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 <linux/error-injection.h>
11#include "messages.h"
12#include "ctree.h"
13#include "disk-io.h"
14#include "transaction.h"
15#include "print-tree.h"
16#include "locking.h"
17#include "volumes.h"
18#include "qgroup.h"
19#include "tree-mod-log.h"
20#include "tree-checker.h"
21#include "fs.h"
22#include "accessors.h"
23#include "extent-tree.h"
24#include "relocation.h"
25#include "file-item.h"
26
27static struct kmem_cache *btrfs_path_cachep;
28
29static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
30 *root, struct btrfs_path *path, int level);
31static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
32 const struct btrfs_key *ins_key, struct btrfs_path *path,
33 int data_size, int extend);
34static int push_node_left(struct btrfs_trans_handle *trans,
35 struct extent_buffer *dst,
36 struct extent_buffer *src, int empty);
37static int balance_node_right(struct btrfs_trans_handle *trans,
38 struct extent_buffer *dst_buf,
39 struct extent_buffer *src_buf);
40
41static const struct btrfs_csums {
42 u16 size;
43 const char name[10];
44 const char driver[12];
45} btrfs_csums[] = {
46 [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
47 [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
48 [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
49 [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
50 .driver = "blake2b-256" },
51};
52
53/*
54 * The leaf data grows from end-to-front in the node. this returns the address
55 * of the start of the last item, which is the stop of the leaf data stack.
56 */
57static unsigned int leaf_data_end(const struct extent_buffer *leaf)
58{
59 u32 nr = btrfs_header_nritems(leaf);
60
61 if (nr == 0)
62 return BTRFS_LEAF_DATA_SIZE(leaf->fs_info);
63 return btrfs_item_offset(leaf, nr - 1);
64}
65
66/*
67 * Move data in a @leaf (using memmove, safe for overlapping ranges).
68 *
69 * @leaf: leaf that we're doing a memmove on
70 * @dst_offset: item data offset we're moving to
71 * @src_offset: item data offset were' moving from
72 * @len: length of the data we're moving
73 *
74 * Wrapper around memmove_extent_buffer() that takes into account the header on
75 * the leaf. The btrfs_item offset's start directly after the header, so we
76 * have to adjust any offsets to account for the header in the leaf. This
77 * handles that math to simplify the callers.
78 */
79static inline void memmove_leaf_data(const struct extent_buffer *leaf,
80 unsigned long dst_offset,
81 unsigned long src_offset,
82 unsigned long len)
83{
84 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, 0) + dst_offset,
85 btrfs_item_nr_offset(leaf, 0) + src_offset, len);
86}
87
88/*
89 * Copy item data from @src into @dst at the given @offset.
90 *
91 * @dst: destination leaf that we're copying into
92 * @src: source leaf that we're copying from
93 * @dst_offset: item data offset we're copying to
94 * @src_offset: item data offset were' copying from
95 * @len: length of the data we're copying
96 *
97 * Wrapper around copy_extent_buffer() that takes into account the header on
98 * the leaf. The btrfs_item offset's start directly after the header, so we
99 * have to adjust any offsets to account for the header in the leaf. This
100 * handles that math to simplify the callers.
101 */
102static inline void copy_leaf_data(const struct extent_buffer *dst,
103 const struct extent_buffer *src,
104 unsigned long dst_offset,
105 unsigned long src_offset, unsigned long len)
106{
107 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, 0) + dst_offset,
108 btrfs_item_nr_offset(src, 0) + src_offset, len);
109}
110
111/*
112 * Move items in a @leaf (using memmove).
113 *
114 * @dst: destination leaf for the items
115 * @dst_item: the item nr we're copying into
116 * @src_item: the item nr we're copying from
117 * @nr_items: the number of items to copy
118 *
119 * Wrapper around memmove_extent_buffer() that does the math to get the
120 * appropriate offsets into the leaf from the item numbers.
121 */
122static inline void memmove_leaf_items(const struct extent_buffer *leaf,
123 int dst_item, int src_item, int nr_items)
124{
125 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, dst_item),
126 btrfs_item_nr_offset(leaf, src_item),
127 nr_items * sizeof(struct btrfs_item));
128}
129
130/*
131 * Copy items from @src into @dst at the given @offset.
132 *
133 * @dst: destination leaf for the items
134 * @src: source leaf for the items
135 * @dst_item: the item nr we're copying into
136 * @src_item: the item nr we're copying from
137 * @nr_items: the number of items to copy
138 *
139 * Wrapper around copy_extent_buffer() that does the math to get the
140 * appropriate offsets into the leaf from the item numbers.
141 */
142static inline void copy_leaf_items(const struct extent_buffer *dst,
143 const struct extent_buffer *src,
144 int dst_item, int src_item, int nr_items)
145{
146 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, dst_item),
147 btrfs_item_nr_offset(src, src_item),
148 nr_items * sizeof(struct btrfs_item));
149}
150
151/* This exists for btrfs-progs usages. */
152u16 btrfs_csum_type_size(u16 type)
153{
154 return btrfs_csums[type].size;
155}
156
157int btrfs_super_csum_size(const struct btrfs_super_block *s)
158{
159 u16 t = btrfs_super_csum_type(s);
160 /*
161 * csum type is validated at mount time
162 */
163 return btrfs_csum_type_size(t);
164}
165
166const char *btrfs_super_csum_name(u16 csum_type)
167{
168 /* csum type is validated at mount time */
169 return btrfs_csums[csum_type].name;
170}
171
172/*
173 * Return driver name if defined, otherwise the name that's also a valid driver
174 * name
175 */
176const char *btrfs_super_csum_driver(u16 csum_type)
177{
178 /* csum type is validated at mount time */
179 return btrfs_csums[csum_type].driver[0] ?
180 btrfs_csums[csum_type].driver :
181 btrfs_csums[csum_type].name;
182}
183
184size_t __attribute_const__ btrfs_get_num_csums(void)
185{
186 return ARRAY_SIZE(btrfs_csums);
187}
188
189struct btrfs_path *btrfs_alloc_path(void)
190{
191 might_sleep();
192
193 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
194}
195
196/* this also releases the path */
197void btrfs_free_path(struct btrfs_path *p)
198{
199 if (!p)
200 return;
201 btrfs_release_path(p);
202 kmem_cache_free(btrfs_path_cachep, p);
203}
204
205/*
206 * path release drops references on the extent buffers in the path
207 * and it drops any locks held by this path
208 *
209 * It is safe to call this on paths that no locks or extent buffers held.
210 */
211noinline void btrfs_release_path(struct btrfs_path *p)
212{
213 int i;
214
215 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
216 p->slots[i] = 0;
217 if (!p->nodes[i])
218 continue;
219 if (p->locks[i]) {
220 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
221 p->locks[i] = 0;
222 }
223 free_extent_buffer(p->nodes[i]);
224 p->nodes[i] = NULL;
225 }
226}
227
228/*
229 * We want the transaction abort to print stack trace only for errors where the
230 * cause could be a bug, eg. due to ENOSPC, and not for common errors that are
231 * caused by external factors.
232 */
233bool __cold abort_should_print_stack(int error)
234{
235 switch (error) {
236 case -EIO:
237 case -EROFS:
238 case -ENOMEM:
239 return false;
240 }
241 return true;
242}
243
244/*
245 * safely gets a reference on the root node of a tree. A lock
246 * is not taken, so a concurrent writer may put a different node
247 * at the root of the tree. See btrfs_lock_root_node for the
248 * looping required.
249 *
250 * The extent buffer returned by this has a reference taken, so
251 * it won't disappear. It may stop being the root of the tree
252 * at any time because there are no locks held.
253 */
254struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
255{
256 struct extent_buffer *eb;
257
258 while (1) {
259 rcu_read_lock();
260 eb = rcu_dereference(root->node);
261
262 /*
263 * RCU really hurts here, we could free up the root node because
264 * it was COWed but we may not get the new root node yet so do
265 * the inc_not_zero dance and if it doesn't work then
266 * synchronize_rcu and try again.
267 */
268 if (atomic_inc_not_zero(&eb->refs)) {
269 rcu_read_unlock();
270 break;
271 }
272 rcu_read_unlock();
273 synchronize_rcu();
274 }
275 return eb;
276}
277
278/*
279 * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
280 * just get put onto a simple dirty list. Transaction walks this list to make
281 * sure they get properly updated on disk.
282 */
283static void add_root_to_dirty_list(struct btrfs_root *root)
284{
285 struct btrfs_fs_info *fs_info = root->fs_info;
286
287 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
288 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
289 return;
290
291 spin_lock(&fs_info->trans_lock);
292 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
293 /* Want the extent tree to be the last on the list */
294 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
295 list_move_tail(&root->dirty_list,
296 &fs_info->dirty_cowonly_roots);
297 else
298 list_move(&root->dirty_list,
299 &fs_info->dirty_cowonly_roots);
300 }
301 spin_unlock(&fs_info->trans_lock);
302}
303
304/*
305 * used by snapshot creation to make a copy of a root for a tree with
306 * a given objectid. The buffer with the new root node is returned in
307 * cow_ret, and this func returns zero on success or a negative error code.
308 */
309int btrfs_copy_root(struct btrfs_trans_handle *trans,
310 struct btrfs_root *root,
311 struct extent_buffer *buf,
312 struct extent_buffer **cow_ret, u64 new_root_objectid)
313{
314 struct btrfs_fs_info *fs_info = root->fs_info;
315 struct extent_buffer *cow;
316 int ret = 0;
317 int level;
318 struct btrfs_disk_key disk_key;
319 u64 reloc_src_root = 0;
320
321 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
322 trans->transid != fs_info->running_transaction->transid);
323 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
324 trans->transid != root->last_trans);
325
326 level = btrfs_header_level(buf);
327 if (level == 0)
328 btrfs_item_key(buf, &disk_key, 0);
329 else
330 btrfs_node_key(buf, &disk_key, 0);
331
332 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
333 reloc_src_root = btrfs_header_owner(buf);
334 cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
335 &disk_key, level, buf->start, 0,
336 reloc_src_root, BTRFS_NESTING_NEW_ROOT);
337 if (IS_ERR(cow))
338 return PTR_ERR(cow);
339
340 copy_extent_buffer_full(cow, buf);
341 btrfs_set_header_bytenr(cow, cow->start);
342 btrfs_set_header_generation(cow, trans->transid);
343 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
344 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
345 BTRFS_HEADER_FLAG_RELOC);
346 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
347 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
348 else
349 btrfs_set_header_owner(cow, new_root_objectid);
350
351 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
352
353 WARN_ON(btrfs_header_generation(buf) > trans->transid);
354 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
355 ret = btrfs_inc_ref(trans, root, cow, 1);
356 else
357 ret = btrfs_inc_ref(trans, root, cow, 0);
358 if (ret) {
359 btrfs_tree_unlock(cow);
360 free_extent_buffer(cow);
361 btrfs_abort_transaction(trans, ret);
362 return ret;
363 }
364
365 btrfs_mark_buffer_dirty(trans, cow);
366 *cow_ret = cow;
367 return 0;
368}
369
370/*
371 * check if the tree block can be shared by multiple trees
372 */
373bool btrfs_block_can_be_shared(struct btrfs_trans_handle *trans,
374 struct btrfs_root *root,
375 struct extent_buffer *buf)
376{
377 const u64 buf_gen = btrfs_header_generation(buf);
378
379 /*
380 * Tree blocks not in shareable trees and tree roots are never shared.
381 * If a block was allocated after the last snapshot and the block was
382 * not allocated by tree relocation, we know the block is not shared.
383 */
384
385 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
386 return false;
387
388 if (buf == root->node)
389 return false;
390
391 if (buf_gen > btrfs_root_last_snapshot(&root->root_item) &&
392 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC))
393 return false;
394
395 if (buf != root->commit_root)
396 return true;
397
398 /*
399 * An extent buffer that used to be the commit root may still be shared
400 * because the tree height may have increased and it became a child of a
401 * higher level root. This can happen when snapshotting a subvolume
402 * created in the current transaction.
403 */
404 if (buf_gen == trans->transid)
405 return true;
406
407 return false;
408}
409
410static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
411 struct btrfs_root *root,
412 struct extent_buffer *buf,
413 struct extent_buffer *cow,
414 int *last_ref)
415{
416 struct btrfs_fs_info *fs_info = root->fs_info;
417 u64 refs;
418 u64 owner;
419 u64 flags;
420 u64 new_flags = 0;
421 int ret;
422
423 /*
424 * Backrefs update rules:
425 *
426 * Always use full backrefs for extent pointers in tree block
427 * allocated by tree relocation.
428 *
429 * If a shared tree block is no longer referenced by its owner
430 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
431 * use full backrefs for extent pointers in tree block.
432 *
433 * If a tree block is been relocating
434 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
435 * use full backrefs for extent pointers in tree block.
436 * The reason for this is some operations (such as drop tree)
437 * are only allowed for blocks use full backrefs.
438 */
439
440 if (btrfs_block_can_be_shared(trans, root, buf)) {
441 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
442 btrfs_header_level(buf), 1,
443 &refs, &flags, NULL);
444 if (ret)
445 return ret;
446 if (unlikely(refs == 0)) {
447 btrfs_crit(fs_info,
448 "found 0 references for tree block at bytenr %llu level %d root %llu",
449 buf->start, btrfs_header_level(buf),
450 btrfs_root_id(root));
451 ret = -EUCLEAN;
452 btrfs_abort_transaction(trans, ret);
453 return ret;
454 }
455 } else {
456 refs = 1;
457 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
458 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
459 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
460 else
461 flags = 0;
462 }
463
464 owner = btrfs_header_owner(buf);
465 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
466 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
467
468 if (refs > 1) {
469 if ((owner == root->root_key.objectid ||
470 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
471 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
472 ret = btrfs_inc_ref(trans, root, buf, 1);
473 if (ret)
474 return ret;
475
476 if (root->root_key.objectid ==
477 BTRFS_TREE_RELOC_OBJECTID) {
478 ret = btrfs_dec_ref(trans, root, buf, 0);
479 if (ret)
480 return ret;
481 ret = btrfs_inc_ref(trans, root, cow, 1);
482 if (ret)
483 return ret;
484 }
485 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
486 } else {
487
488 if (root->root_key.objectid ==
489 BTRFS_TREE_RELOC_OBJECTID)
490 ret = btrfs_inc_ref(trans, root, cow, 1);
491 else
492 ret = btrfs_inc_ref(trans, root, cow, 0);
493 if (ret)
494 return ret;
495 }
496 if (new_flags != 0) {
497 ret = btrfs_set_disk_extent_flags(trans, buf, new_flags);
498 if (ret)
499 return ret;
500 }
501 } else {
502 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
503 if (root->root_key.objectid ==
504 BTRFS_TREE_RELOC_OBJECTID)
505 ret = btrfs_inc_ref(trans, root, cow, 1);
506 else
507 ret = btrfs_inc_ref(trans, root, cow, 0);
508 if (ret)
509 return ret;
510 ret = btrfs_dec_ref(trans, root, buf, 1);
511 if (ret)
512 return ret;
513 }
514 btrfs_clear_buffer_dirty(trans, buf);
515 *last_ref = 1;
516 }
517 return 0;
518}
519
520/*
521 * does the dirty work in cow of a single block. The parent block (if
522 * supplied) is updated to point to the new cow copy. The new buffer is marked
523 * dirty and returned locked. If you modify the block it needs to be marked
524 * dirty again.
525 *
526 * search_start -- an allocation hint for the new block
527 *
528 * empty_size -- a hint that you plan on doing more cow. This is the size in
529 * bytes the allocator should try to find free next to the block it returns.
530 * This is just a hint and may be ignored by the allocator.
531 */
532int btrfs_force_cow_block(struct btrfs_trans_handle *trans,
533 struct btrfs_root *root,
534 struct extent_buffer *buf,
535 struct extent_buffer *parent, int parent_slot,
536 struct extent_buffer **cow_ret,
537 u64 search_start, u64 empty_size,
538 enum btrfs_lock_nesting nest)
539{
540 struct btrfs_fs_info *fs_info = root->fs_info;
541 struct btrfs_disk_key disk_key;
542 struct extent_buffer *cow;
543 int level, ret;
544 int last_ref = 0;
545 int unlock_orig = 0;
546 u64 parent_start = 0;
547 u64 reloc_src_root = 0;
548
549 if (*cow_ret == buf)
550 unlock_orig = 1;
551
552 btrfs_assert_tree_write_locked(buf);
553
554 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
555 trans->transid != fs_info->running_transaction->transid);
556 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
557 trans->transid != root->last_trans);
558
559 level = btrfs_header_level(buf);
560
561 if (level == 0)
562 btrfs_item_key(buf, &disk_key, 0);
563 else
564 btrfs_node_key(buf, &disk_key, 0);
565
566 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
567 if (parent)
568 parent_start = parent->start;
569 reloc_src_root = btrfs_header_owner(buf);
570 }
571 cow = btrfs_alloc_tree_block(trans, root, parent_start,
572 root->root_key.objectid, &disk_key, level,
573 search_start, empty_size, reloc_src_root, nest);
574 if (IS_ERR(cow))
575 return PTR_ERR(cow);
576
577 /* cow is set to blocking by btrfs_init_new_buffer */
578
579 copy_extent_buffer_full(cow, buf);
580 btrfs_set_header_bytenr(cow, cow->start);
581 btrfs_set_header_generation(cow, trans->transid);
582 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
583 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
584 BTRFS_HEADER_FLAG_RELOC);
585 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
586 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
587 else
588 btrfs_set_header_owner(cow, root->root_key.objectid);
589
590 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
591
592 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
593 if (ret) {
594 btrfs_tree_unlock(cow);
595 free_extent_buffer(cow);
596 btrfs_abort_transaction(trans, ret);
597 return ret;
598 }
599
600 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
601 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
602 if (ret) {
603 btrfs_tree_unlock(cow);
604 free_extent_buffer(cow);
605 btrfs_abort_transaction(trans, ret);
606 return ret;
607 }
608 }
609
610 if (buf == root->node) {
611 WARN_ON(parent && parent != buf);
612 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
613 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
614 parent_start = buf->start;
615
616 ret = btrfs_tree_mod_log_insert_root(root->node, cow, true);
617 if (ret < 0) {
618 btrfs_tree_unlock(cow);
619 free_extent_buffer(cow);
620 btrfs_abort_transaction(trans, ret);
621 return ret;
622 }
623 atomic_inc(&cow->refs);
624 rcu_assign_pointer(root->node, cow);
625
626 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
627 parent_start, last_ref);
628 free_extent_buffer(buf);
629 add_root_to_dirty_list(root);
630 } else {
631 WARN_ON(trans->transid != btrfs_header_generation(parent));
632 ret = btrfs_tree_mod_log_insert_key(parent, parent_slot,
633 BTRFS_MOD_LOG_KEY_REPLACE);
634 if (ret) {
635 btrfs_tree_unlock(cow);
636 free_extent_buffer(cow);
637 btrfs_abort_transaction(trans, ret);
638 return ret;
639 }
640 btrfs_set_node_blockptr(parent, parent_slot,
641 cow->start);
642 btrfs_set_node_ptr_generation(parent, parent_slot,
643 trans->transid);
644 btrfs_mark_buffer_dirty(trans, parent);
645 if (last_ref) {
646 ret = btrfs_tree_mod_log_free_eb(buf);
647 if (ret) {
648 btrfs_tree_unlock(cow);
649 free_extent_buffer(cow);
650 btrfs_abort_transaction(trans, ret);
651 return ret;
652 }
653 }
654 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
655 parent_start, last_ref);
656 }
657 if (unlock_orig)
658 btrfs_tree_unlock(buf);
659 free_extent_buffer_stale(buf);
660 btrfs_mark_buffer_dirty(trans, cow);
661 *cow_ret = cow;
662 return 0;
663}
664
665static inline int should_cow_block(struct btrfs_trans_handle *trans,
666 struct btrfs_root *root,
667 struct extent_buffer *buf)
668{
669 if (btrfs_is_testing(root->fs_info))
670 return 0;
671
672 /* Ensure we can see the FORCE_COW bit */
673 smp_mb__before_atomic();
674
675 /*
676 * We do not need to cow a block if
677 * 1) this block is not created or changed in this transaction;
678 * 2) this block does not belong to TREE_RELOC tree;
679 * 3) the root is not forced COW.
680 *
681 * What is forced COW:
682 * when we create snapshot during committing the transaction,
683 * after we've finished copying src root, we must COW the shared
684 * block to ensure the metadata consistency.
685 */
686 if (btrfs_header_generation(buf) == trans->transid &&
687 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
688 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
689 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
690 !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
691 return 0;
692 return 1;
693}
694
695/*
696 * COWs a single block, see btrfs_force_cow_block() for the real work.
697 * This version of it has extra checks so that a block isn't COWed more than
698 * once per transaction, as long as it hasn't been written yet
699 */
700int btrfs_cow_block(struct btrfs_trans_handle *trans,
701 struct btrfs_root *root, struct extent_buffer *buf,
702 struct extent_buffer *parent, int parent_slot,
703 struct extent_buffer **cow_ret,
704 enum btrfs_lock_nesting nest)
705{
706 struct btrfs_fs_info *fs_info = root->fs_info;
707 u64 search_start;
708 int ret;
709
710 if (unlikely(test_bit(BTRFS_ROOT_DELETING, &root->state))) {
711 btrfs_abort_transaction(trans, -EUCLEAN);
712 btrfs_crit(fs_info,
713 "attempt to COW block %llu on root %llu that is being deleted",
714 buf->start, btrfs_root_id(root));
715 return -EUCLEAN;
716 }
717
718 /*
719 * COWing must happen through a running transaction, which always
720 * matches the current fs generation (it's a transaction with a state
721 * less than TRANS_STATE_UNBLOCKED). If it doesn't, then turn the fs
722 * into error state to prevent the commit of any transaction.
723 */
724 if (unlikely(trans->transaction != fs_info->running_transaction ||
725 trans->transid != fs_info->generation)) {
726 btrfs_abort_transaction(trans, -EUCLEAN);
727 btrfs_crit(fs_info,
728"unexpected transaction when attempting to COW block %llu on root %llu, transaction %llu running transaction %llu fs generation %llu",
729 buf->start, btrfs_root_id(root), trans->transid,
730 fs_info->running_transaction->transid,
731 fs_info->generation);
732 return -EUCLEAN;
733 }
734
735 if (!should_cow_block(trans, root, buf)) {
736 *cow_ret = buf;
737 return 0;
738 }
739
740 search_start = round_down(buf->start, SZ_1G);
741
742 /*
743 * Before CoWing this block for later modification, check if it's
744 * the subtree root and do the delayed subtree trace if needed.
745 *
746 * Also We don't care about the error, as it's handled internally.
747 */
748 btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
749 ret = btrfs_force_cow_block(trans, root, buf, parent, parent_slot,
750 cow_ret, search_start, 0, nest);
751
752 trace_btrfs_cow_block(root, buf, *cow_ret);
753
754 return ret;
755}
756ALLOW_ERROR_INJECTION(btrfs_cow_block, ERRNO);
757
758/*
759 * same as comp_keys only with two btrfs_key's
760 */
761int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
762{
763 if (k1->objectid > k2->objectid)
764 return 1;
765 if (k1->objectid < k2->objectid)
766 return -1;
767 if (k1->type > k2->type)
768 return 1;
769 if (k1->type < k2->type)
770 return -1;
771 if (k1->offset > k2->offset)
772 return 1;
773 if (k1->offset < k2->offset)
774 return -1;
775 return 0;
776}
777
778/*
779 * Search for a key in the given extent_buffer.
780 *
781 * The lower boundary for the search is specified by the slot number @first_slot.
782 * Use a value of 0 to search over the whole extent buffer. Works for both
783 * leaves and nodes.
784 *
785 * The slot in the extent buffer is returned via @slot. If the key exists in the
786 * extent buffer, then @slot will point to the slot where the key is, otherwise
787 * it points to the slot where you would insert the key.
788 *
789 * Slot may point to the total number of items (i.e. one position beyond the last
790 * key) if the key is bigger than the last key in the extent buffer.
791 */
792int btrfs_bin_search(struct extent_buffer *eb, int first_slot,
793 const struct btrfs_key *key, int *slot)
794{
795 unsigned long p;
796 int item_size;
797 /*
798 * Use unsigned types for the low and high slots, so that we get a more
799 * efficient division in the search loop below.
800 */
801 u32 low = first_slot;
802 u32 high = btrfs_header_nritems(eb);
803 int ret;
804 const int key_size = sizeof(struct btrfs_disk_key);
805
806 if (unlikely(low > high)) {
807 btrfs_err(eb->fs_info,
808 "%s: low (%u) > high (%u) eb %llu owner %llu level %d",
809 __func__, low, high, eb->start,
810 btrfs_header_owner(eb), btrfs_header_level(eb));
811 return -EINVAL;
812 }
813
814 if (btrfs_header_level(eb) == 0) {
815 p = offsetof(struct btrfs_leaf, items);
816 item_size = sizeof(struct btrfs_item);
817 } else {
818 p = offsetof(struct btrfs_node, ptrs);
819 item_size = sizeof(struct btrfs_key_ptr);
820 }
821
822 while (low < high) {
823 const int unit_size = folio_size(eb->folios[0]);
824 unsigned long oil;
825 unsigned long offset;
826 struct btrfs_disk_key *tmp;
827 struct btrfs_disk_key unaligned;
828 int mid;
829
830 mid = (low + high) / 2;
831 offset = p + mid * item_size;
832 oil = get_eb_offset_in_folio(eb, offset);
833
834 if (oil + key_size <= unit_size) {
835 const unsigned long idx = get_eb_folio_index(eb, offset);
836 char *kaddr = folio_address(eb->folios[idx]);
837
838 oil = get_eb_offset_in_folio(eb, offset);
839 tmp = (struct btrfs_disk_key *)(kaddr + oil);
840 } else {
841 read_extent_buffer(eb, &unaligned, offset, key_size);
842 tmp = &unaligned;
843 }
844
845 ret = btrfs_comp_keys(tmp, key);
846
847 if (ret < 0)
848 low = mid + 1;
849 else if (ret > 0)
850 high = mid;
851 else {
852 *slot = mid;
853 return 0;
854 }
855 }
856 *slot = low;
857 return 1;
858}
859
860static void root_add_used_bytes(struct btrfs_root *root)
861{
862 spin_lock(&root->accounting_lock);
863 btrfs_set_root_used(&root->root_item,
864 btrfs_root_used(&root->root_item) + root->fs_info->nodesize);
865 spin_unlock(&root->accounting_lock);
866}
867
868static void root_sub_used_bytes(struct btrfs_root *root)
869{
870 spin_lock(&root->accounting_lock);
871 btrfs_set_root_used(&root->root_item,
872 btrfs_root_used(&root->root_item) - root->fs_info->nodesize);
873 spin_unlock(&root->accounting_lock);
874}
875
876/* given a node and slot number, this reads the blocks it points to. The
877 * extent buffer is returned with a reference taken (but unlocked).
878 */
879struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
880 int slot)
881{
882 int level = btrfs_header_level(parent);
883 struct btrfs_tree_parent_check check = { 0 };
884 struct extent_buffer *eb;
885
886 if (slot < 0 || slot >= btrfs_header_nritems(parent))
887 return ERR_PTR(-ENOENT);
888
889 ASSERT(level);
890
891 check.level = level - 1;
892 check.transid = btrfs_node_ptr_generation(parent, slot);
893 check.owner_root = btrfs_header_owner(parent);
894 check.has_first_key = true;
895 btrfs_node_key_to_cpu(parent, &check.first_key, slot);
896
897 eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
898 &check);
899 if (IS_ERR(eb))
900 return eb;
901 if (!extent_buffer_uptodate(eb)) {
902 free_extent_buffer(eb);
903 return ERR_PTR(-EIO);
904 }
905
906 return eb;
907}
908
909/*
910 * node level balancing, used to make sure nodes are in proper order for
911 * item deletion. We balance from the top down, so we have to make sure
912 * that a deletion won't leave an node completely empty later on.
913 */
914static noinline int balance_level(struct btrfs_trans_handle *trans,
915 struct btrfs_root *root,
916 struct btrfs_path *path, int level)
917{
918 struct btrfs_fs_info *fs_info = root->fs_info;
919 struct extent_buffer *right = NULL;
920 struct extent_buffer *mid;
921 struct extent_buffer *left = NULL;
922 struct extent_buffer *parent = NULL;
923 int ret = 0;
924 int wret;
925 int pslot;
926 int orig_slot = path->slots[level];
927 u64 orig_ptr;
928
929 ASSERT(level > 0);
930
931 mid = path->nodes[level];
932
933 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK);
934 WARN_ON(btrfs_header_generation(mid) != trans->transid);
935
936 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
937
938 if (level < BTRFS_MAX_LEVEL - 1) {
939 parent = path->nodes[level + 1];
940 pslot = path->slots[level + 1];
941 }
942
943 /*
944 * deal with the case where there is only one pointer in the root
945 * by promoting the node below to a root
946 */
947 if (!parent) {
948 struct extent_buffer *child;
949
950 if (btrfs_header_nritems(mid) != 1)
951 return 0;
952
953 /* promote the child to a root */
954 child = btrfs_read_node_slot(mid, 0);
955 if (IS_ERR(child)) {
956 ret = PTR_ERR(child);
957 goto out;
958 }
959
960 btrfs_tree_lock(child);
961 ret = btrfs_cow_block(trans, root, child, mid, 0, &child,
962 BTRFS_NESTING_COW);
963 if (ret) {
964 btrfs_tree_unlock(child);
965 free_extent_buffer(child);
966 goto out;
967 }
968
969 ret = btrfs_tree_mod_log_insert_root(root->node, child, true);
970 if (ret < 0) {
971 btrfs_tree_unlock(child);
972 free_extent_buffer(child);
973 btrfs_abort_transaction(trans, ret);
974 goto out;
975 }
976 rcu_assign_pointer(root->node, child);
977
978 add_root_to_dirty_list(root);
979 btrfs_tree_unlock(child);
980
981 path->locks[level] = 0;
982 path->nodes[level] = NULL;
983 btrfs_clear_buffer_dirty(trans, mid);
984 btrfs_tree_unlock(mid);
985 /* once for the path */
986 free_extent_buffer(mid);
987
988 root_sub_used_bytes(root);
989 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
990 /* once for the root ptr */
991 free_extent_buffer_stale(mid);
992 return 0;
993 }
994 if (btrfs_header_nritems(mid) >
995 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
996 return 0;
997
998 if (pslot) {
999 left = btrfs_read_node_slot(parent, pslot - 1);
1000 if (IS_ERR(left)) {
1001 ret = PTR_ERR(left);
1002 left = NULL;
1003 goto out;
1004 }
1005
1006 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
1007 wret = btrfs_cow_block(trans, root, left,
1008 parent, pslot - 1, &left,
1009 BTRFS_NESTING_LEFT_COW);
1010 if (wret) {
1011 ret = wret;
1012 goto out;
1013 }
1014 }
1015
1016 if (pslot + 1 < btrfs_header_nritems(parent)) {
1017 right = btrfs_read_node_slot(parent, pslot + 1);
1018 if (IS_ERR(right)) {
1019 ret = PTR_ERR(right);
1020 right = NULL;
1021 goto out;
1022 }
1023
1024 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1025 wret = btrfs_cow_block(trans, root, right,
1026 parent, pslot + 1, &right,
1027 BTRFS_NESTING_RIGHT_COW);
1028 if (wret) {
1029 ret = wret;
1030 goto out;
1031 }
1032 }
1033
1034 /* first, try to make some room in the middle buffer */
1035 if (left) {
1036 orig_slot += btrfs_header_nritems(left);
1037 wret = push_node_left(trans, left, mid, 1);
1038 if (wret < 0)
1039 ret = wret;
1040 }
1041
1042 /*
1043 * then try to empty the right most buffer into the middle
1044 */
1045 if (right) {
1046 wret = push_node_left(trans, mid, right, 1);
1047 if (wret < 0 && wret != -ENOSPC)
1048 ret = wret;
1049 if (btrfs_header_nritems(right) == 0) {
1050 btrfs_clear_buffer_dirty(trans, right);
1051 btrfs_tree_unlock(right);
1052 ret = btrfs_del_ptr(trans, root, path, level + 1, pslot + 1);
1053 if (ret < 0) {
1054 free_extent_buffer_stale(right);
1055 right = NULL;
1056 goto out;
1057 }
1058 root_sub_used_bytes(root);
1059 btrfs_free_tree_block(trans, btrfs_root_id(root), right,
1060 0, 1);
1061 free_extent_buffer_stale(right);
1062 right = NULL;
1063 } else {
1064 struct btrfs_disk_key right_key;
1065 btrfs_node_key(right, &right_key, 0);
1066 ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1067 BTRFS_MOD_LOG_KEY_REPLACE);
1068 if (ret < 0) {
1069 btrfs_abort_transaction(trans, ret);
1070 goto out;
1071 }
1072 btrfs_set_node_key(parent, &right_key, pslot + 1);
1073 btrfs_mark_buffer_dirty(trans, parent);
1074 }
1075 }
1076 if (btrfs_header_nritems(mid) == 1) {
1077 /*
1078 * we're not allowed to leave a node with one item in the
1079 * tree during a delete. A deletion from lower in the tree
1080 * could try to delete the only pointer in this node.
1081 * So, pull some keys from the left.
1082 * There has to be a left pointer at this point because
1083 * otherwise we would have pulled some pointers from the
1084 * right
1085 */
1086 if (unlikely(!left)) {
1087 btrfs_crit(fs_info,
1088"missing left child when middle child only has 1 item, parent bytenr %llu level %d mid bytenr %llu root %llu",
1089 parent->start, btrfs_header_level(parent),
1090 mid->start, btrfs_root_id(root));
1091 ret = -EUCLEAN;
1092 btrfs_abort_transaction(trans, ret);
1093 goto out;
1094 }
1095 wret = balance_node_right(trans, mid, left);
1096 if (wret < 0) {
1097 ret = wret;
1098 goto out;
1099 }
1100 if (wret == 1) {
1101 wret = push_node_left(trans, left, mid, 1);
1102 if (wret < 0)
1103 ret = wret;
1104 }
1105 BUG_ON(wret == 1);
1106 }
1107 if (btrfs_header_nritems(mid) == 0) {
1108 btrfs_clear_buffer_dirty(trans, mid);
1109 btrfs_tree_unlock(mid);
1110 ret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
1111 if (ret < 0) {
1112 free_extent_buffer_stale(mid);
1113 mid = NULL;
1114 goto out;
1115 }
1116 root_sub_used_bytes(root);
1117 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
1118 free_extent_buffer_stale(mid);
1119 mid = NULL;
1120 } else {
1121 /* update the parent key to reflect our changes */
1122 struct btrfs_disk_key mid_key;
1123 btrfs_node_key(mid, &mid_key, 0);
1124 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1125 BTRFS_MOD_LOG_KEY_REPLACE);
1126 if (ret < 0) {
1127 btrfs_abort_transaction(trans, ret);
1128 goto out;
1129 }
1130 btrfs_set_node_key(parent, &mid_key, pslot);
1131 btrfs_mark_buffer_dirty(trans, parent);
1132 }
1133
1134 /* update the path */
1135 if (left) {
1136 if (btrfs_header_nritems(left) > orig_slot) {
1137 atomic_inc(&left->refs);
1138 /* left was locked after cow */
1139 path->nodes[level] = left;
1140 path->slots[level + 1] -= 1;
1141 path->slots[level] = orig_slot;
1142 if (mid) {
1143 btrfs_tree_unlock(mid);
1144 free_extent_buffer(mid);
1145 }
1146 } else {
1147 orig_slot -= btrfs_header_nritems(left);
1148 path->slots[level] = orig_slot;
1149 }
1150 }
1151 /* double check we haven't messed things up */
1152 if (orig_ptr !=
1153 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1154 BUG();
1155out:
1156 if (right) {
1157 btrfs_tree_unlock(right);
1158 free_extent_buffer(right);
1159 }
1160 if (left) {
1161 if (path->nodes[level] != left)
1162 btrfs_tree_unlock(left);
1163 free_extent_buffer(left);
1164 }
1165 return ret;
1166}
1167
1168/* Node balancing for insertion. Here we only split or push nodes around
1169 * when they are completely full. This is also done top down, so we
1170 * have to be pessimistic.
1171 */
1172static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1173 struct btrfs_root *root,
1174 struct btrfs_path *path, int level)
1175{
1176 struct btrfs_fs_info *fs_info = root->fs_info;
1177 struct extent_buffer *right = NULL;
1178 struct extent_buffer *mid;
1179 struct extent_buffer *left = NULL;
1180 struct extent_buffer *parent = NULL;
1181 int ret = 0;
1182 int wret;
1183 int pslot;
1184 int orig_slot = path->slots[level];
1185
1186 if (level == 0)
1187 return 1;
1188
1189 mid = path->nodes[level];
1190 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1191
1192 if (level < BTRFS_MAX_LEVEL - 1) {
1193 parent = path->nodes[level + 1];
1194 pslot = path->slots[level + 1];
1195 }
1196
1197 if (!parent)
1198 return 1;
1199
1200 /* first, try to make some room in the middle buffer */
1201 if (pslot) {
1202 u32 left_nr;
1203
1204 left = btrfs_read_node_slot(parent, pslot - 1);
1205 if (IS_ERR(left))
1206 return PTR_ERR(left);
1207
1208 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
1209
1210 left_nr = btrfs_header_nritems(left);
1211 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1212 wret = 1;
1213 } else {
1214 ret = btrfs_cow_block(trans, root, left, parent,
1215 pslot - 1, &left,
1216 BTRFS_NESTING_LEFT_COW);
1217 if (ret)
1218 wret = 1;
1219 else {
1220 wret = push_node_left(trans, left, mid, 0);
1221 }
1222 }
1223 if (wret < 0)
1224 ret = wret;
1225 if (wret == 0) {
1226 struct btrfs_disk_key disk_key;
1227 orig_slot += left_nr;
1228 btrfs_node_key(mid, &disk_key, 0);
1229 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1230 BTRFS_MOD_LOG_KEY_REPLACE);
1231 if (ret < 0) {
1232 btrfs_tree_unlock(left);
1233 free_extent_buffer(left);
1234 btrfs_abort_transaction(trans, ret);
1235 return ret;
1236 }
1237 btrfs_set_node_key(parent, &disk_key, pslot);
1238 btrfs_mark_buffer_dirty(trans, parent);
1239 if (btrfs_header_nritems(left) > orig_slot) {
1240 path->nodes[level] = left;
1241 path->slots[level + 1] -= 1;
1242 path->slots[level] = orig_slot;
1243 btrfs_tree_unlock(mid);
1244 free_extent_buffer(mid);
1245 } else {
1246 orig_slot -=
1247 btrfs_header_nritems(left);
1248 path->slots[level] = orig_slot;
1249 btrfs_tree_unlock(left);
1250 free_extent_buffer(left);
1251 }
1252 return 0;
1253 }
1254 btrfs_tree_unlock(left);
1255 free_extent_buffer(left);
1256 }
1257
1258 /*
1259 * then try to empty the right most buffer into the middle
1260 */
1261 if (pslot + 1 < btrfs_header_nritems(parent)) {
1262 u32 right_nr;
1263
1264 right = btrfs_read_node_slot(parent, pslot + 1);
1265 if (IS_ERR(right))
1266 return PTR_ERR(right);
1267
1268 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1269
1270 right_nr = btrfs_header_nritems(right);
1271 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1272 wret = 1;
1273 } else {
1274 ret = btrfs_cow_block(trans, root, right,
1275 parent, pslot + 1,
1276 &right, BTRFS_NESTING_RIGHT_COW);
1277 if (ret)
1278 wret = 1;
1279 else {
1280 wret = balance_node_right(trans, right, mid);
1281 }
1282 }
1283 if (wret < 0)
1284 ret = wret;
1285 if (wret == 0) {
1286 struct btrfs_disk_key disk_key;
1287
1288 btrfs_node_key(right, &disk_key, 0);
1289 ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1290 BTRFS_MOD_LOG_KEY_REPLACE);
1291 if (ret < 0) {
1292 btrfs_tree_unlock(right);
1293 free_extent_buffer(right);
1294 btrfs_abort_transaction(trans, ret);
1295 return ret;
1296 }
1297 btrfs_set_node_key(parent, &disk_key, pslot + 1);
1298 btrfs_mark_buffer_dirty(trans, parent);
1299
1300 if (btrfs_header_nritems(mid) <= orig_slot) {
1301 path->nodes[level] = right;
1302 path->slots[level + 1] += 1;
1303 path->slots[level] = orig_slot -
1304 btrfs_header_nritems(mid);
1305 btrfs_tree_unlock(mid);
1306 free_extent_buffer(mid);
1307 } else {
1308 btrfs_tree_unlock(right);
1309 free_extent_buffer(right);
1310 }
1311 return 0;
1312 }
1313 btrfs_tree_unlock(right);
1314 free_extent_buffer(right);
1315 }
1316 return 1;
1317}
1318
1319/*
1320 * readahead one full node of leaves, finding things that are close
1321 * to the block in 'slot', and triggering ra on them.
1322 */
1323static void reada_for_search(struct btrfs_fs_info *fs_info,
1324 struct btrfs_path *path,
1325 int level, int slot, u64 objectid)
1326{
1327 struct extent_buffer *node;
1328 struct btrfs_disk_key disk_key;
1329 u32 nritems;
1330 u64 search;
1331 u64 target;
1332 u64 nread = 0;
1333 u64 nread_max;
1334 u32 nr;
1335 u32 blocksize;
1336 u32 nscan = 0;
1337
1338 if (level != 1 && path->reada != READA_FORWARD_ALWAYS)
1339 return;
1340
1341 if (!path->nodes[level])
1342 return;
1343
1344 node = path->nodes[level];
1345
1346 /*
1347 * Since the time between visiting leaves is much shorter than the time
1348 * between visiting nodes, limit read ahead of nodes to 1, to avoid too
1349 * much IO at once (possibly random).
1350 */
1351 if (path->reada == READA_FORWARD_ALWAYS) {
1352 if (level > 1)
1353 nread_max = node->fs_info->nodesize;
1354 else
1355 nread_max = SZ_128K;
1356 } else {
1357 nread_max = SZ_64K;
1358 }
1359
1360 search = btrfs_node_blockptr(node, slot);
1361 blocksize = fs_info->nodesize;
1362 if (path->reada != READA_FORWARD_ALWAYS) {
1363 struct extent_buffer *eb;
1364
1365 eb = find_extent_buffer(fs_info, search);
1366 if (eb) {
1367 free_extent_buffer(eb);
1368 return;
1369 }
1370 }
1371
1372 target = search;
1373
1374 nritems = btrfs_header_nritems(node);
1375 nr = slot;
1376
1377 while (1) {
1378 if (path->reada == READA_BACK) {
1379 if (nr == 0)
1380 break;
1381 nr--;
1382 } else if (path->reada == READA_FORWARD ||
1383 path->reada == READA_FORWARD_ALWAYS) {
1384 nr++;
1385 if (nr >= nritems)
1386 break;
1387 }
1388 if (path->reada == READA_BACK && objectid) {
1389 btrfs_node_key(node, &disk_key, nr);
1390 if (btrfs_disk_key_objectid(&disk_key) != objectid)
1391 break;
1392 }
1393 search = btrfs_node_blockptr(node, nr);
1394 if (path->reada == READA_FORWARD_ALWAYS ||
1395 (search <= target && target - search <= 65536) ||
1396 (search > target && search - target <= 65536)) {
1397 btrfs_readahead_node_child(node, nr);
1398 nread += blocksize;
1399 }
1400 nscan++;
1401 if (nread > nread_max || nscan > 32)
1402 break;
1403 }
1404}
1405
1406static noinline void reada_for_balance(struct btrfs_path *path, int level)
1407{
1408 struct extent_buffer *parent;
1409 int slot;
1410 int nritems;
1411
1412 parent = path->nodes[level + 1];
1413 if (!parent)
1414 return;
1415
1416 nritems = btrfs_header_nritems(parent);
1417 slot = path->slots[level + 1];
1418
1419 if (slot > 0)
1420 btrfs_readahead_node_child(parent, slot - 1);
1421 if (slot + 1 < nritems)
1422 btrfs_readahead_node_child(parent, slot + 1);
1423}
1424
1425
1426/*
1427 * when we walk down the tree, it is usually safe to unlock the higher layers
1428 * in the tree. The exceptions are when our path goes through slot 0, because
1429 * operations on the tree might require changing key pointers higher up in the
1430 * tree.
1431 *
1432 * callers might also have set path->keep_locks, which tells this code to keep
1433 * the lock if the path points to the last slot in the block. This is part of
1434 * walking through the tree, and selecting the next slot in the higher block.
1435 *
1436 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
1437 * if lowest_unlock is 1, level 0 won't be unlocked
1438 */
1439static noinline void unlock_up(struct btrfs_path *path, int level,
1440 int lowest_unlock, int min_write_lock_level,
1441 int *write_lock_level)
1442{
1443 int i;
1444 int skip_level = level;
1445 bool check_skip = true;
1446
1447 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1448 if (!path->nodes[i])
1449 break;
1450 if (!path->locks[i])
1451 break;
1452
1453 if (check_skip) {
1454 if (path->slots[i] == 0) {
1455 skip_level = i + 1;
1456 continue;
1457 }
1458
1459 if (path->keep_locks) {
1460 u32 nritems;
1461
1462 nritems = btrfs_header_nritems(path->nodes[i]);
1463 if (nritems < 1 || path->slots[i] >= nritems - 1) {
1464 skip_level = i + 1;
1465 continue;
1466 }
1467 }
1468 }
1469
1470 if (i >= lowest_unlock && i > skip_level) {
1471 check_skip = false;
1472 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
1473 path->locks[i] = 0;
1474 if (write_lock_level &&
1475 i > min_write_lock_level &&
1476 i <= *write_lock_level) {
1477 *write_lock_level = i - 1;
1478 }
1479 }
1480 }
1481}
1482
1483/*
1484 * Helper function for btrfs_search_slot() and other functions that do a search
1485 * on a btree. The goal is to find a tree block in the cache (the radix tree at
1486 * fs_info->buffer_radix), but if we can't find it, or it's not up to date, read
1487 * its pages from disk.
1488 *
1489 * Returns -EAGAIN, with the path unlocked, if the caller needs to repeat the
1490 * whole btree search, starting again from the current root node.
1491 */
1492static int
1493read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
1494 struct extent_buffer **eb_ret, int level, int slot,
1495 const struct btrfs_key *key)
1496{
1497 struct btrfs_fs_info *fs_info = root->fs_info;
1498 struct btrfs_tree_parent_check check = { 0 };
1499 u64 blocknr;
1500 u64 gen;
1501 struct extent_buffer *tmp;
1502 int ret;
1503 int parent_level;
1504 bool unlock_up;
1505
1506 unlock_up = ((level + 1 < BTRFS_MAX_LEVEL) && p->locks[level + 1]);
1507 blocknr = btrfs_node_blockptr(*eb_ret, slot);
1508 gen = btrfs_node_ptr_generation(*eb_ret, slot);
1509 parent_level = btrfs_header_level(*eb_ret);
1510 btrfs_node_key_to_cpu(*eb_ret, &check.first_key, slot);
1511 check.has_first_key = true;
1512 check.level = parent_level - 1;
1513 check.transid = gen;
1514 check.owner_root = root->root_key.objectid;
1515
1516 /*
1517 * If we need to read an extent buffer from disk and we are holding locks
1518 * on upper level nodes, we unlock all the upper nodes before reading the
1519 * extent buffer, and then return -EAGAIN to the caller as it needs to
1520 * restart the search. We don't release the lock on the current level
1521 * because we need to walk this node to figure out which blocks to read.
1522 */
1523 tmp = find_extent_buffer(fs_info, blocknr);
1524 if (tmp) {
1525 if (p->reada == READA_FORWARD_ALWAYS)
1526 reada_for_search(fs_info, p, level, slot, key->objectid);
1527
1528 /* first we do an atomic uptodate check */
1529 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
1530 /*
1531 * Do extra check for first_key, eb can be stale due to
1532 * being cached, read from scrub, or have multiple
1533 * parents (shared tree blocks).
1534 */
1535 if (btrfs_verify_level_key(tmp,
1536 parent_level - 1, &check.first_key, gen)) {
1537 free_extent_buffer(tmp);
1538 return -EUCLEAN;
1539 }
1540 *eb_ret = tmp;
1541 return 0;
1542 }
1543
1544 if (p->nowait) {
1545 free_extent_buffer(tmp);
1546 return -EAGAIN;
1547 }
1548
1549 if (unlock_up)
1550 btrfs_unlock_up_safe(p, level + 1);
1551
1552 /* now we're allowed to do a blocking uptodate check */
1553 ret = btrfs_read_extent_buffer(tmp, &check);
1554 if (ret) {
1555 free_extent_buffer(tmp);
1556 btrfs_release_path(p);
1557 return -EIO;
1558 }
1559 if (btrfs_check_eb_owner(tmp, root->root_key.objectid)) {
1560 free_extent_buffer(tmp);
1561 btrfs_release_path(p);
1562 return -EUCLEAN;
1563 }
1564
1565 if (unlock_up)
1566 ret = -EAGAIN;
1567
1568 goto out;
1569 } else if (p->nowait) {
1570 return -EAGAIN;
1571 }
1572
1573 if (unlock_up) {
1574 btrfs_unlock_up_safe(p, level + 1);
1575 ret = -EAGAIN;
1576 } else {
1577 ret = 0;
1578 }
1579
1580 if (p->reada != READA_NONE)
1581 reada_for_search(fs_info, p, level, slot, key->objectid);
1582
1583 tmp = read_tree_block(fs_info, blocknr, &check);
1584 if (IS_ERR(tmp)) {
1585 btrfs_release_path(p);
1586 return PTR_ERR(tmp);
1587 }
1588 /*
1589 * If the read above didn't mark this buffer up to date,
1590 * it will never end up being up to date. Set ret to EIO now
1591 * and give up so that our caller doesn't loop forever
1592 * on our EAGAINs.
1593 */
1594 if (!extent_buffer_uptodate(tmp))
1595 ret = -EIO;
1596
1597out:
1598 if (ret == 0) {
1599 *eb_ret = tmp;
1600 } else {
1601 free_extent_buffer(tmp);
1602 btrfs_release_path(p);
1603 }
1604
1605 return ret;
1606}
1607
1608/*
1609 * helper function for btrfs_search_slot. This does all of the checks
1610 * for node-level blocks and does any balancing required based on
1611 * the ins_len.
1612 *
1613 * If no extra work was required, zero is returned. If we had to
1614 * drop the path, -EAGAIN is returned and btrfs_search_slot must
1615 * start over
1616 */
1617static int
1618setup_nodes_for_search(struct btrfs_trans_handle *trans,
1619 struct btrfs_root *root, struct btrfs_path *p,
1620 struct extent_buffer *b, int level, int ins_len,
1621 int *write_lock_level)
1622{
1623 struct btrfs_fs_info *fs_info = root->fs_info;
1624 int ret = 0;
1625
1626 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1627 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
1628
1629 if (*write_lock_level < level + 1) {
1630 *write_lock_level = level + 1;
1631 btrfs_release_path(p);
1632 return -EAGAIN;
1633 }
1634
1635 reada_for_balance(p, level);
1636 ret = split_node(trans, root, p, level);
1637
1638 b = p->nodes[level];
1639 } else if (ins_len < 0 && btrfs_header_nritems(b) <
1640 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
1641
1642 if (*write_lock_level < level + 1) {
1643 *write_lock_level = level + 1;
1644 btrfs_release_path(p);
1645 return -EAGAIN;
1646 }
1647
1648 reada_for_balance(p, level);
1649 ret = balance_level(trans, root, p, level);
1650 if (ret)
1651 return ret;
1652
1653 b = p->nodes[level];
1654 if (!b) {
1655 btrfs_release_path(p);
1656 return -EAGAIN;
1657 }
1658 BUG_ON(btrfs_header_nritems(b) == 1);
1659 }
1660 return ret;
1661}
1662
1663int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
1664 u64 iobjectid, u64 ioff, u8 key_type,
1665 struct btrfs_key *found_key)
1666{
1667 int ret;
1668 struct btrfs_key key;
1669 struct extent_buffer *eb;
1670
1671 ASSERT(path);
1672 ASSERT(found_key);
1673
1674 key.type = key_type;
1675 key.objectid = iobjectid;
1676 key.offset = ioff;
1677
1678 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1679 if (ret < 0)
1680 return ret;
1681
1682 eb = path->nodes[0];
1683 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1684 ret = btrfs_next_leaf(fs_root, path);
1685 if (ret)
1686 return ret;
1687 eb = path->nodes[0];
1688 }
1689
1690 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1691 if (found_key->type != key.type ||
1692 found_key->objectid != key.objectid)
1693 return 1;
1694
1695 return 0;
1696}
1697
1698static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
1699 struct btrfs_path *p,
1700 int write_lock_level)
1701{
1702 struct extent_buffer *b;
1703 int root_lock = 0;
1704 int level = 0;
1705
1706 if (p->search_commit_root) {
1707 b = root->commit_root;
1708 atomic_inc(&b->refs);
1709 level = btrfs_header_level(b);
1710 /*
1711 * Ensure that all callers have set skip_locking when
1712 * p->search_commit_root = 1.
1713 */
1714 ASSERT(p->skip_locking == 1);
1715
1716 goto out;
1717 }
1718
1719 if (p->skip_locking) {
1720 b = btrfs_root_node(root);
1721 level = btrfs_header_level(b);
1722 goto out;
1723 }
1724
1725 /* We try very hard to do read locks on the root */
1726 root_lock = BTRFS_READ_LOCK;
1727
1728 /*
1729 * If the level is set to maximum, we can skip trying to get the read
1730 * lock.
1731 */
1732 if (write_lock_level < BTRFS_MAX_LEVEL) {
1733 /*
1734 * We don't know the level of the root node until we actually
1735 * have it read locked
1736 */
1737 if (p->nowait) {
1738 b = btrfs_try_read_lock_root_node(root);
1739 if (IS_ERR(b))
1740 return b;
1741 } else {
1742 b = btrfs_read_lock_root_node(root);
1743 }
1744 level = btrfs_header_level(b);
1745 if (level > write_lock_level)
1746 goto out;
1747
1748 /* Whoops, must trade for write lock */
1749 btrfs_tree_read_unlock(b);
1750 free_extent_buffer(b);
1751 }
1752
1753 b = btrfs_lock_root_node(root);
1754 root_lock = BTRFS_WRITE_LOCK;
1755
1756 /* The level might have changed, check again */
1757 level = btrfs_header_level(b);
1758
1759out:
1760 /*
1761 * The root may have failed to write out at some point, and thus is no
1762 * longer valid, return an error in this case.
1763 */
1764 if (!extent_buffer_uptodate(b)) {
1765 if (root_lock)
1766 btrfs_tree_unlock_rw(b, root_lock);
1767 free_extent_buffer(b);
1768 return ERR_PTR(-EIO);
1769 }
1770
1771 p->nodes[level] = b;
1772 if (!p->skip_locking)
1773 p->locks[level] = root_lock;
1774 /*
1775 * Callers are responsible for dropping b's references.
1776 */
1777 return b;
1778}
1779
1780/*
1781 * Replace the extent buffer at the lowest level of the path with a cloned
1782 * version. The purpose is to be able to use it safely, after releasing the
1783 * commit root semaphore, even if relocation is happening in parallel, the
1784 * transaction used for relocation is committed and the extent buffer is
1785 * reallocated in the next transaction.
1786 *
1787 * This is used in a context where the caller does not prevent transaction
1788 * commits from happening, either by holding a transaction handle or holding
1789 * some lock, while it's doing searches through a commit root.
1790 * At the moment it's only used for send operations.
1791 */
1792static int finish_need_commit_sem_search(struct btrfs_path *path)
1793{
1794 const int i = path->lowest_level;
1795 const int slot = path->slots[i];
1796 struct extent_buffer *lowest = path->nodes[i];
1797 struct extent_buffer *clone;
1798
1799 ASSERT(path->need_commit_sem);
1800
1801 if (!lowest)
1802 return 0;
1803
1804 lockdep_assert_held_read(&lowest->fs_info->commit_root_sem);
1805
1806 clone = btrfs_clone_extent_buffer(lowest);
1807 if (!clone)
1808 return -ENOMEM;
1809
1810 btrfs_release_path(path);
1811 path->nodes[i] = clone;
1812 path->slots[i] = slot;
1813
1814 return 0;
1815}
1816
1817static inline int search_for_key_slot(struct extent_buffer *eb,
1818 int search_low_slot,
1819 const struct btrfs_key *key,
1820 int prev_cmp,
1821 int *slot)
1822{
1823 /*
1824 * If a previous call to btrfs_bin_search() on a parent node returned an
1825 * exact match (prev_cmp == 0), we can safely assume the target key will
1826 * always be at slot 0 on lower levels, since each key pointer
1827 * (struct btrfs_key_ptr) refers to the lowest key accessible from the
1828 * subtree it points to. Thus we can skip searching lower levels.
1829 */
1830 if (prev_cmp == 0) {
1831 *slot = 0;
1832 return 0;
1833 }
1834
1835 return btrfs_bin_search(eb, search_low_slot, key, slot);
1836}
1837
1838static int search_leaf(struct btrfs_trans_handle *trans,
1839 struct btrfs_root *root,
1840 const struct btrfs_key *key,
1841 struct btrfs_path *path,
1842 int ins_len,
1843 int prev_cmp)
1844{
1845 struct extent_buffer *leaf = path->nodes[0];
1846 int leaf_free_space = -1;
1847 int search_low_slot = 0;
1848 int ret;
1849 bool do_bin_search = true;
1850
1851 /*
1852 * If we are doing an insertion, the leaf has enough free space and the
1853 * destination slot for the key is not slot 0, then we can unlock our
1854 * write lock on the parent, and any other upper nodes, before doing the
1855 * binary search on the leaf (with search_for_key_slot()), allowing other
1856 * tasks to lock the parent and any other upper nodes.
1857 */
1858 if (ins_len > 0) {
1859 /*
1860 * Cache the leaf free space, since we will need it later and it
1861 * will not change until then.
1862 */
1863 leaf_free_space = btrfs_leaf_free_space(leaf);
1864
1865 /*
1866 * !path->locks[1] means we have a single node tree, the leaf is
1867 * the root of the tree.
1868 */
1869 if (path->locks[1] && leaf_free_space >= ins_len) {
1870 struct btrfs_disk_key first_key;
1871
1872 ASSERT(btrfs_header_nritems(leaf) > 0);
1873 btrfs_item_key(leaf, &first_key, 0);
1874
1875 /*
1876 * Doing the extra comparison with the first key is cheap,
1877 * taking into account that the first key is very likely
1878 * already in a cache line because it immediately follows
1879 * the extent buffer's header and we have recently accessed
1880 * the header's level field.
1881 */
1882 ret = btrfs_comp_keys(&first_key, key);
1883 if (ret < 0) {
1884 /*
1885 * The first key is smaller than the key we want
1886 * to insert, so we are safe to unlock all upper
1887 * nodes and we have to do the binary search.
1888 *
1889 * We do use btrfs_unlock_up_safe() and not
1890 * unlock_up() because the later does not unlock
1891 * nodes with a slot of 0 - we can safely unlock
1892 * any node even if its slot is 0 since in this
1893 * case the key does not end up at slot 0 of the
1894 * leaf and there's no need to split the leaf.
1895 */
1896 btrfs_unlock_up_safe(path, 1);
1897 search_low_slot = 1;
1898 } else {
1899 /*
1900 * The first key is >= then the key we want to
1901 * insert, so we can skip the binary search as
1902 * the target key will be at slot 0.
1903 *
1904 * We can not unlock upper nodes when the key is
1905 * less than the first key, because we will need
1906 * to update the key at slot 0 of the parent node
1907 * and possibly of other upper nodes too.
1908 * If the key matches the first key, then we can
1909 * unlock all the upper nodes, using
1910 * btrfs_unlock_up_safe() instead of unlock_up()
1911 * as stated above.
1912 */
1913 if (ret == 0)
1914 btrfs_unlock_up_safe(path, 1);
1915 /*
1916 * ret is already 0 or 1, matching the result of
1917 * a btrfs_bin_search() call, so there is no need
1918 * to adjust it.
1919 */
1920 do_bin_search = false;
1921 path->slots[0] = 0;
1922 }
1923 }
1924 }
1925
1926 if (do_bin_search) {
1927 ret = search_for_key_slot(leaf, search_low_slot, key,
1928 prev_cmp, &path->slots[0]);
1929 if (ret < 0)
1930 return ret;
1931 }
1932
1933 if (ins_len > 0) {
1934 /*
1935 * Item key already exists. In this case, if we are allowed to
1936 * insert the item (for example, in dir_item case, item key
1937 * collision is allowed), it will be merged with the original
1938 * item. Only the item size grows, no new btrfs item will be
1939 * added. If search_for_extension is not set, ins_len already
1940 * accounts the size btrfs_item, deduct it here so leaf space
1941 * check will be correct.
1942 */
1943 if (ret == 0 && !path->search_for_extension) {
1944 ASSERT(ins_len >= sizeof(struct btrfs_item));
1945 ins_len -= sizeof(struct btrfs_item);
1946 }
1947
1948 ASSERT(leaf_free_space >= 0);
1949
1950 if (leaf_free_space < ins_len) {
1951 int err;
1952
1953 err = split_leaf(trans, root, key, path, ins_len,
1954 (ret == 0));
1955 ASSERT(err <= 0);
1956 if (WARN_ON(err > 0))
1957 err = -EUCLEAN;
1958 if (err)
1959 ret = err;
1960 }
1961 }
1962
1963 return ret;
1964}
1965
1966/*
1967 * Look for a key in a tree and perform necessary modifications to preserve
1968 * tree invariants.
1969 *
1970 * @trans: Handle of transaction, used when modifying the tree
1971 * @p: Holds all btree nodes along the search path
1972 * @root: The root node of the tree
1973 * @key: The key we are looking for
1974 * @ins_len: Indicates purpose of search:
1975 * >0 for inserts it's size of item inserted (*)
1976 * <0 for deletions
1977 * 0 for plain searches, not modifying the tree
1978 *
1979 * (*) If size of item inserted doesn't include
1980 * sizeof(struct btrfs_item), then p->search_for_extension must
1981 * be set.
1982 * @cow: boolean should CoW operations be performed. Must always be 1
1983 * when modifying the tree.
1984 *
1985 * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
1986 * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
1987 *
1988 * If @key is found, 0 is returned and you can find the item in the leaf level
1989 * of the path (level 0)
1990 *
1991 * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
1992 * points to the slot where it should be inserted
1993 *
1994 * If an error is encountered while searching the tree a negative error number
1995 * is returned
1996 */
1997int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1998 const struct btrfs_key *key, struct btrfs_path *p,
1999 int ins_len, int cow)
2000{
2001 struct btrfs_fs_info *fs_info = root->fs_info;
2002 struct extent_buffer *b;
2003 int slot;
2004 int ret;
2005 int err;
2006 int level;
2007 int lowest_unlock = 1;
2008 /* everything at write_lock_level or lower must be write locked */
2009 int write_lock_level = 0;
2010 u8 lowest_level = 0;
2011 int min_write_lock_level;
2012 int prev_cmp;
2013
2014 might_sleep();
2015
2016 lowest_level = p->lowest_level;
2017 WARN_ON(lowest_level && ins_len > 0);
2018 WARN_ON(p->nodes[0] != NULL);
2019 BUG_ON(!cow && ins_len);
2020
2021 /*
2022 * For now only allow nowait for read only operations. There's no
2023 * strict reason why we can't, we just only need it for reads so it's
2024 * only implemented for reads.
2025 */
2026 ASSERT(!p->nowait || !cow);
2027
2028 if (ins_len < 0) {
2029 lowest_unlock = 2;
2030
2031 /* when we are removing items, we might have to go up to level
2032 * two as we update tree pointers Make sure we keep write
2033 * for those levels as well
2034 */
2035 write_lock_level = 2;
2036 } else if (ins_len > 0) {
2037 /*
2038 * for inserting items, make sure we have a write lock on
2039 * level 1 so we can update keys
2040 */
2041 write_lock_level = 1;
2042 }
2043
2044 if (!cow)
2045 write_lock_level = -1;
2046
2047 if (cow && (p->keep_locks || p->lowest_level))
2048 write_lock_level = BTRFS_MAX_LEVEL;
2049
2050 min_write_lock_level = write_lock_level;
2051
2052 if (p->need_commit_sem) {
2053 ASSERT(p->search_commit_root);
2054 if (p->nowait) {
2055 if (!down_read_trylock(&fs_info->commit_root_sem))
2056 return -EAGAIN;
2057 } else {
2058 down_read(&fs_info->commit_root_sem);
2059 }
2060 }
2061
2062again:
2063 prev_cmp = -1;
2064 b = btrfs_search_slot_get_root(root, p, write_lock_level);
2065 if (IS_ERR(b)) {
2066 ret = PTR_ERR(b);
2067 goto done;
2068 }
2069
2070 while (b) {
2071 int dec = 0;
2072
2073 level = btrfs_header_level(b);
2074
2075 if (cow) {
2076 bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2077
2078 /*
2079 * if we don't really need to cow this block
2080 * then we don't want to set the path blocking,
2081 * so we test it here
2082 */
2083 if (!should_cow_block(trans, root, b))
2084 goto cow_done;
2085
2086 /*
2087 * must have write locks on this node and the
2088 * parent
2089 */
2090 if (level > write_lock_level ||
2091 (level + 1 > write_lock_level &&
2092 level + 1 < BTRFS_MAX_LEVEL &&
2093 p->nodes[level + 1])) {
2094 write_lock_level = level + 1;
2095 btrfs_release_path(p);
2096 goto again;
2097 }
2098
2099 if (last_level)
2100 err = btrfs_cow_block(trans, root, b, NULL, 0,
2101 &b,
2102 BTRFS_NESTING_COW);
2103 else
2104 err = btrfs_cow_block(trans, root, b,
2105 p->nodes[level + 1],
2106 p->slots[level + 1], &b,
2107 BTRFS_NESTING_COW);
2108 if (err) {
2109 ret = err;
2110 goto done;
2111 }
2112 }
2113cow_done:
2114 p->nodes[level] = b;
2115
2116 /*
2117 * we have a lock on b and as long as we aren't changing
2118 * the tree, there is no way to for the items in b to change.
2119 * It is safe to drop the lock on our parent before we
2120 * go through the expensive btree search on b.
2121 *
2122 * If we're inserting or deleting (ins_len != 0), then we might
2123 * be changing slot zero, which may require changing the parent.
2124 * So, we can't drop the lock until after we know which slot
2125 * we're operating on.
2126 */
2127 if (!ins_len && !p->keep_locks) {
2128 int u = level + 1;
2129
2130 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2131 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2132 p->locks[u] = 0;
2133 }
2134 }
2135
2136 if (level == 0) {
2137 if (ins_len > 0)
2138 ASSERT(write_lock_level >= 1);
2139
2140 ret = search_leaf(trans, root, key, p, ins_len, prev_cmp);
2141 if (!p->search_for_split)
2142 unlock_up(p, level, lowest_unlock,
2143 min_write_lock_level, NULL);
2144 goto done;
2145 }
2146
2147 ret = search_for_key_slot(b, 0, key, prev_cmp, &slot);
2148 if (ret < 0)
2149 goto done;
2150 prev_cmp = ret;
2151
2152 if (ret && slot > 0) {
2153 dec = 1;
2154 slot--;
2155 }
2156 p->slots[level] = slot;
2157 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
2158 &write_lock_level);
2159 if (err == -EAGAIN)
2160 goto again;
2161 if (err) {
2162 ret = err;
2163 goto done;
2164 }
2165 b = p->nodes[level];
2166 slot = p->slots[level];
2167
2168 /*
2169 * Slot 0 is special, if we change the key we have to update
2170 * the parent pointer which means we must have a write lock on
2171 * the parent
2172 */
2173 if (slot == 0 && ins_len && write_lock_level < level + 1) {
2174 write_lock_level = level + 1;
2175 btrfs_release_path(p);
2176 goto again;
2177 }
2178
2179 unlock_up(p, level, lowest_unlock, min_write_lock_level,
2180 &write_lock_level);
2181
2182 if (level == lowest_level) {
2183 if (dec)
2184 p->slots[level]++;
2185 goto done;
2186 }
2187
2188 err = read_block_for_search(root, p, &b, level, slot, key);
2189 if (err == -EAGAIN)
2190 goto again;
2191 if (err) {
2192 ret = err;
2193 goto done;
2194 }
2195
2196 if (!p->skip_locking) {
2197 level = btrfs_header_level(b);
2198
2199 btrfs_maybe_reset_lockdep_class(root, b);
2200
2201 if (level <= write_lock_level) {
2202 btrfs_tree_lock(b);
2203 p->locks[level] = BTRFS_WRITE_LOCK;
2204 } else {
2205 if (p->nowait) {
2206 if (!btrfs_try_tree_read_lock(b)) {
2207 free_extent_buffer(b);
2208 ret = -EAGAIN;
2209 goto done;
2210 }
2211 } else {
2212 btrfs_tree_read_lock(b);
2213 }
2214 p->locks[level] = BTRFS_READ_LOCK;
2215 }
2216 p->nodes[level] = b;
2217 }
2218 }
2219 ret = 1;
2220done:
2221 if (ret < 0 && !p->skip_release_on_error)
2222 btrfs_release_path(p);
2223
2224 if (p->need_commit_sem) {
2225 int ret2;
2226
2227 ret2 = finish_need_commit_sem_search(p);
2228 up_read(&fs_info->commit_root_sem);
2229 if (ret2)
2230 ret = ret2;
2231 }
2232
2233 return ret;
2234}
2235ALLOW_ERROR_INJECTION(btrfs_search_slot, ERRNO);
2236
2237/*
2238 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2239 * current state of the tree together with the operations recorded in the tree
2240 * modification log to search for the key in a previous version of this tree, as
2241 * denoted by the time_seq parameter.
2242 *
2243 * Naturally, there is no support for insert, delete or cow operations.
2244 *
2245 * The resulting path and return value will be set up as if we called
2246 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2247 */
2248int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2249 struct btrfs_path *p, u64 time_seq)
2250{
2251 struct btrfs_fs_info *fs_info = root->fs_info;
2252 struct extent_buffer *b;
2253 int slot;
2254 int ret;
2255 int err;
2256 int level;
2257 int lowest_unlock = 1;
2258 u8 lowest_level = 0;
2259
2260 lowest_level = p->lowest_level;
2261 WARN_ON(p->nodes[0] != NULL);
2262 ASSERT(!p->nowait);
2263
2264 if (p->search_commit_root) {
2265 BUG_ON(time_seq);
2266 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2267 }
2268
2269again:
2270 b = btrfs_get_old_root(root, time_seq);
2271 if (!b) {
2272 ret = -EIO;
2273 goto done;
2274 }
2275 level = btrfs_header_level(b);
2276 p->locks[level] = BTRFS_READ_LOCK;
2277
2278 while (b) {
2279 int dec = 0;
2280
2281 level = btrfs_header_level(b);
2282 p->nodes[level] = b;
2283
2284 /*
2285 * we have a lock on b and as long as we aren't changing
2286 * the tree, there is no way to for the items in b to change.
2287 * It is safe to drop the lock on our parent before we
2288 * go through the expensive btree search on b.
2289 */
2290 btrfs_unlock_up_safe(p, level + 1);
2291
2292 ret = btrfs_bin_search(b, 0, key, &slot);
2293 if (ret < 0)
2294 goto done;
2295
2296 if (level == 0) {
2297 p->slots[level] = slot;
2298 unlock_up(p, level, lowest_unlock, 0, NULL);
2299 goto done;
2300 }
2301
2302 if (ret && slot > 0) {
2303 dec = 1;
2304 slot--;
2305 }
2306 p->slots[level] = slot;
2307 unlock_up(p, level, lowest_unlock, 0, NULL);
2308
2309 if (level == lowest_level) {
2310 if (dec)
2311 p->slots[level]++;
2312 goto done;
2313 }
2314
2315 err = read_block_for_search(root, p, &b, level, slot, key);
2316 if (err == -EAGAIN)
2317 goto again;
2318 if (err) {
2319 ret = err;
2320 goto done;
2321 }
2322
2323 level = btrfs_header_level(b);
2324 btrfs_tree_read_lock(b);
2325 b = btrfs_tree_mod_log_rewind(fs_info, p, b, time_seq);
2326 if (!b) {
2327 ret = -ENOMEM;
2328 goto done;
2329 }
2330 p->locks[level] = BTRFS_READ_LOCK;
2331 p->nodes[level] = b;
2332 }
2333 ret = 1;
2334done:
2335 if (ret < 0)
2336 btrfs_release_path(p);
2337
2338 return ret;
2339}
2340
2341/*
2342 * Search the tree again to find a leaf with smaller keys.
2343 * Returns 0 if it found something.
2344 * Returns 1 if there are no smaller keys.
2345 * Returns < 0 on error.
2346 *
2347 * This may release the path, and so you may lose any locks held at the
2348 * time you call it.
2349 */
2350static int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2351{
2352 struct btrfs_key key;
2353 struct btrfs_key orig_key;
2354 struct btrfs_disk_key found_key;
2355 int ret;
2356
2357 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
2358 orig_key = key;
2359
2360 if (key.offset > 0) {
2361 key.offset--;
2362 } else if (key.type > 0) {
2363 key.type--;
2364 key.offset = (u64)-1;
2365 } else if (key.objectid > 0) {
2366 key.objectid--;
2367 key.type = (u8)-1;
2368 key.offset = (u64)-1;
2369 } else {
2370 return 1;
2371 }
2372
2373 btrfs_release_path(path);
2374 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2375 if (ret <= 0)
2376 return ret;
2377
2378 /*
2379 * Previous key not found. Even if we were at slot 0 of the leaf we had
2380 * before releasing the path and calling btrfs_search_slot(), we now may
2381 * be in a slot pointing to the same original key - this can happen if
2382 * after we released the path, one of more items were moved from a
2383 * sibling leaf into the front of the leaf we had due to an insertion
2384 * (see push_leaf_right()).
2385 * If we hit this case and our slot is > 0 and just decrement the slot
2386 * so that the caller does not process the same key again, which may or
2387 * may not break the caller, depending on its logic.
2388 */
2389 if (path->slots[0] < btrfs_header_nritems(path->nodes[0])) {
2390 btrfs_item_key(path->nodes[0], &found_key, path->slots[0]);
2391 ret = btrfs_comp_keys(&found_key, &orig_key);
2392 if (ret == 0) {
2393 if (path->slots[0] > 0) {
2394 path->slots[0]--;
2395 return 0;
2396 }
2397 /*
2398 * At slot 0, same key as before, it means orig_key is
2399 * the lowest, leftmost, key in the tree. We're done.
2400 */
2401 return 1;
2402 }
2403 }
2404
2405 btrfs_item_key(path->nodes[0], &found_key, 0);
2406 ret = btrfs_comp_keys(&found_key, &key);
2407 /*
2408 * We might have had an item with the previous key in the tree right
2409 * before we released our path. And after we released our path, that
2410 * item might have been pushed to the first slot (0) of the leaf we
2411 * were holding due to a tree balance. Alternatively, an item with the
2412 * previous key can exist as the only element of a leaf (big fat item).
2413 * Therefore account for these 2 cases, so that our callers (like
2414 * btrfs_previous_item) don't miss an existing item with a key matching
2415 * the previous key we computed above.
2416 */
2417 if (ret <= 0)
2418 return 0;
2419 return 1;
2420}
2421
2422/*
2423 * helper to use instead of search slot if no exact match is needed but
2424 * instead the next or previous item should be returned.
2425 * When find_higher is true, the next higher item is returned, the next lower
2426 * otherwise.
2427 * When return_any and find_higher are both true, and no higher item is found,
2428 * return the next lower instead.
2429 * When return_any is true and find_higher is false, and no lower item is found,
2430 * return the next higher instead.
2431 * It returns 0 if any item is found, 1 if none is found (tree empty), and
2432 * < 0 on error
2433 */
2434int btrfs_search_slot_for_read(struct btrfs_root *root,
2435 const struct btrfs_key *key,
2436 struct btrfs_path *p, int find_higher,
2437 int return_any)
2438{
2439 int ret;
2440 struct extent_buffer *leaf;
2441
2442again:
2443 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
2444 if (ret <= 0)
2445 return ret;
2446 /*
2447 * a return value of 1 means the path is at the position where the
2448 * item should be inserted. Normally this is the next bigger item,
2449 * but in case the previous item is the last in a leaf, path points
2450 * to the first free slot in the previous leaf, i.e. at an invalid
2451 * item.
2452 */
2453 leaf = p->nodes[0];
2454
2455 if (find_higher) {
2456 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
2457 ret = btrfs_next_leaf(root, p);
2458 if (ret <= 0)
2459 return ret;
2460 if (!return_any)
2461 return 1;
2462 /*
2463 * no higher item found, return the next
2464 * lower instead
2465 */
2466 return_any = 0;
2467 find_higher = 0;
2468 btrfs_release_path(p);
2469 goto again;
2470 }
2471 } else {
2472 if (p->slots[0] == 0) {
2473 ret = btrfs_prev_leaf(root, p);
2474 if (ret < 0)
2475 return ret;
2476 if (!ret) {
2477 leaf = p->nodes[0];
2478 if (p->slots[0] == btrfs_header_nritems(leaf))
2479 p->slots[0]--;
2480 return 0;
2481 }
2482 if (!return_any)
2483 return 1;
2484 /*
2485 * no lower item found, return the next
2486 * higher instead
2487 */
2488 return_any = 0;
2489 find_higher = 1;
2490 btrfs_release_path(p);
2491 goto again;
2492 } else {
2493 --p->slots[0];
2494 }
2495 }
2496 return 0;
2497}
2498
2499/*
2500 * Execute search and call btrfs_previous_item to traverse backwards if the item
2501 * was not found.
2502 *
2503 * Return 0 if found, 1 if not found and < 0 if error.
2504 */
2505int btrfs_search_backwards(struct btrfs_root *root, struct btrfs_key *key,
2506 struct btrfs_path *path)
2507{
2508 int ret;
2509
2510 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
2511 if (ret > 0)
2512 ret = btrfs_previous_item(root, path, key->objectid, key->type);
2513
2514 if (ret == 0)
2515 btrfs_item_key_to_cpu(path->nodes[0], key, path->slots[0]);
2516
2517 return ret;
2518}
2519
2520/*
2521 * Search for a valid slot for the given path.
2522 *
2523 * @root: The root node of the tree.
2524 * @key: Will contain a valid item if found.
2525 * @path: The starting point to validate the slot.
2526 *
2527 * Return: 0 if the item is valid
2528 * 1 if not found
2529 * <0 if error.
2530 */
2531int btrfs_get_next_valid_item(struct btrfs_root *root, struct btrfs_key *key,
2532 struct btrfs_path *path)
2533{
2534 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2535 int ret;
2536
2537 ret = btrfs_next_leaf(root, path);
2538 if (ret)
2539 return ret;
2540 }
2541
2542 btrfs_item_key_to_cpu(path->nodes[0], key, path->slots[0]);
2543 return 0;
2544}
2545
2546/*
2547 * adjust the pointers going up the tree, starting at level
2548 * making sure the right key of each node is points to 'key'.
2549 * This is used after shifting pointers to the left, so it stops
2550 * fixing up pointers when a given leaf/node is not in slot 0 of the
2551 * higher levels
2552 *
2553 */
2554static void fixup_low_keys(struct btrfs_trans_handle *trans,
2555 struct btrfs_path *path,
2556 struct btrfs_disk_key *key, int level)
2557{
2558 int i;
2559 struct extent_buffer *t;
2560 int ret;
2561
2562 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2563 int tslot = path->slots[i];
2564
2565 if (!path->nodes[i])
2566 break;
2567 t = path->nodes[i];
2568 ret = btrfs_tree_mod_log_insert_key(t, tslot,
2569 BTRFS_MOD_LOG_KEY_REPLACE);
2570 BUG_ON(ret < 0);
2571 btrfs_set_node_key(t, key, tslot);
2572 btrfs_mark_buffer_dirty(trans, path->nodes[i]);
2573 if (tslot != 0)
2574 break;
2575 }
2576}
2577
2578/*
2579 * update item key.
2580 *
2581 * This function isn't completely safe. It's the caller's responsibility
2582 * that the new key won't break the order
2583 */
2584void btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
2585 struct btrfs_path *path,
2586 const struct btrfs_key *new_key)
2587{
2588 struct btrfs_fs_info *fs_info = trans->fs_info;
2589 struct btrfs_disk_key disk_key;
2590 struct extent_buffer *eb;
2591 int slot;
2592
2593 eb = path->nodes[0];
2594 slot = path->slots[0];
2595 if (slot > 0) {
2596 btrfs_item_key(eb, &disk_key, slot - 1);
2597 if (unlikely(btrfs_comp_keys(&disk_key, new_key) >= 0)) {
2598 btrfs_print_leaf(eb);
2599 btrfs_crit(fs_info,
2600 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2601 slot, btrfs_disk_key_objectid(&disk_key),
2602 btrfs_disk_key_type(&disk_key),
2603 btrfs_disk_key_offset(&disk_key),
2604 new_key->objectid, new_key->type,
2605 new_key->offset);
2606 BUG();
2607 }
2608 }
2609 if (slot < btrfs_header_nritems(eb) - 1) {
2610 btrfs_item_key(eb, &disk_key, slot + 1);
2611 if (unlikely(btrfs_comp_keys(&disk_key, new_key) <= 0)) {
2612 btrfs_print_leaf(eb);
2613 btrfs_crit(fs_info,
2614 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2615 slot, btrfs_disk_key_objectid(&disk_key),
2616 btrfs_disk_key_type(&disk_key),
2617 btrfs_disk_key_offset(&disk_key),
2618 new_key->objectid, new_key->type,
2619 new_key->offset);
2620 BUG();
2621 }
2622 }
2623
2624 btrfs_cpu_key_to_disk(&disk_key, new_key);
2625 btrfs_set_item_key(eb, &disk_key, slot);
2626 btrfs_mark_buffer_dirty(trans, eb);
2627 if (slot == 0)
2628 fixup_low_keys(trans, path, &disk_key, 1);
2629}
2630
2631/*
2632 * Check key order of two sibling extent buffers.
2633 *
2634 * Return true if something is wrong.
2635 * Return false if everything is fine.
2636 *
2637 * Tree-checker only works inside one tree block, thus the following
2638 * corruption can not be detected by tree-checker:
2639 *
2640 * Leaf @left | Leaf @right
2641 * --------------------------------------------------------------
2642 * | 1 | 2 | 3 | 4 | 5 | f6 | | 7 | 8 |
2643 *
2644 * Key f6 in leaf @left itself is valid, but not valid when the next
2645 * key in leaf @right is 7.
2646 * This can only be checked at tree block merge time.
2647 * And since tree checker has ensured all key order in each tree block
2648 * is correct, we only need to bother the last key of @left and the first
2649 * key of @right.
2650 */
2651static bool check_sibling_keys(struct extent_buffer *left,
2652 struct extent_buffer *right)
2653{
2654 struct btrfs_key left_last;
2655 struct btrfs_key right_first;
2656 int level = btrfs_header_level(left);
2657 int nr_left = btrfs_header_nritems(left);
2658 int nr_right = btrfs_header_nritems(right);
2659
2660 /* No key to check in one of the tree blocks */
2661 if (!nr_left || !nr_right)
2662 return false;
2663
2664 if (level) {
2665 btrfs_node_key_to_cpu(left, &left_last, nr_left - 1);
2666 btrfs_node_key_to_cpu(right, &right_first, 0);
2667 } else {
2668 btrfs_item_key_to_cpu(left, &left_last, nr_left - 1);
2669 btrfs_item_key_to_cpu(right, &right_first, 0);
2670 }
2671
2672 if (unlikely(btrfs_comp_cpu_keys(&left_last, &right_first) >= 0)) {
2673 btrfs_crit(left->fs_info, "left extent buffer:");
2674 btrfs_print_tree(left, false);
2675 btrfs_crit(left->fs_info, "right extent buffer:");
2676 btrfs_print_tree(right, false);
2677 btrfs_crit(left->fs_info,
2678"bad key order, sibling blocks, left last (%llu %u %llu) right first (%llu %u %llu)",
2679 left_last.objectid, left_last.type,
2680 left_last.offset, right_first.objectid,
2681 right_first.type, right_first.offset);
2682 return true;
2683 }
2684 return false;
2685}
2686
2687/*
2688 * try to push data from one node into the next node left in the
2689 * tree.
2690 *
2691 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2692 * error, and > 0 if there was no room in the left hand block.
2693 */
2694static int push_node_left(struct btrfs_trans_handle *trans,
2695 struct extent_buffer *dst,
2696 struct extent_buffer *src, int empty)
2697{
2698 struct btrfs_fs_info *fs_info = trans->fs_info;
2699 int push_items = 0;
2700 int src_nritems;
2701 int dst_nritems;
2702 int ret = 0;
2703
2704 src_nritems = btrfs_header_nritems(src);
2705 dst_nritems = btrfs_header_nritems(dst);
2706 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2707 WARN_ON(btrfs_header_generation(src) != trans->transid);
2708 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2709
2710 if (!empty && src_nritems <= 8)
2711 return 1;
2712
2713 if (push_items <= 0)
2714 return 1;
2715
2716 if (empty) {
2717 push_items = min(src_nritems, push_items);
2718 if (push_items < src_nritems) {
2719 /* leave at least 8 pointers in the node if
2720 * we aren't going to empty it
2721 */
2722 if (src_nritems - push_items < 8) {
2723 if (push_items <= 8)
2724 return 1;
2725 push_items -= 8;
2726 }
2727 }
2728 } else
2729 push_items = min(src_nritems - 8, push_items);
2730
2731 /* dst is the left eb, src is the middle eb */
2732 if (check_sibling_keys(dst, src)) {
2733 ret = -EUCLEAN;
2734 btrfs_abort_transaction(trans, ret);
2735 return ret;
2736 }
2737 ret = btrfs_tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
2738 if (ret) {
2739 btrfs_abort_transaction(trans, ret);
2740 return ret;
2741 }
2742 copy_extent_buffer(dst, src,
2743 btrfs_node_key_ptr_offset(dst, dst_nritems),
2744 btrfs_node_key_ptr_offset(src, 0),
2745 push_items * sizeof(struct btrfs_key_ptr));
2746
2747 if (push_items < src_nritems) {
2748 /*
2749 * btrfs_tree_mod_log_eb_copy handles logging the move, so we
2750 * don't need to do an explicit tree mod log operation for it.
2751 */
2752 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(src, 0),
2753 btrfs_node_key_ptr_offset(src, push_items),
2754 (src_nritems - push_items) *
2755 sizeof(struct btrfs_key_ptr));
2756 }
2757 btrfs_set_header_nritems(src, src_nritems - push_items);
2758 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2759 btrfs_mark_buffer_dirty(trans, src);
2760 btrfs_mark_buffer_dirty(trans, dst);
2761
2762 return ret;
2763}
2764
2765/*
2766 * try to push data from one node into the next node right in the
2767 * tree.
2768 *
2769 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2770 * error, and > 0 if there was no room in the right hand block.
2771 *
2772 * this will only push up to 1/2 the contents of the left node over
2773 */
2774static int balance_node_right(struct btrfs_trans_handle *trans,
2775 struct extent_buffer *dst,
2776 struct extent_buffer *src)
2777{
2778 struct btrfs_fs_info *fs_info = trans->fs_info;
2779 int push_items = 0;
2780 int max_push;
2781 int src_nritems;
2782 int dst_nritems;
2783 int ret = 0;
2784
2785 WARN_ON(btrfs_header_generation(src) != trans->transid);
2786 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2787
2788 src_nritems = btrfs_header_nritems(src);
2789 dst_nritems = btrfs_header_nritems(dst);
2790 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2791 if (push_items <= 0)
2792 return 1;
2793
2794 if (src_nritems < 4)
2795 return 1;
2796
2797 max_push = src_nritems / 2 + 1;
2798 /* don't try to empty the node */
2799 if (max_push >= src_nritems)
2800 return 1;
2801
2802 if (max_push < push_items)
2803 push_items = max_push;
2804
2805 /* dst is the right eb, src is the middle eb */
2806 if (check_sibling_keys(src, dst)) {
2807 ret = -EUCLEAN;
2808 btrfs_abort_transaction(trans, ret);
2809 return ret;
2810 }
2811
2812 /*
2813 * btrfs_tree_mod_log_eb_copy handles logging the move, so we don't
2814 * need to do an explicit tree mod log operation for it.
2815 */
2816 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(dst, push_items),
2817 btrfs_node_key_ptr_offset(dst, 0),
2818 (dst_nritems) *
2819 sizeof(struct btrfs_key_ptr));
2820
2821 ret = btrfs_tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
2822 push_items);
2823 if (ret) {
2824 btrfs_abort_transaction(trans, ret);
2825 return ret;
2826 }
2827 copy_extent_buffer(dst, src,
2828 btrfs_node_key_ptr_offset(dst, 0),
2829 btrfs_node_key_ptr_offset(src, src_nritems - push_items),
2830 push_items * sizeof(struct btrfs_key_ptr));
2831
2832 btrfs_set_header_nritems(src, src_nritems - push_items);
2833 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2834
2835 btrfs_mark_buffer_dirty(trans, src);
2836 btrfs_mark_buffer_dirty(trans, dst);
2837
2838 return ret;
2839}
2840
2841/*
2842 * helper function to insert a new root level in the tree.
2843 * A new node is allocated, and a single item is inserted to
2844 * point to the existing root
2845 *
2846 * returns zero on success or < 0 on failure.
2847 */
2848static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2849 struct btrfs_root *root,
2850 struct btrfs_path *path, int level)
2851{
2852 u64 lower_gen;
2853 struct extent_buffer *lower;
2854 struct extent_buffer *c;
2855 struct extent_buffer *old;
2856 struct btrfs_disk_key lower_key;
2857 int ret;
2858
2859 BUG_ON(path->nodes[level]);
2860 BUG_ON(path->nodes[level-1] != root->node);
2861
2862 lower = path->nodes[level-1];
2863 if (level == 1)
2864 btrfs_item_key(lower, &lower_key, 0);
2865 else
2866 btrfs_node_key(lower, &lower_key, 0);
2867
2868 c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
2869 &lower_key, level, root->node->start, 0,
2870 0, BTRFS_NESTING_NEW_ROOT);
2871 if (IS_ERR(c))
2872 return PTR_ERR(c);
2873
2874 root_add_used_bytes(root);
2875
2876 btrfs_set_header_nritems(c, 1);
2877 btrfs_set_node_key(c, &lower_key, 0);
2878 btrfs_set_node_blockptr(c, 0, lower->start);
2879 lower_gen = btrfs_header_generation(lower);
2880 WARN_ON(lower_gen != trans->transid);
2881
2882 btrfs_set_node_ptr_generation(c, 0, lower_gen);
2883
2884 btrfs_mark_buffer_dirty(trans, c);
2885
2886 old = root->node;
2887 ret = btrfs_tree_mod_log_insert_root(root->node, c, false);
2888 if (ret < 0) {
2889 btrfs_free_tree_block(trans, btrfs_root_id(root), c, 0, 1);
2890 btrfs_tree_unlock(c);
2891 free_extent_buffer(c);
2892 return ret;
2893 }
2894 rcu_assign_pointer(root->node, c);
2895
2896 /* the super has an extra ref to root->node */
2897 free_extent_buffer(old);
2898
2899 add_root_to_dirty_list(root);
2900 atomic_inc(&c->refs);
2901 path->nodes[level] = c;
2902 path->locks[level] = BTRFS_WRITE_LOCK;
2903 path->slots[level] = 0;
2904 return 0;
2905}
2906
2907/*
2908 * worker function to insert a single pointer in a node.
2909 * the node should have enough room for the pointer already
2910 *
2911 * slot and level indicate where you want the key to go, and
2912 * blocknr is the block the key points to.
2913 */
2914static int insert_ptr(struct btrfs_trans_handle *trans,
2915 struct btrfs_path *path,
2916 struct btrfs_disk_key *key, u64 bytenr,
2917 int slot, int level)
2918{
2919 struct extent_buffer *lower;
2920 int nritems;
2921 int ret;
2922
2923 BUG_ON(!path->nodes[level]);
2924 btrfs_assert_tree_write_locked(path->nodes[level]);
2925 lower = path->nodes[level];
2926 nritems = btrfs_header_nritems(lower);
2927 BUG_ON(slot > nritems);
2928 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
2929 if (slot != nritems) {
2930 if (level) {
2931 ret = btrfs_tree_mod_log_insert_move(lower, slot + 1,
2932 slot, nritems - slot);
2933 if (ret < 0) {
2934 btrfs_abort_transaction(trans, ret);
2935 return ret;
2936 }
2937 }
2938 memmove_extent_buffer(lower,
2939 btrfs_node_key_ptr_offset(lower, slot + 1),
2940 btrfs_node_key_ptr_offset(lower, slot),
2941 (nritems - slot) * sizeof(struct btrfs_key_ptr));
2942 }
2943 if (level) {
2944 ret = btrfs_tree_mod_log_insert_key(lower, slot,
2945 BTRFS_MOD_LOG_KEY_ADD);
2946 if (ret < 0) {
2947 btrfs_abort_transaction(trans, ret);
2948 return ret;
2949 }
2950 }
2951 btrfs_set_node_key(lower, key, slot);
2952 btrfs_set_node_blockptr(lower, slot, bytenr);
2953 WARN_ON(trans->transid == 0);
2954 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2955 btrfs_set_header_nritems(lower, nritems + 1);
2956 btrfs_mark_buffer_dirty(trans, lower);
2957
2958 return 0;
2959}
2960
2961/*
2962 * split the node at the specified level in path in two.
2963 * The path is corrected to point to the appropriate node after the split
2964 *
2965 * Before splitting this tries to make some room in the node by pushing
2966 * left and right, if either one works, it returns right away.
2967 *
2968 * returns 0 on success and < 0 on failure
2969 */
2970static noinline int split_node(struct btrfs_trans_handle *trans,
2971 struct btrfs_root *root,
2972 struct btrfs_path *path, int level)
2973{
2974 struct btrfs_fs_info *fs_info = root->fs_info;
2975 struct extent_buffer *c;
2976 struct extent_buffer *split;
2977 struct btrfs_disk_key disk_key;
2978 int mid;
2979 int ret;
2980 u32 c_nritems;
2981
2982 c = path->nodes[level];
2983 WARN_ON(btrfs_header_generation(c) != trans->transid);
2984 if (c == root->node) {
2985 /*
2986 * trying to split the root, lets make a new one
2987 *
2988 * tree mod log: We don't log_removal old root in
2989 * insert_new_root, because that root buffer will be kept as a
2990 * normal node. We are going to log removal of half of the
2991 * elements below with btrfs_tree_mod_log_eb_copy(). We're
2992 * holding a tree lock on the buffer, which is why we cannot
2993 * race with other tree_mod_log users.
2994 */
2995 ret = insert_new_root(trans, root, path, level + 1);
2996 if (ret)
2997 return ret;
2998 } else {
2999 ret = push_nodes_for_insert(trans, root, path, level);
3000 c = path->nodes[level];
3001 if (!ret && btrfs_header_nritems(c) <
3002 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3003 return 0;
3004 if (ret < 0)
3005 return ret;
3006 }
3007
3008 c_nritems = btrfs_header_nritems(c);
3009 mid = (c_nritems + 1) / 2;
3010 btrfs_node_key(c, &disk_key, mid);
3011
3012 split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3013 &disk_key, level, c->start, 0,
3014 0, BTRFS_NESTING_SPLIT);
3015 if (IS_ERR(split))
3016 return PTR_ERR(split);
3017
3018 root_add_used_bytes(root);
3019 ASSERT(btrfs_header_level(c) == level);
3020
3021 ret = btrfs_tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
3022 if (ret) {
3023 btrfs_tree_unlock(split);
3024 free_extent_buffer(split);
3025 btrfs_abort_transaction(trans, ret);
3026 return ret;
3027 }
3028 copy_extent_buffer(split, c,
3029 btrfs_node_key_ptr_offset(split, 0),
3030 btrfs_node_key_ptr_offset(c, mid),
3031 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3032 btrfs_set_header_nritems(split, c_nritems - mid);
3033 btrfs_set_header_nritems(c, mid);
3034
3035 btrfs_mark_buffer_dirty(trans, c);
3036 btrfs_mark_buffer_dirty(trans, split);
3037
3038 ret = insert_ptr(trans, path, &disk_key, split->start,
3039 path->slots[level + 1] + 1, level + 1);
3040 if (ret < 0) {
3041 btrfs_tree_unlock(split);
3042 free_extent_buffer(split);
3043 return ret;
3044 }
3045
3046 if (path->slots[level] >= mid) {
3047 path->slots[level] -= mid;
3048 btrfs_tree_unlock(c);
3049 free_extent_buffer(c);
3050 path->nodes[level] = split;
3051 path->slots[level + 1] += 1;
3052 } else {
3053 btrfs_tree_unlock(split);
3054 free_extent_buffer(split);
3055 }
3056 return 0;
3057}
3058
3059/*
3060 * how many bytes are required to store the items in a leaf. start
3061 * and nr indicate which items in the leaf to check. This totals up the
3062 * space used both by the item structs and the item data
3063 */
3064static int leaf_space_used(const struct extent_buffer *l, int start, int nr)
3065{
3066 int data_len;
3067 int nritems = btrfs_header_nritems(l);
3068 int end = min(nritems, start + nr) - 1;
3069
3070 if (!nr)
3071 return 0;
3072 data_len = btrfs_item_offset(l, start) + btrfs_item_size(l, start);
3073 data_len = data_len - btrfs_item_offset(l, end);
3074 data_len += sizeof(struct btrfs_item) * nr;
3075 WARN_ON(data_len < 0);
3076 return data_len;
3077}
3078
3079/*
3080 * The space between the end of the leaf items and
3081 * the start of the leaf data. IOW, how much room
3082 * the leaf has left for both items and data
3083 */
3084int btrfs_leaf_free_space(const struct extent_buffer *leaf)
3085{
3086 struct btrfs_fs_info *fs_info = leaf->fs_info;
3087 int nritems = btrfs_header_nritems(leaf);
3088 int ret;
3089
3090 ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3091 if (ret < 0) {
3092 btrfs_crit(fs_info,
3093 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3094 ret,
3095 (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3096 leaf_space_used(leaf, 0, nritems), nritems);
3097 }
3098 return ret;
3099}
3100
3101/*
3102 * min slot controls the lowest index we're willing to push to the
3103 * right. We'll push up to and including min_slot, but no lower
3104 */
3105static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
3106 struct btrfs_path *path,
3107 int data_size, int empty,
3108 struct extent_buffer *right,
3109 int free_space, u32 left_nritems,
3110 u32 min_slot)
3111{
3112 struct btrfs_fs_info *fs_info = right->fs_info;
3113 struct extent_buffer *left = path->nodes[0];
3114 struct extent_buffer *upper = path->nodes[1];
3115 struct btrfs_map_token token;
3116 struct btrfs_disk_key disk_key;
3117 int slot;
3118 u32 i;
3119 int push_space = 0;
3120 int push_items = 0;
3121 u32 nr;
3122 u32 right_nritems;
3123 u32 data_end;
3124 u32 this_item_size;
3125
3126 if (empty)
3127 nr = 0;
3128 else
3129 nr = max_t(u32, 1, min_slot);
3130
3131 if (path->slots[0] >= left_nritems)
3132 push_space += data_size;
3133
3134 slot = path->slots[1];
3135 i = left_nritems - 1;
3136 while (i >= nr) {
3137 if (!empty && push_items > 0) {
3138 if (path->slots[0] > i)
3139 break;
3140 if (path->slots[0] == i) {
3141 int space = btrfs_leaf_free_space(left);
3142
3143 if (space + push_space * 2 > free_space)
3144 break;
3145 }
3146 }
3147
3148 if (path->slots[0] == i)
3149 push_space += data_size;
3150
3151 this_item_size = btrfs_item_size(left, i);
3152 if (this_item_size + sizeof(struct btrfs_item) +
3153 push_space > free_space)
3154 break;
3155
3156 push_items++;
3157 push_space += this_item_size + sizeof(struct btrfs_item);
3158 if (i == 0)
3159 break;
3160 i--;
3161 }
3162
3163 if (push_items == 0)
3164 goto out_unlock;
3165
3166 WARN_ON(!empty && push_items == left_nritems);
3167
3168 /* push left to right */
3169 right_nritems = btrfs_header_nritems(right);
3170
3171 push_space = btrfs_item_data_end(left, left_nritems - push_items);
3172 push_space -= leaf_data_end(left);
3173
3174 /* make room in the right data area */
3175 data_end = leaf_data_end(right);
3176 memmove_leaf_data(right, data_end - push_space, data_end,
3177 BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3178
3179 /* copy from the left data area */
3180 copy_leaf_data(right, left, BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3181 leaf_data_end(left), push_space);
3182
3183 memmove_leaf_items(right, push_items, 0, right_nritems);
3184
3185 /* copy the items from left to right */
3186 copy_leaf_items(right, left, 0, left_nritems - push_items, push_items);
3187
3188 /* update the item pointers */
3189 btrfs_init_map_token(&token, right);
3190 right_nritems += push_items;
3191 btrfs_set_header_nritems(right, right_nritems);
3192 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3193 for (i = 0; i < right_nritems; i++) {
3194 push_space -= btrfs_token_item_size(&token, i);
3195 btrfs_set_token_item_offset(&token, i, push_space);
3196 }
3197
3198 left_nritems -= push_items;
3199 btrfs_set_header_nritems(left, left_nritems);
3200
3201 if (left_nritems)
3202 btrfs_mark_buffer_dirty(trans, left);
3203 else
3204 btrfs_clear_buffer_dirty(trans, left);
3205
3206 btrfs_mark_buffer_dirty(trans, right);
3207
3208 btrfs_item_key(right, &disk_key, 0);
3209 btrfs_set_node_key(upper, &disk_key, slot + 1);
3210 btrfs_mark_buffer_dirty(trans, upper);
3211
3212 /* then fixup the leaf pointer in the path */
3213 if (path->slots[0] >= left_nritems) {
3214 path->slots[0] -= left_nritems;
3215 if (btrfs_header_nritems(path->nodes[0]) == 0)
3216 btrfs_clear_buffer_dirty(trans, path->nodes[0]);
3217 btrfs_tree_unlock(path->nodes[0]);
3218 free_extent_buffer(path->nodes[0]);
3219 path->nodes[0] = right;
3220 path->slots[1] += 1;
3221 } else {
3222 btrfs_tree_unlock(right);
3223 free_extent_buffer(right);
3224 }
3225 return 0;
3226
3227out_unlock:
3228 btrfs_tree_unlock(right);
3229 free_extent_buffer(right);
3230 return 1;
3231}
3232
3233/*
3234 * push some data in the path leaf to the right, trying to free up at
3235 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3236 *
3237 * returns 1 if the push failed because the other node didn't have enough
3238 * room, 0 if everything worked out and < 0 if there were major errors.
3239 *
3240 * this will push starting from min_slot to the end of the leaf. It won't
3241 * push any slot lower than min_slot
3242 */
3243static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3244 *root, struct btrfs_path *path,
3245 int min_data_size, int data_size,
3246 int empty, u32 min_slot)
3247{
3248 struct extent_buffer *left = path->nodes[0];
3249 struct extent_buffer *right;
3250 struct extent_buffer *upper;
3251 int slot;
3252 int free_space;
3253 u32 left_nritems;
3254 int ret;
3255
3256 if (!path->nodes[1])
3257 return 1;
3258
3259 slot = path->slots[1];
3260 upper = path->nodes[1];
3261 if (slot >= btrfs_header_nritems(upper) - 1)
3262 return 1;
3263
3264 btrfs_assert_tree_write_locked(path->nodes[1]);
3265
3266 right = btrfs_read_node_slot(upper, slot + 1);
3267 if (IS_ERR(right))
3268 return PTR_ERR(right);
3269
3270 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
3271
3272 free_space = btrfs_leaf_free_space(right);
3273 if (free_space < data_size)
3274 goto out_unlock;
3275
3276 ret = btrfs_cow_block(trans, root, right, upper,
3277 slot + 1, &right, BTRFS_NESTING_RIGHT_COW);
3278 if (ret)
3279 goto out_unlock;
3280
3281 left_nritems = btrfs_header_nritems(left);
3282 if (left_nritems == 0)
3283 goto out_unlock;
3284
3285 if (check_sibling_keys(left, right)) {
3286 ret = -EUCLEAN;
3287 btrfs_abort_transaction(trans, ret);
3288 btrfs_tree_unlock(right);
3289 free_extent_buffer(right);
3290 return ret;
3291 }
3292 if (path->slots[0] == left_nritems && !empty) {
3293 /* Key greater than all keys in the leaf, right neighbor has
3294 * enough room for it and we're not emptying our leaf to delete
3295 * it, therefore use right neighbor to insert the new item and
3296 * no need to touch/dirty our left leaf. */
3297 btrfs_tree_unlock(left);
3298 free_extent_buffer(left);
3299 path->nodes[0] = right;
3300 path->slots[0] = 0;
3301 path->slots[1]++;
3302 return 0;
3303 }
3304
3305 return __push_leaf_right(trans, path, min_data_size, empty, right,
3306 free_space, left_nritems, min_slot);
3307out_unlock:
3308 btrfs_tree_unlock(right);
3309 free_extent_buffer(right);
3310 return 1;
3311}
3312
3313/*
3314 * push some data in the path leaf to the left, trying to free up at
3315 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3316 *
3317 * max_slot can put a limit on how far into the leaf we'll push items. The
3318 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3319 * items
3320 */
3321static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
3322 struct btrfs_path *path, int data_size,
3323 int empty, struct extent_buffer *left,
3324 int free_space, u32 right_nritems,
3325 u32 max_slot)
3326{
3327 struct btrfs_fs_info *fs_info = left->fs_info;
3328 struct btrfs_disk_key disk_key;
3329 struct extent_buffer *right = path->nodes[0];
3330 int i;
3331 int push_space = 0;
3332 int push_items = 0;
3333 u32 old_left_nritems;
3334 u32 nr;
3335 int ret = 0;
3336 u32 this_item_size;
3337 u32 old_left_item_size;
3338 struct btrfs_map_token token;
3339
3340 if (empty)
3341 nr = min(right_nritems, max_slot);
3342 else
3343 nr = min(right_nritems - 1, max_slot);
3344
3345 for (i = 0; i < nr; i++) {
3346 if (!empty && push_items > 0) {
3347 if (path->slots[0] < i)
3348 break;
3349 if (path->slots[0] == i) {
3350 int space = btrfs_leaf_free_space(right);
3351
3352 if (space + push_space * 2 > free_space)
3353 break;
3354 }
3355 }
3356
3357 if (path->slots[0] == i)
3358 push_space += data_size;
3359
3360 this_item_size = btrfs_item_size(right, i);
3361 if (this_item_size + sizeof(struct btrfs_item) + push_space >
3362 free_space)
3363 break;
3364
3365 push_items++;
3366 push_space += this_item_size + sizeof(struct btrfs_item);
3367 }
3368
3369 if (push_items == 0) {
3370 ret = 1;
3371 goto out;
3372 }
3373 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3374
3375 /* push data from right to left */
3376 copy_leaf_items(left, right, btrfs_header_nritems(left), 0, push_items);
3377
3378 push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3379 btrfs_item_offset(right, push_items - 1);
3380
3381 copy_leaf_data(left, right, leaf_data_end(left) - push_space,
3382 btrfs_item_offset(right, push_items - 1), push_space);
3383 old_left_nritems = btrfs_header_nritems(left);
3384 BUG_ON(old_left_nritems <= 0);
3385
3386 btrfs_init_map_token(&token, left);
3387 old_left_item_size = btrfs_item_offset(left, old_left_nritems - 1);
3388 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3389 u32 ioff;
3390
3391 ioff = btrfs_token_item_offset(&token, i);
3392 btrfs_set_token_item_offset(&token, i,
3393 ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
3394 }
3395 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3396
3397 /* fixup right node */
3398 if (push_items > right_nritems)
3399 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3400 right_nritems);
3401
3402 if (push_items < right_nritems) {
3403 push_space = btrfs_item_offset(right, push_items - 1) -
3404 leaf_data_end(right);
3405 memmove_leaf_data(right,
3406 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3407 leaf_data_end(right), push_space);
3408
3409 memmove_leaf_items(right, 0, push_items,
3410 btrfs_header_nritems(right) - push_items);
3411 }
3412
3413 btrfs_init_map_token(&token, right);
3414 right_nritems -= push_items;
3415 btrfs_set_header_nritems(right, right_nritems);
3416 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3417 for (i = 0; i < right_nritems; i++) {
3418 push_space = push_space - btrfs_token_item_size(&token, i);
3419 btrfs_set_token_item_offset(&token, i, push_space);
3420 }
3421
3422 btrfs_mark_buffer_dirty(trans, left);
3423 if (right_nritems)
3424 btrfs_mark_buffer_dirty(trans, right);
3425 else
3426 btrfs_clear_buffer_dirty(trans, right);
3427
3428 btrfs_item_key(right, &disk_key, 0);
3429 fixup_low_keys(trans, path, &disk_key, 1);
3430
3431 /* then fixup the leaf pointer in the path */
3432 if (path->slots[0] < push_items) {
3433 path->slots[0] += old_left_nritems;
3434 btrfs_tree_unlock(path->nodes[0]);
3435 free_extent_buffer(path->nodes[0]);
3436 path->nodes[0] = left;
3437 path->slots[1] -= 1;
3438 } else {
3439 btrfs_tree_unlock(left);
3440 free_extent_buffer(left);
3441 path->slots[0] -= push_items;
3442 }
3443 BUG_ON(path->slots[0] < 0);
3444 return ret;
3445out:
3446 btrfs_tree_unlock(left);
3447 free_extent_buffer(left);
3448 return ret;
3449}
3450
3451/*
3452 * push some data in the path leaf to the left, trying to free up at
3453 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3454 *
3455 * max_slot can put a limit on how far into the leaf we'll push items. The
3456 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
3457 * items
3458 */
3459static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3460 *root, struct btrfs_path *path, int min_data_size,
3461 int data_size, int empty, u32 max_slot)
3462{
3463 struct extent_buffer *right = path->nodes[0];
3464 struct extent_buffer *left;
3465 int slot;
3466 int free_space;
3467 u32 right_nritems;
3468 int ret = 0;
3469
3470 slot = path->slots[1];
3471 if (slot == 0)
3472 return 1;
3473 if (!path->nodes[1])
3474 return 1;
3475
3476 right_nritems = btrfs_header_nritems(right);
3477 if (right_nritems == 0)
3478 return 1;
3479
3480 btrfs_assert_tree_write_locked(path->nodes[1]);
3481
3482 left = btrfs_read_node_slot(path->nodes[1], slot - 1);
3483 if (IS_ERR(left))
3484 return PTR_ERR(left);
3485
3486 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
3487
3488 free_space = btrfs_leaf_free_space(left);
3489 if (free_space < data_size) {
3490 ret = 1;
3491 goto out;
3492 }
3493
3494 ret = btrfs_cow_block(trans, root, left,
3495 path->nodes[1], slot - 1, &left,
3496 BTRFS_NESTING_LEFT_COW);
3497 if (ret) {
3498 /* we hit -ENOSPC, but it isn't fatal here */
3499 if (ret == -ENOSPC)
3500 ret = 1;
3501 goto out;
3502 }
3503
3504 if (check_sibling_keys(left, right)) {
3505 ret = -EUCLEAN;
3506 btrfs_abort_transaction(trans, ret);
3507 goto out;
3508 }
3509 return __push_leaf_left(trans, path, min_data_size, empty, left,
3510 free_space, right_nritems, max_slot);
3511out:
3512 btrfs_tree_unlock(left);
3513 free_extent_buffer(left);
3514 return ret;
3515}
3516
3517/*
3518 * split the path's leaf in two, making sure there is at least data_size
3519 * available for the resulting leaf level of the path.
3520 */
3521static noinline int copy_for_split(struct btrfs_trans_handle *trans,
3522 struct btrfs_path *path,
3523 struct extent_buffer *l,
3524 struct extent_buffer *right,
3525 int slot, int mid, int nritems)
3526{
3527 struct btrfs_fs_info *fs_info = trans->fs_info;
3528 int data_copy_size;
3529 int rt_data_off;
3530 int i;
3531 int ret;
3532 struct btrfs_disk_key disk_key;
3533 struct btrfs_map_token token;
3534
3535 nritems = nritems - mid;
3536 btrfs_set_header_nritems(right, nritems);
3537 data_copy_size = btrfs_item_data_end(l, mid) - leaf_data_end(l);
3538
3539 copy_leaf_items(right, l, 0, mid, nritems);
3540
3541 copy_leaf_data(right, l, BTRFS_LEAF_DATA_SIZE(fs_info) - data_copy_size,
3542 leaf_data_end(l), data_copy_size);
3543
3544 rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_data_end(l, mid);
3545
3546 btrfs_init_map_token(&token, right);
3547 for (i = 0; i < nritems; i++) {
3548 u32 ioff;
3549
3550 ioff = btrfs_token_item_offset(&token, i);
3551 btrfs_set_token_item_offset(&token, i, ioff + rt_data_off);
3552 }
3553
3554 btrfs_set_header_nritems(l, mid);
3555 btrfs_item_key(right, &disk_key, 0);
3556 ret = insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
3557 if (ret < 0)
3558 return ret;
3559
3560 btrfs_mark_buffer_dirty(trans, right);
3561 btrfs_mark_buffer_dirty(trans, l);
3562 BUG_ON(path->slots[0] != slot);
3563
3564 if (mid <= slot) {
3565 btrfs_tree_unlock(path->nodes[0]);
3566 free_extent_buffer(path->nodes[0]);
3567 path->nodes[0] = right;
3568 path->slots[0] -= mid;
3569 path->slots[1] += 1;
3570 } else {
3571 btrfs_tree_unlock(right);
3572 free_extent_buffer(right);
3573 }
3574
3575 BUG_ON(path->slots[0] < 0);
3576
3577 return 0;
3578}
3579
3580/*
3581 * double splits happen when we need to insert a big item in the middle
3582 * of a leaf. A double split can leave us with 3 mostly empty leaves:
3583 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
3584 * A B C
3585 *
3586 * We avoid this by trying to push the items on either side of our target
3587 * into the adjacent leaves. If all goes well we can avoid the double split
3588 * completely.
3589 */
3590static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
3591 struct btrfs_root *root,
3592 struct btrfs_path *path,
3593 int data_size)
3594{
3595 int ret;
3596 int progress = 0;
3597 int slot;
3598 u32 nritems;
3599 int space_needed = data_size;
3600
3601 slot = path->slots[0];
3602 if (slot < btrfs_header_nritems(path->nodes[0]))
3603 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3604
3605 /*
3606 * try to push all the items after our slot into the
3607 * right leaf
3608 */
3609 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
3610 if (ret < 0)
3611 return ret;
3612
3613 if (ret == 0)
3614 progress++;
3615
3616 nritems = btrfs_header_nritems(path->nodes[0]);
3617 /*
3618 * our goal is to get our slot at the start or end of a leaf. If
3619 * we've done so we're done
3620 */
3621 if (path->slots[0] == 0 || path->slots[0] == nritems)
3622 return 0;
3623
3624 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3625 return 0;
3626
3627 /* try to push all the items before our slot into the next leaf */
3628 slot = path->slots[0];
3629 space_needed = data_size;
3630 if (slot > 0)
3631 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3632 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
3633 if (ret < 0)
3634 return ret;
3635
3636 if (ret == 0)
3637 progress++;
3638
3639 if (progress)
3640 return 0;
3641 return 1;
3642}
3643
3644/*
3645 * split the path's leaf in two, making sure there is at least data_size
3646 * available for the resulting leaf level of the path.
3647 *
3648 * returns 0 if all went well and < 0 on failure.
3649 */
3650static noinline int split_leaf(struct btrfs_trans_handle *trans,
3651 struct btrfs_root *root,
3652 const struct btrfs_key *ins_key,
3653 struct btrfs_path *path, int data_size,
3654 int extend)
3655{
3656 struct btrfs_disk_key disk_key;
3657 struct extent_buffer *l;
3658 u32 nritems;
3659 int mid;
3660 int slot;
3661 struct extent_buffer *right;
3662 struct btrfs_fs_info *fs_info = root->fs_info;
3663 int ret = 0;
3664 int wret;
3665 int split;
3666 int num_doubles = 0;
3667 int tried_avoid_double = 0;
3668
3669 l = path->nodes[0];
3670 slot = path->slots[0];
3671 if (extend && data_size + btrfs_item_size(l, slot) +
3672 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
3673 return -EOVERFLOW;
3674
3675 /* first try to make some room by pushing left and right */
3676 if (data_size && path->nodes[1]) {
3677 int space_needed = data_size;
3678
3679 if (slot < btrfs_header_nritems(l))
3680 space_needed -= btrfs_leaf_free_space(l);
3681
3682 wret = push_leaf_right(trans, root, path, space_needed,
3683 space_needed, 0, 0);
3684 if (wret < 0)
3685 return wret;
3686 if (wret) {
3687 space_needed = data_size;
3688 if (slot > 0)
3689 space_needed -= btrfs_leaf_free_space(l);
3690 wret = push_leaf_left(trans, root, path, space_needed,
3691 space_needed, 0, (u32)-1);
3692 if (wret < 0)
3693 return wret;
3694 }
3695 l = path->nodes[0];
3696
3697 /* did the pushes work? */
3698 if (btrfs_leaf_free_space(l) >= data_size)
3699 return 0;
3700 }
3701
3702 if (!path->nodes[1]) {
3703 ret = insert_new_root(trans, root, path, 1);
3704 if (ret)
3705 return ret;
3706 }
3707again:
3708 split = 1;
3709 l = path->nodes[0];
3710 slot = path->slots[0];
3711 nritems = btrfs_header_nritems(l);
3712 mid = (nritems + 1) / 2;
3713
3714 if (mid <= slot) {
3715 if (nritems == 1 ||
3716 leaf_space_used(l, mid, nritems - mid) + data_size >
3717 BTRFS_LEAF_DATA_SIZE(fs_info)) {
3718 if (slot >= nritems) {
3719 split = 0;
3720 } else {
3721 mid = slot;
3722 if (mid != nritems &&
3723 leaf_space_used(l, mid, nritems - mid) +
3724 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3725 if (data_size && !tried_avoid_double)
3726 goto push_for_double;
3727 split = 2;
3728 }
3729 }
3730 }
3731 } else {
3732 if (leaf_space_used(l, 0, mid) + data_size >
3733 BTRFS_LEAF_DATA_SIZE(fs_info)) {
3734 if (!extend && data_size && slot == 0) {
3735 split = 0;
3736 } else if ((extend || !data_size) && slot == 0) {
3737 mid = 1;
3738 } else {
3739 mid = slot;
3740 if (mid != nritems &&
3741 leaf_space_used(l, mid, nritems - mid) +
3742 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3743 if (data_size && !tried_avoid_double)
3744 goto push_for_double;
3745 split = 2;
3746 }
3747 }
3748 }
3749 }
3750
3751 if (split == 0)
3752 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3753 else
3754 btrfs_item_key(l, &disk_key, mid);
3755
3756 /*
3757 * We have to about BTRFS_NESTING_NEW_ROOT here if we've done a double
3758 * split, because we're only allowed to have MAX_LOCKDEP_SUBCLASSES
3759 * subclasses, which is 8 at the time of this patch, and we've maxed it
3760 * out. In the future we could add a
3761 * BTRFS_NESTING_SPLIT_THE_SPLITTENING if we need to, but for now just
3762 * use BTRFS_NESTING_NEW_ROOT.
3763 */
3764 right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3765 &disk_key, 0, l->start, 0, 0,
3766 num_doubles ? BTRFS_NESTING_NEW_ROOT :
3767 BTRFS_NESTING_SPLIT);
3768 if (IS_ERR(right))
3769 return PTR_ERR(right);
3770
3771 root_add_used_bytes(root);
3772
3773 if (split == 0) {
3774 if (mid <= slot) {
3775 btrfs_set_header_nritems(right, 0);
3776 ret = insert_ptr(trans, path, &disk_key,
3777 right->start, path->slots[1] + 1, 1);
3778 if (ret < 0) {
3779 btrfs_tree_unlock(right);
3780 free_extent_buffer(right);
3781 return ret;
3782 }
3783 btrfs_tree_unlock(path->nodes[0]);
3784 free_extent_buffer(path->nodes[0]);
3785 path->nodes[0] = right;
3786 path->slots[0] = 0;
3787 path->slots[1] += 1;
3788 } else {
3789 btrfs_set_header_nritems(right, 0);
3790 ret = insert_ptr(trans, path, &disk_key,
3791 right->start, path->slots[1], 1);
3792 if (ret < 0) {
3793 btrfs_tree_unlock(right);
3794 free_extent_buffer(right);
3795 return ret;
3796 }
3797 btrfs_tree_unlock(path->nodes[0]);
3798 free_extent_buffer(path->nodes[0]);
3799 path->nodes[0] = right;
3800 path->slots[0] = 0;
3801 if (path->slots[1] == 0)
3802 fixup_low_keys(trans, path, &disk_key, 1);
3803 }
3804 /*
3805 * We create a new leaf 'right' for the required ins_len and
3806 * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
3807 * the content of ins_len to 'right'.
3808 */
3809 return ret;
3810 }
3811
3812 ret = copy_for_split(trans, path, l, right, slot, mid, nritems);
3813 if (ret < 0) {
3814 btrfs_tree_unlock(right);
3815 free_extent_buffer(right);
3816 return ret;
3817 }
3818
3819 if (split == 2) {
3820 BUG_ON(num_doubles != 0);
3821 num_doubles++;
3822 goto again;
3823 }
3824
3825 return 0;
3826
3827push_for_double:
3828 push_for_double_split(trans, root, path, data_size);
3829 tried_avoid_double = 1;
3830 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3831 return 0;
3832 goto again;
3833}
3834
3835static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3836 struct btrfs_root *root,
3837 struct btrfs_path *path, int ins_len)
3838{
3839 struct btrfs_key key;
3840 struct extent_buffer *leaf;
3841 struct btrfs_file_extent_item *fi;
3842 u64 extent_len = 0;
3843 u32 item_size;
3844 int ret;
3845
3846 leaf = path->nodes[0];
3847 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3848
3849 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3850 key.type != BTRFS_EXTENT_CSUM_KEY);
3851
3852 if (btrfs_leaf_free_space(leaf) >= ins_len)
3853 return 0;
3854
3855 item_size = btrfs_item_size(leaf, path->slots[0]);
3856 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3857 fi = btrfs_item_ptr(leaf, path->slots[0],
3858 struct btrfs_file_extent_item);
3859 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3860 }
3861 btrfs_release_path(path);
3862
3863 path->keep_locks = 1;
3864 path->search_for_split = 1;
3865 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3866 path->search_for_split = 0;
3867 if (ret > 0)
3868 ret = -EAGAIN;
3869 if (ret < 0)
3870 goto err;
3871
3872 ret = -EAGAIN;
3873 leaf = path->nodes[0];
3874 /* if our item isn't there, return now */
3875 if (item_size != btrfs_item_size(leaf, path->slots[0]))
3876 goto err;
3877
3878 /* the leaf has changed, it now has room. return now */
3879 if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
3880 goto err;
3881
3882 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3883 fi = btrfs_item_ptr(leaf, path->slots[0],
3884 struct btrfs_file_extent_item);
3885 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3886 goto err;
3887 }
3888
3889 ret = split_leaf(trans, root, &key, path, ins_len, 1);
3890 if (ret)
3891 goto err;
3892
3893 path->keep_locks = 0;
3894 btrfs_unlock_up_safe(path, 1);
3895 return 0;
3896err:
3897 path->keep_locks = 0;
3898 return ret;
3899}
3900
3901static noinline int split_item(struct btrfs_trans_handle *trans,
3902 struct btrfs_path *path,
3903 const struct btrfs_key *new_key,
3904 unsigned long split_offset)
3905{
3906 struct extent_buffer *leaf;
3907 int orig_slot, slot;
3908 char *buf;
3909 u32 nritems;
3910 u32 item_size;
3911 u32 orig_offset;
3912 struct btrfs_disk_key disk_key;
3913
3914 leaf = path->nodes[0];
3915 /*
3916 * Shouldn't happen because the caller must have previously called
3917 * setup_leaf_for_split() to make room for the new item in the leaf.
3918 */
3919 if (WARN_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item)))
3920 return -ENOSPC;
3921
3922 orig_slot = path->slots[0];
3923 orig_offset = btrfs_item_offset(leaf, path->slots[0]);
3924 item_size = btrfs_item_size(leaf, path->slots[0]);
3925
3926 buf = kmalloc(item_size, GFP_NOFS);
3927 if (!buf)
3928 return -ENOMEM;
3929
3930 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3931 path->slots[0]), item_size);
3932
3933 slot = path->slots[0] + 1;
3934 nritems = btrfs_header_nritems(leaf);
3935 if (slot != nritems) {
3936 /* shift the items */
3937 memmove_leaf_items(leaf, slot + 1, slot, nritems - slot);
3938 }
3939
3940 btrfs_cpu_key_to_disk(&disk_key, new_key);
3941 btrfs_set_item_key(leaf, &disk_key, slot);
3942
3943 btrfs_set_item_offset(leaf, slot, orig_offset);
3944 btrfs_set_item_size(leaf, slot, item_size - split_offset);
3945
3946 btrfs_set_item_offset(leaf, orig_slot,
3947 orig_offset + item_size - split_offset);
3948 btrfs_set_item_size(leaf, orig_slot, split_offset);
3949
3950 btrfs_set_header_nritems(leaf, nritems + 1);
3951
3952 /* write the data for the start of the original item */
3953 write_extent_buffer(leaf, buf,
3954 btrfs_item_ptr_offset(leaf, path->slots[0]),
3955 split_offset);
3956
3957 /* write the data for the new item */
3958 write_extent_buffer(leaf, buf + split_offset,
3959 btrfs_item_ptr_offset(leaf, slot),
3960 item_size - split_offset);
3961 btrfs_mark_buffer_dirty(trans, leaf);
3962
3963 BUG_ON(btrfs_leaf_free_space(leaf) < 0);
3964 kfree(buf);
3965 return 0;
3966}
3967
3968/*
3969 * This function splits a single item into two items,
3970 * giving 'new_key' to the new item and splitting the
3971 * old one at split_offset (from the start of the item).
3972 *
3973 * The path may be released by this operation. After
3974 * the split, the path is pointing to the old item. The
3975 * new item is going to be in the same node as the old one.
3976 *
3977 * Note, the item being split must be smaller enough to live alone on
3978 * a tree block with room for one extra struct btrfs_item
3979 *
3980 * This allows us to split the item in place, keeping a lock on the
3981 * leaf the entire time.
3982 */
3983int btrfs_split_item(struct btrfs_trans_handle *trans,
3984 struct btrfs_root *root,
3985 struct btrfs_path *path,
3986 const struct btrfs_key *new_key,
3987 unsigned long split_offset)
3988{
3989 int ret;
3990 ret = setup_leaf_for_split(trans, root, path,
3991 sizeof(struct btrfs_item));
3992 if (ret)
3993 return ret;
3994
3995 ret = split_item(trans, path, new_key, split_offset);
3996 return ret;
3997}
3998
3999/*
4000 * make the item pointed to by the path smaller. new_size indicates
4001 * how small to make it, and from_end tells us if we just chop bytes
4002 * off the end of the item or if we shift the item to chop bytes off
4003 * the front.
4004 */
4005void btrfs_truncate_item(struct btrfs_trans_handle *trans,
4006 struct btrfs_path *path, u32 new_size, int from_end)
4007{
4008 int slot;
4009 struct extent_buffer *leaf;
4010 u32 nritems;
4011 unsigned int data_end;
4012 unsigned int old_data_start;
4013 unsigned int old_size;
4014 unsigned int size_diff;
4015 int i;
4016 struct btrfs_map_token token;
4017
4018 leaf = path->nodes[0];
4019 slot = path->slots[0];
4020
4021 old_size = btrfs_item_size(leaf, slot);
4022 if (old_size == new_size)
4023 return;
4024
4025 nritems = btrfs_header_nritems(leaf);
4026 data_end = leaf_data_end(leaf);
4027
4028 old_data_start = btrfs_item_offset(leaf, slot);
4029
4030 size_diff = old_size - new_size;
4031
4032 BUG_ON(slot < 0);
4033 BUG_ON(slot >= nritems);
4034
4035 /*
4036 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4037 */
4038 /* first correct the data pointers */
4039 btrfs_init_map_token(&token, leaf);
4040 for (i = slot; i < nritems; i++) {
4041 u32 ioff;
4042
4043 ioff = btrfs_token_item_offset(&token, i);
4044 btrfs_set_token_item_offset(&token, i, ioff + size_diff);
4045 }
4046
4047 /* shift the data */
4048 if (from_end) {
4049 memmove_leaf_data(leaf, data_end + size_diff, data_end,
4050 old_data_start + new_size - data_end);
4051 } else {
4052 struct btrfs_disk_key disk_key;
4053 u64 offset;
4054
4055 btrfs_item_key(leaf, &disk_key, slot);
4056
4057 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4058 unsigned long ptr;
4059 struct btrfs_file_extent_item *fi;
4060
4061 fi = btrfs_item_ptr(leaf, slot,
4062 struct btrfs_file_extent_item);
4063 fi = (struct btrfs_file_extent_item *)(
4064 (unsigned long)fi - size_diff);
4065
4066 if (btrfs_file_extent_type(leaf, fi) ==
4067 BTRFS_FILE_EXTENT_INLINE) {
4068 ptr = btrfs_item_ptr_offset(leaf, slot);
4069 memmove_extent_buffer(leaf, ptr,
4070 (unsigned long)fi,
4071 BTRFS_FILE_EXTENT_INLINE_DATA_START);
4072 }
4073 }
4074
4075 memmove_leaf_data(leaf, data_end + size_diff, data_end,
4076 old_data_start - data_end);
4077
4078 offset = btrfs_disk_key_offset(&disk_key);
4079 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4080 btrfs_set_item_key(leaf, &disk_key, slot);
4081 if (slot == 0)
4082 fixup_low_keys(trans, path, &disk_key, 1);
4083 }
4084
4085 btrfs_set_item_size(leaf, slot, new_size);
4086 btrfs_mark_buffer_dirty(trans, leaf);
4087
4088 if (btrfs_leaf_free_space(leaf) < 0) {
4089 btrfs_print_leaf(leaf);
4090 BUG();
4091 }
4092}
4093
4094/*
4095 * make the item pointed to by the path bigger, data_size is the added size.
4096 */
4097void btrfs_extend_item(struct btrfs_trans_handle *trans,
4098 struct btrfs_path *path, u32 data_size)
4099{
4100 int slot;
4101 struct extent_buffer *leaf;
4102 u32 nritems;
4103 unsigned int data_end;
4104 unsigned int old_data;
4105 unsigned int old_size;
4106 int i;
4107 struct btrfs_map_token token;
4108
4109 leaf = path->nodes[0];
4110
4111 nritems = btrfs_header_nritems(leaf);
4112 data_end = leaf_data_end(leaf);
4113
4114 if (btrfs_leaf_free_space(leaf) < data_size) {
4115 btrfs_print_leaf(leaf);
4116 BUG();
4117 }
4118 slot = path->slots[0];
4119 old_data = btrfs_item_data_end(leaf, slot);
4120
4121 BUG_ON(slot < 0);
4122 if (slot >= nritems) {
4123 btrfs_print_leaf(leaf);
4124 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
4125 slot, nritems);
4126 BUG();
4127 }
4128
4129 /*
4130 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4131 */
4132 /* first correct the data pointers */
4133 btrfs_init_map_token(&token, leaf);
4134 for (i = slot; i < nritems; i++) {
4135 u32 ioff;
4136
4137 ioff = btrfs_token_item_offset(&token, i);
4138 btrfs_set_token_item_offset(&token, i, ioff - data_size);
4139 }
4140
4141 /* shift the data */
4142 memmove_leaf_data(leaf, data_end - data_size, data_end,
4143 old_data - data_end);
4144
4145 data_end = old_data;
4146 old_size = btrfs_item_size(leaf, slot);
4147 btrfs_set_item_size(leaf, slot, old_size + data_size);
4148 btrfs_mark_buffer_dirty(trans, leaf);
4149
4150 if (btrfs_leaf_free_space(leaf) < 0) {
4151 btrfs_print_leaf(leaf);
4152 BUG();
4153 }
4154}
4155
4156/*
4157 * Make space in the node before inserting one or more items.
4158 *
4159 * @trans: transaction handle
4160 * @root: root we are inserting items to
4161 * @path: points to the leaf/slot where we are going to insert new items
4162 * @batch: information about the batch of items to insert
4163 *
4164 * Main purpose is to save stack depth by doing the bulk of the work in a
4165 * function that doesn't call btrfs_search_slot
4166 */
4167static void setup_items_for_insert(struct btrfs_trans_handle *trans,
4168 struct btrfs_root *root, struct btrfs_path *path,
4169 const struct btrfs_item_batch *batch)
4170{
4171 struct btrfs_fs_info *fs_info = root->fs_info;
4172 int i;
4173 u32 nritems;
4174 unsigned int data_end;
4175 struct btrfs_disk_key disk_key;
4176 struct extent_buffer *leaf;
4177 int slot;
4178 struct btrfs_map_token token;
4179 u32 total_size;
4180
4181 /*
4182 * Before anything else, update keys in the parent and other ancestors
4183 * if needed, then release the write locks on them, so that other tasks
4184 * can use them while we modify the leaf.
4185 */
4186 if (path->slots[0] == 0) {
4187 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[0]);
4188 fixup_low_keys(trans, path, &disk_key, 1);
4189 }
4190 btrfs_unlock_up_safe(path, 1);
4191
4192 leaf = path->nodes[0];
4193 slot = path->slots[0];
4194
4195 nritems = btrfs_header_nritems(leaf);
4196 data_end = leaf_data_end(leaf);
4197 total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4198
4199 if (btrfs_leaf_free_space(leaf) < total_size) {
4200 btrfs_print_leaf(leaf);
4201 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4202 total_size, btrfs_leaf_free_space(leaf));
4203 BUG();
4204 }
4205
4206 btrfs_init_map_token(&token, leaf);
4207 if (slot != nritems) {
4208 unsigned int old_data = btrfs_item_data_end(leaf, slot);
4209
4210 if (old_data < data_end) {
4211 btrfs_print_leaf(leaf);
4212 btrfs_crit(fs_info,
4213 "item at slot %d with data offset %u beyond data end of leaf %u",
4214 slot, old_data, data_end);
4215 BUG();
4216 }
4217 /*
4218 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4219 */
4220 /* first correct the data pointers */
4221 for (i = slot; i < nritems; i++) {
4222 u32 ioff;
4223
4224 ioff = btrfs_token_item_offset(&token, i);
4225 btrfs_set_token_item_offset(&token, i,
4226 ioff - batch->total_data_size);
4227 }
4228 /* shift the items */
4229 memmove_leaf_items(leaf, slot + batch->nr, slot, nritems - slot);
4230
4231 /* shift the data */
4232 memmove_leaf_data(leaf, data_end - batch->total_data_size,
4233 data_end, old_data - data_end);
4234 data_end = old_data;
4235 }
4236
4237 /* setup the item for the new data */
4238 for (i = 0; i < batch->nr; i++) {
4239 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[i]);
4240 btrfs_set_item_key(leaf, &disk_key, slot + i);
4241 data_end -= batch->data_sizes[i];
4242 btrfs_set_token_item_offset(&token, slot + i, data_end);
4243 btrfs_set_token_item_size(&token, slot + i, batch->data_sizes[i]);
4244 }
4245
4246 btrfs_set_header_nritems(leaf, nritems + batch->nr);
4247 btrfs_mark_buffer_dirty(trans, leaf);
4248
4249 if (btrfs_leaf_free_space(leaf) < 0) {
4250 btrfs_print_leaf(leaf);
4251 BUG();
4252 }
4253}
4254
4255/*
4256 * Insert a new item into a leaf.
4257 *
4258 * @trans: Transaction handle.
4259 * @root: The root of the btree.
4260 * @path: A path pointing to the target leaf and slot.
4261 * @key: The key of the new item.
4262 * @data_size: The size of the data associated with the new key.
4263 */
4264void btrfs_setup_item_for_insert(struct btrfs_trans_handle *trans,
4265 struct btrfs_root *root,
4266 struct btrfs_path *path,
4267 const struct btrfs_key *key,
4268 u32 data_size)
4269{
4270 struct btrfs_item_batch batch;
4271
4272 batch.keys = key;
4273 batch.data_sizes = &data_size;
4274 batch.total_data_size = data_size;
4275 batch.nr = 1;
4276
4277 setup_items_for_insert(trans, root, path, &batch);
4278}
4279
4280/*
4281 * Given a key and some data, insert items into the tree.
4282 * This does all the path init required, making room in the tree if needed.
4283 */
4284int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4285 struct btrfs_root *root,
4286 struct btrfs_path *path,
4287 const struct btrfs_item_batch *batch)
4288{
4289 int ret = 0;
4290 int slot;
4291 u32 total_size;
4292
4293 total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4294 ret = btrfs_search_slot(trans, root, &batch->keys[0], path, total_size, 1);
4295 if (ret == 0)
4296 return -EEXIST;
4297 if (ret < 0)
4298 return ret;
4299
4300 slot = path->slots[0];
4301 BUG_ON(slot < 0);
4302
4303 setup_items_for_insert(trans, root, path, batch);
4304 return 0;
4305}
4306
4307/*
4308 * Given a key and some data, insert an item into the tree.
4309 * This does all the path init required, making room in the tree if needed.
4310 */
4311int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4312 const struct btrfs_key *cpu_key, void *data,
4313 u32 data_size)
4314{
4315 int ret = 0;
4316 struct btrfs_path *path;
4317 struct extent_buffer *leaf;
4318 unsigned long ptr;
4319
4320 path = btrfs_alloc_path();
4321 if (!path)
4322 return -ENOMEM;
4323 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4324 if (!ret) {
4325 leaf = path->nodes[0];
4326 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4327 write_extent_buffer(leaf, data, ptr, data_size);
4328 btrfs_mark_buffer_dirty(trans, leaf);
4329 }
4330 btrfs_free_path(path);
4331 return ret;
4332}
4333
4334/*
4335 * This function duplicates an item, giving 'new_key' to the new item.
4336 * It guarantees both items live in the same tree leaf and the new item is
4337 * contiguous with the original item.
4338 *
4339 * This allows us to split a file extent in place, keeping a lock on the leaf
4340 * the entire time.
4341 */
4342int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4343 struct btrfs_root *root,
4344 struct btrfs_path *path,
4345 const struct btrfs_key *new_key)
4346{
4347 struct extent_buffer *leaf;
4348 int ret;
4349 u32 item_size;
4350
4351 leaf = path->nodes[0];
4352 item_size = btrfs_item_size(leaf, path->slots[0]);
4353 ret = setup_leaf_for_split(trans, root, path,
4354 item_size + sizeof(struct btrfs_item));
4355 if (ret)
4356 return ret;
4357
4358 path->slots[0]++;
4359 btrfs_setup_item_for_insert(trans, root, path, new_key, item_size);
4360 leaf = path->nodes[0];
4361 memcpy_extent_buffer(leaf,
4362 btrfs_item_ptr_offset(leaf, path->slots[0]),
4363 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4364 item_size);
4365 return 0;
4366}
4367
4368/*
4369 * delete the pointer from a given node.
4370 *
4371 * the tree should have been previously balanced so the deletion does not
4372 * empty a node.
4373 *
4374 * This is exported for use inside btrfs-progs, don't un-export it.
4375 */
4376int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4377 struct btrfs_path *path, int level, int slot)
4378{
4379 struct extent_buffer *parent = path->nodes[level];
4380 u32 nritems;
4381 int ret;
4382
4383 nritems = btrfs_header_nritems(parent);
4384 if (slot != nritems - 1) {
4385 if (level) {
4386 ret = btrfs_tree_mod_log_insert_move(parent, slot,
4387 slot + 1, nritems - slot - 1);
4388 if (ret < 0) {
4389 btrfs_abort_transaction(trans, ret);
4390 return ret;
4391 }
4392 }
4393 memmove_extent_buffer(parent,
4394 btrfs_node_key_ptr_offset(parent, slot),
4395 btrfs_node_key_ptr_offset(parent, slot + 1),
4396 sizeof(struct btrfs_key_ptr) *
4397 (nritems - slot - 1));
4398 } else if (level) {
4399 ret = btrfs_tree_mod_log_insert_key(parent, slot,
4400 BTRFS_MOD_LOG_KEY_REMOVE);
4401 if (ret < 0) {
4402 btrfs_abort_transaction(trans, ret);
4403 return ret;
4404 }
4405 }
4406
4407 nritems--;
4408 btrfs_set_header_nritems(parent, nritems);
4409 if (nritems == 0 && parent == root->node) {
4410 BUG_ON(btrfs_header_level(root->node) != 1);
4411 /* just turn the root into a leaf and break */
4412 btrfs_set_header_level(root->node, 0);
4413 } else if (slot == 0) {
4414 struct btrfs_disk_key disk_key;
4415
4416 btrfs_node_key(parent, &disk_key, 0);
4417 fixup_low_keys(trans, path, &disk_key, level + 1);
4418 }
4419 btrfs_mark_buffer_dirty(trans, parent);
4420 return 0;
4421}
4422
4423/*
4424 * a helper function to delete the leaf pointed to by path->slots[1] and
4425 * path->nodes[1].
4426 *
4427 * This deletes the pointer in path->nodes[1] and frees the leaf
4428 * block extent. zero is returned if it all worked out, < 0 otherwise.
4429 *
4430 * The path must have already been setup for deleting the leaf, including
4431 * all the proper balancing. path->nodes[1] must be locked.
4432 */
4433static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
4434 struct btrfs_root *root,
4435 struct btrfs_path *path,
4436 struct extent_buffer *leaf)
4437{
4438 int ret;
4439
4440 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4441 ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
4442 if (ret < 0)
4443 return ret;
4444
4445 /*
4446 * btrfs_free_extent is expensive, we want to make sure we
4447 * aren't holding any locks when we call it
4448 */
4449 btrfs_unlock_up_safe(path, 0);
4450
4451 root_sub_used_bytes(root);
4452
4453 atomic_inc(&leaf->refs);
4454 btrfs_free_tree_block(trans, btrfs_root_id(root), leaf, 0, 1);
4455 free_extent_buffer_stale(leaf);
4456 return 0;
4457}
4458/*
4459 * delete the item at the leaf level in path. If that empties
4460 * the leaf, remove it from the tree
4461 */
4462int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4463 struct btrfs_path *path, int slot, int nr)
4464{
4465 struct btrfs_fs_info *fs_info = root->fs_info;
4466 struct extent_buffer *leaf;
4467 int ret = 0;
4468 int wret;
4469 u32 nritems;
4470
4471 leaf = path->nodes[0];
4472 nritems = btrfs_header_nritems(leaf);
4473
4474 if (slot + nr != nritems) {
4475 const u32 last_off = btrfs_item_offset(leaf, slot + nr - 1);
4476 const int data_end = leaf_data_end(leaf);
4477 struct btrfs_map_token token;
4478 u32 dsize = 0;
4479 int i;
4480
4481 for (i = 0; i < nr; i++)
4482 dsize += btrfs_item_size(leaf, slot + i);
4483
4484 memmove_leaf_data(leaf, data_end + dsize, data_end,
4485 last_off - data_end);
4486
4487 btrfs_init_map_token(&token, leaf);
4488 for (i = slot + nr; i < nritems; i++) {
4489 u32 ioff;
4490
4491 ioff = btrfs_token_item_offset(&token, i);
4492 btrfs_set_token_item_offset(&token, i, ioff + dsize);
4493 }
4494
4495 memmove_leaf_items(leaf, slot, slot + nr, nritems - slot - nr);
4496 }
4497 btrfs_set_header_nritems(leaf, nritems - nr);
4498 nritems -= nr;
4499
4500 /* delete the leaf if we've emptied it */
4501 if (nritems == 0) {
4502 if (leaf == root->node) {
4503 btrfs_set_header_level(leaf, 0);
4504 } else {
4505 btrfs_clear_buffer_dirty(trans, leaf);
4506 ret = btrfs_del_leaf(trans, root, path, leaf);
4507 if (ret < 0)
4508 return ret;
4509 }
4510 } else {
4511 int used = leaf_space_used(leaf, 0, nritems);
4512 if (slot == 0) {
4513 struct btrfs_disk_key disk_key;
4514
4515 btrfs_item_key(leaf, &disk_key, 0);
4516 fixup_low_keys(trans, path, &disk_key, 1);
4517 }
4518
4519 /*
4520 * Try to delete the leaf if it is mostly empty. We do this by
4521 * trying to move all its items into its left and right neighbours.
4522 * If we can't move all the items, then we don't delete it - it's
4523 * not ideal, but future insertions might fill the leaf with more
4524 * items, or items from other leaves might be moved later into our
4525 * leaf due to deletions on those leaves.
4526 */
4527 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4528 u32 min_push_space;
4529
4530 /* push_leaf_left fixes the path.
4531 * make sure the path still points to our leaf
4532 * for possible call to btrfs_del_ptr below
4533 */
4534 slot = path->slots[1];
4535 atomic_inc(&leaf->refs);
4536 /*
4537 * We want to be able to at least push one item to the
4538 * left neighbour leaf, and that's the first item.
4539 */
4540 min_push_space = sizeof(struct btrfs_item) +
4541 btrfs_item_size(leaf, 0);
4542 wret = push_leaf_left(trans, root, path, 0,
4543 min_push_space, 1, (u32)-1);
4544 if (wret < 0 && wret != -ENOSPC)
4545 ret = wret;
4546
4547 if (path->nodes[0] == leaf &&
4548 btrfs_header_nritems(leaf)) {
4549 /*
4550 * If we were not able to push all items from our
4551 * leaf to its left neighbour, then attempt to
4552 * either push all the remaining items to the
4553 * right neighbour or none. There's no advantage
4554 * in pushing only some items, instead of all, as
4555 * it's pointless to end up with a leaf having
4556 * too few items while the neighbours can be full
4557 * or nearly full.
4558 */
4559 nritems = btrfs_header_nritems(leaf);
4560 min_push_space = leaf_space_used(leaf, 0, nritems);
4561 wret = push_leaf_right(trans, root, path, 0,
4562 min_push_space, 1, 0);
4563 if (wret < 0 && wret != -ENOSPC)
4564 ret = wret;
4565 }
4566
4567 if (btrfs_header_nritems(leaf) == 0) {
4568 path->slots[1] = slot;
4569 ret = btrfs_del_leaf(trans, root, path, leaf);
4570 if (ret < 0)
4571 return ret;
4572 free_extent_buffer(leaf);
4573 ret = 0;
4574 } else {
4575 /* if we're still in the path, make sure
4576 * we're dirty. Otherwise, one of the
4577 * push_leaf functions must have already
4578 * dirtied this buffer
4579 */
4580 if (path->nodes[0] == leaf)
4581 btrfs_mark_buffer_dirty(trans, leaf);
4582 free_extent_buffer(leaf);
4583 }
4584 } else {
4585 btrfs_mark_buffer_dirty(trans, leaf);
4586 }
4587 }
4588 return ret;
4589}
4590
4591/*
4592 * A helper function to walk down the tree starting at min_key, and looking
4593 * for nodes or leaves that are have a minimum transaction id.
4594 * This is used by the btree defrag code, and tree logging
4595 *
4596 * This does not cow, but it does stuff the starting key it finds back
4597 * into min_key, so you can call btrfs_search_slot with cow=1 on the
4598 * key and get a writable path.
4599 *
4600 * This honors path->lowest_level to prevent descent past a given level
4601 * of the tree.
4602 *
4603 * min_trans indicates the oldest transaction that you are interested
4604 * in walking through. Any nodes or leaves older than min_trans are
4605 * skipped over (without reading them).
4606 *
4607 * returns zero if something useful was found, < 0 on error and 1 if there
4608 * was nothing in the tree that matched the search criteria.
4609 */
4610int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4611 struct btrfs_path *path,
4612 u64 min_trans)
4613{
4614 struct extent_buffer *cur;
4615 struct btrfs_key found_key;
4616 int slot;
4617 int sret;
4618 u32 nritems;
4619 int level;
4620 int ret = 1;
4621 int keep_locks = path->keep_locks;
4622
4623 ASSERT(!path->nowait);
4624 path->keep_locks = 1;
4625again:
4626 cur = btrfs_read_lock_root_node(root);
4627 level = btrfs_header_level(cur);
4628 WARN_ON(path->nodes[level]);
4629 path->nodes[level] = cur;
4630 path->locks[level] = BTRFS_READ_LOCK;
4631
4632 if (btrfs_header_generation(cur) < min_trans) {
4633 ret = 1;
4634 goto out;
4635 }
4636 while (1) {
4637 nritems = btrfs_header_nritems(cur);
4638 level = btrfs_header_level(cur);
4639 sret = btrfs_bin_search(cur, 0, min_key, &slot);
4640 if (sret < 0) {
4641 ret = sret;
4642 goto out;
4643 }
4644
4645 /* at the lowest level, we're done, setup the path and exit */
4646 if (level == path->lowest_level) {
4647 if (slot >= nritems)
4648 goto find_next_key;
4649 ret = 0;
4650 path->slots[level] = slot;
4651 btrfs_item_key_to_cpu(cur, &found_key, slot);
4652 goto out;
4653 }
4654 if (sret && slot > 0)
4655 slot--;
4656 /*
4657 * check this node pointer against the min_trans parameters.
4658 * If it is too old, skip to the next one.
4659 */
4660 while (slot < nritems) {
4661 u64 gen;
4662
4663 gen = btrfs_node_ptr_generation(cur, slot);
4664 if (gen < min_trans) {
4665 slot++;
4666 continue;
4667 }
4668 break;
4669 }
4670find_next_key:
4671 /*
4672 * we didn't find a candidate key in this node, walk forward
4673 * and find another one
4674 */
4675 if (slot >= nritems) {
4676 path->slots[level] = slot;
4677 sret = btrfs_find_next_key(root, path, min_key, level,
4678 min_trans);
4679 if (sret == 0) {
4680 btrfs_release_path(path);
4681 goto again;
4682 } else {
4683 goto out;
4684 }
4685 }
4686 /* save our key for returning back */
4687 btrfs_node_key_to_cpu(cur, &found_key, slot);
4688 path->slots[level] = slot;
4689 if (level == path->lowest_level) {
4690 ret = 0;
4691 goto out;
4692 }
4693 cur = btrfs_read_node_slot(cur, slot);
4694 if (IS_ERR(cur)) {
4695 ret = PTR_ERR(cur);
4696 goto out;
4697 }
4698
4699 btrfs_tree_read_lock(cur);
4700
4701 path->locks[level - 1] = BTRFS_READ_LOCK;
4702 path->nodes[level - 1] = cur;
4703 unlock_up(path, level, 1, 0, NULL);
4704 }
4705out:
4706 path->keep_locks = keep_locks;
4707 if (ret == 0) {
4708 btrfs_unlock_up_safe(path, path->lowest_level + 1);
4709 memcpy(min_key, &found_key, sizeof(found_key));
4710 }
4711 return ret;
4712}
4713
4714/*
4715 * this is similar to btrfs_next_leaf, but does not try to preserve
4716 * and fixup the path. It looks for and returns the next key in the
4717 * tree based on the current path and the min_trans parameters.
4718 *
4719 * 0 is returned if another key is found, < 0 if there are any errors
4720 * and 1 is returned if there are no higher keys in the tree
4721 *
4722 * path->keep_locks should be set to 1 on the search made before
4723 * calling this function.
4724 */
4725int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4726 struct btrfs_key *key, int level, u64 min_trans)
4727{
4728 int slot;
4729 struct extent_buffer *c;
4730
4731 WARN_ON(!path->keep_locks && !path->skip_locking);
4732 while (level < BTRFS_MAX_LEVEL) {
4733 if (!path->nodes[level])
4734 return 1;
4735
4736 slot = path->slots[level] + 1;
4737 c = path->nodes[level];
4738next:
4739 if (slot >= btrfs_header_nritems(c)) {
4740 int ret;
4741 int orig_lowest;
4742 struct btrfs_key cur_key;
4743 if (level + 1 >= BTRFS_MAX_LEVEL ||
4744 !path->nodes[level + 1])
4745 return 1;
4746
4747 if (path->locks[level + 1] || path->skip_locking) {
4748 level++;
4749 continue;
4750 }
4751
4752 slot = btrfs_header_nritems(c) - 1;
4753 if (level == 0)
4754 btrfs_item_key_to_cpu(c, &cur_key, slot);
4755 else
4756 btrfs_node_key_to_cpu(c, &cur_key, slot);
4757
4758 orig_lowest = path->lowest_level;
4759 btrfs_release_path(path);
4760 path->lowest_level = level;
4761 ret = btrfs_search_slot(NULL, root, &cur_key, path,
4762 0, 0);
4763 path->lowest_level = orig_lowest;
4764 if (ret < 0)
4765 return ret;
4766
4767 c = path->nodes[level];
4768 slot = path->slots[level];
4769 if (ret == 0)
4770 slot++;
4771 goto next;
4772 }
4773
4774 if (level == 0)
4775 btrfs_item_key_to_cpu(c, key, slot);
4776 else {
4777 u64 gen = btrfs_node_ptr_generation(c, slot);
4778
4779 if (gen < min_trans) {
4780 slot++;
4781 goto next;
4782 }
4783 btrfs_node_key_to_cpu(c, key, slot);
4784 }
4785 return 0;
4786 }
4787 return 1;
4788}
4789
4790int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
4791 u64 time_seq)
4792{
4793 int slot;
4794 int level;
4795 struct extent_buffer *c;
4796 struct extent_buffer *next;
4797 struct btrfs_fs_info *fs_info = root->fs_info;
4798 struct btrfs_key key;
4799 bool need_commit_sem = false;
4800 u32 nritems;
4801 int ret;
4802 int i;
4803
4804 /*
4805 * The nowait semantics are used only for write paths, where we don't
4806 * use the tree mod log and sequence numbers.
4807 */
4808 if (time_seq)
4809 ASSERT(!path->nowait);
4810
4811 nritems = btrfs_header_nritems(path->nodes[0]);
4812 if (nritems == 0)
4813 return 1;
4814
4815 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4816again:
4817 level = 1;
4818 next = NULL;
4819 btrfs_release_path(path);
4820
4821 path->keep_locks = 1;
4822
4823 if (time_seq) {
4824 ret = btrfs_search_old_slot(root, &key, path, time_seq);
4825 } else {
4826 if (path->need_commit_sem) {
4827 path->need_commit_sem = 0;
4828 need_commit_sem = true;
4829 if (path->nowait) {
4830 if (!down_read_trylock(&fs_info->commit_root_sem)) {
4831 ret = -EAGAIN;
4832 goto done;
4833 }
4834 } else {
4835 down_read(&fs_info->commit_root_sem);
4836 }
4837 }
4838 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4839 }
4840 path->keep_locks = 0;
4841
4842 if (ret < 0)
4843 goto done;
4844
4845 nritems = btrfs_header_nritems(path->nodes[0]);
4846 /*
4847 * by releasing the path above we dropped all our locks. A balance
4848 * could have added more items next to the key that used to be
4849 * at the very end of the block. So, check again here and
4850 * advance the path if there are now more items available.
4851 */
4852 if (nritems > 0 && path->slots[0] < nritems - 1) {
4853 if (ret == 0)
4854 path->slots[0]++;
4855 ret = 0;
4856 goto done;
4857 }
4858 /*
4859 * So the above check misses one case:
4860 * - after releasing the path above, someone has removed the item that
4861 * used to be at the very end of the block, and balance between leafs
4862 * gets another one with bigger key.offset to replace it.
4863 *
4864 * This one should be returned as well, or we can get leaf corruption
4865 * later(esp. in __btrfs_drop_extents()).
4866 *
4867 * And a bit more explanation about this check,
4868 * with ret > 0, the key isn't found, the path points to the slot
4869 * where it should be inserted, so the path->slots[0] item must be the
4870 * bigger one.
4871 */
4872 if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
4873 ret = 0;
4874 goto done;
4875 }
4876
4877 while (level < BTRFS_MAX_LEVEL) {
4878 if (!path->nodes[level]) {
4879 ret = 1;
4880 goto done;
4881 }
4882
4883 slot = path->slots[level] + 1;
4884 c = path->nodes[level];
4885 if (slot >= btrfs_header_nritems(c)) {
4886 level++;
4887 if (level == BTRFS_MAX_LEVEL) {
4888 ret = 1;
4889 goto done;
4890 }
4891 continue;
4892 }
4893
4894
4895 /*
4896 * Our current level is where we're going to start from, and to
4897 * make sure lockdep doesn't complain we need to drop our locks
4898 * and nodes from 0 to our current level.
4899 */
4900 for (i = 0; i < level; i++) {
4901 if (path->locks[level]) {
4902 btrfs_tree_read_unlock(path->nodes[i]);
4903 path->locks[i] = 0;
4904 }
4905 free_extent_buffer(path->nodes[i]);
4906 path->nodes[i] = NULL;
4907 }
4908
4909 next = c;
4910 ret = read_block_for_search(root, path, &next, level,
4911 slot, &key);
4912 if (ret == -EAGAIN && !path->nowait)
4913 goto again;
4914
4915 if (ret < 0) {
4916 btrfs_release_path(path);
4917 goto done;
4918 }
4919
4920 if (!path->skip_locking) {
4921 ret = btrfs_try_tree_read_lock(next);
4922 if (!ret && path->nowait) {
4923 ret = -EAGAIN;
4924 goto done;
4925 }
4926 if (!ret && time_seq) {
4927 /*
4928 * If we don't get the lock, we may be racing
4929 * with push_leaf_left, holding that lock while
4930 * itself waiting for the leaf we've currently
4931 * locked. To solve this situation, we give up
4932 * on our lock and cycle.
4933 */
4934 free_extent_buffer(next);
4935 btrfs_release_path(path);
4936 cond_resched();
4937 goto again;
4938 }
4939 if (!ret)
4940 btrfs_tree_read_lock(next);
4941 }
4942 break;
4943 }
4944 path->slots[level] = slot;
4945 while (1) {
4946 level--;
4947 path->nodes[level] = next;
4948 path->slots[level] = 0;
4949 if (!path->skip_locking)
4950 path->locks[level] = BTRFS_READ_LOCK;
4951 if (!level)
4952 break;
4953
4954 ret = read_block_for_search(root, path, &next, level,
4955 0, &key);
4956 if (ret == -EAGAIN && !path->nowait)
4957 goto again;
4958
4959 if (ret < 0) {
4960 btrfs_release_path(path);
4961 goto done;
4962 }
4963
4964 if (!path->skip_locking) {
4965 if (path->nowait) {
4966 if (!btrfs_try_tree_read_lock(next)) {
4967 ret = -EAGAIN;
4968 goto done;
4969 }
4970 } else {
4971 btrfs_tree_read_lock(next);
4972 }
4973 }
4974 }
4975 ret = 0;
4976done:
4977 unlock_up(path, 0, 1, 0, NULL);
4978 if (need_commit_sem) {
4979 int ret2;
4980
4981 path->need_commit_sem = 1;
4982 ret2 = finish_need_commit_sem_search(path);
4983 up_read(&fs_info->commit_root_sem);
4984 if (ret2)
4985 ret = ret2;
4986 }
4987
4988 return ret;
4989}
4990
4991int btrfs_next_old_item(struct btrfs_root *root, struct btrfs_path *path, u64 time_seq)
4992{
4993 path->slots[0]++;
4994 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0]))
4995 return btrfs_next_old_leaf(root, path, time_seq);
4996 return 0;
4997}
4998
4999/*
5000 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5001 * searching until it gets past min_objectid or finds an item of 'type'
5002 *
5003 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5004 */
5005int btrfs_previous_item(struct btrfs_root *root,
5006 struct btrfs_path *path, u64 min_objectid,
5007 int type)
5008{
5009 struct btrfs_key found_key;
5010 struct extent_buffer *leaf;
5011 u32 nritems;
5012 int ret;
5013
5014 while (1) {
5015 if (path->slots[0] == 0) {
5016 ret = btrfs_prev_leaf(root, path);
5017 if (ret != 0)
5018 return ret;
5019 } else {
5020 path->slots[0]--;
5021 }
5022 leaf = path->nodes[0];
5023 nritems = btrfs_header_nritems(leaf);
5024 if (nritems == 0)
5025 return 1;
5026 if (path->slots[0] == nritems)
5027 path->slots[0]--;
5028
5029 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5030 if (found_key.objectid < min_objectid)
5031 break;
5032 if (found_key.type == type)
5033 return 0;
5034 if (found_key.objectid == min_objectid &&
5035 found_key.type < type)
5036 break;
5037 }
5038 return 1;
5039}
5040
5041/*
5042 * search in extent tree to find a previous Metadata/Data extent item with
5043 * min objecitd.
5044 *
5045 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5046 */
5047int btrfs_previous_extent_item(struct btrfs_root *root,
5048 struct btrfs_path *path, u64 min_objectid)
5049{
5050 struct btrfs_key found_key;
5051 struct extent_buffer *leaf;
5052 u32 nritems;
5053 int ret;
5054
5055 while (1) {
5056 if (path->slots[0] == 0) {
5057 ret = btrfs_prev_leaf(root, path);
5058 if (ret != 0)
5059 return ret;
5060 } else {
5061 path->slots[0]--;
5062 }
5063 leaf = path->nodes[0];
5064 nritems = btrfs_header_nritems(leaf);
5065 if (nritems == 0)
5066 return 1;
5067 if (path->slots[0] == nritems)
5068 path->slots[0]--;
5069
5070 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5071 if (found_key.objectid < min_objectid)
5072 break;
5073 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5074 found_key.type == BTRFS_METADATA_ITEM_KEY)
5075 return 0;
5076 if (found_key.objectid == min_objectid &&
5077 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5078 break;
5079 }
5080 return 1;
5081}
5082
5083int __init btrfs_ctree_init(void)
5084{
5085 btrfs_path_cachep = kmem_cache_create("btrfs_path",
5086 sizeof(struct btrfs_path), 0,
5087 SLAB_MEM_SPREAD, NULL);
5088 if (!btrfs_path_cachep)
5089 return -ENOMEM;
5090 return 0;
5091}
5092
5093void __cold btrfs_ctree_exit(void)
5094{
5095 kmem_cache_destroy(btrfs_path_cachep);
5096}
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 <linux/rbtree.h>
22#include "ctree.h"
23#include "disk-io.h"
24#include "transaction.h"
25#include "print-tree.h"
26#include "locking.h"
27
28static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
29 *root, struct btrfs_path *path, int level);
30static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
31 *root, struct btrfs_key *ins_key,
32 struct btrfs_path *path, int data_size, int extend);
33static int push_node_left(struct btrfs_trans_handle *trans,
34 struct btrfs_root *root, struct extent_buffer *dst,
35 struct extent_buffer *src, int empty);
36static int balance_node_right(struct btrfs_trans_handle *trans,
37 struct btrfs_root *root,
38 struct extent_buffer *dst_buf,
39 struct extent_buffer *src_buf);
40static void del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
41 struct btrfs_path *path, int level, int slot,
42 int tree_mod_log);
43static void tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
44 struct extent_buffer *eb);
45struct extent_buffer *read_old_tree_block(struct btrfs_root *root, u64 bytenr,
46 u32 blocksize, u64 parent_transid,
47 u64 time_seq);
48struct extent_buffer *btrfs_find_old_tree_block(struct btrfs_root *root,
49 u64 bytenr, u32 blocksize,
50 u64 time_seq);
51
52struct btrfs_path *btrfs_alloc_path(void)
53{
54 struct btrfs_path *path;
55 path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
56 return path;
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 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
70 if (p->locks[i] == BTRFS_READ_LOCK)
71 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
72 else if (p->locks[i] == BTRFS_WRITE_LOCK)
73 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
74 }
75}
76
77/*
78 * reset all the locked nodes in the patch to spinning locks.
79 *
80 * held is used to keep lockdep happy, when lockdep is enabled
81 * we set held to a blocking lock before we go around and
82 * retake all the spinlocks in the path. You can safely use NULL
83 * for held
84 */
85noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
86 struct extent_buffer *held, int held_rw)
87{
88 int i;
89
90#ifdef CONFIG_DEBUG_LOCK_ALLOC
91 /* lockdep really cares that we take all of these spinlocks
92 * in the right order. If any of the locks in the path are not
93 * currently blocking, it is going to complain. So, make really
94 * really sure by forcing the path to blocking before we clear
95 * the path blocking.
96 */
97 if (held) {
98 btrfs_set_lock_blocking_rw(held, held_rw);
99 if (held_rw == BTRFS_WRITE_LOCK)
100 held_rw = BTRFS_WRITE_LOCK_BLOCKING;
101 else if (held_rw == BTRFS_READ_LOCK)
102 held_rw = BTRFS_READ_LOCK_BLOCKING;
103 }
104 btrfs_set_path_blocking(p);
105#endif
106
107 for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
108 if (p->nodes[i] && p->locks[i]) {
109 btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
110 if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
111 p->locks[i] = BTRFS_WRITE_LOCK;
112 else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
113 p->locks[i] = BTRFS_READ_LOCK;
114 }
115 }
116
117#ifdef CONFIG_DEBUG_LOCK_ALLOC
118 if (held)
119 btrfs_clear_lock_blocking_rw(held, held_rw);
120#endif
121}
122
123/* this also releases the path */
124void btrfs_free_path(struct btrfs_path *p)
125{
126 if (!p)
127 return;
128 btrfs_release_path(p);
129 kmem_cache_free(btrfs_path_cachep, p);
130}
131
132/*
133 * path release drops references on the extent buffers in the path
134 * and it drops any locks held by this path
135 *
136 * It is safe to call this on paths that no locks or extent buffers held.
137 */
138noinline void btrfs_release_path(struct btrfs_path *p)
139{
140 int i;
141
142 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
143 p->slots[i] = 0;
144 if (!p->nodes[i])
145 continue;
146 if (p->locks[i]) {
147 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
148 p->locks[i] = 0;
149 }
150 free_extent_buffer(p->nodes[i]);
151 p->nodes[i] = NULL;
152 }
153}
154
155/*
156 * safely gets a reference on the root node of a tree. A lock
157 * is not taken, so a concurrent writer may put a different node
158 * at the root of the tree. See btrfs_lock_root_node for the
159 * looping required.
160 *
161 * The extent buffer returned by this has a reference taken, so
162 * it won't disappear. It may stop being the root of the tree
163 * at any time because there are no locks held.
164 */
165struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
166{
167 struct extent_buffer *eb;
168
169 while (1) {
170 rcu_read_lock();
171 eb = rcu_dereference(root->node);
172
173 /*
174 * RCU really hurts here, we could free up the root node because
175 * it was cow'ed but we may not get the new root node yet so do
176 * the inc_not_zero dance and if it doesn't work then
177 * synchronize_rcu and try again.
178 */
179 if (atomic_inc_not_zero(&eb->refs)) {
180 rcu_read_unlock();
181 break;
182 }
183 rcu_read_unlock();
184 synchronize_rcu();
185 }
186 return eb;
187}
188
189/* loop around taking references on and locking the root node of the
190 * tree until you end up with a lock on the root. A locked buffer
191 * is returned, with a reference held.
192 */
193struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
194{
195 struct extent_buffer *eb;
196
197 while (1) {
198 eb = btrfs_root_node(root);
199 btrfs_tree_lock(eb);
200 if (eb == root->node)
201 break;
202 btrfs_tree_unlock(eb);
203 free_extent_buffer(eb);
204 }
205 return eb;
206}
207
208/* loop around taking references on and locking the root node of the
209 * tree until you end up with a lock on the root. A locked buffer
210 * is returned, with a reference held.
211 */
212struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
213{
214 struct extent_buffer *eb;
215
216 while (1) {
217 eb = btrfs_root_node(root);
218 btrfs_tree_read_lock(eb);
219 if (eb == root->node)
220 break;
221 btrfs_tree_read_unlock(eb);
222 free_extent_buffer(eb);
223 }
224 return eb;
225}
226
227/* cowonly root (everything not a reference counted cow subvolume), just get
228 * put onto a simple dirty list. transaction.c walks this to make sure they
229 * get properly updated on disk.
230 */
231static void add_root_to_dirty_list(struct btrfs_root *root)
232{
233 spin_lock(&root->fs_info->trans_lock);
234 if (root->track_dirty && list_empty(&root->dirty_list)) {
235 list_add(&root->dirty_list,
236 &root->fs_info->dirty_cowonly_roots);
237 }
238 spin_unlock(&root->fs_info->trans_lock);
239}
240
241/*
242 * used by snapshot creation to make a copy of a root for a tree with
243 * a given objectid. The buffer with the new root node is returned in
244 * cow_ret, and this func returns zero on success or a negative error code.
245 */
246int btrfs_copy_root(struct btrfs_trans_handle *trans,
247 struct btrfs_root *root,
248 struct extent_buffer *buf,
249 struct extent_buffer **cow_ret, u64 new_root_objectid)
250{
251 struct extent_buffer *cow;
252 int ret = 0;
253 int level;
254 struct btrfs_disk_key disk_key;
255
256 WARN_ON(root->ref_cows && trans->transid !=
257 root->fs_info->running_transaction->transid);
258 WARN_ON(root->ref_cows && trans->transid != root->last_trans);
259
260 level = btrfs_header_level(buf);
261 if (level == 0)
262 btrfs_item_key(buf, &disk_key, 0);
263 else
264 btrfs_node_key(buf, &disk_key, 0);
265
266 cow = btrfs_alloc_free_block(trans, root, buf->len, 0,
267 new_root_objectid, &disk_key, level,
268 buf->start, 0);
269 if (IS_ERR(cow))
270 return PTR_ERR(cow);
271
272 copy_extent_buffer(cow, buf, 0, 0, cow->len);
273 btrfs_set_header_bytenr(cow, cow->start);
274 btrfs_set_header_generation(cow, trans->transid);
275 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
276 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
277 BTRFS_HEADER_FLAG_RELOC);
278 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
279 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
280 else
281 btrfs_set_header_owner(cow, new_root_objectid);
282
283 write_extent_buffer(cow, root->fs_info->fsid,
284 (unsigned long)btrfs_header_fsid(cow),
285 BTRFS_FSID_SIZE);
286
287 WARN_ON(btrfs_header_generation(buf) > trans->transid);
288 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
289 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
290 else
291 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
292
293 if (ret)
294 return ret;
295
296 btrfs_mark_buffer_dirty(cow);
297 *cow_ret = cow;
298 return 0;
299}
300
301enum mod_log_op {
302 MOD_LOG_KEY_REPLACE,
303 MOD_LOG_KEY_ADD,
304 MOD_LOG_KEY_REMOVE,
305 MOD_LOG_KEY_REMOVE_WHILE_FREEING,
306 MOD_LOG_KEY_REMOVE_WHILE_MOVING,
307 MOD_LOG_MOVE_KEYS,
308 MOD_LOG_ROOT_REPLACE,
309};
310
311struct tree_mod_move {
312 int dst_slot;
313 int nr_items;
314};
315
316struct tree_mod_root {
317 u64 logical;
318 u8 level;
319};
320
321struct tree_mod_elem {
322 struct rb_node node;
323 u64 index; /* shifted logical */
324 struct seq_list elem;
325 enum mod_log_op op;
326
327 /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
328 int slot;
329
330 /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
331 u64 generation;
332
333 /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
334 struct btrfs_disk_key key;
335 u64 blockptr;
336
337 /* this is used for op == MOD_LOG_MOVE_KEYS */
338 struct tree_mod_move move;
339
340 /* this is used for op == MOD_LOG_ROOT_REPLACE */
341 struct tree_mod_root old_root;
342};
343
344static inline void
345__get_tree_mod_seq(struct btrfs_fs_info *fs_info, struct seq_list *elem)
346{
347 elem->seq = atomic_inc_return(&fs_info->tree_mod_seq);
348 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
349}
350
351void btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
352 struct seq_list *elem)
353{
354 elem->flags = 1;
355 spin_lock(&fs_info->tree_mod_seq_lock);
356 __get_tree_mod_seq(fs_info, elem);
357 spin_unlock(&fs_info->tree_mod_seq_lock);
358}
359
360void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
361 struct seq_list *elem)
362{
363 struct rb_root *tm_root;
364 struct rb_node *node;
365 struct rb_node *next;
366 struct seq_list *cur_elem;
367 struct tree_mod_elem *tm;
368 u64 min_seq = (u64)-1;
369 u64 seq_putting = elem->seq;
370
371 if (!seq_putting)
372 return;
373
374 BUG_ON(!(elem->flags & 1));
375 spin_lock(&fs_info->tree_mod_seq_lock);
376 list_del(&elem->list);
377
378 list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
379 if ((cur_elem->flags & 1) && cur_elem->seq < min_seq) {
380 if (seq_putting > cur_elem->seq) {
381 /*
382 * blocker with lower sequence number exists, we
383 * cannot remove anything from the log
384 */
385 goto out;
386 }
387 min_seq = cur_elem->seq;
388 }
389 }
390
391 /*
392 * anything that's lower than the lowest existing (read: blocked)
393 * sequence number can be removed from the tree.
394 */
395 write_lock(&fs_info->tree_mod_log_lock);
396 tm_root = &fs_info->tree_mod_log;
397 for (node = rb_first(tm_root); node; node = next) {
398 next = rb_next(node);
399 tm = container_of(node, struct tree_mod_elem, node);
400 if (tm->elem.seq > min_seq)
401 continue;
402 rb_erase(node, tm_root);
403 list_del(&tm->elem.list);
404 kfree(tm);
405 }
406 write_unlock(&fs_info->tree_mod_log_lock);
407out:
408 spin_unlock(&fs_info->tree_mod_seq_lock);
409}
410
411/*
412 * key order of the log:
413 * index -> sequence
414 *
415 * the index is the shifted logical of the *new* root node for root replace
416 * operations, or the shifted logical of the affected block for all other
417 * operations.
418 */
419static noinline int
420__tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
421{
422 struct rb_root *tm_root;
423 struct rb_node **new;
424 struct rb_node *parent = NULL;
425 struct tree_mod_elem *cur;
426 int ret = 0;
427
428 BUG_ON(!tm || !tm->elem.seq);
429
430 write_lock(&fs_info->tree_mod_log_lock);
431 tm_root = &fs_info->tree_mod_log;
432 new = &tm_root->rb_node;
433 while (*new) {
434 cur = container_of(*new, struct tree_mod_elem, node);
435 parent = *new;
436 if (cur->index < tm->index)
437 new = &((*new)->rb_left);
438 else if (cur->index > tm->index)
439 new = &((*new)->rb_right);
440 else if (cur->elem.seq < tm->elem.seq)
441 new = &((*new)->rb_left);
442 else if (cur->elem.seq > tm->elem.seq)
443 new = &((*new)->rb_right);
444 else {
445 kfree(tm);
446 ret = -EEXIST;
447 goto unlock;
448 }
449 }
450
451 rb_link_node(&tm->node, parent, new);
452 rb_insert_color(&tm->node, tm_root);
453unlock:
454 write_unlock(&fs_info->tree_mod_log_lock);
455 return ret;
456}
457
458static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
459 struct extent_buffer *eb) {
460 smp_mb();
461 if (list_empty(&(fs_info)->tree_mod_seq_list))
462 return 1;
463 if (!eb)
464 return 0;
465 if (btrfs_header_level(eb) == 0)
466 return 1;
467 return 0;
468}
469
470/*
471 * This allocates memory and gets a tree modification sequence number when
472 * needed.
473 *
474 * Returns 0 when no sequence number is needed, < 0 on error.
475 * Returns 1 when a sequence number was added. In this case,
476 * fs_info->tree_mod_seq_lock was acquired and must be released by the caller
477 * after inserting into the rb tree.
478 */
479static inline int tree_mod_alloc(struct btrfs_fs_info *fs_info, gfp_t flags,
480 struct tree_mod_elem **tm_ret)
481{
482 struct tree_mod_elem *tm;
483 int seq;
484
485 if (tree_mod_dont_log(fs_info, NULL))
486 return 0;
487
488 tm = *tm_ret = kzalloc(sizeof(*tm), flags);
489 if (!tm)
490 return -ENOMEM;
491
492 tm->elem.flags = 0;
493 spin_lock(&fs_info->tree_mod_seq_lock);
494 if (list_empty(&fs_info->tree_mod_seq_list)) {
495 /*
496 * someone emptied the list while we were waiting for the lock.
497 * we must not add to the list, because no blocker exists. items
498 * are removed from the list only when the existing blocker is
499 * removed from the list.
500 */
501 kfree(tm);
502 seq = 0;
503 spin_unlock(&fs_info->tree_mod_seq_lock);
504 } else {
505 __get_tree_mod_seq(fs_info, &tm->elem);
506 seq = tm->elem.seq;
507 }
508
509 return seq;
510}
511
512static noinline int
513tree_mod_log_insert_key_mask(struct btrfs_fs_info *fs_info,
514 struct extent_buffer *eb, int slot,
515 enum mod_log_op op, gfp_t flags)
516{
517 struct tree_mod_elem *tm;
518 int ret;
519
520 ret = tree_mod_alloc(fs_info, flags, &tm);
521 if (ret <= 0)
522 return ret;
523
524 tm->index = eb->start >> PAGE_CACHE_SHIFT;
525 if (op != MOD_LOG_KEY_ADD) {
526 btrfs_node_key(eb, &tm->key, slot);
527 tm->blockptr = btrfs_node_blockptr(eb, slot);
528 }
529 tm->op = op;
530 tm->slot = slot;
531 tm->generation = btrfs_node_ptr_generation(eb, slot);
532
533 ret = __tree_mod_log_insert(fs_info, tm);
534 spin_unlock(&fs_info->tree_mod_seq_lock);
535 return ret;
536}
537
538static noinline int
539tree_mod_log_insert_key(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
540 int slot, enum mod_log_op op)
541{
542 return tree_mod_log_insert_key_mask(fs_info, eb, slot, op, GFP_NOFS);
543}
544
545static noinline int
546tree_mod_log_insert_move(struct btrfs_fs_info *fs_info,
547 struct extent_buffer *eb, int dst_slot, int src_slot,
548 int nr_items, gfp_t flags)
549{
550 struct tree_mod_elem *tm;
551 int ret;
552 int i;
553
554 if (tree_mod_dont_log(fs_info, eb))
555 return 0;
556
557 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
558 ret = tree_mod_log_insert_key(fs_info, eb, i + dst_slot,
559 MOD_LOG_KEY_REMOVE_WHILE_MOVING);
560 BUG_ON(ret < 0);
561 }
562
563 ret = tree_mod_alloc(fs_info, flags, &tm);
564 if (ret <= 0)
565 return ret;
566
567 tm->index = eb->start >> PAGE_CACHE_SHIFT;
568 tm->slot = src_slot;
569 tm->move.dst_slot = dst_slot;
570 tm->move.nr_items = nr_items;
571 tm->op = MOD_LOG_MOVE_KEYS;
572
573 ret = __tree_mod_log_insert(fs_info, tm);
574 spin_unlock(&fs_info->tree_mod_seq_lock);
575 return ret;
576}
577
578static noinline int
579tree_mod_log_insert_root(struct btrfs_fs_info *fs_info,
580 struct extent_buffer *old_root,
581 struct extent_buffer *new_root, gfp_t flags)
582{
583 struct tree_mod_elem *tm;
584 int ret;
585
586 ret = tree_mod_alloc(fs_info, flags, &tm);
587 if (ret <= 0)
588 return ret;
589
590 tm->index = new_root->start >> PAGE_CACHE_SHIFT;
591 tm->old_root.logical = old_root->start;
592 tm->old_root.level = btrfs_header_level(old_root);
593 tm->generation = btrfs_header_generation(old_root);
594 tm->op = MOD_LOG_ROOT_REPLACE;
595
596 ret = __tree_mod_log_insert(fs_info, tm);
597 spin_unlock(&fs_info->tree_mod_seq_lock);
598 return ret;
599}
600
601static struct tree_mod_elem *
602__tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
603 int smallest)
604{
605 struct rb_root *tm_root;
606 struct rb_node *node;
607 struct tree_mod_elem *cur = NULL;
608 struct tree_mod_elem *found = NULL;
609 u64 index = start >> PAGE_CACHE_SHIFT;
610
611 read_lock(&fs_info->tree_mod_log_lock);
612 tm_root = &fs_info->tree_mod_log;
613 node = tm_root->rb_node;
614 while (node) {
615 cur = container_of(node, struct tree_mod_elem, node);
616 if (cur->index < index) {
617 node = node->rb_left;
618 } else if (cur->index > index) {
619 node = node->rb_right;
620 } else if (cur->elem.seq < min_seq) {
621 node = node->rb_left;
622 } else if (!smallest) {
623 /* we want the node with the highest seq */
624 if (found)
625 BUG_ON(found->elem.seq > cur->elem.seq);
626 found = cur;
627 node = node->rb_left;
628 } else if (cur->elem.seq > min_seq) {
629 /* we want the node with the smallest seq */
630 if (found)
631 BUG_ON(found->elem.seq < cur->elem.seq);
632 found = cur;
633 node = node->rb_right;
634 } else {
635 found = cur;
636 break;
637 }
638 }
639 read_unlock(&fs_info->tree_mod_log_lock);
640
641 return found;
642}
643
644/*
645 * this returns the element from the log with the smallest time sequence
646 * value that's in the log (the oldest log item). any element with a time
647 * sequence lower than min_seq will be ignored.
648 */
649static struct tree_mod_elem *
650tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
651 u64 min_seq)
652{
653 return __tree_mod_log_search(fs_info, start, min_seq, 1);
654}
655
656/*
657 * this returns the element from the log with the largest time sequence
658 * value that's in the log (the most recent log item). any element with
659 * a time sequence lower than min_seq will be ignored.
660 */
661static struct tree_mod_elem *
662tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
663{
664 return __tree_mod_log_search(fs_info, start, min_seq, 0);
665}
666
667static inline void
668tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
669 struct extent_buffer *src, unsigned long dst_offset,
670 unsigned long src_offset, int nr_items)
671{
672 int ret;
673 int i;
674
675 if (tree_mod_dont_log(fs_info, NULL))
676 return;
677
678 if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
679 return;
680
681 /* speed this up by single seq for all operations? */
682 for (i = 0; i < nr_items; i++) {
683 ret = tree_mod_log_insert_key(fs_info, src, i + src_offset,
684 MOD_LOG_KEY_REMOVE);
685 BUG_ON(ret < 0);
686 ret = tree_mod_log_insert_key(fs_info, dst, i + dst_offset,
687 MOD_LOG_KEY_ADD);
688 BUG_ON(ret < 0);
689 }
690}
691
692static inline void
693tree_mod_log_eb_move(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
694 int dst_offset, int src_offset, int nr_items)
695{
696 int ret;
697 ret = tree_mod_log_insert_move(fs_info, dst, dst_offset, src_offset,
698 nr_items, GFP_NOFS);
699 BUG_ON(ret < 0);
700}
701
702static inline void
703tree_mod_log_set_node_key(struct btrfs_fs_info *fs_info,
704 struct extent_buffer *eb,
705 struct btrfs_disk_key *disk_key, int slot, int atomic)
706{
707 int ret;
708
709 ret = tree_mod_log_insert_key_mask(fs_info, eb, slot,
710 MOD_LOG_KEY_REPLACE,
711 atomic ? GFP_ATOMIC : GFP_NOFS);
712 BUG_ON(ret < 0);
713}
714
715static void tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
716 struct extent_buffer *eb)
717{
718 int i;
719 int ret;
720 u32 nritems;
721
722 if (tree_mod_dont_log(fs_info, eb))
723 return;
724
725 nritems = btrfs_header_nritems(eb);
726 for (i = nritems - 1; i >= 0; i--) {
727 ret = tree_mod_log_insert_key(fs_info, eb, i,
728 MOD_LOG_KEY_REMOVE_WHILE_FREEING);
729 BUG_ON(ret < 0);
730 }
731}
732
733static inline void
734tree_mod_log_set_root_pointer(struct btrfs_root *root,
735 struct extent_buffer *new_root_node)
736{
737 int ret;
738 tree_mod_log_free_eb(root->fs_info, root->node);
739 ret = tree_mod_log_insert_root(root->fs_info, root->node,
740 new_root_node, GFP_NOFS);
741 BUG_ON(ret < 0);
742}
743
744/*
745 * check if the tree block can be shared by multiple trees
746 */
747int btrfs_block_can_be_shared(struct btrfs_root *root,
748 struct extent_buffer *buf)
749{
750 /*
751 * Tree blocks not in refernece counted trees and tree roots
752 * are never shared. If a block was allocated after the last
753 * snapshot and the block was not allocated by tree relocation,
754 * we know the block is not shared.
755 */
756 if (root->ref_cows &&
757 buf != root->node && buf != root->commit_root &&
758 (btrfs_header_generation(buf) <=
759 btrfs_root_last_snapshot(&root->root_item) ||
760 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
761 return 1;
762#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
763 if (root->ref_cows &&
764 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
765 return 1;
766#endif
767 return 0;
768}
769
770static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
771 struct btrfs_root *root,
772 struct extent_buffer *buf,
773 struct extent_buffer *cow,
774 int *last_ref)
775{
776 u64 refs;
777 u64 owner;
778 u64 flags;
779 u64 new_flags = 0;
780 int ret;
781
782 /*
783 * Backrefs update rules:
784 *
785 * Always use full backrefs for extent pointers in tree block
786 * allocated by tree relocation.
787 *
788 * If a shared tree block is no longer referenced by its owner
789 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
790 * use full backrefs for extent pointers in tree block.
791 *
792 * If a tree block is been relocating
793 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
794 * use full backrefs for extent pointers in tree block.
795 * The reason for this is some operations (such as drop tree)
796 * are only allowed for blocks use full backrefs.
797 */
798
799 if (btrfs_block_can_be_shared(root, buf)) {
800 ret = btrfs_lookup_extent_info(trans, root, buf->start,
801 buf->len, &refs, &flags);
802 if (ret)
803 return ret;
804 if (refs == 0) {
805 ret = -EROFS;
806 btrfs_std_error(root->fs_info, ret);
807 return ret;
808 }
809 } else {
810 refs = 1;
811 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
812 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
813 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
814 else
815 flags = 0;
816 }
817
818 owner = btrfs_header_owner(buf);
819 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
820 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
821
822 if (refs > 1) {
823 if ((owner == root->root_key.objectid ||
824 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
825 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
826 ret = btrfs_inc_ref(trans, root, buf, 1, 1);
827 BUG_ON(ret); /* -ENOMEM */
828
829 if (root->root_key.objectid ==
830 BTRFS_TREE_RELOC_OBJECTID) {
831 ret = btrfs_dec_ref(trans, root, buf, 0, 1);
832 BUG_ON(ret); /* -ENOMEM */
833 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
834 BUG_ON(ret); /* -ENOMEM */
835 }
836 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
837 } else {
838
839 if (root->root_key.objectid ==
840 BTRFS_TREE_RELOC_OBJECTID)
841 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
842 else
843 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
844 BUG_ON(ret); /* -ENOMEM */
845 }
846 if (new_flags != 0) {
847 ret = btrfs_set_disk_extent_flags(trans, root,
848 buf->start,
849 buf->len,
850 new_flags, 0);
851 if (ret)
852 return ret;
853 }
854 } else {
855 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
856 if (root->root_key.objectid ==
857 BTRFS_TREE_RELOC_OBJECTID)
858 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
859 else
860 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
861 BUG_ON(ret); /* -ENOMEM */
862 ret = btrfs_dec_ref(trans, root, buf, 1, 1);
863 BUG_ON(ret); /* -ENOMEM */
864 }
865 /*
866 * don't log freeing in case we're freeing the root node, this
867 * is done by tree_mod_log_set_root_pointer later
868 */
869 if (buf != root->node && btrfs_header_level(buf) != 0)
870 tree_mod_log_free_eb(root->fs_info, buf);
871 clean_tree_block(trans, root, buf);
872 *last_ref = 1;
873 }
874 return 0;
875}
876
877/*
878 * does the dirty work in cow of a single block. The parent block (if
879 * supplied) is updated to point to the new cow copy. The new buffer is marked
880 * dirty and returned locked. If you modify the block it needs to be marked
881 * dirty again.
882 *
883 * search_start -- an allocation hint for the new block
884 *
885 * empty_size -- a hint that you plan on doing more cow. This is the size in
886 * bytes the allocator should try to find free next to the block it returns.
887 * This is just a hint and may be ignored by the allocator.
888 */
889static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
890 struct btrfs_root *root,
891 struct extent_buffer *buf,
892 struct extent_buffer *parent, int parent_slot,
893 struct extent_buffer **cow_ret,
894 u64 search_start, u64 empty_size)
895{
896 struct btrfs_disk_key disk_key;
897 struct extent_buffer *cow;
898 int level, ret;
899 int last_ref = 0;
900 int unlock_orig = 0;
901 u64 parent_start;
902
903 if (*cow_ret == buf)
904 unlock_orig = 1;
905
906 btrfs_assert_tree_locked(buf);
907
908 WARN_ON(root->ref_cows && trans->transid !=
909 root->fs_info->running_transaction->transid);
910 WARN_ON(root->ref_cows && trans->transid != root->last_trans);
911
912 level = btrfs_header_level(buf);
913
914 if (level == 0)
915 btrfs_item_key(buf, &disk_key, 0);
916 else
917 btrfs_node_key(buf, &disk_key, 0);
918
919 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
920 if (parent)
921 parent_start = parent->start;
922 else
923 parent_start = 0;
924 } else
925 parent_start = 0;
926
927 cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
928 root->root_key.objectid, &disk_key,
929 level, search_start, empty_size);
930 if (IS_ERR(cow))
931 return PTR_ERR(cow);
932
933 /* cow is set to blocking by btrfs_init_new_buffer */
934
935 copy_extent_buffer(cow, buf, 0, 0, cow->len);
936 btrfs_set_header_bytenr(cow, cow->start);
937 btrfs_set_header_generation(cow, trans->transid);
938 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
939 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
940 BTRFS_HEADER_FLAG_RELOC);
941 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
942 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
943 else
944 btrfs_set_header_owner(cow, root->root_key.objectid);
945
946 write_extent_buffer(cow, root->fs_info->fsid,
947 (unsigned long)btrfs_header_fsid(cow),
948 BTRFS_FSID_SIZE);
949
950 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
951 if (ret) {
952 btrfs_abort_transaction(trans, root, ret);
953 return ret;
954 }
955
956 if (root->ref_cows)
957 btrfs_reloc_cow_block(trans, root, buf, cow);
958
959 if (buf == root->node) {
960 WARN_ON(parent && parent != buf);
961 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
962 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
963 parent_start = buf->start;
964 else
965 parent_start = 0;
966
967 extent_buffer_get(cow);
968 tree_mod_log_set_root_pointer(root, cow);
969 rcu_assign_pointer(root->node, cow);
970
971 btrfs_free_tree_block(trans, root, buf, parent_start,
972 last_ref);
973 free_extent_buffer(buf);
974 add_root_to_dirty_list(root);
975 } else {
976 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
977 parent_start = parent->start;
978 else
979 parent_start = 0;
980
981 WARN_ON(trans->transid != btrfs_header_generation(parent));
982 tree_mod_log_insert_key(root->fs_info, parent, parent_slot,
983 MOD_LOG_KEY_REPLACE);
984 btrfs_set_node_blockptr(parent, parent_slot,
985 cow->start);
986 btrfs_set_node_ptr_generation(parent, parent_slot,
987 trans->transid);
988 btrfs_mark_buffer_dirty(parent);
989 btrfs_free_tree_block(trans, root, buf, parent_start,
990 last_ref);
991 }
992 if (unlock_orig)
993 btrfs_tree_unlock(buf);
994 free_extent_buffer_stale(buf);
995 btrfs_mark_buffer_dirty(cow);
996 *cow_ret = cow;
997 return 0;
998}
999
1000/*
1001 * returns the logical address of the oldest predecessor of the given root.
1002 * entries older than time_seq are ignored.
1003 */
1004static struct tree_mod_elem *
1005__tree_mod_log_oldest_root(struct btrfs_fs_info *fs_info,
1006 struct btrfs_root *root, u64 time_seq)
1007{
1008 struct tree_mod_elem *tm;
1009 struct tree_mod_elem *found = NULL;
1010 u64 root_logical = root->node->start;
1011 int looped = 0;
1012
1013 if (!time_seq)
1014 return 0;
1015
1016 /*
1017 * the very last operation that's logged for a root is the replacement
1018 * operation (if it is replaced at all). this has the index of the *new*
1019 * root, making it the very first operation that's logged for this root.
1020 */
1021 while (1) {
1022 tm = tree_mod_log_search_oldest(fs_info, root_logical,
1023 time_seq);
1024 if (!looped && !tm)
1025 return 0;
1026 /*
1027 * if there are no tree operation for the oldest root, we simply
1028 * return it. this should only happen if that (old) root is at
1029 * level 0.
1030 */
1031 if (!tm)
1032 break;
1033
1034 /*
1035 * if there's an operation that's not a root replacement, we
1036 * found the oldest version of our root. normally, we'll find a
1037 * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1038 */
1039 if (tm->op != MOD_LOG_ROOT_REPLACE)
1040 break;
1041
1042 found = tm;
1043 root_logical = tm->old_root.logical;
1044 BUG_ON(root_logical == root->node->start);
1045 looped = 1;
1046 }
1047
1048 /* if there's no old root to return, return what we found instead */
1049 if (!found)
1050 found = tm;
1051
1052 return found;
1053}
1054
1055/*
1056 * tm is a pointer to the first operation to rewind within eb. then, all
1057 * previous operations will be rewinded (until we reach something older than
1058 * time_seq).
1059 */
1060static void
1061__tree_mod_log_rewind(struct extent_buffer *eb, u64 time_seq,
1062 struct tree_mod_elem *first_tm)
1063{
1064 u32 n;
1065 struct rb_node *next;
1066 struct tree_mod_elem *tm = first_tm;
1067 unsigned long o_dst;
1068 unsigned long o_src;
1069 unsigned long p_size = sizeof(struct btrfs_key_ptr);
1070
1071 n = btrfs_header_nritems(eb);
1072 while (tm && tm->elem.seq >= time_seq) {
1073 /*
1074 * all the operations are recorded with the operator used for
1075 * the modification. as we're going backwards, we do the
1076 * opposite of each operation here.
1077 */
1078 switch (tm->op) {
1079 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1080 BUG_ON(tm->slot < n);
1081 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1082 case MOD_LOG_KEY_REMOVE:
1083 btrfs_set_node_key(eb, &tm->key, tm->slot);
1084 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1085 btrfs_set_node_ptr_generation(eb, tm->slot,
1086 tm->generation);
1087 n++;
1088 break;
1089 case MOD_LOG_KEY_REPLACE:
1090 BUG_ON(tm->slot >= n);
1091 btrfs_set_node_key(eb, &tm->key, tm->slot);
1092 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1093 btrfs_set_node_ptr_generation(eb, tm->slot,
1094 tm->generation);
1095 break;
1096 case MOD_LOG_KEY_ADD:
1097 /* if a move operation is needed it's in the log */
1098 n--;
1099 break;
1100 case MOD_LOG_MOVE_KEYS:
1101 o_dst = btrfs_node_key_ptr_offset(tm->slot);
1102 o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1103 memmove_extent_buffer(eb, o_dst, o_src,
1104 tm->move.nr_items * p_size);
1105 break;
1106 case MOD_LOG_ROOT_REPLACE:
1107 /*
1108 * this operation is special. for roots, this must be
1109 * handled explicitly before rewinding.
1110 * for non-roots, this operation may exist if the node
1111 * was a root: root A -> child B; then A gets empty and
1112 * B is promoted to the new root. in the mod log, we'll
1113 * have a root-replace operation for B, a tree block
1114 * that is no root. we simply ignore that operation.
1115 */
1116 break;
1117 }
1118 next = rb_next(&tm->node);
1119 if (!next)
1120 break;
1121 tm = container_of(next, struct tree_mod_elem, node);
1122 if (tm->index != first_tm->index)
1123 break;
1124 }
1125 btrfs_set_header_nritems(eb, n);
1126}
1127
1128static struct extent_buffer *
1129tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1130 u64 time_seq)
1131{
1132 struct extent_buffer *eb_rewin;
1133 struct tree_mod_elem *tm;
1134
1135 if (!time_seq)
1136 return eb;
1137
1138 if (btrfs_header_level(eb) == 0)
1139 return eb;
1140
1141 tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1142 if (!tm)
1143 return eb;
1144
1145 if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1146 BUG_ON(tm->slot != 0);
1147 eb_rewin = alloc_dummy_extent_buffer(eb->start,
1148 fs_info->tree_root->nodesize);
1149 BUG_ON(!eb_rewin);
1150 btrfs_set_header_bytenr(eb_rewin, eb->start);
1151 btrfs_set_header_backref_rev(eb_rewin,
1152 btrfs_header_backref_rev(eb));
1153 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1154 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1155 } else {
1156 eb_rewin = btrfs_clone_extent_buffer(eb);
1157 BUG_ON(!eb_rewin);
1158 }
1159
1160 extent_buffer_get(eb_rewin);
1161 free_extent_buffer(eb);
1162
1163 __tree_mod_log_rewind(eb_rewin, time_seq, tm);
1164
1165 return eb_rewin;
1166}
1167
1168/*
1169 * get_old_root() rewinds the state of @root's root node to the given @time_seq
1170 * value. If there are no changes, the current root->root_node is returned. If
1171 * anything changed in between, there's a fresh buffer allocated on which the
1172 * rewind operations are done. In any case, the returned buffer is read locked.
1173 * Returns NULL on error (with no locks held).
1174 */
1175static inline struct extent_buffer *
1176get_old_root(struct btrfs_root *root, u64 time_seq)
1177{
1178 struct tree_mod_elem *tm;
1179 struct extent_buffer *eb;
1180 struct tree_mod_root *old_root = NULL;
1181 u64 old_generation = 0;
1182 u64 logical;
1183
1184 eb = btrfs_read_lock_root_node(root);
1185 tm = __tree_mod_log_oldest_root(root->fs_info, root, time_seq);
1186 if (!tm)
1187 return root->node;
1188
1189 if (tm->op == MOD_LOG_ROOT_REPLACE) {
1190 old_root = &tm->old_root;
1191 old_generation = tm->generation;
1192 logical = old_root->logical;
1193 } else {
1194 logical = root->node->start;
1195 }
1196
1197 tm = tree_mod_log_search(root->fs_info, logical, time_seq);
1198 if (old_root)
1199 eb = alloc_dummy_extent_buffer(logical, root->nodesize);
1200 else
1201 eb = btrfs_clone_extent_buffer(root->node);
1202 btrfs_tree_read_unlock(root->node);
1203 free_extent_buffer(root->node);
1204 if (!eb)
1205 return NULL;
1206 btrfs_tree_read_lock(eb);
1207 if (old_root) {
1208 btrfs_set_header_bytenr(eb, eb->start);
1209 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1210 btrfs_set_header_owner(eb, root->root_key.objectid);
1211 btrfs_set_header_level(eb, old_root->level);
1212 btrfs_set_header_generation(eb, old_generation);
1213 }
1214 if (tm)
1215 __tree_mod_log_rewind(eb, time_seq, tm);
1216 else
1217 WARN_ON(btrfs_header_level(eb) != 0);
1218 extent_buffer_get(eb);
1219
1220 return eb;
1221}
1222
1223static inline int should_cow_block(struct btrfs_trans_handle *trans,
1224 struct btrfs_root *root,
1225 struct extent_buffer *buf)
1226{
1227 /* ensure we can see the force_cow */
1228 smp_rmb();
1229
1230 /*
1231 * We do not need to cow a block if
1232 * 1) this block is not created or changed in this transaction;
1233 * 2) this block does not belong to TREE_RELOC tree;
1234 * 3) the root is not forced COW.
1235 *
1236 * What is forced COW:
1237 * when we create snapshot during commiting the transaction,
1238 * after we've finished coping src root, we must COW the shared
1239 * block to ensure the metadata consistency.
1240 */
1241 if (btrfs_header_generation(buf) == trans->transid &&
1242 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1243 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1244 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1245 !root->force_cow)
1246 return 0;
1247 return 1;
1248}
1249
1250/*
1251 * cows a single block, see __btrfs_cow_block for the real work.
1252 * This version of it has extra checks so that a block isn't cow'd more than
1253 * once per transaction, as long as it hasn't been written yet
1254 */
1255noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1256 struct btrfs_root *root, struct extent_buffer *buf,
1257 struct extent_buffer *parent, int parent_slot,
1258 struct extent_buffer **cow_ret)
1259{
1260 u64 search_start;
1261 int ret;
1262
1263 if (trans->transaction != root->fs_info->running_transaction) {
1264 printk(KERN_CRIT "trans %llu running %llu\n",
1265 (unsigned long long)trans->transid,
1266 (unsigned long long)
1267 root->fs_info->running_transaction->transid);
1268 WARN_ON(1);
1269 }
1270 if (trans->transid != root->fs_info->generation) {
1271 printk(KERN_CRIT "trans %llu running %llu\n",
1272 (unsigned long long)trans->transid,
1273 (unsigned long long)root->fs_info->generation);
1274 WARN_ON(1);
1275 }
1276
1277 if (!should_cow_block(trans, root, buf)) {
1278 *cow_ret = buf;
1279 return 0;
1280 }
1281
1282 search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
1283
1284 if (parent)
1285 btrfs_set_lock_blocking(parent);
1286 btrfs_set_lock_blocking(buf);
1287
1288 ret = __btrfs_cow_block(trans, root, buf, parent,
1289 parent_slot, cow_ret, search_start, 0);
1290
1291 trace_btrfs_cow_block(root, buf, *cow_ret);
1292
1293 return ret;
1294}
1295
1296/*
1297 * helper function for defrag to decide if two blocks pointed to by a
1298 * node are actually close by
1299 */
1300static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1301{
1302 if (blocknr < other && other - (blocknr + blocksize) < 32768)
1303 return 1;
1304 if (blocknr > other && blocknr - (other + blocksize) < 32768)
1305 return 1;
1306 return 0;
1307}
1308
1309/*
1310 * compare two keys in a memcmp fashion
1311 */
1312static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
1313{
1314 struct btrfs_key k1;
1315
1316 btrfs_disk_key_to_cpu(&k1, disk);
1317
1318 return btrfs_comp_cpu_keys(&k1, k2);
1319}
1320
1321/*
1322 * same as comp_keys only with two btrfs_key's
1323 */
1324int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
1325{
1326 if (k1->objectid > k2->objectid)
1327 return 1;
1328 if (k1->objectid < k2->objectid)
1329 return -1;
1330 if (k1->type > k2->type)
1331 return 1;
1332 if (k1->type < k2->type)
1333 return -1;
1334 if (k1->offset > k2->offset)
1335 return 1;
1336 if (k1->offset < k2->offset)
1337 return -1;
1338 return 0;
1339}
1340
1341/*
1342 * this is used by the defrag code to go through all the
1343 * leaves pointed to by a node and reallocate them so that
1344 * disk order is close to key order
1345 */
1346int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1347 struct btrfs_root *root, struct extent_buffer *parent,
1348 int start_slot, int cache_only, u64 *last_ret,
1349 struct btrfs_key *progress)
1350{
1351 struct extent_buffer *cur;
1352 u64 blocknr;
1353 u64 gen;
1354 u64 search_start = *last_ret;
1355 u64 last_block = 0;
1356 u64 other;
1357 u32 parent_nritems;
1358 int end_slot;
1359 int i;
1360 int err = 0;
1361 int parent_level;
1362 int uptodate;
1363 u32 blocksize;
1364 int progress_passed = 0;
1365 struct btrfs_disk_key disk_key;
1366
1367 parent_level = btrfs_header_level(parent);
1368 if (cache_only && parent_level != 1)
1369 return 0;
1370
1371 if (trans->transaction != root->fs_info->running_transaction)
1372 WARN_ON(1);
1373 if (trans->transid != root->fs_info->generation)
1374 WARN_ON(1);
1375
1376 parent_nritems = btrfs_header_nritems(parent);
1377 blocksize = btrfs_level_size(root, parent_level - 1);
1378 end_slot = parent_nritems;
1379
1380 if (parent_nritems == 1)
1381 return 0;
1382
1383 btrfs_set_lock_blocking(parent);
1384
1385 for (i = start_slot; i < end_slot; i++) {
1386 int close = 1;
1387
1388 btrfs_node_key(parent, &disk_key, i);
1389 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1390 continue;
1391
1392 progress_passed = 1;
1393 blocknr = btrfs_node_blockptr(parent, i);
1394 gen = btrfs_node_ptr_generation(parent, i);
1395 if (last_block == 0)
1396 last_block = blocknr;
1397
1398 if (i > 0) {
1399 other = btrfs_node_blockptr(parent, i - 1);
1400 close = close_blocks(blocknr, other, blocksize);
1401 }
1402 if (!close && i < end_slot - 2) {
1403 other = btrfs_node_blockptr(parent, i + 1);
1404 close = close_blocks(blocknr, other, blocksize);
1405 }
1406 if (close) {
1407 last_block = blocknr;
1408 continue;
1409 }
1410
1411 cur = btrfs_find_tree_block(root, blocknr, blocksize);
1412 if (cur)
1413 uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1414 else
1415 uptodate = 0;
1416 if (!cur || !uptodate) {
1417 if (cache_only) {
1418 free_extent_buffer(cur);
1419 continue;
1420 }
1421 if (!cur) {
1422 cur = read_tree_block(root, blocknr,
1423 blocksize, gen);
1424 if (!cur)
1425 return -EIO;
1426 } else if (!uptodate) {
1427 err = btrfs_read_buffer(cur, gen);
1428 if (err) {
1429 free_extent_buffer(cur);
1430 return err;
1431 }
1432 }
1433 }
1434 if (search_start == 0)
1435 search_start = last_block;
1436
1437 btrfs_tree_lock(cur);
1438 btrfs_set_lock_blocking(cur);
1439 err = __btrfs_cow_block(trans, root, cur, parent, i,
1440 &cur, search_start,
1441 min(16 * blocksize,
1442 (end_slot - i) * blocksize));
1443 if (err) {
1444 btrfs_tree_unlock(cur);
1445 free_extent_buffer(cur);
1446 break;
1447 }
1448 search_start = cur->start;
1449 last_block = cur->start;
1450 *last_ret = search_start;
1451 btrfs_tree_unlock(cur);
1452 free_extent_buffer(cur);
1453 }
1454 return err;
1455}
1456
1457/*
1458 * The leaf data grows from end-to-front in the node.
1459 * this returns the address of the start of the last item,
1460 * which is the stop of the leaf data stack
1461 */
1462static inline unsigned int leaf_data_end(struct btrfs_root *root,
1463 struct extent_buffer *leaf)
1464{
1465 u32 nr = btrfs_header_nritems(leaf);
1466 if (nr == 0)
1467 return BTRFS_LEAF_DATA_SIZE(root);
1468 return btrfs_item_offset_nr(leaf, nr - 1);
1469}
1470
1471
1472/*
1473 * search for key in the extent_buffer. The items start at offset p,
1474 * and they are item_size apart. There are 'max' items in p.
1475 *
1476 * the slot in the array is returned via slot, and it points to
1477 * the place where you would insert key if it is not found in
1478 * the array.
1479 *
1480 * slot may point to max if the key is bigger than all of the keys
1481 */
1482static noinline int generic_bin_search(struct extent_buffer *eb,
1483 unsigned long p,
1484 int item_size, struct btrfs_key *key,
1485 int max, int *slot)
1486{
1487 int low = 0;
1488 int high = max;
1489 int mid;
1490 int ret;
1491 struct btrfs_disk_key *tmp = NULL;
1492 struct btrfs_disk_key unaligned;
1493 unsigned long offset;
1494 char *kaddr = NULL;
1495 unsigned long map_start = 0;
1496 unsigned long map_len = 0;
1497 int err;
1498
1499 while (low < high) {
1500 mid = (low + high) / 2;
1501 offset = p + mid * item_size;
1502
1503 if (!kaddr || offset < map_start ||
1504 (offset + sizeof(struct btrfs_disk_key)) >
1505 map_start + map_len) {
1506
1507 err = map_private_extent_buffer(eb, offset,
1508 sizeof(struct btrfs_disk_key),
1509 &kaddr, &map_start, &map_len);
1510
1511 if (!err) {
1512 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1513 map_start);
1514 } else {
1515 read_extent_buffer(eb, &unaligned,
1516 offset, sizeof(unaligned));
1517 tmp = &unaligned;
1518 }
1519
1520 } else {
1521 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1522 map_start);
1523 }
1524 ret = comp_keys(tmp, key);
1525
1526 if (ret < 0)
1527 low = mid + 1;
1528 else if (ret > 0)
1529 high = mid;
1530 else {
1531 *slot = mid;
1532 return 0;
1533 }
1534 }
1535 *slot = low;
1536 return 1;
1537}
1538
1539/*
1540 * simple bin_search frontend that does the right thing for
1541 * leaves vs nodes
1542 */
1543static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1544 int level, int *slot)
1545{
1546 if (level == 0)
1547 return generic_bin_search(eb,
1548 offsetof(struct btrfs_leaf, items),
1549 sizeof(struct btrfs_item),
1550 key, btrfs_header_nritems(eb),
1551 slot);
1552 else
1553 return generic_bin_search(eb,
1554 offsetof(struct btrfs_node, ptrs),
1555 sizeof(struct btrfs_key_ptr),
1556 key, btrfs_header_nritems(eb),
1557 slot);
1558}
1559
1560int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1561 int level, int *slot)
1562{
1563 return bin_search(eb, key, level, slot);
1564}
1565
1566static void root_add_used(struct btrfs_root *root, u32 size)
1567{
1568 spin_lock(&root->accounting_lock);
1569 btrfs_set_root_used(&root->root_item,
1570 btrfs_root_used(&root->root_item) + size);
1571 spin_unlock(&root->accounting_lock);
1572}
1573
1574static void root_sub_used(struct btrfs_root *root, u32 size)
1575{
1576 spin_lock(&root->accounting_lock);
1577 btrfs_set_root_used(&root->root_item,
1578 btrfs_root_used(&root->root_item) - size);
1579 spin_unlock(&root->accounting_lock);
1580}
1581
1582/* given a node and slot number, this reads the blocks it points to. The
1583 * extent buffer is returned with a reference taken (but unlocked).
1584 * NULL is returned on error.
1585 */
1586static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
1587 struct extent_buffer *parent, int slot)
1588{
1589 int level = btrfs_header_level(parent);
1590 if (slot < 0)
1591 return NULL;
1592 if (slot >= btrfs_header_nritems(parent))
1593 return NULL;
1594
1595 BUG_ON(level == 0);
1596
1597 return read_tree_block(root, btrfs_node_blockptr(parent, slot),
1598 btrfs_level_size(root, level - 1),
1599 btrfs_node_ptr_generation(parent, slot));
1600}
1601
1602/*
1603 * node level balancing, used to make sure nodes are in proper order for
1604 * item deletion. We balance from the top down, so we have to make sure
1605 * that a deletion won't leave an node completely empty later on.
1606 */
1607static noinline int balance_level(struct btrfs_trans_handle *trans,
1608 struct btrfs_root *root,
1609 struct btrfs_path *path, int level)
1610{
1611 struct extent_buffer *right = NULL;
1612 struct extent_buffer *mid;
1613 struct extent_buffer *left = NULL;
1614 struct extent_buffer *parent = NULL;
1615 int ret = 0;
1616 int wret;
1617 int pslot;
1618 int orig_slot = path->slots[level];
1619 u64 orig_ptr;
1620
1621 if (level == 0)
1622 return 0;
1623
1624 mid = path->nodes[level];
1625
1626 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1627 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1628 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1629
1630 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1631
1632 if (level < BTRFS_MAX_LEVEL - 1) {
1633 parent = path->nodes[level + 1];
1634 pslot = path->slots[level + 1];
1635 }
1636
1637 /*
1638 * deal with the case where there is only one pointer in the root
1639 * by promoting the node below to a root
1640 */
1641 if (!parent) {
1642 struct extent_buffer *child;
1643
1644 if (btrfs_header_nritems(mid) != 1)
1645 return 0;
1646
1647 /* promote the child to a root */
1648 child = read_node_slot(root, mid, 0);
1649 if (!child) {
1650 ret = -EROFS;
1651 btrfs_std_error(root->fs_info, ret);
1652 goto enospc;
1653 }
1654
1655 btrfs_tree_lock(child);
1656 btrfs_set_lock_blocking(child);
1657 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1658 if (ret) {
1659 btrfs_tree_unlock(child);
1660 free_extent_buffer(child);
1661 goto enospc;
1662 }
1663
1664 tree_mod_log_set_root_pointer(root, child);
1665 rcu_assign_pointer(root->node, child);
1666
1667 add_root_to_dirty_list(root);
1668 btrfs_tree_unlock(child);
1669
1670 path->locks[level] = 0;
1671 path->nodes[level] = NULL;
1672 clean_tree_block(trans, root, mid);
1673 btrfs_tree_unlock(mid);
1674 /* once for the path */
1675 free_extent_buffer(mid);
1676
1677 root_sub_used(root, mid->len);
1678 btrfs_free_tree_block(trans, root, mid, 0, 1);
1679 /* once for the root ptr */
1680 free_extent_buffer_stale(mid);
1681 return 0;
1682 }
1683 if (btrfs_header_nritems(mid) >
1684 BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
1685 return 0;
1686
1687 left = read_node_slot(root, parent, pslot - 1);
1688 if (left) {
1689 btrfs_tree_lock(left);
1690 btrfs_set_lock_blocking(left);
1691 wret = btrfs_cow_block(trans, root, left,
1692 parent, pslot - 1, &left);
1693 if (wret) {
1694 ret = wret;
1695 goto enospc;
1696 }
1697 }
1698 right = read_node_slot(root, parent, pslot + 1);
1699 if (right) {
1700 btrfs_tree_lock(right);
1701 btrfs_set_lock_blocking(right);
1702 wret = btrfs_cow_block(trans, root, right,
1703 parent, pslot + 1, &right);
1704 if (wret) {
1705 ret = wret;
1706 goto enospc;
1707 }
1708 }
1709
1710 /* first, try to make some room in the middle buffer */
1711 if (left) {
1712 orig_slot += btrfs_header_nritems(left);
1713 wret = push_node_left(trans, root, left, mid, 1);
1714 if (wret < 0)
1715 ret = wret;
1716 }
1717
1718 /*
1719 * then try to empty the right most buffer into the middle
1720 */
1721 if (right) {
1722 wret = push_node_left(trans, root, mid, right, 1);
1723 if (wret < 0 && wret != -ENOSPC)
1724 ret = wret;
1725 if (btrfs_header_nritems(right) == 0) {
1726 clean_tree_block(trans, root, right);
1727 btrfs_tree_unlock(right);
1728 del_ptr(trans, root, path, level + 1, pslot + 1, 1);
1729 root_sub_used(root, right->len);
1730 btrfs_free_tree_block(trans, root, right, 0, 1);
1731 free_extent_buffer_stale(right);
1732 right = NULL;
1733 } else {
1734 struct btrfs_disk_key right_key;
1735 btrfs_node_key(right, &right_key, 0);
1736 tree_mod_log_set_node_key(root->fs_info, parent,
1737 &right_key, pslot + 1, 0);
1738 btrfs_set_node_key(parent, &right_key, pslot + 1);
1739 btrfs_mark_buffer_dirty(parent);
1740 }
1741 }
1742 if (btrfs_header_nritems(mid) == 1) {
1743 /*
1744 * we're not allowed to leave a node with one item in the
1745 * tree during a delete. A deletion from lower in the tree
1746 * could try to delete the only pointer in this node.
1747 * So, pull some keys from the left.
1748 * There has to be a left pointer at this point because
1749 * otherwise we would have pulled some pointers from the
1750 * right
1751 */
1752 if (!left) {
1753 ret = -EROFS;
1754 btrfs_std_error(root->fs_info, ret);
1755 goto enospc;
1756 }
1757 wret = balance_node_right(trans, root, mid, left);
1758 if (wret < 0) {
1759 ret = wret;
1760 goto enospc;
1761 }
1762 if (wret == 1) {
1763 wret = push_node_left(trans, root, left, mid, 1);
1764 if (wret < 0)
1765 ret = wret;
1766 }
1767 BUG_ON(wret == 1);
1768 }
1769 if (btrfs_header_nritems(mid) == 0) {
1770 clean_tree_block(trans, root, mid);
1771 btrfs_tree_unlock(mid);
1772 del_ptr(trans, root, path, level + 1, pslot, 1);
1773 root_sub_used(root, mid->len);
1774 btrfs_free_tree_block(trans, root, mid, 0, 1);
1775 free_extent_buffer_stale(mid);
1776 mid = NULL;
1777 } else {
1778 /* update the parent key to reflect our changes */
1779 struct btrfs_disk_key mid_key;
1780 btrfs_node_key(mid, &mid_key, 0);
1781 tree_mod_log_set_node_key(root->fs_info, parent, &mid_key,
1782 pslot, 0);
1783 btrfs_set_node_key(parent, &mid_key, pslot);
1784 btrfs_mark_buffer_dirty(parent);
1785 }
1786
1787 /* update the path */
1788 if (left) {
1789 if (btrfs_header_nritems(left) > orig_slot) {
1790 extent_buffer_get(left);
1791 /* left was locked after cow */
1792 path->nodes[level] = left;
1793 path->slots[level + 1] -= 1;
1794 path->slots[level] = orig_slot;
1795 if (mid) {
1796 btrfs_tree_unlock(mid);
1797 free_extent_buffer(mid);
1798 }
1799 } else {
1800 orig_slot -= btrfs_header_nritems(left);
1801 path->slots[level] = orig_slot;
1802 }
1803 }
1804 /* double check we haven't messed things up */
1805 if (orig_ptr !=
1806 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1807 BUG();
1808enospc:
1809 if (right) {
1810 btrfs_tree_unlock(right);
1811 free_extent_buffer(right);
1812 }
1813 if (left) {
1814 if (path->nodes[level] != left)
1815 btrfs_tree_unlock(left);
1816 free_extent_buffer(left);
1817 }
1818 return ret;
1819}
1820
1821/* Node balancing for insertion. Here we only split or push nodes around
1822 * when they are completely full. This is also done top down, so we
1823 * have to be pessimistic.
1824 */
1825static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1826 struct btrfs_root *root,
1827 struct btrfs_path *path, int level)
1828{
1829 struct extent_buffer *right = NULL;
1830 struct extent_buffer *mid;
1831 struct extent_buffer *left = NULL;
1832 struct extent_buffer *parent = NULL;
1833 int ret = 0;
1834 int wret;
1835 int pslot;
1836 int orig_slot = path->slots[level];
1837
1838 if (level == 0)
1839 return 1;
1840
1841 mid = path->nodes[level];
1842 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1843
1844 if (level < BTRFS_MAX_LEVEL - 1) {
1845 parent = path->nodes[level + 1];
1846 pslot = path->slots[level + 1];
1847 }
1848
1849 if (!parent)
1850 return 1;
1851
1852 left = read_node_slot(root, parent, pslot - 1);
1853
1854 /* first, try to make some room in the middle buffer */
1855 if (left) {
1856 u32 left_nr;
1857
1858 btrfs_tree_lock(left);
1859 btrfs_set_lock_blocking(left);
1860
1861 left_nr = btrfs_header_nritems(left);
1862 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1863 wret = 1;
1864 } else {
1865 ret = btrfs_cow_block(trans, root, left, parent,
1866 pslot - 1, &left);
1867 if (ret)
1868 wret = 1;
1869 else {
1870 wret = push_node_left(trans, root,
1871 left, mid, 0);
1872 }
1873 }
1874 if (wret < 0)
1875 ret = wret;
1876 if (wret == 0) {
1877 struct btrfs_disk_key disk_key;
1878 orig_slot += left_nr;
1879 btrfs_node_key(mid, &disk_key, 0);
1880 tree_mod_log_set_node_key(root->fs_info, parent,
1881 &disk_key, pslot, 0);
1882 btrfs_set_node_key(parent, &disk_key, pslot);
1883 btrfs_mark_buffer_dirty(parent);
1884 if (btrfs_header_nritems(left) > orig_slot) {
1885 path->nodes[level] = left;
1886 path->slots[level + 1] -= 1;
1887 path->slots[level] = orig_slot;
1888 btrfs_tree_unlock(mid);
1889 free_extent_buffer(mid);
1890 } else {
1891 orig_slot -=
1892 btrfs_header_nritems(left);
1893 path->slots[level] = orig_slot;
1894 btrfs_tree_unlock(left);
1895 free_extent_buffer(left);
1896 }
1897 return 0;
1898 }
1899 btrfs_tree_unlock(left);
1900 free_extent_buffer(left);
1901 }
1902 right = read_node_slot(root, parent, pslot + 1);
1903
1904 /*
1905 * then try to empty the right most buffer into the middle
1906 */
1907 if (right) {
1908 u32 right_nr;
1909
1910 btrfs_tree_lock(right);
1911 btrfs_set_lock_blocking(right);
1912
1913 right_nr = btrfs_header_nritems(right);
1914 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1915 wret = 1;
1916 } else {
1917 ret = btrfs_cow_block(trans, root, right,
1918 parent, pslot + 1,
1919 &right);
1920 if (ret)
1921 wret = 1;
1922 else {
1923 wret = balance_node_right(trans, root,
1924 right, mid);
1925 }
1926 }
1927 if (wret < 0)
1928 ret = wret;
1929 if (wret == 0) {
1930 struct btrfs_disk_key disk_key;
1931
1932 btrfs_node_key(right, &disk_key, 0);
1933 tree_mod_log_set_node_key(root->fs_info, parent,
1934 &disk_key, pslot + 1, 0);
1935 btrfs_set_node_key(parent, &disk_key, pslot + 1);
1936 btrfs_mark_buffer_dirty(parent);
1937
1938 if (btrfs_header_nritems(mid) <= orig_slot) {
1939 path->nodes[level] = right;
1940 path->slots[level + 1] += 1;
1941 path->slots[level] = orig_slot -
1942 btrfs_header_nritems(mid);
1943 btrfs_tree_unlock(mid);
1944 free_extent_buffer(mid);
1945 } else {
1946 btrfs_tree_unlock(right);
1947 free_extent_buffer(right);
1948 }
1949 return 0;
1950 }
1951 btrfs_tree_unlock(right);
1952 free_extent_buffer(right);
1953 }
1954 return 1;
1955}
1956
1957/*
1958 * readahead one full node of leaves, finding things that are close
1959 * to the block in 'slot', and triggering ra on them.
1960 */
1961static void reada_for_search(struct btrfs_root *root,
1962 struct btrfs_path *path,
1963 int level, int slot, u64 objectid)
1964{
1965 struct extent_buffer *node;
1966 struct btrfs_disk_key disk_key;
1967 u32 nritems;
1968 u64 search;
1969 u64 target;
1970 u64 nread = 0;
1971 u64 gen;
1972 int direction = path->reada;
1973 struct extent_buffer *eb;
1974 u32 nr;
1975 u32 blocksize;
1976 u32 nscan = 0;
1977
1978 if (level != 1)
1979 return;
1980
1981 if (!path->nodes[level])
1982 return;
1983
1984 node = path->nodes[level];
1985
1986 search = btrfs_node_blockptr(node, slot);
1987 blocksize = btrfs_level_size(root, level - 1);
1988 eb = btrfs_find_tree_block(root, search, blocksize);
1989 if (eb) {
1990 free_extent_buffer(eb);
1991 return;
1992 }
1993
1994 target = search;
1995
1996 nritems = btrfs_header_nritems(node);
1997 nr = slot;
1998
1999 while (1) {
2000 if (direction < 0) {
2001 if (nr == 0)
2002 break;
2003 nr--;
2004 } else if (direction > 0) {
2005 nr++;
2006 if (nr >= nritems)
2007 break;
2008 }
2009 if (path->reada < 0 && objectid) {
2010 btrfs_node_key(node, &disk_key, nr);
2011 if (btrfs_disk_key_objectid(&disk_key) != objectid)
2012 break;
2013 }
2014 search = btrfs_node_blockptr(node, nr);
2015 if ((search <= target && target - search <= 65536) ||
2016 (search > target && search - target <= 65536)) {
2017 gen = btrfs_node_ptr_generation(node, nr);
2018 readahead_tree_block(root, search, blocksize, gen);
2019 nread += blocksize;
2020 }
2021 nscan++;
2022 if ((nread > 65536 || nscan > 32))
2023 break;
2024 }
2025}
2026
2027/*
2028 * returns -EAGAIN if it had to drop the path, or zero if everything was in
2029 * cache
2030 */
2031static noinline int reada_for_balance(struct btrfs_root *root,
2032 struct btrfs_path *path, int level)
2033{
2034 int slot;
2035 int nritems;
2036 struct extent_buffer *parent;
2037 struct extent_buffer *eb;
2038 u64 gen;
2039 u64 block1 = 0;
2040 u64 block2 = 0;
2041 int ret = 0;
2042 int blocksize;
2043
2044 parent = path->nodes[level + 1];
2045 if (!parent)
2046 return 0;
2047
2048 nritems = btrfs_header_nritems(parent);
2049 slot = path->slots[level + 1];
2050 blocksize = btrfs_level_size(root, level);
2051
2052 if (slot > 0) {
2053 block1 = btrfs_node_blockptr(parent, slot - 1);
2054 gen = btrfs_node_ptr_generation(parent, slot - 1);
2055 eb = btrfs_find_tree_block(root, block1, blocksize);
2056 /*
2057 * if we get -eagain from btrfs_buffer_uptodate, we
2058 * don't want to return eagain here. That will loop
2059 * forever
2060 */
2061 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2062 block1 = 0;
2063 free_extent_buffer(eb);
2064 }
2065 if (slot + 1 < nritems) {
2066 block2 = btrfs_node_blockptr(parent, slot + 1);
2067 gen = btrfs_node_ptr_generation(parent, slot + 1);
2068 eb = btrfs_find_tree_block(root, block2, blocksize);
2069 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2070 block2 = 0;
2071 free_extent_buffer(eb);
2072 }
2073 if (block1 || block2) {
2074 ret = -EAGAIN;
2075
2076 /* release the whole path */
2077 btrfs_release_path(path);
2078
2079 /* read the blocks */
2080 if (block1)
2081 readahead_tree_block(root, block1, blocksize, 0);
2082 if (block2)
2083 readahead_tree_block(root, block2, blocksize, 0);
2084
2085 if (block1) {
2086 eb = read_tree_block(root, block1, blocksize, 0);
2087 free_extent_buffer(eb);
2088 }
2089 if (block2) {
2090 eb = read_tree_block(root, block2, blocksize, 0);
2091 free_extent_buffer(eb);
2092 }
2093 }
2094 return ret;
2095}
2096
2097
2098/*
2099 * when we walk down the tree, it is usually safe to unlock the higher layers
2100 * in the tree. The exceptions are when our path goes through slot 0, because
2101 * operations on the tree might require changing key pointers higher up in the
2102 * tree.
2103 *
2104 * callers might also have set path->keep_locks, which tells this code to keep
2105 * the lock if the path points to the last slot in the block. This is part of
2106 * walking through the tree, and selecting the next slot in the higher block.
2107 *
2108 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
2109 * if lowest_unlock is 1, level 0 won't be unlocked
2110 */
2111static noinline void unlock_up(struct btrfs_path *path, int level,
2112 int lowest_unlock, int min_write_lock_level,
2113 int *write_lock_level)
2114{
2115 int i;
2116 int skip_level = level;
2117 int no_skips = 0;
2118 struct extent_buffer *t;
2119
2120 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2121 if (!path->nodes[i])
2122 break;
2123 if (!path->locks[i])
2124 break;
2125 if (!no_skips && path->slots[i] == 0) {
2126 skip_level = i + 1;
2127 continue;
2128 }
2129 if (!no_skips && path->keep_locks) {
2130 u32 nritems;
2131 t = path->nodes[i];
2132 nritems = btrfs_header_nritems(t);
2133 if (nritems < 1 || path->slots[i] >= nritems - 1) {
2134 skip_level = i + 1;
2135 continue;
2136 }
2137 }
2138 if (skip_level < i && i >= lowest_unlock)
2139 no_skips = 1;
2140
2141 t = path->nodes[i];
2142 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2143 btrfs_tree_unlock_rw(t, path->locks[i]);
2144 path->locks[i] = 0;
2145 if (write_lock_level &&
2146 i > min_write_lock_level &&
2147 i <= *write_lock_level) {
2148 *write_lock_level = i - 1;
2149 }
2150 }
2151 }
2152}
2153
2154/*
2155 * This releases any locks held in the path starting at level and
2156 * going all the way up to the root.
2157 *
2158 * btrfs_search_slot will keep the lock held on higher nodes in a few
2159 * corner cases, such as COW of the block at slot zero in the node. This
2160 * ignores those rules, and it should only be called when there are no
2161 * more updates to be done higher up in the tree.
2162 */
2163noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2164{
2165 int i;
2166
2167 if (path->keep_locks)
2168 return;
2169
2170 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2171 if (!path->nodes[i])
2172 continue;
2173 if (!path->locks[i])
2174 continue;
2175 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2176 path->locks[i] = 0;
2177 }
2178}
2179
2180/*
2181 * helper function for btrfs_search_slot. The goal is to find a block
2182 * in cache without setting the path to blocking. If we find the block
2183 * we return zero and the path is unchanged.
2184 *
2185 * If we can't find the block, we set the path blocking and do some
2186 * reada. -EAGAIN is returned and the search must be repeated.
2187 */
2188static int
2189read_block_for_search(struct btrfs_trans_handle *trans,
2190 struct btrfs_root *root, struct btrfs_path *p,
2191 struct extent_buffer **eb_ret, int level, int slot,
2192 struct btrfs_key *key, u64 time_seq)
2193{
2194 u64 blocknr;
2195 u64 gen;
2196 u32 blocksize;
2197 struct extent_buffer *b = *eb_ret;
2198 struct extent_buffer *tmp;
2199 int ret;
2200
2201 blocknr = btrfs_node_blockptr(b, slot);
2202 gen = btrfs_node_ptr_generation(b, slot);
2203 blocksize = btrfs_level_size(root, level - 1);
2204
2205 tmp = btrfs_find_tree_block(root, blocknr, blocksize);
2206 if (tmp) {
2207 /* first we do an atomic uptodate check */
2208 if (btrfs_buffer_uptodate(tmp, 0, 1) > 0) {
2209 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2210 /*
2211 * we found an up to date block without
2212 * sleeping, return
2213 * right away
2214 */
2215 *eb_ret = tmp;
2216 return 0;
2217 }
2218 /* the pages were up to date, but we failed
2219 * the generation number check. Do a full
2220 * read for the generation number that is correct.
2221 * We must do this without dropping locks so
2222 * we can trust our generation number
2223 */
2224 free_extent_buffer(tmp);
2225 btrfs_set_path_blocking(p);
2226
2227 /* now we're allowed to do a blocking uptodate check */
2228 tmp = read_tree_block(root, blocknr, blocksize, gen);
2229 if (tmp && btrfs_buffer_uptodate(tmp, gen, 0) > 0) {
2230 *eb_ret = tmp;
2231 return 0;
2232 }
2233 free_extent_buffer(tmp);
2234 btrfs_release_path(p);
2235 return -EIO;
2236 }
2237 }
2238
2239 /*
2240 * reduce lock contention at high levels
2241 * of the btree by dropping locks before
2242 * we read. Don't release the lock on the current
2243 * level because we need to walk this node to figure
2244 * out which blocks to read.
2245 */
2246 btrfs_unlock_up_safe(p, level + 1);
2247 btrfs_set_path_blocking(p);
2248
2249 free_extent_buffer(tmp);
2250 if (p->reada)
2251 reada_for_search(root, p, level, slot, key->objectid);
2252
2253 btrfs_release_path(p);
2254
2255 ret = -EAGAIN;
2256 tmp = read_tree_block(root, blocknr, blocksize, 0);
2257 if (tmp) {
2258 /*
2259 * If the read above didn't mark this buffer up to date,
2260 * it will never end up being up to date. Set ret to EIO now
2261 * and give up so that our caller doesn't loop forever
2262 * on our EAGAINs.
2263 */
2264 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2265 ret = -EIO;
2266 free_extent_buffer(tmp);
2267 }
2268 return ret;
2269}
2270
2271/*
2272 * helper function for btrfs_search_slot. This does all of the checks
2273 * for node-level blocks and does any balancing required based on
2274 * the ins_len.
2275 *
2276 * If no extra work was required, zero is returned. If we had to
2277 * drop the path, -EAGAIN is returned and btrfs_search_slot must
2278 * start over
2279 */
2280static int
2281setup_nodes_for_search(struct btrfs_trans_handle *trans,
2282 struct btrfs_root *root, struct btrfs_path *p,
2283 struct extent_buffer *b, int level, int ins_len,
2284 int *write_lock_level)
2285{
2286 int ret;
2287 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2288 BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
2289 int sret;
2290
2291 if (*write_lock_level < level + 1) {
2292 *write_lock_level = level + 1;
2293 btrfs_release_path(p);
2294 goto again;
2295 }
2296
2297 sret = reada_for_balance(root, p, level);
2298 if (sret)
2299 goto again;
2300
2301 btrfs_set_path_blocking(p);
2302 sret = split_node(trans, root, p, level);
2303 btrfs_clear_path_blocking(p, NULL, 0);
2304
2305 BUG_ON(sret > 0);
2306 if (sret) {
2307 ret = sret;
2308 goto done;
2309 }
2310 b = p->nodes[level];
2311 } else if (ins_len < 0 && btrfs_header_nritems(b) <
2312 BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
2313 int sret;
2314
2315 if (*write_lock_level < level + 1) {
2316 *write_lock_level = level + 1;
2317 btrfs_release_path(p);
2318 goto again;
2319 }
2320
2321 sret = reada_for_balance(root, p, level);
2322 if (sret)
2323 goto again;
2324
2325 btrfs_set_path_blocking(p);
2326 sret = balance_level(trans, root, p, level);
2327 btrfs_clear_path_blocking(p, NULL, 0);
2328
2329 if (sret) {
2330 ret = sret;
2331 goto done;
2332 }
2333 b = p->nodes[level];
2334 if (!b) {
2335 btrfs_release_path(p);
2336 goto again;
2337 }
2338 BUG_ON(btrfs_header_nritems(b) == 1);
2339 }
2340 return 0;
2341
2342again:
2343 ret = -EAGAIN;
2344done:
2345 return ret;
2346}
2347
2348/*
2349 * look for key in the tree. path is filled in with nodes along the way
2350 * if key is found, we return zero and you can find the item in the leaf
2351 * level of the path (level 0)
2352 *
2353 * If the key isn't found, the path points to the slot where it should
2354 * be inserted, and 1 is returned. If there are other errors during the
2355 * search a negative error number is returned.
2356 *
2357 * if ins_len > 0, nodes and leaves will be split as we walk down the
2358 * tree. if ins_len < 0, nodes will be merged as we walk down the tree (if
2359 * possible)
2360 */
2361int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
2362 *root, struct btrfs_key *key, struct btrfs_path *p, int
2363 ins_len, int cow)
2364{
2365 struct extent_buffer *b;
2366 int slot;
2367 int ret;
2368 int err;
2369 int level;
2370 int lowest_unlock = 1;
2371 int root_lock;
2372 /* everything at write_lock_level or lower must be write locked */
2373 int write_lock_level = 0;
2374 u8 lowest_level = 0;
2375 int min_write_lock_level;
2376
2377 lowest_level = p->lowest_level;
2378 WARN_ON(lowest_level && ins_len > 0);
2379 WARN_ON(p->nodes[0] != NULL);
2380
2381 if (ins_len < 0) {
2382 lowest_unlock = 2;
2383
2384 /* when we are removing items, we might have to go up to level
2385 * two as we update tree pointers Make sure we keep write
2386 * for those levels as well
2387 */
2388 write_lock_level = 2;
2389 } else if (ins_len > 0) {
2390 /*
2391 * for inserting items, make sure we have a write lock on
2392 * level 1 so we can update keys
2393 */
2394 write_lock_level = 1;
2395 }
2396
2397 if (!cow)
2398 write_lock_level = -1;
2399
2400 if (cow && (p->keep_locks || p->lowest_level))
2401 write_lock_level = BTRFS_MAX_LEVEL;
2402
2403 min_write_lock_level = write_lock_level;
2404
2405again:
2406 /*
2407 * we try very hard to do read locks on the root
2408 */
2409 root_lock = BTRFS_READ_LOCK;
2410 level = 0;
2411 if (p->search_commit_root) {
2412 /*
2413 * the commit roots are read only
2414 * so we always do read locks
2415 */
2416 b = root->commit_root;
2417 extent_buffer_get(b);
2418 level = btrfs_header_level(b);
2419 if (!p->skip_locking)
2420 btrfs_tree_read_lock(b);
2421 } else {
2422 if (p->skip_locking) {
2423 b = btrfs_root_node(root);
2424 level = btrfs_header_level(b);
2425 } else {
2426 /* we don't know the level of the root node
2427 * until we actually have it read locked
2428 */
2429 b = btrfs_read_lock_root_node(root);
2430 level = btrfs_header_level(b);
2431 if (level <= write_lock_level) {
2432 /* whoops, must trade for write lock */
2433 btrfs_tree_read_unlock(b);
2434 free_extent_buffer(b);
2435 b = btrfs_lock_root_node(root);
2436 root_lock = BTRFS_WRITE_LOCK;
2437
2438 /* the level might have changed, check again */
2439 level = btrfs_header_level(b);
2440 }
2441 }
2442 }
2443 p->nodes[level] = b;
2444 if (!p->skip_locking)
2445 p->locks[level] = root_lock;
2446
2447 while (b) {
2448 level = btrfs_header_level(b);
2449
2450 /*
2451 * setup the path here so we can release it under lock
2452 * contention with the cow code
2453 */
2454 if (cow) {
2455 /*
2456 * if we don't really need to cow this block
2457 * then we don't want to set the path blocking,
2458 * so we test it here
2459 */
2460 if (!should_cow_block(trans, root, b))
2461 goto cow_done;
2462
2463 btrfs_set_path_blocking(p);
2464
2465 /*
2466 * must have write locks on this node and the
2467 * parent
2468 */
2469 if (level + 1 > write_lock_level) {
2470 write_lock_level = level + 1;
2471 btrfs_release_path(p);
2472 goto again;
2473 }
2474
2475 err = btrfs_cow_block(trans, root, b,
2476 p->nodes[level + 1],
2477 p->slots[level + 1], &b);
2478 if (err) {
2479 ret = err;
2480 goto done;
2481 }
2482 }
2483cow_done:
2484 BUG_ON(!cow && ins_len);
2485
2486 p->nodes[level] = b;
2487 btrfs_clear_path_blocking(p, NULL, 0);
2488
2489 /*
2490 * we have a lock on b and as long as we aren't changing
2491 * the tree, there is no way to for the items in b to change.
2492 * It is safe to drop the lock on our parent before we
2493 * go through the expensive btree search on b.
2494 *
2495 * If cow is true, then we might be changing slot zero,
2496 * which may require changing the parent. So, we can't
2497 * drop the lock until after we know which slot we're
2498 * operating on.
2499 */
2500 if (!cow)
2501 btrfs_unlock_up_safe(p, level + 1);
2502
2503 ret = bin_search(b, key, level, &slot);
2504
2505 if (level != 0) {
2506 int dec = 0;
2507 if (ret && slot > 0) {
2508 dec = 1;
2509 slot -= 1;
2510 }
2511 p->slots[level] = slot;
2512 err = setup_nodes_for_search(trans, root, p, b, level,
2513 ins_len, &write_lock_level);
2514 if (err == -EAGAIN)
2515 goto again;
2516 if (err) {
2517 ret = err;
2518 goto done;
2519 }
2520 b = p->nodes[level];
2521 slot = p->slots[level];
2522
2523 /*
2524 * slot 0 is special, if we change the key
2525 * we have to update the parent pointer
2526 * which means we must have a write lock
2527 * on the parent
2528 */
2529 if (slot == 0 && cow &&
2530 write_lock_level < level + 1) {
2531 write_lock_level = level + 1;
2532 btrfs_release_path(p);
2533 goto again;
2534 }
2535
2536 unlock_up(p, level, lowest_unlock,
2537 min_write_lock_level, &write_lock_level);
2538
2539 if (level == lowest_level) {
2540 if (dec)
2541 p->slots[level]++;
2542 goto done;
2543 }
2544
2545 err = read_block_for_search(trans, root, p,
2546 &b, level, slot, key, 0);
2547 if (err == -EAGAIN)
2548 goto again;
2549 if (err) {
2550 ret = err;
2551 goto done;
2552 }
2553
2554 if (!p->skip_locking) {
2555 level = btrfs_header_level(b);
2556 if (level <= write_lock_level) {
2557 err = btrfs_try_tree_write_lock(b);
2558 if (!err) {
2559 btrfs_set_path_blocking(p);
2560 btrfs_tree_lock(b);
2561 btrfs_clear_path_blocking(p, b,
2562 BTRFS_WRITE_LOCK);
2563 }
2564 p->locks[level] = BTRFS_WRITE_LOCK;
2565 } else {
2566 err = btrfs_try_tree_read_lock(b);
2567 if (!err) {
2568 btrfs_set_path_blocking(p);
2569 btrfs_tree_read_lock(b);
2570 btrfs_clear_path_blocking(p, b,
2571 BTRFS_READ_LOCK);
2572 }
2573 p->locks[level] = BTRFS_READ_LOCK;
2574 }
2575 p->nodes[level] = b;
2576 }
2577 } else {
2578 p->slots[level] = slot;
2579 if (ins_len > 0 &&
2580 btrfs_leaf_free_space(root, b) < ins_len) {
2581 if (write_lock_level < 1) {
2582 write_lock_level = 1;
2583 btrfs_release_path(p);
2584 goto again;
2585 }
2586
2587 btrfs_set_path_blocking(p);
2588 err = split_leaf(trans, root, key,
2589 p, ins_len, ret == 0);
2590 btrfs_clear_path_blocking(p, NULL, 0);
2591
2592 BUG_ON(err > 0);
2593 if (err) {
2594 ret = err;
2595 goto done;
2596 }
2597 }
2598 if (!p->search_for_split)
2599 unlock_up(p, level, lowest_unlock,
2600 min_write_lock_level, &write_lock_level);
2601 goto done;
2602 }
2603 }
2604 ret = 1;
2605done:
2606 /*
2607 * we don't really know what they plan on doing with the path
2608 * from here on, so for now just mark it as blocking
2609 */
2610 if (!p->leave_spinning)
2611 btrfs_set_path_blocking(p);
2612 if (ret < 0)
2613 btrfs_release_path(p);
2614 return ret;
2615}
2616
2617/*
2618 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2619 * current state of the tree together with the operations recorded in the tree
2620 * modification log to search for the key in a previous version of this tree, as
2621 * denoted by the time_seq parameter.
2622 *
2623 * Naturally, there is no support for insert, delete or cow operations.
2624 *
2625 * The resulting path and return value will be set up as if we called
2626 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2627 */
2628int btrfs_search_old_slot(struct btrfs_root *root, struct btrfs_key *key,
2629 struct btrfs_path *p, u64 time_seq)
2630{
2631 struct extent_buffer *b;
2632 int slot;
2633 int ret;
2634 int err;
2635 int level;
2636 int lowest_unlock = 1;
2637 u8 lowest_level = 0;
2638
2639 lowest_level = p->lowest_level;
2640 WARN_ON(p->nodes[0] != NULL);
2641
2642 if (p->search_commit_root) {
2643 BUG_ON(time_seq);
2644 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2645 }
2646
2647again:
2648 b = get_old_root(root, time_seq);
2649 level = btrfs_header_level(b);
2650 p->locks[level] = BTRFS_READ_LOCK;
2651
2652 while (b) {
2653 level = btrfs_header_level(b);
2654 p->nodes[level] = b;
2655 btrfs_clear_path_blocking(p, NULL, 0);
2656
2657 /*
2658 * we have a lock on b and as long as we aren't changing
2659 * the tree, there is no way to for the items in b to change.
2660 * It is safe to drop the lock on our parent before we
2661 * go through the expensive btree search on b.
2662 */
2663 btrfs_unlock_up_safe(p, level + 1);
2664
2665 ret = bin_search(b, key, level, &slot);
2666
2667 if (level != 0) {
2668 int dec = 0;
2669 if (ret && slot > 0) {
2670 dec = 1;
2671 slot -= 1;
2672 }
2673 p->slots[level] = slot;
2674 unlock_up(p, level, lowest_unlock, 0, NULL);
2675
2676 if (level == lowest_level) {
2677 if (dec)
2678 p->slots[level]++;
2679 goto done;
2680 }
2681
2682 err = read_block_for_search(NULL, root, p, &b, level,
2683 slot, key, time_seq);
2684 if (err == -EAGAIN)
2685 goto again;
2686 if (err) {
2687 ret = err;
2688 goto done;
2689 }
2690
2691 level = btrfs_header_level(b);
2692 err = btrfs_try_tree_read_lock(b);
2693 if (!err) {
2694 btrfs_set_path_blocking(p);
2695 btrfs_tree_read_lock(b);
2696 btrfs_clear_path_blocking(p, b,
2697 BTRFS_READ_LOCK);
2698 }
2699 p->locks[level] = BTRFS_READ_LOCK;
2700 p->nodes[level] = b;
2701 b = tree_mod_log_rewind(root->fs_info, b, time_seq);
2702 if (b != p->nodes[level]) {
2703 btrfs_tree_unlock_rw(p->nodes[level],
2704 p->locks[level]);
2705 p->locks[level] = 0;
2706 p->nodes[level] = b;
2707 }
2708 } else {
2709 p->slots[level] = slot;
2710 unlock_up(p, level, lowest_unlock, 0, NULL);
2711 goto done;
2712 }
2713 }
2714 ret = 1;
2715done:
2716 if (!p->leave_spinning)
2717 btrfs_set_path_blocking(p);
2718 if (ret < 0)
2719 btrfs_release_path(p);
2720
2721 return ret;
2722}
2723
2724/*
2725 * adjust the pointers going up the tree, starting at level
2726 * making sure the right key of each node is points to 'key'.
2727 * This is used after shifting pointers to the left, so it stops
2728 * fixing up pointers when a given leaf/node is not in slot 0 of the
2729 * higher levels
2730 *
2731 */
2732static void fixup_low_keys(struct btrfs_trans_handle *trans,
2733 struct btrfs_root *root, struct btrfs_path *path,
2734 struct btrfs_disk_key *key, int level)
2735{
2736 int i;
2737 struct extent_buffer *t;
2738
2739 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2740 int tslot = path->slots[i];
2741 if (!path->nodes[i])
2742 break;
2743 t = path->nodes[i];
2744 tree_mod_log_set_node_key(root->fs_info, t, key, tslot, 1);
2745 btrfs_set_node_key(t, key, tslot);
2746 btrfs_mark_buffer_dirty(path->nodes[i]);
2747 if (tslot != 0)
2748 break;
2749 }
2750}
2751
2752/*
2753 * update item key.
2754 *
2755 * This function isn't completely safe. It's the caller's responsibility
2756 * that the new key won't break the order
2757 */
2758void btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
2759 struct btrfs_root *root, struct btrfs_path *path,
2760 struct btrfs_key *new_key)
2761{
2762 struct btrfs_disk_key disk_key;
2763 struct extent_buffer *eb;
2764 int slot;
2765
2766 eb = path->nodes[0];
2767 slot = path->slots[0];
2768 if (slot > 0) {
2769 btrfs_item_key(eb, &disk_key, slot - 1);
2770 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
2771 }
2772 if (slot < btrfs_header_nritems(eb) - 1) {
2773 btrfs_item_key(eb, &disk_key, slot + 1);
2774 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
2775 }
2776
2777 btrfs_cpu_key_to_disk(&disk_key, new_key);
2778 btrfs_set_item_key(eb, &disk_key, slot);
2779 btrfs_mark_buffer_dirty(eb);
2780 if (slot == 0)
2781 fixup_low_keys(trans, root, path, &disk_key, 1);
2782}
2783
2784/*
2785 * try to push data from one node into the next node left in the
2786 * tree.
2787 *
2788 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2789 * error, and > 0 if there was no room in the left hand block.
2790 */
2791static int push_node_left(struct btrfs_trans_handle *trans,
2792 struct btrfs_root *root, struct extent_buffer *dst,
2793 struct extent_buffer *src, int empty)
2794{
2795 int push_items = 0;
2796 int src_nritems;
2797 int dst_nritems;
2798 int ret = 0;
2799
2800 src_nritems = btrfs_header_nritems(src);
2801 dst_nritems = btrfs_header_nritems(dst);
2802 push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2803 WARN_ON(btrfs_header_generation(src) != trans->transid);
2804 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2805
2806 if (!empty && src_nritems <= 8)
2807 return 1;
2808
2809 if (push_items <= 0)
2810 return 1;
2811
2812 if (empty) {
2813 push_items = min(src_nritems, push_items);
2814 if (push_items < src_nritems) {
2815 /* leave at least 8 pointers in the node if
2816 * we aren't going to empty it
2817 */
2818 if (src_nritems - push_items < 8) {
2819 if (push_items <= 8)
2820 return 1;
2821 push_items -= 8;
2822 }
2823 }
2824 } else
2825 push_items = min(src_nritems - 8, push_items);
2826
2827 tree_mod_log_eb_copy(root->fs_info, dst, src, dst_nritems, 0,
2828 push_items);
2829 copy_extent_buffer(dst, src,
2830 btrfs_node_key_ptr_offset(dst_nritems),
2831 btrfs_node_key_ptr_offset(0),
2832 push_items * sizeof(struct btrfs_key_ptr));
2833
2834 if (push_items < src_nritems) {
2835 tree_mod_log_eb_move(root->fs_info, src, 0, push_items,
2836 src_nritems - push_items);
2837 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
2838 btrfs_node_key_ptr_offset(push_items),
2839 (src_nritems - push_items) *
2840 sizeof(struct btrfs_key_ptr));
2841 }
2842 btrfs_set_header_nritems(src, src_nritems - push_items);
2843 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2844 btrfs_mark_buffer_dirty(src);
2845 btrfs_mark_buffer_dirty(dst);
2846
2847 return ret;
2848}
2849
2850/*
2851 * try to push data from one node into the next node right in the
2852 * tree.
2853 *
2854 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2855 * error, and > 0 if there was no room in the right hand block.
2856 *
2857 * this will only push up to 1/2 the contents of the left node over
2858 */
2859static int balance_node_right(struct btrfs_trans_handle *trans,
2860 struct btrfs_root *root,
2861 struct extent_buffer *dst,
2862 struct extent_buffer *src)
2863{
2864 int push_items = 0;
2865 int max_push;
2866 int src_nritems;
2867 int dst_nritems;
2868 int ret = 0;
2869
2870 WARN_ON(btrfs_header_generation(src) != trans->transid);
2871 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2872
2873 src_nritems = btrfs_header_nritems(src);
2874 dst_nritems = btrfs_header_nritems(dst);
2875 push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2876 if (push_items <= 0)
2877 return 1;
2878
2879 if (src_nritems < 4)
2880 return 1;
2881
2882 max_push = src_nritems / 2 + 1;
2883 /* don't try to empty the node */
2884 if (max_push >= src_nritems)
2885 return 1;
2886
2887 if (max_push < push_items)
2888 push_items = max_push;
2889
2890 tree_mod_log_eb_move(root->fs_info, dst, push_items, 0, dst_nritems);
2891 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2892 btrfs_node_key_ptr_offset(0),
2893 (dst_nritems) *
2894 sizeof(struct btrfs_key_ptr));
2895
2896 tree_mod_log_eb_copy(root->fs_info, dst, src, 0,
2897 src_nritems - push_items, push_items);
2898 copy_extent_buffer(dst, src,
2899 btrfs_node_key_ptr_offset(0),
2900 btrfs_node_key_ptr_offset(src_nritems - push_items),
2901 push_items * sizeof(struct btrfs_key_ptr));
2902
2903 btrfs_set_header_nritems(src, src_nritems - push_items);
2904 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2905
2906 btrfs_mark_buffer_dirty(src);
2907 btrfs_mark_buffer_dirty(dst);
2908
2909 return ret;
2910}
2911
2912/*
2913 * helper function to insert a new root level in the tree.
2914 * A new node is allocated, and a single item is inserted to
2915 * point to the existing root
2916 *
2917 * returns zero on success or < 0 on failure.
2918 */
2919static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2920 struct btrfs_root *root,
2921 struct btrfs_path *path, int level)
2922{
2923 u64 lower_gen;
2924 struct extent_buffer *lower;
2925 struct extent_buffer *c;
2926 struct extent_buffer *old;
2927 struct btrfs_disk_key lower_key;
2928
2929 BUG_ON(path->nodes[level]);
2930 BUG_ON(path->nodes[level-1] != root->node);
2931
2932 lower = path->nodes[level-1];
2933 if (level == 1)
2934 btrfs_item_key(lower, &lower_key, 0);
2935 else
2936 btrfs_node_key(lower, &lower_key, 0);
2937
2938 c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2939 root->root_key.objectid, &lower_key,
2940 level, root->node->start, 0);
2941 if (IS_ERR(c))
2942 return PTR_ERR(c);
2943
2944 root_add_used(root, root->nodesize);
2945
2946 memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
2947 btrfs_set_header_nritems(c, 1);
2948 btrfs_set_header_level(c, level);
2949 btrfs_set_header_bytenr(c, c->start);
2950 btrfs_set_header_generation(c, trans->transid);
2951 btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
2952 btrfs_set_header_owner(c, root->root_key.objectid);
2953
2954 write_extent_buffer(c, root->fs_info->fsid,
2955 (unsigned long)btrfs_header_fsid(c),
2956 BTRFS_FSID_SIZE);
2957
2958 write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
2959 (unsigned long)btrfs_header_chunk_tree_uuid(c),
2960 BTRFS_UUID_SIZE);
2961
2962 btrfs_set_node_key(c, &lower_key, 0);
2963 btrfs_set_node_blockptr(c, 0, lower->start);
2964 lower_gen = btrfs_header_generation(lower);
2965 WARN_ON(lower_gen != trans->transid);
2966
2967 btrfs_set_node_ptr_generation(c, 0, lower_gen);
2968
2969 btrfs_mark_buffer_dirty(c);
2970
2971 old = root->node;
2972 tree_mod_log_set_root_pointer(root, c);
2973 rcu_assign_pointer(root->node, c);
2974
2975 /* the super has an extra ref to root->node */
2976 free_extent_buffer(old);
2977
2978 add_root_to_dirty_list(root);
2979 extent_buffer_get(c);
2980 path->nodes[level] = c;
2981 path->locks[level] = BTRFS_WRITE_LOCK;
2982 path->slots[level] = 0;
2983 return 0;
2984}
2985
2986/*
2987 * worker function to insert a single pointer in a node.
2988 * the node should have enough room for the pointer already
2989 *
2990 * slot and level indicate where you want the key to go, and
2991 * blocknr is the block the key points to.
2992 */
2993static void insert_ptr(struct btrfs_trans_handle *trans,
2994 struct btrfs_root *root, struct btrfs_path *path,
2995 struct btrfs_disk_key *key, u64 bytenr,
2996 int slot, int level)
2997{
2998 struct extent_buffer *lower;
2999 int nritems;
3000 int ret;
3001
3002 BUG_ON(!path->nodes[level]);
3003 btrfs_assert_tree_locked(path->nodes[level]);
3004 lower = path->nodes[level];
3005 nritems = btrfs_header_nritems(lower);
3006 BUG_ON(slot > nritems);
3007 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(root));
3008 if (slot != nritems) {
3009 if (level)
3010 tree_mod_log_eb_move(root->fs_info, lower, slot + 1,
3011 slot, nritems - slot);
3012 memmove_extent_buffer(lower,
3013 btrfs_node_key_ptr_offset(slot + 1),
3014 btrfs_node_key_ptr_offset(slot),
3015 (nritems - slot) * sizeof(struct btrfs_key_ptr));
3016 }
3017 if (level) {
3018 ret = tree_mod_log_insert_key(root->fs_info, lower, slot,
3019 MOD_LOG_KEY_ADD);
3020 BUG_ON(ret < 0);
3021 }
3022 btrfs_set_node_key(lower, key, slot);
3023 btrfs_set_node_blockptr(lower, slot, bytenr);
3024 WARN_ON(trans->transid == 0);
3025 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3026 btrfs_set_header_nritems(lower, nritems + 1);
3027 btrfs_mark_buffer_dirty(lower);
3028}
3029
3030/*
3031 * split the node at the specified level in path in two.
3032 * The path is corrected to point to the appropriate node after the split
3033 *
3034 * Before splitting this tries to make some room in the node by pushing
3035 * left and right, if either one works, it returns right away.
3036 *
3037 * returns 0 on success and < 0 on failure
3038 */
3039static noinline int split_node(struct btrfs_trans_handle *trans,
3040 struct btrfs_root *root,
3041 struct btrfs_path *path, int level)
3042{
3043 struct extent_buffer *c;
3044 struct extent_buffer *split;
3045 struct btrfs_disk_key disk_key;
3046 int mid;
3047 int ret;
3048 u32 c_nritems;
3049
3050 c = path->nodes[level];
3051 WARN_ON(btrfs_header_generation(c) != trans->transid);
3052 if (c == root->node) {
3053 /* trying to split the root, lets make a new one */
3054 ret = insert_new_root(trans, root, path, level + 1);
3055 if (ret)
3056 return ret;
3057 } else {
3058 ret = push_nodes_for_insert(trans, root, path, level);
3059 c = path->nodes[level];
3060 if (!ret && btrfs_header_nritems(c) <
3061 BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
3062 return 0;
3063 if (ret < 0)
3064 return ret;
3065 }
3066
3067 c_nritems = btrfs_header_nritems(c);
3068 mid = (c_nritems + 1) / 2;
3069 btrfs_node_key(c, &disk_key, mid);
3070
3071 split = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
3072 root->root_key.objectid,
3073 &disk_key, level, c->start, 0);
3074 if (IS_ERR(split))
3075 return PTR_ERR(split);
3076
3077 root_add_used(root, root->nodesize);
3078
3079 memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
3080 btrfs_set_header_level(split, btrfs_header_level(c));
3081 btrfs_set_header_bytenr(split, split->start);
3082 btrfs_set_header_generation(split, trans->transid);
3083 btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
3084 btrfs_set_header_owner(split, root->root_key.objectid);
3085 write_extent_buffer(split, root->fs_info->fsid,
3086 (unsigned long)btrfs_header_fsid(split),
3087 BTRFS_FSID_SIZE);
3088 write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
3089 (unsigned long)btrfs_header_chunk_tree_uuid(split),
3090 BTRFS_UUID_SIZE);
3091
3092 tree_mod_log_eb_copy(root->fs_info, split, c, 0, mid, c_nritems - mid);
3093 copy_extent_buffer(split, c,
3094 btrfs_node_key_ptr_offset(0),
3095 btrfs_node_key_ptr_offset(mid),
3096 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3097 btrfs_set_header_nritems(split, c_nritems - mid);
3098 btrfs_set_header_nritems(c, mid);
3099 ret = 0;
3100
3101 btrfs_mark_buffer_dirty(c);
3102 btrfs_mark_buffer_dirty(split);
3103
3104 insert_ptr(trans, root, path, &disk_key, split->start,
3105 path->slots[level + 1] + 1, level + 1);
3106
3107 if (path->slots[level] >= mid) {
3108 path->slots[level] -= mid;
3109 btrfs_tree_unlock(c);
3110 free_extent_buffer(c);
3111 path->nodes[level] = split;
3112 path->slots[level + 1] += 1;
3113 } else {
3114 btrfs_tree_unlock(split);
3115 free_extent_buffer(split);
3116 }
3117 return ret;
3118}
3119
3120/*
3121 * how many bytes are required to store the items in a leaf. start
3122 * and nr indicate which items in the leaf to check. This totals up the
3123 * space used both by the item structs and the item data
3124 */
3125static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3126{
3127 int data_len;
3128 int nritems = btrfs_header_nritems(l);
3129 int end = min(nritems, start + nr) - 1;
3130
3131 if (!nr)
3132 return 0;
3133 data_len = btrfs_item_end_nr(l, start);
3134 data_len = data_len - btrfs_item_offset_nr(l, end);
3135 data_len += sizeof(struct btrfs_item) * nr;
3136 WARN_ON(data_len < 0);
3137 return data_len;
3138}
3139
3140/*
3141 * The space between the end of the leaf items and
3142 * the start of the leaf data. IOW, how much room
3143 * the leaf has left for both items and data
3144 */
3145noinline int btrfs_leaf_free_space(struct btrfs_root *root,
3146 struct extent_buffer *leaf)
3147{
3148 int nritems = btrfs_header_nritems(leaf);
3149 int ret;
3150 ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
3151 if (ret < 0) {
3152 printk(KERN_CRIT "leaf free space ret %d, leaf data size %lu, "
3153 "used %d nritems %d\n",
3154 ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
3155 leaf_space_used(leaf, 0, nritems), nritems);
3156 }
3157 return ret;
3158}
3159
3160/*
3161 * min slot controls the lowest index we're willing to push to the
3162 * right. We'll push up to and including min_slot, but no lower
3163 */
3164static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
3165 struct btrfs_root *root,
3166 struct btrfs_path *path,
3167 int data_size, int empty,
3168 struct extent_buffer *right,
3169 int free_space, u32 left_nritems,
3170 u32 min_slot)
3171{
3172 struct extent_buffer *left = path->nodes[0];
3173 struct extent_buffer *upper = path->nodes[1];
3174 struct btrfs_map_token token;
3175 struct btrfs_disk_key disk_key;
3176 int slot;
3177 u32 i;
3178 int push_space = 0;
3179 int push_items = 0;
3180 struct btrfs_item *item;
3181 u32 nr;
3182 u32 right_nritems;
3183 u32 data_end;
3184 u32 this_item_size;
3185
3186 btrfs_init_map_token(&token);
3187
3188 if (empty)
3189 nr = 0;
3190 else
3191 nr = max_t(u32, 1, min_slot);
3192
3193 if (path->slots[0] >= left_nritems)
3194 push_space += data_size;
3195
3196 slot = path->slots[1];
3197 i = left_nritems - 1;
3198 while (i >= nr) {
3199 item = btrfs_item_nr(left, i);
3200
3201 if (!empty && push_items > 0) {
3202 if (path->slots[0] > i)
3203 break;
3204 if (path->slots[0] == i) {
3205 int space = btrfs_leaf_free_space(root, left);
3206 if (space + push_space * 2 > free_space)
3207 break;
3208 }
3209 }
3210
3211 if (path->slots[0] == i)
3212 push_space += data_size;
3213
3214 this_item_size = btrfs_item_size(left, item);
3215 if (this_item_size + sizeof(*item) + push_space > free_space)
3216 break;
3217
3218 push_items++;
3219 push_space += this_item_size + sizeof(*item);
3220 if (i == 0)
3221 break;
3222 i--;
3223 }
3224
3225 if (push_items == 0)
3226 goto out_unlock;
3227
3228 if (!empty && push_items == left_nritems)
3229 WARN_ON(1);
3230
3231 /* push left to right */
3232 right_nritems = btrfs_header_nritems(right);
3233
3234 push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3235 push_space -= leaf_data_end(root, left);
3236
3237 /* make room in the right data area */
3238 data_end = leaf_data_end(root, right);
3239 memmove_extent_buffer(right,
3240 btrfs_leaf_data(right) + data_end - push_space,
3241 btrfs_leaf_data(right) + data_end,
3242 BTRFS_LEAF_DATA_SIZE(root) - data_end);
3243
3244 /* copy from the left data area */
3245 copy_extent_buffer(right, left, btrfs_leaf_data(right) +
3246 BTRFS_LEAF_DATA_SIZE(root) - push_space,
3247 btrfs_leaf_data(left) + leaf_data_end(root, left),
3248 push_space);
3249
3250 memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3251 btrfs_item_nr_offset(0),
3252 right_nritems * sizeof(struct btrfs_item));
3253
3254 /* copy the items from left to right */
3255 copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3256 btrfs_item_nr_offset(left_nritems - push_items),
3257 push_items * sizeof(struct btrfs_item));
3258
3259 /* update the item pointers */
3260 right_nritems += push_items;
3261 btrfs_set_header_nritems(right, right_nritems);
3262 push_space = BTRFS_LEAF_DATA_SIZE(root);
3263 for (i = 0; i < right_nritems; i++) {
3264 item = btrfs_item_nr(right, i);
3265 push_space -= btrfs_token_item_size(right, item, &token);
3266 btrfs_set_token_item_offset(right, item, push_space, &token);
3267 }
3268
3269 left_nritems -= push_items;
3270 btrfs_set_header_nritems(left, left_nritems);
3271
3272 if (left_nritems)
3273 btrfs_mark_buffer_dirty(left);
3274 else
3275 clean_tree_block(trans, root, left);
3276
3277 btrfs_mark_buffer_dirty(right);
3278
3279 btrfs_item_key(right, &disk_key, 0);
3280 btrfs_set_node_key(upper, &disk_key, slot + 1);
3281 btrfs_mark_buffer_dirty(upper);
3282
3283 /* then fixup the leaf pointer in the path */
3284 if (path->slots[0] >= left_nritems) {
3285 path->slots[0] -= left_nritems;
3286 if (btrfs_header_nritems(path->nodes[0]) == 0)
3287 clean_tree_block(trans, root, path->nodes[0]);
3288 btrfs_tree_unlock(path->nodes[0]);
3289 free_extent_buffer(path->nodes[0]);
3290 path->nodes[0] = right;
3291 path->slots[1] += 1;
3292 } else {
3293 btrfs_tree_unlock(right);
3294 free_extent_buffer(right);
3295 }
3296 return 0;
3297
3298out_unlock:
3299 btrfs_tree_unlock(right);
3300 free_extent_buffer(right);
3301 return 1;
3302}
3303
3304/*
3305 * push some data in the path leaf to the right, trying to free up at
3306 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3307 *
3308 * returns 1 if the push failed because the other node didn't have enough
3309 * room, 0 if everything worked out and < 0 if there were major errors.
3310 *
3311 * this will push starting from min_slot to the end of the leaf. It won't
3312 * push any slot lower than min_slot
3313 */
3314static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3315 *root, struct btrfs_path *path,
3316 int min_data_size, int data_size,
3317 int empty, u32 min_slot)
3318{
3319 struct extent_buffer *left = path->nodes[0];
3320 struct extent_buffer *right;
3321 struct extent_buffer *upper;
3322 int slot;
3323 int free_space;
3324 u32 left_nritems;
3325 int ret;
3326
3327 if (!path->nodes[1])
3328 return 1;
3329
3330 slot = path->slots[1];
3331 upper = path->nodes[1];
3332 if (slot >= btrfs_header_nritems(upper) - 1)
3333 return 1;
3334
3335 btrfs_assert_tree_locked(path->nodes[1]);
3336
3337 right = read_node_slot(root, upper, slot + 1);
3338 if (right == NULL)
3339 return 1;
3340
3341 btrfs_tree_lock(right);
3342 btrfs_set_lock_blocking(right);
3343
3344 free_space = btrfs_leaf_free_space(root, right);
3345 if (free_space < data_size)
3346 goto out_unlock;
3347
3348 /* cow and double check */
3349 ret = btrfs_cow_block(trans, root, right, upper,
3350 slot + 1, &right);
3351 if (ret)
3352 goto out_unlock;
3353
3354 free_space = btrfs_leaf_free_space(root, right);
3355 if (free_space < data_size)
3356 goto out_unlock;
3357
3358 left_nritems = btrfs_header_nritems(left);
3359 if (left_nritems == 0)
3360 goto out_unlock;
3361
3362 return __push_leaf_right(trans, root, path, min_data_size, empty,
3363 right, free_space, left_nritems, min_slot);
3364out_unlock:
3365 btrfs_tree_unlock(right);
3366 free_extent_buffer(right);
3367 return 1;
3368}
3369
3370/*
3371 * push some data in the path leaf to the left, trying to free up at
3372 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3373 *
3374 * max_slot can put a limit on how far into the leaf we'll push items. The
3375 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3376 * items
3377 */
3378static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
3379 struct btrfs_root *root,
3380 struct btrfs_path *path, int data_size,
3381 int empty, struct extent_buffer *left,
3382 int free_space, u32 right_nritems,
3383 u32 max_slot)
3384{
3385 struct btrfs_disk_key disk_key;
3386 struct extent_buffer *right = path->nodes[0];
3387 int i;
3388 int push_space = 0;
3389 int push_items = 0;
3390 struct btrfs_item *item;
3391 u32 old_left_nritems;
3392 u32 nr;
3393 int ret = 0;
3394 u32 this_item_size;
3395 u32 old_left_item_size;
3396 struct btrfs_map_token token;
3397
3398 btrfs_init_map_token(&token);
3399
3400 if (empty)
3401 nr = min(right_nritems, max_slot);
3402 else
3403 nr = min(right_nritems - 1, max_slot);
3404
3405 for (i = 0; i < nr; i++) {
3406 item = btrfs_item_nr(right, i);
3407
3408 if (!empty && push_items > 0) {
3409 if (path->slots[0] < i)
3410 break;
3411 if (path->slots[0] == i) {
3412 int space = btrfs_leaf_free_space(root, right);
3413 if (space + push_space * 2 > free_space)
3414 break;
3415 }
3416 }
3417
3418 if (path->slots[0] == i)
3419 push_space += data_size;
3420
3421 this_item_size = btrfs_item_size(right, item);
3422 if (this_item_size + sizeof(*item) + push_space > free_space)
3423 break;
3424
3425 push_items++;
3426 push_space += this_item_size + sizeof(*item);
3427 }
3428
3429 if (push_items == 0) {
3430 ret = 1;
3431 goto out;
3432 }
3433 if (!empty && push_items == btrfs_header_nritems(right))
3434 WARN_ON(1);
3435
3436 /* push data from right to left */
3437 copy_extent_buffer(left, right,
3438 btrfs_item_nr_offset(btrfs_header_nritems(left)),
3439 btrfs_item_nr_offset(0),
3440 push_items * sizeof(struct btrfs_item));
3441
3442 push_space = BTRFS_LEAF_DATA_SIZE(root) -
3443 btrfs_item_offset_nr(right, push_items - 1);
3444
3445 copy_extent_buffer(left, right, btrfs_leaf_data(left) +
3446 leaf_data_end(root, left) - push_space,
3447 btrfs_leaf_data(right) +
3448 btrfs_item_offset_nr(right, push_items - 1),
3449 push_space);
3450 old_left_nritems = btrfs_header_nritems(left);
3451 BUG_ON(old_left_nritems <= 0);
3452
3453 old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3454 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3455 u32 ioff;
3456
3457 item = btrfs_item_nr(left, i);
3458
3459 ioff = btrfs_token_item_offset(left, item, &token);
3460 btrfs_set_token_item_offset(left, item,
3461 ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size),
3462 &token);
3463 }
3464 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3465
3466 /* fixup right node */
3467 if (push_items > right_nritems) {
3468 printk(KERN_CRIT "push items %d nr %u\n", push_items,
3469 right_nritems);
3470 WARN_ON(1);
3471 }
3472
3473 if (push_items < right_nritems) {
3474 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3475 leaf_data_end(root, right);
3476 memmove_extent_buffer(right, btrfs_leaf_data(right) +
3477 BTRFS_LEAF_DATA_SIZE(root) - push_space,
3478 btrfs_leaf_data(right) +
3479 leaf_data_end(root, right), push_space);
3480
3481 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3482 btrfs_item_nr_offset(push_items),
3483 (btrfs_header_nritems(right) - push_items) *
3484 sizeof(struct btrfs_item));
3485 }
3486 right_nritems -= push_items;
3487 btrfs_set_header_nritems(right, right_nritems);
3488 push_space = BTRFS_LEAF_DATA_SIZE(root);
3489 for (i = 0; i < right_nritems; i++) {
3490 item = btrfs_item_nr(right, i);
3491
3492 push_space = push_space - btrfs_token_item_size(right,
3493 item, &token);
3494 btrfs_set_token_item_offset(right, item, push_space, &token);
3495 }
3496
3497 btrfs_mark_buffer_dirty(left);
3498 if (right_nritems)
3499 btrfs_mark_buffer_dirty(right);
3500 else
3501 clean_tree_block(trans, root, right);
3502
3503 btrfs_item_key(right, &disk_key, 0);
3504 fixup_low_keys(trans, root, path, &disk_key, 1);
3505
3506 /* then fixup the leaf pointer in the path */
3507 if (path->slots[0] < push_items) {
3508 path->slots[0] += old_left_nritems;
3509 btrfs_tree_unlock(path->nodes[0]);
3510 free_extent_buffer(path->nodes[0]);
3511 path->nodes[0] = left;
3512 path->slots[1] -= 1;
3513 } else {
3514 btrfs_tree_unlock(left);
3515 free_extent_buffer(left);
3516 path->slots[0] -= push_items;
3517 }
3518 BUG_ON(path->slots[0] < 0);
3519 return ret;
3520out:
3521 btrfs_tree_unlock(left);
3522 free_extent_buffer(left);
3523 return ret;
3524}
3525
3526/*
3527 * push some data in the path leaf to the left, trying to free up at
3528 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3529 *
3530 * max_slot can put a limit on how far into the leaf we'll push items. The
3531 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
3532 * items
3533 */
3534static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3535 *root, struct btrfs_path *path, int min_data_size,
3536 int data_size, int empty, u32 max_slot)
3537{
3538 struct extent_buffer *right = path->nodes[0];
3539 struct extent_buffer *left;
3540 int slot;
3541 int free_space;
3542 u32 right_nritems;
3543 int ret = 0;
3544
3545 slot = path->slots[1];
3546 if (slot == 0)
3547 return 1;
3548 if (!path->nodes[1])
3549 return 1;
3550
3551 right_nritems = btrfs_header_nritems(right);
3552 if (right_nritems == 0)
3553 return 1;
3554
3555 btrfs_assert_tree_locked(path->nodes[1]);
3556
3557 left = read_node_slot(root, path->nodes[1], slot - 1);
3558 if (left == NULL)
3559 return 1;
3560
3561 btrfs_tree_lock(left);
3562 btrfs_set_lock_blocking(left);
3563
3564 free_space = btrfs_leaf_free_space(root, left);
3565 if (free_space < data_size) {
3566 ret = 1;
3567 goto out;
3568 }
3569
3570 /* cow and double check */
3571 ret = btrfs_cow_block(trans, root, left,
3572 path->nodes[1], slot - 1, &left);
3573 if (ret) {
3574 /* we hit -ENOSPC, but it isn't fatal here */
3575 if (ret == -ENOSPC)
3576 ret = 1;
3577 goto out;
3578 }
3579
3580 free_space = btrfs_leaf_free_space(root, left);
3581 if (free_space < data_size) {
3582 ret = 1;
3583 goto out;
3584 }
3585
3586 return __push_leaf_left(trans, root, path, min_data_size,
3587 empty, left, free_space, right_nritems,
3588 max_slot);
3589out:
3590 btrfs_tree_unlock(left);
3591 free_extent_buffer(left);
3592 return ret;
3593}
3594
3595/*
3596 * split the path's leaf in two, making sure there is at least data_size
3597 * available for the resulting leaf level of the path.
3598 */
3599static noinline void copy_for_split(struct btrfs_trans_handle *trans,
3600 struct btrfs_root *root,
3601 struct btrfs_path *path,
3602 struct extent_buffer *l,
3603 struct extent_buffer *right,
3604 int slot, int mid, int nritems)
3605{
3606 int data_copy_size;
3607 int rt_data_off;
3608 int i;
3609 struct btrfs_disk_key disk_key;
3610 struct btrfs_map_token token;
3611
3612 btrfs_init_map_token(&token);
3613
3614 nritems = nritems - mid;
3615 btrfs_set_header_nritems(right, nritems);
3616 data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
3617
3618 copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
3619 btrfs_item_nr_offset(mid),
3620 nritems * sizeof(struct btrfs_item));
3621
3622 copy_extent_buffer(right, l,
3623 btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
3624 data_copy_size, btrfs_leaf_data(l) +
3625 leaf_data_end(root, l), data_copy_size);
3626
3627 rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
3628 btrfs_item_end_nr(l, mid);
3629
3630 for (i = 0; i < nritems; i++) {
3631 struct btrfs_item *item = btrfs_item_nr(right, i);
3632 u32 ioff;
3633
3634 ioff = btrfs_token_item_offset(right, item, &token);
3635 btrfs_set_token_item_offset(right, item,
3636 ioff + rt_data_off, &token);
3637 }
3638
3639 btrfs_set_header_nritems(l, mid);
3640 btrfs_item_key(right, &disk_key, 0);
3641 insert_ptr(trans, root, path, &disk_key, right->start,
3642 path->slots[1] + 1, 1);
3643
3644 btrfs_mark_buffer_dirty(right);
3645 btrfs_mark_buffer_dirty(l);
3646 BUG_ON(path->slots[0] != slot);
3647
3648 if (mid <= slot) {
3649 btrfs_tree_unlock(path->nodes[0]);
3650 free_extent_buffer(path->nodes[0]);
3651 path->nodes[0] = right;
3652 path->slots[0] -= mid;
3653 path->slots[1] += 1;
3654 } else {
3655 btrfs_tree_unlock(right);
3656 free_extent_buffer(right);
3657 }
3658
3659 BUG_ON(path->slots[0] < 0);
3660}
3661
3662/*
3663 * double splits happen when we need to insert a big item in the middle
3664 * of a leaf. A double split can leave us with 3 mostly empty leaves:
3665 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
3666 * A B C
3667 *
3668 * We avoid this by trying to push the items on either side of our target
3669 * into the adjacent leaves. If all goes well we can avoid the double split
3670 * completely.
3671 */
3672static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
3673 struct btrfs_root *root,
3674 struct btrfs_path *path,
3675 int data_size)
3676{
3677 int ret;
3678 int progress = 0;
3679 int slot;
3680 u32 nritems;
3681
3682 slot = path->slots[0];
3683
3684 /*
3685 * try to push all the items after our slot into the
3686 * right leaf
3687 */
3688 ret = push_leaf_right(trans, root, path, 1, data_size, 0, slot);
3689 if (ret < 0)
3690 return ret;
3691
3692 if (ret == 0)
3693 progress++;
3694
3695 nritems = btrfs_header_nritems(path->nodes[0]);
3696 /*
3697 * our goal is to get our slot at the start or end of a leaf. If
3698 * we've done so we're done
3699 */
3700 if (path->slots[0] == 0 || path->slots[0] == nritems)
3701 return 0;
3702
3703 if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
3704 return 0;
3705
3706 /* try to push all the items before our slot into the next leaf */
3707 slot = path->slots[0];
3708 ret = push_leaf_left(trans, root, path, 1, data_size, 0, slot);
3709 if (ret < 0)
3710 return ret;
3711
3712 if (ret == 0)
3713 progress++;
3714
3715 if (progress)
3716 return 0;
3717 return 1;
3718}
3719
3720/*
3721 * split the path's leaf in two, making sure there is at least data_size
3722 * available for the resulting leaf level of the path.
3723 *
3724 * returns 0 if all went well and < 0 on failure.
3725 */
3726static noinline int split_leaf(struct btrfs_trans_handle *trans,
3727 struct btrfs_root *root,
3728 struct btrfs_key *ins_key,
3729 struct btrfs_path *path, int data_size,
3730 int extend)
3731{
3732 struct btrfs_disk_key disk_key;
3733 struct extent_buffer *l;
3734 u32 nritems;
3735 int mid;
3736 int slot;
3737 struct extent_buffer *right;
3738 int ret = 0;
3739 int wret;
3740 int split;
3741 int num_doubles = 0;
3742 int tried_avoid_double = 0;
3743
3744 l = path->nodes[0];
3745 slot = path->slots[0];
3746 if (extend && data_size + btrfs_item_size_nr(l, slot) +
3747 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
3748 return -EOVERFLOW;
3749
3750 /* first try to make some room by pushing left and right */
3751 if (data_size) {
3752 wret = push_leaf_right(trans, root, path, data_size,
3753 data_size, 0, 0);
3754 if (wret < 0)
3755 return wret;
3756 if (wret) {
3757 wret = push_leaf_left(trans, root, path, data_size,
3758 data_size, 0, (u32)-1);
3759 if (wret < 0)
3760 return wret;
3761 }
3762 l = path->nodes[0];
3763
3764 /* did the pushes work? */
3765 if (btrfs_leaf_free_space(root, l) >= data_size)
3766 return 0;
3767 }
3768
3769 if (!path->nodes[1]) {
3770 ret = insert_new_root(trans, root, path, 1);
3771 if (ret)
3772 return ret;
3773 }
3774again:
3775 split = 1;
3776 l = path->nodes[0];
3777 slot = path->slots[0];
3778 nritems = btrfs_header_nritems(l);
3779 mid = (nritems + 1) / 2;
3780
3781 if (mid <= slot) {
3782 if (nritems == 1 ||
3783 leaf_space_used(l, mid, nritems - mid) + data_size >
3784 BTRFS_LEAF_DATA_SIZE(root)) {
3785 if (slot >= nritems) {
3786 split = 0;
3787 } else {
3788 mid = slot;
3789 if (mid != nritems &&
3790 leaf_space_used(l, mid, nritems - mid) +
3791 data_size > BTRFS_LEAF_DATA_SIZE(root)) {
3792 if (data_size && !tried_avoid_double)
3793 goto push_for_double;
3794 split = 2;
3795 }
3796 }
3797 }
3798 } else {
3799 if (leaf_space_used(l, 0, mid) + data_size >
3800 BTRFS_LEAF_DATA_SIZE(root)) {
3801 if (!extend && data_size && slot == 0) {
3802 split = 0;
3803 } else if ((extend || !data_size) && slot == 0) {
3804 mid = 1;
3805 } else {
3806 mid = slot;
3807 if (mid != nritems &&
3808 leaf_space_used(l, mid, nritems - mid) +
3809 data_size > BTRFS_LEAF_DATA_SIZE(root)) {
3810 if (data_size && !tried_avoid_double)
3811 goto push_for_double;
3812 split = 2 ;
3813 }
3814 }
3815 }
3816 }
3817
3818 if (split == 0)
3819 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3820 else
3821 btrfs_item_key(l, &disk_key, mid);
3822
3823 right = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
3824 root->root_key.objectid,
3825 &disk_key, 0, l->start, 0);
3826 if (IS_ERR(right))
3827 return PTR_ERR(right);
3828
3829 root_add_used(root, root->leafsize);
3830
3831 memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
3832 btrfs_set_header_bytenr(right, right->start);
3833 btrfs_set_header_generation(right, trans->transid);
3834 btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
3835 btrfs_set_header_owner(right, root->root_key.objectid);
3836 btrfs_set_header_level(right, 0);
3837 write_extent_buffer(right, root->fs_info->fsid,
3838 (unsigned long)btrfs_header_fsid(right),
3839 BTRFS_FSID_SIZE);
3840
3841 write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
3842 (unsigned long)btrfs_header_chunk_tree_uuid(right),
3843 BTRFS_UUID_SIZE);
3844
3845 if (split == 0) {
3846 if (mid <= slot) {
3847 btrfs_set_header_nritems(right, 0);
3848 insert_ptr(trans, root, path, &disk_key, right->start,
3849 path->slots[1] + 1, 1);
3850 btrfs_tree_unlock(path->nodes[0]);
3851 free_extent_buffer(path->nodes[0]);
3852 path->nodes[0] = right;
3853 path->slots[0] = 0;
3854 path->slots[1] += 1;
3855 } else {
3856 btrfs_set_header_nritems(right, 0);
3857 insert_ptr(trans, root, path, &disk_key, right->start,
3858 path->slots[1], 1);
3859 btrfs_tree_unlock(path->nodes[0]);
3860 free_extent_buffer(path->nodes[0]);
3861 path->nodes[0] = right;
3862 path->slots[0] = 0;
3863 if (path->slots[1] == 0)
3864 fixup_low_keys(trans, root, path,
3865 &disk_key, 1);
3866 }
3867 btrfs_mark_buffer_dirty(right);
3868 return ret;
3869 }
3870
3871 copy_for_split(trans, root, path, l, right, slot, mid, nritems);
3872
3873 if (split == 2) {
3874 BUG_ON(num_doubles != 0);
3875 num_doubles++;
3876 goto again;
3877 }
3878
3879 return 0;
3880
3881push_for_double:
3882 push_for_double_split(trans, root, path, data_size);
3883 tried_avoid_double = 1;
3884 if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
3885 return 0;
3886 goto again;
3887}
3888
3889static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3890 struct btrfs_root *root,
3891 struct btrfs_path *path, int ins_len)
3892{
3893 struct btrfs_key key;
3894 struct extent_buffer *leaf;
3895 struct btrfs_file_extent_item *fi;
3896 u64 extent_len = 0;
3897 u32 item_size;
3898 int ret;
3899
3900 leaf = path->nodes[0];
3901 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3902
3903 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3904 key.type != BTRFS_EXTENT_CSUM_KEY);
3905
3906 if (btrfs_leaf_free_space(root, leaf) >= ins_len)
3907 return 0;
3908
3909 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3910 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3911 fi = btrfs_item_ptr(leaf, path->slots[0],
3912 struct btrfs_file_extent_item);
3913 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3914 }
3915 btrfs_release_path(path);
3916
3917 path->keep_locks = 1;
3918 path->search_for_split = 1;
3919 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3920 path->search_for_split = 0;
3921 if (ret < 0)
3922 goto err;
3923
3924 ret = -EAGAIN;
3925 leaf = path->nodes[0];
3926 /* if our item isn't there or got smaller, return now */
3927 if (ret > 0 || item_size != btrfs_item_size_nr(leaf, path->slots[0]))
3928 goto err;
3929
3930 /* the leaf has changed, it now has room. return now */
3931 if (btrfs_leaf_free_space(root, path->nodes[0]) >= ins_len)
3932 goto err;
3933
3934 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3935 fi = btrfs_item_ptr(leaf, path->slots[0],
3936 struct btrfs_file_extent_item);
3937 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3938 goto err;
3939 }
3940
3941 btrfs_set_path_blocking(path);
3942 ret = split_leaf(trans, root, &key, path, ins_len, 1);
3943 if (ret)
3944 goto err;
3945
3946 path->keep_locks = 0;
3947 btrfs_unlock_up_safe(path, 1);
3948 return 0;
3949err:
3950 path->keep_locks = 0;
3951 return ret;
3952}
3953
3954static noinline int split_item(struct btrfs_trans_handle *trans,
3955 struct btrfs_root *root,
3956 struct btrfs_path *path,
3957 struct btrfs_key *new_key,
3958 unsigned long split_offset)
3959{
3960 struct extent_buffer *leaf;
3961 struct btrfs_item *item;
3962 struct btrfs_item *new_item;
3963 int slot;
3964 char *buf;
3965 u32 nritems;
3966 u32 item_size;
3967 u32 orig_offset;
3968 struct btrfs_disk_key disk_key;
3969
3970 leaf = path->nodes[0];
3971 BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
3972
3973 btrfs_set_path_blocking(path);
3974
3975 item = btrfs_item_nr(leaf, path->slots[0]);
3976 orig_offset = btrfs_item_offset(leaf, item);
3977 item_size = btrfs_item_size(leaf, item);
3978
3979 buf = kmalloc(item_size, GFP_NOFS);
3980 if (!buf)
3981 return -ENOMEM;
3982
3983 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3984 path->slots[0]), item_size);
3985
3986 slot = path->slots[0] + 1;
3987 nritems = btrfs_header_nritems(leaf);
3988 if (slot != nritems) {
3989 /* shift the items */
3990 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3991 btrfs_item_nr_offset(slot),
3992 (nritems - slot) * sizeof(struct btrfs_item));
3993 }
3994
3995 btrfs_cpu_key_to_disk(&disk_key, new_key);
3996 btrfs_set_item_key(leaf, &disk_key, slot);
3997
3998 new_item = btrfs_item_nr(leaf, slot);
3999
4000 btrfs_set_item_offset(leaf, new_item, orig_offset);
4001 btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4002
4003 btrfs_set_item_offset(leaf, item,
4004 orig_offset + item_size - split_offset);
4005 btrfs_set_item_size(leaf, item, split_offset);
4006
4007 btrfs_set_header_nritems(leaf, nritems + 1);
4008
4009 /* write the data for the start of the original item */
4010 write_extent_buffer(leaf, buf,
4011 btrfs_item_ptr_offset(leaf, path->slots[0]),
4012 split_offset);
4013
4014 /* write the data for the new item */
4015 write_extent_buffer(leaf, buf + split_offset,
4016 btrfs_item_ptr_offset(leaf, slot),
4017 item_size - split_offset);
4018 btrfs_mark_buffer_dirty(leaf);
4019
4020 BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
4021 kfree(buf);
4022 return 0;
4023}
4024
4025/*
4026 * This function splits a single item into two items,
4027 * giving 'new_key' to the new item and splitting the
4028 * old one at split_offset (from the start of the item).
4029 *
4030 * The path may be released by this operation. After
4031 * the split, the path is pointing to the old item. The
4032 * new item is going to be in the same node as the old one.
4033 *
4034 * Note, the item being split must be smaller enough to live alone on
4035 * a tree block with room for one extra struct btrfs_item
4036 *
4037 * This allows us to split the item in place, keeping a lock on the
4038 * leaf the entire time.
4039 */
4040int btrfs_split_item(struct btrfs_trans_handle *trans,
4041 struct btrfs_root *root,
4042 struct btrfs_path *path,
4043 struct btrfs_key *new_key,
4044 unsigned long split_offset)
4045{
4046 int ret;
4047 ret = setup_leaf_for_split(trans, root, path,
4048 sizeof(struct btrfs_item));
4049 if (ret)
4050 return ret;
4051
4052 ret = split_item(trans, root, path, new_key, split_offset);
4053 return ret;
4054}
4055
4056/*
4057 * This function duplicate a item, giving 'new_key' to the new item.
4058 * It guarantees both items live in the same tree leaf and the new item
4059 * is contiguous with the original item.
4060 *
4061 * This allows us to split file extent in place, keeping a lock on the
4062 * leaf the entire time.
4063 */
4064int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4065 struct btrfs_root *root,
4066 struct btrfs_path *path,
4067 struct btrfs_key *new_key)
4068{
4069 struct extent_buffer *leaf;
4070 int ret;
4071 u32 item_size;
4072
4073 leaf = path->nodes[0];
4074 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4075 ret = setup_leaf_for_split(trans, root, path,
4076 item_size + sizeof(struct btrfs_item));
4077 if (ret)
4078 return ret;
4079
4080 path->slots[0]++;
4081 setup_items_for_insert(trans, root, path, new_key, &item_size,
4082 item_size, item_size +
4083 sizeof(struct btrfs_item), 1);
4084 leaf = path->nodes[0];
4085 memcpy_extent_buffer(leaf,
4086 btrfs_item_ptr_offset(leaf, path->slots[0]),
4087 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4088 item_size);
4089 return 0;
4090}
4091
4092/*
4093 * make the item pointed to by the path smaller. new_size indicates
4094 * how small to make it, and from_end tells us if we just chop bytes
4095 * off the end of the item or if we shift the item to chop bytes off
4096 * the front.
4097 */
4098void btrfs_truncate_item(struct btrfs_trans_handle *trans,
4099 struct btrfs_root *root,
4100 struct btrfs_path *path,
4101 u32 new_size, int from_end)
4102{
4103 int slot;
4104 struct extent_buffer *leaf;
4105 struct btrfs_item *item;
4106 u32 nritems;
4107 unsigned int data_end;
4108 unsigned int old_data_start;
4109 unsigned int old_size;
4110 unsigned int size_diff;
4111 int i;
4112 struct btrfs_map_token token;
4113
4114 btrfs_init_map_token(&token);
4115
4116 leaf = path->nodes[0];
4117 slot = path->slots[0];
4118
4119 old_size = btrfs_item_size_nr(leaf, slot);
4120 if (old_size == new_size)
4121 return;
4122
4123 nritems = btrfs_header_nritems(leaf);
4124 data_end = leaf_data_end(root, leaf);
4125
4126 old_data_start = btrfs_item_offset_nr(leaf, slot);
4127
4128 size_diff = old_size - new_size;
4129
4130 BUG_ON(slot < 0);
4131 BUG_ON(slot >= nritems);
4132
4133 /*
4134 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4135 */
4136 /* first correct the data pointers */
4137 for (i = slot; i < nritems; i++) {
4138 u32 ioff;
4139 item = btrfs_item_nr(leaf, i);
4140
4141 ioff = btrfs_token_item_offset(leaf, item, &token);
4142 btrfs_set_token_item_offset(leaf, item,
4143 ioff + size_diff, &token);
4144 }
4145
4146 /* shift the data */
4147 if (from_end) {
4148 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4149 data_end + size_diff, btrfs_leaf_data(leaf) +
4150 data_end, old_data_start + new_size - data_end);
4151 } else {
4152 struct btrfs_disk_key disk_key;
4153 u64 offset;
4154
4155 btrfs_item_key(leaf, &disk_key, slot);
4156
4157 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4158 unsigned long ptr;
4159 struct btrfs_file_extent_item *fi;
4160
4161 fi = btrfs_item_ptr(leaf, slot,
4162 struct btrfs_file_extent_item);
4163 fi = (struct btrfs_file_extent_item *)(
4164 (unsigned long)fi - size_diff);
4165
4166 if (btrfs_file_extent_type(leaf, fi) ==
4167 BTRFS_FILE_EXTENT_INLINE) {
4168 ptr = btrfs_item_ptr_offset(leaf, slot);
4169 memmove_extent_buffer(leaf, ptr,
4170 (unsigned long)fi,
4171 offsetof(struct btrfs_file_extent_item,
4172 disk_bytenr));
4173 }
4174 }
4175
4176 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4177 data_end + size_diff, btrfs_leaf_data(leaf) +
4178 data_end, old_data_start - data_end);
4179
4180 offset = btrfs_disk_key_offset(&disk_key);
4181 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4182 btrfs_set_item_key(leaf, &disk_key, slot);
4183 if (slot == 0)
4184 fixup_low_keys(trans, root, path, &disk_key, 1);
4185 }
4186
4187 item = btrfs_item_nr(leaf, slot);
4188 btrfs_set_item_size(leaf, item, new_size);
4189 btrfs_mark_buffer_dirty(leaf);
4190
4191 if (btrfs_leaf_free_space(root, leaf) < 0) {
4192 btrfs_print_leaf(root, leaf);
4193 BUG();
4194 }
4195}
4196
4197/*
4198 * make the item pointed to by the path bigger, data_size is the new size.
4199 */
4200void btrfs_extend_item(struct btrfs_trans_handle *trans,
4201 struct btrfs_root *root, struct btrfs_path *path,
4202 u32 data_size)
4203{
4204 int slot;
4205 struct extent_buffer *leaf;
4206 struct btrfs_item *item;
4207 u32 nritems;
4208 unsigned int data_end;
4209 unsigned int old_data;
4210 unsigned int old_size;
4211 int i;
4212 struct btrfs_map_token token;
4213
4214 btrfs_init_map_token(&token);
4215
4216 leaf = path->nodes[0];
4217
4218 nritems = btrfs_header_nritems(leaf);
4219 data_end = leaf_data_end(root, leaf);
4220
4221 if (btrfs_leaf_free_space(root, leaf) < data_size) {
4222 btrfs_print_leaf(root, leaf);
4223 BUG();
4224 }
4225 slot = path->slots[0];
4226 old_data = btrfs_item_end_nr(leaf, slot);
4227
4228 BUG_ON(slot < 0);
4229 if (slot >= nritems) {
4230 btrfs_print_leaf(root, leaf);
4231 printk(KERN_CRIT "slot %d too large, nritems %d\n",
4232 slot, nritems);
4233 BUG_ON(1);
4234 }
4235
4236 /*
4237 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4238 */
4239 /* first correct the data pointers */
4240 for (i = slot; i < nritems; i++) {
4241 u32 ioff;
4242 item = btrfs_item_nr(leaf, i);
4243
4244 ioff = btrfs_token_item_offset(leaf, item, &token);
4245 btrfs_set_token_item_offset(leaf, item,
4246 ioff - data_size, &token);
4247 }
4248
4249 /* shift the data */
4250 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4251 data_end - data_size, btrfs_leaf_data(leaf) +
4252 data_end, old_data - data_end);
4253
4254 data_end = old_data;
4255 old_size = btrfs_item_size_nr(leaf, slot);
4256 item = btrfs_item_nr(leaf, slot);
4257 btrfs_set_item_size(leaf, item, old_size + data_size);
4258 btrfs_mark_buffer_dirty(leaf);
4259
4260 if (btrfs_leaf_free_space(root, leaf) < 0) {
4261 btrfs_print_leaf(root, leaf);
4262 BUG();
4263 }
4264}
4265
4266/*
4267 * Given a key and some data, insert items into the tree.
4268 * This does all the path init required, making room in the tree if needed.
4269 * Returns the number of keys that were inserted.
4270 */
4271int btrfs_insert_some_items(struct btrfs_trans_handle *trans,
4272 struct btrfs_root *root,
4273 struct btrfs_path *path,
4274 struct btrfs_key *cpu_key, u32 *data_size,
4275 int nr)
4276{
4277 struct extent_buffer *leaf;
4278 struct btrfs_item *item;
4279 int ret = 0;
4280 int slot;
4281 int i;
4282 u32 nritems;
4283 u32 total_data = 0;
4284 u32 total_size = 0;
4285 unsigned int data_end;
4286 struct btrfs_disk_key disk_key;
4287 struct btrfs_key found_key;
4288 struct btrfs_map_token token;
4289
4290 btrfs_init_map_token(&token);
4291
4292 for (i = 0; i < nr; i++) {
4293 if (total_size + data_size[i] + sizeof(struct btrfs_item) >
4294 BTRFS_LEAF_DATA_SIZE(root)) {
4295 break;
4296 nr = i;
4297 }
4298 total_data += data_size[i];
4299 total_size += data_size[i] + sizeof(struct btrfs_item);
4300 }
4301 BUG_ON(nr == 0);
4302
4303 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4304 if (ret == 0)
4305 return -EEXIST;
4306 if (ret < 0)
4307 goto out;
4308
4309 leaf = path->nodes[0];
4310
4311 nritems = btrfs_header_nritems(leaf);
4312 data_end = leaf_data_end(root, leaf);
4313
4314 if (btrfs_leaf_free_space(root, leaf) < total_size) {
4315 for (i = nr; i >= 0; i--) {
4316 total_data -= data_size[i];
4317 total_size -= data_size[i] + sizeof(struct btrfs_item);
4318 if (total_size < btrfs_leaf_free_space(root, leaf))
4319 break;
4320 }
4321 nr = i;
4322 }
4323
4324 slot = path->slots[0];
4325 BUG_ON(slot < 0);
4326
4327 if (slot != nritems) {
4328 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4329
4330 item = btrfs_item_nr(leaf, slot);
4331 btrfs_item_key_to_cpu(leaf, &found_key, slot);
4332
4333 /* figure out how many keys we can insert in here */
4334 total_data = data_size[0];
4335 for (i = 1; i < nr; i++) {
4336 if (btrfs_comp_cpu_keys(&found_key, cpu_key + i) <= 0)
4337 break;
4338 total_data += data_size[i];
4339 }
4340 nr = i;
4341
4342 if (old_data < data_end) {
4343 btrfs_print_leaf(root, leaf);
4344 printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
4345 slot, old_data, data_end);
4346 BUG_ON(1);
4347 }
4348 /*
4349 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4350 */
4351 /* first correct the data pointers */
4352 for (i = slot; i < nritems; i++) {
4353 u32 ioff;
4354
4355 item = btrfs_item_nr(leaf, i);
4356 ioff = btrfs_token_item_offset(leaf, item, &token);
4357 btrfs_set_token_item_offset(leaf, item,
4358 ioff - total_data, &token);
4359 }
4360 /* shift the items */
4361 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4362 btrfs_item_nr_offset(slot),
4363 (nritems - slot) * sizeof(struct btrfs_item));
4364
4365 /* shift the data */
4366 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4367 data_end - total_data, btrfs_leaf_data(leaf) +
4368 data_end, old_data - data_end);
4369 data_end = old_data;
4370 } else {
4371 /*
4372 * this sucks but it has to be done, if we are inserting at
4373 * the end of the leaf only insert 1 of the items, since we
4374 * have no way of knowing whats on the next leaf and we'd have
4375 * to drop our current locks to figure it out
4376 */
4377 nr = 1;
4378 }
4379
4380 /* setup the item for the new data */
4381 for (i = 0; i < nr; i++) {
4382 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4383 btrfs_set_item_key(leaf, &disk_key, slot + i);
4384 item = btrfs_item_nr(leaf, slot + i);
4385 btrfs_set_token_item_offset(leaf, item,
4386 data_end - data_size[i], &token);
4387 data_end -= data_size[i];
4388 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4389 }
4390 btrfs_set_header_nritems(leaf, nritems + nr);
4391 btrfs_mark_buffer_dirty(leaf);
4392
4393 ret = 0;
4394 if (slot == 0) {
4395 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4396 fixup_low_keys(trans, root, path, &disk_key, 1);
4397 }
4398
4399 if (btrfs_leaf_free_space(root, leaf) < 0) {
4400 btrfs_print_leaf(root, leaf);
4401 BUG();
4402 }
4403out:
4404 if (!ret)
4405 ret = nr;
4406 return ret;
4407}
4408
4409/*
4410 * this is a helper for btrfs_insert_empty_items, the main goal here is
4411 * to save stack depth by doing the bulk of the work in a function
4412 * that doesn't call btrfs_search_slot
4413 */
4414void setup_items_for_insert(struct btrfs_trans_handle *trans,
4415 struct btrfs_root *root, struct btrfs_path *path,
4416 struct btrfs_key *cpu_key, u32 *data_size,
4417 u32 total_data, u32 total_size, int nr)
4418{
4419 struct btrfs_item *item;
4420 int i;
4421 u32 nritems;
4422 unsigned int data_end;
4423 struct btrfs_disk_key disk_key;
4424 struct extent_buffer *leaf;
4425 int slot;
4426 struct btrfs_map_token token;
4427
4428 btrfs_init_map_token(&token);
4429
4430 leaf = path->nodes[0];
4431 slot = path->slots[0];
4432
4433 nritems = btrfs_header_nritems(leaf);
4434 data_end = leaf_data_end(root, leaf);
4435
4436 if (btrfs_leaf_free_space(root, leaf) < total_size) {
4437 btrfs_print_leaf(root, leaf);
4438 printk(KERN_CRIT "not enough freespace need %u have %d\n",
4439 total_size, btrfs_leaf_free_space(root, leaf));
4440 BUG();
4441 }
4442
4443 if (slot != nritems) {
4444 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4445
4446 if (old_data < data_end) {
4447 btrfs_print_leaf(root, leaf);
4448 printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
4449 slot, old_data, data_end);
4450 BUG_ON(1);
4451 }
4452 /*
4453 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4454 */
4455 /* first correct the data pointers */
4456 for (i = slot; i < nritems; i++) {
4457 u32 ioff;
4458
4459 item = btrfs_item_nr(leaf, i);
4460 ioff = btrfs_token_item_offset(leaf, item, &token);
4461 btrfs_set_token_item_offset(leaf, item,
4462 ioff - total_data, &token);
4463 }
4464 /* shift the items */
4465 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4466 btrfs_item_nr_offset(slot),
4467 (nritems - slot) * sizeof(struct btrfs_item));
4468
4469 /* shift the data */
4470 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4471 data_end - total_data, btrfs_leaf_data(leaf) +
4472 data_end, old_data - data_end);
4473 data_end = old_data;
4474 }
4475
4476 /* setup the item for the new data */
4477 for (i = 0; i < nr; i++) {
4478 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4479 btrfs_set_item_key(leaf, &disk_key, slot + i);
4480 item = btrfs_item_nr(leaf, slot + i);
4481 btrfs_set_token_item_offset(leaf, item,
4482 data_end - data_size[i], &token);
4483 data_end -= data_size[i];
4484 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4485 }
4486
4487 btrfs_set_header_nritems(leaf, nritems + nr);
4488
4489 if (slot == 0) {
4490 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4491 fixup_low_keys(trans, root, path, &disk_key, 1);
4492 }
4493 btrfs_unlock_up_safe(path, 1);
4494 btrfs_mark_buffer_dirty(leaf);
4495
4496 if (btrfs_leaf_free_space(root, leaf) < 0) {
4497 btrfs_print_leaf(root, leaf);
4498 BUG();
4499 }
4500}
4501
4502/*
4503 * Given a key and some data, insert items into the tree.
4504 * This does all the path init required, making room in the tree if needed.
4505 */
4506int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4507 struct btrfs_root *root,
4508 struct btrfs_path *path,
4509 struct btrfs_key *cpu_key, u32 *data_size,
4510 int nr)
4511{
4512 int ret = 0;
4513 int slot;
4514 int i;
4515 u32 total_size = 0;
4516 u32 total_data = 0;
4517
4518 for (i = 0; i < nr; i++)
4519 total_data += data_size[i];
4520
4521 total_size = total_data + (nr * sizeof(struct btrfs_item));
4522 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4523 if (ret == 0)
4524 return -EEXIST;
4525 if (ret < 0)
4526 return ret;
4527
4528 slot = path->slots[0];
4529 BUG_ON(slot < 0);
4530
4531 setup_items_for_insert(trans, root, path, cpu_key, data_size,
4532 total_data, total_size, nr);
4533 return 0;
4534}
4535
4536/*
4537 * Given a key and some data, insert an item into the tree.
4538 * This does all the path init required, making room in the tree if needed.
4539 */
4540int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
4541 *root, struct btrfs_key *cpu_key, void *data, u32
4542 data_size)
4543{
4544 int ret = 0;
4545 struct btrfs_path *path;
4546 struct extent_buffer *leaf;
4547 unsigned long ptr;
4548
4549 path = btrfs_alloc_path();
4550 if (!path)
4551 return -ENOMEM;
4552 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4553 if (!ret) {
4554 leaf = path->nodes[0];
4555 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4556 write_extent_buffer(leaf, data, ptr, data_size);
4557 btrfs_mark_buffer_dirty(leaf);
4558 }
4559 btrfs_free_path(path);
4560 return ret;
4561}
4562
4563/*
4564 * delete the pointer from a given node.
4565 *
4566 * the tree should have been previously balanced so the deletion does not
4567 * empty a node.
4568 */
4569static void del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4570 struct btrfs_path *path, int level, int slot,
4571 int tree_mod_log)
4572{
4573 struct extent_buffer *parent = path->nodes[level];
4574 u32 nritems;
4575 int ret;
4576
4577 nritems = btrfs_header_nritems(parent);
4578 if (slot != nritems - 1) {
4579 if (tree_mod_log && level)
4580 tree_mod_log_eb_move(root->fs_info, parent, slot,
4581 slot + 1, nritems - slot - 1);
4582 memmove_extent_buffer(parent,
4583 btrfs_node_key_ptr_offset(slot),
4584 btrfs_node_key_ptr_offset(slot + 1),
4585 sizeof(struct btrfs_key_ptr) *
4586 (nritems - slot - 1));
4587 } else if (tree_mod_log && level) {
4588 ret = tree_mod_log_insert_key(root->fs_info, parent, slot,
4589 MOD_LOG_KEY_REMOVE);
4590 BUG_ON(ret < 0);
4591 }
4592
4593 nritems--;
4594 btrfs_set_header_nritems(parent, nritems);
4595 if (nritems == 0 && parent == root->node) {
4596 BUG_ON(btrfs_header_level(root->node) != 1);
4597 /* just turn the root into a leaf and break */
4598 btrfs_set_header_level(root->node, 0);
4599 } else if (slot == 0) {
4600 struct btrfs_disk_key disk_key;
4601
4602 btrfs_node_key(parent, &disk_key, 0);
4603 fixup_low_keys(trans, root, path, &disk_key, level + 1);
4604 }
4605 btrfs_mark_buffer_dirty(parent);
4606}
4607
4608/*
4609 * a helper function to delete the leaf pointed to by path->slots[1] and
4610 * path->nodes[1].
4611 *
4612 * This deletes the pointer in path->nodes[1] and frees the leaf
4613 * block extent. zero is returned if it all worked out, < 0 otherwise.
4614 *
4615 * The path must have already been setup for deleting the leaf, including
4616 * all the proper balancing. path->nodes[1] must be locked.
4617 */
4618static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4619 struct btrfs_root *root,
4620 struct btrfs_path *path,
4621 struct extent_buffer *leaf)
4622{
4623 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4624 del_ptr(trans, root, path, 1, path->slots[1], 1);
4625
4626 /*
4627 * btrfs_free_extent is expensive, we want to make sure we
4628 * aren't holding any locks when we call it
4629 */
4630 btrfs_unlock_up_safe(path, 0);
4631
4632 root_sub_used(root, leaf->len);
4633
4634 extent_buffer_get(leaf);
4635 btrfs_free_tree_block(trans, root, leaf, 0, 1);
4636 free_extent_buffer_stale(leaf);
4637}
4638/*
4639 * delete the item at the leaf level in path. If that empties
4640 * the leaf, remove it from the tree
4641 */
4642int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4643 struct btrfs_path *path, int slot, int nr)
4644{
4645 struct extent_buffer *leaf;
4646 struct btrfs_item *item;
4647 int last_off;
4648 int dsize = 0;
4649 int ret = 0;
4650 int wret;
4651 int i;
4652 u32 nritems;
4653 struct btrfs_map_token token;
4654
4655 btrfs_init_map_token(&token);
4656
4657 leaf = path->nodes[0];
4658 last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4659
4660 for (i = 0; i < nr; i++)
4661 dsize += btrfs_item_size_nr(leaf, slot + i);
4662
4663 nritems = btrfs_header_nritems(leaf);
4664
4665 if (slot + nr != nritems) {
4666 int data_end = leaf_data_end(root, leaf);
4667
4668 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4669 data_end + dsize,
4670 btrfs_leaf_data(leaf) + data_end,
4671 last_off - data_end);
4672
4673 for (i = slot + nr; i < nritems; i++) {
4674 u32 ioff;
4675
4676 item = btrfs_item_nr(leaf, i);
4677 ioff = btrfs_token_item_offset(leaf, item, &token);
4678 btrfs_set_token_item_offset(leaf, item,
4679 ioff + dsize, &token);
4680 }
4681
4682 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4683 btrfs_item_nr_offset(slot + nr),
4684 sizeof(struct btrfs_item) *
4685 (nritems - slot - nr));
4686 }
4687 btrfs_set_header_nritems(leaf, nritems - nr);
4688 nritems -= nr;
4689
4690 /* delete the leaf if we've emptied it */
4691 if (nritems == 0) {
4692 if (leaf == root->node) {
4693 btrfs_set_header_level(leaf, 0);
4694 } else {
4695 btrfs_set_path_blocking(path);
4696 clean_tree_block(trans, root, leaf);
4697 btrfs_del_leaf(trans, root, path, leaf);
4698 }
4699 } else {
4700 int used = leaf_space_used(leaf, 0, nritems);
4701 if (slot == 0) {
4702 struct btrfs_disk_key disk_key;
4703
4704 btrfs_item_key(leaf, &disk_key, 0);
4705 fixup_low_keys(trans, root, path, &disk_key, 1);
4706 }
4707
4708 /* delete the leaf if it is mostly empty */
4709 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
4710 /* push_leaf_left fixes the path.
4711 * make sure the path still points to our leaf
4712 * for possible call to del_ptr below
4713 */
4714 slot = path->slots[1];
4715 extent_buffer_get(leaf);
4716
4717 btrfs_set_path_blocking(path);
4718 wret = push_leaf_left(trans, root, path, 1, 1,
4719 1, (u32)-1);
4720 if (wret < 0 && wret != -ENOSPC)
4721 ret = wret;
4722
4723 if (path->nodes[0] == leaf &&
4724 btrfs_header_nritems(leaf)) {
4725 wret = push_leaf_right(trans, root, path, 1,
4726 1, 1, 0);
4727 if (wret < 0 && wret != -ENOSPC)
4728 ret = wret;
4729 }
4730
4731 if (btrfs_header_nritems(leaf) == 0) {
4732 path->slots[1] = slot;
4733 btrfs_del_leaf(trans, root, path, leaf);
4734 free_extent_buffer(leaf);
4735 ret = 0;
4736 } else {
4737 /* if we're still in the path, make sure
4738 * we're dirty. Otherwise, one of the
4739 * push_leaf functions must have already
4740 * dirtied this buffer
4741 */
4742 if (path->nodes[0] == leaf)
4743 btrfs_mark_buffer_dirty(leaf);
4744 free_extent_buffer(leaf);
4745 }
4746 } else {
4747 btrfs_mark_buffer_dirty(leaf);
4748 }
4749 }
4750 return ret;
4751}
4752
4753/*
4754 * search the tree again to find a leaf with lesser keys
4755 * returns 0 if it found something or 1 if there are no lesser leaves.
4756 * returns < 0 on io errors.
4757 *
4758 * This may release the path, and so you may lose any locks held at the
4759 * time you call it.
4760 */
4761int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
4762{
4763 struct btrfs_key key;
4764 struct btrfs_disk_key found_key;
4765 int ret;
4766
4767 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
4768
4769 if (key.offset > 0)
4770 key.offset--;
4771 else if (key.type > 0)
4772 key.type--;
4773 else if (key.objectid > 0)
4774 key.objectid--;
4775 else
4776 return 1;
4777
4778 btrfs_release_path(path);
4779 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4780 if (ret < 0)
4781 return ret;
4782 btrfs_item_key(path->nodes[0], &found_key, 0);
4783 ret = comp_keys(&found_key, &key);
4784 if (ret < 0)
4785 return 0;
4786 return 1;
4787}
4788
4789/*
4790 * A helper function to walk down the tree starting at min_key, and looking
4791 * for nodes or leaves that are either in cache or have a minimum
4792 * transaction id. This is used by the btree defrag code, and tree logging
4793 *
4794 * This does not cow, but it does stuff the starting key it finds back
4795 * into min_key, so you can call btrfs_search_slot with cow=1 on the
4796 * key and get a writable path.
4797 *
4798 * This does lock as it descends, and path->keep_locks should be set
4799 * to 1 by the caller.
4800 *
4801 * This honors path->lowest_level to prevent descent past a given level
4802 * of the tree.
4803 *
4804 * min_trans indicates the oldest transaction that you are interested
4805 * in walking through. Any nodes or leaves older than min_trans are
4806 * skipped over (without reading them).
4807 *
4808 * returns zero if something useful was found, < 0 on error and 1 if there
4809 * was nothing in the tree that matched the search criteria.
4810 */
4811int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4812 struct btrfs_key *max_key,
4813 struct btrfs_path *path, int cache_only,
4814 u64 min_trans)
4815{
4816 struct extent_buffer *cur;
4817 struct btrfs_key found_key;
4818 int slot;
4819 int sret;
4820 u32 nritems;
4821 int level;
4822 int ret = 1;
4823
4824 WARN_ON(!path->keep_locks);
4825again:
4826 cur = btrfs_read_lock_root_node(root);
4827 level = btrfs_header_level(cur);
4828 WARN_ON(path->nodes[level]);
4829 path->nodes[level] = cur;
4830 path->locks[level] = BTRFS_READ_LOCK;
4831
4832 if (btrfs_header_generation(cur) < min_trans) {
4833 ret = 1;
4834 goto out;
4835 }
4836 while (1) {
4837 nritems = btrfs_header_nritems(cur);
4838 level = btrfs_header_level(cur);
4839 sret = bin_search(cur, min_key, level, &slot);
4840
4841 /* at the lowest level, we're done, setup the path and exit */
4842 if (level == path->lowest_level) {
4843 if (slot >= nritems)
4844 goto find_next_key;
4845 ret = 0;
4846 path->slots[level] = slot;
4847 btrfs_item_key_to_cpu(cur, &found_key, slot);
4848 goto out;
4849 }
4850 if (sret && slot > 0)
4851 slot--;
4852 /*
4853 * check this node pointer against the cache_only and
4854 * min_trans parameters. If it isn't in cache or is too
4855 * old, skip to the next one.
4856 */
4857 while (slot < nritems) {
4858 u64 blockptr;
4859 u64 gen;
4860 struct extent_buffer *tmp;
4861 struct btrfs_disk_key disk_key;
4862
4863 blockptr = btrfs_node_blockptr(cur, slot);
4864 gen = btrfs_node_ptr_generation(cur, slot);
4865 if (gen < min_trans) {
4866 slot++;
4867 continue;
4868 }
4869 if (!cache_only)
4870 break;
4871
4872 if (max_key) {
4873 btrfs_node_key(cur, &disk_key, slot);
4874 if (comp_keys(&disk_key, max_key) >= 0) {
4875 ret = 1;
4876 goto out;
4877 }
4878 }
4879
4880 tmp = btrfs_find_tree_block(root, blockptr,
4881 btrfs_level_size(root, level - 1));
4882
4883 if (tmp && btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
4884 free_extent_buffer(tmp);
4885 break;
4886 }
4887 if (tmp)
4888 free_extent_buffer(tmp);
4889 slot++;
4890 }
4891find_next_key:
4892 /*
4893 * we didn't find a candidate key in this node, walk forward
4894 * and find another one
4895 */
4896 if (slot >= nritems) {
4897 path->slots[level] = slot;
4898 btrfs_set_path_blocking(path);
4899 sret = btrfs_find_next_key(root, path, min_key, level,
4900 cache_only, min_trans);
4901 if (sret == 0) {
4902 btrfs_release_path(path);
4903 goto again;
4904 } else {
4905 goto out;
4906 }
4907 }
4908 /* save our key for returning back */
4909 btrfs_node_key_to_cpu(cur, &found_key, slot);
4910 path->slots[level] = slot;
4911 if (level == path->lowest_level) {
4912 ret = 0;
4913 unlock_up(path, level, 1, 0, NULL);
4914 goto out;
4915 }
4916 btrfs_set_path_blocking(path);
4917 cur = read_node_slot(root, cur, slot);
4918 BUG_ON(!cur); /* -ENOMEM */
4919
4920 btrfs_tree_read_lock(cur);
4921
4922 path->locks[level - 1] = BTRFS_READ_LOCK;
4923 path->nodes[level - 1] = cur;
4924 unlock_up(path, level, 1, 0, NULL);
4925 btrfs_clear_path_blocking(path, NULL, 0);
4926 }
4927out:
4928 if (ret == 0)
4929 memcpy(min_key, &found_key, sizeof(found_key));
4930 btrfs_set_path_blocking(path);
4931 return ret;
4932}
4933
4934/*
4935 * this is similar to btrfs_next_leaf, but does not try to preserve
4936 * and fixup the path. It looks for and returns the next key in the
4937 * tree based on the current path and the cache_only and min_trans
4938 * parameters.
4939 *
4940 * 0 is returned if another key is found, < 0 if there are any errors
4941 * and 1 is returned if there are no higher keys in the tree
4942 *
4943 * path->keep_locks should be set to 1 on the search made before
4944 * calling this function.
4945 */
4946int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4947 struct btrfs_key *key, int level,
4948 int cache_only, u64 min_trans)
4949{
4950 int slot;
4951 struct extent_buffer *c;
4952
4953 WARN_ON(!path->keep_locks);
4954 while (level < BTRFS_MAX_LEVEL) {
4955 if (!path->nodes[level])
4956 return 1;
4957
4958 slot = path->slots[level] + 1;
4959 c = path->nodes[level];
4960next:
4961 if (slot >= btrfs_header_nritems(c)) {
4962 int ret;
4963 int orig_lowest;
4964 struct btrfs_key cur_key;
4965 if (level + 1 >= BTRFS_MAX_LEVEL ||
4966 !path->nodes[level + 1])
4967 return 1;
4968
4969 if (path->locks[level + 1]) {
4970 level++;
4971 continue;
4972 }
4973
4974 slot = btrfs_header_nritems(c) - 1;
4975 if (level == 0)
4976 btrfs_item_key_to_cpu(c, &cur_key, slot);
4977 else
4978 btrfs_node_key_to_cpu(c, &cur_key, slot);
4979
4980 orig_lowest = path->lowest_level;
4981 btrfs_release_path(path);
4982 path->lowest_level = level;
4983 ret = btrfs_search_slot(NULL, root, &cur_key, path,
4984 0, 0);
4985 path->lowest_level = orig_lowest;
4986 if (ret < 0)
4987 return ret;
4988
4989 c = path->nodes[level];
4990 slot = path->slots[level];
4991 if (ret == 0)
4992 slot++;
4993 goto next;
4994 }
4995
4996 if (level == 0)
4997 btrfs_item_key_to_cpu(c, key, slot);
4998 else {
4999 u64 blockptr = btrfs_node_blockptr(c, slot);
5000 u64 gen = btrfs_node_ptr_generation(c, slot);
5001
5002 if (cache_only) {
5003 struct extent_buffer *cur;
5004 cur = btrfs_find_tree_block(root, blockptr,
5005 btrfs_level_size(root, level - 1));
5006 if (!cur ||
5007 btrfs_buffer_uptodate(cur, gen, 1) <= 0) {
5008 slot++;
5009 if (cur)
5010 free_extent_buffer(cur);
5011 goto next;
5012 }
5013 free_extent_buffer(cur);
5014 }
5015 if (gen < min_trans) {
5016 slot++;
5017 goto next;
5018 }
5019 btrfs_node_key_to_cpu(c, key, slot);
5020 }
5021 return 0;
5022 }
5023 return 1;
5024}
5025
5026/*
5027 * search the tree again to find a leaf with greater keys
5028 * returns 0 if it found something or 1 if there are no greater leaves.
5029 * returns < 0 on io errors.
5030 */
5031int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5032{
5033 return btrfs_next_old_leaf(root, path, 0);
5034}
5035
5036int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5037 u64 time_seq)
5038{
5039 int slot;
5040 int level;
5041 struct extent_buffer *c;
5042 struct extent_buffer *next;
5043 struct btrfs_key key;
5044 u32 nritems;
5045 int ret;
5046 int old_spinning = path->leave_spinning;
5047 int next_rw_lock = 0;
5048
5049 nritems = btrfs_header_nritems(path->nodes[0]);
5050 if (nritems == 0)
5051 return 1;
5052
5053 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5054again:
5055 level = 1;
5056 next = NULL;
5057 next_rw_lock = 0;
5058 btrfs_release_path(path);
5059
5060 path->keep_locks = 1;
5061 path->leave_spinning = 1;
5062
5063 if (time_seq)
5064 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5065 else
5066 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5067 path->keep_locks = 0;
5068
5069 if (ret < 0)
5070 return ret;
5071
5072 nritems = btrfs_header_nritems(path->nodes[0]);
5073 /*
5074 * by releasing the path above we dropped all our locks. A balance
5075 * could have added more items next to the key that used to be
5076 * at the very end of the block. So, check again here and
5077 * advance the path if there are now more items available.
5078 */
5079 if (nritems > 0 && path->slots[0] < nritems - 1) {
5080 if (ret == 0)
5081 path->slots[0]++;
5082 ret = 0;
5083 goto done;
5084 }
5085
5086 while (level < BTRFS_MAX_LEVEL) {
5087 if (!path->nodes[level]) {
5088 ret = 1;
5089 goto done;
5090 }
5091
5092 slot = path->slots[level] + 1;
5093 c = path->nodes[level];
5094 if (slot >= btrfs_header_nritems(c)) {
5095 level++;
5096 if (level == BTRFS_MAX_LEVEL) {
5097 ret = 1;
5098 goto done;
5099 }
5100 continue;
5101 }
5102
5103 if (next) {
5104 btrfs_tree_unlock_rw(next, next_rw_lock);
5105 free_extent_buffer(next);
5106 }
5107
5108 next = c;
5109 next_rw_lock = path->locks[level];
5110 ret = read_block_for_search(NULL, root, path, &next, level,
5111 slot, &key, 0);
5112 if (ret == -EAGAIN)
5113 goto again;
5114
5115 if (ret < 0) {
5116 btrfs_release_path(path);
5117 goto done;
5118 }
5119
5120 if (!path->skip_locking) {
5121 ret = btrfs_try_tree_read_lock(next);
5122 if (!ret && time_seq) {
5123 /*
5124 * If we don't get the lock, we may be racing
5125 * with push_leaf_left, holding that lock while
5126 * itself waiting for the leaf we've currently
5127 * locked. To solve this situation, we give up
5128 * on our lock and cycle.
5129 */
5130 btrfs_release_path(path);
5131 cond_resched();
5132 goto again;
5133 }
5134 if (!ret) {
5135 btrfs_set_path_blocking(path);
5136 btrfs_tree_read_lock(next);
5137 btrfs_clear_path_blocking(path, next,
5138 BTRFS_READ_LOCK);
5139 }
5140 next_rw_lock = BTRFS_READ_LOCK;
5141 }
5142 break;
5143 }
5144 path->slots[level] = slot;
5145 while (1) {
5146 level--;
5147 c = path->nodes[level];
5148 if (path->locks[level])
5149 btrfs_tree_unlock_rw(c, path->locks[level]);
5150
5151 free_extent_buffer(c);
5152 path->nodes[level] = next;
5153 path->slots[level] = 0;
5154 if (!path->skip_locking)
5155 path->locks[level] = next_rw_lock;
5156 if (!level)
5157 break;
5158
5159 ret = read_block_for_search(NULL, root, path, &next, level,
5160 0, &key, 0);
5161 if (ret == -EAGAIN)
5162 goto again;
5163
5164 if (ret < 0) {
5165 btrfs_release_path(path);
5166 goto done;
5167 }
5168
5169 if (!path->skip_locking) {
5170 ret = btrfs_try_tree_read_lock(next);
5171 if (!ret) {
5172 btrfs_set_path_blocking(path);
5173 btrfs_tree_read_lock(next);
5174 btrfs_clear_path_blocking(path, next,
5175 BTRFS_READ_LOCK);
5176 }
5177 next_rw_lock = BTRFS_READ_LOCK;
5178 }
5179 }
5180 ret = 0;
5181done:
5182 unlock_up(path, 0, 1, 0, NULL);
5183 path->leave_spinning = old_spinning;
5184 if (!old_spinning)
5185 btrfs_set_path_blocking(path);
5186
5187 return ret;
5188}
5189
5190/*
5191 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5192 * searching until it gets past min_objectid or finds an item of 'type'
5193 *
5194 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5195 */
5196int btrfs_previous_item(struct btrfs_root *root,
5197 struct btrfs_path *path, u64 min_objectid,
5198 int type)
5199{
5200 struct btrfs_key found_key;
5201 struct extent_buffer *leaf;
5202 u32 nritems;
5203 int ret;
5204
5205 while (1) {
5206 if (path->slots[0] == 0) {
5207 btrfs_set_path_blocking(path);
5208 ret = btrfs_prev_leaf(root, path);
5209 if (ret != 0)
5210 return ret;
5211 } else {
5212 path->slots[0]--;
5213 }
5214 leaf = path->nodes[0];
5215 nritems = btrfs_header_nritems(leaf);
5216 if (nritems == 0)
5217 return 1;
5218 if (path->slots[0] == nritems)
5219 path->slots[0]--;
5220
5221 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5222 if (found_key.objectid < min_objectid)
5223 break;
5224 if (found_key.type == type)
5225 return 0;
5226 if (found_key.objectid == min_objectid &&
5227 found_key.type < type)
5228 break;
5229 }
5230 return 1;
5231}