Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007,2008 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/rbtree.h>
9#include <linux/mm.h>
10#include <linux/error-injection.h>
11#include "messages.h"
12#include "ctree.h"
13#include "disk-io.h"
14#include "transaction.h"
15#include "print-tree.h"
16#include "locking.h"
17#include "volumes.h"
18#include "qgroup.h"
19#include "tree-mod-log.h"
20#include "tree-checker.h"
21#include "fs.h"
22#include "accessors.h"
23#include "extent-tree.h"
24#include "relocation.h"
25#include "file-item.h"
26
27static struct kmem_cache *btrfs_path_cachep;
28
29static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
30 *root, struct btrfs_path *path, int level);
31static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
32 const struct btrfs_key *ins_key, struct btrfs_path *path,
33 int data_size, int extend);
34static int push_node_left(struct btrfs_trans_handle *trans,
35 struct extent_buffer *dst,
36 struct extent_buffer *src, int empty);
37static int balance_node_right(struct btrfs_trans_handle *trans,
38 struct extent_buffer *dst_buf,
39 struct extent_buffer *src_buf);
40
41static const struct btrfs_csums {
42 u16 size;
43 const char name[10];
44 const char driver[12];
45} btrfs_csums[] = {
46 [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
47 [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
48 [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
49 [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
50 .driver = "blake2b-256" },
51};
52
53/*
54 * The leaf data grows from end-to-front in the node. this returns the address
55 * of the start of the last item, which is the stop of the leaf data stack.
56 */
57static unsigned int leaf_data_end(const struct extent_buffer *leaf)
58{
59 u32 nr = btrfs_header_nritems(leaf);
60
61 if (nr == 0)
62 return BTRFS_LEAF_DATA_SIZE(leaf->fs_info);
63 return btrfs_item_offset(leaf, nr - 1);
64}
65
66/*
67 * Move data in a @leaf (using memmove, safe for overlapping ranges).
68 *
69 * @leaf: leaf that we're doing a memmove on
70 * @dst_offset: item data offset we're moving to
71 * @src_offset: item data offset were' moving from
72 * @len: length of the data we're moving
73 *
74 * Wrapper around memmove_extent_buffer() that takes into account the header on
75 * the leaf. The btrfs_item offset's start directly after the header, so we
76 * have to adjust any offsets to account for the header in the leaf. This
77 * handles that math to simplify the callers.
78 */
79static inline void memmove_leaf_data(const struct extent_buffer *leaf,
80 unsigned long dst_offset,
81 unsigned long src_offset,
82 unsigned long len)
83{
84 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, 0) + dst_offset,
85 btrfs_item_nr_offset(leaf, 0) + src_offset, len);
86}
87
88/*
89 * Copy item data from @src into @dst at the given @offset.
90 *
91 * @dst: destination leaf that we're copying into
92 * @src: source leaf that we're copying from
93 * @dst_offset: item data offset we're copying to
94 * @src_offset: item data offset were' copying from
95 * @len: length of the data we're copying
96 *
97 * Wrapper around copy_extent_buffer() that takes into account the header on
98 * the leaf. The btrfs_item offset's start directly after the header, so we
99 * have to adjust any offsets to account for the header in the leaf. This
100 * handles that math to simplify the callers.
101 */
102static inline void copy_leaf_data(const struct extent_buffer *dst,
103 const struct extent_buffer *src,
104 unsigned long dst_offset,
105 unsigned long src_offset, unsigned long len)
106{
107 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, 0) + dst_offset,
108 btrfs_item_nr_offset(src, 0) + src_offset, len);
109}
110
111/*
112 * Move items in a @leaf (using memmove).
113 *
114 * @dst: destination leaf for the items
115 * @dst_item: the item nr we're copying into
116 * @src_item: the item nr we're copying from
117 * @nr_items: the number of items to copy
118 *
119 * Wrapper around memmove_extent_buffer() that does the math to get the
120 * appropriate offsets into the leaf from the item numbers.
121 */
122static inline void memmove_leaf_items(const struct extent_buffer *leaf,
123 int dst_item, int src_item, int nr_items)
124{
125 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, dst_item),
126 btrfs_item_nr_offset(leaf, src_item),
127 nr_items * sizeof(struct btrfs_item));
128}
129
130/*
131 * Copy items from @src into @dst at the given @offset.
132 *
133 * @dst: destination leaf for the items
134 * @src: source leaf for the items
135 * @dst_item: the item nr we're copying into
136 * @src_item: the item nr we're copying from
137 * @nr_items: the number of items to copy
138 *
139 * Wrapper around copy_extent_buffer() that does the math to get the
140 * appropriate offsets into the leaf from the item numbers.
141 */
142static inline void copy_leaf_items(const struct extent_buffer *dst,
143 const struct extent_buffer *src,
144 int dst_item, int src_item, int nr_items)
145{
146 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, dst_item),
147 btrfs_item_nr_offset(src, src_item),
148 nr_items * sizeof(struct btrfs_item));
149}
150
151/* This exists for btrfs-progs usages. */
152u16 btrfs_csum_type_size(u16 type)
153{
154 return btrfs_csums[type].size;
155}
156
157int btrfs_super_csum_size(const struct btrfs_super_block *s)
158{
159 u16 t = btrfs_super_csum_type(s);
160 /*
161 * csum type is validated at mount time
162 */
163 return btrfs_csum_type_size(t);
164}
165
166const char *btrfs_super_csum_name(u16 csum_type)
167{
168 /* csum type is validated at mount time */
169 return btrfs_csums[csum_type].name;
170}
171
172/*
173 * Return driver name if defined, otherwise the name that's also a valid driver
174 * name
175 */
176const char *btrfs_super_csum_driver(u16 csum_type)
177{
178 /* csum type is validated at mount time */
179 return btrfs_csums[csum_type].driver[0] ?
180 btrfs_csums[csum_type].driver :
181 btrfs_csums[csum_type].name;
182}
183
184size_t __attribute_const__ btrfs_get_num_csums(void)
185{
186 return ARRAY_SIZE(btrfs_csums);
187}
188
189struct btrfs_path *btrfs_alloc_path(void)
190{
191 might_sleep();
192
193 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
194}
195
196/* this also releases the path */
197void btrfs_free_path(struct btrfs_path *p)
198{
199 if (!p)
200 return;
201 btrfs_release_path(p);
202 kmem_cache_free(btrfs_path_cachep, p);
203}
204
205/*
206 * path release drops references on the extent buffers in the path
207 * and it drops any locks held by this path
208 *
209 * It is safe to call this on paths that no locks or extent buffers held.
210 */
211noinline void btrfs_release_path(struct btrfs_path *p)
212{
213 int i;
214
215 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
216 p->slots[i] = 0;
217 if (!p->nodes[i])
218 continue;
219 if (p->locks[i]) {
220 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
221 p->locks[i] = 0;
222 }
223 free_extent_buffer(p->nodes[i]);
224 p->nodes[i] = NULL;
225 }
226}
227
228/*
229 * We want the transaction abort to print stack trace only for errors where the
230 * cause could be a bug, eg. due to ENOSPC, and not for common errors that are
231 * caused by external factors.
232 */
233bool __cold abort_should_print_stack(int error)
234{
235 switch (error) {
236 case -EIO:
237 case -EROFS:
238 case -ENOMEM:
239 return false;
240 }
241 return true;
242}
243
244/*
245 * safely gets a reference on the root node of a tree. A lock
246 * is not taken, so a concurrent writer may put a different node
247 * at the root of the tree. See btrfs_lock_root_node for the
248 * looping required.
249 *
250 * The extent buffer returned by this has a reference taken, so
251 * it won't disappear. It may stop being the root of the tree
252 * at any time because there are no locks held.
253 */
254struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
255{
256 struct extent_buffer *eb;
257
258 while (1) {
259 rcu_read_lock();
260 eb = rcu_dereference(root->node);
261
262 /*
263 * RCU really hurts here, we could free up the root node because
264 * it was COWed but we may not get the new root node yet so do
265 * the inc_not_zero dance and if it doesn't work then
266 * synchronize_rcu and try again.
267 */
268 if (atomic_inc_not_zero(&eb->refs)) {
269 rcu_read_unlock();
270 break;
271 }
272 rcu_read_unlock();
273 synchronize_rcu();
274 }
275 return eb;
276}
277
278/*
279 * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
280 * just get put onto a simple dirty list. Transaction walks this list to make
281 * sure they get properly updated on disk.
282 */
283static void add_root_to_dirty_list(struct btrfs_root *root)
284{
285 struct btrfs_fs_info *fs_info = root->fs_info;
286
287 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
288 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
289 return;
290
291 spin_lock(&fs_info->trans_lock);
292 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
293 /* Want the extent tree to be the last on the list */
294 if (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}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007,2008 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/rbtree.h>
9#include <linux/mm.h>
10#include "ctree.h"
11#include "disk-io.h"
12#include "transaction.h"
13#include "print-tree.h"
14#include "locking.h"
15#include "volumes.h"
16#include "qgroup.h"
17
18static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
19 *root, struct btrfs_path *path, int level);
20static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
21 const struct btrfs_key *ins_key, struct btrfs_path *path,
22 int data_size, int extend);
23static int push_node_left(struct btrfs_trans_handle *trans,
24 struct extent_buffer *dst,
25 struct extent_buffer *src, int empty);
26static int balance_node_right(struct btrfs_trans_handle *trans,
27 struct extent_buffer *dst_buf,
28 struct extent_buffer *src_buf);
29static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
30 int level, int slot);
31
32static const struct btrfs_csums {
33 u16 size;
34 const char name[10];
35 const char driver[12];
36} btrfs_csums[] = {
37 [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
38 [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
39 [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
40 [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
41 .driver = "blake2b-256" },
42};
43
44int btrfs_super_csum_size(const struct btrfs_super_block *s)
45{
46 u16 t = btrfs_super_csum_type(s);
47 /*
48 * csum type is validated at mount time
49 */
50 return btrfs_csums[t].size;
51}
52
53const char *btrfs_super_csum_name(u16 csum_type)
54{
55 /* csum type is validated at mount time */
56 return btrfs_csums[csum_type].name;
57}
58
59/*
60 * Return driver name if defined, otherwise the name that's also a valid driver
61 * name
62 */
63const char *btrfs_super_csum_driver(u16 csum_type)
64{
65 /* csum type is validated at mount time */
66 return btrfs_csums[csum_type].driver[0] ?
67 btrfs_csums[csum_type].driver :
68 btrfs_csums[csum_type].name;
69}
70
71size_t __attribute_const__ btrfs_get_num_csums(void)
72{
73 return ARRAY_SIZE(btrfs_csums);
74}
75
76struct btrfs_path *btrfs_alloc_path(void)
77{
78 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
79}
80
81/* this also releases the path */
82void btrfs_free_path(struct btrfs_path *p)
83{
84 if (!p)
85 return;
86 btrfs_release_path(p);
87 kmem_cache_free(btrfs_path_cachep, p);
88}
89
90/*
91 * path release drops references on the extent buffers in the path
92 * and it drops any locks held by this path
93 *
94 * It is safe to call this on paths that no locks or extent buffers held.
95 */
96noinline void btrfs_release_path(struct btrfs_path *p)
97{
98 int i;
99
100 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
101 p->slots[i] = 0;
102 if (!p->nodes[i])
103 continue;
104 if (p->locks[i]) {
105 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
106 p->locks[i] = 0;
107 }
108 free_extent_buffer(p->nodes[i]);
109 p->nodes[i] = NULL;
110 }
111}
112
113/*
114 * safely gets a reference on the root node of a tree. A lock
115 * is not taken, so a concurrent writer may put a different node
116 * at the root of the tree. See btrfs_lock_root_node for the
117 * looping required.
118 *
119 * The extent buffer returned by this has a reference taken, so
120 * it won't disappear. It may stop being the root of the tree
121 * at any time because there are no locks held.
122 */
123struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
124{
125 struct extent_buffer *eb;
126
127 while (1) {
128 rcu_read_lock();
129 eb = rcu_dereference(root->node);
130
131 /*
132 * RCU really hurts here, we could free up the root node because
133 * it was COWed but we may not get the new root node yet so do
134 * the inc_not_zero dance and if it doesn't work then
135 * synchronize_rcu and try again.
136 */
137 if (atomic_inc_not_zero(&eb->refs)) {
138 rcu_read_unlock();
139 break;
140 }
141 rcu_read_unlock();
142 synchronize_rcu();
143 }
144 return eb;
145}
146
147/*
148 * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
149 * just get put onto a simple dirty list. Transaction walks this list to make
150 * sure they get properly updated on disk.
151 */
152static void add_root_to_dirty_list(struct btrfs_root *root)
153{
154 struct btrfs_fs_info *fs_info = root->fs_info;
155
156 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
157 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
158 return;
159
160 spin_lock(&fs_info->trans_lock);
161 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
162 /* Want the extent tree to be the last on the list */
163 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
164 list_move_tail(&root->dirty_list,
165 &fs_info->dirty_cowonly_roots);
166 else
167 list_move(&root->dirty_list,
168 &fs_info->dirty_cowonly_roots);
169 }
170 spin_unlock(&fs_info->trans_lock);
171}
172
173/*
174 * used by snapshot creation to make a copy of a root for a tree with
175 * a given objectid. The buffer with the new root node is returned in
176 * cow_ret, and this func returns zero on success or a negative error code.
177 */
178int btrfs_copy_root(struct btrfs_trans_handle *trans,
179 struct btrfs_root *root,
180 struct extent_buffer *buf,
181 struct extent_buffer **cow_ret, u64 new_root_objectid)
182{
183 struct btrfs_fs_info *fs_info = root->fs_info;
184 struct extent_buffer *cow;
185 int ret = 0;
186 int level;
187 struct btrfs_disk_key disk_key;
188
189 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
190 trans->transid != fs_info->running_transaction->transid);
191 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
192 trans->transid != root->last_trans);
193
194 level = btrfs_header_level(buf);
195 if (level == 0)
196 btrfs_item_key(buf, &disk_key, 0);
197 else
198 btrfs_node_key(buf, &disk_key, 0);
199
200 cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
201 &disk_key, level, buf->start, 0);
202 if (IS_ERR(cow))
203 return PTR_ERR(cow);
204
205 copy_extent_buffer_full(cow, buf);
206 btrfs_set_header_bytenr(cow, cow->start);
207 btrfs_set_header_generation(cow, trans->transid);
208 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
209 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
210 BTRFS_HEADER_FLAG_RELOC);
211 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
212 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
213 else
214 btrfs_set_header_owner(cow, new_root_objectid);
215
216 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
217
218 WARN_ON(btrfs_header_generation(buf) > trans->transid);
219 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
220 ret = btrfs_inc_ref(trans, root, cow, 1);
221 else
222 ret = btrfs_inc_ref(trans, root, cow, 0);
223
224 if (ret)
225 return ret;
226
227 btrfs_mark_buffer_dirty(cow);
228 *cow_ret = cow;
229 return 0;
230}
231
232enum mod_log_op {
233 MOD_LOG_KEY_REPLACE,
234 MOD_LOG_KEY_ADD,
235 MOD_LOG_KEY_REMOVE,
236 MOD_LOG_KEY_REMOVE_WHILE_FREEING,
237 MOD_LOG_KEY_REMOVE_WHILE_MOVING,
238 MOD_LOG_MOVE_KEYS,
239 MOD_LOG_ROOT_REPLACE,
240};
241
242struct tree_mod_root {
243 u64 logical;
244 u8 level;
245};
246
247struct tree_mod_elem {
248 struct rb_node node;
249 u64 logical;
250 u64 seq;
251 enum mod_log_op op;
252
253 /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
254 int slot;
255
256 /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
257 u64 generation;
258
259 /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
260 struct btrfs_disk_key key;
261 u64 blockptr;
262
263 /* this is used for op == MOD_LOG_MOVE_KEYS */
264 struct {
265 int dst_slot;
266 int nr_items;
267 } move;
268
269 /* this is used for op == MOD_LOG_ROOT_REPLACE */
270 struct tree_mod_root old_root;
271};
272
273/*
274 * Pull a new tree mod seq number for our operation.
275 */
276static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
277{
278 return atomic64_inc_return(&fs_info->tree_mod_seq);
279}
280
281/*
282 * This adds a new blocker to the tree mod log's blocker list if the @elem
283 * passed does not already have a sequence number set. So when a caller expects
284 * to record tree modifications, it should ensure to set elem->seq to zero
285 * before calling btrfs_get_tree_mod_seq.
286 * Returns a fresh, unused tree log modification sequence number, even if no new
287 * blocker was added.
288 */
289u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
290 struct seq_list *elem)
291{
292 write_lock(&fs_info->tree_mod_log_lock);
293 if (!elem->seq) {
294 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
295 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
296 }
297 write_unlock(&fs_info->tree_mod_log_lock);
298
299 return elem->seq;
300}
301
302void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
303 struct seq_list *elem)
304{
305 struct rb_root *tm_root;
306 struct rb_node *node;
307 struct rb_node *next;
308 struct tree_mod_elem *tm;
309 u64 min_seq = (u64)-1;
310 u64 seq_putting = elem->seq;
311
312 if (!seq_putting)
313 return;
314
315 write_lock(&fs_info->tree_mod_log_lock);
316 list_del(&elem->list);
317 elem->seq = 0;
318
319 if (!list_empty(&fs_info->tree_mod_seq_list)) {
320 struct seq_list *first;
321
322 first = list_first_entry(&fs_info->tree_mod_seq_list,
323 struct seq_list, list);
324 if (seq_putting > first->seq) {
325 /*
326 * Blocker with lower sequence number exists, we
327 * cannot remove anything from the log.
328 */
329 write_unlock(&fs_info->tree_mod_log_lock);
330 return;
331 }
332 min_seq = first->seq;
333 }
334
335 /*
336 * anything that's lower than the lowest existing (read: blocked)
337 * sequence number can be removed from the tree.
338 */
339 tm_root = &fs_info->tree_mod_log;
340 for (node = rb_first(tm_root); node; node = next) {
341 next = rb_next(node);
342 tm = rb_entry(node, struct tree_mod_elem, node);
343 if (tm->seq >= min_seq)
344 continue;
345 rb_erase(node, tm_root);
346 kfree(tm);
347 }
348 write_unlock(&fs_info->tree_mod_log_lock);
349}
350
351/*
352 * key order of the log:
353 * node/leaf start address -> sequence
354 *
355 * The 'start address' is the logical address of the *new* root node
356 * for root replace operations, or the logical address of the affected
357 * block for all other operations.
358 */
359static noinline int
360__tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
361{
362 struct rb_root *tm_root;
363 struct rb_node **new;
364 struct rb_node *parent = NULL;
365 struct tree_mod_elem *cur;
366
367 lockdep_assert_held_write(&fs_info->tree_mod_log_lock);
368
369 tm->seq = btrfs_inc_tree_mod_seq(fs_info);
370
371 tm_root = &fs_info->tree_mod_log;
372 new = &tm_root->rb_node;
373 while (*new) {
374 cur = rb_entry(*new, struct tree_mod_elem, node);
375 parent = *new;
376 if (cur->logical < tm->logical)
377 new = &((*new)->rb_left);
378 else if (cur->logical > tm->logical)
379 new = &((*new)->rb_right);
380 else if (cur->seq < tm->seq)
381 new = &((*new)->rb_left);
382 else if (cur->seq > tm->seq)
383 new = &((*new)->rb_right);
384 else
385 return -EEXIST;
386 }
387
388 rb_link_node(&tm->node, parent, new);
389 rb_insert_color(&tm->node, tm_root);
390 return 0;
391}
392
393/*
394 * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
395 * returns zero with the tree_mod_log_lock acquired. The caller must hold
396 * this until all tree mod log insertions are recorded in the rb tree and then
397 * write unlock fs_info::tree_mod_log_lock.
398 */
399static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
400 struct extent_buffer *eb) {
401 smp_mb();
402 if (list_empty(&(fs_info)->tree_mod_seq_list))
403 return 1;
404 if (eb && btrfs_header_level(eb) == 0)
405 return 1;
406
407 write_lock(&fs_info->tree_mod_log_lock);
408 if (list_empty(&(fs_info)->tree_mod_seq_list)) {
409 write_unlock(&fs_info->tree_mod_log_lock);
410 return 1;
411 }
412
413 return 0;
414}
415
416/* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
417static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
418 struct extent_buffer *eb)
419{
420 smp_mb();
421 if (list_empty(&(fs_info)->tree_mod_seq_list))
422 return 0;
423 if (eb && btrfs_header_level(eb) == 0)
424 return 0;
425
426 return 1;
427}
428
429static struct tree_mod_elem *
430alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
431 enum mod_log_op op, gfp_t flags)
432{
433 struct tree_mod_elem *tm;
434
435 tm = kzalloc(sizeof(*tm), flags);
436 if (!tm)
437 return NULL;
438
439 tm->logical = eb->start;
440 if (op != MOD_LOG_KEY_ADD) {
441 btrfs_node_key(eb, &tm->key, slot);
442 tm->blockptr = btrfs_node_blockptr(eb, slot);
443 }
444 tm->op = op;
445 tm->slot = slot;
446 tm->generation = btrfs_node_ptr_generation(eb, slot);
447 RB_CLEAR_NODE(&tm->node);
448
449 return tm;
450}
451
452static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
453 enum mod_log_op op, gfp_t flags)
454{
455 struct tree_mod_elem *tm;
456 int ret;
457
458 if (!tree_mod_need_log(eb->fs_info, eb))
459 return 0;
460
461 tm = alloc_tree_mod_elem(eb, slot, op, flags);
462 if (!tm)
463 return -ENOMEM;
464
465 if (tree_mod_dont_log(eb->fs_info, eb)) {
466 kfree(tm);
467 return 0;
468 }
469
470 ret = __tree_mod_log_insert(eb->fs_info, tm);
471 write_unlock(&eb->fs_info->tree_mod_log_lock);
472 if (ret)
473 kfree(tm);
474
475 return ret;
476}
477
478static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
479 int dst_slot, int src_slot, int nr_items)
480{
481 struct tree_mod_elem *tm = NULL;
482 struct tree_mod_elem **tm_list = NULL;
483 int ret = 0;
484 int i;
485 int locked = 0;
486
487 if (!tree_mod_need_log(eb->fs_info, eb))
488 return 0;
489
490 tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
491 if (!tm_list)
492 return -ENOMEM;
493
494 tm = kzalloc(sizeof(*tm), GFP_NOFS);
495 if (!tm) {
496 ret = -ENOMEM;
497 goto free_tms;
498 }
499
500 tm->logical = eb->start;
501 tm->slot = src_slot;
502 tm->move.dst_slot = dst_slot;
503 tm->move.nr_items = nr_items;
504 tm->op = MOD_LOG_MOVE_KEYS;
505
506 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
507 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
508 MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
509 if (!tm_list[i]) {
510 ret = -ENOMEM;
511 goto free_tms;
512 }
513 }
514
515 if (tree_mod_dont_log(eb->fs_info, eb))
516 goto free_tms;
517 locked = 1;
518
519 /*
520 * When we override something during the move, we log these removals.
521 * This can only happen when we move towards the beginning of the
522 * buffer, i.e. dst_slot < src_slot.
523 */
524 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
525 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
526 if (ret)
527 goto free_tms;
528 }
529
530 ret = __tree_mod_log_insert(eb->fs_info, tm);
531 if (ret)
532 goto free_tms;
533 write_unlock(&eb->fs_info->tree_mod_log_lock);
534 kfree(tm_list);
535
536 return 0;
537free_tms:
538 for (i = 0; i < nr_items; i++) {
539 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
540 rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
541 kfree(tm_list[i]);
542 }
543 if (locked)
544 write_unlock(&eb->fs_info->tree_mod_log_lock);
545 kfree(tm_list);
546 kfree(tm);
547
548 return ret;
549}
550
551static inline int
552__tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
553 struct tree_mod_elem **tm_list,
554 int nritems)
555{
556 int i, j;
557 int ret;
558
559 for (i = nritems - 1; i >= 0; i--) {
560 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
561 if (ret) {
562 for (j = nritems - 1; j > i; j--)
563 rb_erase(&tm_list[j]->node,
564 &fs_info->tree_mod_log);
565 return ret;
566 }
567 }
568
569 return 0;
570}
571
572static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
573 struct extent_buffer *new_root, int log_removal)
574{
575 struct btrfs_fs_info *fs_info = old_root->fs_info;
576 struct tree_mod_elem *tm = NULL;
577 struct tree_mod_elem **tm_list = NULL;
578 int nritems = 0;
579 int ret = 0;
580 int i;
581
582 if (!tree_mod_need_log(fs_info, NULL))
583 return 0;
584
585 if (log_removal && btrfs_header_level(old_root) > 0) {
586 nritems = btrfs_header_nritems(old_root);
587 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
588 GFP_NOFS);
589 if (!tm_list) {
590 ret = -ENOMEM;
591 goto free_tms;
592 }
593 for (i = 0; i < nritems; i++) {
594 tm_list[i] = alloc_tree_mod_elem(old_root, i,
595 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
596 if (!tm_list[i]) {
597 ret = -ENOMEM;
598 goto free_tms;
599 }
600 }
601 }
602
603 tm = kzalloc(sizeof(*tm), GFP_NOFS);
604 if (!tm) {
605 ret = -ENOMEM;
606 goto free_tms;
607 }
608
609 tm->logical = new_root->start;
610 tm->old_root.logical = old_root->start;
611 tm->old_root.level = btrfs_header_level(old_root);
612 tm->generation = btrfs_header_generation(old_root);
613 tm->op = MOD_LOG_ROOT_REPLACE;
614
615 if (tree_mod_dont_log(fs_info, NULL))
616 goto free_tms;
617
618 if (tm_list)
619 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
620 if (!ret)
621 ret = __tree_mod_log_insert(fs_info, tm);
622
623 write_unlock(&fs_info->tree_mod_log_lock);
624 if (ret)
625 goto free_tms;
626 kfree(tm_list);
627
628 return ret;
629
630free_tms:
631 if (tm_list) {
632 for (i = 0; i < nritems; i++)
633 kfree(tm_list[i]);
634 kfree(tm_list);
635 }
636 kfree(tm);
637
638 return ret;
639}
640
641static struct tree_mod_elem *
642__tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
643 int smallest)
644{
645 struct rb_root *tm_root;
646 struct rb_node *node;
647 struct tree_mod_elem *cur = NULL;
648 struct tree_mod_elem *found = NULL;
649
650 read_lock(&fs_info->tree_mod_log_lock);
651 tm_root = &fs_info->tree_mod_log;
652 node = tm_root->rb_node;
653 while (node) {
654 cur = rb_entry(node, struct tree_mod_elem, node);
655 if (cur->logical < start) {
656 node = node->rb_left;
657 } else if (cur->logical > start) {
658 node = node->rb_right;
659 } else if (cur->seq < min_seq) {
660 node = node->rb_left;
661 } else if (!smallest) {
662 /* we want the node with the highest seq */
663 if (found)
664 BUG_ON(found->seq > cur->seq);
665 found = cur;
666 node = node->rb_left;
667 } else if (cur->seq > min_seq) {
668 /* we want the node with the smallest seq */
669 if (found)
670 BUG_ON(found->seq < cur->seq);
671 found = cur;
672 node = node->rb_right;
673 } else {
674 found = cur;
675 break;
676 }
677 }
678 read_unlock(&fs_info->tree_mod_log_lock);
679
680 return found;
681}
682
683/*
684 * this returns the element from the log with the smallest time sequence
685 * value that's in the log (the oldest log item). any element with a time
686 * sequence lower than min_seq will be ignored.
687 */
688static struct tree_mod_elem *
689tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
690 u64 min_seq)
691{
692 return __tree_mod_log_search(fs_info, start, min_seq, 1);
693}
694
695/*
696 * this returns the element from the log with the largest time sequence
697 * value that's in the log (the most recent log item). any element with
698 * a time sequence lower than min_seq will be ignored.
699 */
700static struct tree_mod_elem *
701tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
702{
703 return __tree_mod_log_search(fs_info, start, min_seq, 0);
704}
705
706static noinline int tree_mod_log_eb_copy(struct extent_buffer *dst,
707 struct extent_buffer *src, unsigned long dst_offset,
708 unsigned long src_offset, int nr_items)
709{
710 struct btrfs_fs_info *fs_info = dst->fs_info;
711 int ret = 0;
712 struct tree_mod_elem **tm_list = NULL;
713 struct tree_mod_elem **tm_list_add, **tm_list_rem;
714 int i;
715 int locked = 0;
716
717 if (!tree_mod_need_log(fs_info, NULL))
718 return 0;
719
720 if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
721 return 0;
722
723 tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
724 GFP_NOFS);
725 if (!tm_list)
726 return -ENOMEM;
727
728 tm_list_add = tm_list;
729 tm_list_rem = tm_list + nr_items;
730 for (i = 0; i < nr_items; i++) {
731 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
732 MOD_LOG_KEY_REMOVE, GFP_NOFS);
733 if (!tm_list_rem[i]) {
734 ret = -ENOMEM;
735 goto free_tms;
736 }
737
738 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
739 MOD_LOG_KEY_ADD, GFP_NOFS);
740 if (!tm_list_add[i]) {
741 ret = -ENOMEM;
742 goto free_tms;
743 }
744 }
745
746 if (tree_mod_dont_log(fs_info, NULL))
747 goto free_tms;
748 locked = 1;
749
750 for (i = 0; i < nr_items; i++) {
751 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
752 if (ret)
753 goto free_tms;
754 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
755 if (ret)
756 goto free_tms;
757 }
758
759 write_unlock(&fs_info->tree_mod_log_lock);
760 kfree(tm_list);
761
762 return 0;
763
764free_tms:
765 for (i = 0; i < nr_items * 2; i++) {
766 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
767 rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
768 kfree(tm_list[i]);
769 }
770 if (locked)
771 write_unlock(&fs_info->tree_mod_log_lock);
772 kfree(tm_list);
773
774 return ret;
775}
776
777static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
778{
779 struct tree_mod_elem **tm_list = NULL;
780 int nritems = 0;
781 int i;
782 int ret = 0;
783
784 if (btrfs_header_level(eb) == 0)
785 return 0;
786
787 if (!tree_mod_need_log(eb->fs_info, NULL))
788 return 0;
789
790 nritems = btrfs_header_nritems(eb);
791 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
792 if (!tm_list)
793 return -ENOMEM;
794
795 for (i = 0; i < nritems; i++) {
796 tm_list[i] = alloc_tree_mod_elem(eb, i,
797 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
798 if (!tm_list[i]) {
799 ret = -ENOMEM;
800 goto free_tms;
801 }
802 }
803
804 if (tree_mod_dont_log(eb->fs_info, eb))
805 goto free_tms;
806
807 ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
808 write_unlock(&eb->fs_info->tree_mod_log_lock);
809 if (ret)
810 goto free_tms;
811 kfree(tm_list);
812
813 return 0;
814
815free_tms:
816 for (i = 0; i < nritems; i++)
817 kfree(tm_list[i]);
818 kfree(tm_list);
819
820 return ret;
821}
822
823/*
824 * check if the tree block can be shared by multiple trees
825 */
826int btrfs_block_can_be_shared(struct btrfs_root *root,
827 struct extent_buffer *buf)
828{
829 /*
830 * Tree blocks not in shareable trees and tree roots are never shared.
831 * If a block was allocated after the last snapshot and the block was
832 * not allocated by tree relocation, we know the block is not shared.
833 */
834 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
835 buf != root->node && buf != root->commit_root &&
836 (btrfs_header_generation(buf) <=
837 btrfs_root_last_snapshot(&root->root_item) ||
838 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
839 return 1;
840
841 return 0;
842}
843
844static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
845 struct btrfs_root *root,
846 struct extent_buffer *buf,
847 struct extent_buffer *cow,
848 int *last_ref)
849{
850 struct btrfs_fs_info *fs_info = root->fs_info;
851 u64 refs;
852 u64 owner;
853 u64 flags;
854 u64 new_flags = 0;
855 int ret;
856
857 /*
858 * Backrefs update rules:
859 *
860 * Always use full backrefs for extent pointers in tree block
861 * allocated by tree relocation.
862 *
863 * If a shared tree block is no longer referenced by its owner
864 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
865 * use full backrefs for extent pointers in tree block.
866 *
867 * If a tree block is been relocating
868 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
869 * use full backrefs for extent pointers in tree block.
870 * The reason for this is some operations (such as drop tree)
871 * are only allowed for blocks use full backrefs.
872 */
873
874 if (btrfs_block_can_be_shared(root, buf)) {
875 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
876 btrfs_header_level(buf), 1,
877 &refs, &flags);
878 if (ret)
879 return ret;
880 if (refs == 0) {
881 ret = -EROFS;
882 btrfs_handle_fs_error(fs_info, ret, NULL);
883 return ret;
884 }
885 } else {
886 refs = 1;
887 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
888 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
889 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
890 else
891 flags = 0;
892 }
893
894 owner = btrfs_header_owner(buf);
895 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
896 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
897
898 if (refs > 1) {
899 if ((owner == root->root_key.objectid ||
900 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
901 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
902 ret = btrfs_inc_ref(trans, root, buf, 1);
903 if (ret)
904 return ret;
905
906 if (root->root_key.objectid ==
907 BTRFS_TREE_RELOC_OBJECTID) {
908 ret = btrfs_dec_ref(trans, root, buf, 0);
909 if (ret)
910 return ret;
911 ret = btrfs_inc_ref(trans, root, cow, 1);
912 if (ret)
913 return ret;
914 }
915 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
916 } else {
917
918 if (root->root_key.objectid ==
919 BTRFS_TREE_RELOC_OBJECTID)
920 ret = btrfs_inc_ref(trans, root, cow, 1);
921 else
922 ret = btrfs_inc_ref(trans, root, cow, 0);
923 if (ret)
924 return ret;
925 }
926 if (new_flags != 0) {
927 int level = btrfs_header_level(buf);
928
929 ret = btrfs_set_disk_extent_flags(trans, buf,
930 new_flags, level, 0);
931 if (ret)
932 return ret;
933 }
934 } else {
935 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
936 if (root->root_key.objectid ==
937 BTRFS_TREE_RELOC_OBJECTID)
938 ret = btrfs_inc_ref(trans, root, cow, 1);
939 else
940 ret = btrfs_inc_ref(trans, root, cow, 0);
941 if (ret)
942 return ret;
943 ret = btrfs_dec_ref(trans, root, buf, 1);
944 if (ret)
945 return ret;
946 }
947 btrfs_clean_tree_block(buf);
948 *last_ref = 1;
949 }
950 return 0;
951}
952
953static struct extent_buffer *alloc_tree_block_no_bg_flush(
954 struct btrfs_trans_handle *trans,
955 struct btrfs_root *root,
956 u64 parent_start,
957 const struct btrfs_disk_key *disk_key,
958 int level,
959 u64 hint,
960 u64 empty_size)
961{
962 struct btrfs_fs_info *fs_info = root->fs_info;
963 struct extent_buffer *ret;
964
965 /*
966 * If we are COWing a node/leaf from the extent, chunk, device or free
967 * space trees, make sure that we do not finish block group creation of
968 * pending block groups. We do this to avoid a deadlock.
969 * COWing can result in allocation of a new chunk, and flushing pending
970 * block groups (btrfs_create_pending_block_groups()) can be triggered
971 * when finishing allocation of a new chunk. Creation of a pending block
972 * group modifies the extent, chunk, device and free space trees,
973 * therefore we could deadlock with ourselves since we are holding a
974 * lock on an extent buffer that btrfs_create_pending_block_groups() may
975 * try to COW later.
976 * For similar reasons, we also need to delay flushing pending block
977 * groups when splitting a leaf or node, from one of those trees, since
978 * we are holding a write lock on it and its parent or when inserting a
979 * new root node for one of those trees.
980 */
981 if (root == fs_info->extent_root ||
982 root == fs_info->chunk_root ||
983 root == fs_info->dev_root ||
984 root == fs_info->free_space_root)
985 trans->can_flush_pending_bgs = false;
986
987 ret = btrfs_alloc_tree_block(trans, root, parent_start,
988 root->root_key.objectid, disk_key, level,
989 hint, empty_size);
990 trans->can_flush_pending_bgs = true;
991
992 return ret;
993}
994
995/*
996 * does the dirty work in cow of a single block. The parent block (if
997 * supplied) is updated to point to the new cow copy. The new buffer is marked
998 * dirty and returned locked. If you modify the block it needs to be marked
999 * dirty again.
1000 *
1001 * search_start -- an allocation hint for the new block
1002 *
1003 * empty_size -- a hint that you plan on doing more cow. This is the size in
1004 * bytes the allocator should try to find free next to the block it returns.
1005 * This is just a hint and may be ignored by the allocator.
1006 */
1007static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1008 struct btrfs_root *root,
1009 struct extent_buffer *buf,
1010 struct extent_buffer *parent, int parent_slot,
1011 struct extent_buffer **cow_ret,
1012 u64 search_start, u64 empty_size)
1013{
1014 struct btrfs_fs_info *fs_info = root->fs_info;
1015 struct btrfs_disk_key disk_key;
1016 struct extent_buffer *cow;
1017 int level, ret;
1018 int last_ref = 0;
1019 int unlock_orig = 0;
1020 u64 parent_start = 0;
1021
1022 if (*cow_ret == buf)
1023 unlock_orig = 1;
1024
1025 btrfs_assert_tree_locked(buf);
1026
1027 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
1028 trans->transid != fs_info->running_transaction->transid);
1029 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
1030 trans->transid != root->last_trans);
1031
1032 level = btrfs_header_level(buf);
1033
1034 if (level == 0)
1035 btrfs_item_key(buf, &disk_key, 0);
1036 else
1037 btrfs_node_key(buf, &disk_key, 0);
1038
1039 if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1040 parent_start = parent->start;
1041
1042 cow = alloc_tree_block_no_bg_flush(trans, root, parent_start, &disk_key,
1043 level, search_start, empty_size);
1044 if (IS_ERR(cow))
1045 return PTR_ERR(cow);
1046
1047 /* cow is set to blocking by btrfs_init_new_buffer */
1048
1049 copy_extent_buffer_full(cow, buf);
1050 btrfs_set_header_bytenr(cow, cow->start);
1051 btrfs_set_header_generation(cow, trans->transid);
1052 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1053 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1054 BTRFS_HEADER_FLAG_RELOC);
1055 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1056 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1057 else
1058 btrfs_set_header_owner(cow, root->root_key.objectid);
1059
1060 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
1061
1062 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1063 if (ret) {
1064 btrfs_abort_transaction(trans, ret);
1065 return ret;
1066 }
1067
1068 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
1069 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1070 if (ret) {
1071 btrfs_abort_transaction(trans, ret);
1072 return ret;
1073 }
1074 }
1075
1076 if (buf == root->node) {
1077 WARN_ON(parent && parent != buf);
1078 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1079 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1080 parent_start = buf->start;
1081
1082 atomic_inc(&cow->refs);
1083 ret = tree_mod_log_insert_root(root->node, cow, 1);
1084 BUG_ON(ret < 0);
1085 rcu_assign_pointer(root->node, cow);
1086
1087 btrfs_free_tree_block(trans, root, buf, parent_start,
1088 last_ref);
1089 free_extent_buffer(buf);
1090 add_root_to_dirty_list(root);
1091 } else {
1092 WARN_ON(trans->transid != btrfs_header_generation(parent));
1093 tree_mod_log_insert_key(parent, parent_slot,
1094 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1095 btrfs_set_node_blockptr(parent, parent_slot,
1096 cow->start);
1097 btrfs_set_node_ptr_generation(parent, parent_slot,
1098 trans->transid);
1099 btrfs_mark_buffer_dirty(parent);
1100 if (last_ref) {
1101 ret = tree_mod_log_free_eb(buf);
1102 if (ret) {
1103 btrfs_abort_transaction(trans, ret);
1104 return ret;
1105 }
1106 }
1107 btrfs_free_tree_block(trans, root, buf, parent_start,
1108 last_ref);
1109 }
1110 if (unlock_orig)
1111 btrfs_tree_unlock(buf);
1112 free_extent_buffer_stale(buf);
1113 btrfs_mark_buffer_dirty(cow);
1114 *cow_ret = cow;
1115 return 0;
1116}
1117
1118/*
1119 * returns the logical address of the oldest predecessor of the given root.
1120 * entries older than time_seq are ignored.
1121 */
1122static struct tree_mod_elem *__tree_mod_log_oldest_root(
1123 struct extent_buffer *eb_root, u64 time_seq)
1124{
1125 struct tree_mod_elem *tm;
1126 struct tree_mod_elem *found = NULL;
1127 u64 root_logical = eb_root->start;
1128 int looped = 0;
1129
1130 if (!time_seq)
1131 return NULL;
1132
1133 /*
1134 * the very last operation that's logged for a root is the
1135 * replacement operation (if it is replaced at all). this has
1136 * the logical address of the *new* root, making it the very
1137 * first operation that's logged for this root.
1138 */
1139 while (1) {
1140 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1141 time_seq);
1142 if (!looped && !tm)
1143 return NULL;
1144 /*
1145 * if there are no tree operation for the oldest root, we simply
1146 * return it. this should only happen if that (old) root is at
1147 * level 0.
1148 */
1149 if (!tm)
1150 break;
1151
1152 /*
1153 * if there's an operation that's not a root replacement, we
1154 * found the oldest version of our root. normally, we'll find a
1155 * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1156 */
1157 if (tm->op != MOD_LOG_ROOT_REPLACE)
1158 break;
1159
1160 found = tm;
1161 root_logical = tm->old_root.logical;
1162 looped = 1;
1163 }
1164
1165 /* if there's no old root to return, return what we found instead */
1166 if (!found)
1167 found = tm;
1168
1169 return found;
1170}
1171
1172/*
1173 * tm is a pointer to the first operation to rewind within eb. then, all
1174 * previous operations will be rewound (until we reach something older than
1175 * time_seq).
1176 */
1177static void
1178__tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1179 u64 time_seq, struct tree_mod_elem *first_tm)
1180{
1181 u32 n;
1182 struct rb_node *next;
1183 struct tree_mod_elem *tm = first_tm;
1184 unsigned long o_dst;
1185 unsigned long o_src;
1186 unsigned long p_size = sizeof(struct btrfs_key_ptr);
1187
1188 n = btrfs_header_nritems(eb);
1189 read_lock(&fs_info->tree_mod_log_lock);
1190 while (tm && tm->seq >= time_seq) {
1191 /*
1192 * all the operations are recorded with the operator used for
1193 * the modification. as we're going backwards, we do the
1194 * opposite of each operation here.
1195 */
1196 switch (tm->op) {
1197 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1198 BUG_ON(tm->slot < n);
1199 fallthrough;
1200 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1201 case MOD_LOG_KEY_REMOVE:
1202 btrfs_set_node_key(eb, &tm->key, tm->slot);
1203 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1204 btrfs_set_node_ptr_generation(eb, tm->slot,
1205 tm->generation);
1206 n++;
1207 break;
1208 case MOD_LOG_KEY_REPLACE:
1209 BUG_ON(tm->slot >= n);
1210 btrfs_set_node_key(eb, &tm->key, tm->slot);
1211 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1212 btrfs_set_node_ptr_generation(eb, tm->slot,
1213 tm->generation);
1214 break;
1215 case MOD_LOG_KEY_ADD:
1216 /* if a move operation is needed it's in the log */
1217 n--;
1218 break;
1219 case MOD_LOG_MOVE_KEYS:
1220 o_dst = btrfs_node_key_ptr_offset(tm->slot);
1221 o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1222 memmove_extent_buffer(eb, o_dst, o_src,
1223 tm->move.nr_items * p_size);
1224 break;
1225 case MOD_LOG_ROOT_REPLACE:
1226 /*
1227 * this operation is special. for roots, this must be
1228 * handled explicitly before rewinding.
1229 * for non-roots, this operation may exist if the node
1230 * was a root: root A -> child B; then A gets empty and
1231 * B is promoted to the new root. in the mod log, we'll
1232 * have a root-replace operation for B, a tree block
1233 * that is no root. we simply ignore that operation.
1234 */
1235 break;
1236 }
1237 next = rb_next(&tm->node);
1238 if (!next)
1239 break;
1240 tm = rb_entry(next, struct tree_mod_elem, node);
1241 if (tm->logical != first_tm->logical)
1242 break;
1243 }
1244 read_unlock(&fs_info->tree_mod_log_lock);
1245 btrfs_set_header_nritems(eb, n);
1246}
1247
1248/*
1249 * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1250 * is returned. If rewind operations happen, a fresh buffer is returned. The
1251 * returned buffer is always read-locked. If the returned buffer is not the
1252 * input buffer, the lock on the input buffer is released and the input buffer
1253 * is freed (its refcount is decremented).
1254 */
1255static struct extent_buffer *
1256tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1257 struct extent_buffer *eb, u64 time_seq)
1258{
1259 struct extent_buffer *eb_rewin;
1260 struct tree_mod_elem *tm;
1261
1262 if (!time_seq)
1263 return eb;
1264
1265 if (btrfs_header_level(eb) == 0)
1266 return eb;
1267
1268 tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1269 if (!tm)
1270 return eb;
1271
1272 btrfs_set_path_blocking(path);
1273 btrfs_set_lock_blocking_read(eb);
1274
1275 if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1276 BUG_ON(tm->slot != 0);
1277 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1278 if (!eb_rewin) {
1279 btrfs_tree_read_unlock_blocking(eb);
1280 free_extent_buffer(eb);
1281 return NULL;
1282 }
1283 btrfs_set_header_bytenr(eb_rewin, eb->start);
1284 btrfs_set_header_backref_rev(eb_rewin,
1285 btrfs_header_backref_rev(eb));
1286 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1287 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1288 } else {
1289 eb_rewin = btrfs_clone_extent_buffer(eb);
1290 if (!eb_rewin) {
1291 btrfs_tree_read_unlock_blocking(eb);
1292 free_extent_buffer(eb);
1293 return NULL;
1294 }
1295 }
1296
1297 btrfs_tree_read_unlock_blocking(eb);
1298 free_extent_buffer(eb);
1299
1300 btrfs_set_buffer_lockdep_class(btrfs_header_owner(eb_rewin),
1301 eb_rewin, btrfs_header_level(eb_rewin));
1302 btrfs_tree_read_lock(eb_rewin);
1303 __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1304 WARN_ON(btrfs_header_nritems(eb_rewin) >
1305 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1306
1307 return eb_rewin;
1308}
1309
1310/*
1311 * get_old_root() rewinds the state of @root's root node to the given @time_seq
1312 * value. If there are no changes, the current root->root_node is returned. If
1313 * anything changed in between, there's a fresh buffer allocated on which the
1314 * rewind operations are done. In any case, the returned buffer is read locked.
1315 * Returns NULL on error (with no locks held).
1316 */
1317static inline struct extent_buffer *
1318get_old_root(struct btrfs_root *root, u64 time_seq)
1319{
1320 struct btrfs_fs_info *fs_info = root->fs_info;
1321 struct tree_mod_elem *tm;
1322 struct extent_buffer *eb = NULL;
1323 struct extent_buffer *eb_root;
1324 u64 eb_root_owner = 0;
1325 struct extent_buffer *old;
1326 struct tree_mod_root *old_root = NULL;
1327 u64 old_generation = 0;
1328 u64 logical;
1329 int level;
1330
1331 eb_root = btrfs_read_lock_root_node(root);
1332 tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1333 if (!tm)
1334 return eb_root;
1335
1336 if (tm->op == MOD_LOG_ROOT_REPLACE) {
1337 old_root = &tm->old_root;
1338 old_generation = tm->generation;
1339 logical = old_root->logical;
1340 level = old_root->level;
1341 } else {
1342 logical = eb_root->start;
1343 level = btrfs_header_level(eb_root);
1344 }
1345
1346 tm = tree_mod_log_search(fs_info, logical, time_seq);
1347 if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1348 btrfs_tree_read_unlock(eb_root);
1349 free_extent_buffer(eb_root);
1350 old = read_tree_block(fs_info, logical, 0, level, NULL);
1351 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1352 if (!IS_ERR(old))
1353 free_extent_buffer(old);
1354 btrfs_warn(fs_info,
1355 "failed to read tree block %llu from get_old_root",
1356 logical);
1357 } else {
1358 eb = btrfs_clone_extent_buffer(old);
1359 free_extent_buffer(old);
1360 }
1361 } else if (old_root) {
1362 eb_root_owner = btrfs_header_owner(eb_root);
1363 btrfs_tree_read_unlock(eb_root);
1364 free_extent_buffer(eb_root);
1365 eb = alloc_dummy_extent_buffer(fs_info, logical);
1366 } else {
1367 btrfs_set_lock_blocking_read(eb_root);
1368 eb = btrfs_clone_extent_buffer(eb_root);
1369 btrfs_tree_read_unlock_blocking(eb_root);
1370 free_extent_buffer(eb_root);
1371 }
1372
1373 if (!eb)
1374 return NULL;
1375 if (old_root) {
1376 btrfs_set_header_bytenr(eb, eb->start);
1377 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1378 btrfs_set_header_owner(eb, eb_root_owner);
1379 btrfs_set_header_level(eb, old_root->level);
1380 btrfs_set_header_generation(eb, old_generation);
1381 }
1382 btrfs_set_buffer_lockdep_class(btrfs_header_owner(eb), eb,
1383 btrfs_header_level(eb));
1384 btrfs_tree_read_lock(eb);
1385 if (tm)
1386 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1387 else
1388 WARN_ON(btrfs_header_level(eb) != 0);
1389 WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1390
1391 return eb;
1392}
1393
1394int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1395{
1396 struct tree_mod_elem *tm;
1397 int level;
1398 struct extent_buffer *eb_root = btrfs_root_node(root);
1399
1400 tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1401 if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1402 level = tm->old_root.level;
1403 } else {
1404 level = btrfs_header_level(eb_root);
1405 }
1406 free_extent_buffer(eb_root);
1407
1408 return level;
1409}
1410
1411static inline int should_cow_block(struct btrfs_trans_handle *trans,
1412 struct btrfs_root *root,
1413 struct extent_buffer *buf)
1414{
1415 if (btrfs_is_testing(root->fs_info))
1416 return 0;
1417
1418 /* Ensure we can see the FORCE_COW bit */
1419 smp_mb__before_atomic();
1420
1421 /*
1422 * We do not need to cow a block if
1423 * 1) this block is not created or changed in this transaction;
1424 * 2) this block does not belong to TREE_RELOC tree;
1425 * 3) the root is not forced COW.
1426 *
1427 * What is forced COW:
1428 * when we create snapshot during committing the transaction,
1429 * after we've finished copying src root, we must COW the shared
1430 * block to ensure the metadata consistency.
1431 */
1432 if (btrfs_header_generation(buf) == trans->transid &&
1433 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1434 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1435 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1436 !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1437 return 0;
1438 return 1;
1439}
1440
1441/*
1442 * cows a single block, see __btrfs_cow_block for the real work.
1443 * This version of it has extra checks so that a block isn't COWed more than
1444 * once per transaction, as long as it hasn't been written yet
1445 */
1446noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1447 struct btrfs_root *root, struct extent_buffer *buf,
1448 struct extent_buffer *parent, int parent_slot,
1449 struct extent_buffer **cow_ret)
1450{
1451 struct btrfs_fs_info *fs_info = root->fs_info;
1452 u64 search_start;
1453 int ret;
1454
1455 if (test_bit(BTRFS_ROOT_DELETING, &root->state))
1456 btrfs_err(fs_info,
1457 "COW'ing blocks on a fs root that's being dropped");
1458
1459 if (trans->transaction != fs_info->running_transaction)
1460 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1461 trans->transid,
1462 fs_info->running_transaction->transid);
1463
1464 if (trans->transid != fs_info->generation)
1465 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1466 trans->transid, fs_info->generation);
1467
1468 if (!should_cow_block(trans, root, buf)) {
1469 trans->dirty = true;
1470 *cow_ret = buf;
1471 return 0;
1472 }
1473
1474 search_start = buf->start & ~((u64)SZ_1G - 1);
1475
1476 if (parent)
1477 btrfs_set_lock_blocking_write(parent);
1478 btrfs_set_lock_blocking_write(buf);
1479
1480 /*
1481 * Before CoWing this block for later modification, check if it's
1482 * the subtree root and do the delayed subtree trace if needed.
1483 *
1484 * Also We don't care about the error, as it's handled internally.
1485 */
1486 btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
1487 ret = __btrfs_cow_block(trans, root, buf, parent,
1488 parent_slot, cow_ret, search_start, 0);
1489
1490 trace_btrfs_cow_block(root, buf, *cow_ret);
1491
1492 return ret;
1493}
1494
1495/*
1496 * helper function for defrag to decide if two blocks pointed to by a
1497 * node are actually close by
1498 */
1499static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1500{
1501 if (blocknr < other && other - (blocknr + blocksize) < 32768)
1502 return 1;
1503 if (blocknr > other && blocknr - (other + blocksize) < 32768)
1504 return 1;
1505 return 0;
1506}
1507
1508#ifdef __LITTLE_ENDIAN
1509
1510/*
1511 * Compare two keys, on little-endian the disk order is same as CPU order and
1512 * we can avoid the conversion.
1513 */
1514static int comp_keys(const struct btrfs_disk_key *disk_key,
1515 const struct btrfs_key *k2)
1516{
1517 const struct btrfs_key *k1 = (const struct btrfs_key *)disk_key;
1518
1519 return btrfs_comp_cpu_keys(k1, k2);
1520}
1521
1522#else
1523
1524/*
1525 * compare two keys in a memcmp fashion
1526 */
1527static int comp_keys(const struct btrfs_disk_key *disk,
1528 const struct btrfs_key *k2)
1529{
1530 struct btrfs_key k1;
1531
1532 btrfs_disk_key_to_cpu(&k1, disk);
1533
1534 return btrfs_comp_cpu_keys(&k1, k2);
1535}
1536#endif
1537
1538/*
1539 * same as comp_keys only with two btrfs_key's
1540 */
1541int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1542{
1543 if (k1->objectid > k2->objectid)
1544 return 1;
1545 if (k1->objectid < k2->objectid)
1546 return -1;
1547 if (k1->type > k2->type)
1548 return 1;
1549 if (k1->type < k2->type)
1550 return -1;
1551 if (k1->offset > k2->offset)
1552 return 1;
1553 if (k1->offset < k2->offset)
1554 return -1;
1555 return 0;
1556}
1557
1558/*
1559 * this is used by the defrag code to go through all the
1560 * leaves pointed to by a node and reallocate them so that
1561 * disk order is close to key order
1562 */
1563int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1564 struct btrfs_root *root, struct extent_buffer *parent,
1565 int start_slot, u64 *last_ret,
1566 struct btrfs_key *progress)
1567{
1568 struct btrfs_fs_info *fs_info = root->fs_info;
1569 struct extent_buffer *cur;
1570 u64 blocknr;
1571 u64 gen;
1572 u64 search_start = *last_ret;
1573 u64 last_block = 0;
1574 u64 other;
1575 u32 parent_nritems;
1576 int end_slot;
1577 int i;
1578 int err = 0;
1579 int parent_level;
1580 int uptodate;
1581 u32 blocksize;
1582 int progress_passed = 0;
1583 struct btrfs_disk_key disk_key;
1584
1585 parent_level = btrfs_header_level(parent);
1586
1587 WARN_ON(trans->transaction != fs_info->running_transaction);
1588 WARN_ON(trans->transid != fs_info->generation);
1589
1590 parent_nritems = btrfs_header_nritems(parent);
1591 blocksize = fs_info->nodesize;
1592 end_slot = parent_nritems - 1;
1593
1594 if (parent_nritems <= 1)
1595 return 0;
1596
1597 btrfs_set_lock_blocking_write(parent);
1598
1599 for (i = start_slot; i <= end_slot; i++) {
1600 struct btrfs_key first_key;
1601 int close = 1;
1602
1603 btrfs_node_key(parent, &disk_key, i);
1604 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1605 continue;
1606
1607 progress_passed = 1;
1608 blocknr = btrfs_node_blockptr(parent, i);
1609 gen = btrfs_node_ptr_generation(parent, i);
1610 btrfs_node_key_to_cpu(parent, &first_key, i);
1611 if (last_block == 0)
1612 last_block = blocknr;
1613
1614 if (i > 0) {
1615 other = btrfs_node_blockptr(parent, i - 1);
1616 close = close_blocks(blocknr, other, blocksize);
1617 }
1618 if (!close && i < end_slot) {
1619 other = btrfs_node_blockptr(parent, i + 1);
1620 close = close_blocks(blocknr, other, blocksize);
1621 }
1622 if (close) {
1623 last_block = blocknr;
1624 continue;
1625 }
1626
1627 cur = find_extent_buffer(fs_info, blocknr);
1628 if (cur)
1629 uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1630 else
1631 uptodate = 0;
1632 if (!cur || !uptodate) {
1633 if (!cur) {
1634 cur = read_tree_block(fs_info, blocknr, gen,
1635 parent_level - 1,
1636 &first_key);
1637 if (IS_ERR(cur)) {
1638 return PTR_ERR(cur);
1639 } else if (!extent_buffer_uptodate(cur)) {
1640 free_extent_buffer(cur);
1641 return -EIO;
1642 }
1643 } else if (!uptodate) {
1644 err = btrfs_read_buffer(cur, gen,
1645 parent_level - 1,&first_key);
1646 if (err) {
1647 free_extent_buffer(cur);
1648 return err;
1649 }
1650 }
1651 }
1652 if (search_start == 0)
1653 search_start = last_block;
1654
1655 btrfs_tree_lock(cur);
1656 btrfs_set_lock_blocking_write(cur);
1657 err = __btrfs_cow_block(trans, root, cur, parent, i,
1658 &cur, search_start,
1659 min(16 * blocksize,
1660 (end_slot - i) * blocksize));
1661 if (err) {
1662 btrfs_tree_unlock(cur);
1663 free_extent_buffer(cur);
1664 break;
1665 }
1666 search_start = cur->start;
1667 last_block = cur->start;
1668 *last_ret = search_start;
1669 btrfs_tree_unlock(cur);
1670 free_extent_buffer(cur);
1671 }
1672 return err;
1673}
1674
1675/*
1676 * search for key in the extent_buffer. The items start at offset p,
1677 * and they are item_size apart. There are 'max' items in p.
1678 *
1679 * the slot in the array is returned via slot, and it points to
1680 * the place where you would insert key if it is not found in
1681 * the array.
1682 *
1683 * slot may point to max if the key is bigger than all of the keys
1684 */
1685static noinline int generic_bin_search(struct extent_buffer *eb,
1686 unsigned long p, int item_size,
1687 const struct btrfs_key *key,
1688 int max, int *slot)
1689{
1690 int low = 0;
1691 int high = max;
1692 int ret;
1693 const int key_size = sizeof(struct btrfs_disk_key);
1694
1695 if (low > high) {
1696 btrfs_err(eb->fs_info,
1697 "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1698 __func__, low, high, eb->start,
1699 btrfs_header_owner(eb), btrfs_header_level(eb));
1700 return -EINVAL;
1701 }
1702
1703 while (low < high) {
1704 unsigned long oip;
1705 unsigned long offset;
1706 struct btrfs_disk_key *tmp;
1707 struct btrfs_disk_key unaligned;
1708 int mid;
1709
1710 mid = (low + high) / 2;
1711 offset = p + mid * item_size;
1712 oip = offset_in_page(offset);
1713
1714 if (oip + key_size <= PAGE_SIZE) {
1715 const unsigned long idx = offset >> PAGE_SHIFT;
1716 char *kaddr = page_address(eb->pages[idx]);
1717
1718 tmp = (struct btrfs_disk_key *)(kaddr + oip);
1719 } else {
1720 read_extent_buffer(eb, &unaligned, offset, key_size);
1721 tmp = &unaligned;
1722 }
1723
1724 ret = comp_keys(tmp, key);
1725
1726 if (ret < 0)
1727 low = mid + 1;
1728 else if (ret > 0)
1729 high = mid;
1730 else {
1731 *slot = mid;
1732 return 0;
1733 }
1734 }
1735 *slot = low;
1736 return 1;
1737}
1738
1739/*
1740 * simple bin_search frontend that does the right thing for
1741 * leaves vs nodes
1742 */
1743int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1744 int *slot)
1745{
1746 if (btrfs_header_level(eb) == 0)
1747 return generic_bin_search(eb,
1748 offsetof(struct btrfs_leaf, items),
1749 sizeof(struct btrfs_item),
1750 key, btrfs_header_nritems(eb),
1751 slot);
1752 else
1753 return generic_bin_search(eb,
1754 offsetof(struct btrfs_node, ptrs),
1755 sizeof(struct btrfs_key_ptr),
1756 key, btrfs_header_nritems(eb),
1757 slot);
1758}
1759
1760static void root_add_used(struct btrfs_root *root, u32 size)
1761{
1762 spin_lock(&root->accounting_lock);
1763 btrfs_set_root_used(&root->root_item,
1764 btrfs_root_used(&root->root_item) + size);
1765 spin_unlock(&root->accounting_lock);
1766}
1767
1768static void root_sub_used(struct btrfs_root *root, u32 size)
1769{
1770 spin_lock(&root->accounting_lock);
1771 btrfs_set_root_used(&root->root_item,
1772 btrfs_root_used(&root->root_item) - size);
1773 spin_unlock(&root->accounting_lock);
1774}
1775
1776/* given a node and slot number, this reads the blocks it points to. The
1777 * extent buffer is returned with a reference taken (but unlocked).
1778 */
1779struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
1780 int slot)
1781{
1782 int level = btrfs_header_level(parent);
1783 struct extent_buffer *eb;
1784 struct btrfs_key first_key;
1785
1786 if (slot < 0 || slot >= btrfs_header_nritems(parent))
1787 return ERR_PTR(-ENOENT);
1788
1789 BUG_ON(level == 0);
1790
1791 btrfs_node_key_to_cpu(parent, &first_key, slot);
1792 eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
1793 btrfs_node_ptr_generation(parent, slot),
1794 level - 1, &first_key);
1795 if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1796 free_extent_buffer(eb);
1797 eb = ERR_PTR(-EIO);
1798 }
1799
1800 return eb;
1801}
1802
1803/*
1804 * node level balancing, used to make sure nodes are in proper order for
1805 * item deletion. We balance from the top down, so we have to make sure
1806 * that a deletion won't leave an node completely empty later on.
1807 */
1808static noinline int balance_level(struct btrfs_trans_handle *trans,
1809 struct btrfs_root *root,
1810 struct btrfs_path *path, int level)
1811{
1812 struct btrfs_fs_info *fs_info = root->fs_info;
1813 struct extent_buffer *right = NULL;
1814 struct extent_buffer *mid;
1815 struct extent_buffer *left = NULL;
1816 struct extent_buffer *parent = NULL;
1817 int ret = 0;
1818 int wret;
1819 int pslot;
1820 int orig_slot = path->slots[level];
1821 u64 orig_ptr;
1822
1823 ASSERT(level > 0);
1824
1825 mid = path->nodes[level];
1826
1827 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1828 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1829 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1830
1831 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1832
1833 if (level < BTRFS_MAX_LEVEL - 1) {
1834 parent = path->nodes[level + 1];
1835 pslot = path->slots[level + 1];
1836 }
1837
1838 /*
1839 * deal with the case where there is only one pointer in the root
1840 * by promoting the node below to a root
1841 */
1842 if (!parent) {
1843 struct extent_buffer *child;
1844
1845 if (btrfs_header_nritems(mid) != 1)
1846 return 0;
1847
1848 /* promote the child to a root */
1849 child = btrfs_read_node_slot(mid, 0);
1850 if (IS_ERR(child)) {
1851 ret = PTR_ERR(child);
1852 btrfs_handle_fs_error(fs_info, ret, NULL);
1853 goto enospc;
1854 }
1855
1856 btrfs_tree_lock(child);
1857 btrfs_set_lock_blocking_write(child);
1858 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1859 if (ret) {
1860 btrfs_tree_unlock(child);
1861 free_extent_buffer(child);
1862 goto enospc;
1863 }
1864
1865 ret = tree_mod_log_insert_root(root->node, child, 1);
1866 BUG_ON(ret < 0);
1867 rcu_assign_pointer(root->node, child);
1868
1869 add_root_to_dirty_list(root);
1870 btrfs_tree_unlock(child);
1871
1872 path->locks[level] = 0;
1873 path->nodes[level] = NULL;
1874 btrfs_clean_tree_block(mid);
1875 btrfs_tree_unlock(mid);
1876 /* once for the path */
1877 free_extent_buffer(mid);
1878
1879 root_sub_used(root, mid->len);
1880 btrfs_free_tree_block(trans, root, mid, 0, 1);
1881 /* once for the root ptr */
1882 free_extent_buffer_stale(mid);
1883 return 0;
1884 }
1885 if (btrfs_header_nritems(mid) >
1886 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1887 return 0;
1888
1889 left = btrfs_read_node_slot(parent, pslot - 1);
1890 if (IS_ERR(left))
1891 left = NULL;
1892
1893 if (left) {
1894 btrfs_tree_lock(left);
1895 btrfs_set_lock_blocking_write(left);
1896 wret = btrfs_cow_block(trans, root, left,
1897 parent, pslot - 1, &left);
1898 if (wret) {
1899 ret = wret;
1900 goto enospc;
1901 }
1902 }
1903
1904 right = btrfs_read_node_slot(parent, pslot + 1);
1905 if (IS_ERR(right))
1906 right = NULL;
1907
1908 if (right) {
1909 btrfs_tree_lock(right);
1910 btrfs_set_lock_blocking_write(right);
1911 wret = btrfs_cow_block(trans, root, right,
1912 parent, pslot + 1, &right);
1913 if (wret) {
1914 ret = wret;
1915 goto enospc;
1916 }
1917 }
1918
1919 /* first, try to make some room in the middle buffer */
1920 if (left) {
1921 orig_slot += btrfs_header_nritems(left);
1922 wret = push_node_left(trans, left, mid, 1);
1923 if (wret < 0)
1924 ret = wret;
1925 }
1926
1927 /*
1928 * then try to empty the right most buffer into the middle
1929 */
1930 if (right) {
1931 wret = push_node_left(trans, mid, right, 1);
1932 if (wret < 0 && wret != -ENOSPC)
1933 ret = wret;
1934 if (btrfs_header_nritems(right) == 0) {
1935 btrfs_clean_tree_block(right);
1936 btrfs_tree_unlock(right);
1937 del_ptr(root, path, level + 1, pslot + 1);
1938 root_sub_used(root, right->len);
1939 btrfs_free_tree_block(trans, root, right, 0, 1);
1940 free_extent_buffer_stale(right);
1941 right = NULL;
1942 } else {
1943 struct btrfs_disk_key right_key;
1944 btrfs_node_key(right, &right_key, 0);
1945 ret = tree_mod_log_insert_key(parent, pslot + 1,
1946 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1947 BUG_ON(ret < 0);
1948 btrfs_set_node_key(parent, &right_key, pslot + 1);
1949 btrfs_mark_buffer_dirty(parent);
1950 }
1951 }
1952 if (btrfs_header_nritems(mid) == 1) {
1953 /*
1954 * we're not allowed to leave a node with one item in the
1955 * tree during a delete. A deletion from lower in the tree
1956 * could try to delete the only pointer in this node.
1957 * So, pull some keys from the left.
1958 * There has to be a left pointer at this point because
1959 * otherwise we would have pulled some pointers from the
1960 * right
1961 */
1962 if (!left) {
1963 ret = -EROFS;
1964 btrfs_handle_fs_error(fs_info, ret, NULL);
1965 goto enospc;
1966 }
1967 wret = balance_node_right(trans, mid, left);
1968 if (wret < 0) {
1969 ret = wret;
1970 goto enospc;
1971 }
1972 if (wret == 1) {
1973 wret = push_node_left(trans, left, mid, 1);
1974 if (wret < 0)
1975 ret = wret;
1976 }
1977 BUG_ON(wret == 1);
1978 }
1979 if (btrfs_header_nritems(mid) == 0) {
1980 btrfs_clean_tree_block(mid);
1981 btrfs_tree_unlock(mid);
1982 del_ptr(root, path, level + 1, pslot);
1983 root_sub_used(root, mid->len);
1984 btrfs_free_tree_block(trans, root, mid, 0, 1);
1985 free_extent_buffer_stale(mid);
1986 mid = NULL;
1987 } else {
1988 /* update the parent key to reflect our changes */
1989 struct btrfs_disk_key mid_key;
1990 btrfs_node_key(mid, &mid_key, 0);
1991 ret = tree_mod_log_insert_key(parent, pslot,
1992 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1993 BUG_ON(ret < 0);
1994 btrfs_set_node_key(parent, &mid_key, pslot);
1995 btrfs_mark_buffer_dirty(parent);
1996 }
1997
1998 /* update the path */
1999 if (left) {
2000 if (btrfs_header_nritems(left) > orig_slot) {
2001 atomic_inc(&left->refs);
2002 /* left was locked after cow */
2003 path->nodes[level] = left;
2004 path->slots[level + 1] -= 1;
2005 path->slots[level] = orig_slot;
2006 if (mid) {
2007 btrfs_tree_unlock(mid);
2008 free_extent_buffer(mid);
2009 }
2010 } else {
2011 orig_slot -= btrfs_header_nritems(left);
2012 path->slots[level] = orig_slot;
2013 }
2014 }
2015 /* double check we haven't messed things up */
2016 if (orig_ptr !=
2017 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2018 BUG();
2019enospc:
2020 if (right) {
2021 btrfs_tree_unlock(right);
2022 free_extent_buffer(right);
2023 }
2024 if (left) {
2025 if (path->nodes[level] != left)
2026 btrfs_tree_unlock(left);
2027 free_extent_buffer(left);
2028 }
2029 return ret;
2030}
2031
2032/* Node balancing for insertion. Here we only split or push nodes around
2033 * when they are completely full. This is also done top down, so we
2034 * have to be pessimistic.
2035 */
2036static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2037 struct btrfs_root *root,
2038 struct btrfs_path *path, int level)
2039{
2040 struct btrfs_fs_info *fs_info = root->fs_info;
2041 struct extent_buffer *right = NULL;
2042 struct extent_buffer *mid;
2043 struct extent_buffer *left = NULL;
2044 struct extent_buffer *parent = NULL;
2045 int ret = 0;
2046 int wret;
2047 int pslot;
2048 int orig_slot = path->slots[level];
2049
2050 if (level == 0)
2051 return 1;
2052
2053 mid = path->nodes[level];
2054 WARN_ON(btrfs_header_generation(mid) != trans->transid);
2055
2056 if (level < BTRFS_MAX_LEVEL - 1) {
2057 parent = path->nodes[level + 1];
2058 pslot = path->slots[level + 1];
2059 }
2060
2061 if (!parent)
2062 return 1;
2063
2064 left = btrfs_read_node_slot(parent, pslot - 1);
2065 if (IS_ERR(left))
2066 left = NULL;
2067
2068 /* first, try to make some room in the middle buffer */
2069 if (left) {
2070 u32 left_nr;
2071
2072 btrfs_tree_lock(left);
2073 btrfs_set_lock_blocking_write(left);
2074
2075 left_nr = btrfs_header_nritems(left);
2076 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2077 wret = 1;
2078 } else {
2079 ret = btrfs_cow_block(trans, root, left, parent,
2080 pslot - 1, &left);
2081 if (ret)
2082 wret = 1;
2083 else {
2084 wret = push_node_left(trans, left, mid, 0);
2085 }
2086 }
2087 if (wret < 0)
2088 ret = wret;
2089 if (wret == 0) {
2090 struct btrfs_disk_key disk_key;
2091 orig_slot += left_nr;
2092 btrfs_node_key(mid, &disk_key, 0);
2093 ret = tree_mod_log_insert_key(parent, pslot,
2094 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2095 BUG_ON(ret < 0);
2096 btrfs_set_node_key(parent, &disk_key, pslot);
2097 btrfs_mark_buffer_dirty(parent);
2098 if (btrfs_header_nritems(left) > orig_slot) {
2099 path->nodes[level] = left;
2100 path->slots[level + 1] -= 1;
2101 path->slots[level] = orig_slot;
2102 btrfs_tree_unlock(mid);
2103 free_extent_buffer(mid);
2104 } else {
2105 orig_slot -=
2106 btrfs_header_nritems(left);
2107 path->slots[level] = orig_slot;
2108 btrfs_tree_unlock(left);
2109 free_extent_buffer(left);
2110 }
2111 return 0;
2112 }
2113 btrfs_tree_unlock(left);
2114 free_extent_buffer(left);
2115 }
2116 right = btrfs_read_node_slot(parent, pslot + 1);
2117 if (IS_ERR(right))
2118 right = NULL;
2119
2120 /*
2121 * then try to empty the right most buffer into the middle
2122 */
2123 if (right) {
2124 u32 right_nr;
2125
2126 btrfs_tree_lock(right);
2127 btrfs_set_lock_blocking_write(right);
2128
2129 right_nr = btrfs_header_nritems(right);
2130 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2131 wret = 1;
2132 } else {
2133 ret = btrfs_cow_block(trans, root, right,
2134 parent, pslot + 1,
2135 &right);
2136 if (ret)
2137 wret = 1;
2138 else {
2139 wret = balance_node_right(trans, right, mid);
2140 }
2141 }
2142 if (wret < 0)
2143 ret = wret;
2144 if (wret == 0) {
2145 struct btrfs_disk_key disk_key;
2146
2147 btrfs_node_key(right, &disk_key, 0);
2148 ret = tree_mod_log_insert_key(parent, pslot + 1,
2149 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2150 BUG_ON(ret < 0);
2151 btrfs_set_node_key(parent, &disk_key, pslot + 1);
2152 btrfs_mark_buffer_dirty(parent);
2153
2154 if (btrfs_header_nritems(mid) <= orig_slot) {
2155 path->nodes[level] = right;
2156 path->slots[level + 1] += 1;
2157 path->slots[level] = orig_slot -
2158 btrfs_header_nritems(mid);
2159 btrfs_tree_unlock(mid);
2160 free_extent_buffer(mid);
2161 } else {
2162 btrfs_tree_unlock(right);
2163 free_extent_buffer(right);
2164 }
2165 return 0;
2166 }
2167 btrfs_tree_unlock(right);
2168 free_extent_buffer(right);
2169 }
2170 return 1;
2171}
2172
2173/*
2174 * readahead one full node of leaves, finding things that are close
2175 * to the block in 'slot', and triggering ra on them.
2176 */
2177static void reada_for_search(struct btrfs_fs_info *fs_info,
2178 struct btrfs_path *path,
2179 int level, int slot, u64 objectid)
2180{
2181 struct extent_buffer *node;
2182 struct btrfs_disk_key disk_key;
2183 u32 nritems;
2184 u64 search;
2185 u64 target;
2186 u64 nread = 0;
2187 struct extent_buffer *eb;
2188 u32 nr;
2189 u32 blocksize;
2190 u32 nscan = 0;
2191
2192 if (level != 1)
2193 return;
2194
2195 if (!path->nodes[level])
2196 return;
2197
2198 node = path->nodes[level];
2199
2200 search = btrfs_node_blockptr(node, slot);
2201 blocksize = fs_info->nodesize;
2202 eb = find_extent_buffer(fs_info, search);
2203 if (eb) {
2204 free_extent_buffer(eb);
2205 return;
2206 }
2207
2208 target = search;
2209
2210 nritems = btrfs_header_nritems(node);
2211 nr = slot;
2212
2213 while (1) {
2214 if (path->reada == READA_BACK) {
2215 if (nr == 0)
2216 break;
2217 nr--;
2218 } else if (path->reada == READA_FORWARD) {
2219 nr++;
2220 if (nr >= nritems)
2221 break;
2222 }
2223 if (path->reada == READA_BACK && objectid) {
2224 btrfs_node_key(node, &disk_key, nr);
2225 if (btrfs_disk_key_objectid(&disk_key) != objectid)
2226 break;
2227 }
2228 search = btrfs_node_blockptr(node, nr);
2229 if ((search <= target && target - search <= 65536) ||
2230 (search > target && search - target <= 65536)) {
2231 readahead_tree_block(fs_info, search);
2232 nread += blocksize;
2233 }
2234 nscan++;
2235 if ((nread > 65536 || nscan > 32))
2236 break;
2237 }
2238}
2239
2240static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2241 struct btrfs_path *path, int level)
2242{
2243 int slot;
2244 int nritems;
2245 struct extent_buffer *parent;
2246 struct extent_buffer *eb;
2247 u64 gen;
2248 u64 block1 = 0;
2249 u64 block2 = 0;
2250
2251 parent = path->nodes[level + 1];
2252 if (!parent)
2253 return;
2254
2255 nritems = btrfs_header_nritems(parent);
2256 slot = path->slots[level + 1];
2257
2258 if (slot > 0) {
2259 block1 = btrfs_node_blockptr(parent, slot - 1);
2260 gen = btrfs_node_ptr_generation(parent, slot - 1);
2261 eb = find_extent_buffer(fs_info, block1);
2262 /*
2263 * if we get -eagain from btrfs_buffer_uptodate, we
2264 * don't want to return eagain here. That will loop
2265 * forever
2266 */
2267 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2268 block1 = 0;
2269 free_extent_buffer(eb);
2270 }
2271 if (slot + 1 < nritems) {
2272 block2 = btrfs_node_blockptr(parent, slot + 1);
2273 gen = btrfs_node_ptr_generation(parent, slot + 1);
2274 eb = find_extent_buffer(fs_info, block2);
2275 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2276 block2 = 0;
2277 free_extent_buffer(eb);
2278 }
2279
2280 if (block1)
2281 readahead_tree_block(fs_info, block1);
2282 if (block2)
2283 readahead_tree_block(fs_info, block2);
2284}
2285
2286
2287/*
2288 * when we walk down the tree, it is usually safe to unlock the higher layers
2289 * in the tree. The exceptions are when our path goes through slot 0, because
2290 * operations on the tree might require changing key pointers higher up in the
2291 * tree.
2292 *
2293 * callers might also have set path->keep_locks, which tells this code to keep
2294 * the lock if the path points to the last slot in the block. This is part of
2295 * walking through the tree, and selecting the next slot in the higher block.
2296 *
2297 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
2298 * if lowest_unlock is 1, level 0 won't be unlocked
2299 */
2300static noinline void unlock_up(struct btrfs_path *path, int level,
2301 int lowest_unlock, int min_write_lock_level,
2302 int *write_lock_level)
2303{
2304 int i;
2305 int skip_level = level;
2306 int no_skips = 0;
2307 struct extent_buffer *t;
2308
2309 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2310 if (!path->nodes[i])
2311 break;
2312 if (!path->locks[i])
2313 break;
2314 if (!no_skips && path->slots[i] == 0) {
2315 skip_level = i + 1;
2316 continue;
2317 }
2318 if (!no_skips && path->keep_locks) {
2319 u32 nritems;
2320 t = path->nodes[i];
2321 nritems = btrfs_header_nritems(t);
2322 if (nritems < 1 || path->slots[i] >= nritems - 1) {
2323 skip_level = i + 1;
2324 continue;
2325 }
2326 }
2327 if (skip_level < i && i >= lowest_unlock)
2328 no_skips = 1;
2329
2330 t = path->nodes[i];
2331 if (i >= lowest_unlock && i > skip_level) {
2332 btrfs_tree_unlock_rw(t, path->locks[i]);
2333 path->locks[i] = 0;
2334 if (write_lock_level &&
2335 i > min_write_lock_level &&
2336 i <= *write_lock_level) {
2337 *write_lock_level = i - 1;
2338 }
2339 }
2340 }
2341}
2342
2343/*
2344 * helper function for btrfs_search_slot. The goal is to find a block
2345 * in cache without setting the path to blocking. If we find the block
2346 * we return zero and the path is unchanged.
2347 *
2348 * If we can't find the block, we set the path blocking and do some
2349 * reada. -EAGAIN is returned and the search must be repeated.
2350 */
2351static int
2352read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2353 struct extent_buffer **eb_ret, int level, int slot,
2354 const struct btrfs_key *key)
2355{
2356 struct btrfs_fs_info *fs_info = root->fs_info;
2357 u64 blocknr;
2358 u64 gen;
2359 struct extent_buffer *tmp;
2360 struct btrfs_key first_key;
2361 int ret;
2362 int parent_level;
2363
2364 blocknr = btrfs_node_blockptr(*eb_ret, slot);
2365 gen = btrfs_node_ptr_generation(*eb_ret, slot);
2366 parent_level = btrfs_header_level(*eb_ret);
2367 btrfs_node_key_to_cpu(*eb_ret, &first_key, slot);
2368
2369 tmp = find_extent_buffer(fs_info, blocknr);
2370 if (tmp) {
2371 /* first we do an atomic uptodate check */
2372 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2373 /*
2374 * Do extra check for first_key, eb can be stale due to
2375 * being cached, read from scrub, or have multiple
2376 * parents (shared tree blocks).
2377 */
2378 if (btrfs_verify_level_key(tmp,
2379 parent_level - 1, &first_key, gen)) {
2380 free_extent_buffer(tmp);
2381 return -EUCLEAN;
2382 }
2383 *eb_ret = tmp;
2384 return 0;
2385 }
2386
2387 /* the pages were up to date, but we failed
2388 * the generation number check. Do a full
2389 * read for the generation number that is correct.
2390 * We must do this without dropping locks so
2391 * we can trust our generation number
2392 */
2393 btrfs_set_path_blocking(p);
2394
2395 /* now we're allowed to do a blocking uptodate check */
2396 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2397 if (!ret) {
2398 *eb_ret = tmp;
2399 return 0;
2400 }
2401 free_extent_buffer(tmp);
2402 btrfs_release_path(p);
2403 return -EIO;
2404 }
2405
2406 /*
2407 * reduce lock contention at high levels
2408 * of the btree by dropping locks before
2409 * we read. Don't release the lock on the current
2410 * level because we need to walk this node to figure
2411 * out which blocks to read.
2412 */
2413 btrfs_unlock_up_safe(p, level + 1);
2414 btrfs_set_path_blocking(p);
2415
2416 if (p->reada != READA_NONE)
2417 reada_for_search(fs_info, p, level, slot, key->objectid);
2418
2419 ret = -EAGAIN;
2420 tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2421 &first_key);
2422 if (!IS_ERR(tmp)) {
2423 /*
2424 * If the read above didn't mark this buffer up to date,
2425 * it will never end up being up to date. Set ret to EIO now
2426 * and give up so that our caller doesn't loop forever
2427 * on our EAGAINs.
2428 */
2429 if (!extent_buffer_uptodate(tmp))
2430 ret = -EIO;
2431 free_extent_buffer(tmp);
2432 } else {
2433 ret = PTR_ERR(tmp);
2434 }
2435
2436 btrfs_release_path(p);
2437 return ret;
2438}
2439
2440/*
2441 * helper function for btrfs_search_slot. This does all of the checks
2442 * for node-level blocks and does any balancing required based on
2443 * the ins_len.
2444 *
2445 * If no extra work was required, zero is returned. If we had to
2446 * drop the path, -EAGAIN is returned and btrfs_search_slot must
2447 * start over
2448 */
2449static int
2450setup_nodes_for_search(struct btrfs_trans_handle *trans,
2451 struct btrfs_root *root, struct btrfs_path *p,
2452 struct extent_buffer *b, int level, int ins_len,
2453 int *write_lock_level)
2454{
2455 struct btrfs_fs_info *fs_info = root->fs_info;
2456 int ret;
2457
2458 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2459 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2460 int sret;
2461
2462 if (*write_lock_level < level + 1) {
2463 *write_lock_level = level + 1;
2464 btrfs_release_path(p);
2465 goto again;
2466 }
2467
2468 btrfs_set_path_blocking(p);
2469 reada_for_balance(fs_info, p, level);
2470 sret = split_node(trans, root, p, level);
2471
2472 BUG_ON(sret > 0);
2473 if (sret) {
2474 ret = sret;
2475 goto done;
2476 }
2477 b = p->nodes[level];
2478 } else if (ins_len < 0 && btrfs_header_nritems(b) <
2479 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2480 int sret;
2481
2482 if (*write_lock_level < level + 1) {
2483 *write_lock_level = level + 1;
2484 btrfs_release_path(p);
2485 goto again;
2486 }
2487
2488 btrfs_set_path_blocking(p);
2489 reada_for_balance(fs_info, p, level);
2490 sret = balance_level(trans, root, p, level);
2491
2492 if (sret) {
2493 ret = sret;
2494 goto done;
2495 }
2496 b = p->nodes[level];
2497 if (!b) {
2498 btrfs_release_path(p);
2499 goto again;
2500 }
2501 BUG_ON(btrfs_header_nritems(b) == 1);
2502 }
2503 return 0;
2504
2505again:
2506 ret = -EAGAIN;
2507done:
2508 return ret;
2509}
2510
2511int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2512 u64 iobjectid, u64 ioff, u8 key_type,
2513 struct btrfs_key *found_key)
2514{
2515 int ret;
2516 struct btrfs_key key;
2517 struct extent_buffer *eb;
2518
2519 ASSERT(path);
2520 ASSERT(found_key);
2521
2522 key.type = key_type;
2523 key.objectid = iobjectid;
2524 key.offset = ioff;
2525
2526 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2527 if (ret < 0)
2528 return ret;
2529
2530 eb = path->nodes[0];
2531 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2532 ret = btrfs_next_leaf(fs_root, path);
2533 if (ret)
2534 return ret;
2535 eb = path->nodes[0];
2536 }
2537
2538 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2539 if (found_key->type != key.type ||
2540 found_key->objectid != key.objectid)
2541 return 1;
2542
2543 return 0;
2544}
2545
2546static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2547 struct btrfs_path *p,
2548 int write_lock_level)
2549{
2550 struct btrfs_fs_info *fs_info = root->fs_info;
2551 struct extent_buffer *b;
2552 int root_lock;
2553 int level = 0;
2554
2555 /* We try very hard to do read locks on the root */
2556 root_lock = BTRFS_READ_LOCK;
2557
2558 if (p->search_commit_root) {
2559 /*
2560 * The commit roots are read only so we always do read locks,
2561 * and we always must hold the commit_root_sem when doing
2562 * searches on them, the only exception is send where we don't
2563 * want to block transaction commits for a long time, so
2564 * we need to clone the commit root in order to avoid races
2565 * with transaction commits that create a snapshot of one of
2566 * the roots used by a send operation.
2567 */
2568 if (p->need_commit_sem) {
2569 down_read(&fs_info->commit_root_sem);
2570 b = btrfs_clone_extent_buffer(root->commit_root);
2571 up_read(&fs_info->commit_root_sem);
2572 if (!b)
2573 return ERR_PTR(-ENOMEM);
2574
2575 } else {
2576 b = root->commit_root;
2577 atomic_inc(&b->refs);
2578 }
2579 level = btrfs_header_level(b);
2580 /*
2581 * Ensure that all callers have set skip_locking when
2582 * p->search_commit_root = 1.
2583 */
2584 ASSERT(p->skip_locking == 1);
2585
2586 goto out;
2587 }
2588
2589 if (p->skip_locking) {
2590 b = btrfs_root_node(root);
2591 level = btrfs_header_level(b);
2592 goto out;
2593 }
2594
2595 /*
2596 * If the level is set to maximum, we can skip trying to get the read
2597 * lock.
2598 */
2599 if (write_lock_level < BTRFS_MAX_LEVEL) {
2600 /*
2601 * We don't know the level of the root node until we actually
2602 * have it read locked
2603 */
2604 b = btrfs_read_lock_root_node(root);
2605 level = btrfs_header_level(b);
2606 if (level > write_lock_level)
2607 goto out;
2608
2609 /* Whoops, must trade for write lock */
2610 btrfs_tree_read_unlock(b);
2611 free_extent_buffer(b);
2612 }
2613
2614 b = btrfs_lock_root_node(root);
2615 root_lock = BTRFS_WRITE_LOCK;
2616
2617 /* The level might have changed, check again */
2618 level = btrfs_header_level(b);
2619
2620out:
2621 p->nodes[level] = b;
2622 if (!p->skip_locking)
2623 p->locks[level] = root_lock;
2624 /*
2625 * Callers are responsible for dropping b's references.
2626 */
2627 return b;
2628}
2629
2630
2631/*
2632 * btrfs_search_slot - look for a key in a tree and perform necessary
2633 * modifications to preserve tree invariants.
2634 *
2635 * @trans: Handle of transaction, used when modifying the tree
2636 * @p: Holds all btree nodes along the search path
2637 * @root: The root node of the tree
2638 * @key: The key we are looking for
2639 * @ins_len: Indicates purpose of search, for inserts it is 1, for
2640 * deletions it's -1. 0 for plain searches
2641 * @cow: boolean should CoW operations be performed. Must always be 1
2642 * when modifying the tree.
2643 *
2644 * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2645 * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2646 *
2647 * If @key is found, 0 is returned and you can find the item in the leaf level
2648 * of the path (level 0)
2649 *
2650 * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2651 * points to the slot where it should be inserted
2652 *
2653 * If an error is encountered while searching the tree a negative error number
2654 * is returned
2655 */
2656int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2657 const struct btrfs_key *key, struct btrfs_path *p,
2658 int ins_len, int cow)
2659{
2660 struct extent_buffer *b;
2661 int slot;
2662 int ret;
2663 int err;
2664 int level;
2665 int lowest_unlock = 1;
2666 /* everything at write_lock_level or lower must be write locked */
2667 int write_lock_level = 0;
2668 u8 lowest_level = 0;
2669 int min_write_lock_level;
2670 int prev_cmp;
2671
2672 lowest_level = p->lowest_level;
2673 WARN_ON(lowest_level && ins_len > 0);
2674 WARN_ON(p->nodes[0] != NULL);
2675 BUG_ON(!cow && ins_len);
2676
2677 if (ins_len < 0) {
2678 lowest_unlock = 2;
2679
2680 /* when we are removing items, we might have to go up to level
2681 * two as we update tree pointers Make sure we keep write
2682 * for those levels as well
2683 */
2684 write_lock_level = 2;
2685 } else if (ins_len > 0) {
2686 /*
2687 * for inserting items, make sure we have a write lock on
2688 * level 1 so we can update keys
2689 */
2690 write_lock_level = 1;
2691 }
2692
2693 if (!cow)
2694 write_lock_level = -1;
2695
2696 if (cow && (p->keep_locks || p->lowest_level))
2697 write_lock_level = BTRFS_MAX_LEVEL;
2698
2699 min_write_lock_level = write_lock_level;
2700
2701again:
2702 prev_cmp = -1;
2703 b = btrfs_search_slot_get_root(root, p, write_lock_level);
2704 if (IS_ERR(b)) {
2705 ret = PTR_ERR(b);
2706 goto done;
2707 }
2708
2709 while (b) {
2710 int dec = 0;
2711
2712 level = btrfs_header_level(b);
2713
2714 if (cow) {
2715 bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2716
2717 /*
2718 * if we don't really need to cow this block
2719 * then we don't want to set the path blocking,
2720 * so we test it here
2721 */
2722 if (!should_cow_block(trans, root, b)) {
2723 trans->dirty = true;
2724 goto cow_done;
2725 }
2726
2727 /*
2728 * must have write locks on this node and the
2729 * parent
2730 */
2731 if (level > write_lock_level ||
2732 (level + 1 > write_lock_level &&
2733 level + 1 < BTRFS_MAX_LEVEL &&
2734 p->nodes[level + 1])) {
2735 write_lock_level = level + 1;
2736 btrfs_release_path(p);
2737 goto again;
2738 }
2739
2740 btrfs_set_path_blocking(p);
2741 if (last_level)
2742 err = btrfs_cow_block(trans, root, b, NULL, 0,
2743 &b);
2744 else
2745 err = btrfs_cow_block(trans, root, b,
2746 p->nodes[level + 1],
2747 p->slots[level + 1], &b);
2748 if (err) {
2749 ret = err;
2750 goto done;
2751 }
2752 }
2753cow_done:
2754 p->nodes[level] = b;
2755 /*
2756 * Leave path with blocking locks to avoid massive
2757 * lock context switch, this is made on purpose.
2758 */
2759
2760 /*
2761 * we have a lock on b and as long as we aren't changing
2762 * the tree, there is no way to for the items in b to change.
2763 * It is safe to drop the lock on our parent before we
2764 * go through the expensive btree search on b.
2765 *
2766 * If we're inserting or deleting (ins_len != 0), then we might
2767 * be changing slot zero, which may require changing the parent.
2768 * So, we can't drop the lock until after we know which slot
2769 * we're operating on.
2770 */
2771 if (!ins_len && !p->keep_locks) {
2772 int u = level + 1;
2773
2774 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2775 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2776 p->locks[u] = 0;
2777 }
2778 }
2779
2780 /*
2781 * If btrfs_bin_search returns an exact match (prev_cmp == 0)
2782 * we can safely assume the target key will always be in slot 0
2783 * on lower levels due to the invariants BTRFS' btree provides,
2784 * namely that a btrfs_key_ptr entry always points to the
2785 * lowest key in the child node, thus we can skip searching
2786 * lower levels
2787 */
2788 if (prev_cmp == 0) {
2789 slot = 0;
2790 ret = 0;
2791 } else {
2792 ret = btrfs_bin_search(b, key, &slot);
2793 prev_cmp = ret;
2794 if (ret < 0)
2795 goto done;
2796 }
2797
2798 if (level == 0) {
2799 p->slots[level] = slot;
2800 if (ins_len > 0 &&
2801 btrfs_leaf_free_space(b) < ins_len) {
2802 if (write_lock_level < 1) {
2803 write_lock_level = 1;
2804 btrfs_release_path(p);
2805 goto again;
2806 }
2807
2808 btrfs_set_path_blocking(p);
2809 err = split_leaf(trans, root, key,
2810 p, ins_len, ret == 0);
2811
2812 BUG_ON(err > 0);
2813 if (err) {
2814 ret = err;
2815 goto done;
2816 }
2817 }
2818 if (!p->search_for_split)
2819 unlock_up(p, level, lowest_unlock,
2820 min_write_lock_level, NULL);
2821 goto done;
2822 }
2823 if (ret && slot > 0) {
2824 dec = 1;
2825 slot--;
2826 }
2827 p->slots[level] = slot;
2828 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
2829 &write_lock_level);
2830 if (err == -EAGAIN)
2831 goto again;
2832 if (err) {
2833 ret = err;
2834 goto done;
2835 }
2836 b = p->nodes[level];
2837 slot = p->slots[level];
2838
2839 /*
2840 * Slot 0 is special, if we change the key we have to update
2841 * the parent pointer which means we must have a write lock on
2842 * the parent
2843 */
2844 if (slot == 0 && ins_len && write_lock_level < level + 1) {
2845 write_lock_level = level + 1;
2846 btrfs_release_path(p);
2847 goto again;
2848 }
2849
2850 unlock_up(p, level, lowest_unlock, min_write_lock_level,
2851 &write_lock_level);
2852
2853 if (level == lowest_level) {
2854 if (dec)
2855 p->slots[level]++;
2856 goto done;
2857 }
2858
2859 err = read_block_for_search(root, p, &b, level, slot, key);
2860 if (err == -EAGAIN)
2861 goto again;
2862 if (err) {
2863 ret = err;
2864 goto done;
2865 }
2866
2867 if (!p->skip_locking) {
2868 level = btrfs_header_level(b);
2869 if (level <= write_lock_level) {
2870 if (!btrfs_try_tree_write_lock(b)) {
2871 btrfs_set_path_blocking(p);
2872 btrfs_tree_lock(b);
2873 }
2874 p->locks[level] = BTRFS_WRITE_LOCK;
2875 } else {
2876 if (!btrfs_tree_read_lock_atomic(b)) {
2877 btrfs_set_path_blocking(p);
2878 btrfs_tree_read_lock(b);
2879 }
2880 p->locks[level] = BTRFS_READ_LOCK;
2881 }
2882 p->nodes[level] = b;
2883 }
2884 }
2885 ret = 1;
2886done:
2887 /*
2888 * we don't really know what they plan on doing with the path
2889 * from here on, so for now just mark it as blocking
2890 */
2891 if (!p->leave_spinning)
2892 btrfs_set_path_blocking(p);
2893 if (ret < 0 && !p->skip_release_on_error)
2894 btrfs_release_path(p);
2895 return ret;
2896}
2897
2898/*
2899 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2900 * current state of the tree together with the operations recorded in the tree
2901 * modification log to search for the key in a previous version of this tree, as
2902 * denoted by the time_seq parameter.
2903 *
2904 * Naturally, there is no support for insert, delete or cow operations.
2905 *
2906 * The resulting path and return value will be set up as if we called
2907 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2908 */
2909int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2910 struct btrfs_path *p, u64 time_seq)
2911{
2912 struct btrfs_fs_info *fs_info = root->fs_info;
2913 struct extent_buffer *b;
2914 int slot;
2915 int ret;
2916 int err;
2917 int level;
2918 int lowest_unlock = 1;
2919 u8 lowest_level = 0;
2920
2921 lowest_level = p->lowest_level;
2922 WARN_ON(p->nodes[0] != NULL);
2923
2924 if (p->search_commit_root) {
2925 BUG_ON(time_seq);
2926 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2927 }
2928
2929again:
2930 b = get_old_root(root, time_seq);
2931 if (!b) {
2932 ret = -EIO;
2933 goto done;
2934 }
2935 level = btrfs_header_level(b);
2936 p->locks[level] = BTRFS_READ_LOCK;
2937
2938 while (b) {
2939 int dec = 0;
2940
2941 level = btrfs_header_level(b);
2942 p->nodes[level] = b;
2943
2944 /*
2945 * we have a lock on b and as long as we aren't changing
2946 * the tree, there is no way to for the items in b to change.
2947 * It is safe to drop the lock on our parent before we
2948 * go through the expensive btree search on b.
2949 */
2950 btrfs_unlock_up_safe(p, level + 1);
2951
2952 ret = btrfs_bin_search(b, key, &slot);
2953 if (ret < 0)
2954 goto done;
2955
2956 if (level == 0) {
2957 p->slots[level] = slot;
2958 unlock_up(p, level, lowest_unlock, 0, NULL);
2959 goto done;
2960 }
2961
2962 if (ret && slot > 0) {
2963 dec = 1;
2964 slot--;
2965 }
2966 p->slots[level] = slot;
2967 unlock_up(p, level, lowest_unlock, 0, NULL);
2968
2969 if (level == lowest_level) {
2970 if (dec)
2971 p->slots[level]++;
2972 goto done;
2973 }
2974
2975 err = read_block_for_search(root, p, &b, level, slot, key);
2976 if (err == -EAGAIN)
2977 goto again;
2978 if (err) {
2979 ret = err;
2980 goto done;
2981 }
2982
2983 level = btrfs_header_level(b);
2984 if (!btrfs_tree_read_lock_atomic(b)) {
2985 btrfs_set_path_blocking(p);
2986 btrfs_tree_read_lock(b);
2987 }
2988 b = tree_mod_log_rewind(fs_info, p, b, time_seq);
2989 if (!b) {
2990 ret = -ENOMEM;
2991 goto done;
2992 }
2993 p->locks[level] = BTRFS_READ_LOCK;
2994 p->nodes[level] = b;
2995 }
2996 ret = 1;
2997done:
2998 if (!p->leave_spinning)
2999 btrfs_set_path_blocking(p);
3000 if (ret < 0)
3001 btrfs_release_path(p);
3002
3003 return ret;
3004}
3005
3006/*
3007 * helper to use instead of search slot if no exact match is needed but
3008 * instead the next or previous item should be returned.
3009 * When find_higher is true, the next higher item is returned, the next lower
3010 * otherwise.
3011 * When return_any and find_higher are both true, and no higher item is found,
3012 * return the next lower instead.
3013 * When return_any is true and find_higher is false, and no lower item is found,
3014 * return the next higher instead.
3015 * It returns 0 if any item is found, 1 if none is found (tree empty), and
3016 * < 0 on error
3017 */
3018int btrfs_search_slot_for_read(struct btrfs_root *root,
3019 const struct btrfs_key *key,
3020 struct btrfs_path *p, int find_higher,
3021 int return_any)
3022{
3023 int ret;
3024 struct extent_buffer *leaf;
3025
3026again:
3027 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3028 if (ret <= 0)
3029 return ret;
3030 /*
3031 * a return value of 1 means the path is at the position where the
3032 * item should be inserted. Normally this is the next bigger item,
3033 * but in case the previous item is the last in a leaf, path points
3034 * to the first free slot in the previous leaf, i.e. at an invalid
3035 * item.
3036 */
3037 leaf = p->nodes[0];
3038
3039 if (find_higher) {
3040 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3041 ret = btrfs_next_leaf(root, p);
3042 if (ret <= 0)
3043 return ret;
3044 if (!return_any)
3045 return 1;
3046 /*
3047 * no higher item found, return the next
3048 * lower instead
3049 */
3050 return_any = 0;
3051 find_higher = 0;
3052 btrfs_release_path(p);
3053 goto again;
3054 }
3055 } else {
3056 if (p->slots[0] == 0) {
3057 ret = btrfs_prev_leaf(root, p);
3058 if (ret < 0)
3059 return ret;
3060 if (!ret) {
3061 leaf = p->nodes[0];
3062 if (p->slots[0] == btrfs_header_nritems(leaf))
3063 p->slots[0]--;
3064 return 0;
3065 }
3066 if (!return_any)
3067 return 1;
3068 /*
3069 * no lower item found, return the next
3070 * higher instead
3071 */
3072 return_any = 0;
3073 find_higher = 1;
3074 btrfs_release_path(p);
3075 goto again;
3076 } else {
3077 --p->slots[0];
3078 }
3079 }
3080 return 0;
3081}
3082
3083/*
3084 * adjust the pointers going up the tree, starting at level
3085 * making sure the right key of each node is points to 'key'.
3086 * This is used after shifting pointers to the left, so it stops
3087 * fixing up pointers when a given leaf/node is not in slot 0 of the
3088 * higher levels
3089 *
3090 */
3091static void fixup_low_keys(struct btrfs_path *path,
3092 struct btrfs_disk_key *key, int level)
3093{
3094 int i;
3095 struct extent_buffer *t;
3096 int ret;
3097
3098 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3099 int tslot = path->slots[i];
3100
3101 if (!path->nodes[i])
3102 break;
3103 t = path->nodes[i];
3104 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3105 GFP_ATOMIC);
3106 BUG_ON(ret < 0);
3107 btrfs_set_node_key(t, key, tslot);
3108 btrfs_mark_buffer_dirty(path->nodes[i]);
3109 if (tslot != 0)
3110 break;
3111 }
3112}
3113
3114/*
3115 * update item key.
3116 *
3117 * This function isn't completely safe. It's the caller's responsibility
3118 * that the new key won't break the order
3119 */
3120void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3121 struct btrfs_path *path,
3122 const struct btrfs_key *new_key)
3123{
3124 struct btrfs_disk_key disk_key;
3125 struct extent_buffer *eb;
3126 int slot;
3127
3128 eb = path->nodes[0];
3129 slot = path->slots[0];
3130 if (slot > 0) {
3131 btrfs_item_key(eb, &disk_key, slot - 1);
3132 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
3133 btrfs_crit(fs_info,
3134 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3135 slot, btrfs_disk_key_objectid(&disk_key),
3136 btrfs_disk_key_type(&disk_key),
3137 btrfs_disk_key_offset(&disk_key),
3138 new_key->objectid, new_key->type,
3139 new_key->offset);
3140 btrfs_print_leaf(eb);
3141 BUG();
3142 }
3143 }
3144 if (slot < btrfs_header_nritems(eb) - 1) {
3145 btrfs_item_key(eb, &disk_key, slot + 1);
3146 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
3147 btrfs_crit(fs_info,
3148 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3149 slot, btrfs_disk_key_objectid(&disk_key),
3150 btrfs_disk_key_type(&disk_key),
3151 btrfs_disk_key_offset(&disk_key),
3152 new_key->objectid, new_key->type,
3153 new_key->offset);
3154 btrfs_print_leaf(eb);
3155 BUG();
3156 }
3157 }
3158
3159 btrfs_cpu_key_to_disk(&disk_key, new_key);
3160 btrfs_set_item_key(eb, &disk_key, slot);
3161 btrfs_mark_buffer_dirty(eb);
3162 if (slot == 0)
3163 fixup_low_keys(path, &disk_key, 1);
3164}
3165
3166/*
3167 * try to push data from one node into the next node left in the
3168 * tree.
3169 *
3170 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3171 * error, and > 0 if there was no room in the left hand block.
3172 */
3173static int push_node_left(struct btrfs_trans_handle *trans,
3174 struct extent_buffer *dst,
3175 struct extent_buffer *src, int empty)
3176{
3177 struct btrfs_fs_info *fs_info = trans->fs_info;
3178 int push_items = 0;
3179 int src_nritems;
3180 int dst_nritems;
3181 int ret = 0;
3182
3183 src_nritems = btrfs_header_nritems(src);
3184 dst_nritems = btrfs_header_nritems(dst);
3185 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3186 WARN_ON(btrfs_header_generation(src) != trans->transid);
3187 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3188
3189 if (!empty && src_nritems <= 8)
3190 return 1;
3191
3192 if (push_items <= 0)
3193 return 1;
3194
3195 if (empty) {
3196 push_items = min(src_nritems, push_items);
3197 if (push_items < src_nritems) {
3198 /* leave at least 8 pointers in the node if
3199 * we aren't going to empty it
3200 */
3201 if (src_nritems - push_items < 8) {
3202 if (push_items <= 8)
3203 return 1;
3204 push_items -= 8;
3205 }
3206 }
3207 } else
3208 push_items = min(src_nritems - 8, push_items);
3209
3210 ret = tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
3211 if (ret) {
3212 btrfs_abort_transaction(trans, ret);
3213 return ret;
3214 }
3215 copy_extent_buffer(dst, src,
3216 btrfs_node_key_ptr_offset(dst_nritems),
3217 btrfs_node_key_ptr_offset(0),
3218 push_items * sizeof(struct btrfs_key_ptr));
3219
3220 if (push_items < src_nritems) {
3221 /*
3222 * Don't call tree_mod_log_insert_move here, key removal was
3223 * already fully logged by tree_mod_log_eb_copy above.
3224 */
3225 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3226 btrfs_node_key_ptr_offset(push_items),
3227 (src_nritems - push_items) *
3228 sizeof(struct btrfs_key_ptr));
3229 }
3230 btrfs_set_header_nritems(src, src_nritems - push_items);
3231 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3232 btrfs_mark_buffer_dirty(src);
3233 btrfs_mark_buffer_dirty(dst);
3234
3235 return ret;
3236}
3237
3238/*
3239 * try to push data from one node into the next node right in the
3240 * tree.
3241 *
3242 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3243 * error, and > 0 if there was no room in the right hand block.
3244 *
3245 * this will only push up to 1/2 the contents of the left node over
3246 */
3247static int balance_node_right(struct btrfs_trans_handle *trans,
3248 struct extent_buffer *dst,
3249 struct extent_buffer *src)
3250{
3251 struct btrfs_fs_info *fs_info = trans->fs_info;
3252 int push_items = 0;
3253 int max_push;
3254 int src_nritems;
3255 int dst_nritems;
3256 int ret = 0;
3257
3258 WARN_ON(btrfs_header_generation(src) != trans->transid);
3259 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3260
3261 src_nritems = btrfs_header_nritems(src);
3262 dst_nritems = btrfs_header_nritems(dst);
3263 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3264 if (push_items <= 0)
3265 return 1;
3266
3267 if (src_nritems < 4)
3268 return 1;
3269
3270 max_push = src_nritems / 2 + 1;
3271 /* don't try to empty the node */
3272 if (max_push >= src_nritems)
3273 return 1;
3274
3275 if (max_push < push_items)
3276 push_items = max_push;
3277
3278 ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3279 BUG_ON(ret < 0);
3280 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3281 btrfs_node_key_ptr_offset(0),
3282 (dst_nritems) *
3283 sizeof(struct btrfs_key_ptr));
3284
3285 ret = tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
3286 push_items);
3287 if (ret) {
3288 btrfs_abort_transaction(trans, ret);
3289 return ret;
3290 }
3291 copy_extent_buffer(dst, src,
3292 btrfs_node_key_ptr_offset(0),
3293 btrfs_node_key_ptr_offset(src_nritems - push_items),
3294 push_items * sizeof(struct btrfs_key_ptr));
3295
3296 btrfs_set_header_nritems(src, src_nritems - push_items);
3297 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3298
3299 btrfs_mark_buffer_dirty(src);
3300 btrfs_mark_buffer_dirty(dst);
3301
3302 return ret;
3303}
3304
3305/*
3306 * helper function to insert a new root level in the tree.
3307 * A new node is allocated, and a single item is inserted to
3308 * point to the existing root
3309 *
3310 * returns zero on success or < 0 on failure.
3311 */
3312static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3313 struct btrfs_root *root,
3314 struct btrfs_path *path, int level)
3315{
3316 struct btrfs_fs_info *fs_info = root->fs_info;
3317 u64 lower_gen;
3318 struct extent_buffer *lower;
3319 struct extent_buffer *c;
3320 struct extent_buffer *old;
3321 struct btrfs_disk_key lower_key;
3322 int ret;
3323
3324 BUG_ON(path->nodes[level]);
3325 BUG_ON(path->nodes[level-1] != root->node);
3326
3327 lower = path->nodes[level-1];
3328 if (level == 1)
3329 btrfs_item_key(lower, &lower_key, 0);
3330 else
3331 btrfs_node_key(lower, &lower_key, 0);
3332
3333 c = alloc_tree_block_no_bg_flush(trans, root, 0, &lower_key, level,
3334 root->node->start, 0);
3335 if (IS_ERR(c))
3336 return PTR_ERR(c);
3337
3338 root_add_used(root, fs_info->nodesize);
3339
3340 btrfs_set_header_nritems(c, 1);
3341 btrfs_set_node_key(c, &lower_key, 0);
3342 btrfs_set_node_blockptr(c, 0, lower->start);
3343 lower_gen = btrfs_header_generation(lower);
3344 WARN_ON(lower_gen != trans->transid);
3345
3346 btrfs_set_node_ptr_generation(c, 0, lower_gen);
3347
3348 btrfs_mark_buffer_dirty(c);
3349
3350 old = root->node;
3351 ret = tree_mod_log_insert_root(root->node, c, 0);
3352 BUG_ON(ret < 0);
3353 rcu_assign_pointer(root->node, c);
3354
3355 /* the super has an extra ref to root->node */
3356 free_extent_buffer(old);
3357
3358 add_root_to_dirty_list(root);
3359 atomic_inc(&c->refs);
3360 path->nodes[level] = c;
3361 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3362 path->slots[level] = 0;
3363 return 0;
3364}
3365
3366/*
3367 * worker function to insert a single pointer in a node.
3368 * the node should have enough room for the pointer already
3369 *
3370 * slot and level indicate where you want the key to go, and
3371 * blocknr is the block the key points to.
3372 */
3373static void insert_ptr(struct btrfs_trans_handle *trans,
3374 struct btrfs_path *path,
3375 struct btrfs_disk_key *key, u64 bytenr,
3376 int slot, int level)
3377{
3378 struct extent_buffer *lower;
3379 int nritems;
3380 int ret;
3381
3382 BUG_ON(!path->nodes[level]);
3383 btrfs_assert_tree_locked(path->nodes[level]);
3384 lower = path->nodes[level];
3385 nritems = btrfs_header_nritems(lower);
3386 BUG_ON(slot > nritems);
3387 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
3388 if (slot != nritems) {
3389 if (level) {
3390 ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3391 nritems - slot);
3392 BUG_ON(ret < 0);
3393 }
3394 memmove_extent_buffer(lower,
3395 btrfs_node_key_ptr_offset(slot + 1),
3396 btrfs_node_key_ptr_offset(slot),
3397 (nritems - slot) * sizeof(struct btrfs_key_ptr));
3398 }
3399 if (level) {
3400 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3401 GFP_NOFS);
3402 BUG_ON(ret < 0);
3403 }
3404 btrfs_set_node_key(lower, key, slot);
3405 btrfs_set_node_blockptr(lower, slot, bytenr);
3406 WARN_ON(trans->transid == 0);
3407 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3408 btrfs_set_header_nritems(lower, nritems + 1);
3409 btrfs_mark_buffer_dirty(lower);
3410}
3411
3412/*
3413 * split the node at the specified level in path in two.
3414 * The path is corrected to point to the appropriate node after the split
3415 *
3416 * Before splitting this tries to make some room in the node by pushing
3417 * left and right, if either one works, it returns right away.
3418 *
3419 * returns 0 on success and < 0 on failure
3420 */
3421static noinline int split_node(struct btrfs_trans_handle *trans,
3422 struct btrfs_root *root,
3423 struct btrfs_path *path, int level)
3424{
3425 struct btrfs_fs_info *fs_info = root->fs_info;
3426 struct extent_buffer *c;
3427 struct extent_buffer *split;
3428 struct btrfs_disk_key disk_key;
3429 int mid;
3430 int ret;
3431 u32 c_nritems;
3432
3433 c = path->nodes[level];
3434 WARN_ON(btrfs_header_generation(c) != trans->transid);
3435 if (c == root->node) {
3436 /*
3437 * trying to split the root, lets make a new one
3438 *
3439 * tree mod log: We don't log_removal old root in
3440 * insert_new_root, because that root buffer will be kept as a
3441 * normal node. We are going to log removal of half of the
3442 * elements below with tree_mod_log_eb_copy. We're holding a
3443 * tree lock on the buffer, which is why we cannot race with
3444 * other tree_mod_log users.
3445 */
3446 ret = insert_new_root(trans, root, path, level + 1);
3447 if (ret)
3448 return ret;
3449 } else {
3450 ret = push_nodes_for_insert(trans, root, path, level);
3451 c = path->nodes[level];
3452 if (!ret && btrfs_header_nritems(c) <
3453 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3454 return 0;
3455 if (ret < 0)
3456 return ret;
3457 }
3458
3459 c_nritems = btrfs_header_nritems(c);
3460 mid = (c_nritems + 1) / 2;
3461 btrfs_node_key(c, &disk_key, mid);
3462
3463 split = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, level,
3464 c->start, 0);
3465 if (IS_ERR(split))
3466 return PTR_ERR(split);
3467
3468 root_add_used(root, fs_info->nodesize);
3469 ASSERT(btrfs_header_level(c) == level);
3470
3471 ret = tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
3472 if (ret) {
3473 btrfs_abort_transaction(trans, ret);
3474 return ret;
3475 }
3476 copy_extent_buffer(split, c,
3477 btrfs_node_key_ptr_offset(0),
3478 btrfs_node_key_ptr_offset(mid),
3479 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3480 btrfs_set_header_nritems(split, c_nritems - mid);
3481 btrfs_set_header_nritems(c, mid);
3482 ret = 0;
3483
3484 btrfs_mark_buffer_dirty(c);
3485 btrfs_mark_buffer_dirty(split);
3486
3487 insert_ptr(trans, path, &disk_key, split->start,
3488 path->slots[level + 1] + 1, level + 1);
3489
3490 if (path->slots[level] >= mid) {
3491 path->slots[level] -= mid;
3492 btrfs_tree_unlock(c);
3493 free_extent_buffer(c);
3494 path->nodes[level] = split;
3495 path->slots[level + 1] += 1;
3496 } else {
3497 btrfs_tree_unlock(split);
3498 free_extent_buffer(split);
3499 }
3500 return ret;
3501}
3502
3503/*
3504 * how many bytes are required to store the items in a leaf. start
3505 * and nr indicate which items in the leaf to check. This totals up the
3506 * space used both by the item structs and the item data
3507 */
3508static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3509{
3510 struct btrfs_item *start_item;
3511 struct btrfs_item *end_item;
3512 int data_len;
3513 int nritems = btrfs_header_nritems(l);
3514 int end = min(nritems, start + nr) - 1;
3515
3516 if (!nr)
3517 return 0;
3518 start_item = btrfs_item_nr(start);
3519 end_item = btrfs_item_nr(end);
3520 data_len = btrfs_item_offset(l, start_item) +
3521 btrfs_item_size(l, start_item);
3522 data_len = data_len - btrfs_item_offset(l, end_item);
3523 data_len += sizeof(struct btrfs_item) * nr;
3524 WARN_ON(data_len < 0);
3525 return data_len;
3526}
3527
3528/*
3529 * The space between the end of the leaf items and
3530 * the start of the leaf data. IOW, how much room
3531 * the leaf has left for both items and data
3532 */
3533noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
3534{
3535 struct btrfs_fs_info *fs_info = leaf->fs_info;
3536 int nritems = btrfs_header_nritems(leaf);
3537 int ret;
3538
3539 ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3540 if (ret < 0) {
3541 btrfs_crit(fs_info,
3542 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3543 ret,
3544 (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3545 leaf_space_used(leaf, 0, nritems), nritems);
3546 }
3547 return ret;
3548}
3549
3550/*
3551 * min slot controls the lowest index we're willing to push to the
3552 * right. We'll push up to and including min_slot, but no lower
3553 */
3554static noinline int __push_leaf_right(struct btrfs_path *path,
3555 int data_size, int empty,
3556 struct extent_buffer *right,
3557 int free_space, u32 left_nritems,
3558 u32 min_slot)
3559{
3560 struct btrfs_fs_info *fs_info = right->fs_info;
3561 struct extent_buffer *left = path->nodes[0];
3562 struct extent_buffer *upper = path->nodes[1];
3563 struct btrfs_map_token token;
3564 struct btrfs_disk_key disk_key;
3565 int slot;
3566 u32 i;
3567 int push_space = 0;
3568 int push_items = 0;
3569 struct btrfs_item *item;
3570 u32 nr;
3571 u32 right_nritems;
3572 u32 data_end;
3573 u32 this_item_size;
3574
3575 if (empty)
3576 nr = 0;
3577 else
3578 nr = max_t(u32, 1, min_slot);
3579
3580 if (path->slots[0] >= left_nritems)
3581 push_space += data_size;
3582
3583 slot = path->slots[1];
3584 i = left_nritems - 1;
3585 while (i >= nr) {
3586 item = btrfs_item_nr(i);
3587
3588 if (!empty && push_items > 0) {
3589 if (path->slots[0] > i)
3590 break;
3591 if (path->slots[0] == i) {
3592 int space = btrfs_leaf_free_space(left);
3593
3594 if (space + push_space * 2 > free_space)
3595 break;
3596 }
3597 }
3598
3599 if (path->slots[0] == i)
3600 push_space += data_size;
3601
3602 this_item_size = btrfs_item_size(left, item);
3603 if (this_item_size + sizeof(*item) + push_space > free_space)
3604 break;
3605
3606 push_items++;
3607 push_space += this_item_size + sizeof(*item);
3608 if (i == 0)
3609 break;
3610 i--;
3611 }
3612
3613 if (push_items == 0)
3614 goto out_unlock;
3615
3616 WARN_ON(!empty && push_items == left_nritems);
3617
3618 /* push left to right */
3619 right_nritems = btrfs_header_nritems(right);
3620
3621 push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3622 push_space -= leaf_data_end(left);
3623
3624 /* make room in the right data area */
3625 data_end = leaf_data_end(right);
3626 memmove_extent_buffer(right,
3627 BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
3628 BTRFS_LEAF_DATA_OFFSET + data_end,
3629 BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3630
3631 /* copy from the left data area */
3632 copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
3633 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3634 BTRFS_LEAF_DATA_OFFSET + leaf_data_end(left),
3635 push_space);
3636
3637 memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3638 btrfs_item_nr_offset(0),
3639 right_nritems * sizeof(struct btrfs_item));
3640
3641 /* copy the items from left to right */
3642 copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3643 btrfs_item_nr_offset(left_nritems - push_items),
3644 push_items * sizeof(struct btrfs_item));
3645
3646 /* update the item pointers */
3647 btrfs_init_map_token(&token, right);
3648 right_nritems += push_items;
3649 btrfs_set_header_nritems(right, right_nritems);
3650 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3651 for (i = 0; i < right_nritems; i++) {
3652 item = btrfs_item_nr(i);
3653 push_space -= btrfs_token_item_size(&token, item);
3654 btrfs_set_token_item_offset(&token, item, push_space);
3655 }
3656
3657 left_nritems -= push_items;
3658 btrfs_set_header_nritems(left, left_nritems);
3659
3660 if (left_nritems)
3661 btrfs_mark_buffer_dirty(left);
3662 else
3663 btrfs_clean_tree_block(left);
3664
3665 btrfs_mark_buffer_dirty(right);
3666
3667 btrfs_item_key(right, &disk_key, 0);
3668 btrfs_set_node_key(upper, &disk_key, slot + 1);
3669 btrfs_mark_buffer_dirty(upper);
3670
3671 /* then fixup the leaf pointer in the path */
3672 if (path->slots[0] >= left_nritems) {
3673 path->slots[0] -= left_nritems;
3674 if (btrfs_header_nritems(path->nodes[0]) == 0)
3675 btrfs_clean_tree_block(path->nodes[0]);
3676 btrfs_tree_unlock(path->nodes[0]);
3677 free_extent_buffer(path->nodes[0]);
3678 path->nodes[0] = right;
3679 path->slots[1] += 1;
3680 } else {
3681 btrfs_tree_unlock(right);
3682 free_extent_buffer(right);
3683 }
3684 return 0;
3685
3686out_unlock:
3687 btrfs_tree_unlock(right);
3688 free_extent_buffer(right);
3689 return 1;
3690}
3691
3692/*
3693 * push some data in the path leaf to the right, trying to free up at
3694 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3695 *
3696 * returns 1 if the push failed because the other node didn't have enough
3697 * room, 0 if everything worked out and < 0 if there were major errors.
3698 *
3699 * this will push starting from min_slot to the end of the leaf. It won't
3700 * push any slot lower than min_slot
3701 */
3702static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3703 *root, struct btrfs_path *path,
3704 int min_data_size, int data_size,
3705 int empty, u32 min_slot)
3706{
3707 struct extent_buffer *left = path->nodes[0];
3708 struct extent_buffer *right;
3709 struct extent_buffer *upper;
3710 int slot;
3711 int free_space;
3712 u32 left_nritems;
3713 int ret;
3714
3715 if (!path->nodes[1])
3716 return 1;
3717
3718 slot = path->slots[1];
3719 upper = path->nodes[1];
3720 if (slot >= btrfs_header_nritems(upper) - 1)
3721 return 1;
3722
3723 btrfs_assert_tree_locked(path->nodes[1]);
3724
3725 right = btrfs_read_node_slot(upper, slot + 1);
3726 /*
3727 * slot + 1 is not valid or we fail to read the right node,
3728 * no big deal, just return.
3729 */
3730 if (IS_ERR(right))
3731 return 1;
3732
3733 btrfs_tree_lock(right);
3734 btrfs_set_lock_blocking_write(right);
3735
3736 free_space = btrfs_leaf_free_space(right);
3737 if (free_space < data_size)
3738 goto out_unlock;
3739
3740 /* cow and double check */
3741 ret = btrfs_cow_block(trans, root, right, upper,
3742 slot + 1, &right);
3743 if (ret)
3744 goto out_unlock;
3745
3746 free_space = btrfs_leaf_free_space(right);
3747 if (free_space < data_size)
3748 goto out_unlock;
3749
3750 left_nritems = btrfs_header_nritems(left);
3751 if (left_nritems == 0)
3752 goto out_unlock;
3753
3754 if (path->slots[0] == left_nritems && !empty) {
3755 /* Key greater than all keys in the leaf, right neighbor has
3756 * enough room for it and we're not emptying our leaf to delete
3757 * it, therefore use right neighbor to insert the new item and
3758 * no need to touch/dirty our left leaf. */
3759 btrfs_tree_unlock(left);
3760 free_extent_buffer(left);
3761 path->nodes[0] = right;
3762 path->slots[0] = 0;
3763 path->slots[1]++;
3764 return 0;
3765 }
3766
3767 return __push_leaf_right(path, min_data_size, empty,
3768 right, free_space, left_nritems, min_slot);
3769out_unlock:
3770 btrfs_tree_unlock(right);
3771 free_extent_buffer(right);
3772 return 1;
3773}
3774
3775/*
3776 * push some data in the path leaf to the left, trying to free up at
3777 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3778 *
3779 * max_slot can put a limit on how far into the leaf we'll push items. The
3780 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3781 * items
3782 */
3783static noinline int __push_leaf_left(struct btrfs_path *path, int data_size,
3784 int empty, struct extent_buffer *left,
3785 int free_space, u32 right_nritems,
3786 u32 max_slot)
3787{
3788 struct btrfs_fs_info *fs_info = left->fs_info;
3789 struct btrfs_disk_key disk_key;
3790 struct extent_buffer *right = path->nodes[0];
3791 int i;
3792 int push_space = 0;
3793 int push_items = 0;
3794 struct btrfs_item *item;
3795 u32 old_left_nritems;
3796 u32 nr;
3797 int ret = 0;
3798 u32 this_item_size;
3799 u32 old_left_item_size;
3800 struct btrfs_map_token token;
3801
3802 if (empty)
3803 nr = min(right_nritems, max_slot);
3804 else
3805 nr = min(right_nritems - 1, max_slot);
3806
3807 for (i = 0; i < nr; i++) {
3808 item = btrfs_item_nr(i);
3809
3810 if (!empty && push_items > 0) {
3811 if (path->slots[0] < i)
3812 break;
3813 if (path->slots[0] == i) {
3814 int space = btrfs_leaf_free_space(right);
3815
3816 if (space + push_space * 2 > free_space)
3817 break;
3818 }
3819 }
3820
3821 if (path->slots[0] == i)
3822 push_space += data_size;
3823
3824 this_item_size = btrfs_item_size(right, item);
3825 if (this_item_size + sizeof(*item) + push_space > free_space)
3826 break;
3827
3828 push_items++;
3829 push_space += this_item_size + sizeof(*item);
3830 }
3831
3832 if (push_items == 0) {
3833 ret = 1;
3834 goto out;
3835 }
3836 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3837
3838 /* push data from right to left */
3839 copy_extent_buffer(left, right,
3840 btrfs_item_nr_offset(btrfs_header_nritems(left)),
3841 btrfs_item_nr_offset(0),
3842 push_items * sizeof(struct btrfs_item));
3843
3844 push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3845 btrfs_item_offset_nr(right, push_items - 1);
3846
3847 copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
3848 leaf_data_end(left) - push_space,
3849 BTRFS_LEAF_DATA_OFFSET +
3850 btrfs_item_offset_nr(right, push_items - 1),
3851 push_space);
3852 old_left_nritems = btrfs_header_nritems(left);
3853 BUG_ON(old_left_nritems <= 0);
3854
3855 btrfs_init_map_token(&token, left);
3856 old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3857 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3858 u32 ioff;
3859
3860 item = btrfs_item_nr(i);
3861
3862 ioff = btrfs_token_item_offset(&token, item);
3863 btrfs_set_token_item_offset(&token, item,
3864 ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
3865 }
3866 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3867
3868 /* fixup right node */
3869 if (push_items > right_nritems)
3870 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3871 right_nritems);
3872
3873 if (push_items < right_nritems) {
3874 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3875 leaf_data_end(right);
3876 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
3877 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3878 BTRFS_LEAF_DATA_OFFSET +
3879 leaf_data_end(right), push_space);
3880
3881 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3882 btrfs_item_nr_offset(push_items),
3883 (btrfs_header_nritems(right) - push_items) *
3884 sizeof(struct btrfs_item));
3885 }
3886
3887 btrfs_init_map_token(&token, right);
3888 right_nritems -= push_items;
3889 btrfs_set_header_nritems(right, right_nritems);
3890 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3891 for (i = 0; i < right_nritems; i++) {
3892 item = btrfs_item_nr(i);
3893
3894 push_space = push_space - btrfs_token_item_size(&token, item);
3895 btrfs_set_token_item_offset(&token, item, push_space);
3896 }
3897
3898 btrfs_mark_buffer_dirty(left);
3899 if (right_nritems)
3900 btrfs_mark_buffer_dirty(right);
3901 else
3902 btrfs_clean_tree_block(right);
3903
3904 btrfs_item_key(right, &disk_key, 0);
3905 fixup_low_keys(path, &disk_key, 1);
3906
3907 /* then fixup the leaf pointer in the path */
3908 if (path->slots[0] < push_items) {
3909 path->slots[0] += old_left_nritems;
3910 btrfs_tree_unlock(path->nodes[0]);
3911 free_extent_buffer(path->nodes[0]);
3912 path->nodes[0] = left;
3913 path->slots[1] -= 1;
3914 } else {
3915 btrfs_tree_unlock(left);
3916 free_extent_buffer(left);
3917 path->slots[0] -= push_items;
3918 }
3919 BUG_ON(path->slots[0] < 0);
3920 return ret;
3921out:
3922 btrfs_tree_unlock(left);
3923 free_extent_buffer(left);
3924 return ret;
3925}
3926
3927/*
3928 * push some data in the path leaf to the left, trying to free up at
3929 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3930 *
3931 * max_slot can put a limit on how far into the leaf we'll push items. The
3932 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
3933 * items
3934 */
3935static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3936 *root, struct btrfs_path *path, int min_data_size,
3937 int data_size, int empty, u32 max_slot)
3938{
3939 struct extent_buffer *right = path->nodes[0];
3940 struct extent_buffer *left;
3941 int slot;
3942 int free_space;
3943 u32 right_nritems;
3944 int ret = 0;
3945
3946 slot = path->slots[1];
3947 if (slot == 0)
3948 return 1;
3949 if (!path->nodes[1])
3950 return 1;
3951
3952 right_nritems = btrfs_header_nritems(right);
3953 if (right_nritems == 0)
3954 return 1;
3955
3956 btrfs_assert_tree_locked(path->nodes[1]);
3957
3958 left = btrfs_read_node_slot(path->nodes[1], slot - 1);
3959 /*
3960 * slot - 1 is not valid or we fail to read the left node,
3961 * no big deal, just return.
3962 */
3963 if (IS_ERR(left))
3964 return 1;
3965
3966 btrfs_tree_lock(left);
3967 btrfs_set_lock_blocking_write(left);
3968
3969 free_space = btrfs_leaf_free_space(left);
3970 if (free_space < data_size) {
3971 ret = 1;
3972 goto out;
3973 }
3974
3975 /* cow and double check */
3976 ret = btrfs_cow_block(trans, root, left,
3977 path->nodes[1], slot - 1, &left);
3978 if (ret) {
3979 /* we hit -ENOSPC, but it isn't fatal here */
3980 if (ret == -ENOSPC)
3981 ret = 1;
3982 goto out;
3983 }
3984
3985 free_space = btrfs_leaf_free_space(left);
3986 if (free_space < data_size) {
3987 ret = 1;
3988 goto out;
3989 }
3990
3991 return __push_leaf_left(path, min_data_size,
3992 empty, left, free_space, right_nritems,
3993 max_slot);
3994out:
3995 btrfs_tree_unlock(left);
3996 free_extent_buffer(left);
3997 return ret;
3998}
3999
4000/*
4001 * split the path's leaf in two, making sure there is at least data_size
4002 * available for the resulting leaf level of the path.
4003 */
4004static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4005 struct btrfs_path *path,
4006 struct extent_buffer *l,
4007 struct extent_buffer *right,
4008 int slot, int mid, int nritems)
4009{
4010 struct btrfs_fs_info *fs_info = trans->fs_info;
4011 int data_copy_size;
4012 int rt_data_off;
4013 int i;
4014 struct btrfs_disk_key disk_key;
4015 struct btrfs_map_token token;
4016
4017 nritems = nritems - mid;
4018 btrfs_set_header_nritems(right, nritems);
4019 data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(l);
4020
4021 copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4022 btrfs_item_nr_offset(mid),
4023 nritems * sizeof(struct btrfs_item));
4024
4025 copy_extent_buffer(right, l,
4026 BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
4027 data_copy_size, BTRFS_LEAF_DATA_OFFSET +
4028 leaf_data_end(l), data_copy_size);
4029
4030 rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4031
4032 btrfs_init_map_token(&token, right);
4033 for (i = 0; i < nritems; i++) {
4034 struct btrfs_item *item = btrfs_item_nr(i);
4035 u32 ioff;
4036
4037 ioff = btrfs_token_item_offset(&token, item);
4038 btrfs_set_token_item_offset(&token, item, ioff + rt_data_off);
4039 }
4040
4041 btrfs_set_header_nritems(l, mid);
4042 btrfs_item_key(right, &disk_key, 0);
4043 insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
4044
4045 btrfs_mark_buffer_dirty(right);
4046 btrfs_mark_buffer_dirty(l);
4047 BUG_ON(path->slots[0] != slot);
4048
4049 if (mid <= slot) {
4050 btrfs_tree_unlock(path->nodes[0]);
4051 free_extent_buffer(path->nodes[0]);
4052 path->nodes[0] = right;
4053 path->slots[0] -= mid;
4054 path->slots[1] += 1;
4055 } else {
4056 btrfs_tree_unlock(right);
4057 free_extent_buffer(right);
4058 }
4059
4060 BUG_ON(path->slots[0] < 0);
4061}
4062
4063/*
4064 * double splits happen when we need to insert a big item in the middle
4065 * of a leaf. A double split can leave us with 3 mostly empty leaves:
4066 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4067 * A B C
4068 *
4069 * We avoid this by trying to push the items on either side of our target
4070 * into the adjacent leaves. If all goes well we can avoid the double split
4071 * completely.
4072 */
4073static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4074 struct btrfs_root *root,
4075 struct btrfs_path *path,
4076 int data_size)
4077{
4078 int ret;
4079 int progress = 0;
4080 int slot;
4081 u32 nritems;
4082 int space_needed = data_size;
4083
4084 slot = path->slots[0];
4085 if (slot < btrfs_header_nritems(path->nodes[0]))
4086 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4087
4088 /*
4089 * try to push all the items after our slot into the
4090 * right leaf
4091 */
4092 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4093 if (ret < 0)
4094 return ret;
4095
4096 if (ret == 0)
4097 progress++;
4098
4099 nritems = btrfs_header_nritems(path->nodes[0]);
4100 /*
4101 * our goal is to get our slot at the start or end of a leaf. If
4102 * we've done so we're done
4103 */
4104 if (path->slots[0] == 0 || path->slots[0] == nritems)
4105 return 0;
4106
4107 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4108 return 0;
4109
4110 /* try to push all the items before our slot into the next leaf */
4111 slot = path->slots[0];
4112 space_needed = data_size;
4113 if (slot > 0)
4114 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4115 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4116 if (ret < 0)
4117 return ret;
4118
4119 if (ret == 0)
4120 progress++;
4121
4122 if (progress)
4123 return 0;
4124 return 1;
4125}
4126
4127/*
4128 * split the path's leaf in two, making sure there is at least data_size
4129 * available for the resulting leaf level of the path.
4130 *
4131 * returns 0 if all went well and < 0 on failure.
4132 */
4133static noinline int split_leaf(struct btrfs_trans_handle *trans,
4134 struct btrfs_root *root,
4135 const struct btrfs_key *ins_key,
4136 struct btrfs_path *path, int data_size,
4137 int extend)
4138{
4139 struct btrfs_disk_key disk_key;
4140 struct extent_buffer *l;
4141 u32 nritems;
4142 int mid;
4143 int slot;
4144 struct extent_buffer *right;
4145 struct btrfs_fs_info *fs_info = root->fs_info;
4146 int ret = 0;
4147 int wret;
4148 int split;
4149 int num_doubles = 0;
4150 int tried_avoid_double = 0;
4151
4152 l = path->nodes[0];
4153 slot = path->slots[0];
4154 if (extend && data_size + btrfs_item_size_nr(l, slot) +
4155 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4156 return -EOVERFLOW;
4157
4158 /* first try to make some room by pushing left and right */
4159 if (data_size && path->nodes[1]) {
4160 int space_needed = data_size;
4161
4162 if (slot < btrfs_header_nritems(l))
4163 space_needed -= btrfs_leaf_free_space(l);
4164
4165 wret = push_leaf_right(trans, root, path, space_needed,
4166 space_needed, 0, 0);
4167 if (wret < 0)
4168 return wret;
4169 if (wret) {
4170 space_needed = data_size;
4171 if (slot > 0)
4172 space_needed -= btrfs_leaf_free_space(l);
4173 wret = push_leaf_left(trans, root, path, space_needed,
4174 space_needed, 0, (u32)-1);
4175 if (wret < 0)
4176 return wret;
4177 }
4178 l = path->nodes[0];
4179
4180 /* did the pushes work? */
4181 if (btrfs_leaf_free_space(l) >= data_size)
4182 return 0;
4183 }
4184
4185 if (!path->nodes[1]) {
4186 ret = insert_new_root(trans, root, path, 1);
4187 if (ret)
4188 return ret;
4189 }
4190again:
4191 split = 1;
4192 l = path->nodes[0];
4193 slot = path->slots[0];
4194 nritems = btrfs_header_nritems(l);
4195 mid = (nritems + 1) / 2;
4196
4197 if (mid <= slot) {
4198 if (nritems == 1 ||
4199 leaf_space_used(l, mid, nritems - mid) + data_size >
4200 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4201 if (slot >= nritems) {
4202 split = 0;
4203 } else {
4204 mid = slot;
4205 if (mid != nritems &&
4206 leaf_space_used(l, mid, nritems - mid) +
4207 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4208 if (data_size && !tried_avoid_double)
4209 goto push_for_double;
4210 split = 2;
4211 }
4212 }
4213 }
4214 } else {
4215 if (leaf_space_used(l, 0, mid) + data_size >
4216 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4217 if (!extend && data_size && slot == 0) {
4218 split = 0;
4219 } else if ((extend || !data_size) && slot == 0) {
4220 mid = 1;
4221 } else {
4222 mid = slot;
4223 if (mid != nritems &&
4224 leaf_space_used(l, mid, nritems - mid) +
4225 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4226 if (data_size && !tried_avoid_double)
4227 goto push_for_double;
4228 split = 2;
4229 }
4230 }
4231 }
4232 }
4233
4234 if (split == 0)
4235 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4236 else
4237 btrfs_item_key(l, &disk_key, mid);
4238
4239 right = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, 0,
4240 l->start, 0);
4241 if (IS_ERR(right))
4242 return PTR_ERR(right);
4243
4244 root_add_used(root, fs_info->nodesize);
4245
4246 if (split == 0) {
4247 if (mid <= slot) {
4248 btrfs_set_header_nritems(right, 0);
4249 insert_ptr(trans, path, &disk_key,
4250 right->start, path->slots[1] + 1, 1);
4251 btrfs_tree_unlock(path->nodes[0]);
4252 free_extent_buffer(path->nodes[0]);
4253 path->nodes[0] = right;
4254 path->slots[0] = 0;
4255 path->slots[1] += 1;
4256 } else {
4257 btrfs_set_header_nritems(right, 0);
4258 insert_ptr(trans, path, &disk_key,
4259 right->start, path->slots[1], 1);
4260 btrfs_tree_unlock(path->nodes[0]);
4261 free_extent_buffer(path->nodes[0]);
4262 path->nodes[0] = right;
4263 path->slots[0] = 0;
4264 if (path->slots[1] == 0)
4265 fixup_low_keys(path, &disk_key, 1);
4266 }
4267 /*
4268 * We create a new leaf 'right' for the required ins_len and
4269 * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4270 * the content of ins_len to 'right'.
4271 */
4272 return ret;
4273 }
4274
4275 copy_for_split(trans, path, l, right, slot, mid, nritems);
4276
4277 if (split == 2) {
4278 BUG_ON(num_doubles != 0);
4279 num_doubles++;
4280 goto again;
4281 }
4282
4283 return 0;
4284
4285push_for_double:
4286 push_for_double_split(trans, root, path, data_size);
4287 tried_avoid_double = 1;
4288 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4289 return 0;
4290 goto again;
4291}
4292
4293static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4294 struct btrfs_root *root,
4295 struct btrfs_path *path, int ins_len)
4296{
4297 struct btrfs_key key;
4298 struct extent_buffer *leaf;
4299 struct btrfs_file_extent_item *fi;
4300 u64 extent_len = 0;
4301 u32 item_size;
4302 int ret;
4303
4304 leaf = path->nodes[0];
4305 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4306
4307 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4308 key.type != BTRFS_EXTENT_CSUM_KEY);
4309
4310 if (btrfs_leaf_free_space(leaf) >= ins_len)
4311 return 0;
4312
4313 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4314 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4315 fi = btrfs_item_ptr(leaf, path->slots[0],
4316 struct btrfs_file_extent_item);
4317 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4318 }
4319 btrfs_release_path(path);
4320
4321 path->keep_locks = 1;
4322 path->search_for_split = 1;
4323 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4324 path->search_for_split = 0;
4325 if (ret > 0)
4326 ret = -EAGAIN;
4327 if (ret < 0)
4328 goto err;
4329
4330 ret = -EAGAIN;
4331 leaf = path->nodes[0];
4332 /* if our item isn't there, return now */
4333 if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4334 goto err;
4335
4336 /* the leaf has changed, it now has room. return now */
4337 if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
4338 goto err;
4339
4340 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4341 fi = btrfs_item_ptr(leaf, path->slots[0],
4342 struct btrfs_file_extent_item);
4343 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4344 goto err;
4345 }
4346
4347 btrfs_set_path_blocking(path);
4348 ret = split_leaf(trans, root, &key, path, ins_len, 1);
4349 if (ret)
4350 goto err;
4351
4352 path->keep_locks = 0;
4353 btrfs_unlock_up_safe(path, 1);
4354 return 0;
4355err:
4356 path->keep_locks = 0;
4357 return ret;
4358}
4359
4360static noinline int split_item(struct btrfs_path *path,
4361 const struct btrfs_key *new_key,
4362 unsigned long split_offset)
4363{
4364 struct extent_buffer *leaf;
4365 struct btrfs_item *item;
4366 struct btrfs_item *new_item;
4367 int slot;
4368 char *buf;
4369 u32 nritems;
4370 u32 item_size;
4371 u32 orig_offset;
4372 struct btrfs_disk_key disk_key;
4373
4374 leaf = path->nodes[0];
4375 BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
4376
4377 btrfs_set_path_blocking(path);
4378
4379 item = btrfs_item_nr(path->slots[0]);
4380 orig_offset = btrfs_item_offset(leaf, item);
4381 item_size = btrfs_item_size(leaf, item);
4382
4383 buf = kmalloc(item_size, GFP_NOFS);
4384 if (!buf)
4385 return -ENOMEM;
4386
4387 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4388 path->slots[0]), item_size);
4389
4390 slot = path->slots[0] + 1;
4391 nritems = btrfs_header_nritems(leaf);
4392 if (slot != nritems) {
4393 /* shift the items */
4394 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4395 btrfs_item_nr_offset(slot),
4396 (nritems - slot) * sizeof(struct btrfs_item));
4397 }
4398
4399 btrfs_cpu_key_to_disk(&disk_key, new_key);
4400 btrfs_set_item_key(leaf, &disk_key, slot);
4401
4402 new_item = btrfs_item_nr(slot);
4403
4404 btrfs_set_item_offset(leaf, new_item, orig_offset);
4405 btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4406
4407 btrfs_set_item_offset(leaf, item,
4408 orig_offset + item_size - split_offset);
4409 btrfs_set_item_size(leaf, item, split_offset);
4410
4411 btrfs_set_header_nritems(leaf, nritems + 1);
4412
4413 /* write the data for the start of the original item */
4414 write_extent_buffer(leaf, buf,
4415 btrfs_item_ptr_offset(leaf, path->slots[0]),
4416 split_offset);
4417
4418 /* write the data for the new item */
4419 write_extent_buffer(leaf, buf + split_offset,
4420 btrfs_item_ptr_offset(leaf, slot),
4421 item_size - split_offset);
4422 btrfs_mark_buffer_dirty(leaf);
4423
4424 BUG_ON(btrfs_leaf_free_space(leaf) < 0);
4425 kfree(buf);
4426 return 0;
4427}
4428
4429/*
4430 * This function splits a single item into two items,
4431 * giving 'new_key' to the new item and splitting the
4432 * old one at split_offset (from the start of the item).
4433 *
4434 * The path may be released by this operation. After
4435 * the split, the path is pointing to the old item. The
4436 * new item is going to be in the same node as the old one.
4437 *
4438 * Note, the item being split must be smaller enough to live alone on
4439 * a tree block with room for one extra struct btrfs_item
4440 *
4441 * This allows us to split the item in place, keeping a lock on the
4442 * leaf the entire time.
4443 */
4444int btrfs_split_item(struct btrfs_trans_handle *trans,
4445 struct btrfs_root *root,
4446 struct btrfs_path *path,
4447 const struct btrfs_key *new_key,
4448 unsigned long split_offset)
4449{
4450 int ret;
4451 ret = setup_leaf_for_split(trans, root, path,
4452 sizeof(struct btrfs_item));
4453 if (ret)
4454 return ret;
4455
4456 ret = split_item(path, new_key, split_offset);
4457 return ret;
4458}
4459
4460/*
4461 * This function duplicate a item, giving 'new_key' to the new item.
4462 * It guarantees both items live in the same tree leaf and the new item
4463 * is contiguous with the original item.
4464 *
4465 * This allows us to split file extent in place, keeping a lock on the
4466 * leaf the entire time.
4467 */
4468int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4469 struct btrfs_root *root,
4470 struct btrfs_path *path,
4471 const struct btrfs_key *new_key)
4472{
4473 struct extent_buffer *leaf;
4474 int ret;
4475 u32 item_size;
4476
4477 leaf = path->nodes[0];
4478 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4479 ret = setup_leaf_for_split(trans, root, path,
4480 item_size + sizeof(struct btrfs_item));
4481 if (ret)
4482 return ret;
4483
4484 path->slots[0]++;
4485 setup_items_for_insert(root, path, new_key, &item_size,
4486 item_size, item_size +
4487 sizeof(struct btrfs_item), 1);
4488 leaf = path->nodes[0];
4489 memcpy_extent_buffer(leaf,
4490 btrfs_item_ptr_offset(leaf, path->slots[0]),
4491 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4492 item_size);
4493 return 0;
4494}
4495
4496/*
4497 * make the item pointed to by the path smaller. new_size indicates
4498 * how small to make it, and from_end tells us if we just chop bytes
4499 * off the end of the item or if we shift the item to chop bytes off
4500 * the front.
4501 */
4502void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
4503{
4504 int slot;
4505 struct extent_buffer *leaf;
4506 struct btrfs_item *item;
4507 u32 nritems;
4508 unsigned int data_end;
4509 unsigned int old_data_start;
4510 unsigned int old_size;
4511 unsigned int size_diff;
4512 int i;
4513 struct btrfs_map_token token;
4514
4515 leaf = path->nodes[0];
4516 slot = path->slots[0];
4517
4518 old_size = btrfs_item_size_nr(leaf, slot);
4519 if (old_size == new_size)
4520 return;
4521
4522 nritems = btrfs_header_nritems(leaf);
4523 data_end = leaf_data_end(leaf);
4524
4525 old_data_start = btrfs_item_offset_nr(leaf, slot);
4526
4527 size_diff = old_size - new_size;
4528
4529 BUG_ON(slot < 0);
4530 BUG_ON(slot >= nritems);
4531
4532 /*
4533 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4534 */
4535 /* first correct the data pointers */
4536 btrfs_init_map_token(&token, leaf);
4537 for (i = slot; i < nritems; i++) {
4538 u32 ioff;
4539 item = btrfs_item_nr(i);
4540
4541 ioff = btrfs_token_item_offset(&token, item);
4542 btrfs_set_token_item_offset(&token, item, ioff + size_diff);
4543 }
4544
4545 /* shift the data */
4546 if (from_end) {
4547 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4548 data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4549 data_end, old_data_start + new_size - data_end);
4550 } else {
4551 struct btrfs_disk_key disk_key;
4552 u64 offset;
4553
4554 btrfs_item_key(leaf, &disk_key, slot);
4555
4556 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4557 unsigned long ptr;
4558 struct btrfs_file_extent_item *fi;
4559
4560 fi = btrfs_item_ptr(leaf, slot,
4561 struct btrfs_file_extent_item);
4562 fi = (struct btrfs_file_extent_item *)(
4563 (unsigned long)fi - size_diff);
4564
4565 if (btrfs_file_extent_type(leaf, fi) ==
4566 BTRFS_FILE_EXTENT_INLINE) {
4567 ptr = btrfs_item_ptr_offset(leaf, slot);
4568 memmove_extent_buffer(leaf, ptr,
4569 (unsigned long)fi,
4570 BTRFS_FILE_EXTENT_INLINE_DATA_START);
4571 }
4572 }
4573
4574 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4575 data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4576 data_end, old_data_start - data_end);
4577
4578 offset = btrfs_disk_key_offset(&disk_key);
4579 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4580 btrfs_set_item_key(leaf, &disk_key, slot);
4581 if (slot == 0)
4582 fixup_low_keys(path, &disk_key, 1);
4583 }
4584
4585 item = btrfs_item_nr(slot);
4586 btrfs_set_item_size(leaf, item, new_size);
4587 btrfs_mark_buffer_dirty(leaf);
4588
4589 if (btrfs_leaf_free_space(leaf) < 0) {
4590 btrfs_print_leaf(leaf);
4591 BUG();
4592 }
4593}
4594
4595/*
4596 * make the item pointed to by the path bigger, data_size is the added size.
4597 */
4598void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
4599{
4600 int slot;
4601 struct extent_buffer *leaf;
4602 struct btrfs_item *item;
4603 u32 nritems;
4604 unsigned int data_end;
4605 unsigned int old_data;
4606 unsigned int old_size;
4607 int i;
4608 struct btrfs_map_token token;
4609
4610 leaf = path->nodes[0];
4611
4612 nritems = btrfs_header_nritems(leaf);
4613 data_end = leaf_data_end(leaf);
4614
4615 if (btrfs_leaf_free_space(leaf) < data_size) {
4616 btrfs_print_leaf(leaf);
4617 BUG();
4618 }
4619 slot = path->slots[0];
4620 old_data = btrfs_item_end_nr(leaf, slot);
4621
4622 BUG_ON(slot < 0);
4623 if (slot >= nritems) {
4624 btrfs_print_leaf(leaf);
4625 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
4626 slot, nritems);
4627 BUG();
4628 }
4629
4630 /*
4631 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4632 */
4633 /* first correct the data pointers */
4634 btrfs_init_map_token(&token, leaf);
4635 for (i = slot; i < nritems; i++) {
4636 u32 ioff;
4637 item = btrfs_item_nr(i);
4638
4639 ioff = btrfs_token_item_offset(&token, item);
4640 btrfs_set_token_item_offset(&token, item, ioff - data_size);
4641 }
4642
4643 /* shift the data */
4644 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4645 data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
4646 data_end, old_data - data_end);
4647
4648 data_end = old_data;
4649 old_size = btrfs_item_size_nr(leaf, slot);
4650 item = btrfs_item_nr(slot);
4651 btrfs_set_item_size(leaf, item, old_size + data_size);
4652 btrfs_mark_buffer_dirty(leaf);
4653
4654 if (btrfs_leaf_free_space(leaf) < 0) {
4655 btrfs_print_leaf(leaf);
4656 BUG();
4657 }
4658}
4659
4660/*
4661 * this is a helper for btrfs_insert_empty_items, the main goal here is
4662 * to save stack depth by doing the bulk of the work in a function
4663 * that doesn't call btrfs_search_slot
4664 */
4665void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4666 const struct btrfs_key *cpu_key, u32 *data_size,
4667 u32 total_data, u32 total_size, int nr)
4668{
4669 struct btrfs_fs_info *fs_info = root->fs_info;
4670 struct btrfs_item *item;
4671 int i;
4672 u32 nritems;
4673 unsigned int data_end;
4674 struct btrfs_disk_key disk_key;
4675 struct extent_buffer *leaf;
4676 int slot;
4677 struct btrfs_map_token token;
4678
4679 if (path->slots[0] == 0) {
4680 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4681 fixup_low_keys(path, &disk_key, 1);
4682 }
4683 btrfs_unlock_up_safe(path, 1);
4684
4685 leaf = path->nodes[0];
4686 slot = path->slots[0];
4687
4688 nritems = btrfs_header_nritems(leaf);
4689 data_end = leaf_data_end(leaf);
4690
4691 if (btrfs_leaf_free_space(leaf) < total_size) {
4692 btrfs_print_leaf(leaf);
4693 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4694 total_size, btrfs_leaf_free_space(leaf));
4695 BUG();
4696 }
4697
4698 btrfs_init_map_token(&token, leaf);
4699 if (slot != nritems) {
4700 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4701
4702 if (old_data < data_end) {
4703 btrfs_print_leaf(leaf);
4704 btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4705 slot, old_data, data_end);
4706 BUG();
4707 }
4708 /*
4709 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4710 */
4711 /* first correct the data pointers */
4712 for (i = slot; i < nritems; i++) {
4713 u32 ioff;
4714
4715 item = btrfs_item_nr(i);
4716 ioff = btrfs_token_item_offset(&token, item);
4717 btrfs_set_token_item_offset(&token, item,
4718 ioff - total_data);
4719 }
4720 /* shift the items */
4721 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4722 btrfs_item_nr_offset(slot),
4723 (nritems - slot) * sizeof(struct btrfs_item));
4724
4725 /* shift the data */
4726 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4727 data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
4728 data_end, old_data - data_end);
4729 data_end = old_data;
4730 }
4731
4732 /* setup the item for the new data */
4733 for (i = 0; i < nr; i++) {
4734 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4735 btrfs_set_item_key(leaf, &disk_key, slot + i);
4736 item = btrfs_item_nr(slot + i);
4737 btrfs_set_token_item_offset(&token, item, data_end - data_size[i]);
4738 data_end -= data_size[i];
4739 btrfs_set_token_item_size(&token, item, data_size[i]);
4740 }
4741
4742 btrfs_set_header_nritems(leaf, nritems + nr);
4743 btrfs_mark_buffer_dirty(leaf);
4744
4745 if (btrfs_leaf_free_space(leaf) < 0) {
4746 btrfs_print_leaf(leaf);
4747 BUG();
4748 }
4749}
4750
4751/*
4752 * Given a key and some data, insert items into the tree.
4753 * This does all the path init required, making room in the tree if needed.
4754 */
4755int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4756 struct btrfs_root *root,
4757 struct btrfs_path *path,
4758 const struct btrfs_key *cpu_key, u32 *data_size,
4759 int nr)
4760{
4761 int ret = 0;
4762 int slot;
4763 int i;
4764 u32 total_size = 0;
4765 u32 total_data = 0;
4766
4767 for (i = 0; i < nr; i++)
4768 total_data += data_size[i];
4769
4770 total_size = total_data + (nr * sizeof(struct btrfs_item));
4771 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4772 if (ret == 0)
4773 return -EEXIST;
4774 if (ret < 0)
4775 return ret;
4776
4777 slot = path->slots[0];
4778 BUG_ON(slot < 0);
4779
4780 setup_items_for_insert(root, path, cpu_key, data_size,
4781 total_data, total_size, nr);
4782 return 0;
4783}
4784
4785/*
4786 * Given a key and some data, insert an item into the tree.
4787 * This does all the path init required, making room in the tree if needed.
4788 */
4789int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4790 const struct btrfs_key *cpu_key, void *data,
4791 u32 data_size)
4792{
4793 int ret = 0;
4794 struct btrfs_path *path;
4795 struct extent_buffer *leaf;
4796 unsigned long ptr;
4797
4798 path = btrfs_alloc_path();
4799 if (!path)
4800 return -ENOMEM;
4801 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4802 if (!ret) {
4803 leaf = path->nodes[0];
4804 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4805 write_extent_buffer(leaf, data, ptr, data_size);
4806 btrfs_mark_buffer_dirty(leaf);
4807 }
4808 btrfs_free_path(path);
4809 return ret;
4810}
4811
4812/*
4813 * delete the pointer from a given node.
4814 *
4815 * the tree should have been previously balanced so the deletion does not
4816 * empty a node.
4817 */
4818static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4819 int level, int slot)
4820{
4821 struct extent_buffer *parent = path->nodes[level];
4822 u32 nritems;
4823 int ret;
4824
4825 nritems = btrfs_header_nritems(parent);
4826 if (slot != nritems - 1) {
4827 if (level) {
4828 ret = tree_mod_log_insert_move(parent, slot, slot + 1,
4829 nritems - slot - 1);
4830 BUG_ON(ret < 0);
4831 }
4832 memmove_extent_buffer(parent,
4833 btrfs_node_key_ptr_offset(slot),
4834 btrfs_node_key_ptr_offset(slot + 1),
4835 sizeof(struct btrfs_key_ptr) *
4836 (nritems - slot - 1));
4837 } else if (level) {
4838 ret = tree_mod_log_insert_key(parent, slot, MOD_LOG_KEY_REMOVE,
4839 GFP_NOFS);
4840 BUG_ON(ret < 0);
4841 }
4842
4843 nritems--;
4844 btrfs_set_header_nritems(parent, nritems);
4845 if (nritems == 0 && parent == root->node) {
4846 BUG_ON(btrfs_header_level(root->node) != 1);
4847 /* just turn the root into a leaf and break */
4848 btrfs_set_header_level(root->node, 0);
4849 } else if (slot == 0) {
4850 struct btrfs_disk_key disk_key;
4851
4852 btrfs_node_key(parent, &disk_key, 0);
4853 fixup_low_keys(path, &disk_key, level + 1);
4854 }
4855 btrfs_mark_buffer_dirty(parent);
4856}
4857
4858/*
4859 * a helper function to delete the leaf pointed to by path->slots[1] and
4860 * path->nodes[1].
4861 *
4862 * This deletes the pointer in path->nodes[1] and frees the leaf
4863 * block extent. zero is returned if it all worked out, < 0 otherwise.
4864 *
4865 * The path must have already been setup for deleting the leaf, including
4866 * all the proper balancing. path->nodes[1] must be locked.
4867 */
4868static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4869 struct btrfs_root *root,
4870 struct btrfs_path *path,
4871 struct extent_buffer *leaf)
4872{
4873 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4874 del_ptr(root, path, 1, path->slots[1]);
4875
4876 /*
4877 * btrfs_free_extent is expensive, we want to make sure we
4878 * aren't holding any locks when we call it
4879 */
4880 btrfs_unlock_up_safe(path, 0);
4881
4882 root_sub_used(root, leaf->len);
4883
4884 atomic_inc(&leaf->refs);
4885 btrfs_free_tree_block(trans, root, leaf, 0, 1);
4886 free_extent_buffer_stale(leaf);
4887}
4888/*
4889 * delete the item at the leaf level in path. If that empties
4890 * the leaf, remove it from the tree
4891 */
4892int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4893 struct btrfs_path *path, int slot, int nr)
4894{
4895 struct btrfs_fs_info *fs_info = root->fs_info;
4896 struct extent_buffer *leaf;
4897 struct btrfs_item *item;
4898 u32 last_off;
4899 u32 dsize = 0;
4900 int ret = 0;
4901 int wret;
4902 int i;
4903 u32 nritems;
4904
4905 leaf = path->nodes[0];
4906 last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4907
4908 for (i = 0; i < nr; i++)
4909 dsize += btrfs_item_size_nr(leaf, slot + i);
4910
4911 nritems = btrfs_header_nritems(leaf);
4912
4913 if (slot + nr != nritems) {
4914 int data_end = leaf_data_end(leaf);
4915 struct btrfs_map_token token;
4916
4917 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4918 data_end + dsize,
4919 BTRFS_LEAF_DATA_OFFSET + data_end,
4920 last_off - data_end);
4921
4922 btrfs_init_map_token(&token, leaf);
4923 for (i = slot + nr; i < nritems; i++) {
4924 u32 ioff;
4925
4926 item = btrfs_item_nr(i);
4927 ioff = btrfs_token_item_offset(&token, item);
4928 btrfs_set_token_item_offset(&token, item, ioff + dsize);
4929 }
4930
4931 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4932 btrfs_item_nr_offset(slot + nr),
4933 sizeof(struct btrfs_item) *
4934 (nritems - slot - nr));
4935 }
4936 btrfs_set_header_nritems(leaf, nritems - nr);
4937 nritems -= nr;
4938
4939 /* delete the leaf if we've emptied it */
4940 if (nritems == 0) {
4941 if (leaf == root->node) {
4942 btrfs_set_header_level(leaf, 0);
4943 } else {
4944 btrfs_set_path_blocking(path);
4945 btrfs_clean_tree_block(leaf);
4946 btrfs_del_leaf(trans, root, path, leaf);
4947 }
4948 } else {
4949 int used = leaf_space_used(leaf, 0, nritems);
4950 if (slot == 0) {
4951 struct btrfs_disk_key disk_key;
4952
4953 btrfs_item_key(leaf, &disk_key, 0);
4954 fixup_low_keys(path, &disk_key, 1);
4955 }
4956
4957 /* delete the leaf if it is mostly empty */
4958 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4959 /* push_leaf_left fixes the path.
4960 * make sure the path still points to our leaf
4961 * for possible call to del_ptr below
4962 */
4963 slot = path->slots[1];
4964 atomic_inc(&leaf->refs);
4965
4966 btrfs_set_path_blocking(path);
4967 wret = push_leaf_left(trans, root, path, 1, 1,
4968 1, (u32)-1);
4969 if (wret < 0 && wret != -ENOSPC)
4970 ret = wret;
4971
4972 if (path->nodes[0] == leaf &&
4973 btrfs_header_nritems(leaf)) {
4974 wret = push_leaf_right(trans, root, path, 1,
4975 1, 1, 0);
4976 if (wret < 0 && wret != -ENOSPC)
4977 ret = wret;
4978 }
4979
4980 if (btrfs_header_nritems(leaf) == 0) {
4981 path->slots[1] = slot;
4982 btrfs_del_leaf(trans, root, path, leaf);
4983 free_extent_buffer(leaf);
4984 ret = 0;
4985 } else {
4986 /* if we're still in the path, make sure
4987 * we're dirty. Otherwise, one of the
4988 * push_leaf functions must have already
4989 * dirtied this buffer
4990 */
4991 if (path->nodes[0] == leaf)
4992 btrfs_mark_buffer_dirty(leaf);
4993 free_extent_buffer(leaf);
4994 }
4995 } else {
4996 btrfs_mark_buffer_dirty(leaf);
4997 }
4998 }
4999 return ret;
5000}
5001
5002/*
5003 * search the tree again to find a leaf with lesser keys
5004 * returns 0 if it found something or 1 if there are no lesser leaves.
5005 * returns < 0 on io errors.
5006 *
5007 * This may release the path, and so you may lose any locks held at the
5008 * time you call it.
5009 */
5010int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5011{
5012 struct btrfs_key key;
5013 struct btrfs_disk_key found_key;
5014 int ret;
5015
5016 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5017
5018 if (key.offset > 0) {
5019 key.offset--;
5020 } else if (key.type > 0) {
5021 key.type--;
5022 key.offset = (u64)-1;
5023 } else if (key.objectid > 0) {
5024 key.objectid--;
5025 key.type = (u8)-1;
5026 key.offset = (u64)-1;
5027 } else {
5028 return 1;
5029 }
5030
5031 btrfs_release_path(path);
5032 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5033 if (ret < 0)
5034 return ret;
5035 btrfs_item_key(path->nodes[0], &found_key, 0);
5036 ret = comp_keys(&found_key, &key);
5037 /*
5038 * We might have had an item with the previous key in the tree right
5039 * before we released our path. And after we released our path, that
5040 * item might have been pushed to the first slot (0) of the leaf we
5041 * were holding due to a tree balance. Alternatively, an item with the
5042 * previous key can exist as the only element of a leaf (big fat item).
5043 * Therefore account for these 2 cases, so that our callers (like
5044 * btrfs_previous_item) don't miss an existing item with a key matching
5045 * the previous key we computed above.
5046 */
5047 if (ret <= 0)
5048 return 0;
5049 return 1;
5050}
5051
5052/*
5053 * A helper function to walk down the tree starting at min_key, and looking
5054 * for nodes or leaves that are have a minimum transaction id.
5055 * This is used by the btree defrag code, and tree logging
5056 *
5057 * This does not cow, but it does stuff the starting key it finds back
5058 * into min_key, so you can call btrfs_search_slot with cow=1 on the
5059 * key and get a writable path.
5060 *
5061 * This honors path->lowest_level to prevent descent past a given level
5062 * of the tree.
5063 *
5064 * min_trans indicates the oldest transaction that you are interested
5065 * in walking through. Any nodes or leaves older than min_trans are
5066 * skipped over (without reading them).
5067 *
5068 * returns zero if something useful was found, < 0 on error and 1 if there
5069 * was nothing in the tree that matched the search criteria.
5070 */
5071int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5072 struct btrfs_path *path,
5073 u64 min_trans)
5074{
5075 struct extent_buffer *cur;
5076 struct btrfs_key found_key;
5077 int slot;
5078 int sret;
5079 u32 nritems;
5080 int level;
5081 int ret = 1;
5082 int keep_locks = path->keep_locks;
5083
5084 path->keep_locks = 1;
5085again:
5086 cur = btrfs_read_lock_root_node(root);
5087 level = btrfs_header_level(cur);
5088 WARN_ON(path->nodes[level]);
5089 path->nodes[level] = cur;
5090 path->locks[level] = BTRFS_READ_LOCK;
5091
5092 if (btrfs_header_generation(cur) < min_trans) {
5093 ret = 1;
5094 goto out;
5095 }
5096 while (1) {
5097 nritems = btrfs_header_nritems(cur);
5098 level = btrfs_header_level(cur);
5099 sret = btrfs_bin_search(cur, min_key, &slot);
5100 if (sret < 0) {
5101 ret = sret;
5102 goto out;
5103 }
5104
5105 /* at the lowest level, we're done, setup the path and exit */
5106 if (level == path->lowest_level) {
5107 if (slot >= nritems)
5108 goto find_next_key;
5109 ret = 0;
5110 path->slots[level] = slot;
5111 btrfs_item_key_to_cpu(cur, &found_key, slot);
5112 goto out;
5113 }
5114 if (sret && slot > 0)
5115 slot--;
5116 /*
5117 * check this node pointer against the min_trans parameters.
5118 * If it is too old, old, skip to the next one.
5119 */
5120 while (slot < nritems) {
5121 u64 gen;
5122
5123 gen = btrfs_node_ptr_generation(cur, slot);
5124 if (gen < min_trans) {
5125 slot++;
5126 continue;
5127 }
5128 break;
5129 }
5130find_next_key:
5131 /*
5132 * we didn't find a candidate key in this node, walk forward
5133 * and find another one
5134 */
5135 if (slot >= nritems) {
5136 path->slots[level] = slot;
5137 btrfs_set_path_blocking(path);
5138 sret = btrfs_find_next_key(root, path, min_key, level,
5139 min_trans);
5140 if (sret == 0) {
5141 btrfs_release_path(path);
5142 goto again;
5143 } else {
5144 goto out;
5145 }
5146 }
5147 /* save our key for returning back */
5148 btrfs_node_key_to_cpu(cur, &found_key, slot);
5149 path->slots[level] = slot;
5150 if (level == path->lowest_level) {
5151 ret = 0;
5152 goto out;
5153 }
5154 btrfs_set_path_blocking(path);
5155 cur = btrfs_read_node_slot(cur, slot);
5156 if (IS_ERR(cur)) {
5157 ret = PTR_ERR(cur);
5158 goto out;
5159 }
5160
5161 btrfs_tree_read_lock(cur);
5162
5163 path->locks[level - 1] = BTRFS_READ_LOCK;
5164 path->nodes[level - 1] = cur;
5165 unlock_up(path, level, 1, 0, NULL);
5166 }
5167out:
5168 path->keep_locks = keep_locks;
5169 if (ret == 0) {
5170 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5171 btrfs_set_path_blocking(path);
5172 memcpy(min_key, &found_key, sizeof(found_key));
5173 }
5174 return ret;
5175}
5176
5177/*
5178 * this is similar to btrfs_next_leaf, but does not try to preserve
5179 * and fixup the path. It looks for and returns the next key in the
5180 * tree based on the current path and the min_trans parameters.
5181 *
5182 * 0 is returned if another key is found, < 0 if there are any errors
5183 * and 1 is returned if there are no higher keys in the tree
5184 *
5185 * path->keep_locks should be set to 1 on the search made before
5186 * calling this function.
5187 */
5188int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5189 struct btrfs_key *key, int level, u64 min_trans)
5190{
5191 int slot;
5192 struct extent_buffer *c;
5193
5194 WARN_ON(!path->keep_locks && !path->skip_locking);
5195 while (level < BTRFS_MAX_LEVEL) {
5196 if (!path->nodes[level])
5197 return 1;
5198
5199 slot = path->slots[level] + 1;
5200 c = path->nodes[level];
5201next:
5202 if (slot >= btrfs_header_nritems(c)) {
5203 int ret;
5204 int orig_lowest;
5205 struct btrfs_key cur_key;
5206 if (level + 1 >= BTRFS_MAX_LEVEL ||
5207 !path->nodes[level + 1])
5208 return 1;
5209
5210 if (path->locks[level + 1] || path->skip_locking) {
5211 level++;
5212 continue;
5213 }
5214
5215 slot = btrfs_header_nritems(c) - 1;
5216 if (level == 0)
5217 btrfs_item_key_to_cpu(c, &cur_key, slot);
5218 else
5219 btrfs_node_key_to_cpu(c, &cur_key, slot);
5220
5221 orig_lowest = path->lowest_level;
5222 btrfs_release_path(path);
5223 path->lowest_level = level;
5224 ret = btrfs_search_slot(NULL, root, &cur_key, path,
5225 0, 0);
5226 path->lowest_level = orig_lowest;
5227 if (ret < 0)
5228 return ret;
5229
5230 c = path->nodes[level];
5231 slot = path->slots[level];
5232 if (ret == 0)
5233 slot++;
5234 goto next;
5235 }
5236
5237 if (level == 0)
5238 btrfs_item_key_to_cpu(c, key, slot);
5239 else {
5240 u64 gen = btrfs_node_ptr_generation(c, slot);
5241
5242 if (gen < min_trans) {
5243 slot++;
5244 goto next;
5245 }
5246 btrfs_node_key_to_cpu(c, key, slot);
5247 }
5248 return 0;
5249 }
5250 return 1;
5251}
5252
5253/*
5254 * search the tree again to find a leaf with greater keys
5255 * returns 0 if it found something or 1 if there are no greater leaves.
5256 * returns < 0 on io errors.
5257 */
5258int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5259{
5260 return btrfs_next_old_leaf(root, path, 0);
5261}
5262
5263int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5264 u64 time_seq)
5265{
5266 int slot;
5267 int level;
5268 struct extent_buffer *c;
5269 struct extent_buffer *next;
5270 struct btrfs_key key;
5271 u32 nritems;
5272 int ret;
5273 int old_spinning = path->leave_spinning;
5274 int next_rw_lock = 0;
5275
5276 nritems = btrfs_header_nritems(path->nodes[0]);
5277 if (nritems == 0)
5278 return 1;
5279
5280 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5281again:
5282 level = 1;
5283 next = NULL;
5284 next_rw_lock = 0;
5285 btrfs_release_path(path);
5286
5287 path->keep_locks = 1;
5288 path->leave_spinning = 1;
5289
5290 if (time_seq)
5291 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5292 else
5293 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5294 path->keep_locks = 0;
5295
5296 if (ret < 0)
5297 return ret;
5298
5299 nritems = btrfs_header_nritems(path->nodes[0]);
5300 /*
5301 * by releasing the path above we dropped all our locks. A balance
5302 * could have added more items next to the key that used to be
5303 * at the very end of the block. So, check again here and
5304 * advance the path if there are now more items available.
5305 */
5306 if (nritems > 0 && path->slots[0] < nritems - 1) {
5307 if (ret == 0)
5308 path->slots[0]++;
5309 ret = 0;
5310 goto done;
5311 }
5312 /*
5313 * So the above check misses one case:
5314 * - after releasing the path above, someone has removed the item that
5315 * used to be at the very end of the block, and balance between leafs
5316 * gets another one with bigger key.offset to replace it.
5317 *
5318 * This one should be returned as well, or we can get leaf corruption
5319 * later(esp. in __btrfs_drop_extents()).
5320 *
5321 * And a bit more explanation about this check,
5322 * with ret > 0, the key isn't found, the path points to the slot
5323 * where it should be inserted, so the path->slots[0] item must be the
5324 * bigger one.
5325 */
5326 if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5327 ret = 0;
5328 goto done;
5329 }
5330
5331 while (level < BTRFS_MAX_LEVEL) {
5332 if (!path->nodes[level]) {
5333 ret = 1;
5334 goto done;
5335 }
5336
5337 slot = path->slots[level] + 1;
5338 c = path->nodes[level];
5339 if (slot >= btrfs_header_nritems(c)) {
5340 level++;
5341 if (level == BTRFS_MAX_LEVEL) {
5342 ret = 1;
5343 goto done;
5344 }
5345 continue;
5346 }
5347
5348 if (next) {
5349 btrfs_tree_unlock_rw(next, next_rw_lock);
5350 free_extent_buffer(next);
5351 }
5352
5353 next = c;
5354 next_rw_lock = path->locks[level];
5355 ret = read_block_for_search(root, path, &next, level,
5356 slot, &key);
5357 if (ret == -EAGAIN)
5358 goto again;
5359
5360 if (ret < 0) {
5361 btrfs_release_path(path);
5362 goto done;
5363 }
5364
5365 if (!path->skip_locking) {
5366 ret = btrfs_try_tree_read_lock(next);
5367 if (!ret && time_seq) {
5368 /*
5369 * If we don't get the lock, we may be racing
5370 * with push_leaf_left, holding that lock while
5371 * itself waiting for the leaf we've currently
5372 * locked. To solve this situation, we give up
5373 * on our lock and cycle.
5374 */
5375 free_extent_buffer(next);
5376 btrfs_release_path(path);
5377 cond_resched();
5378 goto again;
5379 }
5380 if (!ret) {
5381 btrfs_set_path_blocking(path);
5382 btrfs_tree_read_lock(next);
5383 }
5384 next_rw_lock = BTRFS_READ_LOCK;
5385 }
5386 break;
5387 }
5388 path->slots[level] = slot;
5389 while (1) {
5390 level--;
5391 c = path->nodes[level];
5392 if (path->locks[level])
5393 btrfs_tree_unlock_rw(c, path->locks[level]);
5394
5395 free_extent_buffer(c);
5396 path->nodes[level] = next;
5397 path->slots[level] = 0;
5398 if (!path->skip_locking)
5399 path->locks[level] = next_rw_lock;
5400 if (!level)
5401 break;
5402
5403 ret = read_block_for_search(root, path, &next, level,
5404 0, &key);
5405 if (ret == -EAGAIN)
5406 goto again;
5407
5408 if (ret < 0) {
5409 btrfs_release_path(path);
5410 goto done;
5411 }
5412
5413 if (!path->skip_locking) {
5414 ret = btrfs_try_tree_read_lock(next);
5415 if (!ret) {
5416 btrfs_set_path_blocking(path);
5417 btrfs_tree_read_lock(next);
5418 }
5419 next_rw_lock = BTRFS_READ_LOCK;
5420 }
5421 }
5422 ret = 0;
5423done:
5424 unlock_up(path, 0, 1, 0, NULL);
5425 path->leave_spinning = old_spinning;
5426 if (!old_spinning)
5427 btrfs_set_path_blocking(path);
5428
5429 return ret;
5430}
5431
5432/*
5433 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5434 * searching until it gets past min_objectid or finds an item of 'type'
5435 *
5436 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5437 */
5438int btrfs_previous_item(struct btrfs_root *root,
5439 struct btrfs_path *path, u64 min_objectid,
5440 int type)
5441{
5442 struct btrfs_key found_key;
5443 struct extent_buffer *leaf;
5444 u32 nritems;
5445 int ret;
5446
5447 while (1) {
5448 if (path->slots[0] == 0) {
5449 btrfs_set_path_blocking(path);
5450 ret = btrfs_prev_leaf(root, path);
5451 if (ret != 0)
5452 return ret;
5453 } else {
5454 path->slots[0]--;
5455 }
5456 leaf = path->nodes[0];
5457 nritems = btrfs_header_nritems(leaf);
5458 if (nritems == 0)
5459 return 1;
5460 if (path->slots[0] == nritems)
5461 path->slots[0]--;
5462
5463 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5464 if (found_key.objectid < min_objectid)
5465 break;
5466 if (found_key.type == type)
5467 return 0;
5468 if (found_key.objectid == min_objectid &&
5469 found_key.type < type)
5470 break;
5471 }
5472 return 1;
5473}
5474
5475/*
5476 * search in extent tree to find a previous Metadata/Data extent item with
5477 * min objecitd.
5478 *
5479 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5480 */
5481int btrfs_previous_extent_item(struct btrfs_root *root,
5482 struct btrfs_path *path, u64 min_objectid)
5483{
5484 struct btrfs_key found_key;
5485 struct extent_buffer *leaf;
5486 u32 nritems;
5487 int ret;
5488
5489 while (1) {
5490 if (path->slots[0] == 0) {
5491 btrfs_set_path_blocking(path);
5492 ret = btrfs_prev_leaf(root, path);
5493 if (ret != 0)
5494 return ret;
5495 } else {
5496 path->slots[0]--;
5497 }
5498 leaf = path->nodes[0];
5499 nritems = btrfs_header_nritems(leaf);
5500 if (nritems == 0)
5501 return 1;
5502 if (path->slots[0] == nritems)
5503 path->slots[0]--;
5504
5505 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5506 if (found_key.objectid < min_objectid)
5507 break;
5508 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5509 found_key.type == BTRFS_METADATA_ITEM_KEY)
5510 return 0;
5511 if (found_key.objectid == min_objectid &&
5512 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5513 break;
5514 }
5515 return 1;
5516}