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
686 int smallest)
687{
688 struct rb_root *tm_root;
689 struct rb_node *node;
690 struct tree_mod_elem *cur = NULL;
691 struct tree_mod_elem *found = NULL;
692
693 read_lock(&fs_info->tree_mod_log_lock);
694 tm_root = &fs_info->tree_mod_log;
695 node = tm_root->rb_node;
696 while (node) {
697 cur = rb_entry(node, struct tree_mod_elem, node);
698 if (cur->logical < start) {
699 node = node->rb_left;
700 } else if (cur->logical > start) {
701 node = node->rb_right;
702 } else if (cur->seq < min_seq) {
703 node = node->rb_left;
704 } else if (!smallest) {
705 /* we want the node with the highest seq */
706 if (found)
707 BUG_ON(found->seq > cur->seq);
708 found = cur;
709 node = node->rb_left;
710 } else if (cur->seq > min_seq) {
711 /* we want the node with the smallest seq */
712 if (found)
713 BUG_ON(found->seq < cur->seq);
714 found = cur;
715 node = node->rb_right;
716 } else {
717 found = cur;
718 break;
719 }
720 }
721 read_unlock(&fs_info->tree_mod_log_lock);
722
723 return found;
724}
725
726/*
727 * this returns the element from the log with the smallest time sequence
728 * value that's in the log (the oldest log item). any element with a time
729 * sequence lower than min_seq will be ignored.
730 */
731static struct tree_mod_elem *
732tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
733 u64 min_seq)
734{
735 return __tree_mod_log_search(fs_info, start, min_seq, 1);
736}
737
738/*
739 * this returns the element from the log with the largest time sequence
740 * value that's in the log (the most recent log item). any element with
741 * a time sequence lower than min_seq will be ignored.
742 */
743static struct tree_mod_elem *
744tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
745{
746 return __tree_mod_log_search(fs_info, start, min_seq, 0);
747}
748
749static noinline int tree_mod_log_eb_copy(struct extent_buffer *dst,
750 struct extent_buffer *src, unsigned long dst_offset,
751 unsigned long src_offset, int nr_items)
752{
753 struct btrfs_fs_info *fs_info = dst->fs_info;
754 int ret = 0;
755 struct tree_mod_elem **tm_list = NULL;
756 struct tree_mod_elem **tm_list_add, **tm_list_rem;
757 int i;
758 int locked = 0;
759
760 if (!tree_mod_need_log(fs_info, NULL))
761 return 0;
762
763 if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
764 return 0;
765
766 tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
767 GFP_NOFS);
768 if (!tm_list)
769 return -ENOMEM;
770
771 tm_list_add = tm_list;
772 tm_list_rem = tm_list + nr_items;
773 for (i = 0; i < nr_items; i++) {
774 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
775 MOD_LOG_KEY_REMOVE, GFP_NOFS);
776 if (!tm_list_rem[i]) {
777 ret = -ENOMEM;
778 goto free_tms;
779 }
780
781 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
782 MOD_LOG_KEY_ADD, GFP_NOFS);
783 if (!tm_list_add[i]) {
784 ret = -ENOMEM;
785 goto free_tms;
786 }
787 }
788
789 if (tree_mod_dont_log(fs_info, NULL))
790 goto free_tms;
791 locked = 1;
792
793 for (i = 0; i < nr_items; i++) {
794 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
795 if (ret)
796 goto free_tms;
797 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
798 if (ret)
799 goto free_tms;
800 }
801
802 write_unlock(&fs_info->tree_mod_log_lock);
803 kfree(tm_list);
804
805 return 0;
806
807free_tms:
808 for (i = 0; i < nr_items * 2; i++) {
809 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
810 rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
811 kfree(tm_list[i]);
812 }
813 if (locked)
814 write_unlock(&fs_info->tree_mod_log_lock);
815 kfree(tm_list);
816
817 return ret;
818}
819
820static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
821{
822 struct tree_mod_elem **tm_list = NULL;
823 int nritems = 0;
824 int i;
825 int ret = 0;
826
827 if (btrfs_header_level(eb) == 0)
828 return 0;
829
830 if (!tree_mod_need_log(eb->fs_info, NULL))
831 return 0;
832
833 nritems = btrfs_header_nritems(eb);
834 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
835 if (!tm_list)
836 return -ENOMEM;
837
838 for (i = 0; i < nritems; i++) {
839 tm_list[i] = alloc_tree_mod_elem(eb, i,
840 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
841 if (!tm_list[i]) {
842 ret = -ENOMEM;
843 goto free_tms;
844 }
845 }
846
847 if (tree_mod_dont_log(eb->fs_info, eb))
848 goto free_tms;
849
850 ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
851 write_unlock(&eb->fs_info->tree_mod_log_lock);
852 if (ret)
853 goto free_tms;
854 kfree(tm_list);
855
856 return 0;
857
858free_tms:
859 for (i = 0; i < nritems; i++)
860 kfree(tm_list[i]);
861 kfree(tm_list);
862
863 return ret;
864}
865
866/*
867 * check if the tree block can be shared by multiple trees
868 */
869int btrfs_block_can_be_shared(struct btrfs_root *root,
870 struct extent_buffer *buf)
871{
872 /*
873 * Tree blocks not in reference counted trees and tree roots
874 * are never shared. If a block was allocated after the last
875 * snapshot and the block was not allocated by tree relocation,
876 * we know the block is not shared.
877 */
878 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
879 buf != root->node && buf != root->commit_root &&
880 (btrfs_header_generation(buf) <=
881 btrfs_root_last_snapshot(&root->root_item) ||
882 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
883 return 1;
884
885 return 0;
886}
887
888static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
889 struct btrfs_root *root,
890 struct extent_buffer *buf,
891 struct extent_buffer *cow,
892 int *last_ref)
893{
894 struct btrfs_fs_info *fs_info = root->fs_info;
895 u64 refs;
896 u64 owner;
897 u64 flags;
898 u64 new_flags = 0;
899 int ret;
900
901 /*
902 * Backrefs update rules:
903 *
904 * Always use full backrefs for extent pointers in tree block
905 * allocated by tree relocation.
906 *
907 * If a shared tree block is no longer referenced by its owner
908 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
909 * use full backrefs for extent pointers in tree block.
910 *
911 * If a tree block is been relocating
912 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
913 * use full backrefs for extent pointers in tree block.
914 * The reason for this is some operations (such as drop tree)
915 * are only allowed for blocks use full backrefs.
916 */
917
918 if (btrfs_block_can_be_shared(root, buf)) {
919 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
920 btrfs_header_level(buf), 1,
921 &refs, &flags);
922 if (ret)
923 return ret;
924 if (refs == 0) {
925 ret = -EROFS;
926 btrfs_handle_fs_error(fs_info, ret, NULL);
927 return ret;
928 }
929 } else {
930 refs = 1;
931 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
932 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
933 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
934 else
935 flags = 0;
936 }
937
938 owner = btrfs_header_owner(buf);
939 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
940 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
941
942 if (refs > 1) {
943 if ((owner == root->root_key.objectid ||
944 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
945 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
946 ret = btrfs_inc_ref(trans, root, buf, 1);
947 if (ret)
948 return ret;
949
950 if (root->root_key.objectid ==
951 BTRFS_TREE_RELOC_OBJECTID) {
952 ret = btrfs_dec_ref(trans, root, buf, 0);
953 if (ret)
954 return ret;
955 ret = btrfs_inc_ref(trans, root, cow, 1);
956 if (ret)
957 return ret;
958 }
959 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
960 } else {
961
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 }
970 if (new_flags != 0) {
971 int level = btrfs_header_level(buf);
972
973 ret = btrfs_set_disk_extent_flags(trans,
974 buf->start,
975 buf->len,
976 new_flags, level, 0);
977 if (ret)
978 return ret;
979 }
980 } else {
981 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
982 if (root->root_key.objectid ==
983 BTRFS_TREE_RELOC_OBJECTID)
984 ret = btrfs_inc_ref(trans, root, cow, 1);
985 else
986 ret = btrfs_inc_ref(trans, root, cow, 0);
987 if (ret)
988 return ret;
989 ret = btrfs_dec_ref(trans, root, buf, 1);
990 if (ret)
991 return ret;
992 }
993 btrfs_clean_tree_block(buf);
994 *last_ref = 1;
995 }
996 return 0;
997}
998
999static struct extent_buffer *alloc_tree_block_no_bg_flush(
1000 struct btrfs_trans_handle *trans,
1001 struct btrfs_root *root,
1002 u64 parent_start,
1003 const struct btrfs_disk_key *disk_key,
1004 int level,
1005 u64 hint,
1006 u64 empty_size)
1007{
1008 struct btrfs_fs_info *fs_info = root->fs_info;
1009 struct extent_buffer *ret;
1010
1011 /*
1012 * If we are COWing a node/leaf from the extent, chunk, device or free
1013 * space trees, make sure that we do not finish block group creation of
1014 * pending block groups. We do this to avoid a deadlock.
1015 * COWing can result in allocation of a new chunk, and flushing pending
1016 * block groups (btrfs_create_pending_block_groups()) can be triggered
1017 * when finishing allocation of a new chunk. Creation of a pending block
1018 * group modifies the extent, chunk, device and free space trees,
1019 * therefore we could deadlock with ourselves since we are holding a
1020 * lock on an extent buffer that btrfs_create_pending_block_groups() may
1021 * try to COW later.
1022 * For similar reasons, we also need to delay flushing pending block
1023 * groups when splitting a leaf or node, from one of those trees, since
1024 * we are holding a write lock on it and its parent or when inserting a
1025 * new root node for one of those trees.
1026 */
1027 if (root == fs_info->extent_root ||
1028 root == fs_info->chunk_root ||
1029 root == fs_info->dev_root ||
1030 root == fs_info->free_space_root)
1031 trans->can_flush_pending_bgs = false;
1032
1033 ret = btrfs_alloc_tree_block(trans, root, parent_start,
1034 root->root_key.objectid, disk_key, level,
1035 hint, empty_size);
1036 trans->can_flush_pending_bgs = true;
1037
1038 return ret;
1039}
1040
1041/*
1042 * does the dirty work in cow of a single block. The parent block (if
1043 * supplied) is updated to point to the new cow copy. The new buffer is marked
1044 * dirty and returned locked. If you modify the block it needs to be marked
1045 * dirty again.
1046 *
1047 * search_start -- an allocation hint for the new block
1048 *
1049 * empty_size -- a hint that you plan on doing more cow. This is the size in
1050 * bytes the allocator should try to find free next to the block it returns.
1051 * This is just a hint and may be ignored by the allocator.
1052 */
1053static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1054 struct btrfs_root *root,
1055 struct extent_buffer *buf,
1056 struct extent_buffer *parent, int parent_slot,
1057 struct extent_buffer **cow_ret,
1058 u64 search_start, u64 empty_size)
1059{
1060 struct btrfs_fs_info *fs_info = root->fs_info;
1061 struct btrfs_disk_key disk_key;
1062 struct extent_buffer *cow;
1063 int level, ret;
1064 int last_ref = 0;
1065 int unlock_orig = 0;
1066 u64 parent_start = 0;
1067
1068 if (*cow_ret == buf)
1069 unlock_orig = 1;
1070
1071 btrfs_assert_tree_locked(buf);
1072
1073 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1074 trans->transid != fs_info->running_transaction->transid);
1075 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1076 trans->transid != root->last_trans);
1077
1078 level = btrfs_header_level(buf);
1079
1080 if (level == 0)
1081 btrfs_item_key(buf, &disk_key, 0);
1082 else
1083 btrfs_node_key(buf, &disk_key, 0);
1084
1085 if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1086 parent_start = parent->start;
1087
1088 cow = alloc_tree_block_no_bg_flush(trans, root, parent_start, &disk_key,
1089 level, search_start, empty_size);
1090 if (IS_ERR(cow))
1091 return PTR_ERR(cow);
1092
1093 /* cow is set to blocking by btrfs_init_new_buffer */
1094
1095 copy_extent_buffer_full(cow, buf);
1096 btrfs_set_header_bytenr(cow, cow->start);
1097 btrfs_set_header_generation(cow, trans->transid);
1098 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1099 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1100 BTRFS_HEADER_FLAG_RELOC);
1101 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1102 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1103 else
1104 btrfs_set_header_owner(cow, root->root_key.objectid);
1105
1106 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
1107
1108 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1109 if (ret) {
1110 btrfs_abort_transaction(trans, ret);
1111 return ret;
1112 }
1113
1114 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1115 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1116 if (ret) {
1117 btrfs_abort_transaction(trans, ret);
1118 return ret;
1119 }
1120 }
1121
1122 if (buf == root->node) {
1123 WARN_ON(parent && parent != buf);
1124 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1125 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1126 parent_start = buf->start;
1127
1128 extent_buffer_get(cow);
1129 ret = tree_mod_log_insert_root(root->node, cow, 1);
1130 BUG_ON(ret < 0);
1131 rcu_assign_pointer(root->node, cow);
1132
1133 btrfs_free_tree_block(trans, root, buf, parent_start,
1134 last_ref);
1135 free_extent_buffer(buf);
1136 add_root_to_dirty_list(root);
1137 } else {
1138 WARN_ON(trans->transid != btrfs_header_generation(parent));
1139 tree_mod_log_insert_key(parent, parent_slot,
1140 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1141 btrfs_set_node_blockptr(parent, parent_slot,
1142 cow->start);
1143 btrfs_set_node_ptr_generation(parent, parent_slot,
1144 trans->transid);
1145 btrfs_mark_buffer_dirty(parent);
1146 if (last_ref) {
1147 ret = tree_mod_log_free_eb(buf);
1148 if (ret) {
1149 btrfs_abort_transaction(trans, ret);
1150 return ret;
1151 }
1152 }
1153 btrfs_free_tree_block(trans, root, buf, parent_start,
1154 last_ref);
1155 }
1156 if (unlock_orig)
1157 btrfs_tree_unlock(buf);
1158 free_extent_buffer_stale(buf);
1159 btrfs_mark_buffer_dirty(cow);
1160 *cow_ret = cow;
1161 return 0;
1162}
1163
1164/*
1165 * returns the logical address of the oldest predecessor of the given root.
1166 * entries older than time_seq are ignored.
1167 */
1168static struct tree_mod_elem *__tree_mod_log_oldest_root(
1169 struct extent_buffer *eb_root, u64 time_seq)
1170{
1171 struct tree_mod_elem *tm;
1172 struct tree_mod_elem *found = NULL;
1173 u64 root_logical = eb_root->start;
1174 int looped = 0;
1175
1176 if (!time_seq)
1177 return NULL;
1178
1179 /*
1180 * the very last operation that's logged for a root is the
1181 * replacement operation (if it is replaced at all). this has
1182 * the logical address of the *new* root, making it the very
1183 * first operation that's logged for this root.
1184 */
1185 while (1) {
1186 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1187 time_seq);
1188 if (!looped && !tm)
1189 return NULL;
1190 /*
1191 * if there are no tree operation for the oldest root, we simply
1192 * return it. this should only happen if that (old) root is at
1193 * level 0.
1194 */
1195 if (!tm)
1196 break;
1197
1198 /*
1199 * if there's an operation that's not a root replacement, we
1200 * found the oldest version of our root. normally, we'll find a
1201 * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1202 */
1203 if (tm->op != MOD_LOG_ROOT_REPLACE)
1204 break;
1205
1206 found = tm;
1207 root_logical = tm->old_root.logical;
1208 looped = 1;
1209 }
1210
1211 /* if there's no old root to return, return what we found instead */
1212 if (!found)
1213 found = tm;
1214
1215 return found;
1216}
1217
1218/*
1219 * tm is a pointer to the first operation to rewind within eb. then, all
1220 * previous operations will be rewound (until we reach something older than
1221 * time_seq).
1222 */
1223static void
1224__tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1225 u64 time_seq, struct tree_mod_elem *first_tm)
1226{
1227 u32 n;
1228 struct rb_node *next;
1229 struct tree_mod_elem *tm = first_tm;
1230 unsigned long o_dst;
1231 unsigned long o_src;
1232 unsigned long p_size = sizeof(struct btrfs_key_ptr);
1233
1234 n = btrfs_header_nritems(eb);
1235 read_lock(&fs_info->tree_mod_log_lock);
1236 while (tm && tm->seq >= time_seq) {
1237 /*
1238 * all the operations are recorded with the operator used for
1239 * the modification. as we're going backwards, we do the
1240 * opposite of each operation here.
1241 */
1242 switch (tm->op) {
1243 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1244 BUG_ON(tm->slot < n);
1245 /* Fallthrough */
1246 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1247 case MOD_LOG_KEY_REMOVE:
1248 btrfs_set_node_key(eb, &tm->key, tm->slot);
1249 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1250 btrfs_set_node_ptr_generation(eb, tm->slot,
1251 tm->generation);
1252 n++;
1253 break;
1254 case MOD_LOG_KEY_REPLACE:
1255 BUG_ON(tm->slot >= n);
1256 btrfs_set_node_key(eb, &tm->key, tm->slot);
1257 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1258 btrfs_set_node_ptr_generation(eb, tm->slot,
1259 tm->generation);
1260 break;
1261 case MOD_LOG_KEY_ADD:
1262 /* if a move operation is needed it's in the log */
1263 n--;
1264 break;
1265 case MOD_LOG_MOVE_KEYS:
1266 o_dst = btrfs_node_key_ptr_offset(tm->slot);
1267 o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1268 memmove_extent_buffer(eb, o_dst, o_src,
1269 tm->move.nr_items * p_size);
1270 break;
1271 case MOD_LOG_ROOT_REPLACE:
1272 /*
1273 * this operation is special. for roots, this must be
1274 * handled explicitly before rewinding.
1275 * for non-roots, this operation may exist if the node
1276 * was a root: root A -> child B; then A gets empty and
1277 * B is promoted to the new root. in the mod log, we'll
1278 * have a root-replace operation for B, a tree block
1279 * that is no root. we simply ignore that operation.
1280 */
1281 break;
1282 }
1283 next = rb_next(&tm->node);
1284 if (!next)
1285 break;
1286 tm = rb_entry(next, struct tree_mod_elem, node);
1287 if (tm->logical != first_tm->logical)
1288 break;
1289 }
1290 read_unlock(&fs_info->tree_mod_log_lock);
1291 btrfs_set_header_nritems(eb, n);
1292}
1293
1294/*
1295 * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1296 * is returned. If rewind operations happen, a fresh buffer is returned. The
1297 * returned buffer is always read-locked. If the returned buffer is not the
1298 * input buffer, the lock on the input buffer is released and the input buffer
1299 * is freed (its refcount is decremented).
1300 */
1301static struct extent_buffer *
1302tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1303 struct extent_buffer *eb, u64 time_seq)
1304{
1305 struct extent_buffer *eb_rewin;
1306 struct tree_mod_elem *tm;
1307
1308 if (!time_seq)
1309 return eb;
1310
1311 if (btrfs_header_level(eb) == 0)
1312 return eb;
1313
1314 tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1315 if (!tm)
1316 return eb;
1317
1318 btrfs_set_path_blocking(path);
1319 btrfs_set_lock_blocking_read(eb);
1320
1321 if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1322 BUG_ON(tm->slot != 0);
1323 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1324 if (!eb_rewin) {
1325 btrfs_tree_read_unlock_blocking(eb);
1326 free_extent_buffer(eb);
1327 return NULL;
1328 }
1329 btrfs_set_header_bytenr(eb_rewin, eb->start);
1330 btrfs_set_header_backref_rev(eb_rewin,
1331 btrfs_header_backref_rev(eb));
1332 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1333 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1334 } else {
1335 eb_rewin = btrfs_clone_extent_buffer(eb);
1336 if (!eb_rewin) {
1337 btrfs_tree_read_unlock_blocking(eb);
1338 free_extent_buffer(eb);
1339 return NULL;
1340 }
1341 }
1342
1343 btrfs_tree_read_unlock_blocking(eb);
1344 free_extent_buffer(eb);
1345
1346 btrfs_tree_read_lock(eb_rewin);
1347 __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1348 WARN_ON(btrfs_header_nritems(eb_rewin) >
1349 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1350
1351 return eb_rewin;
1352}
1353
1354/*
1355 * get_old_root() rewinds the state of @root's root node to the given @time_seq
1356 * value. If there are no changes, the current root->root_node is returned. If
1357 * anything changed in between, there's a fresh buffer allocated on which the
1358 * rewind operations are done. In any case, the returned buffer is read locked.
1359 * Returns NULL on error (with no locks held).
1360 */
1361static inline struct extent_buffer *
1362get_old_root(struct btrfs_root *root, u64 time_seq)
1363{
1364 struct btrfs_fs_info *fs_info = root->fs_info;
1365 struct tree_mod_elem *tm;
1366 struct extent_buffer *eb = NULL;
1367 struct extent_buffer *eb_root;
1368 u64 eb_root_owner = 0;
1369 struct extent_buffer *old;
1370 struct tree_mod_root *old_root = NULL;
1371 u64 old_generation = 0;
1372 u64 logical;
1373 int level;
1374
1375 eb_root = btrfs_read_lock_root_node(root);
1376 tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1377 if (!tm)
1378 return eb_root;
1379
1380 if (tm->op == MOD_LOG_ROOT_REPLACE) {
1381 old_root = &tm->old_root;
1382 old_generation = tm->generation;
1383 logical = old_root->logical;
1384 level = old_root->level;
1385 } else {
1386 logical = eb_root->start;
1387 level = btrfs_header_level(eb_root);
1388 }
1389
1390 tm = tree_mod_log_search(fs_info, logical, time_seq);
1391 if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1392 btrfs_tree_read_unlock(eb_root);
1393 free_extent_buffer(eb_root);
1394 old = read_tree_block(fs_info, logical, 0, level, NULL);
1395 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1396 if (!IS_ERR(old))
1397 free_extent_buffer(old);
1398 btrfs_warn(fs_info,
1399 "failed to read tree block %llu from get_old_root",
1400 logical);
1401 } else {
1402 eb = btrfs_clone_extent_buffer(old);
1403 free_extent_buffer(old);
1404 }
1405 } else if (old_root) {
1406 eb_root_owner = btrfs_header_owner(eb_root);
1407 btrfs_tree_read_unlock(eb_root);
1408 free_extent_buffer(eb_root);
1409 eb = alloc_dummy_extent_buffer(fs_info, logical);
1410 } else {
1411 btrfs_set_lock_blocking_read(eb_root);
1412 eb = btrfs_clone_extent_buffer(eb_root);
1413 btrfs_tree_read_unlock_blocking(eb_root);
1414 free_extent_buffer(eb_root);
1415 }
1416
1417 if (!eb)
1418 return NULL;
1419 btrfs_tree_read_lock(eb);
1420 if (old_root) {
1421 btrfs_set_header_bytenr(eb, eb->start);
1422 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1423 btrfs_set_header_owner(eb, eb_root_owner);
1424 btrfs_set_header_level(eb, old_root->level);
1425 btrfs_set_header_generation(eb, old_generation);
1426 }
1427 if (tm)
1428 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1429 else
1430 WARN_ON(btrfs_header_level(eb) != 0);
1431 WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1432
1433 return eb;
1434}
1435
1436int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1437{
1438 struct tree_mod_elem *tm;
1439 int level;
1440 struct extent_buffer *eb_root = btrfs_root_node(root);
1441
1442 tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1443 if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1444 level = tm->old_root.level;
1445 } else {
1446 level = btrfs_header_level(eb_root);
1447 }
1448 free_extent_buffer(eb_root);
1449
1450 return level;
1451}
1452
1453static inline int should_cow_block(struct btrfs_trans_handle *trans,
1454 struct btrfs_root *root,
1455 struct extent_buffer *buf)
1456{
1457 if (btrfs_is_testing(root->fs_info))
1458 return 0;
1459
1460 /* Ensure we can see the FORCE_COW bit */
1461 smp_mb__before_atomic();
1462
1463 /*
1464 * We do not need to cow a block if
1465 * 1) this block is not created or changed in this transaction;
1466 * 2) this block does not belong to TREE_RELOC tree;
1467 * 3) the root is not forced COW.
1468 *
1469 * What is forced COW:
1470 * when we create snapshot during committing the transaction,
1471 * after we've finished copying src root, we must COW the shared
1472 * block to ensure the metadata consistency.
1473 */
1474 if (btrfs_header_generation(buf) == trans->transid &&
1475 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1476 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1477 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1478 !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1479 return 0;
1480 return 1;
1481}
1482
1483/*
1484 * cows a single block, see __btrfs_cow_block for the real work.
1485 * This version of it has extra checks so that a block isn't COWed more than
1486 * once per transaction, as long as it hasn't been written yet
1487 */
1488noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1489 struct btrfs_root *root, struct extent_buffer *buf,
1490 struct extent_buffer *parent, int parent_slot,
1491 struct extent_buffer **cow_ret)
1492{
1493 struct btrfs_fs_info *fs_info = root->fs_info;
1494 u64 search_start;
1495 int ret;
1496
1497 if (test_bit(BTRFS_ROOT_DELETING, &root->state))
1498 btrfs_err(fs_info,
1499 "COW'ing blocks on a fs root that's being dropped");
1500
1501 if (trans->transaction != fs_info->running_transaction)
1502 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1503 trans->transid,
1504 fs_info->running_transaction->transid);
1505
1506 if (trans->transid != fs_info->generation)
1507 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1508 trans->transid, fs_info->generation);
1509
1510 if (!should_cow_block(trans, root, buf)) {
1511 trans->dirty = true;
1512 *cow_ret = buf;
1513 return 0;
1514 }
1515
1516 search_start = buf->start & ~((u64)SZ_1G - 1);
1517
1518 if (parent)
1519 btrfs_set_lock_blocking_write(parent);
1520 btrfs_set_lock_blocking_write(buf);
1521
1522 /*
1523 * Before CoWing this block for later modification, check if it's
1524 * the subtree root and do the delayed subtree trace if needed.
1525 *
1526 * Also We don't care about the error, as it's handled internally.
1527 */
1528 btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
1529 ret = __btrfs_cow_block(trans, root, buf, parent,
1530 parent_slot, cow_ret, search_start, 0);
1531
1532 trace_btrfs_cow_block(root, buf, *cow_ret);
1533
1534 return ret;
1535}
1536
1537/*
1538 * helper function for defrag to decide if two blocks pointed to by a
1539 * node are actually close by
1540 */
1541static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1542{
1543 if (blocknr < other && other - (blocknr + blocksize) < 32768)
1544 return 1;
1545 if (blocknr > other && blocknr - (other + blocksize) < 32768)
1546 return 1;
1547 return 0;
1548}
1549
1550/*
1551 * compare two keys in a memcmp fashion
1552 */
1553static int comp_keys(const struct btrfs_disk_key *disk,
1554 const struct btrfs_key *k2)
1555{
1556 struct btrfs_key k1;
1557
1558 btrfs_disk_key_to_cpu(&k1, disk);
1559
1560 return btrfs_comp_cpu_keys(&k1, k2);
1561}
1562
1563/*
1564 * same as comp_keys only with two btrfs_key's
1565 */
1566int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1567{
1568 if (k1->objectid > k2->objectid)
1569 return 1;
1570 if (k1->objectid < k2->objectid)
1571 return -1;
1572 if (k1->type > k2->type)
1573 return 1;
1574 if (k1->type < k2->type)
1575 return -1;
1576 if (k1->offset > k2->offset)
1577 return 1;
1578 if (k1->offset < k2->offset)
1579 return -1;
1580 return 0;
1581}
1582
1583/*
1584 * this is used by the defrag code to go through all the
1585 * leaves pointed to by a node and reallocate them so that
1586 * disk order is close to key order
1587 */
1588int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1589 struct btrfs_root *root, struct extent_buffer *parent,
1590 int start_slot, u64 *last_ret,
1591 struct btrfs_key *progress)
1592{
1593 struct btrfs_fs_info *fs_info = root->fs_info;
1594 struct extent_buffer *cur;
1595 u64 blocknr;
1596 u64 gen;
1597 u64 search_start = *last_ret;
1598 u64 last_block = 0;
1599 u64 other;
1600 u32 parent_nritems;
1601 int end_slot;
1602 int i;
1603 int err = 0;
1604 int parent_level;
1605 int uptodate;
1606 u32 blocksize;
1607 int progress_passed = 0;
1608 struct btrfs_disk_key disk_key;
1609
1610 parent_level = btrfs_header_level(parent);
1611
1612 WARN_ON(trans->transaction != fs_info->running_transaction);
1613 WARN_ON(trans->transid != fs_info->generation);
1614
1615 parent_nritems = btrfs_header_nritems(parent);
1616 blocksize = fs_info->nodesize;
1617 end_slot = parent_nritems - 1;
1618
1619 if (parent_nritems <= 1)
1620 return 0;
1621
1622 btrfs_set_lock_blocking_write(parent);
1623
1624 for (i = start_slot; i <= end_slot; i++) {
1625 struct btrfs_key first_key;
1626 int close = 1;
1627
1628 btrfs_node_key(parent, &disk_key, i);
1629 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1630 continue;
1631
1632 progress_passed = 1;
1633 blocknr = btrfs_node_blockptr(parent, i);
1634 gen = btrfs_node_ptr_generation(parent, i);
1635 btrfs_node_key_to_cpu(parent, &first_key, i);
1636 if (last_block == 0)
1637 last_block = blocknr;
1638
1639 if (i > 0) {
1640 other = btrfs_node_blockptr(parent, i - 1);
1641 close = close_blocks(blocknr, other, blocksize);
1642 }
1643 if (!close && i < end_slot) {
1644 other = btrfs_node_blockptr(parent, i + 1);
1645 close = close_blocks(blocknr, other, blocksize);
1646 }
1647 if (close) {
1648 last_block = blocknr;
1649 continue;
1650 }
1651
1652 cur = find_extent_buffer(fs_info, blocknr);
1653 if (cur)
1654 uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1655 else
1656 uptodate = 0;
1657 if (!cur || !uptodate) {
1658 if (!cur) {
1659 cur = read_tree_block(fs_info, blocknr, gen,
1660 parent_level - 1,
1661 &first_key);
1662 if (IS_ERR(cur)) {
1663 return PTR_ERR(cur);
1664 } else if (!extent_buffer_uptodate(cur)) {
1665 free_extent_buffer(cur);
1666 return -EIO;
1667 }
1668 } else if (!uptodate) {
1669 err = btrfs_read_buffer(cur, gen,
1670 parent_level - 1,&first_key);
1671 if (err) {
1672 free_extent_buffer(cur);
1673 return err;
1674 }
1675 }
1676 }
1677 if (search_start == 0)
1678 search_start = last_block;
1679
1680 btrfs_tree_lock(cur);
1681 btrfs_set_lock_blocking_write(cur);
1682 err = __btrfs_cow_block(trans, root, cur, parent, i,
1683 &cur, search_start,
1684 min(16 * blocksize,
1685 (end_slot - i) * blocksize));
1686 if (err) {
1687 btrfs_tree_unlock(cur);
1688 free_extent_buffer(cur);
1689 break;
1690 }
1691 search_start = cur->start;
1692 last_block = cur->start;
1693 *last_ret = search_start;
1694 btrfs_tree_unlock(cur);
1695 free_extent_buffer(cur);
1696 }
1697 return err;
1698}
1699
1700/*
1701 * search for key in the extent_buffer. The items start at offset p,
1702 * and they are item_size apart. There are 'max' items in p.
1703 *
1704 * the slot in the array is returned via slot, and it points to
1705 * the place where you would insert key if it is not found in
1706 * the array.
1707 *
1708 * slot may point to max if the key is bigger than all of the keys
1709 */
1710static noinline int generic_bin_search(struct extent_buffer *eb,
1711 unsigned long p, int item_size,
1712 const struct btrfs_key *key,
1713 int max, int *slot)
1714{
1715 int low = 0;
1716 int high = max;
1717 int mid;
1718 int ret;
1719 struct btrfs_disk_key *tmp = NULL;
1720 struct btrfs_disk_key unaligned;
1721 unsigned long offset;
1722 char *kaddr = NULL;
1723 unsigned long map_start = 0;
1724 unsigned long map_len = 0;
1725 int err;
1726
1727 if (low > high) {
1728 btrfs_err(eb->fs_info,
1729 "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1730 __func__, low, high, eb->start,
1731 btrfs_header_owner(eb), btrfs_header_level(eb));
1732 return -EINVAL;
1733 }
1734
1735 while (low < high) {
1736 mid = (low + high) / 2;
1737 offset = p + mid * item_size;
1738
1739 if (!kaddr || offset < map_start ||
1740 (offset + sizeof(struct btrfs_disk_key)) >
1741 map_start + map_len) {
1742
1743 err = map_private_extent_buffer(eb, offset,
1744 sizeof(struct btrfs_disk_key),
1745 &kaddr, &map_start, &map_len);
1746
1747 if (!err) {
1748 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1749 map_start);
1750 } else if (err == 1) {
1751 read_extent_buffer(eb, &unaligned,
1752 offset, sizeof(unaligned));
1753 tmp = &unaligned;
1754 } else {
1755 return err;
1756 }
1757
1758 } else {
1759 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1760 map_start);
1761 }
1762 ret = comp_keys(tmp, key);
1763
1764 if (ret < 0)
1765 low = mid + 1;
1766 else if (ret > 0)
1767 high = mid;
1768 else {
1769 *slot = mid;
1770 return 0;
1771 }
1772 }
1773 *slot = low;
1774 return 1;
1775}
1776
1777/*
1778 * simple bin_search frontend that does the right thing for
1779 * leaves vs nodes
1780 */
1781int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1782 int level, int *slot)
1783{
1784 if (level == 0)
1785 return generic_bin_search(eb,
1786 offsetof(struct btrfs_leaf, items),
1787 sizeof(struct btrfs_item),
1788 key, btrfs_header_nritems(eb),
1789 slot);
1790 else
1791 return generic_bin_search(eb,
1792 offsetof(struct btrfs_node, ptrs),
1793 sizeof(struct btrfs_key_ptr),
1794 key, btrfs_header_nritems(eb),
1795 slot);
1796}
1797
1798static void root_add_used(struct btrfs_root *root, u32 size)
1799{
1800 spin_lock(&root->accounting_lock);
1801 btrfs_set_root_used(&root->root_item,
1802 btrfs_root_used(&root->root_item) + size);
1803 spin_unlock(&root->accounting_lock);
1804}
1805
1806static void root_sub_used(struct btrfs_root *root, u32 size)
1807{
1808 spin_lock(&root->accounting_lock);
1809 btrfs_set_root_used(&root->root_item,
1810 btrfs_root_used(&root->root_item) - size);
1811 spin_unlock(&root->accounting_lock);
1812}
1813
1814/* given a node and slot number, this reads the blocks it points to. The
1815 * extent buffer is returned with a reference taken (but unlocked).
1816 */
1817struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
1818 int slot)
1819{
1820 int level = btrfs_header_level(parent);
1821 struct extent_buffer *eb;
1822 struct btrfs_key first_key;
1823
1824 if (slot < 0 || slot >= btrfs_header_nritems(parent))
1825 return ERR_PTR(-ENOENT);
1826
1827 BUG_ON(level == 0);
1828
1829 btrfs_node_key_to_cpu(parent, &first_key, slot);
1830 eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
1831 btrfs_node_ptr_generation(parent, slot),
1832 level - 1, &first_key);
1833 if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1834 free_extent_buffer(eb);
1835 eb = ERR_PTR(-EIO);
1836 }
1837
1838 return eb;
1839}
1840
1841/*
1842 * node level balancing, used to make sure nodes are in proper order for
1843 * item deletion. We balance from the top down, so we have to make sure
1844 * that a deletion won't leave an node completely empty later on.
1845 */
1846static noinline int balance_level(struct btrfs_trans_handle *trans,
1847 struct btrfs_root *root,
1848 struct btrfs_path *path, int level)
1849{
1850 struct btrfs_fs_info *fs_info = root->fs_info;
1851 struct extent_buffer *right = NULL;
1852 struct extent_buffer *mid;
1853 struct extent_buffer *left = NULL;
1854 struct extent_buffer *parent = NULL;
1855 int ret = 0;
1856 int wret;
1857 int pslot;
1858 int orig_slot = path->slots[level];
1859 u64 orig_ptr;
1860
1861 ASSERT(level > 0);
1862
1863 mid = path->nodes[level];
1864
1865 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1866 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1867 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1868
1869 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1870
1871 if (level < BTRFS_MAX_LEVEL - 1) {
1872 parent = path->nodes[level + 1];
1873 pslot = path->slots[level + 1];
1874 }
1875
1876 /*
1877 * deal with the case where there is only one pointer in the root
1878 * by promoting the node below to a root
1879 */
1880 if (!parent) {
1881 struct extent_buffer *child;
1882
1883 if (btrfs_header_nritems(mid) != 1)
1884 return 0;
1885
1886 /* promote the child to a root */
1887 child = btrfs_read_node_slot(mid, 0);
1888 if (IS_ERR(child)) {
1889 ret = PTR_ERR(child);
1890 btrfs_handle_fs_error(fs_info, ret, NULL);
1891 goto enospc;
1892 }
1893
1894 btrfs_tree_lock(child);
1895 btrfs_set_lock_blocking_write(child);
1896 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1897 if (ret) {
1898 btrfs_tree_unlock(child);
1899 free_extent_buffer(child);
1900 goto enospc;
1901 }
1902
1903 ret = tree_mod_log_insert_root(root->node, child, 1);
1904 BUG_ON(ret < 0);
1905 rcu_assign_pointer(root->node, child);
1906
1907 add_root_to_dirty_list(root);
1908 btrfs_tree_unlock(child);
1909
1910 path->locks[level] = 0;
1911 path->nodes[level] = NULL;
1912 btrfs_clean_tree_block(mid);
1913 btrfs_tree_unlock(mid);
1914 /* once for the path */
1915 free_extent_buffer(mid);
1916
1917 root_sub_used(root, mid->len);
1918 btrfs_free_tree_block(trans, root, mid, 0, 1);
1919 /* once for the root ptr */
1920 free_extent_buffer_stale(mid);
1921 return 0;
1922 }
1923 if (btrfs_header_nritems(mid) >
1924 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1925 return 0;
1926
1927 left = btrfs_read_node_slot(parent, pslot - 1);
1928 if (IS_ERR(left))
1929 left = NULL;
1930
1931 if (left) {
1932 btrfs_tree_lock(left);
1933 btrfs_set_lock_blocking_write(left);
1934 wret = btrfs_cow_block(trans, root, left,
1935 parent, pslot - 1, &left);
1936 if (wret) {
1937 ret = wret;
1938 goto enospc;
1939 }
1940 }
1941
1942 right = btrfs_read_node_slot(parent, pslot + 1);
1943 if (IS_ERR(right))
1944 right = NULL;
1945
1946 if (right) {
1947 btrfs_tree_lock(right);
1948 btrfs_set_lock_blocking_write(right);
1949 wret = btrfs_cow_block(trans, root, right,
1950 parent, pslot + 1, &right);
1951 if (wret) {
1952 ret = wret;
1953 goto enospc;
1954 }
1955 }
1956
1957 /* first, try to make some room in the middle buffer */
1958 if (left) {
1959 orig_slot += btrfs_header_nritems(left);
1960 wret = push_node_left(trans, left, mid, 1);
1961 if (wret < 0)
1962 ret = wret;
1963 }
1964
1965 /*
1966 * then try to empty the right most buffer into the middle
1967 */
1968 if (right) {
1969 wret = push_node_left(trans, mid, right, 1);
1970 if (wret < 0 && wret != -ENOSPC)
1971 ret = wret;
1972 if (btrfs_header_nritems(right) == 0) {
1973 btrfs_clean_tree_block(right);
1974 btrfs_tree_unlock(right);
1975 del_ptr(root, path, level + 1, pslot + 1);
1976 root_sub_used(root, right->len);
1977 btrfs_free_tree_block(trans, root, right, 0, 1);
1978 free_extent_buffer_stale(right);
1979 right = NULL;
1980 } else {
1981 struct btrfs_disk_key right_key;
1982 btrfs_node_key(right, &right_key, 0);
1983 ret = tree_mod_log_insert_key(parent, pslot + 1,
1984 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1985 BUG_ON(ret < 0);
1986 btrfs_set_node_key(parent, &right_key, pslot + 1);
1987 btrfs_mark_buffer_dirty(parent);
1988 }
1989 }
1990 if (btrfs_header_nritems(mid) == 1) {
1991 /*
1992 * we're not allowed to leave a node with one item in the
1993 * tree during a delete. A deletion from lower in the tree
1994 * could try to delete the only pointer in this node.
1995 * So, pull some keys from the left.
1996 * There has to be a left pointer at this point because
1997 * otherwise we would have pulled some pointers from the
1998 * right
1999 */
2000 if (!left) {
2001 ret = -EROFS;
2002 btrfs_handle_fs_error(fs_info, ret, NULL);
2003 goto enospc;
2004 }
2005 wret = balance_node_right(trans, mid, left);
2006 if (wret < 0) {
2007 ret = wret;
2008 goto enospc;
2009 }
2010 if (wret == 1) {
2011 wret = push_node_left(trans, left, mid, 1);
2012 if (wret < 0)
2013 ret = wret;
2014 }
2015 BUG_ON(wret == 1);
2016 }
2017 if (btrfs_header_nritems(mid) == 0) {
2018 btrfs_clean_tree_block(mid);
2019 btrfs_tree_unlock(mid);
2020 del_ptr(root, path, level + 1, pslot);
2021 root_sub_used(root, mid->len);
2022 btrfs_free_tree_block(trans, root, mid, 0, 1);
2023 free_extent_buffer_stale(mid);
2024 mid = NULL;
2025 } else {
2026 /* update the parent key to reflect our changes */
2027 struct btrfs_disk_key mid_key;
2028 btrfs_node_key(mid, &mid_key, 0);
2029 ret = tree_mod_log_insert_key(parent, pslot,
2030 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2031 BUG_ON(ret < 0);
2032 btrfs_set_node_key(parent, &mid_key, pslot);
2033 btrfs_mark_buffer_dirty(parent);
2034 }
2035
2036 /* update the path */
2037 if (left) {
2038 if (btrfs_header_nritems(left) > orig_slot) {
2039 extent_buffer_get(left);
2040 /* left was locked after cow */
2041 path->nodes[level] = left;
2042 path->slots[level + 1] -= 1;
2043 path->slots[level] = orig_slot;
2044 if (mid) {
2045 btrfs_tree_unlock(mid);
2046 free_extent_buffer(mid);
2047 }
2048 } else {
2049 orig_slot -= btrfs_header_nritems(left);
2050 path->slots[level] = orig_slot;
2051 }
2052 }
2053 /* double check we haven't messed things up */
2054 if (orig_ptr !=
2055 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2056 BUG();
2057enospc:
2058 if (right) {
2059 btrfs_tree_unlock(right);
2060 free_extent_buffer(right);
2061 }
2062 if (left) {
2063 if (path->nodes[level] != left)
2064 btrfs_tree_unlock(left);
2065 free_extent_buffer(left);
2066 }
2067 return ret;
2068}
2069
2070/* Node balancing for insertion. Here we only split or push nodes around
2071 * when they are completely full. This is also done top down, so we
2072 * have to be pessimistic.
2073 */
2074static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2075 struct btrfs_root *root,
2076 struct btrfs_path *path, int level)
2077{
2078 struct btrfs_fs_info *fs_info = root->fs_info;
2079 struct extent_buffer *right = NULL;
2080 struct extent_buffer *mid;
2081 struct extent_buffer *left = NULL;
2082 struct extent_buffer *parent = NULL;
2083 int ret = 0;
2084 int wret;
2085 int pslot;
2086 int orig_slot = path->slots[level];
2087
2088 if (level == 0)
2089 return 1;
2090
2091 mid = path->nodes[level];
2092 WARN_ON(btrfs_header_generation(mid) != trans->transid);
2093
2094 if (level < BTRFS_MAX_LEVEL - 1) {
2095 parent = path->nodes[level + 1];
2096 pslot = path->slots[level + 1];
2097 }
2098
2099 if (!parent)
2100 return 1;
2101
2102 left = btrfs_read_node_slot(parent, pslot - 1);
2103 if (IS_ERR(left))
2104 left = NULL;
2105
2106 /* first, try to make some room in the middle buffer */
2107 if (left) {
2108 u32 left_nr;
2109
2110 btrfs_tree_lock(left);
2111 btrfs_set_lock_blocking_write(left);
2112
2113 left_nr = btrfs_header_nritems(left);
2114 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2115 wret = 1;
2116 } else {
2117 ret = btrfs_cow_block(trans, root, left, parent,
2118 pslot - 1, &left);
2119 if (ret)
2120 wret = 1;
2121 else {
2122 wret = push_node_left(trans, left, mid, 0);
2123 }
2124 }
2125 if (wret < 0)
2126 ret = wret;
2127 if (wret == 0) {
2128 struct btrfs_disk_key disk_key;
2129 orig_slot += left_nr;
2130 btrfs_node_key(mid, &disk_key, 0);
2131 ret = tree_mod_log_insert_key(parent, pslot,
2132 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2133 BUG_ON(ret < 0);
2134 btrfs_set_node_key(parent, &disk_key, pslot);
2135 btrfs_mark_buffer_dirty(parent);
2136 if (btrfs_header_nritems(left) > orig_slot) {
2137 path->nodes[level] = left;
2138 path->slots[level + 1] -= 1;
2139 path->slots[level] = orig_slot;
2140 btrfs_tree_unlock(mid);
2141 free_extent_buffer(mid);
2142 } else {
2143 orig_slot -=
2144 btrfs_header_nritems(left);
2145 path->slots[level] = orig_slot;
2146 btrfs_tree_unlock(left);
2147 free_extent_buffer(left);
2148 }
2149 return 0;
2150 }
2151 btrfs_tree_unlock(left);
2152 free_extent_buffer(left);
2153 }
2154 right = btrfs_read_node_slot(parent, pslot + 1);
2155 if (IS_ERR(right))
2156 right = NULL;
2157
2158 /*
2159 * then try to empty the right most buffer into the middle
2160 */
2161 if (right) {
2162 u32 right_nr;
2163
2164 btrfs_tree_lock(right);
2165 btrfs_set_lock_blocking_write(right);
2166
2167 right_nr = btrfs_header_nritems(right);
2168 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2169 wret = 1;
2170 } else {
2171 ret = btrfs_cow_block(trans, root, right,
2172 parent, pslot + 1,
2173 &right);
2174 if (ret)
2175 wret = 1;
2176 else {
2177 wret = balance_node_right(trans, right, mid);
2178 }
2179 }
2180 if (wret < 0)
2181 ret = wret;
2182 if (wret == 0) {
2183 struct btrfs_disk_key disk_key;
2184
2185 btrfs_node_key(right, &disk_key, 0);
2186 ret = tree_mod_log_insert_key(parent, pslot + 1,
2187 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2188 BUG_ON(ret < 0);
2189 btrfs_set_node_key(parent, &disk_key, pslot + 1);
2190 btrfs_mark_buffer_dirty(parent);
2191
2192 if (btrfs_header_nritems(mid) <= orig_slot) {
2193 path->nodes[level] = right;
2194 path->slots[level + 1] += 1;
2195 path->slots[level] = orig_slot -
2196 btrfs_header_nritems(mid);
2197 btrfs_tree_unlock(mid);
2198 free_extent_buffer(mid);
2199 } else {
2200 btrfs_tree_unlock(right);
2201 free_extent_buffer(right);
2202 }
2203 return 0;
2204 }
2205 btrfs_tree_unlock(right);
2206 free_extent_buffer(right);
2207 }
2208 return 1;
2209}
2210
2211/*
2212 * readahead one full node of leaves, finding things that are close
2213 * to the block in 'slot', and triggering ra on them.
2214 */
2215static void reada_for_search(struct btrfs_fs_info *fs_info,
2216 struct btrfs_path *path,
2217 int level, int slot, u64 objectid)
2218{
2219 struct extent_buffer *node;
2220 struct btrfs_disk_key disk_key;
2221 u32 nritems;
2222 u64 search;
2223 u64 target;
2224 u64 nread = 0;
2225 struct extent_buffer *eb;
2226 u32 nr;
2227 u32 blocksize;
2228 u32 nscan = 0;
2229
2230 if (level != 1)
2231 return;
2232
2233 if (!path->nodes[level])
2234 return;
2235
2236 node = path->nodes[level];
2237
2238 search = btrfs_node_blockptr(node, slot);
2239 blocksize = fs_info->nodesize;
2240 eb = find_extent_buffer(fs_info, search);
2241 if (eb) {
2242 free_extent_buffer(eb);
2243 return;
2244 }
2245
2246 target = search;
2247
2248 nritems = btrfs_header_nritems(node);
2249 nr = slot;
2250
2251 while (1) {
2252 if (path->reada == READA_BACK) {
2253 if (nr == 0)
2254 break;
2255 nr--;
2256 } else if (path->reada == READA_FORWARD) {
2257 nr++;
2258 if (nr >= nritems)
2259 break;
2260 }
2261 if (path->reada == READA_BACK && objectid) {
2262 btrfs_node_key(node, &disk_key, nr);
2263 if (btrfs_disk_key_objectid(&disk_key) != objectid)
2264 break;
2265 }
2266 search = btrfs_node_blockptr(node, nr);
2267 if ((search <= target && target - search <= 65536) ||
2268 (search > target && search - target <= 65536)) {
2269 readahead_tree_block(fs_info, search);
2270 nread += blocksize;
2271 }
2272 nscan++;
2273 if ((nread > 65536 || nscan > 32))
2274 break;
2275 }
2276}
2277
2278static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2279 struct btrfs_path *path, int level)
2280{
2281 int slot;
2282 int nritems;
2283 struct extent_buffer *parent;
2284 struct extent_buffer *eb;
2285 u64 gen;
2286 u64 block1 = 0;
2287 u64 block2 = 0;
2288
2289 parent = path->nodes[level + 1];
2290 if (!parent)
2291 return;
2292
2293 nritems = btrfs_header_nritems(parent);
2294 slot = path->slots[level + 1];
2295
2296 if (slot > 0) {
2297 block1 = btrfs_node_blockptr(parent, slot - 1);
2298 gen = btrfs_node_ptr_generation(parent, slot - 1);
2299 eb = find_extent_buffer(fs_info, block1);
2300 /*
2301 * if we get -eagain from btrfs_buffer_uptodate, we
2302 * don't want to return eagain here. That will loop
2303 * forever
2304 */
2305 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2306 block1 = 0;
2307 free_extent_buffer(eb);
2308 }
2309 if (slot + 1 < nritems) {
2310 block2 = btrfs_node_blockptr(parent, slot + 1);
2311 gen = btrfs_node_ptr_generation(parent, slot + 1);
2312 eb = find_extent_buffer(fs_info, block2);
2313 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2314 block2 = 0;
2315 free_extent_buffer(eb);
2316 }
2317
2318 if (block1)
2319 readahead_tree_block(fs_info, block1);
2320 if (block2)
2321 readahead_tree_block(fs_info, block2);
2322}
2323
2324
2325/*
2326 * when we walk down the tree, it is usually safe to unlock the higher layers
2327 * in the tree. The exceptions are when our path goes through slot 0, because
2328 * operations on the tree might require changing key pointers higher up in the
2329 * tree.
2330 *
2331 * callers might also have set path->keep_locks, which tells this code to keep
2332 * the lock if the path points to the last slot in the block. This is part of
2333 * walking through the tree, and selecting the next slot in the higher block.
2334 *
2335 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
2336 * if lowest_unlock is 1, level 0 won't be unlocked
2337 */
2338static noinline void unlock_up(struct btrfs_path *path, int level,
2339 int lowest_unlock, int min_write_lock_level,
2340 int *write_lock_level)
2341{
2342 int i;
2343 int skip_level = level;
2344 int no_skips = 0;
2345 struct extent_buffer *t;
2346
2347 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2348 if (!path->nodes[i])
2349 break;
2350 if (!path->locks[i])
2351 break;
2352 if (!no_skips && path->slots[i] == 0) {
2353 skip_level = i + 1;
2354 continue;
2355 }
2356 if (!no_skips && path->keep_locks) {
2357 u32 nritems;
2358 t = path->nodes[i];
2359 nritems = btrfs_header_nritems(t);
2360 if (nritems < 1 || path->slots[i] >= nritems - 1) {
2361 skip_level = i + 1;
2362 continue;
2363 }
2364 }
2365 if (skip_level < i && i >= lowest_unlock)
2366 no_skips = 1;
2367
2368 t = path->nodes[i];
2369 if (i >= lowest_unlock && i > skip_level) {
2370 btrfs_tree_unlock_rw(t, path->locks[i]);
2371 path->locks[i] = 0;
2372 if (write_lock_level &&
2373 i > min_write_lock_level &&
2374 i <= *write_lock_level) {
2375 *write_lock_level = i - 1;
2376 }
2377 }
2378 }
2379}
2380
2381/*
2382 * This releases any locks held in the path starting at level and
2383 * going all the way up to the root.
2384 *
2385 * btrfs_search_slot will keep the lock held on higher nodes in a few
2386 * corner cases, such as COW of the block at slot zero in the node. This
2387 * ignores those rules, and it should only be called when there are no
2388 * more updates to be done higher up in the tree.
2389 */
2390noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2391{
2392 int i;
2393
2394 if (path->keep_locks)
2395 return;
2396
2397 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2398 if (!path->nodes[i])
2399 continue;
2400 if (!path->locks[i])
2401 continue;
2402 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2403 path->locks[i] = 0;
2404 }
2405}
2406
2407/*
2408 * helper function for btrfs_search_slot. The goal is to find a block
2409 * in cache without setting the path to blocking. If we find the block
2410 * we return zero and the path is unchanged.
2411 *
2412 * If we can't find the block, we set the path blocking and do some
2413 * reada. -EAGAIN is returned and the search must be repeated.
2414 */
2415static int
2416read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2417 struct extent_buffer **eb_ret, int level, int slot,
2418 const struct btrfs_key *key)
2419{
2420 struct btrfs_fs_info *fs_info = root->fs_info;
2421 u64 blocknr;
2422 u64 gen;
2423 struct extent_buffer *b = *eb_ret;
2424 struct extent_buffer *tmp;
2425 struct btrfs_key first_key;
2426 int ret;
2427 int parent_level;
2428
2429 blocknr = btrfs_node_blockptr(b, slot);
2430 gen = btrfs_node_ptr_generation(b, slot);
2431 parent_level = btrfs_header_level(b);
2432 btrfs_node_key_to_cpu(b, &first_key, slot);
2433
2434 tmp = find_extent_buffer(fs_info, blocknr);
2435 if (tmp) {
2436 /* first we do an atomic uptodate check */
2437 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2438 /*
2439 * Do extra check for first_key, eb can be stale due to
2440 * being cached, read from scrub, or have multiple
2441 * parents (shared tree blocks).
2442 */
2443 if (btrfs_verify_level_key(tmp,
2444 parent_level - 1, &first_key, gen)) {
2445 free_extent_buffer(tmp);
2446 return -EUCLEAN;
2447 }
2448 *eb_ret = tmp;
2449 return 0;
2450 }
2451
2452 /* the pages were up to date, but we failed
2453 * the generation number check. Do a full
2454 * read for the generation number that is correct.
2455 * We must do this without dropping locks so
2456 * we can trust our generation number
2457 */
2458 btrfs_set_path_blocking(p);
2459
2460 /* now we're allowed to do a blocking uptodate check */
2461 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2462 if (!ret) {
2463 *eb_ret = tmp;
2464 return 0;
2465 }
2466 free_extent_buffer(tmp);
2467 btrfs_release_path(p);
2468 return -EIO;
2469 }
2470
2471 /*
2472 * reduce lock contention at high levels
2473 * of the btree by dropping locks before
2474 * we read. Don't release the lock on the current
2475 * level because we need to walk this node to figure
2476 * out which blocks to read.
2477 */
2478 btrfs_unlock_up_safe(p, level + 1);
2479 btrfs_set_path_blocking(p);
2480
2481 if (p->reada != READA_NONE)
2482 reada_for_search(fs_info, p, level, slot, key->objectid);
2483
2484 ret = -EAGAIN;
2485 tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2486 &first_key);
2487 if (!IS_ERR(tmp)) {
2488 /*
2489 * If the read above didn't mark this buffer up to date,
2490 * it will never end up being up to date. Set ret to EIO now
2491 * and give up so that our caller doesn't loop forever
2492 * on our EAGAINs.
2493 */
2494 if (!extent_buffer_uptodate(tmp))
2495 ret = -EIO;
2496 free_extent_buffer(tmp);
2497 } else {
2498 ret = PTR_ERR(tmp);
2499 }
2500
2501 btrfs_release_path(p);
2502 return ret;
2503}
2504
2505/*
2506 * helper function for btrfs_search_slot. This does all of the checks
2507 * for node-level blocks and does any balancing required based on
2508 * the ins_len.
2509 *
2510 * If no extra work was required, zero is returned. If we had to
2511 * drop the path, -EAGAIN is returned and btrfs_search_slot must
2512 * start over
2513 */
2514static int
2515setup_nodes_for_search(struct btrfs_trans_handle *trans,
2516 struct btrfs_root *root, struct btrfs_path *p,
2517 struct extent_buffer *b, int level, int ins_len,
2518 int *write_lock_level)
2519{
2520 struct btrfs_fs_info *fs_info = root->fs_info;
2521 int ret;
2522
2523 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2524 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
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 = split_node(trans, root, p, level);
2536
2537 BUG_ON(sret > 0);
2538 if (sret) {
2539 ret = sret;
2540 goto done;
2541 }
2542 b = p->nodes[level];
2543 } else if (ins_len < 0 && btrfs_header_nritems(b) <
2544 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2545 int sret;
2546
2547 if (*write_lock_level < level + 1) {
2548 *write_lock_level = level + 1;
2549 btrfs_release_path(p);
2550 goto again;
2551 }
2552
2553 btrfs_set_path_blocking(p);
2554 reada_for_balance(fs_info, p, level);
2555 sret = balance_level(trans, root, p, level);
2556
2557 if (sret) {
2558 ret = sret;
2559 goto done;
2560 }
2561 b = p->nodes[level];
2562 if (!b) {
2563 btrfs_release_path(p);
2564 goto again;
2565 }
2566 BUG_ON(btrfs_header_nritems(b) == 1);
2567 }
2568 return 0;
2569
2570again:
2571 ret = -EAGAIN;
2572done:
2573 return ret;
2574}
2575
2576static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2577 int level, int *prev_cmp, int *slot)
2578{
2579 if (*prev_cmp != 0) {
2580 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2581 return *prev_cmp;
2582 }
2583
2584 *slot = 0;
2585
2586 return 0;
2587}
2588
2589int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2590 u64 iobjectid, u64 ioff, u8 key_type,
2591 struct btrfs_key *found_key)
2592{
2593 int ret;
2594 struct btrfs_key key;
2595 struct extent_buffer *eb;
2596
2597 ASSERT(path);
2598 ASSERT(found_key);
2599
2600 key.type = key_type;
2601 key.objectid = iobjectid;
2602 key.offset = ioff;
2603
2604 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2605 if (ret < 0)
2606 return ret;
2607
2608 eb = path->nodes[0];
2609 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2610 ret = btrfs_next_leaf(fs_root, path);
2611 if (ret)
2612 return ret;
2613 eb = path->nodes[0];
2614 }
2615
2616 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2617 if (found_key->type != key.type ||
2618 found_key->objectid != key.objectid)
2619 return 1;
2620
2621 return 0;
2622}
2623
2624static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2625 struct btrfs_path *p,
2626 int write_lock_level)
2627{
2628 struct btrfs_fs_info *fs_info = root->fs_info;
2629 struct extent_buffer *b;
2630 int root_lock;
2631 int level = 0;
2632
2633 /* We try very hard to do read locks on the root */
2634 root_lock = BTRFS_READ_LOCK;
2635
2636 if (p->search_commit_root) {
2637 /*
2638 * The commit roots are read only so we always do read locks,
2639 * and we always must hold the commit_root_sem when doing
2640 * searches on them, the only exception is send where we don't
2641 * want to block transaction commits for a long time, so
2642 * we need to clone the commit root in order to avoid races
2643 * with transaction commits that create a snapshot of one of
2644 * the roots used by a send operation.
2645 */
2646 if (p->need_commit_sem) {
2647 down_read(&fs_info->commit_root_sem);
2648 b = btrfs_clone_extent_buffer(root->commit_root);
2649 up_read(&fs_info->commit_root_sem);
2650 if (!b)
2651 return ERR_PTR(-ENOMEM);
2652
2653 } else {
2654 b = root->commit_root;
2655 extent_buffer_get(b);
2656 }
2657 level = btrfs_header_level(b);
2658 /*
2659 * Ensure that all callers have set skip_locking when
2660 * p->search_commit_root = 1.
2661 */
2662 ASSERT(p->skip_locking == 1);
2663
2664 goto out;
2665 }
2666
2667 if (p->skip_locking) {
2668 b = btrfs_root_node(root);
2669 level = btrfs_header_level(b);
2670 goto out;
2671 }
2672
2673 /*
2674 * If the level is set to maximum, we can skip trying to get the read
2675 * lock.
2676 */
2677 if (write_lock_level < BTRFS_MAX_LEVEL) {
2678 /*
2679 * We don't know the level of the root node until we actually
2680 * have it read locked
2681 */
2682 b = btrfs_read_lock_root_node(root);
2683 level = btrfs_header_level(b);
2684 if (level > write_lock_level)
2685 goto out;
2686
2687 /* Whoops, must trade for write lock */
2688 btrfs_tree_read_unlock(b);
2689 free_extent_buffer(b);
2690 }
2691
2692 b = btrfs_lock_root_node(root);
2693 root_lock = BTRFS_WRITE_LOCK;
2694
2695 /* The level might have changed, check again */
2696 level = btrfs_header_level(b);
2697
2698out:
2699 p->nodes[level] = b;
2700 if (!p->skip_locking)
2701 p->locks[level] = root_lock;
2702 /*
2703 * Callers are responsible for dropping b's references.
2704 */
2705 return b;
2706}
2707
2708
2709/*
2710 * btrfs_search_slot - look for a key in a tree and perform necessary
2711 * modifications to preserve tree invariants.
2712 *
2713 * @trans: Handle of transaction, used when modifying the tree
2714 * @p: Holds all btree nodes along the search path
2715 * @root: The root node of the tree
2716 * @key: The key we are looking for
2717 * @ins_len: Indicates purpose of search, for inserts it is 1, for
2718 * deletions it's -1. 0 for plain searches
2719 * @cow: boolean should CoW operations be performed. Must always be 1
2720 * when modifying the tree.
2721 *
2722 * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2723 * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2724 *
2725 * If @key is found, 0 is returned and you can find the item in the leaf level
2726 * of the path (level 0)
2727 *
2728 * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2729 * points to the slot where it should be inserted
2730 *
2731 * If an error is encountered while searching the tree a negative error number
2732 * is returned
2733 */
2734int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2735 const struct btrfs_key *key, struct btrfs_path *p,
2736 int ins_len, int cow)
2737{
2738 struct extent_buffer *b;
2739 int slot;
2740 int ret;
2741 int err;
2742 int level;
2743 int lowest_unlock = 1;
2744 /* everything at write_lock_level or lower must be write locked */
2745 int write_lock_level = 0;
2746 u8 lowest_level = 0;
2747 int min_write_lock_level;
2748 int prev_cmp;
2749
2750 lowest_level = p->lowest_level;
2751 WARN_ON(lowest_level && ins_len > 0);
2752 WARN_ON(p->nodes[0] != NULL);
2753 BUG_ON(!cow && ins_len);
2754
2755 if (ins_len < 0) {
2756 lowest_unlock = 2;
2757
2758 /* when we are removing items, we might have to go up to level
2759 * two as we update tree pointers Make sure we keep write
2760 * for those levels as well
2761 */
2762 write_lock_level = 2;
2763 } else if (ins_len > 0) {
2764 /*
2765 * for inserting items, make sure we have a write lock on
2766 * level 1 so we can update keys
2767 */
2768 write_lock_level = 1;
2769 }
2770
2771 if (!cow)
2772 write_lock_level = -1;
2773
2774 if (cow && (p->keep_locks || p->lowest_level))
2775 write_lock_level = BTRFS_MAX_LEVEL;
2776
2777 min_write_lock_level = write_lock_level;
2778
2779again:
2780 prev_cmp = -1;
2781 b = btrfs_search_slot_get_root(root, p, write_lock_level);
2782 if (IS_ERR(b)) {
2783 ret = PTR_ERR(b);
2784 goto done;
2785 }
2786
2787 while (b) {
2788 level = btrfs_header_level(b);
2789
2790 /*
2791 * setup the path here so we can release it under lock
2792 * contention with the cow code
2793 */
2794 if (cow) {
2795 bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2796
2797 /*
2798 * if we don't really need to cow this block
2799 * then we don't want to set the path blocking,
2800 * so we test it here
2801 */
2802 if (!should_cow_block(trans, root, b)) {
2803 trans->dirty = true;
2804 goto cow_done;
2805 }
2806
2807 /*
2808 * must have write locks on this node and the
2809 * parent
2810 */
2811 if (level > write_lock_level ||
2812 (level + 1 > write_lock_level &&
2813 level + 1 < BTRFS_MAX_LEVEL &&
2814 p->nodes[level + 1])) {
2815 write_lock_level = level + 1;
2816 btrfs_release_path(p);
2817 goto again;
2818 }
2819
2820 btrfs_set_path_blocking(p);
2821 if (last_level)
2822 err = btrfs_cow_block(trans, root, b, NULL, 0,
2823 &b);
2824 else
2825 err = btrfs_cow_block(trans, root, b,
2826 p->nodes[level + 1],
2827 p->slots[level + 1], &b);
2828 if (err) {
2829 ret = err;
2830 goto done;
2831 }
2832 }
2833cow_done:
2834 p->nodes[level] = b;
2835 /*
2836 * Leave path with blocking locks to avoid massive
2837 * lock context switch, this is made on purpose.
2838 */
2839
2840 /*
2841 * we have a lock on b and as long as we aren't changing
2842 * the tree, there is no way to for the items in b to change.
2843 * It is safe to drop the lock on our parent before we
2844 * go through the expensive btree search on b.
2845 *
2846 * If we're inserting or deleting (ins_len != 0), then we might
2847 * be changing slot zero, which may require changing the parent.
2848 * So, we can't drop the lock until after we know which slot
2849 * we're operating on.
2850 */
2851 if (!ins_len && !p->keep_locks) {
2852 int u = level + 1;
2853
2854 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2855 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2856 p->locks[u] = 0;
2857 }
2858 }
2859
2860 ret = key_search(b, key, level, &prev_cmp, &slot);
2861 if (ret < 0)
2862 goto done;
2863
2864 if (level != 0) {
2865 int dec = 0;
2866 if (ret && slot > 0) {
2867 dec = 1;
2868 slot -= 1;
2869 }
2870 p->slots[level] = slot;
2871 err = setup_nodes_for_search(trans, root, p, b, level,
2872 ins_len, &write_lock_level);
2873 if (err == -EAGAIN)
2874 goto again;
2875 if (err) {
2876 ret = err;
2877 goto done;
2878 }
2879 b = p->nodes[level];
2880 slot = p->slots[level];
2881
2882 /*
2883 * slot 0 is special, if we change the key
2884 * we have to update the parent pointer
2885 * which means we must have a write lock
2886 * on the parent
2887 */
2888 if (slot == 0 && ins_len &&
2889 write_lock_level < level + 1) {
2890 write_lock_level = level + 1;
2891 btrfs_release_path(p);
2892 goto again;
2893 }
2894
2895 unlock_up(p, level, lowest_unlock,
2896 min_write_lock_level, &write_lock_level);
2897
2898 if (level == lowest_level) {
2899 if (dec)
2900 p->slots[level]++;
2901 goto done;
2902 }
2903
2904 err = read_block_for_search(root, p, &b, level,
2905 slot, key);
2906 if (err == -EAGAIN)
2907 goto again;
2908 if (err) {
2909 ret = err;
2910 goto done;
2911 }
2912
2913 if (!p->skip_locking) {
2914 level = btrfs_header_level(b);
2915 if (level <= write_lock_level) {
2916 if (!btrfs_try_tree_write_lock(b)) {
2917 btrfs_set_path_blocking(p);
2918 btrfs_tree_lock(b);
2919 }
2920 p->locks[level] = BTRFS_WRITE_LOCK;
2921 } else {
2922 if (!btrfs_tree_read_lock_atomic(b)) {
2923 btrfs_set_path_blocking(p);
2924 btrfs_tree_read_lock(b);
2925 }
2926 p->locks[level] = BTRFS_READ_LOCK;
2927 }
2928 p->nodes[level] = b;
2929 }
2930 } else {
2931 p->slots[level] = slot;
2932 if (ins_len > 0 &&
2933 btrfs_leaf_free_space(b) < ins_len) {
2934 if (write_lock_level < 1) {
2935 write_lock_level = 1;
2936 btrfs_release_path(p);
2937 goto again;
2938 }
2939
2940 btrfs_set_path_blocking(p);
2941 err = split_leaf(trans, root, key,
2942 p, ins_len, ret == 0);
2943
2944 BUG_ON(err > 0);
2945 if (err) {
2946 ret = err;
2947 goto done;
2948 }
2949 }
2950 if (!p->search_for_split)
2951 unlock_up(p, level, lowest_unlock,
2952 min_write_lock_level, NULL);
2953 goto done;
2954 }
2955 }
2956 ret = 1;
2957done:
2958 /*
2959 * we don't really know what they plan on doing with the path
2960 * from here on, so for now just mark it as blocking
2961 */
2962 if (!p->leave_spinning)
2963 btrfs_set_path_blocking(p);
2964 if (ret < 0 && !p->skip_release_on_error)
2965 btrfs_release_path(p);
2966 return ret;
2967}
2968
2969/*
2970 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2971 * current state of the tree together with the operations recorded in the tree
2972 * modification log to search for the key in a previous version of this tree, as
2973 * denoted by the time_seq parameter.
2974 *
2975 * Naturally, there is no support for insert, delete or cow operations.
2976 *
2977 * The resulting path and return value will be set up as if we called
2978 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2979 */
2980int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2981 struct btrfs_path *p, u64 time_seq)
2982{
2983 struct btrfs_fs_info *fs_info = root->fs_info;
2984 struct extent_buffer *b;
2985 int slot;
2986 int ret;
2987 int err;
2988 int level;
2989 int lowest_unlock = 1;
2990 u8 lowest_level = 0;
2991 int prev_cmp = -1;
2992
2993 lowest_level = p->lowest_level;
2994 WARN_ON(p->nodes[0] != NULL);
2995
2996 if (p->search_commit_root) {
2997 BUG_ON(time_seq);
2998 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2999 }
3000
3001again:
3002 b = get_old_root(root, time_seq);
3003 if (!b) {
3004 ret = -EIO;
3005 goto done;
3006 }
3007 level = btrfs_header_level(b);
3008 p->locks[level] = BTRFS_READ_LOCK;
3009
3010 while (b) {
3011 level = btrfs_header_level(b);
3012 p->nodes[level] = b;
3013
3014 /*
3015 * we have a lock on b and as long as we aren't changing
3016 * the tree, there is no way to for the items in b to change.
3017 * It is safe to drop the lock on our parent before we
3018 * go through the expensive btree search on b.
3019 */
3020 btrfs_unlock_up_safe(p, level + 1);
3021
3022 /*
3023 * Since we can unwind ebs we want to do a real search every
3024 * time.
3025 */
3026 prev_cmp = -1;
3027 ret = key_search(b, key, level, &prev_cmp, &slot);
3028 if (ret < 0)
3029 goto done;
3030
3031 if (level != 0) {
3032 int dec = 0;
3033 if (ret && slot > 0) {
3034 dec = 1;
3035 slot -= 1;
3036 }
3037 p->slots[level] = slot;
3038 unlock_up(p, level, lowest_unlock, 0, NULL);
3039
3040 if (level == lowest_level) {
3041 if (dec)
3042 p->slots[level]++;
3043 goto done;
3044 }
3045
3046 err = read_block_for_search(root, p, &b, level,
3047 slot, key);
3048 if (err == -EAGAIN)
3049 goto again;
3050 if (err) {
3051 ret = err;
3052 goto done;
3053 }
3054
3055 level = btrfs_header_level(b);
3056 if (!btrfs_tree_read_lock_atomic(b)) {
3057 btrfs_set_path_blocking(p);
3058 btrfs_tree_read_lock(b);
3059 }
3060 b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3061 if (!b) {
3062 ret = -ENOMEM;
3063 goto done;
3064 }
3065 p->locks[level] = BTRFS_READ_LOCK;
3066 p->nodes[level] = b;
3067 } else {
3068 p->slots[level] = slot;
3069 unlock_up(p, level, lowest_unlock, 0, NULL);
3070 goto done;
3071 }
3072 }
3073 ret = 1;
3074done:
3075 if (!p->leave_spinning)
3076 btrfs_set_path_blocking(p);
3077 if (ret < 0)
3078 btrfs_release_path(p);
3079
3080 return ret;
3081}
3082
3083/*
3084 * helper to use instead of search slot if no exact match is needed but
3085 * instead the next or previous item should be returned.
3086 * When find_higher is true, the next higher item is returned, the next lower
3087 * otherwise.
3088 * When return_any and find_higher are both true, and no higher item is found,
3089 * return the next lower instead.
3090 * When return_any is true and find_higher is false, and no lower item is found,
3091 * return the next higher instead.
3092 * It returns 0 if any item is found, 1 if none is found (tree empty), and
3093 * < 0 on error
3094 */
3095int btrfs_search_slot_for_read(struct btrfs_root *root,
3096 const struct btrfs_key *key,
3097 struct btrfs_path *p, int find_higher,
3098 int return_any)
3099{
3100 int ret;
3101 struct extent_buffer *leaf;
3102
3103again:
3104 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3105 if (ret <= 0)
3106 return ret;
3107 /*
3108 * a return value of 1 means the path is at the position where the
3109 * item should be inserted. Normally this is the next bigger item,
3110 * but in case the previous item is the last in a leaf, path points
3111 * to the first free slot in the previous leaf, i.e. at an invalid
3112 * item.
3113 */
3114 leaf = p->nodes[0];
3115
3116 if (find_higher) {
3117 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3118 ret = btrfs_next_leaf(root, p);
3119 if (ret <= 0)
3120 return ret;
3121 if (!return_any)
3122 return 1;
3123 /*
3124 * no higher item found, return the next
3125 * lower instead
3126 */
3127 return_any = 0;
3128 find_higher = 0;
3129 btrfs_release_path(p);
3130 goto again;
3131 }
3132 } else {
3133 if (p->slots[0] == 0) {
3134 ret = btrfs_prev_leaf(root, p);
3135 if (ret < 0)
3136 return ret;
3137 if (!ret) {
3138 leaf = p->nodes[0];
3139 if (p->slots[0] == btrfs_header_nritems(leaf))
3140 p->slots[0]--;
3141 return 0;
3142 }
3143 if (!return_any)
3144 return 1;
3145 /*
3146 * no lower item found, return the next
3147 * higher instead
3148 */
3149 return_any = 0;
3150 find_higher = 1;
3151 btrfs_release_path(p);
3152 goto again;
3153 } else {
3154 --p->slots[0];
3155 }
3156 }
3157 return 0;
3158}
3159
3160/*
3161 * adjust the pointers going up the tree, starting at level
3162 * making sure the right key of each node is points to 'key'.
3163 * This is used after shifting pointers to the left, so it stops
3164 * fixing up pointers when a given leaf/node is not in slot 0 of the
3165 * higher levels
3166 *
3167 */
3168static void fixup_low_keys(struct btrfs_path *path,
3169 struct btrfs_disk_key *key, int level)
3170{
3171 int i;
3172 struct extent_buffer *t;
3173 int ret;
3174
3175 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3176 int tslot = path->slots[i];
3177
3178 if (!path->nodes[i])
3179 break;
3180 t = path->nodes[i];
3181 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3182 GFP_ATOMIC);
3183 BUG_ON(ret < 0);
3184 btrfs_set_node_key(t, key, tslot);
3185 btrfs_mark_buffer_dirty(path->nodes[i]);
3186 if (tslot != 0)
3187 break;
3188 }
3189}
3190
3191/*
3192 * update item key.
3193 *
3194 * This function isn't completely safe. It's the caller's responsibility
3195 * that the new key won't break the order
3196 */
3197void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3198 struct btrfs_path *path,
3199 const struct btrfs_key *new_key)
3200{
3201 struct btrfs_disk_key disk_key;
3202 struct extent_buffer *eb;
3203 int slot;
3204
3205 eb = path->nodes[0];
3206 slot = path->slots[0];
3207 if (slot > 0) {
3208 btrfs_item_key(eb, &disk_key, slot - 1);
3209 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
3210 btrfs_crit(fs_info,
3211 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3212 slot, btrfs_disk_key_objectid(&disk_key),
3213 btrfs_disk_key_type(&disk_key),
3214 btrfs_disk_key_offset(&disk_key),
3215 new_key->objectid, new_key->type,
3216 new_key->offset);
3217 btrfs_print_leaf(eb);
3218 BUG();
3219 }
3220 }
3221 if (slot < btrfs_header_nritems(eb) - 1) {
3222 btrfs_item_key(eb, &disk_key, slot + 1);
3223 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
3224 btrfs_crit(fs_info,
3225 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3226 slot, btrfs_disk_key_objectid(&disk_key),
3227 btrfs_disk_key_type(&disk_key),
3228 btrfs_disk_key_offset(&disk_key),
3229 new_key->objectid, new_key->type,
3230 new_key->offset);
3231 btrfs_print_leaf(eb);
3232 BUG();
3233 }
3234 }
3235
3236 btrfs_cpu_key_to_disk(&disk_key, new_key);
3237 btrfs_set_item_key(eb, &disk_key, slot);
3238 btrfs_mark_buffer_dirty(eb);
3239 if (slot == 0)
3240 fixup_low_keys(path, &disk_key, 1);
3241}
3242
3243/*
3244 * try to push data from one node into the next node left in the
3245 * tree.
3246 *
3247 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3248 * error, and > 0 if there was no room in the left hand block.
3249 */
3250static int push_node_left(struct btrfs_trans_handle *trans,
3251 struct extent_buffer *dst,
3252 struct extent_buffer *src, int empty)
3253{
3254 struct btrfs_fs_info *fs_info = trans->fs_info;
3255 int push_items = 0;
3256 int src_nritems;
3257 int dst_nritems;
3258 int ret = 0;
3259
3260 src_nritems = btrfs_header_nritems(src);
3261 dst_nritems = btrfs_header_nritems(dst);
3262 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3263 WARN_ON(btrfs_header_generation(src) != trans->transid);
3264 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3265
3266 if (!empty && src_nritems <= 8)
3267 return 1;
3268
3269 if (push_items <= 0)
3270 return 1;
3271
3272 if (empty) {
3273 push_items = min(src_nritems, push_items);
3274 if (push_items < src_nritems) {
3275 /* leave at least 8 pointers in the node if
3276 * we aren't going to empty it
3277 */
3278 if (src_nritems - push_items < 8) {
3279 if (push_items <= 8)
3280 return 1;
3281 push_items -= 8;
3282 }
3283 }
3284 } else
3285 push_items = min(src_nritems - 8, push_items);
3286
3287 ret = tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
3288 if (ret) {
3289 btrfs_abort_transaction(trans, ret);
3290 return ret;
3291 }
3292 copy_extent_buffer(dst, src,
3293 btrfs_node_key_ptr_offset(dst_nritems),
3294 btrfs_node_key_ptr_offset(0),
3295 push_items * sizeof(struct btrfs_key_ptr));
3296
3297 if (push_items < src_nritems) {
3298 /*
3299 * Don't call tree_mod_log_insert_move here, key removal was
3300 * already fully logged by tree_mod_log_eb_copy above.
3301 */
3302 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3303 btrfs_node_key_ptr_offset(push_items),
3304 (src_nritems - push_items) *
3305 sizeof(struct btrfs_key_ptr));
3306 }
3307 btrfs_set_header_nritems(src, src_nritems - push_items);
3308 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3309 btrfs_mark_buffer_dirty(src);
3310 btrfs_mark_buffer_dirty(dst);
3311
3312 return ret;
3313}
3314
3315/*
3316 * try to push data from one node into the next node right in the
3317 * tree.
3318 *
3319 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3320 * error, and > 0 if there was no room in the right hand block.
3321 *
3322 * this will only push up to 1/2 the contents of the left node over
3323 */
3324static int balance_node_right(struct btrfs_trans_handle *trans,
3325 struct extent_buffer *dst,
3326 struct extent_buffer *src)
3327{
3328 struct btrfs_fs_info *fs_info = trans->fs_info;
3329 int push_items = 0;
3330 int max_push;
3331 int src_nritems;
3332 int dst_nritems;
3333 int ret = 0;
3334
3335 WARN_ON(btrfs_header_generation(src) != trans->transid);
3336 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3337
3338 src_nritems = btrfs_header_nritems(src);
3339 dst_nritems = btrfs_header_nritems(dst);
3340 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3341 if (push_items <= 0)
3342 return 1;
3343
3344 if (src_nritems < 4)
3345 return 1;
3346
3347 max_push = src_nritems / 2 + 1;
3348 /* don't try to empty the node */
3349 if (max_push >= src_nritems)
3350 return 1;
3351
3352 if (max_push < push_items)
3353 push_items = max_push;
3354
3355 ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3356 BUG_ON(ret < 0);
3357 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3358 btrfs_node_key_ptr_offset(0),
3359 (dst_nritems) *
3360 sizeof(struct btrfs_key_ptr));
3361
3362 ret = tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
3363 push_items);
3364 if (ret) {
3365 btrfs_abort_transaction(trans, ret);
3366 return ret;
3367 }
3368 copy_extent_buffer(dst, src,
3369 btrfs_node_key_ptr_offset(0),
3370 btrfs_node_key_ptr_offset(src_nritems - push_items),
3371 push_items * sizeof(struct btrfs_key_ptr));
3372
3373 btrfs_set_header_nritems(src, src_nritems - push_items);
3374 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3375
3376 btrfs_mark_buffer_dirty(src);
3377 btrfs_mark_buffer_dirty(dst);
3378
3379 return ret;
3380}
3381
3382/*
3383 * helper function to insert a new root level in the tree.
3384 * A new node is allocated, and a single item is inserted to
3385 * point to the existing root
3386 *
3387 * returns zero on success or < 0 on failure.
3388 */
3389static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3390 struct btrfs_root *root,
3391 struct btrfs_path *path, int level)
3392{
3393 struct btrfs_fs_info *fs_info = root->fs_info;
3394 u64 lower_gen;
3395 struct extent_buffer *lower;
3396 struct extent_buffer *c;
3397 struct extent_buffer *old;
3398 struct btrfs_disk_key lower_key;
3399 int ret;
3400
3401 BUG_ON(path->nodes[level]);
3402 BUG_ON(path->nodes[level-1] != root->node);
3403
3404 lower = path->nodes[level-1];
3405 if (level == 1)
3406 btrfs_item_key(lower, &lower_key, 0);
3407 else
3408 btrfs_node_key(lower, &lower_key, 0);
3409
3410 c = alloc_tree_block_no_bg_flush(trans, root, 0, &lower_key, level,
3411 root->node->start, 0);
3412 if (IS_ERR(c))
3413 return PTR_ERR(c);
3414
3415 root_add_used(root, fs_info->nodesize);
3416
3417 btrfs_set_header_nritems(c, 1);
3418 btrfs_set_node_key(c, &lower_key, 0);
3419 btrfs_set_node_blockptr(c, 0, lower->start);
3420 lower_gen = btrfs_header_generation(lower);
3421 WARN_ON(lower_gen != trans->transid);
3422
3423 btrfs_set_node_ptr_generation(c, 0, lower_gen);
3424
3425 btrfs_mark_buffer_dirty(c);
3426
3427 old = root->node;
3428 ret = tree_mod_log_insert_root(root->node, c, 0);
3429 BUG_ON(ret < 0);
3430 rcu_assign_pointer(root->node, c);
3431
3432 /* the super has an extra ref to root->node */
3433 free_extent_buffer(old);
3434
3435 add_root_to_dirty_list(root);
3436 extent_buffer_get(c);
3437 path->nodes[level] = c;
3438 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3439 path->slots[level] = 0;
3440 return 0;
3441}
3442
3443/*
3444 * worker function to insert a single pointer in a node.
3445 * the node should have enough room for the pointer already
3446 *
3447 * slot and level indicate where you want the key to go, and
3448 * blocknr is the block the key points to.
3449 */
3450static void insert_ptr(struct btrfs_trans_handle *trans,
3451 struct btrfs_path *path,
3452 struct btrfs_disk_key *key, u64 bytenr,
3453 int slot, int level)
3454{
3455 struct extent_buffer *lower;
3456 int nritems;
3457 int ret;
3458
3459 BUG_ON(!path->nodes[level]);
3460 btrfs_assert_tree_locked(path->nodes[level]);
3461 lower = path->nodes[level];
3462 nritems = btrfs_header_nritems(lower);
3463 BUG_ON(slot > nritems);
3464 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
3465 if (slot != nritems) {
3466 if (level) {
3467 ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3468 nritems - slot);
3469 BUG_ON(ret < 0);
3470 }
3471 memmove_extent_buffer(lower,
3472 btrfs_node_key_ptr_offset(slot + 1),
3473 btrfs_node_key_ptr_offset(slot),
3474 (nritems - slot) * sizeof(struct btrfs_key_ptr));
3475 }
3476 if (level) {
3477 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3478 GFP_NOFS);
3479 BUG_ON(ret < 0);
3480 }
3481 btrfs_set_node_key(lower, key, slot);
3482 btrfs_set_node_blockptr(lower, slot, bytenr);
3483 WARN_ON(trans->transid == 0);
3484 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3485 btrfs_set_header_nritems(lower, nritems + 1);
3486 btrfs_mark_buffer_dirty(lower);
3487}
3488
3489/*
3490 * split the node at the specified level in path in two.
3491 * The path is corrected to point to the appropriate node after the split
3492 *
3493 * Before splitting this tries to make some room in the node by pushing
3494 * left and right, if either one works, it returns right away.
3495 *
3496 * returns 0 on success and < 0 on failure
3497 */
3498static noinline int split_node(struct btrfs_trans_handle *trans,
3499 struct btrfs_root *root,
3500 struct btrfs_path *path, int level)
3501{
3502 struct btrfs_fs_info *fs_info = root->fs_info;
3503 struct extent_buffer *c;
3504 struct extent_buffer *split;
3505 struct btrfs_disk_key disk_key;
3506 int mid;
3507 int ret;
3508 u32 c_nritems;
3509
3510 c = path->nodes[level];
3511 WARN_ON(btrfs_header_generation(c) != trans->transid);
3512 if (c == root->node) {
3513 /*
3514 * trying to split the root, lets make a new one
3515 *
3516 * tree mod log: We don't log_removal old root in
3517 * insert_new_root, because that root buffer will be kept as a
3518 * normal node. We are going to log removal of half of the
3519 * elements below with tree_mod_log_eb_copy. We're holding a
3520 * tree lock on the buffer, which is why we cannot race with
3521 * other tree_mod_log users.
3522 */
3523 ret = insert_new_root(trans, root, path, level + 1);
3524 if (ret)
3525 return ret;
3526 } else {
3527 ret = push_nodes_for_insert(trans, root, path, level);
3528 c = path->nodes[level];
3529 if (!ret && btrfs_header_nritems(c) <
3530 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3531 return 0;
3532 if (ret < 0)
3533 return ret;
3534 }
3535
3536 c_nritems = btrfs_header_nritems(c);
3537 mid = (c_nritems + 1) / 2;
3538 btrfs_node_key(c, &disk_key, mid);
3539
3540 split = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, level,
3541 c->start, 0);
3542 if (IS_ERR(split))
3543 return PTR_ERR(split);
3544
3545 root_add_used(root, fs_info->nodesize);
3546 ASSERT(btrfs_header_level(c) == level);
3547
3548 ret = tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
3549 if (ret) {
3550 btrfs_abort_transaction(trans, ret);
3551 return ret;
3552 }
3553 copy_extent_buffer(split, c,
3554 btrfs_node_key_ptr_offset(0),
3555 btrfs_node_key_ptr_offset(mid),
3556 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3557 btrfs_set_header_nritems(split, c_nritems - mid);
3558 btrfs_set_header_nritems(c, mid);
3559 ret = 0;
3560
3561 btrfs_mark_buffer_dirty(c);
3562 btrfs_mark_buffer_dirty(split);
3563
3564 insert_ptr(trans, path, &disk_key, split->start,
3565 path->slots[level + 1] + 1, level + 1);
3566
3567 if (path->slots[level] >= mid) {
3568 path->slots[level] -= mid;
3569 btrfs_tree_unlock(c);
3570 free_extent_buffer(c);
3571 path->nodes[level] = split;
3572 path->slots[level + 1] += 1;
3573 } else {
3574 btrfs_tree_unlock(split);
3575 free_extent_buffer(split);
3576 }
3577 return ret;
3578}
3579
3580/*
3581 * how many bytes are required to store the items in a leaf. start
3582 * and nr indicate which items in the leaf to check. This totals up the
3583 * space used both by the item structs and the item data
3584 */
3585static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3586{
3587 struct btrfs_item *start_item;
3588 struct btrfs_item *end_item;
3589 struct btrfs_map_token token;
3590 int data_len;
3591 int nritems = btrfs_header_nritems(l);
3592 int end = min(nritems, start + nr) - 1;
3593
3594 if (!nr)
3595 return 0;
3596 btrfs_init_map_token(&token, l);
3597 start_item = btrfs_item_nr(start);
3598 end_item = btrfs_item_nr(end);
3599 data_len = btrfs_token_item_offset(l, start_item, &token) +
3600 btrfs_token_item_size(l, start_item, &token);
3601 data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3602 data_len += sizeof(struct btrfs_item) * nr;
3603 WARN_ON(data_len < 0);
3604 return data_len;
3605}
3606
3607/*
3608 * The space between the end of the leaf items and
3609 * the start of the leaf data. IOW, how much room
3610 * the leaf has left for both items and data
3611 */
3612noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
3613{
3614 struct btrfs_fs_info *fs_info = leaf->fs_info;
3615 int nritems = btrfs_header_nritems(leaf);
3616 int ret;
3617
3618 ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3619 if (ret < 0) {
3620 btrfs_crit(fs_info,
3621 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3622 ret,
3623 (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3624 leaf_space_used(leaf, 0, nritems), nritems);
3625 }
3626 return ret;
3627}
3628
3629/*
3630 * min slot controls the lowest index we're willing to push to the
3631 * right. We'll push up to and including min_slot, but no lower
3632 */
3633static noinline int __push_leaf_right(struct btrfs_path *path,
3634 int data_size, int empty,
3635 struct extent_buffer *right,
3636 int free_space, u32 left_nritems,
3637 u32 min_slot)
3638{
3639 struct btrfs_fs_info *fs_info = right->fs_info;
3640 struct extent_buffer *left = path->nodes[0];
3641 struct extent_buffer *upper = path->nodes[1];
3642 struct btrfs_map_token token;
3643 struct btrfs_disk_key disk_key;
3644 int slot;
3645 u32 i;
3646 int push_space = 0;
3647 int push_items = 0;
3648 struct btrfs_item *item;
3649 u32 nr;
3650 u32 right_nritems;
3651 u32 data_end;
3652 u32 this_item_size;
3653
3654 if (empty)
3655 nr = 0;
3656 else
3657 nr = max_t(u32, 1, min_slot);
3658
3659 if (path->slots[0] >= left_nritems)
3660 push_space += data_size;
3661
3662 slot = path->slots[1];
3663 i = left_nritems - 1;
3664 while (i >= nr) {
3665 item = btrfs_item_nr(i);
3666
3667 if (!empty && push_items > 0) {
3668 if (path->slots[0] > i)
3669 break;
3670 if (path->slots[0] == i) {
3671 int space = btrfs_leaf_free_space(left);
3672
3673 if (space + push_space * 2 > free_space)
3674 break;
3675 }
3676 }
3677
3678 if (path->slots[0] == i)
3679 push_space += data_size;
3680
3681 this_item_size = btrfs_item_size(left, item);
3682 if (this_item_size + sizeof(*item) + push_space > free_space)
3683 break;
3684
3685 push_items++;
3686 push_space += this_item_size + sizeof(*item);
3687 if (i == 0)
3688 break;
3689 i--;
3690 }
3691
3692 if (push_items == 0)
3693 goto out_unlock;
3694
3695 WARN_ON(!empty && push_items == left_nritems);
3696
3697 /* push left to right */
3698 right_nritems = btrfs_header_nritems(right);
3699
3700 push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3701 push_space -= leaf_data_end(left);
3702
3703 /* make room in the right data area */
3704 data_end = leaf_data_end(right);
3705 memmove_extent_buffer(right,
3706 BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
3707 BTRFS_LEAF_DATA_OFFSET + data_end,
3708 BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3709
3710 /* copy from the left data area */
3711 copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
3712 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3713 BTRFS_LEAF_DATA_OFFSET + leaf_data_end(left),
3714 push_space);
3715
3716 memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3717 btrfs_item_nr_offset(0),
3718 right_nritems * sizeof(struct btrfs_item));
3719
3720 /* copy the items from left to right */
3721 copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3722 btrfs_item_nr_offset(left_nritems - push_items),
3723 push_items * sizeof(struct btrfs_item));
3724
3725 /* update the item pointers */
3726 btrfs_init_map_token(&token, right);
3727 right_nritems += push_items;
3728 btrfs_set_header_nritems(right, right_nritems);
3729 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3730 for (i = 0; i < right_nritems; i++) {
3731 item = btrfs_item_nr(i);
3732 push_space -= btrfs_token_item_size(right, item, &token);
3733 btrfs_set_token_item_offset(right, item, push_space, &token);
3734 }
3735
3736 left_nritems -= push_items;
3737 btrfs_set_header_nritems(left, left_nritems);
3738
3739 if (left_nritems)
3740 btrfs_mark_buffer_dirty(left);
3741 else
3742 btrfs_clean_tree_block(left);
3743
3744 btrfs_mark_buffer_dirty(right);
3745
3746 btrfs_item_key(right, &disk_key, 0);
3747 btrfs_set_node_key(upper, &disk_key, slot + 1);
3748 btrfs_mark_buffer_dirty(upper);
3749
3750 /* then fixup the leaf pointer in the path */
3751 if (path->slots[0] >= left_nritems) {
3752 path->slots[0] -= left_nritems;
3753 if (btrfs_header_nritems(path->nodes[0]) == 0)
3754 btrfs_clean_tree_block(path->nodes[0]);
3755 btrfs_tree_unlock(path->nodes[0]);
3756 free_extent_buffer(path->nodes[0]);
3757 path->nodes[0] = right;
3758 path->slots[1] += 1;
3759 } else {
3760 btrfs_tree_unlock(right);
3761 free_extent_buffer(right);
3762 }
3763 return 0;
3764
3765out_unlock:
3766 btrfs_tree_unlock(right);
3767 free_extent_buffer(right);
3768 return 1;
3769}
3770
3771/*
3772 * push some data in the path leaf to the right, trying to free up at
3773 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3774 *
3775 * returns 1 if the push failed because the other node didn't have enough
3776 * room, 0 if everything worked out and < 0 if there were major errors.
3777 *
3778 * this will push starting from min_slot to the end of the leaf. It won't
3779 * push any slot lower than min_slot
3780 */
3781static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3782 *root, struct btrfs_path *path,
3783 int min_data_size, int data_size,
3784 int empty, u32 min_slot)
3785{
3786 struct extent_buffer *left = path->nodes[0];
3787 struct extent_buffer *right;
3788 struct extent_buffer *upper;
3789 int slot;
3790 int free_space;
3791 u32 left_nritems;
3792 int ret;
3793
3794 if (!path->nodes[1])
3795 return 1;
3796
3797 slot = path->slots[1];
3798 upper = path->nodes[1];
3799 if (slot >= btrfs_header_nritems(upper) - 1)
3800 return 1;
3801
3802 btrfs_assert_tree_locked(path->nodes[1]);
3803
3804 right = btrfs_read_node_slot(upper, slot + 1);
3805 /*
3806 * slot + 1 is not valid or we fail to read the right node,
3807 * no big deal, just return.
3808 */
3809 if (IS_ERR(right))
3810 return 1;
3811
3812 btrfs_tree_lock(right);
3813 btrfs_set_lock_blocking_write(right);
3814
3815 free_space = btrfs_leaf_free_space(right);
3816 if (free_space < data_size)
3817 goto out_unlock;
3818
3819 /* cow and double check */
3820 ret = btrfs_cow_block(trans, root, right, upper,
3821 slot + 1, &right);
3822 if (ret)
3823 goto out_unlock;
3824
3825 free_space = btrfs_leaf_free_space(right);
3826 if (free_space < data_size)
3827 goto out_unlock;
3828
3829 left_nritems = btrfs_header_nritems(left);
3830 if (left_nritems == 0)
3831 goto out_unlock;
3832
3833 if (path->slots[0] == left_nritems && !empty) {
3834 /* Key greater than all keys in the leaf, right neighbor has
3835 * enough room for it and we're not emptying our leaf to delete
3836 * it, therefore use right neighbor to insert the new item and
3837 * no need to touch/dirty our left leaf. */
3838 btrfs_tree_unlock(left);
3839 free_extent_buffer(left);
3840 path->nodes[0] = right;
3841 path->slots[0] = 0;
3842 path->slots[1]++;
3843 return 0;
3844 }
3845
3846 return __push_leaf_right(path, min_data_size, empty,
3847 right, free_space, left_nritems, min_slot);
3848out_unlock:
3849 btrfs_tree_unlock(right);
3850 free_extent_buffer(right);
3851 return 1;
3852}
3853
3854/*
3855 * push some data in the path leaf to the left, trying to free up at
3856 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3857 *
3858 * max_slot can put a limit on how far into the leaf we'll push items. The
3859 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3860 * items
3861 */
3862static noinline int __push_leaf_left(struct btrfs_path *path, int data_size,
3863 int empty, struct extent_buffer *left,
3864 int free_space, u32 right_nritems,
3865 u32 max_slot)
3866{
3867 struct btrfs_fs_info *fs_info = left->fs_info;
3868 struct btrfs_disk_key disk_key;
3869 struct extent_buffer *right = path->nodes[0];
3870 int i;
3871 int push_space = 0;
3872 int push_items = 0;
3873 struct btrfs_item *item;
3874 u32 old_left_nritems;
3875 u32 nr;
3876 int ret = 0;
3877 u32 this_item_size;
3878 u32 old_left_item_size;
3879 struct btrfs_map_token token;
3880
3881 if (empty)
3882 nr = min(right_nritems, max_slot);
3883 else
3884 nr = min(right_nritems - 1, max_slot);
3885
3886 for (i = 0; i < nr; i++) {
3887 item = btrfs_item_nr(i);
3888
3889 if (!empty && push_items > 0) {
3890 if (path->slots[0] < i)
3891 break;
3892 if (path->slots[0] == i) {
3893 int space = btrfs_leaf_free_space(right);
3894
3895 if (space + push_space * 2 > free_space)
3896 break;
3897 }
3898 }
3899
3900 if (path->slots[0] == i)
3901 push_space += data_size;
3902
3903 this_item_size = btrfs_item_size(right, item);
3904 if (this_item_size + sizeof(*item) + push_space > free_space)
3905 break;
3906
3907 push_items++;
3908 push_space += this_item_size + sizeof(*item);
3909 }
3910
3911 if (push_items == 0) {
3912 ret = 1;
3913 goto out;
3914 }
3915 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3916
3917 /* push data from right to left */
3918 copy_extent_buffer(left, right,
3919 btrfs_item_nr_offset(btrfs_header_nritems(left)),
3920 btrfs_item_nr_offset(0),
3921 push_items * sizeof(struct btrfs_item));
3922
3923 push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3924 btrfs_item_offset_nr(right, push_items - 1);
3925
3926 copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
3927 leaf_data_end(left) - push_space,
3928 BTRFS_LEAF_DATA_OFFSET +
3929 btrfs_item_offset_nr(right, push_items - 1),
3930 push_space);
3931 old_left_nritems = btrfs_header_nritems(left);
3932 BUG_ON(old_left_nritems <= 0);
3933
3934 btrfs_init_map_token(&token, left);
3935 old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3936 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3937 u32 ioff;
3938
3939 item = btrfs_item_nr(i);
3940
3941 ioff = btrfs_token_item_offset(left, item, &token);
3942 btrfs_set_token_item_offset(left, item,
3943 ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size),
3944 &token);
3945 }
3946 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3947
3948 /* fixup right node */
3949 if (push_items > right_nritems)
3950 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3951 right_nritems);
3952
3953 if (push_items < right_nritems) {
3954 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3955 leaf_data_end(right);
3956 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
3957 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3958 BTRFS_LEAF_DATA_OFFSET +
3959 leaf_data_end(right), push_space);
3960
3961 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3962 btrfs_item_nr_offset(push_items),
3963 (btrfs_header_nritems(right) - push_items) *
3964 sizeof(struct btrfs_item));
3965 }
3966
3967 btrfs_init_map_token(&token, right);
3968 right_nritems -= push_items;
3969 btrfs_set_header_nritems(right, right_nritems);
3970 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3971 for (i = 0; i < right_nritems; i++) {
3972 item = btrfs_item_nr(i);
3973
3974 push_space = push_space - btrfs_token_item_size(right,
3975 item, &token);
3976 btrfs_set_token_item_offset(right, item, push_space, &token);
3977 }
3978
3979 btrfs_mark_buffer_dirty(left);
3980 if (right_nritems)
3981 btrfs_mark_buffer_dirty(right);
3982 else
3983 btrfs_clean_tree_block(right);
3984
3985 btrfs_item_key(right, &disk_key, 0);
3986 fixup_low_keys(path, &disk_key, 1);
3987
3988 /* then fixup the leaf pointer in the path */
3989 if (path->slots[0] < push_items) {
3990 path->slots[0] += old_left_nritems;
3991 btrfs_tree_unlock(path->nodes[0]);
3992 free_extent_buffer(path->nodes[0]);
3993 path->nodes[0] = left;
3994 path->slots[1] -= 1;
3995 } else {
3996 btrfs_tree_unlock(left);
3997 free_extent_buffer(left);
3998 path->slots[0] -= push_items;
3999 }
4000 BUG_ON(path->slots[0] < 0);
4001 return ret;
4002out:
4003 btrfs_tree_unlock(left);
4004 free_extent_buffer(left);
4005 return ret;
4006}
4007
4008/*
4009 * push some data in the path leaf to the left, trying to free up at
4010 * least data_size bytes. returns zero if the push worked, nonzero otherwise
4011 *
4012 * max_slot can put a limit on how far into the leaf we'll push items. The
4013 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
4014 * items
4015 */
4016static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
4017 *root, struct btrfs_path *path, int min_data_size,
4018 int data_size, int empty, u32 max_slot)
4019{
4020 struct extent_buffer *right = path->nodes[0];
4021 struct extent_buffer *left;
4022 int slot;
4023 int free_space;
4024 u32 right_nritems;
4025 int ret = 0;
4026
4027 slot = path->slots[1];
4028 if (slot == 0)
4029 return 1;
4030 if (!path->nodes[1])
4031 return 1;
4032
4033 right_nritems = btrfs_header_nritems(right);
4034 if (right_nritems == 0)
4035 return 1;
4036
4037 btrfs_assert_tree_locked(path->nodes[1]);
4038
4039 left = btrfs_read_node_slot(path->nodes[1], slot - 1);
4040 /*
4041 * slot - 1 is not valid or we fail to read the left node,
4042 * no big deal, just return.
4043 */
4044 if (IS_ERR(left))
4045 return 1;
4046
4047 btrfs_tree_lock(left);
4048 btrfs_set_lock_blocking_write(left);
4049
4050 free_space = btrfs_leaf_free_space(left);
4051 if (free_space < data_size) {
4052 ret = 1;
4053 goto out;
4054 }
4055
4056 /* cow and double check */
4057 ret = btrfs_cow_block(trans, root, left,
4058 path->nodes[1], slot - 1, &left);
4059 if (ret) {
4060 /* we hit -ENOSPC, but it isn't fatal here */
4061 if (ret == -ENOSPC)
4062 ret = 1;
4063 goto out;
4064 }
4065
4066 free_space = btrfs_leaf_free_space(left);
4067 if (free_space < data_size) {
4068 ret = 1;
4069 goto out;
4070 }
4071
4072 return __push_leaf_left(path, min_data_size,
4073 empty, left, free_space, right_nritems,
4074 max_slot);
4075out:
4076 btrfs_tree_unlock(left);
4077 free_extent_buffer(left);
4078 return ret;
4079}
4080
4081/*
4082 * split the path's leaf in two, making sure there is at least data_size
4083 * available for the resulting leaf level of the path.
4084 */
4085static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4086 struct btrfs_path *path,
4087 struct extent_buffer *l,
4088 struct extent_buffer *right,
4089 int slot, int mid, int nritems)
4090{
4091 struct btrfs_fs_info *fs_info = trans->fs_info;
4092 int data_copy_size;
4093 int rt_data_off;
4094 int i;
4095 struct btrfs_disk_key disk_key;
4096 struct btrfs_map_token token;
4097
4098 nritems = nritems - mid;
4099 btrfs_set_header_nritems(right, nritems);
4100 data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(l);
4101
4102 copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4103 btrfs_item_nr_offset(mid),
4104 nritems * sizeof(struct btrfs_item));
4105
4106 copy_extent_buffer(right, l,
4107 BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
4108 data_copy_size, BTRFS_LEAF_DATA_OFFSET +
4109 leaf_data_end(l), data_copy_size);
4110
4111 rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4112
4113 btrfs_init_map_token(&token, right);
4114 for (i = 0; i < nritems; i++) {
4115 struct btrfs_item *item = btrfs_item_nr(i);
4116 u32 ioff;
4117
4118 ioff = btrfs_token_item_offset(right, item, &token);
4119 btrfs_set_token_item_offset(right, item,
4120 ioff + rt_data_off, &token);
4121 }
4122
4123 btrfs_set_header_nritems(l, mid);
4124 btrfs_item_key(right, &disk_key, 0);
4125 insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
4126
4127 btrfs_mark_buffer_dirty(right);
4128 btrfs_mark_buffer_dirty(l);
4129 BUG_ON(path->slots[0] != slot);
4130
4131 if (mid <= slot) {
4132 btrfs_tree_unlock(path->nodes[0]);
4133 free_extent_buffer(path->nodes[0]);
4134 path->nodes[0] = right;
4135 path->slots[0] -= mid;
4136 path->slots[1] += 1;
4137 } else {
4138 btrfs_tree_unlock(right);
4139 free_extent_buffer(right);
4140 }
4141
4142 BUG_ON(path->slots[0] < 0);
4143}
4144
4145/*
4146 * double splits happen when we need to insert a big item in the middle
4147 * of a leaf. A double split can leave us with 3 mostly empty leaves:
4148 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4149 * A B C
4150 *
4151 * We avoid this by trying to push the items on either side of our target
4152 * into the adjacent leaves. If all goes well we can avoid the double split
4153 * completely.
4154 */
4155static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4156 struct btrfs_root *root,
4157 struct btrfs_path *path,
4158 int data_size)
4159{
4160 int ret;
4161 int progress = 0;
4162 int slot;
4163 u32 nritems;
4164 int space_needed = data_size;
4165
4166 slot = path->slots[0];
4167 if (slot < btrfs_header_nritems(path->nodes[0]))
4168 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4169
4170 /*
4171 * try to push all the items after our slot into the
4172 * right leaf
4173 */
4174 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4175 if (ret < 0)
4176 return ret;
4177
4178 if (ret == 0)
4179 progress++;
4180
4181 nritems = btrfs_header_nritems(path->nodes[0]);
4182 /*
4183 * our goal is to get our slot at the start or end of a leaf. If
4184 * we've done so we're done
4185 */
4186 if (path->slots[0] == 0 || path->slots[0] == nritems)
4187 return 0;
4188
4189 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4190 return 0;
4191
4192 /* try to push all the items before our slot into the next leaf */
4193 slot = path->slots[0];
4194 space_needed = data_size;
4195 if (slot > 0)
4196 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4197 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4198 if (ret < 0)
4199 return ret;
4200
4201 if (ret == 0)
4202 progress++;
4203
4204 if (progress)
4205 return 0;
4206 return 1;
4207}
4208
4209/*
4210 * split the path's leaf in two, making sure there is at least data_size
4211 * available for the resulting leaf level of the path.
4212 *
4213 * returns 0 if all went well and < 0 on failure.
4214 */
4215static noinline int split_leaf(struct btrfs_trans_handle *trans,
4216 struct btrfs_root *root,
4217 const struct btrfs_key *ins_key,
4218 struct btrfs_path *path, int data_size,
4219 int extend)
4220{
4221 struct btrfs_disk_key disk_key;
4222 struct extent_buffer *l;
4223 u32 nritems;
4224 int mid;
4225 int slot;
4226 struct extent_buffer *right;
4227 struct btrfs_fs_info *fs_info = root->fs_info;
4228 int ret = 0;
4229 int wret;
4230 int split;
4231 int num_doubles = 0;
4232 int tried_avoid_double = 0;
4233
4234 l = path->nodes[0];
4235 slot = path->slots[0];
4236 if (extend && data_size + btrfs_item_size_nr(l, slot) +
4237 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4238 return -EOVERFLOW;
4239
4240 /* first try to make some room by pushing left and right */
4241 if (data_size && path->nodes[1]) {
4242 int space_needed = data_size;
4243
4244 if (slot < btrfs_header_nritems(l))
4245 space_needed -= btrfs_leaf_free_space(l);
4246
4247 wret = push_leaf_right(trans, root, path, space_needed,
4248 space_needed, 0, 0);
4249 if (wret < 0)
4250 return wret;
4251 if (wret) {
4252 space_needed = data_size;
4253 if (slot > 0)
4254 space_needed -= btrfs_leaf_free_space(l);
4255 wret = push_leaf_left(trans, root, path, space_needed,
4256 space_needed, 0, (u32)-1);
4257 if (wret < 0)
4258 return wret;
4259 }
4260 l = path->nodes[0];
4261
4262 /* did the pushes work? */
4263 if (btrfs_leaf_free_space(l) >= data_size)
4264 return 0;
4265 }
4266
4267 if (!path->nodes[1]) {
4268 ret = insert_new_root(trans, root, path, 1);
4269 if (ret)
4270 return ret;
4271 }
4272again:
4273 split = 1;
4274 l = path->nodes[0];
4275 slot = path->slots[0];
4276 nritems = btrfs_header_nritems(l);
4277 mid = (nritems + 1) / 2;
4278
4279 if (mid <= slot) {
4280 if (nritems == 1 ||
4281 leaf_space_used(l, mid, nritems - mid) + data_size >
4282 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4283 if (slot >= nritems) {
4284 split = 0;
4285 } else {
4286 mid = slot;
4287 if (mid != nritems &&
4288 leaf_space_used(l, mid, nritems - mid) +
4289 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4290 if (data_size && !tried_avoid_double)
4291 goto push_for_double;
4292 split = 2;
4293 }
4294 }
4295 }
4296 } else {
4297 if (leaf_space_used(l, 0, mid) + data_size >
4298 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4299 if (!extend && data_size && slot == 0) {
4300 split = 0;
4301 } else if ((extend || !data_size) && slot == 0) {
4302 mid = 1;
4303 } else {
4304 mid = slot;
4305 if (mid != nritems &&
4306 leaf_space_used(l, mid, nritems - mid) +
4307 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4308 if (data_size && !tried_avoid_double)
4309 goto push_for_double;
4310 split = 2;
4311 }
4312 }
4313 }
4314 }
4315
4316 if (split == 0)
4317 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4318 else
4319 btrfs_item_key(l, &disk_key, mid);
4320
4321 right = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, 0,
4322 l->start, 0);
4323 if (IS_ERR(right))
4324 return PTR_ERR(right);
4325
4326 root_add_used(root, fs_info->nodesize);
4327
4328 if (split == 0) {
4329 if (mid <= slot) {
4330 btrfs_set_header_nritems(right, 0);
4331 insert_ptr(trans, path, &disk_key,
4332 right->start, path->slots[1] + 1, 1);
4333 btrfs_tree_unlock(path->nodes[0]);
4334 free_extent_buffer(path->nodes[0]);
4335 path->nodes[0] = right;
4336 path->slots[0] = 0;
4337 path->slots[1] += 1;
4338 } else {
4339 btrfs_set_header_nritems(right, 0);
4340 insert_ptr(trans, path, &disk_key,
4341 right->start, path->slots[1], 1);
4342 btrfs_tree_unlock(path->nodes[0]);
4343 free_extent_buffer(path->nodes[0]);
4344 path->nodes[0] = right;
4345 path->slots[0] = 0;
4346 if (path->slots[1] == 0)
4347 fixup_low_keys(path, &disk_key, 1);
4348 }
4349 /*
4350 * We create a new leaf 'right' for the required ins_len and
4351 * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4352 * the content of ins_len to 'right'.
4353 */
4354 return ret;
4355 }
4356
4357 copy_for_split(trans, path, l, right, slot, mid, nritems);
4358
4359 if (split == 2) {
4360 BUG_ON(num_doubles != 0);
4361 num_doubles++;
4362 goto again;
4363 }
4364
4365 return 0;
4366
4367push_for_double:
4368 push_for_double_split(trans, root, path, data_size);
4369 tried_avoid_double = 1;
4370 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4371 return 0;
4372 goto again;
4373}
4374
4375static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4376 struct btrfs_root *root,
4377 struct btrfs_path *path, int ins_len)
4378{
4379 struct btrfs_key key;
4380 struct extent_buffer *leaf;
4381 struct btrfs_file_extent_item *fi;
4382 u64 extent_len = 0;
4383 u32 item_size;
4384 int ret;
4385
4386 leaf = path->nodes[0];
4387 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4388
4389 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4390 key.type != BTRFS_EXTENT_CSUM_KEY);
4391
4392 if (btrfs_leaf_free_space(leaf) >= ins_len)
4393 return 0;
4394
4395 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4396 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4397 fi = btrfs_item_ptr(leaf, path->slots[0],
4398 struct btrfs_file_extent_item);
4399 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4400 }
4401 btrfs_release_path(path);
4402
4403 path->keep_locks = 1;
4404 path->search_for_split = 1;
4405 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4406 path->search_for_split = 0;
4407 if (ret > 0)
4408 ret = -EAGAIN;
4409 if (ret < 0)
4410 goto err;
4411
4412 ret = -EAGAIN;
4413 leaf = path->nodes[0];
4414 /* if our item isn't there, return now */
4415 if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4416 goto err;
4417
4418 /* the leaf has changed, it now has room. return now */
4419 if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
4420 goto err;
4421
4422 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4423 fi = btrfs_item_ptr(leaf, path->slots[0],
4424 struct btrfs_file_extent_item);
4425 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4426 goto err;
4427 }
4428
4429 btrfs_set_path_blocking(path);
4430 ret = split_leaf(trans, root, &key, path, ins_len, 1);
4431 if (ret)
4432 goto err;
4433
4434 path->keep_locks = 0;
4435 btrfs_unlock_up_safe(path, 1);
4436 return 0;
4437err:
4438 path->keep_locks = 0;
4439 return ret;
4440}
4441
4442static noinline int split_item(struct btrfs_path *path,
4443 const struct btrfs_key *new_key,
4444 unsigned long split_offset)
4445{
4446 struct extent_buffer *leaf;
4447 struct btrfs_item *item;
4448 struct btrfs_item *new_item;
4449 int slot;
4450 char *buf;
4451 u32 nritems;
4452 u32 item_size;
4453 u32 orig_offset;
4454 struct btrfs_disk_key disk_key;
4455
4456 leaf = path->nodes[0];
4457 BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
4458
4459 btrfs_set_path_blocking(path);
4460
4461 item = btrfs_item_nr(path->slots[0]);
4462 orig_offset = btrfs_item_offset(leaf, item);
4463 item_size = btrfs_item_size(leaf, item);
4464
4465 buf = kmalloc(item_size, GFP_NOFS);
4466 if (!buf)
4467 return -ENOMEM;
4468
4469 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4470 path->slots[0]), item_size);
4471
4472 slot = path->slots[0] + 1;
4473 nritems = btrfs_header_nritems(leaf);
4474 if (slot != nritems) {
4475 /* shift the items */
4476 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4477 btrfs_item_nr_offset(slot),
4478 (nritems - slot) * sizeof(struct btrfs_item));
4479 }
4480
4481 btrfs_cpu_key_to_disk(&disk_key, new_key);
4482 btrfs_set_item_key(leaf, &disk_key, slot);
4483
4484 new_item = btrfs_item_nr(slot);
4485
4486 btrfs_set_item_offset(leaf, new_item, orig_offset);
4487 btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4488
4489 btrfs_set_item_offset(leaf, item,
4490 orig_offset + item_size - split_offset);
4491 btrfs_set_item_size(leaf, item, split_offset);
4492
4493 btrfs_set_header_nritems(leaf, nritems + 1);
4494
4495 /* write the data for the start of the original item */
4496 write_extent_buffer(leaf, buf,
4497 btrfs_item_ptr_offset(leaf, path->slots[0]),
4498 split_offset);
4499
4500 /* write the data for the new item */
4501 write_extent_buffer(leaf, buf + split_offset,
4502 btrfs_item_ptr_offset(leaf, slot),
4503 item_size - split_offset);
4504 btrfs_mark_buffer_dirty(leaf);
4505
4506 BUG_ON(btrfs_leaf_free_space(leaf) < 0);
4507 kfree(buf);
4508 return 0;
4509}
4510
4511/*
4512 * This function splits a single item into two items,
4513 * giving 'new_key' to the new item and splitting the
4514 * old one at split_offset (from the start of the item).
4515 *
4516 * The path may be released by this operation. After
4517 * the split, the path is pointing to the old item. The
4518 * new item is going to be in the same node as the old one.
4519 *
4520 * Note, the item being split must be smaller enough to live alone on
4521 * a tree block with room for one extra struct btrfs_item
4522 *
4523 * This allows us to split the item in place, keeping a lock on the
4524 * leaf the entire time.
4525 */
4526int btrfs_split_item(struct btrfs_trans_handle *trans,
4527 struct btrfs_root *root,
4528 struct btrfs_path *path,
4529 const struct btrfs_key *new_key,
4530 unsigned long split_offset)
4531{
4532 int ret;
4533 ret = setup_leaf_for_split(trans, root, path,
4534 sizeof(struct btrfs_item));
4535 if (ret)
4536 return ret;
4537
4538 ret = split_item(path, new_key, split_offset);
4539 return ret;
4540}
4541
4542/*
4543 * This function duplicate a item, giving 'new_key' to the new item.
4544 * It guarantees both items live in the same tree leaf and the new item
4545 * is contiguous with the original item.
4546 *
4547 * This allows us to split file extent in place, keeping a lock on the
4548 * leaf the entire time.
4549 */
4550int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4551 struct btrfs_root *root,
4552 struct btrfs_path *path,
4553 const struct btrfs_key *new_key)
4554{
4555 struct extent_buffer *leaf;
4556 int ret;
4557 u32 item_size;
4558
4559 leaf = path->nodes[0];
4560 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4561 ret = setup_leaf_for_split(trans, root, path,
4562 item_size + sizeof(struct btrfs_item));
4563 if (ret)
4564 return ret;
4565
4566 path->slots[0]++;
4567 setup_items_for_insert(root, path, new_key, &item_size,
4568 item_size, item_size +
4569 sizeof(struct btrfs_item), 1);
4570 leaf = path->nodes[0];
4571 memcpy_extent_buffer(leaf,
4572 btrfs_item_ptr_offset(leaf, path->slots[0]),
4573 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4574 item_size);
4575 return 0;
4576}
4577
4578/*
4579 * make the item pointed to by the path smaller. new_size indicates
4580 * how small to make it, and from_end tells us if we just chop bytes
4581 * off the end of the item or if we shift the item to chop bytes off
4582 * the front.
4583 */
4584void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
4585{
4586 int slot;
4587 struct extent_buffer *leaf;
4588 struct btrfs_item *item;
4589 u32 nritems;
4590 unsigned int data_end;
4591 unsigned int old_data_start;
4592 unsigned int old_size;
4593 unsigned int size_diff;
4594 int i;
4595 struct btrfs_map_token token;
4596
4597 leaf = path->nodes[0];
4598 slot = path->slots[0];
4599
4600 old_size = btrfs_item_size_nr(leaf, slot);
4601 if (old_size == new_size)
4602 return;
4603
4604 nritems = btrfs_header_nritems(leaf);
4605 data_end = leaf_data_end(leaf);
4606
4607 old_data_start = btrfs_item_offset_nr(leaf, slot);
4608
4609 size_diff = old_size - new_size;
4610
4611 BUG_ON(slot < 0);
4612 BUG_ON(slot >= nritems);
4613
4614 /*
4615 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4616 */
4617 /* first correct the data pointers */
4618 btrfs_init_map_token(&token, leaf);
4619 for (i = slot; i < nritems; i++) {
4620 u32 ioff;
4621 item = btrfs_item_nr(i);
4622
4623 ioff = btrfs_token_item_offset(leaf, item, &token);
4624 btrfs_set_token_item_offset(leaf, item,
4625 ioff + size_diff, &token);
4626 }
4627
4628 /* shift the data */
4629 if (from_end) {
4630 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4631 data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4632 data_end, old_data_start + new_size - data_end);
4633 } else {
4634 struct btrfs_disk_key disk_key;
4635 u64 offset;
4636
4637 btrfs_item_key(leaf, &disk_key, slot);
4638
4639 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4640 unsigned long ptr;
4641 struct btrfs_file_extent_item *fi;
4642
4643 fi = btrfs_item_ptr(leaf, slot,
4644 struct btrfs_file_extent_item);
4645 fi = (struct btrfs_file_extent_item *)(
4646 (unsigned long)fi - size_diff);
4647
4648 if (btrfs_file_extent_type(leaf, fi) ==
4649 BTRFS_FILE_EXTENT_INLINE) {
4650 ptr = btrfs_item_ptr_offset(leaf, slot);
4651 memmove_extent_buffer(leaf, ptr,
4652 (unsigned long)fi,
4653 BTRFS_FILE_EXTENT_INLINE_DATA_START);
4654 }
4655 }
4656
4657 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4658 data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4659 data_end, old_data_start - data_end);
4660
4661 offset = btrfs_disk_key_offset(&disk_key);
4662 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4663 btrfs_set_item_key(leaf, &disk_key, slot);
4664 if (slot == 0)
4665 fixup_low_keys(path, &disk_key, 1);
4666 }
4667
4668 item = btrfs_item_nr(slot);
4669 btrfs_set_item_size(leaf, item, new_size);
4670 btrfs_mark_buffer_dirty(leaf);
4671
4672 if (btrfs_leaf_free_space(leaf) < 0) {
4673 btrfs_print_leaf(leaf);
4674 BUG();
4675 }
4676}
4677
4678/*
4679 * make the item pointed to by the path bigger, data_size is the added size.
4680 */
4681void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
4682{
4683 int slot;
4684 struct extent_buffer *leaf;
4685 struct btrfs_item *item;
4686 u32 nritems;
4687 unsigned int data_end;
4688 unsigned int old_data;
4689 unsigned int old_size;
4690 int i;
4691 struct btrfs_map_token token;
4692
4693 leaf = path->nodes[0];
4694
4695 nritems = btrfs_header_nritems(leaf);
4696 data_end = leaf_data_end(leaf);
4697
4698 if (btrfs_leaf_free_space(leaf) < data_size) {
4699 btrfs_print_leaf(leaf);
4700 BUG();
4701 }
4702 slot = path->slots[0];
4703 old_data = btrfs_item_end_nr(leaf, slot);
4704
4705 BUG_ON(slot < 0);
4706 if (slot >= nritems) {
4707 btrfs_print_leaf(leaf);
4708 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
4709 slot, nritems);
4710 BUG();
4711 }
4712
4713 /*
4714 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4715 */
4716 /* first correct the data pointers */
4717 btrfs_init_map_token(&token, leaf);
4718 for (i = slot; i < nritems; i++) {
4719 u32 ioff;
4720 item = btrfs_item_nr(i);
4721
4722 ioff = btrfs_token_item_offset(leaf, item, &token);
4723 btrfs_set_token_item_offset(leaf, item,
4724 ioff - data_size, &token);
4725 }
4726
4727 /* shift the data */
4728 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4729 data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
4730 data_end, old_data - data_end);
4731
4732 data_end = old_data;
4733 old_size = btrfs_item_size_nr(leaf, slot);
4734 item = btrfs_item_nr(slot);
4735 btrfs_set_item_size(leaf, item, old_size + data_size);
4736 btrfs_mark_buffer_dirty(leaf);
4737
4738 if (btrfs_leaf_free_space(leaf) < 0) {
4739 btrfs_print_leaf(leaf);
4740 BUG();
4741 }
4742}
4743
4744/*
4745 * this is a helper for btrfs_insert_empty_items, the main goal here is
4746 * to save stack depth by doing the bulk of the work in a function
4747 * that doesn't call btrfs_search_slot
4748 */
4749void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4750 const struct btrfs_key *cpu_key, u32 *data_size,
4751 u32 total_data, u32 total_size, int nr)
4752{
4753 struct btrfs_fs_info *fs_info = root->fs_info;
4754 struct btrfs_item *item;
4755 int i;
4756 u32 nritems;
4757 unsigned int data_end;
4758 struct btrfs_disk_key disk_key;
4759 struct extent_buffer *leaf;
4760 int slot;
4761 struct btrfs_map_token token;
4762
4763 if (path->slots[0] == 0) {
4764 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4765 fixup_low_keys(path, &disk_key, 1);
4766 }
4767 btrfs_unlock_up_safe(path, 1);
4768
4769 leaf = path->nodes[0];
4770 slot = path->slots[0];
4771
4772 nritems = btrfs_header_nritems(leaf);
4773 data_end = leaf_data_end(leaf);
4774
4775 if (btrfs_leaf_free_space(leaf) < total_size) {
4776 btrfs_print_leaf(leaf);
4777 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4778 total_size, btrfs_leaf_free_space(leaf));
4779 BUG();
4780 }
4781
4782 btrfs_init_map_token(&token, leaf);
4783 if (slot != nritems) {
4784 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4785
4786 if (old_data < data_end) {
4787 btrfs_print_leaf(leaf);
4788 btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4789 slot, old_data, data_end);
4790 BUG();
4791 }
4792 /*
4793 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4794 */
4795 /* first correct the data pointers */
4796 for (i = slot; i < nritems; i++) {
4797 u32 ioff;
4798
4799 item = btrfs_item_nr(i);
4800 ioff = btrfs_token_item_offset(leaf, item, &token);
4801 btrfs_set_token_item_offset(leaf, item,
4802 ioff - total_data, &token);
4803 }
4804 /* shift the items */
4805 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4806 btrfs_item_nr_offset(slot),
4807 (nritems - slot) * sizeof(struct btrfs_item));
4808
4809 /* shift the data */
4810 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4811 data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
4812 data_end, old_data - data_end);
4813 data_end = old_data;
4814 }
4815
4816 /* setup the item for the new data */
4817 for (i = 0; i < nr; i++) {
4818 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4819 btrfs_set_item_key(leaf, &disk_key, slot + i);
4820 item = btrfs_item_nr(slot + i);
4821 btrfs_set_token_item_offset(leaf, item,
4822 data_end - data_size[i], &token);
4823 data_end -= data_size[i];
4824 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4825 }
4826
4827 btrfs_set_header_nritems(leaf, nritems + nr);
4828 btrfs_mark_buffer_dirty(leaf);
4829
4830 if (btrfs_leaf_free_space(leaf) < 0) {
4831 btrfs_print_leaf(leaf);
4832 BUG();
4833 }
4834}
4835
4836/*
4837 * Given a key and some data, insert items into the tree.
4838 * This does all the path init required, making room in the tree if needed.
4839 */
4840int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4841 struct btrfs_root *root,
4842 struct btrfs_path *path,
4843 const struct btrfs_key *cpu_key, u32 *data_size,
4844 int nr)
4845{
4846 int ret = 0;
4847 int slot;
4848 int i;
4849 u32 total_size = 0;
4850 u32 total_data = 0;
4851
4852 for (i = 0; i < nr; i++)
4853 total_data += data_size[i];
4854
4855 total_size = total_data + (nr * sizeof(struct btrfs_item));
4856 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4857 if (ret == 0)
4858 return -EEXIST;
4859 if (ret < 0)
4860 return ret;
4861
4862 slot = path->slots[0];
4863 BUG_ON(slot < 0);
4864
4865 setup_items_for_insert(root, path, cpu_key, data_size,
4866 total_data, total_size, nr);
4867 return 0;
4868}
4869
4870/*
4871 * Given a key and some data, insert an item into the tree.
4872 * This does all the path init required, making room in the tree if needed.
4873 */
4874int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4875 const struct btrfs_key *cpu_key, void *data,
4876 u32 data_size)
4877{
4878 int ret = 0;
4879 struct btrfs_path *path;
4880 struct extent_buffer *leaf;
4881 unsigned long ptr;
4882
4883 path = btrfs_alloc_path();
4884 if (!path)
4885 return -ENOMEM;
4886 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4887 if (!ret) {
4888 leaf = path->nodes[0];
4889 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4890 write_extent_buffer(leaf, data, ptr, data_size);
4891 btrfs_mark_buffer_dirty(leaf);
4892 }
4893 btrfs_free_path(path);
4894 return ret;
4895}
4896
4897/*
4898 * delete the pointer from a given node.
4899 *
4900 * the tree should have been previously balanced so the deletion does not
4901 * empty a node.
4902 */
4903static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4904 int level, int slot)
4905{
4906 struct extent_buffer *parent = path->nodes[level];
4907 u32 nritems;
4908 int ret;
4909
4910 nritems = btrfs_header_nritems(parent);
4911 if (slot != nritems - 1) {
4912 if (level) {
4913 ret = tree_mod_log_insert_move(parent, slot, slot + 1,
4914 nritems - slot - 1);
4915 BUG_ON(ret < 0);
4916 }
4917 memmove_extent_buffer(parent,
4918 btrfs_node_key_ptr_offset(slot),
4919 btrfs_node_key_ptr_offset(slot + 1),
4920 sizeof(struct btrfs_key_ptr) *
4921 (nritems - slot - 1));
4922 } else if (level) {
4923 ret = tree_mod_log_insert_key(parent, slot, MOD_LOG_KEY_REMOVE,
4924 GFP_NOFS);
4925 BUG_ON(ret < 0);
4926 }
4927
4928 nritems--;
4929 btrfs_set_header_nritems(parent, nritems);
4930 if (nritems == 0 && parent == root->node) {
4931 BUG_ON(btrfs_header_level(root->node) != 1);
4932 /* just turn the root into a leaf and break */
4933 btrfs_set_header_level(root->node, 0);
4934 } else if (slot == 0) {
4935 struct btrfs_disk_key disk_key;
4936
4937 btrfs_node_key(parent, &disk_key, 0);
4938 fixup_low_keys(path, &disk_key, level + 1);
4939 }
4940 btrfs_mark_buffer_dirty(parent);
4941}
4942
4943/*
4944 * a helper function to delete the leaf pointed to by path->slots[1] and
4945 * path->nodes[1].
4946 *
4947 * This deletes the pointer in path->nodes[1] and frees the leaf
4948 * block extent. zero is returned if it all worked out, < 0 otherwise.
4949 *
4950 * The path must have already been setup for deleting the leaf, including
4951 * all the proper balancing. path->nodes[1] must be locked.
4952 */
4953static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4954 struct btrfs_root *root,
4955 struct btrfs_path *path,
4956 struct extent_buffer *leaf)
4957{
4958 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4959 del_ptr(root, path, 1, path->slots[1]);
4960
4961 /*
4962 * btrfs_free_extent is expensive, we want to make sure we
4963 * aren't holding any locks when we call it
4964 */
4965 btrfs_unlock_up_safe(path, 0);
4966
4967 root_sub_used(root, leaf->len);
4968
4969 extent_buffer_get(leaf);
4970 btrfs_free_tree_block(trans, root, leaf, 0, 1);
4971 free_extent_buffer_stale(leaf);
4972}
4973/*
4974 * delete the item at the leaf level in path. If that empties
4975 * the leaf, remove it from the tree
4976 */
4977int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4978 struct btrfs_path *path, int slot, int nr)
4979{
4980 struct btrfs_fs_info *fs_info = root->fs_info;
4981 struct extent_buffer *leaf;
4982 struct btrfs_item *item;
4983 u32 last_off;
4984 u32 dsize = 0;
4985 int ret = 0;
4986 int wret;
4987 int i;
4988 u32 nritems;
4989
4990 leaf = path->nodes[0];
4991 last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4992
4993 for (i = 0; i < nr; i++)
4994 dsize += btrfs_item_size_nr(leaf, slot + i);
4995
4996 nritems = btrfs_header_nritems(leaf);
4997
4998 if (slot + nr != nritems) {
4999 int data_end = leaf_data_end(leaf);
5000 struct btrfs_map_token token;
5001
5002 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
5003 data_end + dsize,
5004 BTRFS_LEAF_DATA_OFFSET + data_end,
5005 last_off - data_end);
5006
5007 btrfs_init_map_token(&token, leaf);
5008 for (i = slot + nr; i < nritems; i++) {
5009 u32 ioff;
5010
5011 item = btrfs_item_nr(i);
5012 ioff = btrfs_token_item_offset(leaf, item, &token);
5013 btrfs_set_token_item_offset(leaf, item,
5014 ioff + dsize, &token);
5015 }
5016
5017 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
5018 btrfs_item_nr_offset(slot + nr),
5019 sizeof(struct btrfs_item) *
5020 (nritems - slot - nr));
5021 }
5022 btrfs_set_header_nritems(leaf, nritems - nr);
5023 nritems -= nr;
5024
5025 /* delete the leaf if we've emptied it */
5026 if (nritems == 0) {
5027 if (leaf == root->node) {
5028 btrfs_set_header_level(leaf, 0);
5029 } else {
5030 btrfs_set_path_blocking(path);
5031 btrfs_clean_tree_block(leaf);
5032 btrfs_del_leaf(trans, root, path, leaf);
5033 }
5034 } else {
5035 int used = leaf_space_used(leaf, 0, nritems);
5036 if (slot == 0) {
5037 struct btrfs_disk_key disk_key;
5038
5039 btrfs_item_key(leaf, &disk_key, 0);
5040 fixup_low_keys(path, &disk_key, 1);
5041 }
5042
5043 /* delete the leaf if it is mostly empty */
5044 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
5045 /* push_leaf_left fixes the path.
5046 * make sure the path still points to our leaf
5047 * for possible call to del_ptr below
5048 */
5049 slot = path->slots[1];
5050 extent_buffer_get(leaf);
5051
5052 btrfs_set_path_blocking(path);
5053 wret = push_leaf_left(trans, root, path, 1, 1,
5054 1, (u32)-1);
5055 if (wret < 0 && wret != -ENOSPC)
5056 ret = wret;
5057
5058 if (path->nodes[0] == leaf &&
5059 btrfs_header_nritems(leaf)) {
5060 wret = push_leaf_right(trans, root, path, 1,
5061 1, 1, 0);
5062 if (wret < 0 && wret != -ENOSPC)
5063 ret = wret;
5064 }
5065
5066 if (btrfs_header_nritems(leaf) == 0) {
5067 path->slots[1] = slot;
5068 btrfs_del_leaf(trans, root, path, leaf);
5069 free_extent_buffer(leaf);
5070 ret = 0;
5071 } else {
5072 /* if we're still in the path, make sure
5073 * we're dirty. Otherwise, one of the
5074 * push_leaf functions must have already
5075 * dirtied this buffer
5076 */
5077 if (path->nodes[0] == leaf)
5078 btrfs_mark_buffer_dirty(leaf);
5079 free_extent_buffer(leaf);
5080 }
5081 } else {
5082 btrfs_mark_buffer_dirty(leaf);
5083 }
5084 }
5085 return ret;
5086}
5087
5088/*
5089 * search the tree again to find a leaf with lesser keys
5090 * returns 0 if it found something or 1 if there are no lesser leaves.
5091 * returns < 0 on io errors.
5092 *
5093 * This may release the path, and so you may lose any locks held at the
5094 * time you call it.
5095 */
5096int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5097{
5098 struct btrfs_key key;
5099 struct btrfs_disk_key found_key;
5100 int ret;
5101
5102 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5103
5104 if (key.offset > 0) {
5105 key.offset--;
5106 } else if (key.type > 0) {
5107 key.type--;
5108 key.offset = (u64)-1;
5109 } else if (key.objectid > 0) {
5110 key.objectid--;
5111 key.type = (u8)-1;
5112 key.offset = (u64)-1;
5113 } else {
5114 return 1;
5115 }
5116
5117 btrfs_release_path(path);
5118 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5119 if (ret < 0)
5120 return ret;
5121 btrfs_item_key(path->nodes[0], &found_key, 0);
5122 ret = comp_keys(&found_key, &key);
5123 /*
5124 * We might have had an item with the previous key in the tree right
5125 * before we released our path. And after we released our path, that
5126 * item might have been pushed to the first slot (0) of the leaf we
5127 * were holding due to a tree balance. Alternatively, an item with the
5128 * previous key can exist as the only element of a leaf (big fat item).
5129 * Therefore account for these 2 cases, so that our callers (like
5130 * btrfs_previous_item) don't miss an existing item with a key matching
5131 * the previous key we computed above.
5132 */
5133 if (ret <= 0)
5134 return 0;
5135 return 1;
5136}
5137
5138/*
5139 * A helper function to walk down the tree starting at min_key, and looking
5140 * for nodes or leaves that are have a minimum transaction id.
5141 * This is used by the btree defrag code, and tree logging
5142 *
5143 * This does not cow, but it does stuff the starting key it finds back
5144 * into min_key, so you can call btrfs_search_slot with cow=1 on the
5145 * key and get a writable path.
5146 *
5147 * This honors path->lowest_level to prevent descent past a given level
5148 * of the tree.
5149 *
5150 * min_trans indicates the oldest transaction that you are interested
5151 * in walking through. Any nodes or leaves older than min_trans are
5152 * skipped over (without reading them).
5153 *
5154 * returns zero if something useful was found, < 0 on error and 1 if there
5155 * was nothing in the tree that matched the search criteria.
5156 */
5157int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5158 struct btrfs_path *path,
5159 u64 min_trans)
5160{
5161 struct extent_buffer *cur;
5162 struct btrfs_key found_key;
5163 int slot;
5164 int sret;
5165 u32 nritems;
5166 int level;
5167 int ret = 1;
5168 int keep_locks = path->keep_locks;
5169
5170 path->keep_locks = 1;
5171again:
5172 cur = btrfs_read_lock_root_node(root);
5173 level = btrfs_header_level(cur);
5174 WARN_ON(path->nodes[level]);
5175 path->nodes[level] = cur;
5176 path->locks[level] = BTRFS_READ_LOCK;
5177
5178 if (btrfs_header_generation(cur) < min_trans) {
5179 ret = 1;
5180 goto out;
5181 }
5182 while (1) {
5183 nritems = btrfs_header_nritems(cur);
5184 level = btrfs_header_level(cur);
5185 sret = btrfs_bin_search(cur, min_key, level, &slot);
5186 if (sret < 0) {
5187 ret = sret;
5188 goto out;
5189 }
5190
5191 /* at the lowest level, we're done, setup the path and exit */
5192 if (level == path->lowest_level) {
5193 if (slot >= nritems)
5194 goto find_next_key;
5195 ret = 0;
5196 path->slots[level] = slot;
5197 btrfs_item_key_to_cpu(cur, &found_key, slot);
5198 goto out;
5199 }
5200 if (sret && slot > 0)
5201 slot--;
5202 /*
5203 * check this node pointer against the min_trans parameters.
5204 * If it is too old, old, skip to the next one.
5205 */
5206 while (slot < nritems) {
5207 u64 gen;
5208
5209 gen = btrfs_node_ptr_generation(cur, slot);
5210 if (gen < min_trans) {
5211 slot++;
5212 continue;
5213 }
5214 break;
5215 }
5216find_next_key:
5217 /*
5218 * we didn't find a candidate key in this node, walk forward
5219 * and find another one
5220 */
5221 if (slot >= nritems) {
5222 path->slots[level] = slot;
5223 btrfs_set_path_blocking(path);
5224 sret = btrfs_find_next_key(root, path, min_key, level,
5225 min_trans);
5226 if (sret == 0) {
5227 btrfs_release_path(path);
5228 goto again;
5229 } else {
5230 goto out;
5231 }
5232 }
5233 /* save our key for returning back */
5234 btrfs_node_key_to_cpu(cur, &found_key, slot);
5235 path->slots[level] = slot;
5236 if (level == path->lowest_level) {
5237 ret = 0;
5238 goto out;
5239 }
5240 btrfs_set_path_blocking(path);
5241 cur = btrfs_read_node_slot(cur, slot);
5242 if (IS_ERR(cur)) {
5243 ret = PTR_ERR(cur);
5244 goto out;
5245 }
5246
5247 btrfs_tree_read_lock(cur);
5248
5249 path->locks[level - 1] = BTRFS_READ_LOCK;
5250 path->nodes[level - 1] = cur;
5251 unlock_up(path, level, 1, 0, NULL);
5252 }
5253out:
5254 path->keep_locks = keep_locks;
5255 if (ret == 0) {
5256 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5257 btrfs_set_path_blocking(path);
5258 memcpy(min_key, &found_key, sizeof(found_key));
5259 }
5260 return ret;
5261}
5262
5263/*
5264 * this is similar to btrfs_next_leaf, but does not try to preserve
5265 * and fixup the path. It looks for and returns the next key in the
5266 * tree based on the current path and the min_trans parameters.
5267 *
5268 * 0 is returned if another key is found, < 0 if there are any errors
5269 * and 1 is returned if there are no higher keys in the tree
5270 *
5271 * path->keep_locks should be set to 1 on the search made before
5272 * calling this function.
5273 */
5274int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5275 struct btrfs_key *key, int level, u64 min_trans)
5276{
5277 int slot;
5278 struct extent_buffer *c;
5279
5280 WARN_ON(!path->keep_locks && !path->skip_locking);
5281 while (level < BTRFS_MAX_LEVEL) {
5282 if (!path->nodes[level])
5283 return 1;
5284
5285 slot = path->slots[level] + 1;
5286 c = path->nodes[level];
5287next:
5288 if (slot >= btrfs_header_nritems(c)) {
5289 int ret;
5290 int orig_lowest;
5291 struct btrfs_key cur_key;
5292 if (level + 1 >= BTRFS_MAX_LEVEL ||
5293 !path->nodes[level + 1])
5294 return 1;
5295
5296 if (path->locks[level + 1] || path->skip_locking) {
5297 level++;
5298 continue;
5299 }
5300
5301 slot = btrfs_header_nritems(c) - 1;
5302 if (level == 0)
5303 btrfs_item_key_to_cpu(c, &cur_key, slot);
5304 else
5305 btrfs_node_key_to_cpu(c, &cur_key, slot);
5306
5307 orig_lowest = path->lowest_level;
5308 btrfs_release_path(path);
5309 path->lowest_level = level;
5310 ret = btrfs_search_slot(NULL, root, &cur_key, path,
5311 0, 0);
5312 path->lowest_level = orig_lowest;
5313 if (ret < 0)
5314 return ret;
5315
5316 c = path->nodes[level];
5317 slot = path->slots[level];
5318 if (ret == 0)
5319 slot++;
5320 goto next;
5321 }
5322
5323 if (level == 0)
5324 btrfs_item_key_to_cpu(c, key, slot);
5325 else {
5326 u64 gen = btrfs_node_ptr_generation(c, slot);
5327
5328 if (gen < min_trans) {
5329 slot++;
5330 goto next;
5331 }
5332 btrfs_node_key_to_cpu(c, key, slot);
5333 }
5334 return 0;
5335 }
5336 return 1;
5337}
5338
5339/*
5340 * search the tree again to find a leaf with greater keys
5341 * returns 0 if it found something or 1 if there are no greater leaves.
5342 * returns < 0 on io errors.
5343 */
5344int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5345{
5346 return btrfs_next_old_leaf(root, path, 0);
5347}
5348
5349int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5350 u64 time_seq)
5351{
5352 int slot;
5353 int level;
5354 struct extent_buffer *c;
5355 struct extent_buffer *next;
5356 struct btrfs_key key;
5357 u32 nritems;
5358 int ret;
5359 int old_spinning = path->leave_spinning;
5360 int next_rw_lock = 0;
5361
5362 nritems = btrfs_header_nritems(path->nodes[0]);
5363 if (nritems == 0)
5364 return 1;
5365
5366 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5367again:
5368 level = 1;
5369 next = NULL;
5370 next_rw_lock = 0;
5371 btrfs_release_path(path);
5372
5373 path->keep_locks = 1;
5374 path->leave_spinning = 1;
5375
5376 if (time_seq)
5377 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5378 else
5379 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5380 path->keep_locks = 0;
5381
5382 if (ret < 0)
5383 return ret;
5384
5385 nritems = btrfs_header_nritems(path->nodes[0]);
5386 /*
5387 * by releasing the path above we dropped all our locks. A balance
5388 * could have added more items next to the key that used to be
5389 * at the very end of the block. So, check again here and
5390 * advance the path if there are now more items available.
5391 */
5392 if (nritems > 0 && path->slots[0] < nritems - 1) {
5393 if (ret == 0)
5394 path->slots[0]++;
5395 ret = 0;
5396 goto done;
5397 }
5398 /*
5399 * So the above check misses one case:
5400 * - after releasing the path above, someone has removed the item that
5401 * used to be at the very end of the block, and balance between leafs
5402 * gets another one with bigger key.offset to replace it.
5403 *
5404 * This one should be returned as well, or we can get leaf corruption
5405 * later(esp. in __btrfs_drop_extents()).
5406 *
5407 * And a bit more explanation about this check,
5408 * with ret > 0, the key isn't found, the path points to the slot
5409 * where it should be inserted, so the path->slots[0] item must be the
5410 * bigger one.
5411 */
5412 if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5413 ret = 0;
5414 goto done;
5415 }
5416
5417 while (level < BTRFS_MAX_LEVEL) {
5418 if (!path->nodes[level]) {
5419 ret = 1;
5420 goto done;
5421 }
5422
5423 slot = path->slots[level] + 1;
5424 c = path->nodes[level];
5425 if (slot >= btrfs_header_nritems(c)) {
5426 level++;
5427 if (level == BTRFS_MAX_LEVEL) {
5428 ret = 1;
5429 goto done;
5430 }
5431 continue;
5432 }
5433
5434 if (next) {
5435 btrfs_tree_unlock_rw(next, next_rw_lock);
5436 free_extent_buffer(next);
5437 }
5438
5439 next = c;
5440 next_rw_lock = path->locks[level];
5441 ret = read_block_for_search(root, path, &next, level,
5442 slot, &key);
5443 if (ret == -EAGAIN)
5444 goto again;
5445
5446 if (ret < 0) {
5447 btrfs_release_path(path);
5448 goto done;
5449 }
5450
5451 if (!path->skip_locking) {
5452 ret = btrfs_try_tree_read_lock(next);
5453 if (!ret && time_seq) {
5454 /*
5455 * If we don't get the lock, we may be racing
5456 * with push_leaf_left, holding that lock while
5457 * itself waiting for the leaf we've currently
5458 * locked. To solve this situation, we give up
5459 * on our lock and cycle.
5460 */
5461 free_extent_buffer(next);
5462 btrfs_release_path(path);
5463 cond_resched();
5464 goto again;
5465 }
5466 if (!ret) {
5467 btrfs_set_path_blocking(path);
5468 btrfs_tree_read_lock(next);
5469 }
5470 next_rw_lock = BTRFS_READ_LOCK;
5471 }
5472 break;
5473 }
5474 path->slots[level] = slot;
5475 while (1) {
5476 level--;
5477 c = path->nodes[level];
5478 if (path->locks[level])
5479 btrfs_tree_unlock_rw(c, path->locks[level]);
5480
5481 free_extent_buffer(c);
5482 path->nodes[level] = next;
5483 path->slots[level] = 0;
5484 if (!path->skip_locking)
5485 path->locks[level] = next_rw_lock;
5486 if (!level)
5487 break;
5488
5489 ret = read_block_for_search(root, path, &next, level,
5490 0, &key);
5491 if (ret == -EAGAIN)
5492 goto again;
5493
5494 if (ret < 0) {
5495 btrfs_release_path(path);
5496 goto done;
5497 }
5498
5499 if (!path->skip_locking) {
5500 ret = btrfs_try_tree_read_lock(next);
5501 if (!ret) {
5502 btrfs_set_path_blocking(path);
5503 btrfs_tree_read_lock(next);
5504 }
5505 next_rw_lock = BTRFS_READ_LOCK;
5506 }
5507 }
5508 ret = 0;
5509done:
5510 unlock_up(path, 0, 1, 0, NULL);
5511 path->leave_spinning = old_spinning;
5512 if (!old_spinning)
5513 btrfs_set_path_blocking(path);
5514
5515 return ret;
5516}
5517
5518/*
5519 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5520 * searching until it gets past min_objectid or finds an item of 'type'
5521 *
5522 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5523 */
5524int btrfs_previous_item(struct btrfs_root *root,
5525 struct btrfs_path *path, u64 min_objectid,
5526 int type)
5527{
5528 struct btrfs_key found_key;
5529 struct extent_buffer *leaf;
5530 u32 nritems;
5531 int ret;
5532
5533 while (1) {
5534 if (path->slots[0] == 0) {
5535 btrfs_set_path_blocking(path);
5536 ret = btrfs_prev_leaf(root, path);
5537 if (ret != 0)
5538 return ret;
5539 } else {
5540 path->slots[0]--;
5541 }
5542 leaf = path->nodes[0];
5543 nritems = btrfs_header_nritems(leaf);
5544 if (nritems == 0)
5545 return 1;
5546 if (path->slots[0] == nritems)
5547 path->slots[0]--;
5548
5549 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5550 if (found_key.objectid < min_objectid)
5551 break;
5552 if (found_key.type == type)
5553 return 0;
5554 if (found_key.objectid == min_objectid &&
5555 found_key.type < type)
5556 break;
5557 }
5558 return 1;
5559}
5560
5561/*
5562 * search in extent tree to find a previous Metadata/Data extent item with
5563 * min objecitd.
5564 *
5565 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5566 */
5567int btrfs_previous_extent_item(struct btrfs_root *root,
5568 struct btrfs_path *path, u64 min_objectid)
5569{
5570 struct btrfs_key found_key;
5571 struct extent_buffer *leaf;
5572 u32 nritems;
5573 int ret;
5574
5575 while (1) {
5576 if (path->slots[0] == 0) {
5577 btrfs_set_path_blocking(path);
5578 ret = btrfs_prev_leaf(root, path);
5579 if (ret != 0)
5580 return ret;
5581 } else {
5582 path->slots[0]--;
5583 }
5584 leaf = path->nodes[0];
5585 nritems = btrfs_header_nritems(leaf);
5586 if (nritems == 0)
5587 return 1;
5588 if (path->slots[0] == nritems)
5589 path->slots[0]--;
5590
5591 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5592 if (found_key.objectid < min_objectid)
5593 break;
5594 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5595 found_key.type == BTRFS_METADATA_ITEM_KEY)
5596 return 0;
5597 if (found_key.objectid == min_objectid &&
5598 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5599 break;
5600 }
5601 return 1;
5602}
1/*
2 * Copyright (C) 2007,2008 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/sched.h>
20#include <linux/slab.h>
21#include "ctree.h"
22#include "disk-io.h"
23#include "transaction.h"
24#include "print-tree.h"
25#include "locking.h"
26
27static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
28 *root, struct btrfs_path *path, int level);
29static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
30 *root, struct btrfs_key *ins_key,
31 struct btrfs_path *path, int data_size, int extend);
32static int push_node_left(struct btrfs_trans_handle *trans,
33 struct btrfs_root *root, struct extent_buffer *dst,
34 struct extent_buffer *src, int empty);
35static int balance_node_right(struct btrfs_trans_handle *trans,
36 struct btrfs_root *root,
37 struct extent_buffer *dst_buf,
38 struct extent_buffer *src_buf);
39static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
40 struct btrfs_path *path, int level, int slot);
41
42struct btrfs_path *btrfs_alloc_path(void)
43{
44 struct btrfs_path *path;
45 path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
46 return path;
47}
48
49/*
50 * set all locked nodes in the path to blocking locks. This should
51 * be done before scheduling
52 */
53noinline void btrfs_set_path_blocking(struct btrfs_path *p)
54{
55 int i;
56 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
57 if (!p->nodes[i] || !p->locks[i])
58 continue;
59 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
60 if (p->locks[i] == BTRFS_READ_LOCK)
61 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
62 else if (p->locks[i] == BTRFS_WRITE_LOCK)
63 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
64 }
65}
66
67/*
68 * reset all the locked nodes in the patch to spinning locks.
69 *
70 * held is used to keep lockdep happy, when lockdep is enabled
71 * we set held to a blocking lock before we go around and
72 * retake all the spinlocks in the path. You can safely use NULL
73 * for held
74 */
75noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
76 struct extent_buffer *held, int held_rw)
77{
78 int i;
79
80#ifdef CONFIG_DEBUG_LOCK_ALLOC
81 /* lockdep really cares that we take all of these spinlocks
82 * in the right order. If any of the locks in the path are not
83 * currently blocking, it is going to complain. So, make really
84 * really sure by forcing the path to blocking before we clear
85 * the path blocking.
86 */
87 if (held) {
88 btrfs_set_lock_blocking_rw(held, held_rw);
89 if (held_rw == BTRFS_WRITE_LOCK)
90 held_rw = BTRFS_WRITE_LOCK_BLOCKING;
91 else if (held_rw == BTRFS_READ_LOCK)
92 held_rw = BTRFS_READ_LOCK_BLOCKING;
93 }
94 btrfs_set_path_blocking(p);
95#endif
96
97 for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
98 if (p->nodes[i] && p->locks[i]) {
99 btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
100 if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
101 p->locks[i] = BTRFS_WRITE_LOCK;
102 else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
103 p->locks[i] = BTRFS_READ_LOCK;
104 }
105 }
106
107#ifdef CONFIG_DEBUG_LOCK_ALLOC
108 if (held)
109 btrfs_clear_lock_blocking_rw(held, held_rw);
110#endif
111}
112
113/* this also releases the path */
114void btrfs_free_path(struct btrfs_path *p)
115{
116 if (!p)
117 return;
118 btrfs_release_path(p);
119 kmem_cache_free(btrfs_path_cachep, p);
120}
121
122/*
123 * path release drops references on the extent buffers in the path
124 * and it drops any locks held by this path
125 *
126 * It is safe to call this on paths that no locks or extent buffers held.
127 */
128noinline void btrfs_release_path(struct btrfs_path *p)
129{
130 int i;
131
132 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
133 p->slots[i] = 0;
134 if (!p->nodes[i])
135 continue;
136 if (p->locks[i]) {
137 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
138 p->locks[i] = 0;
139 }
140 free_extent_buffer(p->nodes[i]);
141 p->nodes[i] = NULL;
142 }
143}
144
145/*
146 * safely gets a reference on the root node of a tree. A lock
147 * is not taken, so a concurrent writer may put a different node
148 * at the root of the tree. See btrfs_lock_root_node for the
149 * looping required.
150 *
151 * The extent buffer returned by this has a reference taken, so
152 * it won't disappear. It may stop being the root of the tree
153 * at any time because there are no locks held.
154 */
155struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
156{
157 struct extent_buffer *eb;
158
159 rcu_read_lock();
160 eb = rcu_dereference(root->node);
161 extent_buffer_get(eb);
162 rcu_read_unlock();
163 return eb;
164}
165
166/* loop around taking references on and locking the root node of the
167 * tree until you end up with a lock on the root. A locked buffer
168 * is returned, with a reference held.
169 */
170struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
171{
172 struct extent_buffer *eb;
173
174 while (1) {
175 eb = btrfs_root_node(root);
176 btrfs_tree_lock(eb);
177 if (eb == root->node)
178 break;
179 btrfs_tree_unlock(eb);
180 free_extent_buffer(eb);
181 }
182 return eb;
183}
184
185/* loop around taking references on and locking the root node of the
186 * tree until you end up with a lock on the root. A locked buffer
187 * is returned, with a reference held.
188 */
189struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
190{
191 struct extent_buffer *eb;
192
193 while (1) {
194 eb = btrfs_root_node(root);
195 btrfs_tree_read_lock(eb);
196 if (eb == root->node)
197 break;
198 btrfs_tree_read_unlock(eb);
199 free_extent_buffer(eb);
200 }
201 return eb;
202}
203
204/* cowonly root (everything not a reference counted cow subvolume), just get
205 * put onto a simple dirty list. transaction.c walks this to make sure they
206 * get properly updated on disk.
207 */
208static void add_root_to_dirty_list(struct btrfs_root *root)
209{
210 if (root->track_dirty && list_empty(&root->dirty_list)) {
211 list_add(&root->dirty_list,
212 &root->fs_info->dirty_cowonly_roots);
213 }
214}
215
216/*
217 * used by snapshot creation to make a copy of a root for a tree with
218 * a given objectid. The buffer with the new root node is returned in
219 * cow_ret, and this func returns zero on success or a negative error code.
220 */
221int btrfs_copy_root(struct btrfs_trans_handle *trans,
222 struct btrfs_root *root,
223 struct extent_buffer *buf,
224 struct extent_buffer **cow_ret, u64 new_root_objectid)
225{
226 struct extent_buffer *cow;
227 int ret = 0;
228 int level;
229 struct btrfs_disk_key disk_key;
230
231 WARN_ON(root->ref_cows && trans->transid !=
232 root->fs_info->running_transaction->transid);
233 WARN_ON(root->ref_cows && trans->transid != root->last_trans);
234
235 level = btrfs_header_level(buf);
236 if (level == 0)
237 btrfs_item_key(buf, &disk_key, 0);
238 else
239 btrfs_node_key(buf, &disk_key, 0);
240
241 cow = btrfs_alloc_free_block(trans, root, buf->len, 0,
242 new_root_objectid, &disk_key, level,
243 buf->start, 0);
244 if (IS_ERR(cow))
245 return PTR_ERR(cow);
246
247 copy_extent_buffer(cow, buf, 0, 0, cow->len);
248 btrfs_set_header_bytenr(cow, cow->start);
249 btrfs_set_header_generation(cow, trans->transid);
250 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
251 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
252 BTRFS_HEADER_FLAG_RELOC);
253 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
254 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
255 else
256 btrfs_set_header_owner(cow, new_root_objectid);
257
258 write_extent_buffer(cow, root->fs_info->fsid,
259 (unsigned long)btrfs_header_fsid(cow),
260 BTRFS_FSID_SIZE);
261
262 WARN_ON(btrfs_header_generation(buf) > trans->transid);
263 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
264 ret = btrfs_inc_ref(trans, root, cow, 1);
265 else
266 ret = btrfs_inc_ref(trans, root, cow, 0);
267
268 if (ret)
269 return ret;
270
271 btrfs_mark_buffer_dirty(cow);
272 *cow_ret = cow;
273 return 0;
274}
275
276/*
277 * check if the tree block can be shared by multiple trees
278 */
279int btrfs_block_can_be_shared(struct btrfs_root *root,
280 struct extent_buffer *buf)
281{
282 /*
283 * Tree blocks not in refernece counted trees and tree roots
284 * are never shared. If a block was allocated after the last
285 * snapshot and the block was not allocated by tree relocation,
286 * we know the block is not shared.
287 */
288 if (root->ref_cows &&
289 buf != root->node && buf != root->commit_root &&
290 (btrfs_header_generation(buf) <=
291 btrfs_root_last_snapshot(&root->root_item) ||
292 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
293 return 1;
294#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
295 if (root->ref_cows &&
296 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
297 return 1;
298#endif
299 return 0;
300}
301
302static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
303 struct btrfs_root *root,
304 struct extent_buffer *buf,
305 struct extent_buffer *cow,
306 int *last_ref)
307{
308 u64 refs;
309 u64 owner;
310 u64 flags;
311 u64 new_flags = 0;
312 int ret;
313
314 /*
315 * Backrefs update rules:
316 *
317 * Always use full backrefs for extent pointers in tree block
318 * allocated by tree relocation.
319 *
320 * If a shared tree block is no longer referenced by its owner
321 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
322 * use full backrefs for extent pointers in tree block.
323 *
324 * If a tree block is been relocating
325 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
326 * use full backrefs for extent pointers in tree block.
327 * The reason for this is some operations (such as drop tree)
328 * are only allowed for blocks use full backrefs.
329 */
330
331 if (btrfs_block_can_be_shared(root, buf)) {
332 ret = btrfs_lookup_extent_info(trans, root, buf->start,
333 buf->len, &refs, &flags);
334 BUG_ON(ret);
335 BUG_ON(refs == 0);
336 } else {
337 refs = 1;
338 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
339 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
340 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
341 else
342 flags = 0;
343 }
344
345 owner = btrfs_header_owner(buf);
346 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
347 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
348
349 if (refs > 1) {
350 if ((owner == root->root_key.objectid ||
351 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
352 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
353 ret = btrfs_inc_ref(trans, root, buf, 1);
354 BUG_ON(ret);
355
356 if (root->root_key.objectid ==
357 BTRFS_TREE_RELOC_OBJECTID) {
358 ret = btrfs_dec_ref(trans, root, buf, 0);
359 BUG_ON(ret);
360 ret = btrfs_inc_ref(trans, root, cow, 1);
361 BUG_ON(ret);
362 }
363 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
364 } else {
365
366 if (root->root_key.objectid ==
367 BTRFS_TREE_RELOC_OBJECTID)
368 ret = btrfs_inc_ref(trans, root, cow, 1);
369 else
370 ret = btrfs_inc_ref(trans, root, cow, 0);
371 BUG_ON(ret);
372 }
373 if (new_flags != 0) {
374 ret = btrfs_set_disk_extent_flags(trans, root,
375 buf->start,
376 buf->len,
377 new_flags, 0);
378 BUG_ON(ret);
379 }
380 } else {
381 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
382 if (root->root_key.objectid ==
383 BTRFS_TREE_RELOC_OBJECTID)
384 ret = btrfs_inc_ref(trans, root, cow, 1);
385 else
386 ret = btrfs_inc_ref(trans, root, cow, 0);
387 BUG_ON(ret);
388 ret = btrfs_dec_ref(trans, root, buf, 1);
389 BUG_ON(ret);
390 }
391 clean_tree_block(trans, root, buf);
392 *last_ref = 1;
393 }
394 return 0;
395}
396
397/*
398 * does the dirty work in cow of a single block. The parent block (if
399 * supplied) is updated to point to the new cow copy. The new buffer is marked
400 * dirty and returned locked. If you modify the block it needs to be marked
401 * dirty again.
402 *
403 * search_start -- an allocation hint for the new block
404 *
405 * empty_size -- a hint that you plan on doing more cow. This is the size in
406 * bytes the allocator should try to find free next to the block it returns.
407 * This is just a hint and may be ignored by the allocator.
408 */
409static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
410 struct btrfs_root *root,
411 struct extent_buffer *buf,
412 struct extent_buffer *parent, int parent_slot,
413 struct extent_buffer **cow_ret,
414 u64 search_start, u64 empty_size)
415{
416 struct btrfs_disk_key disk_key;
417 struct extent_buffer *cow;
418 int level;
419 int last_ref = 0;
420 int unlock_orig = 0;
421 u64 parent_start;
422
423 if (*cow_ret == buf)
424 unlock_orig = 1;
425
426 btrfs_assert_tree_locked(buf);
427
428 WARN_ON(root->ref_cows && trans->transid !=
429 root->fs_info->running_transaction->transid);
430 WARN_ON(root->ref_cows && trans->transid != root->last_trans);
431
432 level = btrfs_header_level(buf);
433
434 if (level == 0)
435 btrfs_item_key(buf, &disk_key, 0);
436 else
437 btrfs_node_key(buf, &disk_key, 0);
438
439 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
440 if (parent)
441 parent_start = parent->start;
442 else
443 parent_start = 0;
444 } else
445 parent_start = 0;
446
447 cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
448 root->root_key.objectid, &disk_key,
449 level, search_start, empty_size);
450 if (IS_ERR(cow))
451 return PTR_ERR(cow);
452
453 /* cow is set to blocking by btrfs_init_new_buffer */
454
455 copy_extent_buffer(cow, buf, 0, 0, cow->len);
456 btrfs_set_header_bytenr(cow, cow->start);
457 btrfs_set_header_generation(cow, trans->transid);
458 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
459 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
460 BTRFS_HEADER_FLAG_RELOC);
461 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
462 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
463 else
464 btrfs_set_header_owner(cow, root->root_key.objectid);
465
466 write_extent_buffer(cow, root->fs_info->fsid,
467 (unsigned long)btrfs_header_fsid(cow),
468 BTRFS_FSID_SIZE);
469
470 update_ref_for_cow(trans, root, buf, cow, &last_ref);
471
472 if (root->ref_cows)
473 btrfs_reloc_cow_block(trans, root, buf, cow);
474
475 if (buf == root->node) {
476 WARN_ON(parent && parent != buf);
477 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
478 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
479 parent_start = buf->start;
480 else
481 parent_start = 0;
482
483 extent_buffer_get(cow);
484 rcu_assign_pointer(root->node, cow);
485
486 btrfs_free_tree_block(trans, root, buf, parent_start,
487 last_ref);
488 free_extent_buffer(buf);
489 add_root_to_dirty_list(root);
490 } else {
491 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
492 parent_start = parent->start;
493 else
494 parent_start = 0;
495
496 WARN_ON(trans->transid != btrfs_header_generation(parent));
497 btrfs_set_node_blockptr(parent, parent_slot,
498 cow->start);
499 btrfs_set_node_ptr_generation(parent, parent_slot,
500 trans->transid);
501 btrfs_mark_buffer_dirty(parent);
502 btrfs_free_tree_block(trans, root, buf, parent_start,
503 last_ref);
504 }
505 if (unlock_orig)
506 btrfs_tree_unlock(buf);
507 free_extent_buffer(buf);
508 btrfs_mark_buffer_dirty(cow);
509 *cow_ret = cow;
510 return 0;
511}
512
513static inline int should_cow_block(struct btrfs_trans_handle *trans,
514 struct btrfs_root *root,
515 struct extent_buffer *buf)
516{
517 if (btrfs_header_generation(buf) == trans->transid &&
518 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
519 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
520 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
521 return 0;
522 return 1;
523}
524
525/*
526 * cows a single block, see __btrfs_cow_block for the real work.
527 * This version of it has extra checks so that a block isn't cow'd more than
528 * once per transaction, as long as it hasn't been written yet
529 */
530noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
531 struct btrfs_root *root, struct extent_buffer *buf,
532 struct extent_buffer *parent, int parent_slot,
533 struct extent_buffer **cow_ret)
534{
535 u64 search_start;
536 int ret;
537
538 if (trans->transaction != root->fs_info->running_transaction) {
539 printk(KERN_CRIT "trans %llu running %llu\n",
540 (unsigned long long)trans->transid,
541 (unsigned long long)
542 root->fs_info->running_transaction->transid);
543 WARN_ON(1);
544 }
545 if (trans->transid != root->fs_info->generation) {
546 printk(KERN_CRIT "trans %llu running %llu\n",
547 (unsigned long long)trans->transid,
548 (unsigned long long)root->fs_info->generation);
549 WARN_ON(1);
550 }
551
552 if (!should_cow_block(trans, root, buf)) {
553 *cow_ret = buf;
554 return 0;
555 }
556
557 search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
558
559 if (parent)
560 btrfs_set_lock_blocking(parent);
561 btrfs_set_lock_blocking(buf);
562
563 ret = __btrfs_cow_block(trans, root, buf, parent,
564 parent_slot, cow_ret, search_start, 0);
565
566 trace_btrfs_cow_block(root, buf, *cow_ret);
567
568 return ret;
569}
570
571/*
572 * helper function for defrag to decide if two blocks pointed to by a
573 * node are actually close by
574 */
575static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
576{
577 if (blocknr < other && other - (blocknr + blocksize) < 32768)
578 return 1;
579 if (blocknr > other && blocknr - (other + blocksize) < 32768)
580 return 1;
581 return 0;
582}
583
584/*
585 * compare two keys in a memcmp fashion
586 */
587static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
588{
589 struct btrfs_key k1;
590
591 btrfs_disk_key_to_cpu(&k1, disk);
592
593 return btrfs_comp_cpu_keys(&k1, k2);
594}
595
596/*
597 * same as comp_keys only with two btrfs_key's
598 */
599int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
600{
601 if (k1->objectid > k2->objectid)
602 return 1;
603 if (k1->objectid < k2->objectid)
604 return -1;
605 if (k1->type > k2->type)
606 return 1;
607 if (k1->type < k2->type)
608 return -1;
609 if (k1->offset > k2->offset)
610 return 1;
611 if (k1->offset < k2->offset)
612 return -1;
613 return 0;
614}
615
616/*
617 * this is used by the defrag code to go through all the
618 * leaves pointed to by a node and reallocate them so that
619 * disk order is close to key order
620 */
621int btrfs_realloc_node(struct btrfs_trans_handle *trans,
622 struct btrfs_root *root, struct extent_buffer *parent,
623 int start_slot, int cache_only, u64 *last_ret,
624 struct btrfs_key *progress)
625{
626 struct extent_buffer *cur;
627 u64 blocknr;
628 u64 gen;
629 u64 search_start = *last_ret;
630 u64 last_block = 0;
631 u64 other;
632 u32 parent_nritems;
633 int end_slot;
634 int i;
635 int err = 0;
636 int parent_level;
637 int uptodate;
638 u32 blocksize;
639 int progress_passed = 0;
640 struct btrfs_disk_key disk_key;
641
642 parent_level = btrfs_header_level(parent);
643 if (cache_only && parent_level != 1)
644 return 0;
645
646 if (trans->transaction != root->fs_info->running_transaction)
647 WARN_ON(1);
648 if (trans->transid != root->fs_info->generation)
649 WARN_ON(1);
650
651 parent_nritems = btrfs_header_nritems(parent);
652 blocksize = btrfs_level_size(root, parent_level - 1);
653 end_slot = parent_nritems;
654
655 if (parent_nritems == 1)
656 return 0;
657
658 btrfs_set_lock_blocking(parent);
659
660 for (i = start_slot; i < end_slot; i++) {
661 int close = 1;
662
663 btrfs_node_key(parent, &disk_key, i);
664 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
665 continue;
666
667 progress_passed = 1;
668 blocknr = btrfs_node_blockptr(parent, i);
669 gen = btrfs_node_ptr_generation(parent, i);
670 if (last_block == 0)
671 last_block = blocknr;
672
673 if (i > 0) {
674 other = btrfs_node_blockptr(parent, i - 1);
675 close = close_blocks(blocknr, other, blocksize);
676 }
677 if (!close && i < end_slot - 2) {
678 other = btrfs_node_blockptr(parent, i + 1);
679 close = close_blocks(blocknr, other, blocksize);
680 }
681 if (close) {
682 last_block = blocknr;
683 continue;
684 }
685
686 cur = btrfs_find_tree_block(root, blocknr, blocksize);
687 if (cur)
688 uptodate = btrfs_buffer_uptodate(cur, gen);
689 else
690 uptodate = 0;
691 if (!cur || !uptodate) {
692 if (cache_only) {
693 free_extent_buffer(cur);
694 continue;
695 }
696 if (!cur) {
697 cur = read_tree_block(root, blocknr,
698 blocksize, gen);
699 if (!cur)
700 return -EIO;
701 } else if (!uptodate) {
702 btrfs_read_buffer(cur, gen);
703 }
704 }
705 if (search_start == 0)
706 search_start = last_block;
707
708 btrfs_tree_lock(cur);
709 btrfs_set_lock_blocking(cur);
710 err = __btrfs_cow_block(trans, root, cur, parent, i,
711 &cur, search_start,
712 min(16 * blocksize,
713 (end_slot - i) * blocksize));
714 if (err) {
715 btrfs_tree_unlock(cur);
716 free_extent_buffer(cur);
717 break;
718 }
719 search_start = cur->start;
720 last_block = cur->start;
721 *last_ret = search_start;
722 btrfs_tree_unlock(cur);
723 free_extent_buffer(cur);
724 }
725 return err;
726}
727
728/*
729 * The leaf data grows from end-to-front in the node.
730 * this returns the address of the start of the last item,
731 * which is the stop of the leaf data stack
732 */
733static inline unsigned int leaf_data_end(struct btrfs_root *root,
734 struct extent_buffer *leaf)
735{
736 u32 nr = btrfs_header_nritems(leaf);
737 if (nr == 0)
738 return BTRFS_LEAF_DATA_SIZE(root);
739 return btrfs_item_offset_nr(leaf, nr - 1);
740}
741
742
743/*
744 * search for key in the extent_buffer. The items start at offset p,
745 * and they are item_size apart. There are 'max' items in p.
746 *
747 * the slot in the array is returned via slot, and it points to
748 * the place where you would insert key if it is not found in
749 * the array.
750 *
751 * slot may point to max if the key is bigger than all of the keys
752 */
753static noinline int generic_bin_search(struct extent_buffer *eb,
754 unsigned long p,
755 int item_size, struct btrfs_key *key,
756 int max, int *slot)
757{
758 int low = 0;
759 int high = max;
760 int mid;
761 int ret;
762 struct btrfs_disk_key *tmp = NULL;
763 struct btrfs_disk_key unaligned;
764 unsigned long offset;
765 char *kaddr = NULL;
766 unsigned long map_start = 0;
767 unsigned long map_len = 0;
768 int err;
769
770 while (low < high) {
771 mid = (low + high) / 2;
772 offset = p + mid * item_size;
773
774 if (!kaddr || offset < map_start ||
775 (offset + sizeof(struct btrfs_disk_key)) >
776 map_start + map_len) {
777
778 err = map_private_extent_buffer(eb, offset,
779 sizeof(struct btrfs_disk_key),
780 &kaddr, &map_start, &map_len);
781
782 if (!err) {
783 tmp = (struct btrfs_disk_key *)(kaddr + offset -
784 map_start);
785 } else {
786 read_extent_buffer(eb, &unaligned,
787 offset, sizeof(unaligned));
788 tmp = &unaligned;
789 }
790
791 } else {
792 tmp = (struct btrfs_disk_key *)(kaddr + offset -
793 map_start);
794 }
795 ret = comp_keys(tmp, key);
796
797 if (ret < 0)
798 low = mid + 1;
799 else if (ret > 0)
800 high = mid;
801 else {
802 *slot = mid;
803 return 0;
804 }
805 }
806 *slot = low;
807 return 1;
808}
809
810/*
811 * simple bin_search frontend that does the right thing for
812 * leaves vs nodes
813 */
814static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
815 int level, int *slot)
816{
817 if (level == 0) {
818 return generic_bin_search(eb,
819 offsetof(struct btrfs_leaf, items),
820 sizeof(struct btrfs_item),
821 key, btrfs_header_nritems(eb),
822 slot);
823 } else {
824 return generic_bin_search(eb,
825 offsetof(struct btrfs_node, ptrs),
826 sizeof(struct btrfs_key_ptr),
827 key, btrfs_header_nritems(eb),
828 slot);
829 }
830 return -1;
831}
832
833int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
834 int level, int *slot)
835{
836 return bin_search(eb, key, level, slot);
837}
838
839static void root_add_used(struct btrfs_root *root, u32 size)
840{
841 spin_lock(&root->accounting_lock);
842 btrfs_set_root_used(&root->root_item,
843 btrfs_root_used(&root->root_item) + size);
844 spin_unlock(&root->accounting_lock);
845}
846
847static void root_sub_used(struct btrfs_root *root, u32 size)
848{
849 spin_lock(&root->accounting_lock);
850 btrfs_set_root_used(&root->root_item,
851 btrfs_root_used(&root->root_item) - size);
852 spin_unlock(&root->accounting_lock);
853}
854
855/* given a node and slot number, this reads the blocks it points to. The
856 * extent buffer is returned with a reference taken (but unlocked).
857 * NULL is returned on error.
858 */
859static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
860 struct extent_buffer *parent, int slot)
861{
862 int level = btrfs_header_level(parent);
863 if (slot < 0)
864 return NULL;
865 if (slot >= btrfs_header_nritems(parent))
866 return NULL;
867
868 BUG_ON(level == 0);
869
870 return read_tree_block(root, btrfs_node_blockptr(parent, slot),
871 btrfs_level_size(root, level - 1),
872 btrfs_node_ptr_generation(parent, slot));
873}
874
875/*
876 * node level balancing, used to make sure nodes are in proper order for
877 * item deletion. We balance from the top down, so we have to make sure
878 * that a deletion won't leave an node completely empty later on.
879 */
880static noinline int balance_level(struct btrfs_trans_handle *trans,
881 struct btrfs_root *root,
882 struct btrfs_path *path, int level)
883{
884 struct extent_buffer *right = NULL;
885 struct extent_buffer *mid;
886 struct extent_buffer *left = NULL;
887 struct extent_buffer *parent = NULL;
888 int ret = 0;
889 int wret;
890 int pslot;
891 int orig_slot = path->slots[level];
892 u64 orig_ptr;
893
894 if (level == 0)
895 return 0;
896
897 mid = path->nodes[level];
898
899 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
900 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
901 WARN_ON(btrfs_header_generation(mid) != trans->transid);
902
903 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
904
905 if (level < BTRFS_MAX_LEVEL - 1)
906 parent = path->nodes[level + 1];
907 pslot = path->slots[level + 1];
908
909 /*
910 * deal with the case where there is only one pointer in the root
911 * by promoting the node below to a root
912 */
913 if (!parent) {
914 struct extent_buffer *child;
915
916 if (btrfs_header_nritems(mid) != 1)
917 return 0;
918
919 /* promote the child to a root */
920 child = read_node_slot(root, mid, 0);
921 BUG_ON(!child);
922 btrfs_tree_lock(child);
923 btrfs_set_lock_blocking(child);
924 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
925 if (ret) {
926 btrfs_tree_unlock(child);
927 free_extent_buffer(child);
928 goto enospc;
929 }
930
931 rcu_assign_pointer(root->node, child);
932
933 add_root_to_dirty_list(root);
934 btrfs_tree_unlock(child);
935
936 path->locks[level] = 0;
937 path->nodes[level] = NULL;
938 clean_tree_block(trans, root, mid);
939 btrfs_tree_unlock(mid);
940 /* once for the path */
941 free_extent_buffer(mid);
942
943 root_sub_used(root, mid->len);
944 btrfs_free_tree_block(trans, root, mid, 0, 1);
945 /* once for the root ptr */
946 free_extent_buffer(mid);
947 return 0;
948 }
949 if (btrfs_header_nritems(mid) >
950 BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
951 return 0;
952
953 btrfs_header_nritems(mid);
954
955 left = read_node_slot(root, parent, pslot - 1);
956 if (left) {
957 btrfs_tree_lock(left);
958 btrfs_set_lock_blocking(left);
959 wret = btrfs_cow_block(trans, root, left,
960 parent, pslot - 1, &left);
961 if (wret) {
962 ret = wret;
963 goto enospc;
964 }
965 }
966 right = read_node_slot(root, parent, pslot + 1);
967 if (right) {
968 btrfs_tree_lock(right);
969 btrfs_set_lock_blocking(right);
970 wret = btrfs_cow_block(trans, root, right,
971 parent, pslot + 1, &right);
972 if (wret) {
973 ret = wret;
974 goto enospc;
975 }
976 }
977
978 /* first, try to make some room in the middle buffer */
979 if (left) {
980 orig_slot += btrfs_header_nritems(left);
981 wret = push_node_left(trans, root, left, mid, 1);
982 if (wret < 0)
983 ret = wret;
984 btrfs_header_nritems(mid);
985 }
986
987 /*
988 * then try to empty the right most buffer into the middle
989 */
990 if (right) {
991 wret = push_node_left(trans, root, mid, right, 1);
992 if (wret < 0 && wret != -ENOSPC)
993 ret = wret;
994 if (btrfs_header_nritems(right) == 0) {
995 clean_tree_block(trans, root, right);
996 btrfs_tree_unlock(right);
997 wret = del_ptr(trans, root, path, level + 1, pslot +
998 1);
999 if (wret)
1000 ret = wret;
1001 root_sub_used(root, right->len);
1002 btrfs_free_tree_block(trans, root, right, 0, 1);
1003 free_extent_buffer(right);
1004 right = NULL;
1005 } else {
1006 struct btrfs_disk_key right_key;
1007 btrfs_node_key(right, &right_key, 0);
1008 btrfs_set_node_key(parent, &right_key, pslot + 1);
1009 btrfs_mark_buffer_dirty(parent);
1010 }
1011 }
1012 if (btrfs_header_nritems(mid) == 1) {
1013 /*
1014 * we're not allowed to leave a node with one item in the
1015 * tree during a delete. A deletion from lower in the tree
1016 * could try to delete the only pointer in this node.
1017 * So, pull some keys from the left.
1018 * There has to be a left pointer at this point because
1019 * otherwise we would have pulled some pointers from the
1020 * right
1021 */
1022 BUG_ON(!left);
1023 wret = balance_node_right(trans, root, mid, left);
1024 if (wret < 0) {
1025 ret = wret;
1026 goto enospc;
1027 }
1028 if (wret == 1) {
1029 wret = push_node_left(trans, root, left, mid, 1);
1030 if (wret < 0)
1031 ret = wret;
1032 }
1033 BUG_ON(wret == 1);
1034 }
1035 if (btrfs_header_nritems(mid) == 0) {
1036 clean_tree_block(trans, root, mid);
1037 btrfs_tree_unlock(mid);
1038 wret = del_ptr(trans, root, path, level + 1, pslot);
1039 if (wret)
1040 ret = wret;
1041 root_sub_used(root, mid->len);
1042 btrfs_free_tree_block(trans, root, mid, 0, 1);
1043 free_extent_buffer(mid);
1044 mid = NULL;
1045 } else {
1046 /* update the parent key to reflect our changes */
1047 struct btrfs_disk_key mid_key;
1048 btrfs_node_key(mid, &mid_key, 0);
1049 btrfs_set_node_key(parent, &mid_key, pslot);
1050 btrfs_mark_buffer_dirty(parent);
1051 }
1052
1053 /* update the path */
1054 if (left) {
1055 if (btrfs_header_nritems(left) > orig_slot) {
1056 extent_buffer_get(left);
1057 /* left was locked after cow */
1058 path->nodes[level] = left;
1059 path->slots[level + 1] -= 1;
1060 path->slots[level] = orig_slot;
1061 if (mid) {
1062 btrfs_tree_unlock(mid);
1063 free_extent_buffer(mid);
1064 }
1065 } else {
1066 orig_slot -= btrfs_header_nritems(left);
1067 path->slots[level] = orig_slot;
1068 }
1069 }
1070 /* double check we haven't messed things up */
1071 if (orig_ptr !=
1072 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1073 BUG();
1074enospc:
1075 if (right) {
1076 btrfs_tree_unlock(right);
1077 free_extent_buffer(right);
1078 }
1079 if (left) {
1080 if (path->nodes[level] != left)
1081 btrfs_tree_unlock(left);
1082 free_extent_buffer(left);
1083 }
1084 return ret;
1085}
1086
1087/* Node balancing for insertion. Here we only split or push nodes around
1088 * when they are completely full. This is also done top down, so we
1089 * have to be pessimistic.
1090 */
1091static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1092 struct btrfs_root *root,
1093 struct btrfs_path *path, int level)
1094{
1095 struct extent_buffer *right = NULL;
1096 struct extent_buffer *mid;
1097 struct extent_buffer *left = NULL;
1098 struct extent_buffer *parent = NULL;
1099 int ret = 0;
1100 int wret;
1101 int pslot;
1102 int orig_slot = path->slots[level];
1103
1104 if (level == 0)
1105 return 1;
1106
1107 mid = path->nodes[level];
1108 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1109
1110 if (level < BTRFS_MAX_LEVEL - 1)
1111 parent = path->nodes[level + 1];
1112 pslot = path->slots[level + 1];
1113
1114 if (!parent)
1115 return 1;
1116
1117 left = read_node_slot(root, parent, pslot - 1);
1118
1119 /* first, try to make some room in the middle buffer */
1120 if (left) {
1121 u32 left_nr;
1122
1123 btrfs_tree_lock(left);
1124 btrfs_set_lock_blocking(left);
1125
1126 left_nr = btrfs_header_nritems(left);
1127 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1128 wret = 1;
1129 } else {
1130 ret = btrfs_cow_block(trans, root, left, parent,
1131 pslot - 1, &left);
1132 if (ret)
1133 wret = 1;
1134 else {
1135 wret = push_node_left(trans, root,
1136 left, mid, 0);
1137 }
1138 }
1139 if (wret < 0)
1140 ret = wret;
1141 if (wret == 0) {
1142 struct btrfs_disk_key disk_key;
1143 orig_slot += left_nr;
1144 btrfs_node_key(mid, &disk_key, 0);
1145 btrfs_set_node_key(parent, &disk_key, pslot);
1146 btrfs_mark_buffer_dirty(parent);
1147 if (btrfs_header_nritems(left) > orig_slot) {
1148 path->nodes[level] = left;
1149 path->slots[level + 1] -= 1;
1150 path->slots[level] = orig_slot;
1151 btrfs_tree_unlock(mid);
1152 free_extent_buffer(mid);
1153 } else {
1154 orig_slot -=
1155 btrfs_header_nritems(left);
1156 path->slots[level] = orig_slot;
1157 btrfs_tree_unlock(left);
1158 free_extent_buffer(left);
1159 }
1160 return 0;
1161 }
1162 btrfs_tree_unlock(left);
1163 free_extent_buffer(left);
1164 }
1165 right = read_node_slot(root, parent, pslot + 1);
1166
1167 /*
1168 * then try to empty the right most buffer into the middle
1169 */
1170 if (right) {
1171 u32 right_nr;
1172
1173 btrfs_tree_lock(right);
1174 btrfs_set_lock_blocking(right);
1175
1176 right_nr = btrfs_header_nritems(right);
1177 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1178 wret = 1;
1179 } else {
1180 ret = btrfs_cow_block(trans, root, right,
1181 parent, pslot + 1,
1182 &right);
1183 if (ret)
1184 wret = 1;
1185 else {
1186 wret = balance_node_right(trans, root,
1187 right, mid);
1188 }
1189 }
1190 if (wret < 0)
1191 ret = wret;
1192 if (wret == 0) {
1193 struct btrfs_disk_key disk_key;
1194
1195 btrfs_node_key(right, &disk_key, 0);
1196 btrfs_set_node_key(parent, &disk_key, pslot + 1);
1197 btrfs_mark_buffer_dirty(parent);
1198
1199 if (btrfs_header_nritems(mid) <= orig_slot) {
1200 path->nodes[level] = right;
1201 path->slots[level + 1] += 1;
1202 path->slots[level] = orig_slot -
1203 btrfs_header_nritems(mid);
1204 btrfs_tree_unlock(mid);
1205 free_extent_buffer(mid);
1206 } else {
1207 btrfs_tree_unlock(right);
1208 free_extent_buffer(right);
1209 }
1210 return 0;
1211 }
1212 btrfs_tree_unlock(right);
1213 free_extent_buffer(right);
1214 }
1215 return 1;
1216}
1217
1218/*
1219 * readahead one full node of leaves, finding things that are close
1220 * to the block in 'slot', and triggering ra on them.
1221 */
1222static void reada_for_search(struct btrfs_root *root,
1223 struct btrfs_path *path,
1224 int level, int slot, u64 objectid)
1225{
1226 struct extent_buffer *node;
1227 struct btrfs_disk_key disk_key;
1228 u32 nritems;
1229 u64 search;
1230 u64 target;
1231 u64 nread = 0;
1232 u64 gen;
1233 int direction = path->reada;
1234 struct extent_buffer *eb;
1235 u32 nr;
1236 u32 blocksize;
1237 u32 nscan = 0;
1238
1239 if (level != 1)
1240 return;
1241
1242 if (!path->nodes[level])
1243 return;
1244
1245 node = path->nodes[level];
1246
1247 search = btrfs_node_blockptr(node, slot);
1248 blocksize = btrfs_level_size(root, level - 1);
1249 eb = btrfs_find_tree_block(root, search, blocksize);
1250 if (eb) {
1251 free_extent_buffer(eb);
1252 return;
1253 }
1254
1255 target = search;
1256
1257 nritems = btrfs_header_nritems(node);
1258 nr = slot;
1259
1260 while (1) {
1261 if (direction < 0) {
1262 if (nr == 0)
1263 break;
1264 nr--;
1265 } else if (direction > 0) {
1266 nr++;
1267 if (nr >= nritems)
1268 break;
1269 }
1270 if (path->reada < 0 && objectid) {
1271 btrfs_node_key(node, &disk_key, nr);
1272 if (btrfs_disk_key_objectid(&disk_key) != objectid)
1273 break;
1274 }
1275 search = btrfs_node_blockptr(node, nr);
1276 if ((search <= target && target - search <= 65536) ||
1277 (search > target && search - target <= 65536)) {
1278 gen = btrfs_node_ptr_generation(node, nr);
1279 readahead_tree_block(root, search, blocksize, gen);
1280 nread += blocksize;
1281 }
1282 nscan++;
1283 if ((nread > 65536 || nscan > 32))
1284 break;
1285 }
1286}
1287
1288/*
1289 * returns -EAGAIN if it had to drop the path, or zero if everything was in
1290 * cache
1291 */
1292static noinline int reada_for_balance(struct btrfs_root *root,
1293 struct btrfs_path *path, int level)
1294{
1295 int slot;
1296 int nritems;
1297 struct extent_buffer *parent;
1298 struct extent_buffer *eb;
1299 u64 gen;
1300 u64 block1 = 0;
1301 u64 block2 = 0;
1302 int ret = 0;
1303 int blocksize;
1304
1305 parent = path->nodes[level + 1];
1306 if (!parent)
1307 return 0;
1308
1309 nritems = btrfs_header_nritems(parent);
1310 slot = path->slots[level + 1];
1311 blocksize = btrfs_level_size(root, level);
1312
1313 if (slot > 0) {
1314 block1 = btrfs_node_blockptr(parent, slot - 1);
1315 gen = btrfs_node_ptr_generation(parent, slot - 1);
1316 eb = btrfs_find_tree_block(root, block1, blocksize);
1317 if (eb && btrfs_buffer_uptodate(eb, gen))
1318 block1 = 0;
1319 free_extent_buffer(eb);
1320 }
1321 if (slot + 1 < nritems) {
1322 block2 = btrfs_node_blockptr(parent, slot + 1);
1323 gen = btrfs_node_ptr_generation(parent, slot + 1);
1324 eb = btrfs_find_tree_block(root, block2, blocksize);
1325 if (eb && btrfs_buffer_uptodate(eb, gen))
1326 block2 = 0;
1327 free_extent_buffer(eb);
1328 }
1329 if (block1 || block2) {
1330 ret = -EAGAIN;
1331
1332 /* release the whole path */
1333 btrfs_release_path(path);
1334
1335 /* read the blocks */
1336 if (block1)
1337 readahead_tree_block(root, block1, blocksize, 0);
1338 if (block2)
1339 readahead_tree_block(root, block2, blocksize, 0);
1340
1341 if (block1) {
1342 eb = read_tree_block(root, block1, blocksize, 0);
1343 free_extent_buffer(eb);
1344 }
1345 if (block2) {
1346 eb = read_tree_block(root, block2, blocksize, 0);
1347 free_extent_buffer(eb);
1348 }
1349 }
1350 return ret;
1351}
1352
1353
1354/*
1355 * when we walk down the tree, it is usually safe to unlock the higher layers
1356 * in the tree. The exceptions are when our path goes through slot 0, because
1357 * operations on the tree might require changing key pointers higher up in the
1358 * tree.
1359 *
1360 * callers might also have set path->keep_locks, which tells this code to keep
1361 * the lock if the path points to the last slot in the block. This is part of
1362 * walking through the tree, and selecting the next slot in the higher block.
1363 *
1364 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
1365 * if lowest_unlock is 1, level 0 won't be unlocked
1366 */
1367static noinline void unlock_up(struct btrfs_path *path, int level,
1368 int lowest_unlock)
1369{
1370 int i;
1371 int skip_level = level;
1372 int no_skips = 0;
1373 struct extent_buffer *t;
1374
1375 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1376 if (!path->nodes[i])
1377 break;
1378 if (!path->locks[i])
1379 break;
1380 if (!no_skips && path->slots[i] == 0) {
1381 skip_level = i + 1;
1382 continue;
1383 }
1384 if (!no_skips && path->keep_locks) {
1385 u32 nritems;
1386 t = path->nodes[i];
1387 nritems = btrfs_header_nritems(t);
1388 if (nritems < 1 || path->slots[i] >= nritems - 1) {
1389 skip_level = i + 1;
1390 continue;
1391 }
1392 }
1393 if (skip_level < i && i >= lowest_unlock)
1394 no_skips = 1;
1395
1396 t = path->nodes[i];
1397 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1398 btrfs_tree_unlock_rw(t, path->locks[i]);
1399 path->locks[i] = 0;
1400 }
1401 }
1402}
1403
1404/*
1405 * This releases any locks held in the path starting at level and
1406 * going all the way up to the root.
1407 *
1408 * btrfs_search_slot will keep the lock held on higher nodes in a few
1409 * corner cases, such as COW of the block at slot zero in the node. This
1410 * ignores those rules, and it should only be called when there are no
1411 * more updates to be done higher up in the tree.
1412 */
1413noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
1414{
1415 int i;
1416
1417 if (path->keep_locks)
1418 return;
1419
1420 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1421 if (!path->nodes[i])
1422 continue;
1423 if (!path->locks[i])
1424 continue;
1425 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
1426 path->locks[i] = 0;
1427 }
1428}
1429
1430/*
1431 * helper function for btrfs_search_slot. The goal is to find a block
1432 * in cache without setting the path to blocking. If we find the block
1433 * we return zero and the path is unchanged.
1434 *
1435 * If we can't find the block, we set the path blocking and do some
1436 * reada. -EAGAIN is returned and the search must be repeated.
1437 */
1438static int
1439read_block_for_search(struct btrfs_trans_handle *trans,
1440 struct btrfs_root *root, struct btrfs_path *p,
1441 struct extent_buffer **eb_ret, int level, int slot,
1442 struct btrfs_key *key)
1443{
1444 u64 blocknr;
1445 u64 gen;
1446 u32 blocksize;
1447 struct extent_buffer *b = *eb_ret;
1448 struct extent_buffer *tmp;
1449 int ret;
1450
1451 blocknr = btrfs_node_blockptr(b, slot);
1452 gen = btrfs_node_ptr_generation(b, slot);
1453 blocksize = btrfs_level_size(root, level - 1);
1454
1455 tmp = btrfs_find_tree_block(root, blocknr, blocksize);
1456 if (tmp) {
1457 if (btrfs_buffer_uptodate(tmp, 0)) {
1458 if (btrfs_buffer_uptodate(tmp, gen)) {
1459 /*
1460 * we found an up to date block without
1461 * sleeping, return
1462 * right away
1463 */
1464 *eb_ret = tmp;
1465 return 0;
1466 }
1467 /* the pages were up to date, but we failed
1468 * the generation number check. Do a full
1469 * read for the generation number that is correct.
1470 * We must do this without dropping locks so
1471 * we can trust our generation number
1472 */
1473 free_extent_buffer(tmp);
1474 btrfs_set_path_blocking(p);
1475
1476 tmp = read_tree_block(root, blocknr, blocksize, gen);
1477 if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
1478 *eb_ret = tmp;
1479 return 0;
1480 }
1481 free_extent_buffer(tmp);
1482 btrfs_release_path(p);
1483 return -EIO;
1484 }
1485 }
1486
1487 /*
1488 * reduce lock contention at high levels
1489 * of the btree by dropping locks before
1490 * we read. Don't release the lock on the current
1491 * level because we need to walk this node to figure
1492 * out which blocks to read.
1493 */
1494 btrfs_unlock_up_safe(p, level + 1);
1495 btrfs_set_path_blocking(p);
1496
1497 free_extent_buffer(tmp);
1498 if (p->reada)
1499 reada_for_search(root, p, level, slot, key->objectid);
1500
1501 btrfs_release_path(p);
1502
1503 ret = -EAGAIN;
1504 tmp = read_tree_block(root, blocknr, blocksize, 0);
1505 if (tmp) {
1506 /*
1507 * If the read above didn't mark this buffer up to date,
1508 * it will never end up being up to date. Set ret to EIO now
1509 * and give up so that our caller doesn't loop forever
1510 * on our EAGAINs.
1511 */
1512 if (!btrfs_buffer_uptodate(tmp, 0))
1513 ret = -EIO;
1514 free_extent_buffer(tmp);
1515 }
1516 return ret;
1517}
1518
1519/*
1520 * helper function for btrfs_search_slot. This does all of the checks
1521 * for node-level blocks and does any balancing required based on
1522 * the ins_len.
1523 *
1524 * If no extra work was required, zero is returned. If we had to
1525 * drop the path, -EAGAIN is returned and btrfs_search_slot must
1526 * start over
1527 */
1528static int
1529setup_nodes_for_search(struct btrfs_trans_handle *trans,
1530 struct btrfs_root *root, struct btrfs_path *p,
1531 struct extent_buffer *b, int level, int ins_len,
1532 int *write_lock_level)
1533{
1534 int ret;
1535 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1536 BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1537 int sret;
1538
1539 if (*write_lock_level < level + 1) {
1540 *write_lock_level = level + 1;
1541 btrfs_release_path(p);
1542 goto again;
1543 }
1544
1545 sret = reada_for_balance(root, p, level);
1546 if (sret)
1547 goto again;
1548
1549 btrfs_set_path_blocking(p);
1550 sret = split_node(trans, root, p, level);
1551 btrfs_clear_path_blocking(p, NULL, 0);
1552
1553 BUG_ON(sret > 0);
1554 if (sret) {
1555 ret = sret;
1556 goto done;
1557 }
1558 b = p->nodes[level];
1559 } else if (ins_len < 0 && btrfs_header_nritems(b) <
1560 BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
1561 int sret;
1562
1563 if (*write_lock_level < level + 1) {
1564 *write_lock_level = level + 1;
1565 btrfs_release_path(p);
1566 goto again;
1567 }
1568
1569 sret = reada_for_balance(root, p, level);
1570 if (sret)
1571 goto again;
1572
1573 btrfs_set_path_blocking(p);
1574 sret = balance_level(trans, root, p, level);
1575 btrfs_clear_path_blocking(p, NULL, 0);
1576
1577 if (sret) {
1578 ret = sret;
1579 goto done;
1580 }
1581 b = p->nodes[level];
1582 if (!b) {
1583 btrfs_release_path(p);
1584 goto again;
1585 }
1586 BUG_ON(btrfs_header_nritems(b) == 1);
1587 }
1588 return 0;
1589
1590again:
1591 ret = -EAGAIN;
1592done:
1593 return ret;
1594}
1595
1596/*
1597 * look for key in the tree. path is filled in with nodes along the way
1598 * if key is found, we return zero and you can find the item in the leaf
1599 * level of the path (level 0)
1600 *
1601 * If the key isn't found, the path points to the slot where it should
1602 * be inserted, and 1 is returned. If there are other errors during the
1603 * search a negative error number is returned.
1604 *
1605 * if ins_len > 0, nodes and leaves will be split as we walk down the
1606 * tree. if ins_len < 0, nodes will be merged as we walk down the tree (if
1607 * possible)
1608 */
1609int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1610 *root, struct btrfs_key *key, struct btrfs_path *p, int
1611 ins_len, int cow)
1612{
1613 struct extent_buffer *b;
1614 int slot;
1615 int ret;
1616 int err;
1617 int level;
1618 int lowest_unlock = 1;
1619 int root_lock;
1620 /* everything at write_lock_level or lower must be write locked */
1621 int write_lock_level = 0;
1622 u8 lowest_level = 0;
1623
1624 lowest_level = p->lowest_level;
1625 WARN_ON(lowest_level && ins_len > 0);
1626 WARN_ON(p->nodes[0] != NULL);
1627
1628 if (ins_len < 0) {
1629 lowest_unlock = 2;
1630
1631 /* when we are removing items, we might have to go up to level
1632 * two as we update tree pointers Make sure we keep write
1633 * for those levels as well
1634 */
1635 write_lock_level = 2;
1636 } else if (ins_len > 0) {
1637 /*
1638 * for inserting items, make sure we have a write lock on
1639 * level 1 so we can update keys
1640 */
1641 write_lock_level = 1;
1642 }
1643
1644 if (!cow)
1645 write_lock_level = -1;
1646
1647 if (cow && (p->keep_locks || p->lowest_level))
1648 write_lock_level = BTRFS_MAX_LEVEL;
1649
1650again:
1651 /*
1652 * we try very hard to do read locks on the root
1653 */
1654 root_lock = BTRFS_READ_LOCK;
1655 level = 0;
1656 if (p->search_commit_root) {
1657 /*
1658 * the commit roots are read only
1659 * so we always do read locks
1660 */
1661 b = root->commit_root;
1662 extent_buffer_get(b);
1663 level = btrfs_header_level(b);
1664 if (!p->skip_locking)
1665 btrfs_tree_read_lock(b);
1666 } else {
1667 if (p->skip_locking) {
1668 b = btrfs_root_node(root);
1669 level = btrfs_header_level(b);
1670 } else {
1671 /* we don't know the level of the root node
1672 * until we actually have it read locked
1673 */
1674 b = btrfs_read_lock_root_node(root);
1675 level = btrfs_header_level(b);
1676 if (level <= write_lock_level) {
1677 /* whoops, must trade for write lock */
1678 btrfs_tree_read_unlock(b);
1679 free_extent_buffer(b);
1680 b = btrfs_lock_root_node(root);
1681 root_lock = BTRFS_WRITE_LOCK;
1682
1683 /* the level might have changed, check again */
1684 level = btrfs_header_level(b);
1685 }
1686 }
1687 }
1688 p->nodes[level] = b;
1689 if (!p->skip_locking)
1690 p->locks[level] = root_lock;
1691
1692 while (b) {
1693 level = btrfs_header_level(b);
1694
1695 /*
1696 * setup the path here so we can release it under lock
1697 * contention with the cow code
1698 */
1699 if (cow) {
1700 /*
1701 * if we don't really need to cow this block
1702 * then we don't want to set the path blocking,
1703 * so we test it here
1704 */
1705 if (!should_cow_block(trans, root, b))
1706 goto cow_done;
1707
1708 btrfs_set_path_blocking(p);
1709
1710 /*
1711 * must have write locks on this node and the
1712 * parent
1713 */
1714 if (level + 1 > write_lock_level) {
1715 write_lock_level = level + 1;
1716 btrfs_release_path(p);
1717 goto again;
1718 }
1719
1720 err = btrfs_cow_block(trans, root, b,
1721 p->nodes[level + 1],
1722 p->slots[level + 1], &b);
1723 if (err) {
1724 ret = err;
1725 goto done;
1726 }
1727 }
1728cow_done:
1729 BUG_ON(!cow && ins_len);
1730
1731 p->nodes[level] = b;
1732 btrfs_clear_path_blocking(p, NULL, 0);
1733
1734 /*
1735 * we have a lock on b and as long as we aren't changing
1736 * the tree, there is no way to for the items in b to change.
1737 * It is safe to drop the lock on our parent before we
1738 * go through the expensive btree search on b.
1739 *
1740 * If cow is true, then we might be changing slot zero,
1741 * which may require changing the parent. So, we can't
1742 * drop the lock until after we know which slot we're
1743 * operating on.
1744 */
1745 if (!cow)
1746 btrfs_unlock_up_safe(p, level + 1);
1747
1748 ret = bin_search(b, key, level, &slot);
1749
1750 if (level != 0) {
1751 int dec = 0;
1752 if (ret && slot > 0) {
1753 dec = 1;
1754 slot -= 1;
1755 }
1756 p->slots[level] = slot;
1757 err = setup_nodes_for_search(trans, root, p, b, level,
1758 ins_len, &write_lock_level);
1759 if (err == -EAGAIN)
1760 goto again;
1761 if (err) {
1762 ret = err;
1763 goto done;
1764 }
1765 b = p->nodes[level];
1766 slot = p->slots[level];
1767
1768 /*
1769 * slot 0 is special, if we change the key
1770 * we have to update the parent pointer
1771 * which means we must have a write lock
1772 * on the parent
1773 */
1774 if (slot == 0 && cow &&
1775 write_lock_level < level + 1) {
1776 write_lock_level = level + 1;
1777 btrfs_release_path(p);
1778 goto again;
1779 }
1780
1781 unlock_up(p, level, lowest_unlock);
1782
1783 if (level == lowest_level) {
1784 if (dec)
1785 p->slots[level]++;
1786 goto done;
1787 }
1788
1789 err = read_block_for_search(trans, root, p,
1790 &b, level, slot, key);
1791 if (err == -EAGAIN)
1792 goto again;
1793 if (err) {
1794 ret = err;
1795 goto done;
1796 }
1797
1798 if (!p->skip_locking) {
1799 level = btrfs_header_level(b);
1800 if (level <= write_lock_level) {
1801 err = btrfs_try_tree_write_lock(b);
1802 if (!err) {
1803 btrfs_set_path_blocking(p);
1804 btrfs_tree_lock(b);
1805 btrfs_clear_path_blocking(p, b,
1806 BTRFS_WRITE_LOCK);
1807 }
1808 p->locks[level] = BTRFS_WRITE_LOCK;
1809 } else {
1810 err = btrfs_try_tree_read_lock(b);
1811 if (!err) {
1812 btrfs_set_path_blocking(p);
1813 btrfs_tree_read_lock(b);
1814 btrfs_clear_path_blocking(p, b,
1815 BTRFS_READ_LOCK);
1816 }
1817 p->locks[level] = BTRFS_READ_LOCK;
1818 }
1819 p->nodes[level] = b;
1820 }
1821 } else {
1822 p->slots[level] = slot;
1823 if (ins_len > 0 &&
1824 btrfs_leaf_free_space(root, b) < ins_len) {
1825 if (write_lock_level < 1) {
1826 write_lock_level = 1;
1827 btrfs_release_path(p);
1828 goto again;
1829 }
1830
1831 btrfs_set_path_blocking(p);
1832 err = split_leaf(trans, root, key,
1833 p, ins_len, ret == 0);
1834 btrfs_clear_path_blocking(p, NULL, 0);
1835
1836 BUG_ON(err > 0);
1837 if (err) {
1838 ret = err;
1839 goto done;
1840 }
1841 }
1842 if (!p->search_for_split)
1843 unlock_up(p, level, lowest_unlock);
1844 goto done;
1845 }
1846 }
1847 ret = 1;
1848done:
1849 /*
1850 * we don't really know what they plan on doing with the path
1851 * from here on, so for now just mark it as blocking
1852 */
1853 if (!p->leave_spinning)
1854 btrfs_set_path_blocking(p);
1855 if (ret < 0)
1856 btrfs_release_path(p);
1857 return ret;
1858}
1859
1860/*
1861 * adjust the pointers going up the tree, starting at level
1862 * making sure the right key of each node is points to 'key'.
1863 * This is used after shifting pointers to the left, so it stops
1864 * fixing up pointers when a given leaf/node is not in slot 0 of the
1865 * higher levels
1866 *
1867 * If this fails to write a tree block, it returns -1, but continues
1868 * fixing up the blocks in ram so the tree is consistent.
1869 */
1870static int fixup_low_keys(struct btrfs_trans_handle *trans,
1871 struct btrfs_root *root, struct btrfs_path *path,
1872 struct btrfs_disk_key *key, int level)
1873{
1874 int i;
1875 int ret = 0;
1876 struct extent_buffer *t;
1877
1878 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1879 int tslot = path->slots[i];
1880 if (!path->nodes[i])
1881 break;
1882 t = path->nodes[i];
1883 btrfs_set_node_key(t, key, tslot);
1884 btrfs_mark_buffer_dirty(path->nodes[i]);
1885 if (tslot != 0)
1886 break;
1887 }
1888 return ret;
1889}
1890
1891/*
1892 * update item key.
1893 *
1894 * This function isn't completely safe. It's the caller's responsibility
1895 * that the new key won't break the order
1896 */
1897int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1898 struct btrfs_root *root, struct btrfs_path *path,
1899 struct btrfs_key *new_key)
1900{
1901 struct btrfs_disk_key disk_key;
1902 struct extent_buffer *eb;
1903 int slot;
1904
1905 eb = path->nodes[0];
1906 slot = path->slots[0];
1907 if (slot > 0) {
1908 btrfs_item_key(eb, &disk_key, slot - 1);
1909 if (comp_keys(&disk_key, new_key) >= 0)
1910 return -1;
1911 }
1912 if (slot < btrfs_header_nritems(eb) - 1) {
1913 btrfs_item_key(eb, &disk_key, slot + 1);
1914 if (comp_keys(&disk_key, new_key) <= 0)
1915 return -1;
1916 }
1917
1918 btrfs_cpu_key_to_disk(&disk_key, new_key);
1919 btrfs_set_item_key(eb, &disk_key, slot);
1920 btrfs_mark_buffer_dirty(eb);
1921 if (slot == 0)
1922 fixup_low_keys(trans, root, path, &disk_key, 1);
1923 return 0;
1924}
1925
1926/*
1927 * try to push data from one node into the next node left in the
1928 * tree.
1929 *
1930 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1931 * error, and > 0 if there was no room in the left hand block.
1932 */
1933static int push_node_left(struct btrfs_trans_handle *trans,
1934 struct btrfs_root *root, struct extent_buffer *dst,
1935 struct extent_buffer *src, int empty)
1936{
1937 int push_items = 0;
1938 int src_nritems;
1939 int dst_nritems;
1940 int ret = 0;
1941
1942 src_nritems = btrfs_header_nritems(src);
1943 dst_nritems = btrfs_header_nritems(dst);
1944 push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1945 WARN_ON(btrfs_header_generation(src) != trans->transid);
1946 WARN_ON(btrfs_header_generation(dst) != trans->transid);
1947
1948 if (!empty && src_nritems <= 8)
1949 return 1;
1950
1951 if (push_items <= 0)
1952 return 1;
1953
1954 if (empty) {
1955 push_items = min(src_nritems, push_items);
1956 if (push_items < src_nritems) {
1957 /* leave at least 8 pointers in the node if
1958 * we aren't going to empty it
1959 */
1960 if (src_nritems - push_items < 8) {
1961 if (push_items <= 8)
1962 return 1;
1963 push_items -= 8;
1964 }
1965 }
1966 } else
1967 push_items = min(src_nritems - 8, push_items);
1968
1969 copy_extent_buffer(dst, src,
1970 btrfs_node_key_ptr_offset(dst_nritems),
1971 btrfs_node_key_ptr_offset(0),
1972 push_items * sizeof(struct btrfs_key_ptr));
1973
1974 if (push_items < src_nritems) {
1975 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1976 btrfs_node_key_ptr_offset(push_items),
1977 (src_nritems - push_items) *
1978 sizeof(struct btrfs_key_ptr));
1979 }
1980 btrfs_set_header_nritems(src, src_nritems - push_items);
1981 btrfs_set_header_nritems(dst, dst_nritems + push_items);
1982 btrfs_mark_buffer_dirty(src);
1983 btrfs_mark_buffer_dirty(dst);
1984
1985 return ret;
1986}
1987
1988/*
1989 * try to push data from one node into the next node right in the
1990 * tree.
1991 *
1992 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1993 * error, and > 0 if there was no room in the right hand block.
1994 *
1995 * this will only push up to 1/2 the contents of the left node over
1996 */
1997static int balance_node_right(struct btrfs_trans_handle *trans,
1998 struct btrfs_root *root,
1999 struct extent_buffer *dst,
2000 struct extent_buffer *src)
2001{
2002 int push_items = 0;
2003 int max_push;
2004 int src_nritems;
2005 int dst_nritems;
2006 int ret = 0;
2007
2008 WARN_ON(btrfs_header_generation(src) != trans->transid);
2009 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2010
2011 src_nritems = btrfs_header_nritems(src);
2012 dst_nritems = btrfs_header_nritems(dst);
2013 push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2014 if (push_items <= 0)
2015 return 1;
2016
2017 if (src_nritems < 4)
2018 return 1;
2019
2020 max_push = src_nritems / 2 + 1;
2021 /* don't try to empty the node */
2022 if (max_push >= src_nritems)
2023 return 1;
2024
2025 if (max_push < push_items)
2026 push_items = max_push;
2027
2028 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2029 btrfs_node_key_ptr_offset(0),
2030 (dst_nritems) *
2031 sizeof(struct btrfs_key_ptr));
2032
2033 copy_extent_buffer(dst, src,
2034 btrfs_node_key_ptr_offset(0),
2035 btrfs_node_key_ptr_offset(src_nritems - push_items),
2036 push_items * sizeof(struct btrfs_key_ptr));
2037
2038 btrfs_set_header_nritems(src, src_nritems - push_items);
2039 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2040
2041 btrfs_mark_buffer_dirty(src);
2042 btrfs_mark_buffer_dirty(dst);
2043
2044 return ret;
2045}
2046
2047/*
2048 * helper function to insert a new root level in the tree.
2049 * A new node is allocated, and a single item is inserted to
2050 * point to the existing root
2051 *
2052 * returns zero on success or < 0 on failure.
2053 */
2054static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2055 struct btrfs_root *root,
2056 struct btrfs_path *path, int level)
2057{
2058 u64 lower_gen;
2059 struct extent_buffer *lower;
2060 struct extent_buffer *c;
2061 struct extent_buffer *old;
2062 struct btrfs_disk_key lower_key;
2063
2064 BUG_ON(path->nodes[level]);
2065 BUG_ON(path->nodes[level-1] != root->node);
2066
2067 lower = path->nodes[level-1];
2068 if (level == 1)
2069 btrfs_item_key(lower, &lower_key, 0);
2070 else
2071 btrfs_node_key(lower, &lower_key, 0);
2072
2073 c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2074 root->root_key.objectid, &lower_key,
2075 level, root->node->start, 0);
2076 if (IS_ERR(c))
2077 return PTR_ERR(c);
2078
2079 root_add_used(root, root->nodesize);
2080
2081 memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
2082 btrfs_set_header_nritems(c, 1);
2083 btrfs_set_header_level(c, level);
2084 btrfs_set_header_bytenr(c, c->start);
2085 btrfs_set_header_generation(c, trans->transid);
2086 btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
2087 btrfs_set_header_owner(c, root->root_key.objectid);
2088
2089 write_extent_buffer(c, root->fs_info->fsid,
2090 (unsigned long)btrfs_header_fsid(c),
2091 BTRFS_FSID_SIZE);
2092
2093 write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
2094 (unsigned long)btrfs_header_chunk_tree_uuid(c),
2095 BTRFS_UUID_SIZE);
2096
2097 btrfs_set_node_key(c, &lower_key, 0);
2098 btrfs_set_node_blockptr(c, 0, lower->start);
2099 lower_gen = btrfs_header_generation(lower);
2100 WARN_ON(lower_gen != trans->transid);
2101
2102 btrfs_set_node_ptr_generation(c, 0, lower_gen);
2103
2104 btrfs_mark_buffer_dirty(c);
2105
2106 old = root->node;
2107 rcu_assign_pointer(root->node, c);
2108
2109 /* the super has an extra ref to root->node */
2110 free_extent_buffer(old);
2111
2112 add_root_to_dirty_list(root);
2113 extent_buffer_get(c);
2114 path->nodes[level] = c;
2115 path->locks[level] = BTRFS_WRITE_LOCK;
2116 path->slots[level] = 0;
2117 return 0;
2118}
2119
2120/*
2121 * worker function to insert a single pointer in a node.
2122 * the node should have enough room for the pointer already
2123 *
2124 * slot and level indicate where you want the key to go, and
2125 * blocknr is the block the key points to.
2126 *
2127 * returns zero on success and < 0 on any error
2128 */
2129static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
2130 *root, struct btrfs_path *path, struct btrfs_disk_key
2131 *key, u64 bytenr, int slot, int level)
2132{
2133 struct extent_buffer *lower;
2134 int nritems;
2135
2136 BUG_ON(!path->nodes[level]);
2137 btrfs_assert_tree_locked(path->nodes[level]);
2138 lower = path->nodes[level];
2139 nritems = btrfs_header_nritems(lower);
2140 BUG_ON(slot > nritems);
2141 if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
2142 BUG();
2143 if (slot != nritems) {
2144 memmove_extent_buffer(lower,
2145 btrfs_node_key_ptr_offset(slot + 1),
2146 btrfs_node_key_ptr_offset(slot),
2147 (nritems - slot) * sizeof(struct btrfs_key_ptr));
2148 }
2149 btrfs_set_node_key(lower, key, slot);
2150 btrfs_set_node_blockptr(lower, slot, bytenr);
2151 WARN_ON(trans->transid == 0);
2152 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2153 btrfs_set_header_nritems(lower, nritems + 1);
2154 btrfs_mark_buffer_dirty(lower);
2155 return 0;
2156}
2157
2158/*
2159 * split the node at the specified level in path in two.
2160 * The path is corrected to point to the appropriate node after the split
2161 *
2162 * Before splitting this tries to make some room in the node by pushing
2163 * left and right, if either one works, it returns right away.
2164 *
2165 * returns 0 on success and < 0 on failure
2166 */
2167static noinline int split_node(struct btrfs_trans_handle *trans,
2168 struct btrfs_root *root,
2169 struct btrfs_path *path, int level)
2170{
2171 struct extent_buffer *c;
2172 struct extent_buffer *split;
2173 struct btrfs_disk_key disk_key;
2174 int mid;
2175 int ret;
2176 int wret;
2177 u32 c_nritems;
2178
2179 c = path->nodes[level];
2180 WARN_ON(btrfs_header_generation(c) != trans->transid);
2181 if (c == root->node) {
2182 /* trying to split the root, lets make a new one */
2183 ret = insert_new_root(trans, root, path, level + 1);
2184 if (ret)
2185 return ret;
2186 } else {
2187 ret = push_nodes_for_insert(trans, root, path, level);
2188 c = path->nodes[level];
2189 if (!ret && btrfs_header_nritems(c) <
2190 BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
2191 return 0;
2192 if (ret < 0)
2193 return ret;
2194 }
2195
2196 c_nritems = btrfs_header_nritems(c);
2197 mid = (c_nritems + 1) / 2;
2198 btrfs_node_key(c, &disk_key, mid);
2199
2200 split = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2201 root->root_key.objectid,
2202 &disk_key, level, c->start, 0);
2203 if (IS_ERR(split))
2204 return PTR_ERR(split);
2205
2206 root_add_used(root, root->nodesize);
2207
2208 memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
2209 btrfs_set_header_level(split, btrfs_header_level(c));
2210 btrfs_set_header_bytenr(split, split->start);
2211 btrfs_set_header_generation(split, trans->transid);
2212 btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
2213 btrfs_set_header_owner(split, root->root_key.objectid);
2214 write_extent_buffer(split, root->fs_info->fsid,
2215 (unsigned long)btrfs_header_fsid(split),
2216 BTRFS_FSID_SIZE);
2217 write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
2218 (unsigned long)btrfs_header_chunk_tree_uuid(split),
2219 BTRFS_UUID_SIZE);
2220
2221
2222 copy_extent_buffer(split, c,
2223 btrfs_node_key_ptr_offset(0),
2224 btrfs_node_key_ptr_offset(mid),
2225 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2226 btrfs_set_header_nritems(split, c_nritems - mid);
2227 btrfs_set_header_nritems(c, mid);
2228 ret = 0;
2229
2230 btrfs_mark_buffer_dirty(c);
2231 btrfs_mark_buffer_dirty(split);
2232
2233 wret = insert_ptr(trans, root, path, &disk_key, split->start,
2234 path->slots[level + 1] + 1,
2235 level + 1);
2236 if (wret)
2237 ret = wret;
2238
2239 if (path->slots[level] >= mid) {
2240 path->slots[level] -= mid;
2241 btrfs_tree_unlock(c);
2242 free_extent_buffer(c);
2243 path->nodes[level] = split;
2244 path->slots[level + 1] += 1;
2245 } else {
2246 btrfs_tree_unlock(split);
2247 free_extent_buffer(split);
2248 }
2249 return ret;
2250}
2251
2252/*
2253 * how many bytes are required to store the items in a leaf. start
2254 * and nr indicate which items in the leaf to check. This totals up the
2255 * space used both by the item structs and the item data
2256 */
2257static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2258{
2259 int data_len;
2260 int nritems = btrfs_header_nritems(l);
2261 int end = min(nritems, start + nr) - 1;
2262
2263 if (!nr)
2264 return 0;
2265 data_len = btrfs_item_end_nr(l, start);
2266 data_len = data_len - btrfs_item_offset_nr(l, end);
2267 data_len += sizeof(struct btrfs_item) * nr;
2268 WARN_ON(data_len < 0);
2269 return data_len;
2270}
2271
2272/*
2273 * The space between the end of the leaf items and
2274 * the start of the leaf data. IOW, how much room
2275 * the leaf has left for both items and data
2276 */
2277noinline int btrfs_leaf_free_space(struct btrfs_root *root,
2278 struct extent_buffer *leaf)
2279{
2280 int nritems = btrfs_header_nritems(leaf);
2281 int ret;
2282 ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
2283 if (ret < 0) {
2284 printk(KERN_CRIT "leaf free space ret %d, leaf data size %lu, "
2285 "used %d nritems %d\n",
2286 ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
2287 leaf_space_used(leaf, 0, nritems), nritems);
2288 }
2289 return ret;
2290}
2291
2292/*
2293 * min slot controls the lowest index we're willing to push to the
2294 * right. We'll push up to and including min_slot, but no lower
2295 */
2296static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
2297 struct btrfs_root *root,
2298 struct btrfs_path *path,
2299 int data_size, int empty,
2300 struct extent_buffer *right,
2301 int free_space, u32 left_nritems,
2302 u32 min_slot)
2303{
2304 struct extent_buffer *left = path->nodes[0];
2305 struct extent_buffer *upper = path->nodes[1];
2306 struct btrfs_disk_key disk_key;
2307 int slot;
2308 u32 i;
2309 int push_space = 0;
2310 int push_items = 0;
2311 struct btrfs_item *item;
2312 u32 nr;
2313 u32 right_nritems;
2314 u32 data_end;
2315 u32 this_item_size;
2316
2317 if (empty)
2318 nr = 0;
2319 else
2320 nr = max_t(u32, 1, min_slot);
2321
2322 if (path->slots[0] >= left_nritems)
2323 push_space += data_size;
2324
2325 slot = path->slots[1];
2326 i = left_nritems - 1;
2327 while (i >= nr) {
2328 item = btrfs_item_nr(left, i);
2329
2330 if (!empty && push_items > 0) {
2331 if (path->slots[0] > i)
2332 break;
2333 if (path->slots[0] == i) {
2334 int space = btrfs_leaf_free_space(root, left);
2335 if (space + push_space * 2 > free_space)
2336 break;
2337 }
2338 }
2339
2340 if (path->slots[0] == i)
2341 push_space += data_size;
2342
2343 this_item_size = btrfs_item_size(left, item);
2344 if (this_item_size + sizeof(*item) + push_space > free_space)
2345 break;
2346
2347 push_items++;
2348 push_space += this_item_size + sizeof(*item);
2349 if (i == 0)
2350 break;
2351 i--;
2352 }
2353
2354 if (push_items == 0)
2355 goto out_unlock;
2356
2357 if (!empty && push_items == left_nritems)
2358 WARN_ON(1);
2359
2360 /* push left to right */
2361 right_nritems = btrfs_header_nritems(right);
2362
2363 push_space = btrfs_item_end_nr(left, left_nritems - push_items);
2364 push_space -= leaf_data_end(root, left);
2365
2366 /* make room in the right data area */
2367 data_end = leaf_data_end(root, right);
2368 memmove_extent_buffer(right,
2369 btrfs_leaf_data(right) + data_end - push_space,
2370 btrfs_leaf_data(right) + data_end,
2371 BTRFS_LEAF_DATA_SIZE(root) - data_end);
2372
2373 /* copy from the left data area */
2374 copy_extent_buffer(right, left, btrfs_leaf_data(right) +
2375 BTRFS_LEAF_DATA_SIZE(root) - push_space,
2376 btrfs_leaf_data(left) + leaf_data_end(root, left),
2377 push_space);
2378
2379 memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2380 btrfs_item_nr_offset(0),
2381 right_nritems * sizeof(struct btrfs_item));
2382
2383 /* copy the items from left to right */
2384 copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2385 btrfs_item_nr_offset(left_nritems - push_items),
2386 push_items * sizeof(struct btrfs_item));
2387
2388 /* update the item pointers */
2389 right_nritems += push_items;
2390 btrfs_set_header_nritems(right, right_nritems);
2391 push_space = BTRFS_LEAF_DATA_SIZE(root);
2392 for (i = 0; i < right_nritems; i++) {
2393 item = btrfs_item_nr(right, i);
2394 push_space -= btrfs_item_size(right, item);
2395 btrfs_set_item_offset(right, item, push_space);
2396 }
2397
2398 left_nritems -= push_items;
2399 btrfs_set_header_nritems(left, left_nritems);
2400
2401 if (left_nritems)
2402 btrfs_mark_buffer_dirty(left);
2403 else
2404 clean_tree_block(trans, root, left);
2405
2406 btrfs_mark_buffer_dirty(right);
2407
2408 btrfs_item_key(right, &disk_key, 0);
2409 btrfs_set_node_key(upper, &disk_key, slot + 1);
2410 btrfs_mark_buffer_dirty(upper);
2411
2412 /* then fixup the leaf pointer in the path */
2413 if (path->slots[0] >= left_nritems) {
2414 path->slots[0] -= left_nritems;
2415 if (btrfs_header_nritems(path->nodes[0]) == 0)
2416 clean_tree_block(trans, root, path->nodes[0]);
2417 btrfs_tree_unlock(path->nodes[0]);
2418 free_extent_buffer(path->nodes[0]);
2419 path->nodes[0] = right;
2420 path->slots[1] += 1;
2421 } else {
2422 btrfs_tree_unlock(right);
2423 free_extent_buffer(right);
2424 }
2425 return 0;
2426
2427out_unlock:
2428 btrfs_tree_unlock(right);
2429 free_extent_buffer(right);
2430 return 1;
2431}
2432
2433/*
2434 * push some data in the path leaf to the right, trying to free up at
2435 * least data_size bytes. returns zero if the push worked, nonzero otherwise
2436 *
2437 * returns 1 if the push failed because the other node didn't have enough
2438 * room, 0 if everything worked out and < 0 if there were major errors.
2439 *
2440 * this will push starting from min_slot to the end of the leaf. It won't
2441 * push any slot lower than min_slot
2442 */
2443static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
2444 *root, struct btrfs_path *path,
2445 int min_data_size, int data_size,
2446 int empty, u32 min_slot)
2447{
2448 struct extent_buffer *left = path->nodes[0];
2449 struct extent_buffer *right;
2450 struct extent_buffer *upper;
2451 int slot;
2452 int free_space;
2453 u32 left_nritems;
2454 int ret;
2455
2456 if (!path->nodes[1])
2457 return 1;
2458
2459 slot = path->slots[1];
2460 upper = path->nodes[1];
2461 if (slot >= btrfs_header_nritems(upper) - 1)
2462 return 1;
2463
2464 btrfs_assert_tree_locked(path->nodes[1]);
2465
2466 right = read_node_slot(root, upper, slot + 1);
2467 if (right == NULL)
2468 return 1;
2469
2470 btrfs_tree_lock(right);
2471 btrfs_set_lock_blocking(right);
2472
2473 free_space = btrfs_leaf_free_space(root, right);
2474 if (free_space < data_size)
2475 goto out_unlock;
2476
2477 /* cow and double check */
2478 ret = btrfs_cow_block(trans, root, right, upper,
2479 slot + 1, &right);
2480 if (ret)
2481 goto out_unlock;
2482
2483 free_space = btrfs_leaf_free_space(root, right);
2484 if (free_space < data_size)
2485 goto out_unlock;
2486
2487 left_nritems = btrfs_header_nritems(left);
2488 if (left_nritems == 0)
2489 goto out_unlock;
2490
2491 return __push_leaf_right(trans, root, path, min_data_size, empty,
2492 right, free_space, left_nritems, min_slot);
2493out_unlock:
2494 btrfs_tree_unlock(right);
2495 free_extent_buffer(right);
2496 return 1;
2497}
2498
2499/*
2500 * push some data in the path leaf to the left, trying to free up at
2501 * least data_size bytes. returns zero if the push worked, nonzero otherwise
2502 *
2503 * max_slot can put a limit on how far into the leaf we'll push items. The
2504 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
2505 * items
2506 */
2507static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
2508 struct btrfs_root *root,
2509 struct btrfs_path *path, int data_size,
2510 int empty, struct extent_buffer *left,
2511 int free_space, u32 right_nritems,
2512 u32 max_slot)
2513{
2514 struct btrfs_disk_key disk_key;
2515 struct extent_buffer *right = path->nodes[0];
2516 int i;
2517 int push_space = 0;
2518 int push_items = 0;
2519 struct btrfs_item *item;
2520 u32 old_left_nritems;
2521 u32 nr;
2522 int ret = 0;
2523 int wret;
2524 u32 this_item_size;
2525 u32 old_left_item_size;
2526
2527 if (empty)
2528 nr = min(right_nritems, max_slot);
2529 else
2530 nr = min(right_nritems - 1, max_slot);
2531
2532 for (i = 0; i < nr; i++) {
2533 item = btrfs_item_nr(right, i);
2534
2535 if (!empty && push_items > 0) {
2536 if (path->slots[0] < i)
2537 break;
2538 if (path->slots[0] == i) {
2539 int space = btrfs_leaf_free_space(root, right);
2540 if (space + push_space * 2 > free_space)
2541 break;
2542 }
2543 }
2544
2545 if (path->slots[0] == i)
2546 push_space += data_size;
2547
2548 this_item_size = btrfs_item_size(right, item);
2549 if (this_item_size + sizeof(*item) + push_space > free_space)
2550 break;
2551
2552 push_items++;
2553 push_space += this_item_size + sizeof(*item);
2554 }
2555
2556 if (push_items == 0) {
2557 ret = 1;
2558 goto out;
2559 }
2560 if (!empty && push_items == btrfs_header_nritems(right))
2561 WARN_ON(1);
2562
2563 /* push data from right to left */
2564 copy_extent_buffer(left, right,
2565 btrfs_item_nr_offset(btrfs_header_nritems(left)),
2566 btrfs_item_nr_offset(0),
2567 push_items * sizeof(struct btrfs_item));
2568
2569 push_space = BTRFS_LEAF_DATA_SIZE(root) -
2570 btrfs_item_offset_nr(right, push_items - 1);
2571
2572 copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2573 leaf_data_end(root, left) - push_space,
2574 btrfs_leaf_data(right) +
2575 btrfs_item_offset_nr(right, push_items - 1),
2576 push_space);
2577 old_left_nritems = btrfs_header_nritems(left);
2578 BUG_ON(old_left_nritems <= 0);
2579
2580 old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2581 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2582 u32 ioff;
2583
2584 item = btrfs_item_nr(left, i);
2585
2586 ioff = btrfs_item_offset(left, item);
2587 btrfs_set_item_offset(left, item,
2588 ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2589 }
2590 btrfs_set_header_nritems(left, old_left_nritems + push_items);
2591
2592 /* fixup right node */
2593 if (push_items > right_nritems) {
2594 printk(KERN_CRIT "push items %d nr %u\n", push_items,
2595 right_nritems);
2596 WARN_ON(1);
2597 }
2598
2599 if (push_items < right_nritems) {
2600 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2601 leaf_data_end(root, right);
2602 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2603 BTRFS_LEAF_DATA_SIZE(root) - push_space,
2604 btrfs_leaf_data(right) +
2605 leaf_data_end(root, right), push_space);
2606
2607 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2608 btrfs_item_nr_offset(push_items),
2609 (btrfs_header_nritems(right) - push_items) *
2610 sizeof(struct btrfs_item));
2611 }
2612 right_nritems -= push_items;
2613 btrfs_set_header_nritems(right, right_nritems);
2614 push_space = BTRFS_LEAF_DATA_SIZE(root);
2615 for (i = 0; i < right_nritems; i++) {
2616 item = btrfs_item_nr(right, i);
2617
2618 push_space = push_space - btrfs_item_size(right, item);
2619 btrfs_set_item_offset(right, item, push_space);
2620 }
2621
2622 btrfs_mark_buffer_dirty(left);
2623 if (right_nritems)
2624 btrfs_mark_buffer_dirty(right);
2625 else
2626 clean_tree_block(trans, root, right);
2627
2628 btrfs_item_key(right, &disk_key, 0);
2629 wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2630 if (wret)
2631 ret = wret;
2632
2633 /* then fixup the leaf pointer in the path */
2634 if (path->slots[0] < push_items) {
2635 path->slots[0] += old_left_nritems;
2636 btrfs_tree_unlock(path->nodes[0]);
2637 free_extent_buffer(path->nodes[0]);
2638 path->nodes[0] = left;
2639 path->slots[1] -= 1;
2640 } else {
2641 btrfs_tree_unlock(left);
2642 free_extent_buffer(left);
2643 path->slots[0] -= push_items;
2644 }
2645 BUG_ON(path->slots[0] < 0);
2646 return ret;
2647out:
2648 btrfs_tree_unlock(left);
2649 free_extent_buffer(left);
2650 return ret;
2651}
2652
2653/*
2654 * push some data in the path leaf to the left, trying to free up at
2655 * least data_size bytes. returns zero if the push worked, nonzero otherwise
2656 *
2657 * max_slot can put a limit on how far into the leaf we'll push items. The
2658 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
2659 * items
2660 */
2661static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
2662 *root, struct btrfs_path *path, int min_data_size,
2663 int data_size, int empty, u32 max_slot)
2664{
2665 struct extent_buffer *right = path->nodes[0];
2666 struct extent_buffer *left;
2667 int slot;
2668 int free_space;
2669 u32 right_nritems;
2670 int ret = 0;
2671
2672 slot = path->slots[1];
2673 if (slot == 0)
2674 return 1;
2675 if (!path->nodes[1])
2676 return 1;
2677
2678 right_nritems = btrfs_header_nritems(right);
2679 if (right_nritems == 0)
2680 return 1;
2681
2682 btrfs_assert_tree_locked(path->nodes[1]);
2683
2684 left = read_node_slot(root, path->nodes[1], slot - 1);
2685 if (left == NULL)
2686 return 1;
2687
2688 btrfs_tree_lock(left);
2689 btrfs_set_lock_blocking(left);
2690
2691 free_space = btrfs_leaf_free_space(root, left);
2692 if (free_space < data_size) {
2693 ret = 1;
2694 goto out;
2695 }
2696
2697 /* cow and double check */
2698 ret = btrfs_cow_block(trans, root, left,
2699 path->nodes[1], slot - 1, &left);
2700 if (ret) {
2701 /* we hit -ENOSPC, but it isn't fatal here */
2702 ret = 1;
2703 goto out;
2704 }
2705
2706 free_space = btrfs_leaf_free_space(root, left);
2707 if (free_space < data_size) {
2708 ret = 1;
2709 goto out;
2710 }
2711
2712 return __push_leaf_left(trans, root, path, min_data_size,
2713 empty, left, free_space, right_nritems,
2714 max_slot);
2715out:
2716 btrfs_tree_unlock(left);
2717 free_extent_buffer(left);
2718 return ret;
2719}
2720
2721/*
2722 * split the path's leaf in two, making sure there is at least data_size
2723 * available for the resulting leaf level of the path.
2724 *
2725 * returns 0 if all went well and < 0 on failure.
2726 */
2727static noinline int copy_for_split(struct btrfs_trans_handle *trans,
2728 struct btrfs_root *root,
2729 struct btrfs_path *path,
2730 struct extent_buffer *l,
2731 struct extent_buffer *right,
2732 int slot, int mid, int nritems)
2733{
2734 int data_copy_size;
2735 int rt_data_off;
2736 int i;
2737 int ret = 0;
2738 int wret;
2739 struct btrfs_disk_key disk_key;
2740
2741 nritems = nritems - mid;
2742 btrfs_set_header_nritems(right, nritems);
2743 data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2744
2745 copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2746 btrfs_item_nr_offset(mid),
2747 nritems * sizeof(struct btrfs_item));
2748
2749 copy_extent_buffer(right, l,
2750 btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2751 data_copy_size, btrfs_leaf_data(l) +
2752 leaf_data_end(root, l), data_copy_size);
2753
2754 rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2755 btrfs_item_end_nr(l, mid);
2756
2757 for (i = 0; i < nritems; i++) {
2758 struct btrfs_item *item = btrfs_item_nr(right, i);
2759 u32 ioff;
2760
2761 ioff = btrfs_item_offset(right, item);
2762 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2763 }
2764
2765 btrfs_set_header_nritems(l, mid);
2766 ret = 0;
2767 btrfs_item_key(right, &disk_key, 0);
2768 wret = insert_ptr(trans, root, path, &disk_key, right->start,
2769 path->slots[1] + 1, 1);
2770 if (wret)
2771 ret = wret;
2772
2773 btrfs_mark_buffer_dirty(right);
2774 btrfs_mark_buffer_dirty(l);
2775 BUG_ON(path->slots[0] != slot);
2776
2777 if (mid <= slot) {
2778 btrfs_tree_unlock(path->nodes[0]);
2779 free_extent_buffer(path->nodes[0]);
2780 path->nodes[0] = right;
2781 path->slots[0] -= mid;
2782 path->slots[1] += 1;
2783 } else {
2784 btrfs_tree_unlock(right);
2785 free_extent_buffer(right);
2786 }
2787
2788 BUG_ON(path->slots[0] < 0);
2789
2790 return ret;
2791}
2792
2793/*
2794 * double splits happen when we need to insert a big item in the middle
2795 * of a leaf. A double split can leave us with 3 mostly empty leaves:
2796 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
2797 * A B C
2798 *
2799 * We avoid this by trying to push the items on either side of our target
2800 * into the adjacent leaves. If all goes well we can avoid the double split
2801 * completely.
2802 */
2803static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
2804 struct btrfs_root *root,
2805 struct btrfs_path *path,
2806 int data_size)
2807{
2808 int ret;
2809 int progress = 0;
2810 int slot;
2811 u32 nritems;
2812
2813 slot = path->slots[0];
2814
2815 /*
2816 * try to push all the items after our slot into the
2817 * right leaf
2818 */
2819 ret = push_leaf_right(trans, root, path, 1, data_size, 0, slot);
2820 if (ret < 0)
2821 return ret;
2822
2823 if (ret == 0)
2824 progress++;
2825
2826 nritems = btrfs_header_nritems(path->nodes[0]);
2827 /*
2828 * our goal is to get our slot at the start or end of a leaf. If
2829 * we've done so we're done
2830 */
2831 if (path->slots[0] == 0 || path->slots[0] == nritems)
2832 return 0;
2833
2834 if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
2835 return 0;
2836
2837 /* try to push all the items before our slot into the next leaf */
2838 slot = path->slots[0];
2839 ret = push_leaf_left(trans, root, path, 1, data_size, 0, slot);
2840 if (ret < 0)
2841 return ret;
2842
2843 if (ret == 0)
2844 progress++;
2845
2846 if (progress)
2847 return 0;
2848 return 1;
2849}
2850
2851/*
2852 * split the path's leaf in two, making sure there is at least data_size
2853 * available for the resulting leaf level of the path.
2854 *
2855 * returns 0 if all went well and < 0 on failure.
2856 */
2857static noinline int split_leaf(struct btrfs_trans_handle *trans,
2858 struct btrfs_root *root,
2859 struct btrfs_key *ins_key,
2860 struct btrfs_path *path, int data_size,
2861 int extend)
2862{
2863 struct btrfs_disk_key disk_key;
2864 struct extent_buffer *l;
2865 u32 nritems;
2866 int mid;
2867 int slot;
2868 struct extent_buffer *right;
2869 int ret = 0;
2870 int wret;
2871 int split;
2872 int num_doubles = 0;
2873 int tried_avoid_double = 0;
2874
2875 l = path->nodes[0];
2876 slot = path->slots[0];
2877 if (extend && data_size + btrfs_item_size_nr(l, slot) +
2878 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2879 return -EOVERFLOW;
2880
2881 /* first try to make some room by pushing left and right */
2882 if (data_size) {
2883 wret = push_leaf_right(trans, root, path, data_size,
2884 data_size, 0, 0);
2885 if (wret < 0)
2886 return wret;
2887 if (wret) {
2888 wret = push_leaf_left(trans, root, path, data_size,
2889 data_size, 0, (u32)-1);
2890 if (wret < 0)
2891 return wret;
2892 }
2893 l = path->nodes[0];
2894
2895 /* did the pushes work? */
2896 if (btrfs_leaf_free_space(root, l) >= data_size)
2897 return 0;
2898 }
2899
2900 if (!path->nodes[1]) {
2901 ret = insert_new_root(trans, root, path, 1);
2902 if (ret)
2903 return ret;
2904 }
2905again:
2906 split = 1;
2907 l = path->nodes[0];
2908 slot = path->slots[0];
2909 nritems = btrfs_header_nritems(l);
2910 mid = (nritems + 1) / 2;
2911
2912 if (mid <= slot) {
2913 if (nritems == 1 ||
2914 leaf_space_used(l, mid, nritems - mid) + data_size >
2915 BTRFS_LEAF_DATA_SIZE(root)) {
2916 if (slot >= nritems) {
2917 split = 0;
2918 } else {
2919 mid = slot;
2920 if (mid != nritems &&
2921 leaf_space_used(l, mid, nritems - mid) +
2922 data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2923 if (data_size && !tried_avoid_double)
2924 goto push_for_double;
2925 split = 2;
2926 }
2927 }
2928 }
2929 } else {
2930 if (leaf_space_used(l, 0, mid) + data_size >
2931 BTRFS_LEAF_DATA_SIZE(root)) {
2932 if (!extend && data_size && slot == 0) {
2933 split = 0;
2934 } else if ((extend || !data_size) && slot == 0) {
2935 mid = 1;
2936 } else {
2937 mid = slot;
2938 if (mid != nritems &&
2939 leaf_space_used(l, mid, nritems - mid) +
2940 data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2941 if (data_size && !tried_avoid_double)
2942 goto push_for_double;
2943 split = 2 ;
2944 }
2945 }
2946 }
2947 }
2948
2949 if (split == 0)
2950 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2951 else
2952 btrfs_item_key(l, &disk_key, mid);
2953
2954 right = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
2955 root->root_key.objectid,
2956 &disk_key, 0, l->start, 0);
2957 if (IS_ERR(right))
2958 return PTR_ERR(right);
2959
2960 root_add_used(root, root->leafsize);
2961
2962 memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2963 btrfs_set_header_bytenr(right, right->start);
2964 btrfs_set_header_generation(right, trans->transid);
2965 btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2966 btrfs_set_header_owner(right, root->root_key.objectid);
2967 btrfs_set_header_level(right, 0);
2968 write_extent_buffer(right, root->fs_info->fsid,
2969 (unsigned long)btrfs_header_fsid(right),
2970 BTRFS_FSID_SIZE);
2971
2972 write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2973 (unsigned long)btrfs_header_chunk_tree_uuid(right),
2974 BTRFS_UUID_SIZE);
2975
2976 if (split == 0) {
2977 if (mid <= slot) {
2978 btrfs_set_header_nritems(right, 0);
2979 wret = insert_ptr(trans, root, path,
2980 &disk_key, right->start,
2981 path->slots[1] + 1, 1);
2982 if (wret)
2983 ret = wret;
2984
2985 btrfs_tree_unlock(path->nodes[0]);
2986 free_extent_buffer(path->nodes[0]);
2987 path->nodes[0] = right;
2988 path->slots[0] = 0;
2989 path->slots[1] += 1;
2990 } else {
2991 btrfs_set_header_nritems(right, 0);
2992 wret = insert_ptr(trans, root, path,
2993 &disk_key,
2994 right->start,
2995 path->slots[1], 1);
2996 if (wret)
2997 ret = wret;
2998 btrfs_tree_unlock(path->nodes[0]);
2999 free_extent_buffer(path->nodes[0]);
3000 path->nodes[0] = right;
3001 path->slots[0] = 0;
3002 if (path->slots[1] == 0) {
3003 wret = fixup_low_keys(trans, root,
3004 path, &disk_key, 1);
3005 if (wret)
3006 ret = wret;
3007 }
3008 }
3009 btrfs_mark_buffer_dirty(right);
3010 return ret;
3011 }
3012
3013 ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
3014 BUG_ON(ret);
3015
3016 if (split == 2) {
3017 BUG_ON(num_doubles != 0);
3018 num_doubles++;
3019 goto again;
3020 }
3021
3022 return ret;
3023
3024push_for_double:
3025 push_for_double_split(trans, root, path, data_size);
3026 tried_avoid_double = 1;
3027 if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
3028 return 0;
3029 goto again;
3030}
3031
3032static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3033 struct btrfs_root *root,
3034 struct btrfs_path *path, int ins_len)
3035{
3036 struct btrfs_key key;
3037 struct extent_buffer *leaf;
3038 struct btrfs_file_extent_item *fi;
3039 u64 extent_len = 0;
3040 u32 item_size;
3041 int ret;
3042
3043 leaf = path->nodes[0];
3044 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3045
3046 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3047 key.type != BTRFS_EXTENT_CSUM_KEY);
3048
3049 if (btrfs_leaf_free_space(root, leaf) >= ins_len)
3050 return 0;
3051
3052 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3053 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3054 fi = btrfs_item_ptr(leaf, path->slots[0],
3055 struct btrfs_file_extent_item);
3056 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3057 }
3058 btrfs_release_path(path);
3059
3060 path->keep_locks = 1;
3061 path->search_for_split = 1;
3062 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3063 path->search_for_split = 0;
3064 if (ret < 0)
3065 goto err;
3066
3067 ret = -EAGAIN;
3068 leaf = path->nodes[0];
3069 /* if our item isn't there or got smaller, return now */
3070 if (ret > 0 || item_size != btrfs_item_size_nr(leaf, path->slots[0]))
3071 goto err;
3072
3073 /* the leaf has changed, it now has room. return now */
3074 if (btrfs_leaf_free_space(root, path->nodes[0]) >= ins_len)
3075 goto err;
3076
3077 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3078 fi = btrfs_item_ptr(leaf, path->slots[0],
3079 struct btrfs_file_extent_item);
3080 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3081 goto err;
3082 }
3083
3084 btrfs_set_path_blocking(path);
3085 ret = split_leaf(trans, root, &key, path, ins_len, 1);
3086 if (ret)
3087 goto err;
3088
3089 path->keep_locks = 0;
3090 btrfs_unlock_up_safe(path, 1);
3091 return 0;
3092err:
3093 path->keep_locks = 0;
3094 return ret;
3095}
3096
3097static noinline int split_item(struct btrfs_trans_handle *trans,
3098 struct btrfs_root *root,
3099 struct btrfs_path *path,
3100 struct btrfs_key *new_key,
3101 unsigned long split_offset)
3102{
3103 struct extent_buffer *leaf;
3104 struct btrfs_item *item;
3105 struct btrfs_item *new_item;
3106 int slot;
3107 char *buf;
3108 u32 nritems;
3109 u32 item_size;
3110 u32 orig_offset;
3111 struct btrfs_disk_key disk_key;
3112
3113 leaf = path->nodes[0];
3114 BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
3115
3116 btrfs_set_path_blocking(path);
3117
3118 item = btrfs_item_nr(leaf, path->slots[0]);
3119 orig_offset = btrfs_item_offset(leaf, item);
3120 item_size = btrfs_item_size(leaf, item);
3121
3122 buf = kmalloc(item_size, GFP_NOFS);
3123 if (!buf)
3124 return -ENOMEM;
3125
3126 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3127 path->slots[0]), item_size);
3128
3129 slot = path->slots[0] + 1;
3130 nritems = btrfs_header_nritems(leaf);
3131 if (slot != nritems) {
3132 /* shift the items */
3133 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3134 btrfs_item_nr_offset(slot),
3135 (nritems - slot) * sizeof(struct btrfs_item));
3136 }
3137
3138 btrfs_cpu_key_to_disk(&disk_key, new_key);
3139 btrfs_set_item_key(leaf, &disk_key, slot);
3140
3141 new_item = btrfs_item_nr(leaf, slot);
3142
3143 btrfs_set_item_offset(leaf, new_item, orig_offset);
3144 btrfs_set_item_size(leaf, new_item, item_size - split_offset);
3145
3146 btrfs_set_item_offset(leaf, item,
3147 orig_offset + item_size - split_offset);
3148 btrfs_set_item_size(leaf, item, split_offset);
3149
3150 btrfs_set_header_nritems(leaf, nritems + 1);
3151
3152 /* write the data for the start of the original item */
3153 write_extent_buffer(leaf, buf,
3154 btrfs_item_ptr_offset(leaf, path->slots[0]),
3155 split_offset);
3156
3157 /* write the data for the new item */
3158 write_extent_buffer(leaf, buf + split_offset,
3159 btrfs_item_ptr_offset(leaf, slot),
3160 item_size - split_offset);
3161 btrfs_mark_buffer_dirty(leaf);
3162
3163 BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
3164 kfree(buf);
3165 return 0;
3166}
3167
3168/*
3169 * This function splits a single item into two items,
3170 * giving 'new_key' to the new item and splitting the
3171 * old one at split_offset (from the start of the item).
3172 *
3173 * The path may be released by this operation. After
3174 * the split, the path is pointing to the old item. The
3175 * new item is going to be in the same node as the old one.
3176 *
3177 * Note, the item being split must be smaller enough to live alone on
3178 * a tree block with room for one extra struct btrfs_item
3179 *
3180 * This allows us to split the item in place, keeping a lock on the
3181 * leaf the entire time.
3182 */
3183int btrfs_split_item(struct btrfs_trans_handle *trans,
3184 struct btrfs_root *root,
3185 struct btrfs_path *path,
3186 struct btrfs_key *new_key,
3187 unsigned long split_offset)
3188{
3189 int ret;
3190 ret = setup_leaf_for_split(trans, root, path,
3191 sizeof(struct btrfs_item));
3192 if (ret)
3193 return ret;
3194
3195 ret = split_item(trans, root, path, new_key, split_offset);
3196 return ret;
3197}
3198
3199/*
3200 * This function duplicate a item, giving 'new_key' to the new item.
3201 * It guarantees both items live in the same tree leaf and the new item
3202 * is contiguous with the original item.
3203 *
3204 * This allows us to split file extent in place, keeping a lock on the
3205 * leaf the entire time.
3206 */
3207int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
3208 struct btrfs_root *root,
3209 struct btrfs_path *path,
3210 struct btrfs_key *new_key)
3211{
3212 struct extent_buffer *leaf;
3213 int ret;
3214 u32 item_size;
3215
3216 leaf = path->nodes[0];
3217 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3218 ret = setup_leaf_for_split(trans, root, path,
3219 item_size + sizeof(struct btrfs_item));
3220 if (ret)
3221 return ret;
3222
3223 path->slots[0]++;
3224 ret = setup_items_for_insert(trans, root, path, new_key, &item_size,
3225 item_size, item_size +
3226 sizeof(struct btrfs_item), 1);
3227 BUG_ON(ret);
3228
3229 leaf = path->nodes[0];
3230 memcpy_extent_buffer(leaf,
3231 btrfs_item_ptr_offset(leaf, path->slots[0]),
3232 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
3233 item_size);
3234 return 0;
3235}
3236
3237/*
3238 * make the item pointed to by the path smaller. new_size indicates
3239 * how small to make it, and from_end tells us if we just chop bytes
3240 * off the end of the item or if we shift the item to chop bytes off
3241 * the front.
3242 */
3243int btrfs_truncate_item(struct btrfs_trans_handle *trans,
3244 struct btrfs_root *root,
3245 struct btrfs_path *path,
3246 u32 new_size, int from_end)
3247{
3248 int slot;
3249 struct extent_buffer *leaf;
3250 struct btrfs_item *item;
3251 u32 nritems;
3252 unsigned int data_end;
3253 unsigned int old_data_start;
3254 unsigned int old_size;
3255 unsigned int size_diff;
3256 int i;
3257
3258 leaf = path->nodes[0];
3259 slot = path->slots[0];
3260
3261 old_size = btrfs_item_size_nr(leaf, slot);
3262 if (old_size == new_size)
3263 return 0;
3264
3265 nritems = btrfs_header_nritems(leaf);
3266 data_end = leaf_data_end(root, leaf);
3267
3268 old_data_start = btrfs_item_offset_nr(leaf, slot);
3269
3270 size_diff = old_size - new_size;
3271
3272 BUG_ON(slot < 0);
3273 BUG_ON(slot >= nritems);
3274
3275 /*
3276 * item0..itemN ... dataN.offset..dataN.size .. data0.size
3277 */
3278 /* first correct the data pointers */
3279 for (i = slot; i < nritems; i++) {
3280 u32 ioff;
3281 item = btrfs_item_nr(leaf, i);
3282
3283 ioff = btrfs_item_offset(leaf, item);
3284 btrfs_set_item_offset(leaf, item, ioff + size_diff);
3285 }
3286
3287 /* shift the data */
3288 if (from_end) {
3289 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3290 data_end + size_diff, btrfs_leaf_data(leaf) +
3291 data_end, old_data_start + new_size - data_end);
3292 } else {
3293 struct btrfs_disk_key disk_key;
3294 u64 offset;
3295
3296 btrfs_item_key(leaf, &disk_key, slot);
3297
3298 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3299 unsigned long ptr;
3300 struct btrfs_file_extent_item *fi;
3301
3302 fi = btrfs_item_ptr(leaf, slot,
3303 struct btrfs_file_extent_item);
3304 fi = (struct btrfs_file_extent_item *)(
3305 (unsigned long)fi - size_diff);
3306
3307 if (btrfs_file_extent_type(leaf, fi) ==
3308 BTRFS_FILE_EXTENT_INLINE) {
3309 ptr = btrfs_item_ptr_offset(leaf, slot);
3310 memmove_extent_buffer(leaf, ptr,
3311 (unsigned long)fi,
3312 offsetof(struct btrfs_file_extent_item,
3313 disk_bytenr));
3314 }
3315 }
3316
3317 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3318 data_end + size_diff, btrfs_leaf_data(leaf) +
3319 data_end, old_data_start - data_end);
3320
3321 offset = btrfs_disk_key_offset(&disk_key);
3322 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3323 btrfs_set_item_key(leaf, &disk_key, slot);
3324 if (slot == 0)
3325 fixup_low_keys(trans, root, path, &disk_key, 1);
3326 }
3327
3328 item = btrfs_item_nr(leaf, slot);
3329 btrfs_set_item_size(leaf, item, new_size);
3330 btrfs_mark_buffer_dirty(leaf);
3331
3332 if (btrfs_leaf_free_space(root, leaf) < 0) {
3333 btrfs_print_leaf(root, leaf);
3334 BUG();
3335 }
3336 return 0;
3337}
3338
3339/*
3340 * make the item pointed to by the path bigger, data_size is the new size.
3341 */
3342int btrfs_extend_item(struct btrfs_trans_handle *trans,
3343 struct btrfs_root *root, struct btrfs_path *path,
3344 u32 data_size)
3345{
3346 int slot;
3347 struct extent_buffer *leaf;
3348 struct btrfs_item *item;
3349 u32 nritems;
3350 unsigned int data_end;
3351 unsigned int old_data;
3352 unsigned int old_size;
3353 int i;
3354
3355 leaf = path->nodes[0];
3356
3357 nritems = btrfs_header_nritems(leaf);
3358 data_end = leaf_data_end(root, leaf);
3359
3360 if (btrfs_leaf_free_space(root, leaf) < data_size) {
3361 btrfs_print_leaf(root, leaf);
3362 BUG();
3363 }
3364 slot = path->slots[0];
3365 old_data = btrfs_item_end_nr(leaf, slot);
3366
3367 BUG_ON(slot < 0);
3368 if (slot >= nritems) {
3369 btrfs_print_leaf(root, leaf);
3370 printk(KERN_CRIT "slot %d too large, nritems %d\n",
3371 slot, nritems);
3372 BUG_ON(1);
3373 }
3374
3375 /*
3376 * item0..itemN ... dataN.offset..dataN.size .. data0.size
3377 */
3378 /* first correct the data pointers */
3379 for (i = slot; i < nritems; i++) {
3380 u32 ioff;
3381 item = btrfs_item_nr(leaf, i);
3382
3383 ioff = btrfs_item_offset(leaf, item);
3384 btrfs_set_item_offset(leaf, item, ioff - data_size);
3385 }
3386
3387 /* shift the data */
3388 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3389 data_end - data_size, btrfs_leaf_data(leaf) +
3390 data_end, old_data - data_end);
3391
3392 data_end = old_data;
3393 old_size = btrfs_item_size_nr(leaf, slot);
3394 item = btrfs_item_nr(leaf, slot);
3395 btrfs_set_item_size(leaf, item, old_size + data_size);
3396 btrfs_mark_buffer_dirty(leaf);
3397
3398 if (btrfs_leaf_free_space(root, leaf) < 0) {
3399 btrfs_print_leaf(root, leaf);
3400 BUG();
3401 }
3402 return 0;
3403}
3404
3405/*
3406 * Given a key and some data, insert items into the tree.
3407 * This does all the path init required, making room in the tree if needed.
3408 * Returns the number of keys that were inserted.
3409 */
3410int btrfs_insert_some_items(struct btrfs_trans_handle *trans,
3411 struct btrfs_root *root,
3412 struct btrfs_path *path,
3413 struct btrfs_key *cpu_key, u32 *data_size,
3414 int nr)
3415{
3416 struct extent_buffer *leaf;
3417 struct btrfs_item *item;
3418 int ret = 0;
3419 int slot;
3420 int i;
3421 u32 nritems;
3422 u32 total_data = 0;
3423 u32 total_size = 0;
3424 unsigned int data_end;
3425 struct btrfs_disk_key disk_key;
3426 struct btrfs_key found_key;
3427
3428 for (i = 0; i < nr; i++) {
3429 if (total_size + data_size[i] + sizeof(struct btrfs_item) >
3430 BTRFS_LEAF_DATA_SIZE(root)) {
3431 break;
3432 nr = i;
3433 }
3434 total_data += data_size[i];
3435 total_size += data_size[i] + sizeof(struct btrfs_item);
3436 }
3437 BUG_ON(nr == 0);
3438
3439 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3440 if (ret == 0)
3441 return -EEXIST;
3442 if (ret < 0)
3443 goto out;
3444
3445 leaf = path->nodes[0];
3446
3447 nritems = btrfs_header_nritems(leaf);
3448 data_end = leaf_data_end(root, leaf);
3449
3450 if (btrfs_leaf_free_space(root, leaf) < total_size) {
3451 for (i = nr; i >= 0; i--) {
3452 total_data -= data_size[i];
3453 total_size -= data_size[i] + sizeof(struct btrfs_item);
3454 if (total_size < btrfs_leaf_free_space(root, leaf))
3455 break;
3456 }
3457 nr = i;
3458 }
3459
3460 slot = path->slots[0];
3461 BUG_ON(slot < 0);
3462
3463 if (slot != nritems) {
3464 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3465
3466 item = btrfs_item_nr(leaf, slot);
3467 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3468
3469 /* figure out how many keys we can insert in here */
3470 total_data = data_size[0];
3471 for (i = 1; i < nr; i++) {
3472 if (btrfs_comp_cpu_keys(&found_key, cpu_key + i) <= 0)
3473 break;
3474 total_data += data_size[i];
3475 }
3476 nr = i;
3477
3478 if (old_data < data_end) {
3479 btrfs_print_leaf(root, leaf);
3480 printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3481 slot, old_data, data_end);
3482 BUG_ON(1);
3483 }
3484 /*
3485 * item0..itemN ... dataN.offset..dataN.size .. data0.size
3486 */
3487 /* first correct the data pointers */
3488 for (i = slot; i < nritems; i++) {
3489 u32 ioff;
3490
3491 item = btrfs_item_nr(leaf, i);
3492 ioff = btrfs_item_offset(leaf, item);
3493 btrfs_set_item_offset(leaf, item, ioff - total_data);
3494 }
3495 /* shift the items */
3496 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3497 btrfs_item_nr_offset(slot),
3498 (nritems - slot) * sizeof(struct btrfs_item));
3499
3500 /* shift the data */
3501 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3502 data_end - total_data, btrfs_leaf_data(leaf) +
3503 data_end, old_data - data_end);
3504 data_end = old_data;
3505 } else {
3506 /*
3507 * this sucks but it has to be done, if we are inserting at
3508 * the end of the leaf only insert 1 of the items, since we
3509 * have no way of knowing whats on the next leaf and we'd have
3510 * to drop our current locks to figure it out
3511 */
3512 nr = 1;
3513 }
3514
3515 /* setup the item for the new data */
3516 for (i = 0; i < nr; i++) {
3517 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3518 btrfs_set_item_key(leaf, &disk_key, slot + i);
3519 item = btrfs_item_nr(leaf, slot + i);
3520 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3521 data_end -= data_size[i];
3522 btrfs_set_item_size(leaf, item, data_size[i]);
3523 }
3524 btrfs_set_header_nritems(leaf, nritems + nr);
3525 btrfs_mark_buffer_dirty(leaf);
3526
3527 ret = 0;
3528 if (slot == 0) {
3529 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3530 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3531 }
3532
3533 if (btrfs_leaf_free_space(root, leaf) < 0) {
3534 btrfs_print_leaf(root, leaf);
3535 BUG();
3536 }
3537out:
3538 if (!ret)
3539 ret = nr;
3540 return ret;
3541}
3542
3543/*
3544 * this is a helper for btrfs_insert_empty_items, the main goal here is
3545 * to save stack depth by doing the bulk of the work in a function
3546 * that doesn't call btrfs_search_slot
3547 */
3548int setup_items_for_insert(struct btrfs_trans_handle *trans,
3549 struct btrfs_root *root, struct btrfs_path *path,
3550 struct btrfs_key *cpu_key, u32 *data_size,
3551 u32 total_data, u32 total_size, int nr)
3552{
3553 struct btrfs_item *item;
3554 int i;
3555 u32 nritems;
3556 unsigned int data_end;
3557 struct btrfs_disk_key disk_key;
3558 int ret;
3559 struct extent_buffer *leaf;
3560 int slot;
3561
3562 leaf = path->nodes[0];
3563 slot = path->slots[0];
3564
3565 nritems = btrfs_header_nritems(leaf);
3566 data_end = leaf_data_end(root, leaf);
3567
3568 if (btrfs_leaf_free_space(root, leaf) < total_size) {
3569 btrfs_print_leaf(root, leaf);
3570 printk(KERN_CRIT "not enough freespace need %u have %d\n",
3571 total_size, btrfs_leaf_free_space(root, leaf));
3572 BUG();
3573 }
3574
3575 if (slot != nritems) {
3576 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3577
3578 if (old_data < data_end) {
3579 btrfs_print_leaf(root, leaf);
3580 printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3581 slot, old_data, data_end);
3582 BUG_ON(1);
3583 }
3584 /*
3585 * item0..itemN ... dataN.offset..dataN.size .. data0.size
3586 */
3587 /* first correct the data pointers */
3588 for (i = slot; i < nritems; i++) {
3589 u32 ioff;
3590
3591 item = btrfs_item_nr(leaf, i);
3592 ioff = btrfs_item_offset(leaf, item);
3593 btrfs_set_item_offset(leaf, item, ioff - total_data);
3594 }
3595 /* shift the items */
3596 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3597 btrfs_item_nr_offset(slot),
3598 (nritems - slot) * sizeof(struct btrfs_item));
3599
3600 /* shift the data */
3601 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3602 data_end - total_data, btrfs_leaf_data(leaf) +
3603 data_end, old_data - data_end);
3604 data_end = old_data;
3605 }
3606
3607 /* setup the item for the new data */
3608 for (i = 0; i < nr; i++) {
3609 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3610 btrfs_set_item_key(leaf, &disk_key, slot + i);
3611 item = btrfs_item_nr(leaf, slot + i);
3612 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3613 data_end -= data_size[i];
3614 btrfs_set_item_size(leaf, item, data_size[i]);
3615 }
3616
3617 btrfs_set_header_nritems(leaf, nritems + nr);
3618
3619 ret = 0;
3620 if (slot == 0) {
3621 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3622 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3623 }
3624 btrfs_unlock_up_safe(path, 1);
3625 btrfs_mark_buffer_dirty(leaf);
3626
3627 if (btrfs_leaf_free_space(root, leaf) < 0) {
3628 btrfs_print_leaf(root, leaf);
3629 BUG();
3630 }
3631 return ret;
3632}
3633
3634/*
3635 * Given a key and some data, insert items into the tree.
3636 * This does all the path init required, making room in the tree if needed.
3637 */
3638int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
3639 struct btrfs_root *root,
3640 struct btrfs_path *path,
3641 struct btrfs_key *cpu_key, u32 *data_size,
3642 int nr)
3643{
3644 int ret = 0;
3645 int slot;
3646 int i;
3647 u32 total_size = 0;
3648 u32 total_data = 0;
3649
3650 for (i = 0; i < nr; i++)
3651 total_data += data_size[i];
3652
3653 total_size = total_data + (nr * sizeof(struct btrfs_item));
3654 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3655 if (ret == 0)
3656 return -EEXIST;
3657 if (ret < 0)
3658 goto out;
3659
3660 slot = path->slots[0];
3661 BUG_ON(slot < 0);
3662
3663 ret = setup_items_for_insert(trans, root, path, cpu_key, data_size,
3664 total_data, total_size, nr);
3665
3666out:
3667 return ret;
3668}
3669
3670/*
3671 * Given a key and some data, insert an item into the tree.
3672 * This does all the path init required, making room in the tree if needed.
3673 */
3674int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
3675 *root, struct btrfs_key *cpu_key, void *data, u32
3676 data_size)
3677{
3678 int ret = 0;
3679 struct btrfs_path *path;
3680 struct extent_buffer *leaf;
3681 unsigned long ptr;
3682
3683 path = btrfs_alloc_path();
3684 if (!path)
3685 return -ENOMEM;
3686 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
3687 if (!ret) {
3688 leaf = path->nodes[0];
3689 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3690 write_extent_buffer(leaf, data, ptr, data_size);
3691 btrfs_mark_buffer_dirty(leaf);
3692 }
3693 btrfs_free_path(path);
3694 return ret;
3695}
3696
3697/*
3698 * delete the pointer from a given node.
3699 *
3700 * the tree should have been previously balanced so the deletion does not
3701 * empty a node.
3702 */
3703static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3704 struct btrfs_path *path, int level, int slot)
3705{
3706 struct extent_buffer *parent = path->nodes[level];
3707 u32 nritems;
3708 int ret = 0;
3709 int wret;
3710
3711 nritems = btrfs_header_nritems(parent);
3712 if (slot != nritems - 1) {
3713 memmove_extent_buffer(parent,
3714 btrfs_node_key_ptr_offset(slot),
3715 btrfs_node_key_ptr_offset(slot + 1),
3716 sizeof(struct btrfs_key_ptr) *
3717 (nritems - slot - 1));
3718 }
3719 nritems--;
3720 btrfs_set_header_nritems(parent, nritems);
3721 if (nritems == 0 && parent == root->node) {
3722 BUG_ON(btrfs_header_level(root->node) != 1);
3723 /* just turn the root into a leaf and break */
3724 btrfs_set_header_level(root->node, 0);
3725 } else if (slot == 0) {
3726 struct btrfs_disk_key disk_key;
3727
3728 btrfs_node_key(parent, &disk_key, 0);
3729 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
3730 if (wret)
3731 ret = wret;
3732 }
3733 btrfs_mark_buffer_dirty(parent);
3734 return ret;
3735}
3736
3737/*
3738 * a helper function to delete the leaf pointed to by path->slots[1] and
3739 * path->nodes[1].
3740 *
3741 * This deletes the pointer in path->nodes[1] and frees the leaf
3742 * block extent. zero is returned if it all worked out, < 0 otherwise.
3743 *
3744 * The path must have already been setup for deleting the leaf, including
3745 * all the proper balancing. path->nodes[1] must be locked.
3746 */
3747static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
3748 struct btrfs_root *root,
3749 struct btrfs_path *path,
3750 struct extent_buffer *leaf)
3751{
3752 int ret;
3753
3754 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
3755 ret = del_ptr(trans, root, path, 1, path->slots[1]);
3756 if (ret)
3757 return ret;
3758
3759 /*
3760 * btrfs_free_extent is expensive, we want to make sure we
3761 * aren't holding any locks when we call it
3762 */
3763 btrfs_unlock_up_safe(path, 0);
3764
3765 root_sub_used(root, leaf->len);
3766
3767 btrfs_free_tree_block(trans, root, leaf, 0, 1);
3768 return 0;
3769}
3770/*
3771 * delete the item at the leaf level in path. If that empties
3772 * the leaf, remove it from the tree
3773 */
3774int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3775 struct btrfs_path *path, int slot, int nr)
3776{
3777 struct extent_buffer *leaf;
3778 struct btrfs_item *item;
3779 int last_off;
3780 int dsize = 0;
3781 int ret = 0;
3782 int wret;
3783 int i;
3784 u32 nritems;
3785
3786 leaf = path->nodes[0];
3787 last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
3788
3789 for (i = 0; i < nr; i++)
3790 dsize += btrfs_item_size_nr(leaf, slot + i);
3791
3792 nritems = btrfs_header_nritems(leaf);
3793
3794 if (slot + nr != nritems) {
3795 int data_end = leaf_data_end(root, leaf);
3796
3797 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3798 data_end + dsize,
3799 btrfs_leaf_data(leaf) + data_end,
3800 last_off - data_end);
3801
3802 for (i = slot + nr; i < nritems; i++) {
3803 u32 ioff;
3804
3805 item = btrfs_item_nr(leaf, i);
3806 ioff = btrfs_item_offset(leaf, item);
3807 btrfs_set_item_offset(leaf, item, ioff + dsize);
3808 }
3809
3810 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
3811 btrfs_item_nr_offset(slot + nr),
3812 sizeof(struct btrfs_item) *
3813 (nritems - slot - nr));
3814 }
3815 btrfs_set_header_nritems(leaf, nritems - nr);
3816 nritems -= nr;
3817
3818 /* delete the leaf if we've emptied it */
3819 if (nritems == 0) {
3820 if (leaf == root->node) {
3821 btrfs_set_header_level(leaf, 0);
3822 } else {
3823 btrfs_set_path_blocking(path);
3824 clean_tree_block(trans, root, leaf);
3825 ret = btrfs_del_leaf(trans, root, path, leaf);
3826 BUG_ON(ret);
3827 }
3828 } else {
3829 int used = leaf_space_used(leaf, 0, nritems);
3830 if (slot == 0) {
3831 struct btrfs_disk_key disk_key;
3832
3833 btrfs_item_key(leaf, &disk_key, 0);
3834 wret = fixup_low_keys(trans, root, path,
3835 &disk_key, 1);
3836 if (wret)
3837 ret = wret;
3838 }
3839
3840 /* delete the leaf if it is mostly empty */
3841 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
3842 /* push_leaf_left fixes the path.
3843 * make sure the path still points to our leaf
3844 * for possible call to del_ptr below
3845 */
3846 slot = path->slots[1];
3847 extent_buffer_get(leaf);
3848
3849 btrfs_set_path_blocking(path);
3850 wret = push_leaf_left(trans, root, path, 1, 1,
3851 1, (u32)-1);
3852 if (wret < 0 && wret != -ENOSPC)
3853 ret = wret;
3854
3855 if (path->nodes[0] == leaf &&
3856 btrfs_header_nritems(leaf)) {
3857 wret = push_leaf_right(trans, root, path, 1,
3858 1, 1, 0);
3859 if (wret < 0 && wret != -ENOSPC)
3860 ret = wret;
3861 }
3862
3863 if (btrfs_header_nritems(leaf) == 0) {
3864 path->slots[1] = slot;
3865 ret = btrfs_del_leaf(trans, root, path, leaf);
3866 BUG_ON(ret);
3867 free_extent_buffer(leaf);
3868 } else {
3869 /* if we're still in the path, make sure
3870 * we're dirty. Otherwise, one of the
3871 * push_leaf functions must have already
3872 * dirtied this buffer
3873 */
3874 if (path->nodes[0] == leaf)
3875 btrfs_mark_buffer_dirty(leaf);
3876 free_extent_buffer(leaf);
3877 }
3878 } else {
3879 btrfs_mark_buffer_dirty(leaf);
3880 }
3881 }
3882 return ret;
3883}
3884
3885/*
3886 * search the tree again to find a leaf with lesser keys
3887 * returns 0 if it found something or 1 if there are no lesser leaves.
3888 * returns < 0 on io errors.
3889 *
3890 * This may release the path, and so you may lose any locks held at the
3891 * time you call it.
3892 */
3893int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
3894{
3895 struct btrfs_key key;
3896 struct btrfs_disk_key found_key;
3897 int ret;
3898
3899 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
3900
3901 if (key.offset > 0)
3902 key.offset--;
3903 else if (key.type > 0)
3904 key.type--;
3905 else if (key.objectid > 0)
3906 key.objectid--;
3907 else
3908 return 1;
3909
3910 btrfs_release_path(path);
3911 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3912 if (ret < 0)
3913 return ret;
3914 btrfs_item_key(path->nodes[0], &found_key, 0);
3915 ret = comp_keys(&found_key, &key);
3916 if (ret < 0)
3917 return 0;
3918 return 1;
3919}
3920
3921/*
3922 * A helper function to walk down the tree starting at min_key, and looking
3923 * for nodes or leaves that are either in cache or have a minimum
3924 * transaction id. This is used by the btree defrag code, and tree logging
3925 *
3926 * This does not cow, but it does stuff the starting key it finds back
3927 * into min_key, so you can call btrfs_search_slot with cow=1 on the
3928 * key and get a writable path.
3929 *
3930 * This does lock as it descends, and path->keep_locks should be set
3931 * to 1 by the caller.
3932 *
3933 * This honors path->lowest_level to prevent descent past a given level
3934 * of the tree.
3935 *
3936 * min_trans indicates the oldest transaction that you are interested
3937 * in walking through. Any nodes or leaves older than min_trans are
3938 * skipped over (without reading them).
3939 *
3940 * returns zero if something useful was found, < 0 on error and 1 if there
3941 * was nothing in the tree that matched the search criteria.
3942 */
3943int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
3944 struct btrfs_key *max_key,
3945 struct btrfs_path *path, int cache_only,
3946 u64 min_trans)
3947{
3948 struct extent_buffer *cur;
3949 struct btrfs_key found_key;
3950 int slot;
3951 int sret;
3952 u32 nritems;
3953 int level;
3954 int ret = 1;
3955
3956 WARN_ON(!path->keep_locks);
3957again:
3958 cur = btrfs_read_lock_root_node(root);
3959 level = btrfs_header_level(cur);
3960 WARN_ON(path->nodes[level]);
3961 path->nodes[level] = cur;
3962 path->locks[level] = BTRFS_READ_LOCK;
3963
3964 if (btrfs_header_generation(cur) < min_trans) {
3965 ret = 1;
3966 goto out;
3967 }
3968 while (1) {
3969 nritems = btrfs_header_nritems(cur);
3970 level = btrfs_header_level(cur);
3971 sret = bin_search(cur, min_key, level, &slot);
3972
3973 /* at the lowest level, we're done, setup the path and exit */
3974 if (level == path->lowest_level) {
3975 if (slot >= nritems)
3976 goto find_next_key;
3977 ret = 0;
3978 path->slots[level] = slot;
3979 btrfs_item_key_to_cpu(cur, &found_key, slot);
3980 goto out;
3981 }
3982 if (sret && slot > 0)
3983 slot--;
3984 /*
3985 * check this node pointer against the cache_only and
3986 * min_trans parameters. If it isn't in cache or is too
3987 * old, skip to the next one.
3988 */
3989 while (slot < nritems) {
3990 u64 blockptr;
3991 u64 gen;
3992 struct extent_buffer *tmp;
3993 struct btrfs_disk_key disk_key;
3994
3995 blockptr = btrfs_node_blockptr(cur, slot);
3996 gen = btrfs_node_ptr_generation(cur, slot);
3997 if (gen < min_trans) {
3998 slot++;
3999 continue;
4000 }
4001 if (!cache_only)
4002 break;
4003
4004 if (max_key) {
4005 btrfs_node_key(cur, &disk_key, slot);
4006 if (comp_keys(&disk_key, max_key) >= 0) {
4007 ret = 1;
4008 goto out;
4009 }
4010 }
4011
4012 tmp = btrfs_find_tree_block(root, blockptr,
4013 btrfs_level_size(root, level - 1));
4014
4015 if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
4016 free_extent_buffer(tmp);
4017 break;
4018 }
4019 if (tmp)
4020 free_extent_buffer(tmp);
4021 slot++;
4022 }
4023find_next_key:
4024 /*
4025 * we didn't find a candidate key in this node, walk forward
4026 * and find another one
4027 */
4028 if (slot >= nritems) {
4029 path->slots[level] = slot;
4030 btrfs_set_path_blocking(path);
4031 sret = btrfs_find_next_key(root, path, min_key, level,
4032 cache_only, min_trans);
4033 if (sret == 0) {
4034 btrfs_release_path(path);
4035 goto again;
4036 } else {
4037 goto out;
4038 }
4039 }
4040 /* save our key for returning back */
4041 btrfs_node_key_to_cpu(cur, &found_key, slot);
4042 path->slots[level] = slot;
4043 if (level == path->lowest_level) {
4044 ret = 0;
4045 unlock_up(path, level, 1);
4046 goto out;
4047 }
4048 btrfs_set_path_blocking(path);
4049 cur = read_node_slot(root, cur, slot);
4050 BUG_ON(!cur);
4051
4052 btrfs_tree_read_lock(cur);
4053
4054 path->locks[level - 1] = BTRFS_READ_LOCK;
4055 path->nodes[level - 1] = cur;
4056 unlock_up(path, level, 1);
4057 btrfs_clear_path_blocking(path, NULL, 0);
4058 }
4059out:
4060 if (ret == 0)
4061 memcpy(min_key, &found_key, sizeof(found_key));
4062 btrfs_set_path_blocking(path);
4063 return ret;
4064}
4065
4066/*
4067 * this is similar to btrfs_next_leaf, but does not try to preserve
4068 * and fixup the path. It looks for and returns the next key in the
4069 * tree based on the current path and the cache_only and min_trans
4070 * parameters.
4071 *
4072 * 0 is returned if another key is found, < 0 if there are any errors
4073 * and 1 is returned if there are no higher keys in the tree
4074 *
4075 * path->keep_locks should be set to 1 on the search made before
4076 * calling this function.
4077 */
4078int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4079 struct btrfs_key *key, int level,
4080 int cache_only, u64 min_trans)
4081{
4082 int slot;
4083 struct extent_buffer *c;
4084
4085 WARN_ON(!path->keep_locks);
4086 while (level < BTRFS_MAX_LEVEL) {
4087 if (!path->nodes[level])
4088 return 1;
4089
4090 slot = path->slots[level] + 1;
4091 c = path->nodes[level];
4092next:
4093 if (slot >= btrfs_header_nritems(c)) {
4094 int ret;
4095 int orig_lowest;
4096 struct btrfs_key cur_key;
4097 if (level + 1 >= BTRFS_MAX_LEVEL ||
4098 !path->nodes[level + 1])
4099 return 1;
4100
4101 if (path->locks[level + 1]) {
4102 level++;
4103 continue;
4104 }
4105
4106 slot = btrfs_header_nritems(c) - 1;
4107 if (level == 0)
4108 btrfs_item_key_to_cpu(c, &cur_key, slot);
4109 else
4110 btrfs_node_key_to_cpu(c, &cur_key, slot);
4111
4112 orig_lowest = path->lowest_level;
4113 btrfs_release_path(path);
4114 path->lowest_level = level;
4115 ret = btrfs_search_slot(NULL, root, &cur_key, path,
4116 0, 0);
4117 path->lowest_level = orig_lowest;
4118 if (ret < 0)
4119 return ret;
4120
4121 c = path->nodes[level];
4122 slot = path->slots[level];
4123 if (ret == 0)
4124 slot++;
4125 goto next;
4126 }
4127
4128 if (level == 0)
4129 btrfs_item_key_to_cpu(c, key, slot);
4130 else {
4131 u64 blockptr = btrfs_node_blockptr(c, slot);
4132 u64 gen = btrfs_node_ptr_generation(c, slot);
4133
4134 if (cache_only) {
4135 struct extent_buffer *cur;
4136 cur = btrfs_find_tree_block(root, blockptr,
4137 btrfs_level_size(root, level - 1));
4138 if (!cur || !btrfs_buffer_uptodate(cur, gen)) {
4139 slot++;
4140 if (cur)
4141 free_extent_buffer(cur);
4142 goto next;
4143 }
4144 free_extent_buffer(cur);
4145 }
4146 if (gen < min_trans) {
4147 slot++;
4148 goto next;
4149 }
4150 btrfs_node_key_to_cpu(c, key, slot);
4151 }
4152 return 0;
4153 }
4154 return 1;
4155}
4156
4157/*
4158 * search the tree again to find a leaf with greater keys
4159 * returns 0 if it found something or 1 if there are no greater leaves.
4160 * returns < 0 on io errors.
4161 */
4162int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
4163{
4164 int slot;
4165 int level;
4166 struct extent_buffer *c;
4167 struct extent_buffer *next;
4168 struct btrfs_key key;
4169 u32 nritems;
4170 int ret;
4171 int old_spinning = path->leave_spinning;
4172 int next_rw_lock = 0;
4173
4174 nritems = btrfs_header_nritems(path->nodes[0]);
4175 if (nritems == 0)
4176 return 1;
4177
4178 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4179again:
4180 level = 1;
4181 next = NULL;
4182 next_rw_lock = 0;
4183 btrfs_release_path(path);
4184
4185 path->keep_locks = 1;
4186 path->leave_spinning = 1;
4187
4188 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4189 path->keep_locks = 0;
4190
4191 if (ret < 0)
4192 return ret;
4193
4194 nritems = btrfs_header_nritems(path->nodes[0]);
4195 /*
4196 * by releasing the path above we dropped all our locks. A balance
4197 * could have added more items next to the key that used to be
4198 * at the very end of the block. So, check again here and
4199 * advance the path if there are now more items available.
4200 */
4201 if (nritems > 0 && path->slots[0] < nritems - 1) {
4202 if (ret == 0)
4203 path->slots[0]++;
4204 ret = 0;
4205 goto done;
4206 }
4207
4208 while (level < BTRFS_MAX_LEVEL) {
4209 if (!path->nodes[level]) {
4210 ret = 1;
4211 goto done;
4212 }
4213
4214 slot = path->slots[level] + 1;
4215 c = path->nodes[level];
4216 if (slot >= btrfs_header_nritems(c)) {
4217 level++;
4218 if (level == BTRFS_MAX_LEVEL) {
4219 ret = 1;
4220 goto done;
4221 }
4222 continue;
4223 }
4224
4225 if (next) {
4226 btrfs_tree_unlock_rw(next, next_rw_lock);
4227 free_extent_buffer(next);
4228 }
4229
4230 next = c;
4231 next_rw_lock = path->locks[level];
4232 ret = read_block_for_search(NULL, root, path, &next, level,
4233 slot, &key);
4234 if (ret == -EAGAIN)
4235 goto again;
4236
4237 if (ret < 0) {
4238 btrfs_release_path(path);
4239 goto done;
4240 }
4241
4242 if (!path->skip_locking) {
4243 ret = btrfs_try_tree_read_lock(next);
4244 if (!ret) {
4245 btrfs_set_path_blocking(path);
4246 btrfs_tree_read_lock(next);
4247 btrfs_clear_path_blocking(path, next,
4248 BTRFS_READ_LOCK);
4249 }
4250 next_rw_lock = BTRFS_READ_LOCK;
4251 }
4252 break;
4253 }
4254 path->slots[level] = slot;
4255 while (1) {
4256 level--;
4257 c = path->nodes[level];
4258 if (path->locks[level])
4259 btrfs_tree_unlock_rw(c, path->locks[level]);
4260
4261 free_extent_buffer(c);
4262 path->nodes[level] = next;
4263 path->slots[level] = 0;
4264 if (!path->skip_locking)
4265 path->locks[level] = next_rw_lock;
4266 if (!level)
4267 break;
4268
4269 ret = read_block_for_search(NULL, root, path, &next, level,
4270 0, &key);
4271 if (ret == -EAGAIN)
4272 goto again;
4273
4274 if (ret < 0) {
4275 btrfs_release_path(path);
4276 goto done;
4277 }
4278
4279 if (!path->skip_locking) {
4280 ret = btrfs_try_tree_read_lock(next);
4281 if (!ret) {
4282 btrfs_set_path_blocking(path);
4283 btrfs_tree_read_lock(next);
4284 btrfs_clear_path_blocking(path, next,
4285 BTRFS_READ_LOCK);
4286 }
4287 next_rw_lock = BTRFS_READ_LOCK;
4288 }
4289 }
4290 ret = 0;
4291done:
4292 unlock_up(path, 0, 1);
4293 path->leave_spinning = old_spinning;
4294 if (!old_spinning)
4295 btrfs_set_path_blocking(path);
4296
4297 return ret;
4298}
4299
4300/*
4301 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4302 * searching until it gets past min_objectid or finds an item of 'type'
4303 *
4304 * returns 0 if something is found, 1 if nothing was found and < 0 on error
4305 */
4306int btrfs_previous_item(struct btrfs_root *root,
4307 struct btrfs_path *path, u64 min_objectid,
4308 int type)
4309{
4310 struct btrfs_key found_key;
4311 struct extent_buffer *leaf;
4312 u32 nritems;
4313 int ret;
4314
4315 while (1) {
4316 if (path->slots[0] == 0) {
4317 btrfs_set_path_blocking(path);
4318 ret = btrfs_prev_leaf(root, path);
4319 if (ret != 0)
4320 return ret;
4321 } else {
4322 path->slots[0]--;
4323 }
4324 leaf = path->nodes[0];
4325 nritems = btrfs_header_nritems(leaf);
4326 if (nritems == 0)
4327 return 1;
4328 if (path->slots[0] == nritems)
4329 path->slots[0]--;
4330
4331 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4332 if (found_key.objectid < min_objectid)
4333 break;
4334 if (found_key.type == type)
4335 return 0;
4336 if (found_key.objectid == min_objectid &&
4337 found_key.type < type)
4338 break;
4339 }
4340 return 1;
4341}