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