Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007,2008 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/rbtree.h>
9#include <linux/mm.h>
10#include <linux/error-injection.h>
11#include "messages.h"
12#include "ctree.h"
13#include "disk-io.h"
14#include "transaction.h"
15#include "print-tree.h"
16#include "locking.h"
17#include "volumes.h"
18#include "qgroup.h"
19#include "tree-mod-log.h"
20#include "tree-checker.h"
21#include "fs.h"
22#include "accessors.h"
23#include "extent-tree.h"
24#include "relocation.h"
25#include "file-item.h"
26
27static struct kmem_cache *btrfs_path_cachep;
28
29static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
30 *root, struct btrfs_path *path, int level);
31static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
32 const struct btrfs_key *ins_key, struct btrfs_path *path,
33 int data_size, int extend);
34static int push_node_left(struct btrfs_trans_handle *trans,
35 struct extent_buffer *dst,
36 struct extent_buffer *src, int empty);
37static int balance_node_right(struct btrfs_trans_handle *trans,
38 struct extent_buffer *dst_buf,
39 struct extent_buffer *src_buf);
40static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
41 int level, int slot);
42
43static const struct btrfs_csums {
44 u16 size;
45 const char name[10];
46 const char driver[12];
47} btrfs_csums[] = {
48 [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
49 [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
50 [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
51 [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
52 .driver = "blake2b-256" },
53};
54
55/*
56 * The leaf data grows from end-to-front in the node. this returns the address
57 * of the start of the last item, which is the stop of the leaf data stack.
58 */
59static unsigned int leaf_data_end(const struct extent_buffer *leaf)
60{
61 u32 nr = btrfs_header_nritems(leaf);
62
63 if (nr == 0)
64 return BTRFS_LEAF_DATA_SIZE(leaf->fs_info);
65 return btrfs_item_offset(leaf, nr - 1);
66}
67
68/*
69 * Move data in a @leaf (using memmove, safe for overlapping ranges).
70 *
71 * @leaf: leaf that we're doing a memmove on
72 * @dst_offset: item data offset we're moving to
73 * @src_offset: item data offset were' moving from
74 * @len: length of the data we're moving
75 *
76 * Wrapper around memmove_extent_buffer() that takes into account the header on
77 * the leaf. The btrfs_item offset's start directly after the header, so we
78 * have to adjust any offsets to account for the header in the leaf. This
79 * handles that math to simplify the callers.
80 */
81static inline void memmove_leaf_data(const struct extent_buffer *leaf,
82 unsigned long dst_offset,
83 unsigned long src_offset,
84 unsigned long len)
85{
86 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, 0) + dst_offset,
87 btrfs_item_nr_offset(leaf, 0) + src_offset, len);
88}
89
90/*
91 * Copy item data from @src into @dst at the given @offset.
92 *
93 * @dst: destination leaf that we're copying into
94 * @src: source leaf that we're copying from
95 * @dst_offset: item data offset we're copying to
96 * @src_offset: item data offset were' copying from
97 * @len: length of the data we're copying
98 *
99 * Wrapper around copy_extent_buffer() that takes into account the header on
100 * the leaf. The btrfs_item offset's start directly after the header, so we
101 * have to adjust any offsets to account for the header in the leaf. This
102 * handles that math to simplify the callers.
103 */
104static inline void copy_leaf_data(const struct extent_buffer *dst,
105 const struct extent_buffer *src,
106 unsigned long dst_offset,
107 unsigned long src_offset, unsigned long len)
108{
109 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, 0) + dst_offset,
110 btrfs_item_nr_offset(src, 0) + src_offset, len);
111}
112
113/*
114 * Move items in a @leaf (using memmove).
115 *
116 * @dst: destination leaf for the items
117 * @dst_item: the item nr we're copying into
118 * @src_item: the item nr we're copying from
119 * @nr_items: the number of items to copy
120 *
121 * Wrapper around memmove_extent_buffer() that does the math to get the
122 * appropriate offsets into the leaf from the item numbers.
123 */
124static inline void memmove_leaf_items(const struct extent_buffer *leaf,
125 int dst_item, int src_item, int nr_items)
126{
127 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, dst_item),
128 btrfs_item_nr_offset(leaf, src_item),
129 nr_items * sizeof(struct btrfs_item));
130}
131
132/*
133 * Copy items from @src into @dst at the given @offset.
134 *
135 * @dst: destination leaf for the items
136 * @src: source leaf for the items
137 * @dst_item: the item nr we're copying into
138 * @src_item: the item nr we're copying from
139 * @nr_items: the number of items to copy
140 *
141 * Wrapper around copy_extent_buffer() that does the math to get the
142 * appropriate offsets into the leaf from the item numbers.
143 */
144static inline void copy_leaf_items(const struct extent_buffer *dst,
145 const struct extent_buffer *src,
146 int dst_item, int src_item, int nr_items)
147{
148 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, dst_item),
149 btrfs_item_nr_offset(src, src_item),
150 nr_items * sizeof(struct btrfs_item));
151}
152
153int btrfs_super_csum_size(const struct btrfs_super_block *s)
154{
155 u16 t = btrfs_super_csum_type(s);
156 /*
157 * csum type is validated at mount time
158 */
159 return btrfs_csums[t].size;
160}
161
162const char *btrfs_super_csum_name(u16 csum_type)
163{
164 /* csum type is validated at mount time */
165 return btrfs_csums[csum_type].name;
166}
167
168/*
169 * Return driver name if defined, otherwise the name that's also a valid driver
170 * name
171 */
172const char *btrfs_super_csum_driver(u16 csum_type)
173{
174 /* csum type is validated at mount time */
175 return btrfs_csums[csum_type].driver[0] ?
176 btrfs_csums[csum_type].driver :
177 btrfs_csums[csum_type].name;
178}
179
180size_t __attribute_const__ btrfs_get_num_csums(void)
181{
182 return ARRAY_SIZE(btrfs_csums);
183}
184
185struct btrfs_path *btrfs_alloc_path(void)
186{
187 might_sleep();
188
189 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
190}
191
192/* this also releases the path */
193void btrfs_free_path(struct btrfs_path *p)
194{
195 if (!p)
196 return;
197 btrfs_release_path(p);
198 kmem_cache_free(btrfs_path_cachep, p);
199}
200
201/*
202 * path release drops references on the extent buffers in the path
203 * and it drops any locks held by this path
204 *
205 * It is safe to call this on paths that no locks or extent buffers held.
206 */
207noinline void btrfs_release_path(struct btrfs_path *p)
208{
209 int i;
210
211 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
212 p->slots[i] = 0;
213 if (!p->nodes[i])
214 continue;
215 if (p->locks[i]) {
216 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
217 p->locks[i] = 0;
218 }
219 free_extent_buffer(p->nodes[i]);
220 p->nodes[i] = NULL;
221 }
222}
223
224/*
225 * We want the transaction abort to print stack trace only for errors where the
226 * cause could be a bug, eg. due to ENOSPC, and not for common errors that are
227 * caused by external factors.
228 */
229bool __cold abort_should_print_stack(int errno)
230{
231 switch (errno) {
232 case -EIO:
233 case -EROFS:
234 case -ENOMEM:
235 return false;
236 }
237 return true;
238}
239
240/*
241 * safely gets a reference on the root node of a tree. A lock
242 * is not taken, so a concurrent writer may put a different node
243 * at the root of the tree. See btrfs_lock_root_node for the
244 * looping required.
245 *
246 * The extent buffer returned by this has a reference taken, so
247 * it won't disappear. It may stop being the root of the tree
248 * at any time because there are no locks held.
249 */
250struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
251{
252 struct extent_buffer *eb;
253
254 while (1) {
255 rcu_read_lock();
256 eb = rcu_dereference(root->node);
257
258 /*
259 * RCU really hurts here, we could free up the root node because
260 * it was COWed but we may not get the new root node yet so do
261 * the inc_not_zero dance and if it doesn't work then
262 * synchronize_rcu and try again.
263 */
264 if (atomic_inc_not_zero(&eb->refs)) {
265 rcu_read_unlock();
266 break;
267 }
268 rcu_read_unlock();
269 synchronize_rcu();
270 }
271 return eb;
272}
273
274/*
275 * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
276 * just get put onto a simple dirty list. Transaction walks this list to make
277 * sure they get properly updated on disk.
278 */
279static void add_root_to_dirty_list(struct btrfs_root *root)
280{
281 struct btrfs_fs_info *fs_info = root->fs_info;
282
283 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
284 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
285 return;
286
287 spin_lock(&fs_info->trans_lock);
288 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
289 /* Want the extent tree to be the last on the list */
290 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
291 list_move_tail(&root->dirty_list,
292 &fs_info->dirty_cowonly_roots);
293 else
294 list_move(&root->dirty_list,
295 &fs_info->dirty_cowonly_roots);
296 }
297 spin_unlock(&fs_info->trans_lock);
298}
299
300/*
301 * used by snapshot creation to make a copy of a root for a tree with
302 * a given objectid. The buffer with the new root node is returned in
303 * cow_ret, and this func returns zero on success or a negative error code.
304 */
305int btrfs_copy_root(struct btrfs_trans_handle *trans,
306 struct btrfs_root *root,
307 struct extent_buffer *buf,
308 struct extent_buffer **cow_ret, u64 new_root_objectid)
309{
310 struct btrfs_fs_info *fs_info = root->fs_info;
311 struct extent_buffer *cow;
312 int ret = 0;
313 int level;
314 struct btrfs_disk_key disk_key;
315
316 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
317 trans->transid != fs_info->running_transaction->transid);
318 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
319 trans->transid != root->last_trans);
320
321 level = btrfs_header_level(buf);
322 if (level == 0)
323 btrfs_item_key(buf, &disk_key, 0);
324 else
325 btrfs_node_key(buf, &disk_key, 0);
326
327 cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
328 &disk_key, level, buf->start, 0,
329 BTRFS_NESTING_NEW_ROOT);
330 if (IS_ERR(cow))
331 return PTR_ERR(cow);
332
333 copy_extent_buffer_full(cow, buf);
334 btrfs_set_header_bytenr(cow, cow->start);
335 btrfs_set_header_generation(cow, trans->transid);
336 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
337 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
338 BTRFS_HEADER_FLAG_RELOC);
339 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
340 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
341 else
342 btrfs_set_header_owner(cow, new_root_objectid);
343
344 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
345
346 WARN_ON(btrfs_header_generation(buf) > trans->transid);
347 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
348 ret = btrfs_inc_ref(trans, root, cow, 1);
349 else
350 ret = btrfs_inc_ref(trans, root, cow, 0);
351 if (ret) {
352 btrfs_tree_unlock(cow);
353 free_extent_buffer(cow);
354 btrfs_abort_transaction(trans, ret);
355 return ret;
356 }
357
358 btrfs_mark_buffer_dirty(cow);
359 *cow_ret = cow;
360 return 0;
361}
362
363/*
364 * check if the tree block can be shared by multiple trees
365 */
366int btrfs_block_can_be_shared(struct btrfs_root *root,
367 struct extent_buffer *buf)
368{
369 /*
370 * Tree blocks not in shareable trees and tree roots are never shared.
371 * If a block was allocated after the last snapshot and the block was
372 * not allocated by tree relocation, we know the block is not shared.
373 */
374 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
375 buf != root->node && buf != root->commit_root &&
376 (btrfs_header_generation(buf) <=
377 btrfs_root_last_snapshot(&root->root_item) ||
378 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
379 return 1;
380
381 return 0;
382}
383
384static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
385 struct btrfs_root *root,
386 struct extent_buffer *buf,
387 struct extent_buffer *cow,
388 int *last_ref)
389{
390 struct btrfs_fs_info *fs_info = root->fs_info;
391 u64 refs;
392 u64 owner;
393 u64 flags;
394 u64 new_flags = 0;
395 int ret;
396
397 /*
398 * Backrefs update rules:
399 *
400 * Always use full backrefs for extent pointers in tree block
401 * allocated by tree relocation.
402 *
403 * If a shared tree block is no longer referenced by its owner
404 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
405 * use full backrefs for extent pointers in tree block.
406 *
407 * If a tree block is been relocating
408 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
409 * use full backrefs for extent pointers in tree block.
410 * The reason for this is some operations (such as drop tree)
411 * are only allowed for blocks use full backrefs.
412 */
413
414 if (btrfs_block_can_be_shared(root, buf)) {
415 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
416 btrfs_header_level(buf), 1,
417 &refs, &flags);
418 if (ret)
419 return ret;
420 if (refs == 0) {
421 ret = -EROFS;
422 btrfs_handle_fs_error(fs_info, ret, NULL);
423 return ret;
424 }
425 } else {
426 refs = 1;
427 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
428 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
429 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
430 else
431 flags = 0;
432 }
433
434 owner = btrfs_header_owner(buf);
435 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
436 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
437
438 if (refs > 1) {
439 if ((owner == root->root_key.objectid ||
440 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
441 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
442 ret = btrfs_inc_ref(trans, root, buf, 1);
443 if (ret)
444 return ret;
445
446 if (root->root_key.objectid ==
447 BTRFS_TREE_RELOC_OBJECTID) {
448 ret = btrfs_dec_ref(trans, root, buf, 0);
449 if (ret)
450 return ret;
451 ret = btrfs_inc_ref(trans, root, cow, 1);
452 if (ret)
453 return ret;
454 }
455 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
456 } else {
457
458 if (root->root_key.objectid ==
459 BTRFS_TREE_RELOC_OBJECTID)
460 ret = btrfs_inc_ref(trans, root, cow, 1);
461 else
462 ret = btrfs_inc_ref(trans, root, cow, 0);
463 if (ret)
464 return ret;
465 }
466 if (new_flags != 0) {
467 int level = btrfs_header_level(buf);
468
469 ret = btrfs_set_disk_extent_flags(trans, buf,
470 new_flags, level);
471 if (ret)
472 return ret;
473 }
474 } else {
475 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
476 if (root->root_key.objectid ==
477 BTRFS_TREE_RELOC_OBJECTID)
478 ret = btrfs_inc_ref(trans, root, cow, 1);
479 else
480 ret = btrfs_inc_ref(trans, root, cow, 0);
481 if (ret)
482 return ret;
483 ret = btrfs_dec_ref(trans, root, buf, 1);
484 if (ret)
485 return ret;
486 }
487 btrfs_clean_tree_block(buf);
488 *last_ref = 1;
489 }
490 return 0;
491}
492
493/*
494 * does the dirty work in cow of a single block. The parent block (if
495 * supplied) is updated to point to the new cow copy. The new buffer is marked
496 * dirty and returned locked. If you modify the block it needs to be marked
497 * dirty again.
498 *
499 * search_start -- an allocation hint for the new block
500 *
501 * empty_size -- a hint that you plan on doing more cow. This is the size in
502 * bytes the allocator should try to find free next to the block it returns.
503 * This is just a hint and may be ignored by the allocator.
504 */
505static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
506 struct btrfs_root *root,
507 struct extent_buffer *buf,
508 struct extent_buffer *parent, int parent_slot,
509 struct extent_buffer **cow_ret,
510 u64 search_start, u64 empty_size,
511 enum btrfs_lock_nesting nest)
512{
513 struct btrfs_fs_info *fs_info = root->fs_info;
514 struct btrfs_disk_key disk_key;
515 struct extent_buffer *cow;
516 int level, ret;
517 int last_ref = 0;
518 int unlock_orig = 0;
519 u64 parent_start = 0;
520
521 if (*cow_ret == buf)
522 unlock_orig = 1;
523
524 btrfs_assert_tree_write_locked(buf);
525
526 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
527 trans->transid != fs_info->running_transaction->transid);
528 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
529 trans->transid != root->last_trans);
530
531 level = btrfs_header_level(buf);
532
533 if (level == 0)
534 btrfs_item_key(buf, &disk_key, 0);
535 else
536 btrfs_node_key(buf, &disk_key, 0);
537
538 if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
539 parent_start = parent->start;
540
541 cow = btrfs_alloc_tree_block(trans, root, parent_start,
542 root->root_key.objectid, &disk_key, level,
543 search_start, empty_size, nest);
544 if (IS_ERR(cow))
545 return PTR_ERR(cow);
546
547 /* cow is set to blocking by btrfs_init_new_buffer */
548
549 copy_extent_buffer_full(cow, buf);
550 btrfs_set_header_bytenr(cow, cow->start);
551 btrfs_set_header_generation(cow, trans->transid);
552 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
553 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
554 BTRFS_HEADER_FLAG_RELOC);
555 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
556 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
557 else
558 btrfs_set_header_owner(cow, root->root_key.objectid);
559
560 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
561
562 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
563 if (ret) {
564 btrfs_tree_unlock(cow);
565 free_extent_buffer(cow);
566 btrfs_abort_transaction(trans, ret);
567 return ret;
568 }
569
570 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
571 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
572 if (ret) {
573 btrfs_tree_unlock(cow);
574 free_extent_buffer(cow);
575 btrfs_abort_transaction(trans, ret);
576 return ret;
577 }
578 }
579
580 if (buf == root->node) {
581 WARN_ON(parent && parent != buf);
582 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
583 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
584 parent_start = buf->start;
585
586 atomic_inc(&cow->refs);
587 ret = btrfs_tree_mod_log_insert_root(root->node, cow, true);
588 BUG_ON(ret < 0);
589 rcu_assign_pointer(root->node, cow);
590
591 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
592 parent_start, last_ref);
593 free_extent_buffer(buf);
594 add_root_to_dirty_list(root);
595 } else {
596 WARN_ON(trans->transid != btrfs_header_generation(parent));
597 btrfs_tree_mod_log_insert_key(parent, parent_slot,
598 BTRFS_MOD_LOG_KEY_REPLACE);
599 btrfs_set_node_blockptr(parent, parent_slot,
600 cow->start);
601 btrfs_set_node_ptr_generation(parent, parent_slot,
602 trans->transid);
603 btrfs_mark_buffer_dirty(parent);
604 if (last_ref) {
605 ret = btrfs_tree_mod_log_free_eb(buf);
606 if (ret) {
607 btrfs_tree_unlock(cow);
608 free_extent_buffer(cow);
609 btrfs_abort_transaction(trans, ret);
610 return ret;
611 }
612 }
613 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
614 parent_start, last_ref);
615 }
616 if (unlock_orig)
617 btrfs_tree_unlock(buf);
618 free_extent_buffer_stale(buf);
619 btrfs_mark_buffer_dirty(cow);
620 *cow_ret = cow;
621 return 0;
622}
623
624static inline int should_cow_block(struct btrfs_trans_handle *trans,
625 struct btrfs_root *root,
626 struct extent_buffer *buf)
627{
628 if (btrfs_is_testing(root->fs_info))
629 return 0;
630
631 /* Ensure we can see the FORCE_COW bit */
632 smp_mb__before_atomic();
633
634 /*
635 * We do not need to cow a block if
636 * 1) this block is not created or changed in this transaction;
637 * 2) this block does not belong to TREE_RELOC tree;
638 * 3) the root is not forced COW.
639 *
640 * What is forced COW:
641 * when we create snapshot during committing the transaction,
642 * after we've finished copying src root, we must COW the shared
643 * block to ensure the metadata consistency.
644 */
645 if (btrfs_header_generation(buf) == trans->transid &&
646 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
647 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
648 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
649 !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
650 return 0;
651 return 1;
652}
653
654/*
655 * cows a single block, see __btrfs_cow_block for the real work.
656 * This version of it has extra checks so that a block isn't COWed more than
657 * once per transaction, as long as it hasn't been written yet
658 */
659noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
660 struct btrfs_root *root, struct extent_buffer *buf,
661 struct extent_buffer *parent, int parent_slot,
662 struct extent_buffer **cow_ret,
663 enum btrfs_lock_nesting nest)
664{
665 struct btrfs_fs_info *fs_info = root->fs_info;
666 u64 search_start;
667 int ret;
668
669 if (test_bit(BTRFS_ROOT_DELETING, &root->state))
670 btrfs_err(fs_info,
671 "COW'ing blocks on a fs root that's being dropped");
672
673 if (trans->transaction != fs_info->running_transaction)
674 WARN(1, KERN_CRIT "trans %llu running %llu\n",
675 trans->transid,
676 fs_info->running_transaction->transid);
677
678 if (trans->transid != fs_info->generation)
679 WARN(1, KERN_CRIT "trans %llu running %llu\n",
680 trans->transid, fs_info->generation);
681
682 if (!should_cow_block(trans, root, buf)) {
683 *cow_ret = buf;
684 return 0;
685 }
686
687 search_start = buf->start & ~((u64)SZ_1G - 1);
688
689 /*
690 * Before CoWing this block for later modification, check if it's
691 * the subtree root and do the delayed subtree trace if needed.
692 *
693 * Also We don't care about the error, as it's handled internally.
694 */
695 btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
696 ret = __btrfs_cow_block(trans, root, buf, parent,
697 parent_slot, cow_ret, search_start, 0, nest);
698
699 trace_btrfs_cow_block(root, buf, *cow_ret);
700
701 return ret;
702}
703ALLOW_ERROR_INJECTION(btrfs_cow_block, ERRNO);
704
705/*
706 * helper function for defrag to decide if two blocks pointed to by a
707 * node are actually close by
708 */
709static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
710{
711 if (blocknr < other && other - (blocknr + blocksize) < 32768)
712 return 1;
713 if (blocknr > other && blocknr - (other + blocksize) < 32768)
714 return 1;
715 return 0;
716}
717
718#ifdef __LITTLE_ENDIAN
719
720/*
721 * Compare two keys, on little-endian the disk order is same as CPU order and
722 * we can avoid the conversion.
723 */
724static int comp_keys(const struct btrfs_disk_key *disk_key,
725 const struct btrfs_key *k2)
726{
727 const struct btrfs_key *k1 = (const struct btrfs_key *)disk_key;
728
729 return btrfs_comp_cpu_keys(k1, k2);
730}
731
732#else
733
734/*
735 * compare two keys in a memcmp fashion
736 */
737static int comp_keys(const struct btrfs_disk_key *disk,
738 const struct btrfs_key *k2)
739{
740 struct btrfs_key k1;
741
742 btrfs_disk_key_to_cpu(&k1, disk);
743
744 return btrfs_comp_cpu_keys(&k1, k2);
745}
746#endif
747
748/*
749 * same as comp_keys only with two btrfs_key's
750 */
751int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
752{
753 if (k1->objectid > k2->objectid)
754 return 1;
755 if (k1->objectid < k2->objectid)
756 return -1;
757 if (k1->type > k2->type)
758 return 1;
759 if (k1->type < k2->type)
760 return -1;
761 if (k1->offset > k2->offset)
762 return 1;
763 if (k1->offset < k2->offset)
764 return -1;
765 return 0;
766}
767
768/*
769 * this is used by the defrag code to go through all the
770 * leaves pointed to by a node and reallocate them so that
771 * disk order is close to key order
772 */
773int btrfs_realloc_node(struct btrfs_trans_handle *trans,
774 struct btrfs_root *root, struct extent_buffer *parent,
775 int start_slot, u64 *last_ret,
776 struct btrfs_key *progress)
777{
778 struct btrfs_fs_info *fs_info = root->fs_info;
779 struct extent_buffer *cur;
780 u64 blocknr;
781 u64 search_start = *last_ret;
782 u64 last_block = 0;
783 u64 other;
784 u32 parent_nritems;
785 int end_slot;
786 int i;
787 int err = 0;
788 u32 blocksize;
789 int progress_passed = 0;
790 struct btrfs_disk_key disk_key;
791
792 WARN_ON(trans->transaction != fs_info->running_transaction);
793 WARN_ON(trans->transid != fs_info->generation);
794
795 parent_nritems = btrfs_header_nritems(parent);
796 blocksize = fs_info->nodesize;
797 end_slot = parent_nritems - 1;
798
799 if (parent_nritems <= 1)
800 return 0;
801
802 for (i = start_slot; i <= end_slot; i++) {
803 int close = 1;
804
805 btrfs_node_key(parent, &disk_key, i);
806 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
807 continue;
808
809 progress_passed = 1;
810 blocknr = btrfs_node_blockptr(parent, i);
811 if (last_block == 0)
812 last_block = blocknr;
813
814 if (i > 0) {
815 other = btrfs_node_blockptr(parent, i - 1);
816 close = close_blocks(blocknr, other, blocksize);
817 }
818 if (!close && i < end_slot) {
819 other = btrfs_node_blockptr(parent, i + 1);
820 close = close_blocks(blocknr, other, blocksize);
821 }
822 if (close) {
823 last_block = blocknr;
824 continue;
825 }
826
827 cur = btrfs_read_node_slot(parent, i);
828 if (IS_ERR(cur))
829 return PTR_ERR(cur);
830 if (search_start == 0)
831 search_start = last_block;
832
833 btrfs_tree_lock(cur);
834 err = __btrfs_cow_block(trans, root, cur, parent, i,
835 &cur, search_start,
836 min(16 * blocksize,
837 (end_slot - i) * blocksize),
838 BTRFS_NESTING_COW);
839 if (err) {
840 btrfs_tree_unlock(cur);
841 free_extent_buffer(cur);
842 break;
843 }
844 search_start = cur->start;
845 last_block = cur->start;
846 *last_ret = search_start;
847 btrfs_tree_unlock(cur);
848 free_extent_buffer(cur);
849 }
850 return err;
851}
852
853/*
854 * Search for a key in the given extent_buffer.
855 *
856 * The lower boundary for the search is specified by the slot number @low. Use a
857 * value of 0 to search over the whole extent buffer.
858 *
859 * The slot in the extent buffer is returned via @slot. If the key exists in the
860 * extent buffer, then @slot will point to the slot where the key is, otherwise
861 * it points to the slot where you would insert the key.
862 *
863 * Slot may point to the total number of items (i.e. one position beyond the last
864 * key) if the key is bigger than the last key in the extent buffer.
865 */
866static noinline int generic_bin_search(struct extent_buffer *eb, int low,
867 const struct btrfs_key *key, int *slot)
868{
869 unsigned long p;
870 int item_size;
871 int high = btrfs_header_nritems(eb);
872 int ret;
873 const int key_size = sizeof(struct btrfs_disk_key);
874
875 if (low > high) {
876 btrfs_err(eb->fs_info,
877 "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
878 __func__, low, high, eb->start,
879 btrfs_header_owner(eb), btrfs_header_level(eb));
880 return -EINVAL;
881 }
882
883 if (btrfs_header_level(eb) == 0) {
884 p = offsetof(struct btrfs_leaf, items);
885 item_size = sizeof(struct btrfs_item);
886 } else {
887 p = offsetof(struct btrfs_node, ptrs);
888 item_size = sizeof(struct btrfs_key_ptr);
889 }
890
891 while (low < high) {
892 unsigned long oip;
893 unsigned long offset;
894 struct btrfs_disk_key *tmp;
895 struct btrfs_disk_key unaligned;
896 int mid;
897
898 mid = (low + high) / 2;
899 offset = p + mid * item_size;
900 oip = offset_in_page(offset);
901
902 if (oip + key_size <= PAGE_SIZE) {
903 const unsigned long idx = get_eb_page_index(offset);
904 char *kaddr = page_address(eb->pages[idx]);
905
906 oip = get_eb_offset_in_page(eb, offset);
907 tmp = (struct btrfs_disk_key *)(kaddr + oip);
908 } else {
909 read_extent_buffer(eb, &unaligned, offset, key_size);
910 tmp = &unaligned;
911 }
912
913 ret = comp_keys(tmp, key);
914
915 if (ret < 0)
916 low = mid + 1;
917 else if (ret > 0)
918 high = mid;
919 else {
920 *slot = mid;
921 return 0;
922 }
923 }
924 *slot = low;
925 return 1;
926}
927
928/*
929 * Simple binary search on an extent buffer. Works for both leaves and nodes, and
930 * always searches over the whole range of keys (slot 0 to slot 'nritems - 1').
931 */
932int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
933 int *slot)
934{
935 return generic_bin_search(eb, 0, key, slot);
936}
937
938static void root_add_used(struct btrfs_root *root, u32 size)
939{
940 spin_lock(&root->accounting_lock);
941 btrfs_set_root_used(&root->root_item,
942 btrfs_root_used(&root->root_item) + size);
943 spin_unlock(&root->accounting_lock);
944}
945
946static void root_sub_used(struct btrfs_root *root, u32 size)
947{
948 spin_lock(&root->accounting_lock);
949 btrfs_set_root_used(&root->root_item,
950 btrfs_root_used(&root->root_item) - size);
951 spin_unlock(&root->accounting_lock);
952}
953
954/* given a node and slot number, this reads the blocks it points to. The
955 * extent buffer is returned with a reference taken (but unlocked).
956 */
957struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
958 int slot)
959{
960 int level = btrfs_header_level(parent);
961 struct btrfs_tree_parent_check check = { 0 };
962 struct extent_buffer *eb;
963
964 if (slot < 0 || slot >= btrfs_header_nritems(parent))
965 return ERR_PTR(-ENOENT);
966
967 BUG_ON(level == 0);
968
969 check.level = level - 1;
970 check.transid = btrfs_node_ptr_generation(parent, slot);
971 check.owner_root = btrfs_header_owner(parent);
972 check.has_first_key = true;
973 btrfs_node_key_to_cpu(parent, &check.first_key, slot);
974
975 eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
976 &check);
977 if (IS_ERR(eb))
978 return eb;
979 if (!extent_buffer_uptodate(eb)) {
980 free_extent_buffer(eb);
981 return ERR_PTR(-EIO);
982 }
983
984 return eb;
985}
986
987/*
988 * node level balancing, used to make sure nodes are in proper order for
989 * item deletion. We balance from the top down, so we have to make sure
990 * that a deletion won't leave an node completely empty later on.
991 */
992static noinline int balance_level(struct btrfs_trans_handle *trans,
993 struct btrfs_root *root,
994 struct btrfs_path *path, int level)
995{
996 struct btrfs_fs_info *fs_info = root->fs_info;
997 struct extent_buffer *right = NULL;
998 struct extent_buffer *mid;
999 struct extent_buffer *left = NULL;
1000 struct extent_buffer *parent = NULL;
1001 int ret = 0;
1002 int wret;
1003 int pslot;
1004 int orig_slot = path->slots[level];
1005 u64 orig_ptr;
1006
1007 ASSERT(level > 0);
1008
1009 mid = path->nodes[level];
1010
1011 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK);
1012 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1013
1014 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1015
1016 if (level < BTRFS_MAX_LEVEL - 1) {
1017 parent = path->nodes[level + 1];
1018 pslot = path->slots[level + 1];
1019 }
1020
1021 /*
1022 * deal with the case where there is only one pointer in the root
1023 * by promoting the node below to a root
1024 */
1025 if (!parent) {
1026 struct extent_buffer *child;
1027
1028 if (btrfs_header_nritems(mid) != 1)
1029 return 0;
1030
1031 /* promote the child to a root */
1032 child = btrfs_read_node_slot(mid, 0);
1033 if (IS_ERR(child)) {
1034 ret = PTR_ERR(child);
1035 btrfs_handle_fs_error(fs_info, ret, NULL);
1036 goto enospc;
1037 }
1038
1039 btrfs_tree_lock(child);
1040 ret = btrfs_cow_block(trans, root, child, mid, 0, &child,
1041 BTRFS_NESTING_COW);
1042 if (ret) {
1043 btrfs_tree_unlock(child);
1044 free_extent_buffer(child);
1045 goto enospc;
1046 }
1047
1048 ret = btrfs_tree_mod_log_insert_root(root->node, child, true);
1049 BUG_ON(ret < 0);
1050 rcu_assign_pointer(root->node, child);
1051
1052 add_root_to_dirty_list(root);
1053 btrfs_tree_unlock(child);
1054
1055 path->locks[level] = 0;
1056 path->nodes[level] = NULL;
1057 btrfs_clean_tree_block(mid);
1058 btrfs_tree_unlock(mid);
1059 /* once for the path */
1060 free_extent_buffer(mid);
1061
1062 root_sub_used(root, mid->len);
1063 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
1064 /* once for the root ptr */
1065 free_extent_buffer_stale(mid);
1066 return 0;
1067 }
1068 if (btrfs_header_nritems(mid) >
1069 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1070 return 0;
1071
1072 left = btrfs_read_node_slot(parent, pslot - 1);
1073 if (IS_ERR(left))
1074 left = NULL;
1075
1076 if (left) {
1077 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
1078 wret = btrfs_cow_block(trans, root, left,
1079 parent, pslot - 1, &left,
1080 BTRFS_NESTING_LEFT_COW);
1081 if (wret) {
1082 ret = wret;
1083 goto enospc;
1084 }
1085 }
1086
1087 right = btrfs_read_node_slot(parent, pslot + 1);
1088 if (IS_ERR(right))
1089 right = NULL;
1090
1091 if (right) {
1092 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1093 wret = btrfs_cow_block(trans, root, right,
1094 parent, pslot + 1, &right,
1095 BTRFS_NESTING_RIGHT_COW);
1096 if (wret) {
1097 ret = wret;
1098 goto enospc;
1099 }
1100 }
1101
1102 /* first, try to make some room in the middle buffer */
1103 if (left) {
1104 orig_slot += btrfs_header_nritems(left);
1105 wret = push_node_left(trans, left, mid, 1);
1106 if (wret < 0)
1107 ret = wret;
1108 }
1109
1110 /*
1111 * then try to empty the right most buffer into the middle
1112 */
1113 if (right) {
1114 wret = push_node_left(trans, mid, right, 1);
1115 if (wret < 0 && wret != -ENOSPC)
1116 ret = wret;
1117 if (btrfs_header_nritems(right) == 0) {
1118 btrfs_clean_tree_block(right);
1119 btrfs_tree_unlock(right);
1120 del_ptr(root, path, level + 1, pslot + 1);
1121 root_sub_used(root, right->len);
1122 btrfs_free_tree_block(trans, btrfs_root_id(root), right,
1123 0, 1);
1124 free_extent_buffer_stale(right);
1125 right = NULL;
1126 } else {
1127 struct btrfs_disk_key right_key;
1128 btrfs_node_key(right, &right_key, 0);
1129 ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1130 BTRFS_MOD_LOG_KEY_REPLACE);
1131 BUG_ON(ret < 0);
1132 btrfs_set_node_key(parent, &right_key, pslot + 1);
1133 btrfs_mark_buffer_dirty(parent);
1134 }
1135 }
1136 if (btrfs_header_nritems(mid) == 1) {
1137 /*
1138 * we're not allowed to leave a node with one item in the
1139 * tree during a delete. A deletion from lower in the tree
1140 * could try to delete the only pointer in this node.
1141 * So, pull some keys from the left.
1142 * There has to be a left pointer at this point because
1143 * otherwise we would have pulled some pointers from the
1144 * right
1145 */
1146 if (!left) {
1147 ret = -EROFS;
1148 btrfs_handle_fs_error(fs_info, ret, NULL);
1149 goto enospc;
1150 }
1151 wret = balance_node_right(trans, mid, left);
1152 if (wret < 0) {
1153 ret = wret;
1154 goto enospc;
1155 }
1156 if (wret == 1) {
1157 wret = push_node_left(trans, left, mid, 1);
1158 if (wret < 0)
1159 ret = wret;
1160 }
1161 BUG_ON(wret == 1);
1162 }
1163 if (btrfs_header_nritems(mid) == 0) {
1164 btrfs_clean_tree_block(mid);
1165 btrfs_tree_unlock(mid);
1166 del_ptr(root, path, level + 1, pslot);
1167 root_sub_used(root, mid->len);
1168 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
1169 free_extent_buffer_stale(mid);
1170 mid = NULL;
1171 } else {
1172 /* update the parent key to reflect our changes */
1173 struct btrfs_disk_key mid_key;
1174 btrfs_node_key(mid, &mid_key, 0);
1175 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1176 BTRFS_MOD_LOG_KEY_REPLACE);
1177 BUG_ON(ret < 0);
1178 btrfs_set_node_key(parent, &mid_key, pslot);
1179 btrfs_mark_buffer_dirty(parent);
1180 }
1181
1182 /* update the path */
1183 if (left) {
1184 if (btrfs_header_nritems(left) > orig_slot) {
1185 atomic_inc(&left->refs);
1186 /* left was locked after cow */
1187 path->nodes[level] = left;
1188 path->slots[level + 1] -= 1;
1189 path->slots[level] = orig_slot;
1190 if (mid) {
1191 btrfs_tree_unlock(mid);
1192 free_extent_buffer(mid);
1193 }
1194 } else {
1195 orig_slot -= btrfs_header_nritems(left);
1196 path->slots[level] = orig_slot;
1197 }
1198 }
1199 /* double check we haven't messed things up */
1200 if (orig_ptr !=
1201 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1202 BUG();
1203enospc:
1204 if (right) {
1205 btrfs_tree_unlock(right);
1206 free_extent_buffer(right);
1207 }
1208 if (left) {
1209 if (path->nodes[level] != left)
1210 btrfs_tree_unlock(left);
1211 free_extent_buffer(left);
1212 }
1213 return ret;
1214}
1215
1216/* Node balancing for insertion. Here we only split or push nodes around
1217 * when they are completely full. This is also done top down, so we
1218 * have to be pessimistic.
1219 */
1220static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1221 struct btrfs_root *root,
1222 struct btrfs_path *path, int level)
1223{
1224 struct btrfs_fs_info *fs_info = root->fs_info;
1225 struct extent_buffer *right = NULL;
1226 struct extent_buffer *mid;
1227 struct extent_buffer *left = NULL;
1228 struct extent_buffer *parent = NULL;
1229 int ret = 0;
1230 int wret;
1231 int pslot;
1232 int orig_slot = path->slots[level];
1233
1234 if (level == 0)
1235 return 1;
1236
1237 mid = path->nodes[level];
1238 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1239
1240 if (level < BTRFS_MAX_LEVEL - 1) {
1241 parent = path->nodes[level + 1];
1242 pslot = path->slots[level + 1];
1243 }
1244
1245 if (!parent)
1246 return 1;
1247
1248 left = btrfs_read_node_slot(parent, pslot - 1);
1249 if (IS_ERR(left))
1250 left = NULL;
1251
1252 /* first, try to make some room in the middle buffer */
1253 if (left) {
1254 u32 left_nr;
1255
1256 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
1257
1258 left_nr = btrfs_header_nritems(left);
1259 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1260 wret = 1;
1261 } else {
1262 ret = btrfs_cow_block(trans, root, left, parent,
1263 pslot - 1, &left,
1264 BTRFS_NESTING_LEFT_COW);
1265 if (ret)
1266 wret = 1;
1267 else {
1268 wret = push_node_left(trans, left, mid, 0);
1269 }
1270 }
1271 if (wret < 0)
1272 ret = wret;
1273 if (wret == 0) {
1274 struct btrfs_disk_key disk_key;
1275 orig_slot += left_nr;
1276 btrfs_node_key(mid, &disk_key, 0);
1277 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1278 BTRFS_MOD_LOG_KEY_REPLACE);
1279 BUG_ON(ret < 0);
1280 btrfs_set_node_key(parent, &disk_key, pslot);
1281 btrfs_mark_buffer_dirty(parent);
1282 if (btrfs_header_nritems(left) > orig_slot) {
1283 path->nodes[level] = left;
1284 path->slots[level + 1] -= 1;
1285 path->slots[level] = orig_slot;
1286 btrfs_tree_unlock(mid);
1287 free_extent_buffer(mid);
1288 } else {
1289 orig_slot -=
1290 btrfs_header_nritems(left);
1291 path->slots[level] = orig_slot;
1292 btrfs_tree_unlock(left);
1293 free_extent_buffer(left);
1294 }
1295 return 0;
1296 }
1297 btrfs_tree_unlock(left);
1298 free_extent_buffer(left);
1299 }
1300 right = btrfs_read_node_slot(parent, pslot + 1);
1301 if (IS_ERR(right))
1302 right = NULL;
1303
1304 /*
1305 * then try to empty the right most buffer into the middle
1306 */
1307 if (right) {
1308 u32 right_nr;
1309
1310 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1311
1312 right_nr = btrfs_header_nritems(right);
1313 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1314 wret = 1;
1315 } else {
1316 ret = btrfs_cow_block(trans, root, right,
1317 parent, pslot + 1,
1318 &right, BTRFS_NESTING_RIGHT_COW);
1319 if (ret)
1320 wret = 1;
1321 else {
1322 wret = balance_node_right(trans, right, mid);
1323 }
1324 }
1325 if (wret < 0)
1326 ret = wret;
1327 if (wret == 0) {
1328 struct btrfs_disk_key disk_key;
1329
1330 btrfs_node_key(right, &disk_key, 0);
1331 ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1332 BTRFS_MOD_LOG_KEY_REPLACE);
1333 BUG_ON(ret < 0);
1334 btrfs_set_node_key(parent, &disk_key, pslot + 1);
1335 btrfs_mark_buffer_dirty(parent);
1336
1337 if (btrfs_header_nritems(mid) <= orig_slot) {
1338 path->nodes[level] = right;
1339 path->slots[level + 1] += 1;
1340 path->slots[level] = orig_slot -
1341 btrfs_header_nritems(mid);
1342 btrfs_tree_unlock(mid);
1343 free_extent_buffer(mid);
1344 } else {
1345 btrfs_tree_unlock(right);
1346 free_extent_buffer(right);
1347 }
1348 return 0;
1349 }
1350 btrfs_tree_unlock(right);
1351 free_extent_buffer(right);
1352 }
1353 return 1;
1354}
1355
1356/*
1357 * readahead one full node of leaves, finding things that are close
1358 * to the block in 'slot', and triggering ra on them.
1359 */
1360static void reada_for_search(struct btrfs_fs_info *fs_info,
1361 struct btrfs_path *path,
1362 int level, int slot, u64 objectid)
1363{
1364 struct extent_buffer *node;
1365 struct btrfs_disk_key disk_key;
1366 u32 nritems;
1367 u64 search;
1368 u64 target;
1369 u64 nread = 0;
1370 u64 nread_max;
1371 u32 nr;
1372 u32 blocksize;
1373 u32 nscan = 0;
1374
1375 if (level != 1 && path->reada != READA_FORWARD_ALWAYS)
1376 return;
1377
1378 if (!path->nodes[level])
1379 return;
1380
1381 node = path->nodes[level];
1382
1383 /*
1384 * Since the time between visiting leaves is much shorter than the time
1385 * between visiting nodes, limit read ahead of nodes to 1, to avoid too
1386 * much IO at once (possibly random).
1387 */
1388 if (path->reada == READA_FORWARD_ALWAYS) {
1389 if (level > 1)
1390 nread_max = node->fs_info->nodesize;
1391 else
1392 nread_max = SZ_128K;
1393 } else {
1394 nread_max = SZ_64K;
1395 }
1396
1397 search = btrfs_node_blockptr(node, slot);
1398 blocksize = fs_info->nodesize;
1399 if (path->reada != READA_FORWARD_ALWAYS) {
1400 struct extent_buffer *eb;
1401
1402 eb = find_extent_buffer(fs_info, search);
1403 if (eb) {
1404 free_extent_buffer(eb);
1405 return;
1406 }
1407 }
1408
1409 target = search;
1410
1411 nritems = btrfs_header_nritems(node);
1412 nr = slot;
1413
1414 while (1) {
1415 if (path->reada == READA_BACK) {
1416 if (nr == 0)
1417 break;
1418 nr--;
1419 } else if (path->reada == READA_FORWARD ||
1420 path->reada == READA_FORWARD_ALWAYS) {
1421 nr++;
1422 if (nr >= nritems)
1423 break;
1424 }
1425 if (path->reada == READA_BACK && objectid) {
1426 btrfs_node_key(node, &disk_key, nr);
1427 if (btrfs_disk_key_objectid(&disk_key) != objectid)
1428 break;
1429 }
1430 search = btrfs_node_blockptr(node, nr);
1431 if (path->reada == READA_FORWARD_ALWAYS ||
1432 (search <= target && target - search <= 65536) ||
1433 (search > target && search - target <= 65536)) {
1434 btrfs_readahead_node_child(node, nr);
1435 nread += blocksize;
1436 }
1437 nscan++;
1438 if (nread > nread_max || nscan > 32)
1439 break;
1440 }
1441}
1442
1443static noinline void reada_for_balance(struct btrfs_path *path, int level)
1444{
1445 struct extent_buffer *parent;
1446 int slot;
1447 int nritems;
1448
1449 parent = path->nodes[level + 1];
1450 if (!parent)
1451 return;
1452
1453 nritems = btrfs_header_nritems(parent);
1454 slot = path->slots[level + 1];
1455
1456 if (slot > 0)
1457 btrfs_readahead_node_child(parent, slot - 1);
1458 if (slot + 1 < nritems)
1459 btrfs_readahead_node_child(parent, slot + 1);
1460}
1461
1462
1463/*
1464 * when we walk down the tree, it is usually safe to unlock the higher layers
1465 * in the tree. The exceptions are when our path goes through slot 0, because
1466 * operations on the tree might require changing key pointers higher up in the
1467 * tree.
1468 *
1469 * callers might also have set path->keep_locks, which tells this code to keep
1470 * the lock if the path points to the last slot in the block. This is part of
1471 * walking through the tree, and selecting the next slot in the higher block.
1472 *
1473 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
1474 * if lowest_unlock is 1, level 0 won't be unlocked
1475 */
1476static noinline void unlock_up(struct btrfs_path *path, int level,
1477 int lowest_unlock, int min_write_lock_level,
1478 int *write_lock_level)
1479{
1480 int i;
1481 int skip_level = level;
1482 bool check_skip = true;
1483
1484 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1485 if (!path->nodes[i])
1486 break;
1487 if (!path->locks[i])
1488 break;
1489
1490 if (check_skip) {
1491 if (path->slots[i] == 0) {
1492 skip_level = i + 1;
1493 continue;
1494 }
1495
1496 if (path->keep_locks) {
1497 u32 nritems;
1498
1499 nritems = btrfs_header_nritems(path->nodes[i]);
1500 if (nritems < 1 || path->slots[i] >= nritems - 1) {
1501 skip_level = i + 1;
1502 continue;
1503 }
1504 }
1505 }
1506
1507 if (i >= lowest_unlock && i > skip_level) {
1508 check_skip = false;
1509 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
1510 path->locks[i] = 0;
1511 if (write_lock_level &&
1512 i > min_write_lock_level &&
1513 i <= *write_lock_level) {
1514 *write_lock_level = i - 1;
1515 }
1516 }
1517 }
1518}
1519
1520/*
1521 * Helper function for btrfs_search_slot() and other functions that do a search
1522 * on a btree. The goal is to find a tree block in the cache (the radix tree at
1523 * fs_info->buffer_radix), but if we can't find it, or it's not up to date, read
1524 * its pages from disk.
1525 *
1526 * Returns -EAGAIN, with the path unlocked, if the caller needs to repeat the
1527 * whole btree search, starting again from the current root node.
1528 */
1529static int
1530read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
1531 struct extent_buffer **eb_ret, int level, int slot,
1532 const struct btrfs_key *key)
1533{
1534 struct btrfs_fs_info *fs_info = root->fs_info;
1535 struct btrfs_tree_parent_check check = { 0 };
1536 u64 blocknr;
1537 u64 gen;
1538 struct extent_buffer *tmp;
1539 int ret;
1540 int parent_level;
1541 bool unlock_up;
1542
1543 unlock_up = ((level + 1 < BTRFS_MAX_LEVEL) && p->locks[level + 1]);
1544 blocknr = btrfs_node_blockptr(*eb_ret, slot);
1545 gen = btrfs_node_ptr_generation(*eb_ret, slot);
1546 parent_level = btrfs_header_level(*eb_ret);
1547 btrfs_node_key_to_cpu(*eb_ret, &check.first_key, slot);
1548 check.has_first_key = true;
1549 check.level = parent_level - 1;
1550 check.transid = gen;
1551 check.owner_root = root->root_key.objectid;
1552
1553 /*
1554 * If we need to read an extent buffer from disk and we are holding locks
1555 * on upper level nodes, we unlock all the upper nodes before reading the
1556 * extent buffer, and then return -EAGAIN to the caller as it needs to
1557 * restart the search. We don't release the lock on the current level
1558 * because we need to walk this node to figure out which blocks to read.
1559 */
1560 tmp = find_extent_buffer(fs_info, blocknr);
1561 if (tmp) {
1562 if (p->reada == READA_FORWARD_ALWAYS)
1563 reada_for_search(fs_info, p, level, slot, key->objectid);
1564
1565 /* first we do an atomic uptodate check */
1566 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
1567 /*
1568 * Do extra check for first_key, eb can be stale due to
1569 * being cached, read from scrub, or have multiple
1570 * parents (shared tree blocks).
1571 */
1572 if (btrfs_verify_level_key(tmp,
1573 parent_level - 1, &check.first_key, gen)) {
1574 free_extent_buffer(tmp);
1575 return -EUCLEAN;
1576 }
1577 *eb_ret = tmp;
1578 return 0;
1579 }
1580
1581 if (p->nowait) {
1582 free_extent_buffer(tmp);
1583 return -EAGAIN;
1584 }
1585
1586 if (unlock_up)
1587 btrfs_unlock_up_safe(p, level + 1);
1588
1589 /* now we're allowed to do a blocking uptodate check */
1590 ret = btrfs_read_extent_buffer(tmp, &check);
1591 if (ret) {
1592 free_extent_buffer(tmp);
1593 btrfs_release_path(p);
1594 return -EIO;
1595 }
1596 if (btrfs_check_eb_owner(tmp, root->root_key.objectid)) {
1597 free_extent_buffer(tmp);
1598 btrfs_release_path(p);
1599 return -EUCLEAN;
1600 }
1601
1602 if (unlock_up)
1603 ret = -EAGAIN;
1604
1605 goto out;
1606 } else if (p->nowait) {
1607 return -EAGAIN;
1608 }
1609
1610 if (unlock_up) {
1611 btrfs_unlock_up_safe(p, level + 1);
1612 ret = -EAGAIN;
1613 } else {
1614 ret = 0;
1615 }
1616
1617 if (p->reada != READA_NONE)
1618 reada_for_search(fs_info, p, level, slot, key->objectid);
1619
1620 tmp = read_tree_block(fs_info, blocknr, &check);
1621 if (IS_ERR(tmp)) {
1622 btrfs_release_path(p);
1623 return PTR_ERR(tmp);
1624 }
1625 /*
1626 * If the read above didn't mark this buffer up to date,
1627 * it will never end up being up to date. Set ret to EIO now
1628 * and give up so that our caller doesn't loop forever
1629 * on our EAGAINs.
1630 */
1631 if (!extent_buffer_uptodate(tmp))
1632 ret = -EIO;
1633
1634out:
1635 if (ret == 0) {
1636 *eb_ret = tmp;
1637 } else {
1638 free_extent_buffer(tmp);
1639 btrfs_release_path(p);
1640 }
1641
1642 return ret;
1643}
1644
1645/*
1646 * helper function for btrfs_search_slot. This does all of the checks
1647 * for node-level blocks and does any balancing required based on
1648 * the ins_len.
1649 *
1650 * If no extra work was required, zero is returned. If we had to
1651 * drop the path, -EAGAIN is returned and btrfs_search_slot must
1652 * start over
1653 */
1654static int
1655setup_nodes_for_search(struct btrfs_trans_handle *trans,
1656 struct btrfs_root *root, struct btrfs_path *p,
1657 struct extent_buffer *b, int level, int ins_len,
1658 int *write_lock_level)
1659{
1660 struct btrfs_fs_info *fs_info = root->fs_info;
1661 int ret = 0;
1662
1663 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1664 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
1665
1666 if (*write_lock_level < level + 1) {
1667 *write_lock_level = level + 1;
1668 btrfs_release_path(p);
1669 return -EAGAIN;
1670 }
1671
1672 reada_for_balance(p, level);
1673 ret = split_node(trans, root, p, level);
1674
1675 b = p->nodes[level];
1676 } else if (ins_len < 0 && btrfs_header_nritems(b) <
1677 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
1678
1679 if (*write_lock_level < level + 1) {
1680 *write_lock_level = level + 1;
1681 btrfs_release_path(p);
1682 return -EAGAIN;
1683 }
1684
1685 reada_for_balance(p, level);
1686 ret = balance_level(trans, root, p, level);
1687 if (ret)
1688 return ret;
1689
1690 b = p->nodes[level];
1691 if (!b) {
1692 btrfs_release_path(p);
1693 return -EAGAIN;
1694 }
1695 BUG_ON(btrfs_header_nritems(b) == 1);
1696 }
1697 return ret;
1698}
1699
1700int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
1701 u64 iobjectid, u64 ioff, u8 key_type,
1702 struct btrfs_key *found_key)
1703{
1704 int ret;
1705 struct btrfs_key key;
1706 struct extent_buffer *eb;
1707
1708 ASSERT(path);
1709 ASSERT(found_key);
1710
1711 key.type = key_type;
1712 key.objectid = iobjectid;
1713 key.offset = ioff;
1714
1715 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1716 if (ret < 0)
1717 return ret;
1718
1719 eb = path->nodes[0];
1720 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1721 ret = btrfs_next_leaf(fs_root, path);
1722 if (ret)
1723 return ret;
1724 eb = path->nodes[0];
1725 }
1726
1727 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1728 if (found_key->type != key.type ||
1729 found_key->objectid != key.objectid)
1730 return 1;
1731
1732 return 0;
1733}
1734
1735static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
1736 struct btrfs_path *p,
1737 int write_lock_level)
1738{
1739 struct extent_buffer *b;
1740 int root_lock = 0;
1741 int level = 0;
1742
1743 if (p->search_commit_root) {
1744 b = root->commit_root;
1745 atomic_inc(&b->refs);
1746 level = btrfs_header_level(b);
1747 /*
1748 * Ensure that all callers have set skip_locking when
1749 * p->search_commit_root = 1.
1750 */
1751 ASSERT(p->skip_locking == 1);
1752
1753 goto out;
1754 }
1755
1756 if (p->skip_locking) {
1757 b = btrfs_root_node(root);
1758 level = btrfs_header_level(b);
1759 goto out;
1760 }
1761
1762 /* We try very hard to do read locks on the root */
1763 root_lock = BTRFS_READ_LOCK;
1764
1765 /*
1766 * If the level is set to maximum, we can skip trying to get the read
1767 * lock.
1768 */
1769 if (write_lock_level < BTRFS_MAX_LEVEL) {
1770 /*
1771 * We don't know the level of the root node until we actually
1772 * have it read locked
1773 */
1774 if (p->nowait) {
1775 b = btrfs_try_read_lock_root_node(root);
1776 if (IS_ERR(b))
1777 return b;
1778 } else {
1779 b = btrfs_read_lock_root_node(root);
1780 }
1781 level = btrfs_header_level(b);
1782 if (level > write_lock_level)
1783 goto out;
1784
1785 /* Whoops, must trade for write lock */
1786 btrfs_tree_read_unlock(b);
1787 free_extent_buffer(b);
1788 }
1789
1790 b = btrfs_lock_root_node(root);
1791 root_lock = BTRFS_WRITE_LOCK;
1792
1793 /* The level might have changed, check again */
1794 level = btrfs_header_level(b);
1795
1796out:
1797 /*
1798 * The root may have failed to write out at some point, and thus is no
1799 * longer valid, return an error in this case.
1800 */
1801 if (!extent_buffer_uptodate(b)) {
1802 if (root_lock)
1803 btrfs_tree_unlock_rw(b, root_lock);
1804 free_extent_buffer(b);
1805 return ERR_PTR(-EIO);
1806 }
1807
1808 p->nodes[level] = b;
1809 if (!p->skip_locking)
1810 p->locks[level] = root_lock;
1811 /*
1812 * Callers are responsible for dropping b's references.
1813 */
1814 return b;
1815}
1816
1817/*
1818 * Replace the extent buffer at the lowest level of the path with a cloned
1819 * version. The purpose is to be able to use it safely, after releasing the
1820 * commit root semaphore, even if relocation is happening in parallel, the
1821 * transaction used for relocation is committed and the extent buffer is
1822 * reallocated in the next transaction.
1823 *
1824 * This is used in a context where the caller does not prevent transaction
1825 * commits from happening, either by holding a transaction handle or holding
1826 * some lock, while it's doing searches through a commit root.
1827 * At the moment it's only used for send operations.
1828 */
1829static int finish_need_commit_sem_search(struct btrfs_path *path)
1830{
1831 const int i = path->lowest_level;
1832 const int slot = path->slots[i];
1833 struct extent_buffer *lowest = path->nodes[i];
1834 struct extent_buffer *clone;
1835
1836 ASSERT(path->need_commit_sem);
1837
1838 if (!lowest)
1839 return 0;
1840
1841 lockdep_assert_held_read(&lowest->fs_info->commit_root_sem);
1842
1843 clone = btrfs_clone_extent_buffer(lowest);
1844 if (!clone)
1845 return -ENOMEM;
1846
1847 btrfs_release_path(path);
1848 path->nodes[i] = clone;
1849 path->slots[i] = slot;
1850
1851 return 0;
1852}
1853
1854static inline int search_for_key_slot(struct extent_buffer *eb,
1855 int search_low_slot,
1856 const struct btrfs_key *key,
1857 int prev_cmp,
1858 int *slot)
1859{
1860 /*
1861 * If a previous call to btrfs_bin_search() on a parent node returned an
1862 * exact match (prev_cmp == 0), we can safely assume the target key will
1863 * always be at slot 0 on lower levels, since each key pointer
1864 * (struct btrfs_key_ptr) refers to the lowest key accessible from the
1865 * subtree it points to. Thus we can skip searching lower levels.
1866 */
1867 if (prev_cmp == 0) {
1868 *slot = 0;
1869 return 0;
1870 }
1871
1872 return generic_bin_search(eb, search_low_slot, key, slot);
1873}
1874
1875static int search_leaf(struct btrfs_trans_handle *trans,
1876 struct btrfs_root *root,
1877 const struct btrfs_key *key,
1878 struct btrfs_path *path,
1879 int ins_len,
1880 int prev_cmp)
1881{
1882 struct extent_buffer *leaf = path->nodes[0];
1883 int leaf_free_space = -1;
1884 int search_low_slot = 0;
1885 int ret;
1886 bool do_bin_search = true;
1887
1888 /*
1889 * If we are doing an insertion, the leaf has enough free space and the
1890 * destination slot for the key is not slot 0, then we can unlock our
1891 * write lock on the parent, and any other upper nodes, before doing the
1892 * binary search on the leaf (with search_for_key_slot()), allowing other
1893 * tasks to lock the parent and any other upper nodes.
1894 */
1895 if (ins_len > 0) {
1896 /*
1897 * Cache the leaf free space, since we will need it later and it
1898 * will not change until then.
1899 */
1900 leaf_free_space = btrfs_leaf_free_space(leaf);
1901
1902 /*
1903 * !path->locks[1] means we have a single node tree, the leaf is
1904 * the root of the tree.
1905 */
1906 if (path->locks[1] && leaf_free_space >= ins_len) {
1907 struct btrfs_disk_key first_key;
1908
1909 ASSERT(btrfs_header_nritems(leaf) > 0);
1910 btrfs_item_key(leaf, &first_key, 0);
1911
1912 /*
1913 * Doing the extra comparison with the first key is cheap,
1914 * taking into account that the first key is very likely
1915 * already in a cache line because it immediately follows
1916 * the extent buffer's header and we have recently accessed
1917 * the header's level field.
1918 */
1919 ret = comp_keys(&first_key, key);
1920 if (ret < 0) {
1921 /*
1922 * The first key is smaller than the key we want
1923 * to insert, so we are safe to unlock all upper
1924 * nodes and we have to do the binary search.
1925 *
1926 * We do use btrfs_unlock_up_safe() and not
1927 * unlock_up() because the later does not unlock
1928 * nodes with a slot of 0 - we can safely unlock
1929 * any node even if its slot is 0 since in this
1930 * case the key does not end up at slot 0 of the
1931 * leaf and there's no need to split the leaf.
1932 */
1933 btrfs_unlock_up_safe(path, 1);
1934 search_low_slot = 1;
1935 } else {
1936 /*
1937 * The first key is >= then the key we want to
1938 * insert, so we can skip the binary search as
1939 * the target key will be at slot 0.
1940 *
1941 * We can not unlock upper nodes when the key is
1942 * less than the first key, because we will need
1943 * to update the key at slot 0 of the parent node
1944 * and possibly of other upper nodes too.
1945 * If the key matches the first key, then we can
1946 * unlock all the upper nodes, using
1947 * btrfs_unlock_up_safe() instead of unlock_up()
1948 * as stated above.
1949 */
1950 if (ret == 0)
1951 btrfs_unlock_up_safe(path, 1);
1952 /*
1953 * ret is already 0 or 1, matching the result of
1954 * a btrfs_bin_search() call, so there is no need
1955 * to adjust it.
1956 */
1957 do_bin_search = false;
1958 path->slots[0] = 0;
1959 }
1960 }
1961 }
1962
1963 if (do_bin_search) {
1964 ret = search_for_key_slot(leaf, search_low_slot, key,
1965 prev_cmp, &path->slots[0]);
1966 if (ret < 0)
1967 return ret;
1968 }
1969
1970 if (ins_len > 0) {
1971 /*
1972 * Item key already exists. In this case, if we are allowed to
1973 * insert the item (for example, in dir_item case, item key
1974 * collision is allowed), it will be merged with the original
1975 * item. Only the item size grows, no new btrfs item will be
1976 * added. If search_for_extension is not set, ins_len already
1977 * accounts the size btrfs_item, deduct it here so leaf space
1978 * check will be correct.
1979 */
1980 if (ret == 0 && !path->search_for_extension) {
1981 ASSERT(ins_len >= sizeof(struct btrfs_item));
1982 ins_len -= sizeof(struct btrfs_item);
1983 }
1984
1985 ASSERT(leaf_free_space >= 0);
1986
1987 if (leaf_free_space < ins_len) {
1988 int err;
1989
1990 err = split_leaf(trans, root, key, path, ins_len,
1991 (ret == 0));
1992 ASSERT(err <= 0);
1993 if (WARN_ON(err > 0))
1994 err = -EUCLEAN;
1995 if (err)
1996 ret = err;
1997 }
1998 }
1999
2000 return ret;
2001}
2002
2003/*
2004 * btrfs_search_slot - look for a key in a tree and perform necessary
2005 * modifications to preserve tree invariants.
2006 *
2007 * @trans: Handle of transaction, used when modifying the tree
2008 * @p: Holds all btree nodes along the search path
2009 * @root: The root node of the tree
2010 * @key: The key we are looking for
2011 * @ins_len: Indicates purpose of search:
2012 * >0 for inserts it's size of item inserted (*)
2013 * <0 for deletions
2014 * 0 for plain searches, not modifying the tree
2015 *
2016 * (*) If size of item inserted doesn't include
2017 * sizeof(struct btrfs_item), then p->search_for_extension must
2018 * be set.
2019 * @cow: boolean should CoW operations be performed. Must always be 1
2020 * when modifying the tree.
2021 *
2022 * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2023 * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2024 *
2025 * If @key is found, 0 is returned and you can find the item in the leaf level
2026 * of the path (level 0)
2027 *
2028 * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2029 * points to the slot where it should be inserted
2030 *
2031 * If an error is encountered while searching the tree a negative error number
2032 * is returned
2033 */
2034int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2035 const struct btrfs_key *key, struct btrfs_path *p,
2036 int ins_len, int cow)
2037{
2038 struct btrfs_fs_info *fs_info = root->fs_info;
2039 struct extent_buffer *b;
2040 int slot;
2041 int ret;
2042 int err;
2043 int level;
2044 int lowest_unlock = 1;
2045 /* everything at write_lock_level or lower must be write locked */
2046 int write_lock_level = 0;
2047 u8 lowest_level = 0;
2048 int min_write_lock_level;
2049 int prev_cmp;
2050
2051 might_sleep();
2052
2053 lowest_level = p->lowest_level;
2054 WARN_ON(lowest_level && ins_len > 0);
2055 WARN_ON(p->nodes[0] != NULL);
2056 BUG_ON(!cow && ins_len);
2057
2058 /*
2059 * For now only allow nowait for read only operations. There's no
2060 * strict reason why we can't, we just only need it for reads so it's
2061 * only implemented for reads.
2062 */
2063 ASSERT(!p->nowait || !cow);
2064
2065 if (ins_len < 0) {
2066 lowest_unlock = 2;
2067
2068 /* when we are removing items, we might have to go up to level
2069 * two as we update tree pointers Make sure we keep write
2070 * for those levels as well
2071 */
2072 write_lock_level = 2;
2073 } else if (ins_len > 0) {
2074 /*
2075 * for inserting items, make sure we have a write lock on
2076 * level 1 so we can update keys
2077 */
2078 write_lock_level = 1;
2079 }
2080
2081 if (!cow)
2082 write_lock_level = -1;
2083
2084 if (cow && (p->keep_locks || p->lowest_level))
2085 write_lock_level = BTRFS_MAX_LEVEL;
2086
2087 min_write_lock_level = write_lock_level;
2088
2089 if (p->need_commit_sem) {
2090 ASSERT(p->search_commit_root);
2091 if (p->nowait) {
2092 if (!down_read_trylock(&fs_info->commit_root_sem))
2093 return -EAGAIN;
2094 } else {
2095 down_read(&fs_info->commit_root_sem);
2096 }
2097 }
2098
2099again:
2100 prev_cmp = -1;
2101 b = btrfs_search_slot_get_root(root, p, write_lock_level);
2102 if (IS_ERR(b)) {
2103 ret = PTR_ERR(b);
2104 goto done;
2105 }
2106
2107 while (b) {
2108 int dec = 0;
2109
2110 level = btrfs_header_level(b);
2111
2112 if (cow) {
2113 bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2114
2115 /*
2116 * if we don't really need to cow this block
2117 * then we don't want to set the path blocking,
2118 * so we test it here
2119 */
2120 if (!should_cow_block(trans, root, b))
2121 goto cow_done;
2122
2123 /*
2124 * must have write locks on this node and the
2125 * parent
2126 */
2127 if (level > write_lock_level ||
2128 (level + 1 > write_lock_level &&
2129 level + 1 < BTRFS_MAX_LEVEL &&
2130 p->nodes[level + 1])) {
2131 write_lock_level = level + 1;
2132 btrfs_release_path(p);
2133 goto again;
2134 }
2135
2136 if (last_level)
2137 err = btrfs_cow_block(trans, root, b, NULL, 0,
2138 &b,
2139 BTRFS_NESTING_COW);
2140 else
2141 err = btrfs_cow_block(trans, root, b,
2142 p->nodes[level + 1],
2143 p->slots[level + 1], &b,
2144 BTRFS_NESTING_COW);
2145 if (err) {
2146 ret = err;
2147 goto done;
2148 }
2149 }
2150cow_done:
2151 p->nodes[level] = b;
2152
2153 /*
2154 * we have a lock on b and as long as we aren't changing
2155 * the tree, there is no way to for the items in b to change.
2156 * It is safe to drop the lock on our parent before we
2157 * go through the expensive btree search on b.
2158 *
2159 * If we're inserting or deleting (ins_len != 0), then we might
2160 * be changing slot zero, which may require changing the parent.
2161 * So, we can't drop the lock until after we know which slot
2162 * we're operating on.
2163 */
2164 if (!ins_len && !p->keep_locks) {
2165 int u = level + 1;
2166
2167 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2168 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2169 p->locks[u] = 0;
2170 }
2171 }
2172
2173 if (level == 0) {
2174 if (ins_len > 0)
2175 ASSERT(write_lock_level >= 1);
2176
2177 ret = search_leaf(trans, root, key, p, ins_len, prev_cmp);
2178 if (!p->search_for_split)
2179 unlock_up(p, level, lowest_unlock,
2180 min_write_lock_level, NULL);
2181 goto done;
2182 }
2183
2184 ret = search_for_key_slot(b, 0, key, prev_cmp, &slot);
2185 if (ret < 0)
2186 goto done;
2187 prev_cmp = ret;
2188
2189 if (ret && slot > 0) {
2190 dec = 1;
2191 slot--;
2192 }
2193 p->slots[level] = slot;
2194 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
2195 &write_lock_level);
2196 if (err == -EAGAIN)
2197 goto again;
2198 if (err) {
2199 ret = err;
2200 goto done;
2201 }
2202 b = p->nodes[level];
2203 slot = p->slots[level];
2204
2205 /*
2206 * Slot 0 is special, if we change the key we have to update
2207 * the parent pointer which means we must have a write lock on
2208 * the parent
2209 */
2210 if (slot == 0 && ins_len && write_lock_level < level + 1) {
2211 write_lock_level = level + 1;
2212 btrfs_release_path(p);
2213 goto again;
2214 }
2215
2216 unlock_up(p, level, lowest_unlock, min_write_lock_level,
2217 &write_lock_level);
2218
2219 if (level == lowest_level) {
2220 if (dec)
2221 p->slots[level]++;
2222 goto done;
2223 }
2224
2225 err = read_block_for_search(root, p, &b, level, slot, key);
2226 if (err == -EAGAIN)
2227 goto again;
2228 if (err) {
2229 ret = err;
2230 goto done;
2231 }
2232
2233 if (!p->skip_locking) {
2234 level = btrfs_header_level(b);
2235
2236 btrfs_maybe_reset_lockdep_class(root, b);
2237
2238 if (level <= write_lock_level) {
2239 btrfs_tree_lock(b);
2240 p->locks[level] = BTRFS_WRITE_LOCK;
2241 } else {
2242 if (p->nowait) {
2243 if (!btrfs_try_tree_read_lock(b)) {
2244 free_extent_buffer(b);
2245 ret = -EAGAIN;
2246 goto done;
2247 }
2248 } else {
2249 btrfs_tree_read_lock(b);
2250 }
2251 p->locks[level] = BTRFS_READ_LOCK;
2252 }
2253 p->nodes[level] = b;
2254 }
2255 }
2256 ret = 1;
2257done:
2258 if (ret < 0 && !p->skip_release_on_error)
2259 btrfs_release_path(p);
2260
2261 if (p->need_commit_sem) {
2262 int ret2;
2263
2264 ret2 = finish_need_commit_sem_search(p);
2265 up_read(&fs_info->commit_root_sem);
2266 if (ret2)
2267 ret = ret2;
2268 }
2269
2270 return ret;
2271}
2272ALLOW_ERROR_INJECTION(btrfs_search_slot, ERRNO);
2273
2274/*
2275 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2276 * current state of the tree together with the operations recorded in the tree
2277 * modification log to search for the key in a previous version of this tree, as
2278 * denoted by the time_seq parameter.
2279 *
2280 * Naturally, there is no support for insert, delete or cow operations.
2281 *
2282 * The resulting path and return value will be set up as if we called
2283 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2284 */
2285int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2286 struct btrfs_path *p, u64 time_seq)
2287{
2288 struct btrfs_fs_info *fs_info = root->fs_info;
2289 struct extent_buffer *b;
2290 int slot;
2291 int ret;
2292 int err;
2293 int level;
2294 int lowest_unlock = 1;
2295 u8 lowest_level = 0;
2296
2297 lowest_level = p->lowest_level;
2298 WARN_ON(p->nodes[0] != NULL);
2299 ASSERT(!p->nowait);
2300
2301 if (p->search_commit_root) {
2302 BUG_ON(time_seq);
2303 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2304 }
2305
2306again:
2307 b = btrfs_get_old_root(root, time_seq);
2308 if (!b) {
2309 ret = -EIO;
2310 goto done;
2311 }
2312 level = btrfs_header_level(b);
2313 p->locks[level] = BTRFS_READ_LOCK;
2314
2315 while (b) {
2316 int dec = 0;
2317
2318 level = btrfs_header_level(b);
2319 p->nodes[level] = b;
2320
2321 /*
2322 * we have a lock on b and as long as we aren't changing
2323 * the tree, there is no way to for the items in b to change.
2324 * It is safe to drop the lock on our parent before we
2325 * go through the expensive btree search on b.
2326 */
2327 btrfs_unlock_up_safe(p, level + 1);
2328
2329 ret = btrfs_bin_search(b, key, &slot);
2330 if (ret < 0)
2331 goto done;
2332
2333 if (level == 0) {
2334 p->slots[level] = slot;
2335 unlock_up(p, level, lowest_unlock, 0, NULL);
2336 goto done;
2337 }
2338
2339 if (ret && slot > 0) {
2340 dec = 1;
2341 slot--;
2342 }
2343 p->slots[level] = slot;
2344 unlock_up(p, level, lowest_unlock, 0, NULL);
2345
2346 if (level == lowest_level) {
2347 if (dec)
2348 p->slots[level]++;
2349 goto done;
2350 }
2351
2352 err = read_block_for_search(root, p, &b, level, slot, key);
2353 if (err == -EAGAIN)
2354 goto again;
2355 if (err) {
2356 ret = err;
2357 goto done;
2358 }
2359
2360 level = btrfs_header_level(b);
2361 btrfs_tree_read_lock(b);
2362 b = btrfs_tree_mod_log_rewind(fs_info, p, b, time_seq);
2363 if (!b) {
2364 ret = -ENOMEM;
2365 goto done;
2366 }
2367 p->locks[level] = BTRFS_READ_LOCK;
2368 p->nodes[level] = b;
2369 }
2370 ret = 1;
2371done:
2372 if (ret < 0)
2373 btrfs_release_path(p);
2374
2375 return ret;
2376}
2377
2378/*
2379 * helper to use instead of search slot if no exact match is needed but
2380 * instead the next or previous item should be returned.
2381 * When find_higher is true, the next higher item is returned, the next lower
2382 * otherwise.
2383 * When return_any and find_higher are both true, and no higher item is found,
2384 * return the next lower instead.
2385 * When return_any is true and find_higher is false, and no lower item is found,
2386 * return the next higher instead.
2387 * It returns 0 if any item is found, 1 if none is found (tree empty), and
2388 * < 0 on error
2389 */
2390int btrfs_search_slot_for_read(struct btrfs_root *root,
2391 const struct btrfs_key *key,
2392 struct btrfs_path *p, int find_higher,
2393 int return_any)
2394{
2395 int ret;
2396 struct extent_buffer *leaf;
2397
2398again:
2399 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
2400 if (ret <= 0)
2401 return ret;
2402 /*
2403 * a return value of 1 means the path is at the position where the
2404 * item should be inserted. Normally this is the next bigger item,
2405 * but in case the previous item is the last in a leaf, path points
2406 * to the first free slot in the previous leaf, i.e. at an invalid
2407 * item.
2408 */
2409 leaf = p->nodes[0];
2410
2411 if (find_higher) {
2412 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
2413 ret = btrfs_next_leaf(root, p);
2414 if (ret <= 0)
2415 return ret;
2416 if (!return_any)
2417 return 1;
2418 /*
2419 * no higher item found, return the next
2420 * lower instead
2421 */
2422 return_any = 0;
2423 find_higher = 0;
2424 btrfs_release_path(p);
2425 goto again;
2426 }
2427 } else {
2428 if (p->slots[0] == 0) {
2429 ret = btrfs_prev_leaf(root, p);
2430 if (ret < 0)
2431 return ret;
2432 if (!ret) {
2433 leaf = p->nodes[0];
2434 if (p->slots[0] == btrfs_header_nritems(leaf))
2435 p->slots[0]--;
2436 return 0;
2437 }
2438 if (!return_any)
2439 return 1;
2440 /*
2441 * no lower item found, return the next
2442 * higher instead
2443 */
2444 return_any = 0;
2445 find_higher = 1;
2446 btrfs_release_path(p);
2447 goto again;
2448 } else {
2449 --p->slots[0];
2450 }
2451 }
2452 return 0;
2453}
2454
2455/*
2456 * Execute search and call btrfs_previous_item to traverse backwards if the item
2457 * was not found.
2458 *
2459 * Return 0 if found, 1 if not found and < 0 if error.
2460 */
2461int btrfs_search_backwards(struct btrfs_root *root, struct btrfs_key *key,
2462 struct btrfs_path *path)
2463{
2464 int ret;
2465
2466 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
2467 if (ret > 0)
2468 ret = btrfs_previous_item(root, path, key->objectid, key->type);
2469
2470 if (ret == 0)
2471 btrfs_item_key_to_cpu(path->nodes[0], key, path->slots[0]);
2472
2473 return ret;
2474}
2475
2476/*
2477 * Search for a valid slot for the given path.
2478 *
2479 * @root: The root node of the tree.
2480 * @key: Will contain a valid item if found.
2481 * @path: The starting point to validate the slot.
2482 *
2483 * Return: 0 if the item is valid
2484 * 1 if not found
2485 * <0 if error.
2486 */
2487int btrfs_get_next_valid_item(struct btrfs_root *root, struct btrfs_key *key,
2488 struct btrfs_path *path)
2489{
2490 while (1) {
2491 int ret;
2492 const int slot = path->slots[0];
2493 const struct extent_buffer *leaf = path->nodes[0];
2494
2495 /* This is where we start walking the path. */
2496 if (slot >= btrfs_header_nritems(leaf)) {
2497 /*
2498 * If we've reached the last slot in this leaf we need
2499 * to go to the next leaf and reset the path.
2500 */
2501 ret = btrfs_next_leaf(root, path);
2502 if (ret)
2503 return ret;
2504 continue;
2505 }
2506 /* Store the found, valid item in @key. */
2507 btrfs_item_key_to_cpu(leaf, key, slot);
2508 break;
2509 }
2510 return 0;
2511}
2512
2513/*
2514 * adjust the pointers going up the tree, starting at level
2515 * making sure the right key of each node is points to 'key'.
2516 * This is used after shifting pointers to the left, so it stops
2517 * fixing up pointers when a given leaf/node is not in slot 0 of the
2518 * higher levels
2519 *
2520 */
2521static void fixup_low_keys(struct btrfs_path *path,
2522 struct btrfs_disk_key *key, int level)
2523{
2524 int i;
2525 struct extent_buffer *t;
2526 int ret;
2527
2528 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2529 int tslot = path->slots[i];
2530
2531 if (!path->nodes[i])
2532 break;
2533 t = path->nodes[i];
2534 ret = btrfs_tree_mod_log_insert_key(t, tslot,
2535 BTRFS_MOD_LOG_KEY_REPLACE);
2536 BUG_ON(ret < 0);
2537 btrfs_set_node_key(t, key, tslot);
2538 btrfs_mark_buffer_dirty(path->nodes[i]);
2539 if (tslot != 0)
2540 break;
2541 }
2542}
2543
2544/*
2545 * update item key.
2546 *
2547 * This function isn't completely safe. It's the caller's responsibility
2548 * that the new key won't break the order
2549 */
2550void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
2551 struct btrfs_path *path,
2552 const struct btrfs_key *new_key)
2553{
2554 struct btrfs_disk_key disk_key;
2555 struct extent_buffer *eb;
2556 int slot;
2557
2558 eb = path->nodes[0];
2559 slot = path->slots[0];
2560 if (slot > 0) {
2561 btrfs_item_key(eb, &disk_key, slot - 1);
2562 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
2563 btrfs_crit(fs_info,
2564 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2565 slot, btrfs_disk_key_objectid(&disk_key),
2566 btrfs_disk_key_type(&disk_key),
2567 btrfs_disk_key_offset(&disk_key),
2568 new_key->objectid, new_key->type,
2569 new_key->offset);
2570 btrfs_print_leaf(eb);
2571 BUG();
2572 }
2573 }
2574 if (slot < btrfs_header_nritems(eb) - 1) {
2575 btrfs_item_key(eb, &disk_key, slot + 1);
2576 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
2577 btrfs_crit(fs_info,
2578 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2579 slot, btrfs_disk_key_objectid(&disk_key),
2580 btrfs_disk_key_type(&disk_key),
2581 btrfs_disk_key_offset(&disk_key),
2582 new_key->objectid, new_key->type,
2583 new_key->offset);
2584 btrfs_print_leaf(eb);
2585 BUG();
2586 }
2587 }
2588
2589 btrfs_cpu_key_to_disk(&disk_key, new_key);
2590 btrfs_set_item_key(eb, &disk_key, slot);
2591 btrfs_mark_buffer_dirty(eb);
2592 if (slot == 0)
2593 fixup_low_keys(path, &disk_key, 1);
2594}
2595
2596/*
2597 * Check key order of two sibling extent buffers.
2598 *
2599 * Return true if something is wrong.
2600 * Return false if everything is fine.
2601 *
2602 * Tree-checker only works inside one tree block, thus the following
2603 * corruption can not be detected by tree-checker:
2604 *
2605 * Leaf @left | Leaf @right
2606 * --------------------------------------------------------------
2607 * | 1 | 2 | 3 | 4 | 5 | f6 | | 7 | 8 |
2608 *
2609 * Key f6 in leaf @left itself is valid, but not valid when the next
2610 * key in leaf @right is 7.
2611 * This can only be checked at tree block merge time.
2612 * And since tree checker has ensured all key order in each tree block
2613 * is correct, we only need to bother the last key of @left and the first
2614 * key of @right.
2615 */
2616static bool check_sibling_keys(struct extent_buffer *left,
2617 struct extent_buffer *right)
2618{
2619 struct btrfs_key left_last;
2620 struct btrfs_key right_first;
2621 int level = btrfs_header_level(left);
2622 int nr_left = btrfs_header_nritems(left);
2623 int nr_right = btrfs_header_nritems(right);
2624
2625 /* No key to check in one of the tree blocks */
2626 if (!nr_left || !nr_right)
2627 return false;
2628
2629 if (level) {
2630 btrfs_node_key_to_cpu(left, &left_last, nr_left - 1);
2631 btrfs_node_key_to_cpu(right, &right_first, 0);
2632 } else {
2633 btrfs_item_key_to_cpu(left, &left_last, nr_left - 1);
2634 btrfs_item_key_to_cpu(right, &right_first, 0);
2635 }
2636
2637 if (btrfs_comp_cpu_keys(&left_last, &right_first) >= 0) {
2638 btrfs_crit(left->fs_info,
2639"bad key order, sibling blocks, left last (%llu %u %llu) right first (%llu %u %llu)",
2640 left_last.objectid, left_last.type,
2641 left_last.offset, right_first.objectid,
2642 right_first.type, right_first.offset);
2643 return true;
2644 }
2645 return false;
2646}
2647
2648/*
2649 * try to push data from one node into the next node left in the
2650 * tree.
2651 *
2652 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2653 * error, and > 0 if there was no room in the left hand block.
2654 */
2655static int push_node_left(struct btrfs_trans_handle *trans,
2656 struct extent_buffer *dst,
2657 struct extent_buffer *src, int empty)
2658{
2659 struct btrfs_fs_info *fs_info = trans->fs_info;
2660 int push_items = 0;
2661 int src_nritems;
2662 int dst_nritems;
2663 int ret = 0;
2664
2665 src_nritems = btrfs_header_nritems(src);
2666 dst_nritems = btrfs_header_nritems(dst);
2667 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2668 WARN_ON(btrfs_header_generation(src) != trans->transid);
2669 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2670
2671 if (!empty && src_nritems <= 8)
2672 return 1;
2673
2674 if (push_items <= 0)
2675 return 1;
2676
2677 if (empty) {
2678 push_items = min(src_nritems, push_items);
2679 if (push_items < src_nritems) {
2680 /* leave at least 8 pointers in the node if
2681 * we aren't going to empty it
2682 */
2683 if (src_nritems - push_items < 8) {
2684 if (push_items <= 8)
2685 return 1;
2686 push_items -= 8;
2687 }
2688 }
2689 } else
2690 push_items = min(src_nritems - 8, push_items);
2691
2692 /* dst is the left eb, src is the middle eb */
2693 if (check_sibling_keys(dst, src)) {
2694 ret = -EUCLEAN;
2695 btrfs_abort_transaction(trans, ret);
2696 return ret;
2697 }
2698 ret = btrfs_tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
2699 if (ret) {
2700 btrfs_abort_transaction(trans, ret);
2701 return ret;
2702 }
2703 copy_extent_buffer(dst, src,
2704 btrfs_node_key_ptr_offset(dst, dst_nritems),
2705 btrfs_node_key_ptr_offset(src, 0),
2706 push_items * sizeof(struct btrfs_key_ptr));
2707
2708 if (push_items < src_nritems) {
2709 /*
2710 * Don't call btrfs_tree_mod_log_insert_move() here, key removal
2711 * was already fully logged by btrfs_tree_mod_log_eb_copy() above.
2712 */
2713 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(src, 0),
2714 btrfs_node_key_ptr_offset(src, push_items),
2715 (src_nritems - push_items) *
2716 sizeof(struct btrfs_key_ptr));
2717 }
2718 btrfs_set_header_nritems(src, src_nritems - push_items);
2719 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2720 btrfs_mark_buffer_dirty(src);
2721 btrfs_mark_buffer_dirty(dst);
2722
2723 return ret;
2724}
2725
2726/*
2727 * try to push data from one node into the next node right in the
2728 * tree.
2729 *
2730 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2731 * error, and > 0 if there was no room in the right hand block.
2732 *
2733 * this will only push up to 1/2 the contents of the left node over
2734 */
2735static int balance_node_right(struct btrfs_trans_handle *trans,
2736 struct extent_buffer *dst,
2737 struct extent_buffer *src)
2738{
2739 struct btrfs_fs_info *fs_info = trans->fs_info;
2740 int push_items = 0;
2741 int max_push;
2742 int src_nritems;
2743 int dst_nritems;
2744 int ret = 0;
2745
2746 WARN_ON(btrfs_header_generation(src) != trans->transid);
2747 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2748
2749 src_nritems = btrfs_header_nritems(src);
2750 dst_nritems = btrfs_header_nritems(dst);
2751 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2752 if (push_items <= 0)
2753 return 1;
2754
2755 if (src_nritems < 4)
2756 return 1;
2757
2758 max_push = src_nritems / 2 + 1;
2759 /* don't try to empty the node */
2760 if (max_push >= src_nritems)
2761 return 1;
2762
2763 if (max_push < push_items)
2764 push_items = max_push;
2765
2766 /* dst is the right eb, src is the middle eb */
2767 if (check_sibling_keys(src, dst)) {
2768 ret = -EUCLEAN;
2769 btrfs_abort_transaction(trans, ret);
2770 return ret;
2771 }
2772 ret = btrfs_tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
2773 BUG_ON(ret < 0);
2774 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(dst, push_items),
2775 btrfs_node_key_ptr_offset(dst, 0),
2776 (dst_nritems) *
2777 sizeof(struct btrfs_key_ptr));
2778
2779 ret = btrfs_tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
2780 push_items);
2781 if (ret) {
2782 btrfs_abort_transaction(trans, ret);
2783 return ret;
2784 }
2785 copy_extent_buffer(dst, src,
2786 btrfs_node_key_ptr_offset(dst, 0),
2787 btrfs_node_key_ptr_offset(src, src_nritems - push_items),
2788 push_items * sizeof(struct btrfs_key_ptr));
2789
2790 btrfs_set_header_nritems(src, src_nritems - push_items);
2791 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2792
2793 btrfs_mark_buffer_dirty(src);
2794 btrfs_mark_buffer_dirty(dst);
2795
2796 return ret;
2797}
2798
2799/*
2800 * helper function to insert a new root level in the tree.
2801 * A new node is allocated, and a single item is inserted to
2802 * point to the existing root
2803 *
2804 * returns zero on success or < 0 on failure.
2805 */
2806static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2807 struct btrfs_root *root,
2808 struct btrfs_path *path, int level)
2809{
2810 struct btrfs_fs_info *fs_info = root->fs_info;
2811 u64 lower_gen;
2812 struct extent_buffer *lower;
2813 struct extent_buffer *c;
2814 struct extent_buffer *old;
2815 struct btrfs_disk_key lower_key;
2816 int ret;
2817
2818 BUG_ON(path->nodes[level]);
2819 BUG_ON(path->nodes[level-1] != root->node);
2820
2821 lower = path->nodes[level-1];
2822 if (level == 1)
2823 btrfs_item_key(lower, &lower_key, 0);
2824 else
2825 btrfs_node_key(lower, &lower_key, 0);
2826
2827 c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
2828 &lower_key, level, root->node->start, 0,
2829 BTRFS_NESTING_NEW_ROOT);
2830 if (IS_ERR(c))
2831 return PTR_ERR(c);
2832
2833 root_add_used(root, fs_info->nodesize);
2834
2835 btrfs_set_header_nritems(c, 1);
2836 btrfs_set_node_key(c, &lower_key, 0);
2837 btrfs_set_node_blockptr(c, 0, lower->start);
2838 lower_gen = btrfs_header_generation(lower);
2839 WARN_ON(lower_gen != trans->transid);
2840
2841 btrfs_set_node_ptr_generation(c, 0, lower_gen);
2842
2843 btrfs_mark_buffer_dirty(c);
2844
2845 old = root->node;
2846 ret = btrfs_tree_mod_log_insert_root(root->node, c, false);
2847 BUG_ON(ret < 0);
2848 rcu_assign_pointer(root->node, c);
2849
2850 /* the super has an extra ref to root->node */
2851 free_extent_buffer(old);
2852
2853 add_root_to_dirty_list(root);
2854 atomic_inc(&c->refs);
2855 path->nodes[level] = c;
2856 path->locks[level] = BTRFS_WRITE_LOCK;
2857 path->slots[level] = 0;
2858 return 0;
2859}
2860
2861/*
2862 * worker function to insert a single pointer in a node.
2863 * the node should have enough room for the pointer already
2864 *
2865 * slot and level indicate where you want the key to go, and
2866 * blocknr is the block the key points to.
2867 */
2868static void insert_ptr(struct btrfs_trans_handle *trans,
2869 struct btrfs_path *path,
2870 struct btrfs_disk_key *key, u64 bytenr,
2871 int slot, int level)
2872{
2873 struct extent_buffer *lower;
2874 int nritems;
2875 int ret;
2876
2877 BUG_ON(!path->nodes[level]);
2878 btrfs_assert_tree_write_locked(path->nodes[level]);
2879 lower = path->nodes[level];
2880 nritems = btrfs_header_nritems(lower);
2881 BUG_ON(slot > nritems);
2882 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
2883 if (slot != nritems) {
2884 if (level) {
2885 ret = btrfs_tree_mod_log_insert_move(lower, slot + 1,
2886 slot, nritems - slot);
2887 BUG_ON(ret < 0);
2888 }
2889 memmove_extent_buffer(lower,
2890 btrfs_node_key_ptr_offset(lower, slot + 1),
2891 btrfs_node_key_ptr_offset(lower, slot),
2892 (nritems - slot) * sizeof(struct btrfs_key_ptr));
2893 }
2894 if (level) {
2895 ret = btrfs_tree_mod_log_insert_key(lower, slot,
2896 BTRFS_MOD_LOG_KEY_ADD);
2897 BUG_ON(ret < 0);
2898 }
2899 btrfs_set_node_key(lower, key, slot);
2900 btrfs_set_node_blockptr(lower, slot, bytenr);
2901 WARN_ON(trans->transid == 0);
2902 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2903 btrfs_set_header_nritems(lower, nritems + 1);
2904 btrfs_mark_buffer_dirty(lower);
2905}
2906
2907/*
2908 * split the node at the specified level in path in two.
2909 * The path is corrected to point to the appropriate node after the split
2910 *
2911 * Before splitting this tries to make some room in the node by pushing
2912 * left and right, if either one works, it returns right away.
2913 *
2914 * returns 0 on success and < 0 on failure
2915 */
2916static noinline int split_node(struct btrfs_trans_handle *trans,
2917 struct btrfs_root *root,
2918 struct btrfs_path *path, int level)
2919{
2920 struct btrfs_fs_info *fs_info = root->fs_info;
2921 struct extent_buffer *c;
2922 struct extent_buffer *split;
2923 struct btrfs_disk_key disk_key;
2924 int mid;
2925 int ret;
2926 u32 c_nritems;
2927
2928 c = path->nodes[level];
2929 WARN_ON(btrfs_header_generation(c) != trans->transid);
2930 if (c == root->node) {
2931 /*
2932 * trying to split the root, lets make a new one
2933 *
2934 * tree mod log: We don't log_removal old root in
2935 * insert_new_root, because that root buffer will be kept as a
2936 * normal node. We are going to log removal of half of the
2937 * elements below with btrfs_tree_mod_log_eb_copy(). We're
2938 * holding a tree lock on the buffer, which is why we cannot
2939 * race with other tree_mod_log users.
2940 */
2941 ret = insert_new_root(trans, root, path, level + 1);
2942 if (ret)
2943 return ret;
2944 } else {
2945 ret = push_nodes_for_insert(trans, root, path, level);
2946 c = path->nodes[level];
2947 if (!ret && btrfs_header_nritems(c) <
2948 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
2949 return 0;
2950 if (ret < 0)
2951 return ret;
2952 }
2953
2954 c_nritems = btrfs_header_nritems(c);
2955 mid = (c_nritems + 1) / 2;
2956 btrfs_node_key(c, &disk_key, mid);
2957
2958 split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
2959 &disk_key, level, c->start, 0,
2960 BTRFS_NESTING_SPLIT);
2961 if (IS_ERR(split))
2962 return PTR_ERR(split);
2963
2964 root_add_used(root, fs_info->nodesize);
2965 ASSERT(btrfs_header_level(c) == level);
2966
2967 ret = btrfs_tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
2968 if (ret) {
2969 btrfs_abort_transaction(trans, ret);
2970 return ret;
2971 }
2972 copy_extent_buffer(split, c,
2973 btrfs_node_key_ptr_offset(split, 0),
2974 btrfs_node_key_ptr_offset(c, mid),
2975 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2976 btrfs_set_header_nritems(split, c_nritems - mid);
2977 btrfs_set_header_nritems(c, mid);
2978
2979 btrfs_mark_buffer_dirty(c);
2980 btrfs_mark_buffer_dirty(split);
2981
2982 insert_ptr(trans, path, &disk_key, split->start,
2983 path->slots[level + 1] + 1, level + 1);
2984
2985 if (path->slots[level] >= mid) {
2986 path->slots[level] -= mid;
2987 btrfs_tree_unlock(c);
2988 free_extent_buffer(c);
2989 path->nodes[level] = split;
2990 path->slots[level + 1] += 1;
2991 } else {
2992 btrfs_tree_unlock(split);
2993 free_extent_buffer(split);
2994 }
2995 return 0;
2996}
2997
2998/*
2999 * how many bytes are required to store the items in a leaf. start
3000 * and nr indicate which items in the leaf to check. This totals up the
3001 * space used both by the item structs and the item data
3002 */
3003static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3004{
3005 int data_len;
3006 int nritems = btrfs_header_nritems(l);
3007 int end = min(nritems, start + nr) - 1;
3008
3009 if (!nr)
3010 return 0;
3011 data_len = btrfs_item_offset(l, start) + btrfs_item_size(l, start);
3012 data_len = data_len - btrfs_item_offset(l, end);
3013 data_len += sizeof(struct btrfs_item) * nr;
3014 WARN_ON(data_len < 0);
3015 return data_len;
3016}
3017
3018/*
3019 * The space between the end of the leaf items and
3020 * the start of the leaf data. IOW, how much room
3021 * the leaf has left for both items and data
3022 */
3023noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
3024{
3025 struct btrfs_fs_info *fs_info = leaf->fs_info;
3026 int nritems = btrfs_header_nritems(leaf);
3027 int ret;
3028
3029 ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3030 if (ret < 0) {
3031 btrfs_crit(fs_info,
3032 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3033 ret,
3034 (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3035 leaf_space_used(leaf, 0, nritems), nritems);
3036 }
3037 return ret;
3038}
3039
3040/*
3041 * min slot controls the lowest index we're willing to push to the
3042 * right. We'll push up to and including min_slot, but no lower
3043 */
3044static noinline int __push_leaf_right(struct btrfs_path *path,
3045 int data_size, int empty,
3046 struct extent_buffer *right,
3047 int free_space, u32 left_nritems,
3048 u32 min_slot)
3049{
3050 struct btrfs_fs_info *fs_info = right->fs_info;
3051 struct extent_buffer *left = path->nodes[0];
3052 struct extent_buffer *upper = path->nodes[1];
3053 struct btrfs_map_token token;
3054 struct btrfs_disk_key disk_key;
3055 int slot;
3056 u32 i;
3057 int push_space = 0;
3058 int push_items = 0;
3059 u32 nr;
3060 u32 right_nritems;
3061 u32 data_end;
3062 u32 this_item_size;
3063
3064 if (empty)
3065 nr = 0;
3066 else
3067 nr = max_t(u32, 1, min_slot);
3068
3069 if (path->slots[0] >= left_nritems)
3070 push_space += data_size;
3071
3072 slot = path->slots[1];
3073 i = left_nritems - 1;
3074 while (i >= nr) {
3075 if (!empty && push_items > 0) {
3076 if (path->slots[0] > i)
3077 break;
3078 if (path->slots[0] == i) {
3079 int space = btrfs_leaf_free_space(left);
3080
3081 if (space + push_space * 2 > free_space)
3082 break;
3083 }
3084 }
3085
3086 if (path->slots[0] == i)
3087 push_space += data_size;
3088
3089 this_item_size = btrfs_item_size(left, i);
3090 if (this_item_size + sizeof(struct btrfs_item) +
3091 push_space > free_space)
3092 break;
3093
3094 push_items++;
3095 push_space += this_item_size + sizeof(struct btrfs_item);
3096 if (i == 0)
3097 break;
3098 i--;
3099 }
3100
3101 if (push_items == 0)
3102 goto out_unlock;
3103
3104 WARN_ON(!empty && push_items == left_nritems);
3105
3106 /* push left to right */
3107 right_nritems = btrfs_header_nritems(right);
3108
3109 push_space = btrfs_item_data_end(left, left_nritems - push_items);
3110 push_space -= leaf_data_end(left);
3111
3112 /* make room in the right data area */
3113 data_end = leaf_data_end(right);
3114 memmove_leaf_data(right, data_end - push_space, data_end,
3115 BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3116
3117 /* copy from the left data area */
3118 copy_leaf_data(right, left, BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3119 leaf_data_end(left), push_space);
3120
3121 memmove_leaf_items(right, push_items, 0, right_nritems);
3122
3123 /* copy the items from left to right */
3124 copy_leaf_items(right, left, 0, left_nritems - push_items, push_items);
3125
3126 /* update the item pointers */
3127 btrfs_init_map_token(&token, right);
3128 right_nritems += push_items;
3129 btrfs_set_header_nritems(right, right_nritems);
3130 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3131 for (i = 0; i < right_nritems; i++) {
3132 push_space -= btrfs_token_item_size(&token, i);
3133 btrfs_set_token_item_offset(&token, i, push_space);
3134 }
3135
3136 left_nritems -= push_items;
3137 btrfs_set_header_nritems(left, left_nritems);
3138
3139 if (left_nritems)
3140 btrfs_mark_buffer_dirty(left);
3141 else
3142 btrfs_clean_tree_block(left);
3143
3144 btrfs_mark_buffer_dirty(right);
3145
3146 btrfs_item_key(right, &disk_key, 0);
3147 btrfs_set_node_key(upper, &disk_key, slot + 1);
3148 btrfs_mark_buffer_dirty(upper);
3149
3150 /* then fixup the leaf pointer in the path */
3151 if (path->slots[0] >= left_nritems) {
3152 path->slots[0] -= left_nritems;
3153 if (btrfs_header_nritems(path->nodes[0]) == 0)
3154 btrfs_clean_tree_block(path->nodes[0]);
3155 btrfs_tree_unlock(path->nodes[0]);
3156 free_extent_buffer(path->nodes[0]);
3157 path->nodes[0] = right;
3158 path->slots[1] += 1;
3159 } else {
3160 btrfs_tree_unlock(right);
3161 free_extent_buffer(right);
3162 }
3163 return 0;
3164
3165out_unlock:
3166 btrfs_tree_unlock(right);
3167 free_extent_buffer(right);
3168 return 1;
3169}
3170
3171/*
3172 * push some data in the path leaf to the right, trying to free up at
3173 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3174 *
3175 * returns 1 if the push failed because the other node didn't have enough
3176 * room, 0 if everything worked out and < 0 if there were major errors.
3177 *
3178 * this will push starting from min_slot to the end of the leaf. It won't
3179 * push any slot lower than min_slot
3180 */
3181static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3182 *root, struct btrfs_path *path,
3183 int min_data_size, int data_size,
3184 int empty, u32 min_slot)
3185{
3186 struct extent_buffer *left = path->nodes[0];
3187 struct extent_buffer *right;
3188 struct extent_buffer *upper;
3189 int slot;
3190 int free_space;
3191 u32 left_nritems;
3192 int ret;
3193
3194 if (!path->nodes[1])
3195 return 1;
3196
3197 slot = path->slots[1];
3198 upper = path->nodes[1];
3199 if (slot >= btrfs_header_nritems(upper) - 1)
3200 return 1;
3201
3202 btrfs_assert_tree_write_locked(path->nodes[1]);
3203
3204 right = btrfs_read_node_slot(upper, slot + 1);
3205 /*
3206 * slot + 1 is not valid or we fail to read the right node,
3207 * no big deal, just return.
3208 */
3209 if (IS_ERR(right))
3210 return 1;
3211
3212 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
3213
3214 free_space = btrfs_leaf_free_space(right);
3215 if (free_space < data_size)
3216 goto out_unlock;
3217
3218 ret = btrfs_cow_block(trans, root, right, upper,
3219 slot + 1, &right, BTRFS_NESTING_RIGHT_COW);
3220 if (ret)
3221 goto out_unlock;
3222
3223 left_nritems = btrfs_header_nritems(left);
3224 if (left_nritems == 0)
3225 goto out_unlock;
3226
3227 if (check_sibling_keys(left, right)) {
3228 ret = -EUCLEAN;
3229 btrfs_tree_unlock(right);
3230 free_extent_buffer(right);
3231 return ret;
3232 }
3233 if (path->slots[0] == left_nritems && !empty) {
3234 /* Key greater than all keys in the leaf, right neighbor has
3235 * enough room for it and we're not emptying our leaf to delete
3236 * it, therefore use right neighbor to insert the new item and
3237 * no need to touch/dirty our left leaf. */
3238 btrfs_tree_unlock(left);
3239 free_extent_buffer(left);
3240 path->nodes[0] = right;
3241 path->slots[0] = 0;
3242 path->slots[1]++;
3243 return 0;
3244 }
3245
3246 return __push_leaf_right(path, min_data_size, empty,
3247 right, free_space, left_nritems, min_slot);
3248out_unlock:
3249 btrfs_tree_unlock(right);
3250 free_extent_buffer(right);
3251 return 1;
3252}
3253
3254/*
3255 * push some data in the path leaf to the left, trying to free up at
3256 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3257 *
3258 * max_slot can put a limit on how far into the leaf we'll push items. The
3259 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3260 * items
3261 */
3262static noinline int __push_leaf_left(struct btrfs_path *path, int data_size,
3263 int empty, struct extent_buffer *left,
3264 int free_space, u32 right_nritems,
3265 u32 max_slot)
3266{
3267 struct btrfs_fs_info *fs_info = left->fs_info;
3268 struct btrfs_disk_key disk_key;
3269 struct extent_buffer *right = path->nodes[0];
3270 int i;
3271 int push_space = 0;
3272 int push_items = 0;
3273 u32 old_left_nritems;
3274 u32 nr;
3275 int ret = 0;
3276 u32 this_item_size;
3277 u32 old_left_item_size;
3278 struct btrfs_map_token token;
3279
3280 if (empty)
3281 nr = min(right_nritems, max_slot);
3282 else
3283 nr = min(right_nritems - 1, max_slot);
3284
3285 for (i = 0; i < nr; i++) {
3286 if (!empty && push_items > 0) {
3287 if (path->slots[0] < i)
3288 break;
3289 if (path->slots[0] == i) {
3290 int space = btrfs_leaf_free_space(right);
3291
3292 if (space + push_space * 2 > free_space)
3293 break;
3294 }
3295 }
3296
3297 if (path->slots[0] == i)
3298 push_space += data_size;
3299
3300 this_item_size = btrfs_item_size(right, i);
3301 if (this_item_size + sizeof(struct btrfs_item) + push_space >
3302 free_space)
3303 break;
3304
3305 push_items++;
3306 push_space += this_item_size + sizeof(struct btrfs_item);
3307 }
3308
3309 if (push_items == 0) {
3310 ret = 1;
3311 goto out;
3312 }
3313 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3314
3315 /* push data from right to left */
3316 copy_leaf_items(left, right, btrfs_header_nritems(left), 0, push_items);
3317
3318 push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3319 btrfs_item_offset(right, push_items - 1);
3320
3321 copy_leaf_data(left, right, leaf_data_end(left) - push_space,
3322 btrfs_item_offset(right, push_items - 1), push_space);
3323 old_left_nritems = btrfs_header_nritems(left);
3324 BUG_ON(old_left_nritems <= 0);
3325
3326 btrfs_init_map_token(&token, left);
3327 old_left_item_size = btrfs_item_offset(left, old_left_nritems - 1);
3328 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3329 u32 ioff;
3330
3331 ioff = btrfs_token_item_offset(&token, i);
3332 btrfs_set_token_item_offset(&token, i,
3333 ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
3334 }
3335 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3336
3337 /* fixup right node */
3338 if (push_items > right_nritems)
3339 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3340 right_nritems);
3341
3342 if (push_items < right_nritems) {
3343 push_space = btrfs_item_offset(right, push_items - 1) -
3344 leaf_data_end(right);
3345 memmove_leaf_data(right,
3346 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3347 leaf_data_end(right), push_space);
3348
3349 memmove_leaf_items(right, 0, push_items,
3350 btrfs_header_nritems(right) - push_items);
3351 }
3352
3353 btrfs_init_map_token(&token, right);
3354 right_nritems -= push_items;
3355 btrfs_set_header_nritems(right, right_nritems);
3356 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3357 for (i = 0; i < right_nritems; i++) {
3358 push_space = push_space - btrfs_token_item_size(&token, i);
3359 btrfs_set_token_item_offset(&token, i, push_space);
3360 }
3361
3362 btrfs_mark_buffer_dirty(left);
3363 if (right_nritems)
3364 btrfs_mark_buffer_dirty(right);
3365 else
3366 btrfs_clean_tree_block(right);
3367
3368 btrfs_item_key(right, &disk_key, 0);
3369 fixup_low_keys(path, &disk_key, 1);
3370
3371 /* then fixup the leaf pointer in the path */
3372 if (path->slots[0] < push_items) {
3373 path->slots[0] += old_left_nritems;
3374 btrfs_tree_unlock(path->nodes[0]);
3375 free_extent_buffer(path->nodes[0]);
3376 path->nodes[0] = left;
3377 path->slots[1] -= 1;
3378 } else {
3379 btrfs_tree_unlock(left);
3380 free_extent_buffer(left);
3381 path->slots[0] -= push_items;
3382 }
3383 BUG_ON(path->slots[0] < 0);
3384 return ret;
3385out:
3386 btrfs_tree_unlock(left);
3387 free_extent_buffer(left);
3388 return ret;
3389}
3390
3391/*
3392 * push some data in the path leaf to the left, trying to free up at
3393 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3394 *
3395 * max_slot can put a limit on how far into the leaf we'll push items. The
3396 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
3397 * items
3398 */
3399static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3400 *root, struct btrfs_path *path, int min_data_size,
3401 int data_size, int empty, u32 max_slot)
3402{
3403 struct extent_buffer *right = path->nodes[0];
3404 struct extent_buffer *left;
3405 int slot;
3406 int free_space;
3407 u32 right_nritems;
3408 int ret = 0;
3409
3410 slot = path->slots[1];
3411 if (slot == 0)
3412 return 1;
3413 if (!path->nodes[1])
3414 return 1;
3415
3416 right_nritems = btrfs_header_nritems(right);
3417 if (right_nritems == 0)
3418 return 1;
3419
3420 btrfs_assert_tree_write_locked(path->nodes[1]);
3421
3422 left = btrfs_read_node_slot(path->nodes[1], slot - 1);
3423 /*
3424 * slot - 1 is not valid or we fail to read the left node,
3425 * no big deal, just return.
3426 */
3427 if (IS_ERR(left))
3428 return 1;
3429
3430 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
3431
3432 free_space = btrfs_leaf_free_space(left);
3433 if (free_space < data_size) {
3434 ret = 1;
3435 goto out;
3436 }
3437
3438 ret = btrfs_cow_block(trans, root, left,
3439 path->nodes[1], slot - 1, &left,
3440 BTRFS_NESTING_LEFT_COW);
3441 if (ret) {
3442 /* we hit -ENOSPC, but it isn't fatal here */
3443 if (ret == -ENOSPC)
3444 ret = 1;
3445 goto out;
3446 }
3447
3448 if (check_sibling_keys(left, right)) {
3449 ret = -EUCLEAN;
3450 goto out;
3451 }
3452 return __push_leaf_left(path, min_data_size,
3453 empty, left, free_space, right_nritems,
3454 max_slot);
3455out:
3456 btrfs_tree_unlock(left);
3457 free_extent_buffer(left);
3458 return ret;
3459}
3460
3461/*
3462 * split the path's leaf in two, making sure there is at least data_size
3463 * available for the resulting leaf level of the path.
3464 */
3465static noinline void copy_for_split(struct btrfs_trans_handle *trans,
3466 struct btrfs_path *path,
3467 struct extent_buffer *l,
3468 struct extent_buffer *right,
3469 int slot, int mid, int nritems)
3470{
3471 struct btrfs_fs_info *fs_info = trans->fs_info;
3472 int data_copy_size;
3473 int rt_data_off;
3474 int i;
3475 struct btrfs_disk_key disk_key;
3476 struct btrfs_map_token token;
3477
3478 nritems = nritems - mid;
3479 btrfs_set_header_nritems(right, nritems);
3480 data_copy_size = btrfs_item_data_end(l, mid) - leaf_data_end(l);
3481
3482 copy_leaf_items(right, l, 0, mid, nritems);
3483
3484 copy_leaf_data(right, l, BTRFS_LEAF_DATA_SIZE(fs_info) - data_copy_size,
3485 leaf_data_end(l), data_copy_size);
3486
3487 rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_data_end(l, mid);
3488
3489 btrfs_init_map_token(&token, right);
3490 for (i = 0; i < nritems; i++) {
3491 u32 ioff;
3492
3493 ioff = btrfs_token_item_offset(&token, i);
3494 btrfs_set_token_item_offset(&token, i, ioff + rt_data_off);
3495 }
3496
3497 btrfs_set_header_nritems(l, mid);
3498 btrfs_item_key(right, &disk_key, 0);
3499 insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
3500
3501 btrfs_mark_buffer_dirty(right);
3502 btrfs_mark_buffer_dirty(l);
3503 BUG_ON(path->slots[0] != slot);
3504
3505 if (mid <= slot) {
3506 btrfs_tree_unlock(path->nodes[0]);
3507 free_extent_buffer(path->nodes[0]);
3508 path->nodes[0] = right;
3509 path->slots[0] -= mid;
3510 path->slots[1] += 1;
3511 } else {
3512 btrfs_tree_unlock(right);
3513 free_extent_buffer(right);
3514 }
3515
3516 BUG_ON(path->slots[0] < 0);
3517}
3518
3519/*
3520 * double splits happen when we need to insert a big item in the middle
3521 * of a leaf. A double split can leave us with 3 mostly empty leaves:
3522 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
3523 * A B C
3524 *
3525 * We avoid this by trying to push the items on either side of our target
3526 * into the adjacent leaves. If all goes well we can avoid the double split
3527 * completely.
3528 */
3529static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
3530 struct btrfs_root *root,
3531 struct btrfs_path *path,
3532 int data_size)
3533{
3534 int ret;
3535 int progress = 0;
3536 int slot;
3537 u32 nritems;
3538 int space_needed = data_size;
3539
3540 slot = path->slots[0];
3541 if (slot < btrfs_header_nritems(path->nodes[0]))
3542 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3543
3544 /*
3545 * try to push all the items after our slot into the
3546 * right leaf
3547 */
3548 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
3549 if (ret < 0)
3550 return ret;
3551
3552 if (ret == 0)
3553 progress++;
3554
3555 nritems = btrfs_header_nritems(path->nodes[0]);
3556 /*
3557 * our goal is to get our slot at the start or end of a leaf. If
3558 * we've done so we're done
3559 */
3560 if (path->slots[0] == 0 || path->slots[0] == nritems)
3561 return 0;
3562
3563 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3564 return 0;
3565
3566 /* try to push all the items before our slot into the next leaf */
3567 slot = path->slots[0];
3568 space_needed = data_size;
3569 if (slot > 0)
3570 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3571 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
3572 if (ret < 0)
3573 return ret;
3574
3575 if (ret == 0)
3576 progress++;
3577
3578 if (progress)
3579 return 0;
3580 return 1;
3581}
3582
3583/*
3584 * split the path's leaf in two, making sure there is at least data_size
3585 * available for the resulting leaf level of the path.
3586 *
3587 * returns 0 if all went well and < 0 on failure.
3588 */
3589static noinline int split_leaf(struct btrfs_trans_handle *trans,
3590 struct btrfs_root *root,
3591 const struct btrfs_key *ins_key,
3592 struct btrfs_path *path, int data_size,
3593 int extend)
3594{
3595 struct btrfs_disk_key disk_key;
3596 struct extent_buffer *l;
3597 u32 nritems;
3598 int mid;
3599 int slot;
3600 struct extent_buffer *right;
3601 struct btrfs_fs_info *fs_info = root->fs_info;
3602 int ret = 0;
3603 int wret;
3604 int split;
3605 int num_doubles = 0;
3606 int tried_avoid_double = 0;
3607
3608 l = path->nodes[0];
3609 slot = path->slots[0];
3610 if (extend && data_size + btrfs_item_size(l, slot) +
3611 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
3612 return -EOVERFLOW;
3613
3614 /* first try to make some room by pushing left and right */
3615 if (data_size && path->nodes[1]) {
3616 int space_needed = data_size;
3617
3618 if (slot < btrfs_header_nritems(l))
3619 space_needed -= btrfs_leaf_free_space(l);
3620
3621 wret = push_leaf_right(trans, root, path, space_needed,
3622 space_needed, 0, 0);
3623 if (wret < 0)
3624 return wret;
3625 if (wret) {
3626 space_needed = data_size;
3627 if (slot > 0)
3628 space_needed -= btrfs_leaf_free_space(l);
3629 wret = push_leaf_left(trans, root, path, space_needed,
3630 space_needed, 0, (u32)-1);
3631 if (wret < 0)
3632 return wret;
3633 }
3634 l = path->nodes[0];
3635
3636 /* did the pushes work? */
3637 if (btrfs_leaf_free_space(l) >= data_size)
3638 return 0;
3639 }
3640
3641 if (!path->nodes[1]) {
3642 ret = insert_new_root(trans, root, path, 1);
3643 if (ret)
3644 return ret;
3645 }
3646again:
3647 split = 1;
3648 l = path->nodes[0];
3649 slot = path->slots[0];
3650 nritems = btrfs_header_nritems(l);
3651 mid = (nritems + 1) / 2;
3652
3653 if (mid <= slot) {
3654 if (nritems == 1 ||
3655 leaf_space_used(l, mid, nritems - mid) + data_size >
3656 BTRFS_LEAF_DATA_SIZE(fs_info)) {
3657 if (slot >= nritems) {
3658 split = 0;
3659 } else {
3660 mid = slot;
3661 if (mid != nritems &&
3662 leaf_space_used(l, mid, nritems - mid) +
3663 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3664 if (data_size && !tried_avoid_double)
3665 goto push_for_double;
3666 split = 2;
3667 }
3668 }
3669 }
3670 } else {
3671 if (leaf_space_used(l, 0, mid) + data_size >
3672 BTRFS_LEAF_DATA_SIZE(fs_info)) {
3673 if (!extend && data_size && slot == 0) {
3674 split = 0;
3675 } else if ((extend || !data_size) && slot == 0) {
3676 mid = 1;
3677 } else {
3678 mid = slot;
3679 if (mid != nritems &&
3680 leaf_space_used(l, mid, nritems - mid) +
3681 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3682 if (data_size && !tried_avoid_double)
3683 goto push_for_double;
3684 split = 2;
3685 }
3686 }
3687 }
3688 }
3689
3690 if (split == 0)
3691 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3692 else
3693 btrfs_item_key(l, &disk_key, mid);
3694
3695 /*
3696 * We have to about BTRFS_NESTING_NEW_ROOT here if we've done a double
3697 * split, because we're only allowed to have MAX_LOCKDEP_SUBCLASSES
3698 * subclasses, which is 8 at the time of this patch, and we've maxed it
3699 * out. In the future we could add a
3700 * BTRFS_NESTING_SPLIT_THE_SPLITTENING if we need to, but for now just
3701 * use BTRFS_NESTING_NEW_ROOT.
3702 */
3703 right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3704 &disk_key, 0, l->start, 0,
3705 num_doubles ? BTRFS_NESTING_NEW_ROOT :
3706 BTRFS_NESTING_SPLIT);
3707 if (IS_ERR(right))
3708 return PTR_ERR(right);
3709
3710 root_add_used(root, fs_info->nodesize);
3711
3712 if (split == 0) {
3713 if (mid <= slot) {
3714 btrfs_set_header_nritems(right, 0);
3715 insert_ptr(trans, path, &disk_key,
3716 right->start, path->slots[1] + 1, 1);
3717 btrfs_tree_unlock(path->nodes[0]);
3718 free_extent_buffer(path->nodes[0]);
3719 path->nodes[0] = right;
3720 path->slots[0] = 0;
3721 path->slots[1] += 1;
3722 } else {
3723 btrfs_set_header_nritems(right, 0);
3724 insert_ptr(trans, path, &disk_key,
3725 right->start, path->slots[1], 1);
3726 btrfs_tree_unlock(path->nodes[0]);
3727 free_extent_buffer(path->nodes[0]);
3728 path->nodes[0] = right;
3729 path->slots[0] = 0;
3730 if (path->slots[1] == 0)
3731 fixup_low_keys(path, &disk_key, 1);
3732 }
3733 /*
3734 * We create a new leaf 'right' for the required ins_len and
3735 * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
3736 * the content of ins_len to 'right'.
3737 */
3738 return ret;
3739 }
3740
3741 copy_for_split(trans, path, l, right, slot, mid, nritems);
3742
3743 if (split == 2) {
3744 BUG_ON(num_doubles != 0);
3745 num_doubles++;
3746 goto again;
3747 }
3748
3749 return 0;
3750
3751push_for_double:
3752 push_for_double_split(trans, root, path, data_size);
3753 tried_avoid_double = 1;
3754 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3755 return 0;
3756 goto again;
3757}
3758
3759static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3760 struct btrfs_root *root,
3761 struct btrfs_path *path, int ins_len)
3762{
3763 struct btrfs_key key;
3764 struct extent_buffer *leaf;
3765 struct btrfs_file_extent_item *fi;
3766 u64 extent_len = 0;
3767 u32 item_size;
3768 int ret;
3769
3770 leaf = path->nodes[0];
3771 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3772
3773 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3774 key.type != BTRFS_EXTENT_CSUM_KEY);
3775
3776 if (btrfs_leaf_free_space(leaf) >= ins_len)
3777 return 0;
3778
3779 item_size = btrfs_item_size(leaf, path->slots[0]);
3780 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3781 fi = btrfs_item_ptr(leaf, path->slots[0],
3782 struct btrfs_file_extent_item);
3783 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3784 }
3785 btrfs_release_path(path);
3786
3787 path->keep_locks = 1;
3788 path->search_for_split = 1;
3789 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3790 path->search_for_split = 0;
3791 if (ret > 0)
3792 ret = -EAGAIN;
3793 if (ret < 0)
3794 goto err;
3795
3796 ret = -EAGAIN;
3797 leaf = path->nodes[0];
3798 /* if our item isn't there, return now */
3799 if (item_size != btrfs_item_size(leaf, path->slots[0]))
3800 goto err;
3801
3802 /* the leaf has changed, it now has room. return now */
3803 if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
3804 goto err;
3805
3806 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3807 fi = btrfs_item_ptr(leaf, path->slots[0],
3808 struct btrfs_file_extent_item);
3809 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3810 goto err;
3811 }
3812
3813 ret = split_leaf(trans, root, &key, path, ins_len, 1);
3814 if (ret)
3815 goto err;
3816
3817 path->keep_locks = 0;
3818 btrfs_unlock_up_safe(path, 1);
3819 return 0;
3820err:
3821 path->keep_locks = 0;
3822 return ret;
3823}
3824
3825static noinline int split_item(struct btrfs_path *path,
3826 const struct btrfs_key *new_key,
3827 unsigned long split_offset)
3828{
3829 struct extent_buffer *leaf;
3830 int orig_slot, slot;
3831 char *buf;
3832 u32 nritems;
3833 u32 item_size;
3834 u32 orig_offset;
3835 struct btrfs_disk_key disk_key;
3836
3837 leaf = path->nodes[0];
3838 BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
3839
3840 orig_slot = path->slots[0];
3841 orig_offset = btrfs_item_offset(leaf, path->slots[0]);
3842 item_size = btrfs_item_size(leaf, path->slots[0]);
3843
3844 buf = kmalloc(item_size, GFP_NOFS);
3845 if (!buf)
3846 return -ENOMEM;
3847
3848 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3849 path->slots[0]), item_size);
3850
3851 slot = path->slots[0] + 1;
3852 nritems = btrfs_header_nritems(leaf);
3853 if (slot != nritems) {
3854 /* shift the items */
3855 memmove_leaf_items(leaf, slot + 1, slot, nritems - slot);
3856 }
3857
3858 btrfs_cpu_key_to_disk(&disk_key, new_key);
3859 btrfs_set_item_key(leaf, &disk_key, slot);
3860
3861 btrfs_set_item_offset(leaf, slot, orig_offset);
3862 btrfs_set_item_size(leaf, slot, item_size - split_offset);
3863
3864 btrfs_set_item_offset(leaf, orig_slot,
3865 orig_offset + item_size - split_offset);
3866 btrfs_set_item_size(leaf, orig_slot, split_offset);
3867
3868 btrfs_set_header_nritems(leaf, nritems + 1);
3869
3870 /* write the data for the start of the original item */
3871 write_extent_buffer(leaf, buf,
3872 btrfs_item_ptr_offset(leaf, path->slots[0]),
3873 split_offset);
3874
3875 /* write the data for the new item */
3876 write_extent_buffer(leaf, buf + split_offset,
3877 btrfs_item_ptr_offset(leaf, slot),
3878 item_size - split_offset);
3879 btrfs_mark_buffer_dirty(leaf);
3880
3881 BUG_ON(btrfs_leaf_free_space(leaf) < 0);
3882 kfree(buf);
3883 return 0;
3884}
3885
3886/*
3887 * This function splits a single item into two items,
3888 * giving 'new_key' to the new item and splitting the
3889 * old one at split_offset (from the start of the item).
3890 *
3891 * The path may be released by this operation. After
3892 * the split, the path is pointing to the old item. The
3893 * new item is going to be in the same node as the old one.
3894 *
3895 * Note, the item being split must be smaller enough to live alone on
3896 * a tree block with room for one extra struct btrfs_item
3897 *
3898 * This allows us to split the item in place, keeping a lock on the
3899 * leaf the entire time.
3900 */
3901int btrfs_split_item(struct btrfs_trans_handle *trans,
3902 struct btrfs_root *root,
3903 struct btrfs_path *path,
3904 const struct btrfs_key *new_key,
3905 unsigned long split_offset)
3906{
3907 int ret;
3908 ret = setup_leaf_for_split(trans, root, path,
3909 sizeof(struct btrfs_item));
3910 if (ret)
3911 return ret;
3912
3913 ret = split_item(path, new_key, split_offset);
3914 return ret;
3915}
3916
3917/*
3918 * make the item pointed to by the path smaller. new_size indicates
3919 * how small to make it, and from_end tells us if we just chop bytes
3920 * off the end of the item or if we shift the item to chop bytes off
3921 * the front.
3922 */
3923void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
3924{
3925 int slot;
3926 struct extent_buffer *leaf;
3927 u32 nritems;
3928 unsigned int data_end;
3929 unsigned int old_data_start;
3930 unsigned int old_size;
3931 unsigned int size_diff;
3932 int i;
3933 struct btrfs_map_token token;
3934
3935 leaf = path->nodes[0];
3936 slot = path->slots[0];
3937
3938 old_size = btrfs_item_size(leaf, slot);
3939 if (old_size == new_size)
3940 return;
3941
3942 nritems = btrfs_header_nritems(leaf);
3943 data_end = leaf_data_end(leaf);
3944
3945 old_data_start = btrfs_item_offset(leaf, slot);
3946
3947 size_diff = old_size - new_size;
3948
3949 BUG_ON(slot < 0);
3950 BUG_ON(slot >= nritems);
3951
3952 /*
3953 * item0..itemN ... dataN.offset..dataN.size .. data0.size
3954 */
3955 /* first correct the data pointers */
3956 btrfs_init_map_token(&token, leaf);
3957 for (i = slot; i < nritems; i++) {
3958 u32 ioff;
3959
3960 ioff = btrfs_token_item_offset(&token, i);
3961 btrfs_set_token_item_offset(&token, i, ioff + size_diff);
3962 }
3963
3964 /* shift the data */
3965 if (from_end) {
3966 memmove_leaf_data(leaf, data_end + size_diff, data_end,
3967 old_data_start + new_size - data_end);
3968 } else {
3969 struct btrfs_disk_key disk_key;
3970 u64 offset;
3971
3972 btrfs_item_key(leaf, &disk_key, slot);
3973
3974 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3975 unsigned long ptr;
3976 struct btrfs_file_extent_item *fi;
3977
3978 fi = btrfs_item_ptr(leaf, slot,
3979 struct btrfs_file_extent_item);
3980 fi = (struct btrfs_file_extent_item *)(
3981 (unsigned long)fi - size_diff);
3982
3983 if (btrfs_file_extent_type(leaf, fi) ==
3984 BTRFS_FILE_EXTENT_INLINE) {
3985 ptr = btrfs_item_ptr_offset(leaf, slot);
3986 memmove_extent_buffer(leaf, ptr,
3987 (unsigned long)fi,
3988 BTRFS_FILE_EXTENT_INLINE_DATA_START);
3989 }
3990 }
3991
3992 memmove_leaf_data(leaf, data_end + size_diff, data_end,
3993 old_data_start - data_end);
3994
3995 offset = btrfs_disk_key_offset(&disk_key);
3996 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3997 btrfs_set_item_key(leaf, &disk_key, slot);
3998 if (slot == 0)
3999 fixup_low_keys(path, &disk_key, 1);
4000 }
4001
4002 btrfs_set_item_size(leaf, slot, new_size);
4003 btrfs_mark_buffer_dirty(leaf);
4004
4005 if (btrfs_leaf_free_space(leaf) < 0) {
4006 btrfs_print_leaf(leaf);
4007 BUG();
4008 }
4009}
4010
4011/*
4012 * make the item pointed to by the path bigger, data_size is the added size.
4013 */
4014void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
4015{
4016 int slot;
4017 struct extent_buffer *leaf;
4018 u32 nritems;
4019 unsigned int data_end;
4020 unsigned int old_data;
4021 unsigned int old_size;
4022 int i;
4023 struct btrfs_map_token token;
4024
4025 leaf = path->nodes[0];
4026
4027 nritems = btrfs_header_nritems(leaf);
4028 data_end = leaf_data_end(leaf);
4029
4030 if (btrfs_leaf_free_space(leaf) < data_size) {
4031 btrfs_print_leaf(leaf);
4032 BUG();
4033 }
4034 slot = path->slots[0];
4035 old_data = btrfs_item_data_end(leaf, slot);
4036
4037 BUG_ON(slot < 0);
4038 if (slot >= nritems) {
4039 btrfs_print_leaf(leaf);
4040 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
4041 slot, nritems);
4042 BUG();
4043 }
4044
4045 /*
4046 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4047 */
4048 /* first correct the data pointers */
4049 btrfs_init_map_token(&token, leaf);
4050 for (i = slot; i < nritems; i++) {
4051 u32 ioff;
4052
4053 ioff = btrfs_token_item_offset(&token, i);
4054 btrfs_set_token_item_offset(&token, i, ioff - data_size);
4055 }
4056
4057 /* shift the data */
4058 memmove_leaf_data(leaf, data_end - data_size, data_end,
4059 old_data - data_end);
4060
4061 data_end = old_data;
4062 old_size = btrfs_item_size(leaf, slot);
4063 btrfs_set_item_size(leaf, slot, old_size + data_size);
4064 btrfs_mark_buffer_dirty(leaf);
4065
4066 if (btrfs_leaf_free_space(leaf) < 0) {
4067 btrfs_print_leaf(leaf);
4068 BUG();
4069 }
4070}
4071
4072/*
4073 * Make space in the node before inserting one or more items.
4074 *
4075 * @root: root we are inserting items to
4076 * @path: points to the leaf/slot where we are going to insert new items
4077 * @batch: information about the batch of items to insert
4078 *
4079 * Main purpose is to save stack depth by doing the bulk of the work in a
4080 * function that doesn't call btrfs_search_slot
4081 */
4082static void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4083 const struct btrfs_item_batch *batch)
4084{
4085 struct btrfs_fs_info *fs_info = root->fs_info;
4086 int i;
4087 u32 nritems;
4088 unsigned int data_end;
4089 struct btrfs_disk_key disk_key;
4090 struct extent_buffer *leaf;
4091 int slot;
4092 struct btrfs_map_token token;
4093 u32 total_size;
4094
4095 /*
4096 * Before anything else, update keys in the parent and other ancestors
4097 * if needed, then release the write locks on them, so that other tasks
4098 * can use them while we modify the leaf.
4099 */
4100 if (path->slots[0] == 0) {
4101 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[0]);
4102 fixup_low_keys(path, &disk_key, 1);
4103 }
4104 btrfs_unlock_up_safe(path, 1);
4105
4106 leaf = path->nodes[0];
4107 slot = path->slots[0];
4108
4109 nritems = btrfs_header_nritems(leaf);
4110 data_end = leaf_data_end(leaf);
4111 total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4112
4113 if (btrfs_leaf_free_space(leaf) < total_size) {
4114 btrfs_print_leaf(leaf);
4115 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4116 total_size, btrfs_leaf_free_space(leaf));
4117 BUG();
4118 }
4119
4120 btrfs_init_map_token(&token, leaf);
4121 if (slot != nritems) {
4122 unsigned int old_data = btrfs_item_data_end(leaf, slot);
4123
4124 if (old_data < data_end) {
4125 btrfs_print_leaf(leaf);
4126 btrfs_crit(fs_info,
4127 "item at slot %d with data offset %u beyond data end of leaf %u",
4128 slot, old_data, data_end);
4129 BUG();
4130 }
4131 /*
4132 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4133 */
4134 /* first correct the data pointers */
4135 for (i = slot; i < nritems; i++) {
4136 u32 ioff;
4137
4138 ioff = btrfs_token_item_offset(&token, i);
4139 btrfs_set_token_item_offset(&token, i,
4140 ioff - batch->total_data_size);
4141 }
4142 /* shift the items */
4143 memmove_leaf_items(leaf, slot + batch->nr, slot, nritems - slot);
4144
4145 /* shift the data */
4146 memmove_leaf_data(leaf, data_end - batch->total_data_size,
4147 data_end, old_data - data_end);
4148 data_end = old_data;
4149 }
4150
4151 /* setup the item for the new data */
4152 for (i = 0; i < batch->nr; i++) {
4153 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[i]);
4154 btrfs_set_item_key(leaf, &disk_key, slot + i);
4155 data_end -= batch->data_sizes[i];
4156 btrfs_set_token_item_offset(&token, slot + i, data_end);
4157 btrfs_set_token_item_size(&token, slot + i, batch->data_sizes[i]);
4158 }
4159
4160 btrfs_set_header_nritems(leaf, nritems + batch->nr);
4161 btrfs_mark_buffer_dirty(leaf);
4162
4163 if (btrfs_leaf_free_space(leaf) < 0) {
4164 btrfs_print_leaf(leaf);
4165 BUG();
4166 }
4167}
4168
4169/*
4170 * Insert a new item into a leaf.
4171 *
4172 * @root: The root of the btree.
4173 * @path: A path pointing to the target leaf and slot.
4174 * @key: The key of the new item.
4175 * @data_size: The size of the data associated with the new key.
4176 */
4177void btrfs_setup_item_for_insert(struct btrfs_root *root,
4178 struct btrfs_path *path,
4179 const struct btrfs_key *key,
4180 u32 data_size)
4181{
4182 struct btrfs_item_batch batch;
4183
4184 batch.keys = key;
4185 batch.data_sizes = &data_size;
4186 batch.total_data_size = data_size;
4187 batch.nr = 1;
4188
4189 setup_items_for_insert(root, path, &batch);
4190}
4191
4192/*
4193 * Given a key and some data, insert items into the tree.
4194 * This does all the path init required, making room in the tree if needed.
4195 */
4196int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4197 struct btrfs_root *root,
4198 struct btrfs_path *path,
4199 const struct btrfs_item_batch *batch)
4200{
4201 int ret = 0;
4202 int slot;
4203 u32 total_size;
4204
4205 total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4206 ret = btrfs_search_slot(trans, root, &batch->keys[0], path, total_size, 1);
4207 if (ret == 0)
4208 return -EEXIST;
4209 if (ret < 0)
4210 return ret;
4211
4212 slot = path->slots[0];
4213 BUG_ON(slot < 0);
4214
4215 setup_items_for_insert(root, path, batch);
4216 return 0;
4217}
4218
4219/*
4220 * Given a key and some data, insert an item into the tree.
4221 * This does all the path init required, making room in the tree if needed.
4222 */
4223int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4224 const struct btrfs_key *cpu_key, void *data,
4225 u32 data_size)
4226{
4227 int ret = 0;
4228 struct btrfs_path *path;
4229 struct extent_buffer *leaf;
4230 unsigned long ptr;
4231
4232 path = btrfs_alloc_path();
4233 if (!path)
4234 return -ENOMEM;
4235 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4236 if (!ret) {
4237 leaf = path->nodes[0];
4238 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4239 write_extent_buffer(leaf, data, ptr, data_size);
4240 btrfs_mark_buffer_dirty(leaf);
4241 }
4242 btrfs_free_path(path);
4243 return ret;
4244}
4245
4246/*
4247 * This function duplicates an item, giving 'new_key' to the new item.
4248 * It guarantees both items live in the same tree leaf and the new item is
4249 * contiguous with the original item.
4250 *
4251 * This allows us to split a file extent in place, keeping a lock on the leaf
4252 * the entire time.
4253 */
4254int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4255 struct btrfs_root *root,
4256 struct btrfs_path *path,
4257 const struct btrfs_key *new_key)
4258{
4259 struct extent_buffer *leaf;
4260 int ret;
4261 u32 item_size;
4262
4263 leaf = path->nodes[0];
4264 item_size = btrfs_item_size(leaf, path->slots[0]);
4265 ret = setup_leaf_for_split(trans, root, path,
4266 item_size + sizeof(struct btrfs_item));
4267 if (ret)
4268 return ret;
4269
4270 path->slots[0]++;
4271 btrfs_setup_item_for_insert(root, path, new_key, item_size);
4272 leaf = path->nodes[0];
4273 memcpy_extent_buffer(leaf,
4274 btrfs_item_ptr_offset(leaf, path->slots[0]),
4275 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4276 item_size);
4277 return 0;
4278}
4279
4280/*
4281 * delete the pointer from a given node.
4282 *
4283 * the tree should have been previously balanced so the deletion does not
4284 * empty a node.
4285 */
4286static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4287 int level, int slot)
4288{
4289 struct extent_buffer *parent = path->nodes[level];
4290 u32 nritems;
4291 int ret;
4292
4293 nritems = btrfs_header_nritems(parent);
4294 if (slot != nritems - 1) {
4295 if (level) {
4296 ret = btrfs_tree_mod_log_insert_move(parent, slot,
4297 slot + 1, nritems - slot - 1);
4298 BUG_ON(ret < 0);
4299 }
4300 memmove_extent_buffer(parent,
4301 btrfs_node_key_ptr_offset(parent, slot),
4302 btrfs_node_key_ptr_offset(parent, slot + 1),
4303 sizeof(struct btrfs_key_ptr) *
4304 (nritems - slot - 1));
4305 } else if (level) {
4306 ret = btrfs_tree_mod_log_insert_key(parent, slot,
4307 BTRFS_MOD_LOG_KEY_REMOVE);
4308 BUG_ON(ret < 0);
4309 }
4310
4311 nritems--;
4312 btrfs_set_header_nritems(parent, nritems);
4313 if (nritems == 0 && parent == root->node) {
4314 BUG_ON(btrfs_header_level(root->node) != 1);
4315 /* just turn the root into a leaf and break */
4316 btrfs_set_header_level(root->node, 0);
4317 } else if (slot == 0) {
4318 struct btrfs_disk_key disk_key;
4319
4320 btrfs_node_key(parent, &disk_key, 0);
4321 fixup_low_keys(path, &disk_key, level + 1);
4322 }
4323 btrfs_mark_buffer_dirty(parent);
4324}
4325
4326/*
4327 * a helper function to delete the leaf pointed to by path->slots[1] and
4328 * path->nodes[1].
4329 *
4330 * This deletes the pointer in path->nodes[1] and frees the leaf
4331 * block extent. zero is returned if it all worked out, < 0 otherwise.
4332 *
4333 * The path must have already been setup for deleting the leaf, including
4334 * all the proper balancing. path->nodes[1] must be locked.
4335 */
4336static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4337 struct btrfs_root *root,
4338 struct btrfs_path *path,
4339 struct extent_buffer *leaf)
4340{
4341 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4342 del_ptr(root, path, 1, path->slots[1]);
4343
4344 /*
4345 * btrfs_free_extent is expensive, we want to make sure we
4346 * aren't holding any locks when we call it
4347 */
4348 btrfs_unlock_up_safe(path, 0);
4349
4350 root_sub_used(root, leaf->len);
4351
4352 atomic_inc(&leaf->refs);
4353 btrfs_free_tree_block(trans, btrfs_root_id(root), leaf, 0, 1);
4354 free_extent_buffer_stale(leaf);
4355}
4356/*
4357 * delete the item at the leaf level in path. If that empties
4358 * the leaf, remove it from the tree
4359 */
4360int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4361 struct btrfs_path *path, int slot, int nr)
4362{
4363 struct btrfs_fs_info *fs_info = root->fs_info;
4364 struct extent_buffer *leaf;
4365 int ret = 0;
4366 int wret;
4367 u32 nritems;
4368
4369 leaf = path->nodes[0];
4370 nritems = btrfs_header_nritems(leaf);
4371
4372 if (slot + nr != nritems) {
4373 const u32 last_off = btrfs_item_offset(leaf, slot + nr - 1);
4374 const int data_end = leaf_data_end(leaf);
4375 struct btrfs_map_token token;
4376 u32 dsize = 0;
4377 int i;
4378
4379 for (i = 0; i < nr; i++)
4380 dsize += btrfs_item_size(leaf, slot + i);
4381
4382 memmove_leaf_data(leaf, data_end + dsize, data_end,
4383 last_off - data_end);
4384
4385 btrfs_init_map_token(&token, leaf);
4386 for (i = slot + nr; i < nritems; i++) {
4387 u32 ioff;
4388
4389 ioff = btrfs_token_item_offset(&token, i);
4390 btrfs_set_token_item_offset(&token, i, ioff + dsize);
4391 }
4392
4393 memmove_leaf_items(leaf, slot, slot + nr, nritems - slot - nr);
4394 }
4395 btrfs_set_header_nritems(leaf, nritems - nr);
4396 nritems -= nr;
4397
4398 /* delete the leaf if we've emptied it */
4399 if (nritems == 0) {
4400 if (leaf == root->node) {
4401 btrfs_set_header_level(leaf, 0);
4402 } else {
4403 btrfs_clean_tree_block(leaf);
4404 btrfs_del_leaf(trans, root, path, leaf);
4405 }
4406 } else {
4407 int used = leaf_space_used(leaf, 0, nritems);
4408 if (slot == 0) {
4409 struct btrfs_disk_key disk_key;
4410
4411 btrfs_item_key(leaf, &disk_key, 0);
4412 fixup_low_keys(path, &disk_key, 1);
4413 }
4414
4415 /*
4416 * Try to delete the leaf if it is mostly empty. We do this by
4417 * trying to move all its items into its left and right neighbours.
4418 * If we can't move all the items, then we don't delete it - it's
4419 * not ideal, but future insertions might fill the leaf with more
4420 * items, or items from other leaves might be moved later into our
4421 * leaf due to deletions on those leaves.
4422 */
4423 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4424 u32 min_push_space;
4425
4426 /* push_leaf_left fixes the path.
4427 * make sure the path still points to our leaf
4428 * for possible call to del_ptr below
4429 */
4430 slot = path->slots[1];
4431 atomic_inc(&leaf->refs);
4432 /*
4433 * We want to be able to at least push one item to the
4434 * left neighbour leaf, and that's the first item.
4435 */
4436 min_push_space = sizeof(struct btrfs_item) +
4437 btrfs_item_size(leaf, 0);
4438 wret = push_leaf_left(trans, root, path, 0,
4439 min_push_space, 1, (u32)-1);
4440 if (wret < 0 && wret != -ENOSPC)
4441 ret = wret;
4442
4443 if (path->nodes[0] == leaf &&
4444 btrfs_header_nritems(leaf)) {
4445 /*
4446 * If we were not able to push all items from our
4447 * leaf to its left neighbour, then attempt to
4448 * either push all the remaining items to the
4449 * right neighbour or none. There's no advantage
4450 * in pushing only some items, instead of all, as
4451 * it's pointless to end up with a leaf having
4452 * too few items while the neighbours can be full
4453 * or nearly full.
4454 */
4455 nritems = btrfs_header_nritems(leaf);
4456 min_push_space = leaf_space_used(leaf, 0, nritems);
4457 wret = push_leaf_right(trans, root, path, 0,
4458 min_push_space, 1, 0);
4459 if (wret < 0 && wret != -ENOSPC)
4460 ret = wret;
4461 }
4462
4463 if (btrfs_header_nritems(leaf) == 0) {
4464 path->slots[1] = slot;
4465 btrfs_del_leaf(trans, root, path, leaf);
4466 free_extent_buffer(leaf);
4467 ret = 0;
4468 } else {
4469 /* if we're still in the path, make sure
4470 * we're dirty. Otherwise, one of the
4471 * push_leaf functions must have already
4472 * dirtied this buffer
4473 */
4474 if (path->nodes[0] == leaf)
4475 btrfs_mark_buffer_dirty(leaf);
4476 free_extent_buffer(leaf);
4477 }
4478 } else {
4479 btrfs_mark_buffer_dirty(leaf);
4480 }
4481 }
4482 return ret;
4483}
4484
4485/*
4486 * search the tree again to find a leaf with lesser keys
4487 * returns 0 if it found something or 1 if there are no lesser leaves.
4488 * returns < 0 on io errors.
4489 *
4490 * This may release the path, and so you may lose any locks held at the
4491 * time you call it.
4492 */
4493int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
4494{
4495 struct btrfs_key key;
4496 struct btrfs_disk_key found_key;
4497 int ret;
4498
4499 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
4500
4501 if (key.offset > 0) {
4502 key.offset--;
4503 } else if (key.type > 0) {
4504 key.type--;
4505 key.offset = (u64)-1;
4506 } else if (key.objectid > 0) {
4507 key.objectid--;
4508 key.type = (u8)-1;
4509 key.offset = (u64)-1;
4510 } else {
4511 return 1;
4512 }
4513
4514 btrfs_release_path(path);
4515 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4516 if (ret < 0)
4517 return ret;
4518 btrfs_item_key(path->nodes[0], &found_key, 0);
4519 ret = comp_keys(&found_key, &key);
4520 /*
4521 * We might have had an item with the previous key in the tree right
4522 * before we released our path. And after we released our path, that
4523 * item might have been pushed to the first slot (0) of the leaf we
4524 * were holding due to a tree balance. Alternatively, an item with the
4525 * previous key can exist as the only element of a leaf (big fat item).
4526 * Therefore account for these 2 cases, so that our callers (like
4527 * btrfs_previous_item) don't miss an existing item with a key matching
4528 * the previous key we computed above.
4529 */
4530 if (ret <= 0)
4531 return 0;
4532 return 1;
4533}
4534
4535/*
4536 * A helper function to walk down the tree starting at min_key, and looking
4537 * for nodes or leaves that are have a minimum transaction id.
4538 * This is used by the btree defrag code, and tree logging
4539 *
4540 * This does not cow, but it does stuff the starting key it finds back
4541 * into min_key, so you can call btrfs_search_slot with cow=1 on the
4542 * key and get a writable path.
4543 *
4544 * This honors path->lowest_level to prevent descent past a given level
4545 * of the tree.
4546 *
4547 * min_trans indicates the oldest transaction that you are interested
4548 * in walking through. Any nodes or leaves older than min_trans are
4549 * skipped over (without reading them).
4550 *
4551 * returns zero if something useful was found, < 0 on error and 1 if there
4552 * was nothing in the tree that matched the search criteria.
4553 */
4554int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4555 struct btrfs_path *path,
4556 u64 min_trans)
4557{
4558 struct extent_buffer *cur;
4559 struct btrfs_key found_key;
4560 int slot;
4561 int sret;
4562 u32 nritems;
4563 int level;
4564 int ret = 1;
4565 int keep_locks = path->keep_locks;
4566
4567 ASSERT(!path->nowait);
4568 path->keep_locks = 1;
4569again:
4570 cur = btrfs_read_lock_root_node(root);
4571 level = btrfs_header_level(cur);
4572 WARN_ON(path->nodes[level]);
4573 path->nodes[level] = cur;
4574 path->locks[level] = BTRFS_READ_LOCK;
4575
4576 if (btrfs_header_generation(cur) < min_trans) {
4577 ret = 1;
4578 goto out;
4579 }
4580 while (1) {
4581 nritems = btrfs_header_nritems(cur);
4582 level = btrfs_header_level(cur);
4583 sret = btrfs_bin_search(cur, min_key, &slot);
4584 if (sret < 0) {
4585 ret = sret;
4586 goto out;
4587 }
4588
4589 /* at the lowest level, we're done, setup the path and exit */
4590 if (level == path->lowest_level) {
4591 if (slot >= nritems)
4592 goto find_next_key;
4593 ret = 0;
4594 path->slots[level] = slot;
4595 btrfs_item_key_to_cpu(cur, &found_key, slot);
4596 goto out;
4597 }
4598 if (sret && slot > 0)
4599 slot--;
4600 /*
4601 * check this node pointer against the min_trans parameters.
4602 * If it is too old, skip to the next one.
4603 */
4604 while (slot < nritems) {
4605 u64 gen;
4606
4607 gen = btrfs_node_ptr_generation(cur, slot);
4608 if (gen < min_trans) {
4609 slot++;
4610 continue;
4611 }
4612 break;
4613 }
4614find_next_key:
4615 /*
4616 * we didn't find a candidate key in this node, walk forward
4617 * and find another one
4618 */
4619 if (slot >= nritems) {
4620 path->slots[level] = slot;
4621 sret = btrfs_find_next_key(root, path, min_key, level,
4622 min_trans);
4623 if (sret == 0) {
4624 btrfs_release_path(path);
4625 goto again;
4626 } else {
4627 goto out;
4628 }
4629 }
4630 /* save our key for returning back */
4631 btrfs_node_key_to_cpu(cur, &found_key, slot);
4632 path->slots[level] = slot;
4633 if (level == path->lowest_level) {
4634 ret = 0;
4635 goto out;
4636 }
4637 cur = btrfs_read_node_slot(cur, slot);
4638 if (IS_ERR(cur)) {
4639 ret = PTR_ERR(cur);
4640 goto out;
4641 }
4642
4643 btrfs_tree_read_lock(cur);
4644
4645 path->locks[level - 1] = BTRFS_READ_LOCK;
4646 path->nodes[level - 1] = cur;
4647 unlock_up(path, level, 1, 0, NULL);
4648 }
4649out:
4650 path->keep_locks = keep_locks;
4651 if (ret == 0) {
4652 btrfs_unlock_up_safe(path, path->lowest_level + 1);
4653 memcpy(min_key, &found_key, sizeof(found_key));
4654 }
4655 return ret;
4656}
4657
4658/*
4659 * this is similar to btrfs_next_leaf, but does not try to preserve
4660 * and fixup the path. It looks for and returns the next key in the
4661 * tree based on the current path and the min_trans parameters.
4662 *
4663 * 0 is returned if another key is found, < 0 if there are any errors
4664 * and 1 is returned if there are no higher keys in the tree
4665 *
4666 * path->keep_locks should be set to 1 on the search made before
4667 * calling this function.
4668 */
4669int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4670 struct btrfs_key *key, int level, u64 min_trans)
4671{
4672 int slot;
4673 struct extent_buffer *c;
4674
4675 WARN_ON(!path->keep_locks && !path->skip_locking);
4676 while (level < BTRFS_MAX_LEVEL) {
4677 if (!path->nodes[level])
4678 return 1;
4679
4680 slot = path->slots[level] + 1;
4681 c = path->nodes[level];
4682next:
4683 if (slot >= btrfs_header_nritems(c)) {
4684 int ret;
4685 int orig_lowest;
4686 struct btrfs_key cur_key;
4687 if (level + 1 >= BTRFS_MAX_LEVEL ||
4688 !path->nodes[level + 1])
4689 return 1;
4690
4691 if (path->locks[level + 1] || path->skip_locking) {
4692 level++;
4693 continue;
4694 }
4695
4696 slot = btrfs_header_nritems(c) - 1;
4697 if (level == 0)
4698 btrfs_item_key_to_cpu(c, &cur_key, slot);
4699 else
4700 btrfs_node_key_to_cpu(c, &cur_key, slot);
4701
4702 orig_lowest = path->lowest_level;
4703 btrfs_release_path(path);
4704 path->lowest_level = level;
4705 ret = btrfs_search_slot(NULL, root, &cur_key, path,
4706 0, 0);
4707 path->lowest_level = orig_lowest;
4708 if (ret < 0)
4709 return ret;
4710
4711 c = path->nodes[level];
4712 slot = path->slots[level];
4713 if (ret == 0)
4714 slot++;
4715 goto next;
4716 }
4717
4718 if (level == 0)
4719 btrfs_item_key_to_cpu(c, key, slot);
4720 else {
4721 u64 gen = btrfs_node_ptr_generation(c, slot);
4722
4723 if (gen < min_trans) {
4724 slot++;
4725 goto next;
4726 }
4727 btrfs_node_key_to_cpu(c, key, slot);
4728 }
4729 return 0;
4730 }
4731 return 1;
4732}
4733
4734int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
4735 u64 time_seq)
4736{
4737 int slot;
4738 int level;
4739 struct extent_buffer *c;
4740 struct extent_buffer *next;
4741 struct btrfs_fs_info *fs_info = root->fs_info;
4742 struct btrfs_key key;
4743 bool need_commit_sem = false;
4744 u32 nritems;
4745 int ret;
4746 int i;
4747
4748 /*
4749 * The nowait semantics are used only for write paths, where we don't
4750 * use the tree mod log and sequence numbers.
4751 */
4752 if (time_seq)
4753 ASSERT(!path->nowait);
4754
4755 nritems = btrfs_header_nritems(path->nodes[0]);
4756 if (nritems == 0)
4757 return 1;
4758
4759 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4760again:
4761 level = 1;
4762 next = NULL;
4763 btrfs_release_path(path);
4764
4765 path->keep_locks = 1;
4766
4767 if (time_seq) {
4768 ret = btrfs_search_old_slot(root, &key, path, time_seq);
4769 } else {
4770 if (path->need_commit_sem) {
4771 path->need_commit_sem = 0;
4772 need_commit_sem = true;
4773 if (path->nowait) {
4774 if (!down_read_trylock(&fs_info->commit_root_sem)) {
4775 ret = -EAGAIN;
4776 goto done;
4777 }
4778 } else {
4779 down_read(&fs_info->commit_root_sem);
4780 }
4781 }
4782 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4783 }
4784 path->keep_locks = 0;
4785
4786 if (ret < 0)
4787 goto done;
4788
4789 nritems = btrfs_header_nritems(path->nodes[0]);
4790 /*
4791 * by releasing the path above we dropped all our locks. A balance
4792 * could have added more items next to the key that used to be
4793 * at the very end of the block. So, check again here and
4794 * advance the path if there are now more items available.
4795 */
4796 if (nritems > 0 && path->slots[0] < nritems - 1) {
4797 if (ret == 0)
4798 path->slots[0]++;
4799 ret = 0;
4800 goto done;
4801 }
4802 /*
4803 * So the above check misses one case:
4804 * - after releasing the path above, someone has removed the item that
4805 * used to be at the very end of the block, and balance between leafs
4806 * gets another one with bigger key.offset to replace it.
4807 *
4808 * This one should be returned as well, or we can get leaf corruption
4809 * later(esp. in __btrfs_drop_extents()).
4810 *
4811 * And a bit more explanation about this check,
4812 * with ret > 0, the key isn't found, the path points to the slot
4813 * where it should be inserted, so the path->slots[0] item must be the
4814 * bigger one.
4815 */
4816 if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
4817 ret = 0;
4818 goto done;
4819 }
4820
4821 while (level < BTRFS_MAX_LEVEL) {
4822 if (!path->nodes[level]) {
4823 ret = 1;
4824 goto done;
4825 }
4826
4827 slot = path->slots[level] + 1;
4828 c = path->nodes[level];
4829 if (slot >= btrfs_header_nritems(c)) {
4830 level++;
4831 if (level == BTRFS_MAX_LEVEL) {
4832 ret = 1;
4833 goto done;
4834 }
4835 continue;
4836 }
4837
4838
4839 /*
4840 * Our current level is where we're going to start from, and to
4841 * make sure lockdep doesn't complain we need to drop our locks
4842 * and nodes from 0 to our current level.
4843 */
4844 for (i = 0; i < level; i++) {
4845 if (path->locks[level]) {
4846 btrfs_tree_read_unlock(path->nodes[i]);
4847 path->locks[i] = 0;
4848 }
4849 free_extent_buffer(path->nodes[i]);
4850 path->nodes[i] = NULL;
4851 }
4852
4853 next = c;
4854 ret = read_block_for_search(root, path, &next, level,
4855 slot, &key);
4856 if (ret == -EAGAIN && !path->nowait)
4857 goto again;
4858
4859 if (ret < 0) {
4860 btrfs_release_path(path);
4861 goto done;
4862 }
4863
4864 if (!path->skip_locking) {
4865 ret = btrfs_try_tree_read_lock(next);
4866 if (!ret && path->nowait) {
4867 ret = -EAGAIN;
4868 goto done;
4869 }
4870 if (!ret && time_seq) {
4871 /*
4872 * If we don't get the lock, we may be racing
4873 * with push_leaf_left, holding that lock while
4874 * itself waiting for the leaf we've currently
4875 * locked. To solve this situation, we give up
4876 * on our lock and cycle.
4877 */
4878 free_extent_buffer(next);
4879 btrfs_release_path(path);
4880 cond_resched();
4881 goto again;
4882 }
4883 if (!ret)
4884 btrfs_tree_read_lock(next);
4885 }
4886 break;
4887 }
4888 path->slots[level] = slot;
4889 while (1) {
4890 level--;
4891 path->nodes[level] = next;
4892 path->slots[level] = 0;
4893 if (!path->skip_locking)
4894 path->locks[level] = BTRFS_READ_LOCK;
4895 if (!level)
4896 break;
4897
4898 ret = read_block_for_search(root, path, &next, level,
4899 0, &key);
4900 if (ret == -EAGAIN && !path->nowait)
4901 goto again;
4902
4903 if (ret < 0) {
4904 btrfs_release_path(path);
4905 goto done;
4906 }
4907
4908 if (!path->skip_locking) {
4909 if (path->nowait) {
4910 if (!btrfs_try_tree_read_lock(next)) {
4911 ret = -EAGAIN;
4912 goto done;
4913 }
4914 } else {
4915 btrfs_tree_read_lock(next);
4916 }
4917 }
4918 }
4919 ret = 0;
4920done:
4921 unlock_up(path, 0, 1, 0, NULL);
4922 if (need_commit_sem) {
4923 int ret2;
4924
4925 path->need_commit_sem = 1;
4926 ret2 = finish_need_commit_sem_search(path);
4927 up_read(&fs_info->commit_root_sem);
4928 if (ret2)
4929 ret = ret2;
4930 }
4931
4932 return ret;
4933}
4934
4935int btrfs_next_old_item(struct btrfs_root *root, struct btrfs_path *path, u64 time_seq)
4936{
4937 path->slots[0]++;
4938 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0]))
4939 return btrfs_next_old_leaf(root, path, time_seq);
4940 return 0;
4941}
4942
4943/*
4944 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4945 * searching until it gets past min_objectid or finds an item of 'type'
4946 *
4947 * returns 0 if something is found, 1 if nothing was found and < 0 on error
4948 */
4949int btrfs_previous_item(struct btrfs_root *root,
4950 struct btrfs_path *path, u64 min_objectid,
4951 int type)
4952{
4953 struct btrfs_key found_key;
4954 struct extent_buffer *leaf;
4955 u32 nritems;
4956 int ret;
4957
4958 while (1) {
4959 if (path->slots[0] == 0) {
4960 ret = btrfs_prev_leaf(root, path);
4961 if (ret != 0)
4962 return ret;
4963 } else {
4964 path->slots[0]--;
4965 }
4966 leaf = path->nodes[0];
4967 nritems = btrfs_header_nritems(leaf);
4968 if (nritems == 0)
4969 return 1;
4970 if (path->slots[0] == nritems)
4971 path->slots[0]--;
4972
4973 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4974 if (found_key.objectid < min_objectid)
4975 break;
4976 if (found_key.type == type)
4977 return 0;
4978 if (found_key.objectid == min_objectid &&
4979 found_key.type < type)
4980 break;
4981 }
4982 return 1;
4983}
4984
4985/*
4986 * search in extent tree to find a previous Metadata/Data extent item with
4987 * min objecitd.
4988 *
4989 * returns 0 if something is found, 1 if nothing was found and < 0 on error
4990 */
4991int btrfs_previous_extent_item(struct btrfs_root *root,
4992 struct btrfs_path *path, u64 min_objectid)
4993{
4994 struct btrfs_key found_key;
4995 struct extent_buffer *leaf;
4996 u32 nritems;
4997 int ret;
4998
4999 while (1) {
5000 if (path->slots[0] == 0) {
5001 ret = btrfs_prev_leaf(root, path);
5002 if (ret != 0)
5003 return ret;
5004 } else {
5005 path->slots[0]--;
5006 }
5007 leaf = path->nodes[0];
5008 nritems = btrfs_header_nritems(leaf);
5009 if (nritems == 0)
5010 return 1;
5011 if (path->slots[0] == nritems)
5012 path->slots[0]--;
5013
5014 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5015 if (found_key.objectid < min_objectid)
5016 break;
5017 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5018 found_key.type == BTRFS_METADATA_ITEM_KEY)
5019 return 0;
5020 if (found_key.objectid == min_objectid &&
5021 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5022 break;
5023 }
5024 return 1;
5025}
5026
5027int __init btrfs_ctree_init(void)
5028{
5029 btrfs_path_cachep = kmem_cache_create("btrfs_path",
5030 sizeof(struct btrfs_path), 0,
5031 SLAB_MEM_SPREAD, NULL);
5032 if (!btrfs_path_cachep)
5033 return -ENOMEM;
5034 return 0;
5035}
5036
5037void __cold btrfs_ctree_exit(void)
5038{
5039 kmem_cache_destroy(btrfs_path_cachep);
5040}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007,2008 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/rbtree.h>
9#include <linux/mm.h>
10#include "ctree.h"
11#include "disk-io.h"
12#include "transaction.h"
13#include "print-tree.h"
14#include "locking.h"
15
16static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
17 *root, struct btrfs_path *path, int level);
18static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
19 const struct btrfs_key *ins_key, struct btrfs_path *path,
20 int data_size, int extend);
21static int push_node_left(struct btrfs_trans_handle *trans,
22 struct btrfs_fs_info *fs_info,
23 struct extent_buffer *dst,
24 struct extent_buffer *src, int empty);
25static int balance_node_right(struct btrfs_trans_handle *trans,
26 struct btrfs_fs_info *fs_info,
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
32struct btrfs_path *btrfs_alloc_path(void)
33{
34 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
35}
36
37/*
38 * set all locked nodes in the path to blocking locks. This should
39 * be done before scheduling
40 */
41noinline void btrfs_set_path_blocking(struct btrfs_path *p)
42{
43 int i;
44 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
45 if (!p->nodes[i] || !p->locks[i])
46 continue;
47 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
48 if (p->locks[i] == BTRFS_READ_LOCK)
49 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
50 else if (p->locks[i] == BTRFS_WRITE_LOCK)
51 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
52 }
53}
54
55/*
56 * reset all the locked nodes in the patch to spinning locks.
57 *
58 * held is used to keep lockdep happy, when lockdep is enabled
59 * we set held to a blocking lock before we go around and
60 * retake all the spinlocks in the path. You can safely use NULL
61 * for held
62 */
63noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
64 struct extent_buffer *held, int held_rw)
65{
66 int i;
67
68 if (held) {
69 btrfs_set_lock_blocking_rw(held, held_rw);
70 if (held_rw == BTRFS_WRITE_LOCK)
71 held_rw = BTRFS_WRITE_LOCK_BLOCKING;
72 else if (held_rw == BTRFS_READ_LOCK)
73 held_rw = BTRFS_READ_LOCK_BLOCKING;
74 }
75 btrfs_set_path_blocking(p);
76
77 for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
78 if (p->nodes[i] && p->locks[i]) {
79 btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
80 if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
81 p->locks[i] = BTRFS_WRITE_LOCK;
82 else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
83 p->locks[i] = BTRFS_READ_LOCK;
84 }
85 }
86
87 if (held)
88 btrfs_clear_lock_blocking_rw(held, held_rw);
89}
90
91/* this also releases the path */
92void btrfs_free_path(struct btrfs_path *p)
93{
94 if (!p)
95 return;
96 btrfs_release_path(p);
97 kmem_cache_free(btrfs_path_cachep, p);
98}
99
100/*
101 * path release drops references on the extent buffers in the path
102 * and it drops any locks held by this path
103 *
104 * It is safe to call this on paths that no locks or extent buffers held.
105 */
106noinline void btrfs_release_path(struct btrfs_path *p)
107{
108 int i;
109
110 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
111 p->slots[i] = 0;
112 if (!p->nodes[i])
113 continue;
114 if (p->locks[i]) {
115 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
116 p->locks[i] = 0;
117 }
118 free_extent_buffer(p->nodes[i]);
119 p->nodes[i] = NULL;
120 }
121}
122
123/*
124 * safely gets a reference on the root node of a tree. A lock
125 * is not taken, so a concurrent writer may put a different node
126 * at the root of the tree. See btrfs_lock_root_node for the
127 * looping required.
128 *
129 * The extent buffer returned by this has a reference taken, so
130 * it won't disappear. It may stop being the root of the tree
131 * at any time because there are no locks held.
132 */
133struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
134{
135 struct extent_buffer *eb;
136
137 while (1) {
138 rcu_read_lock();
139 eb = rcu_dereference(root->node);
140
141 /*
142 * RCU really hurts here, we could free up the root node because
143 * it was COWed but we may not get the new root node yet so do
144 * the inc_not_zero dance and if it doesn't work then
145 * synchronize_rcu and try again.
146 */
147 if (atomic_inc_not_zero(&eb->refs)) {
148 rcu_read_unlock();
149 break;
150 }
151 rcu_read_unlock();
152 synchronize_rcu();
153 }
154 return eb;
155}
156
157/* loop around taking references on and locking the root node of the
158 * tree until you end up with a lock on the root. A locked buffer
159 * is returned, with a reference held.
160 */
161struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
162{
163 struct extent_buffer *eb;
164
165 while (1) {
166 eb = btrfs_root_node(root);
167 btrfs_tree_lock(eb);
168 if (eb == root->node)
169 break;
170 btrfs_tree_unlock(eb);
171 free_extent_buffer(eb);
172 }
173 return eb;
174}
175
176/* loop around taking references on and locking the root node of the
177 * tree until you end up with a lock on the root. A locked buffer
178 * is returned, with a reference held.
179 */
180struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
181{
182 struct extent_buffer *eb;
183
184 while (1) {
185 eb = btrfs_root_node(root);
186 btrfs_tree_read_lock(eb);
187 if (eb == root->node)
188 break;
189 btrfs_tree_read_unlock(eb);
190 free_extent_buffer(eb);
191 }
192 return eb;
193}
194
195/* cowonly root (everything not a reference counted cow subvolume), just get
196 * put onto a simple dirty list. transaction.c walks this to make sure they
197 * get properly updated on disk.
198 */
199static void add_root_to_dirty_list(struct btrfs_root *root)
200{
201 struct btrfs_fs_info *fs_info = root->fs_info;
202
203 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
204 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
205 return;
206
207 spin_lock(&fs_info->trans_lock);
208 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
209 /* Want the extent tree to be the last on the list */
210 if (root->objectid == BTRFS_EXTENT_TREE_OBJECTID)
211 list_move_tail(&root->dirty_list,
212 &fs_info->dirty_cowonly_roots);
213 else
214 list_move(&root->dirty_list,
215 &fs_info->dirty_cowonly_roots);
216 }
217 spin_unlock(&fs_info->trans_lock);
218}
219
220/*
221 * used by snapshot creation to make a copy of a root for a tree with
222 * a given objectid. The buffer with the new root node is returned in
223 * cow_ret, and this func returns zero on success or a negative error code.
224 */
225int btrfs_copy_root(struct btrfs_trans_handle *trans,
226 struct btrfs_root *root,
227 struct extent_buffer *buf,
228 struct extent_buffer **cow_ret, u64 new_root_objectid)
229{
230 struct btrfs_fs_info *fs_info = root->fs_info;
231 struct extent_buffer *cow;
232 int ret = 0;
233 int level;
234 struct btrfs_disk_key disk_key;
235
236 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
237 trans->transid != fs_info->running_transaction->transid);
238 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
239 trans->transid != root->last_trans);
240
241 level = btrfs_header_level(buf);
242 if (level == 0)
243 btrfs_item_key(buf, &disk_key, 0);
244 else
245 btrfs_node_key(buf, &disk_key, 0);
246
247 cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
248 &disk_key, level, buf->start, 0);
249 if (IS_ERR(cow))
250 return PTR_ERR(cow);
251
252 copy_extent_buffer_full(cow, buf);
253 btrfs_set_header_bytenr(cow, cow->start);
254 btrfs_set_header_generation(cow, trans->transid);
255 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
256 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
257 BTRFS_HEADER_FLAG_RELOC);
258 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
259 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
260 else
261 btrfs_set_header_owner(cow, new_root_objectid);
262
263 write_extent_buffer_fsid(cow, fs_info->fsid);
264
265 WARN_ON(btrfs_header_generation(buf) > trans->transid);
266 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
267 ret = btrfs_inc_ref(trans, root, cow, 1);
268 else
269 ret = btrfs_inc_ref(trans, root, cow, 0);
270
271 if (ret)
272 return ret;
273
274 btrfs_mark_buffer_dirty(cow);
275 *cow_ret = cow;
276 return 0;
277}
278
279enum mod_log_op {
280 MOD_LOG_KEY_REPLACE,
281 MOD_LOG_KEY_ADD,
282 MOD_LOG_KEY_REMOVE,
283 MOD_LOG_KEY_REMOVE_WHILE_FREEING,
284 MOD_LOG_KEY_REMOVE_WHILE_MOVING,
285 MOD_LOG_MOVE_KEYS,
286 MOD_LOG_ROOT_REPLACE,
287};
288
289struct tree_mod_root {
290 u64 logical;
291 u8 level;
292};
293
294struct tree_mod_elem {
295 struct rb_node node;
296 u64 logical;
297 u64 seq;
298 enum mod_log_op op;
299
300 /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
301 int slot;
302
303 /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
304 u64 generation;
305
306 /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
307 struct btrfs_disk_key key;
308 u64 blockptr;
309
310 /* this is used for op == MOD_LOG_MOVE_KEYS */
311 struct {
312 int dst_slot;
313 int nr_items;
314 } move;
315
316 /* this is used for op == MOD_LOG_ROOT_REPLACE */
317 struct tree_mod_root old_root;
318};
319
320/*
321 * Pull a new tree mod seq number for our operation.
322 */
323static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
324{
325 return atomic64_inc_return(&fs_info->tree_mod_seq);
326}
327
328/*
329 * This adds a new blocker to the tree mod log's blocker list if the @elem
330 * passed does not already have a sequence number set. So when a caller expects
331 * to record tree modifications, it should ensure to set elem->seq to zero
332 * before calling btrfs_get_tree_mod_seq.
333 * Returns a fresh, unused tree log modification sequence number, even if no new
334 * blocker was added.
335 */
336u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
337 struct seq_list *elem)
338{
339 write_lock(&fs_info->tree_mod_log_lock);
340 spin_lock(&fs_info->tree_mod_seq_lock);
341 if (!elem->seq) {
342 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
343 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
344 }
345 spin_unlock(&fs_info->tree_mod_seq_lock);
346 write_unlock(&fs_info->tree_mod_log_lock);
347
348 return elem->seq;
349}
350
351void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
352 struct seq_list *elem)
353{
354 struct rb_root *tm_root;
355 struct rb_node *node;
356 struct rb_node *next;
357 struct seq_list *cur_elem;
358 struct tree_mod_elem *tm;
359 u64 min_seq = (u64)-1;
360 u64 seq_putting = elem->seq;
361
362 if (!seq_putting)
363 return;
364
365 spin_lock(&fs_info->tree_mod_seq_lock);
366 list_del(&elem->list);
367 elem->seq = 0;
368
369 list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
370 if (cur_elem->seq < min_seq) {
371 if (seq_putting > cur_elem->seq) {
372 /*
373 * blocker with lower sequence number exists, we
374 * cannot remove anything from the log
375 */
376 spin_unlock(&fs_info->tree_mod_seq_lock);
377 return;
378 }
379 min_seq = cur_elem->seq;
380 }
381 }
382 spin_unlock(&fs_info->tree_mod_seq_lock);
383
384 /*
385 * anything that's lower than the lowest existing (read: blocked)
386 * sequence number can be removed from the tree.
387 */
388 write_lock(&fs_info->tree_mod_log_lock);
389 tm_root = &fs_info->tree_mod_log;
390 for (node = rb_first(tm_root); node; node = next) {
391 next = rb_next(node);
392 tm = rb_entry(node, struct tree_mod_elem, node);
393 if (tm->seq > min_seq)
394 continue;
395 rb_erase(node, tm_root);
396 kfree(tm);
397 }
398 write_unlock(&fs_info->tree_mod_log_lock);
399}
400
401/*
402 * key order of the log:
403 * node/leaf start address -> sequence
404 *
405 * The 'start address' is the logical address of the *new* root node
406 * for root replace operations, or the logical address of the affected
407 * block for all other operations.
408 *
409 * Note: must be called with write lock for fs_info::tree_mod_log_lock.
410 */
411static noinline int
412__tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
413{
414 struct rb_root *tm_root;
415 struct rb_node **new;
416 struct rb_node *parent = NULL;
417 struct tree_mod_elem *cur;
418
419 tm->seq = btrfs_inc_tree_mod_seq(fs_info);
420
421 tm_root = &fs_info->tree_mod_log;
422 new = &tm_root->rb_node;
423 while (*new) {
424 cur = rb_entry(*new, struct tree_mod_elem, node);
425 parent = *new;
426 if (cur->logical < tm->logical)
427 new = &((*new)->rb_left);
428 else if (cur->logical > tm->logical)
429 new = &((*new)->rb_right);
430 else if (cur->seq < tm->seq)
431 new = &((*new)->rb_left);
432 else if (cur->seq > tm->seq)
433 new = &((*new)->rb_right);
434 else
435 return -EEXIST;
436 }
437
438 rb_link_node(&tm->node, parent, new);
439 rb_insert_color(&tm->node, tm_root);
440 return 0;
441}
442
443/*
444 * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
445 * returns zero with the tree_mod_log_lock acquired. The caller must hold
446 * this until all tree mod log insertions are recorded in the rb tree and then
447 * write unlock fs_info::tree_mod_log_lock.
448 */
449static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
450 struct extent_buffer *eb) {
451 smp_mb();
452 if (list_empty(&(fs_info)->tree_mod_seq_list))
453 return 1;
454 if (eb && btrfs_header_level(eb) == 0)
455 return 1;
456
457 write_lock(&fs_info->tree_mod_log_lock);
458 if (list_empty(&(fs_info)->tree_mod_seq_list)) {
459 write_unlock(&fs_info->tree_mod_log_lock);
460 return 1;
461 }
462
463 return 0;
464}
465
466/* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
467static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
468 struct extent_buffer *eb)
469{
470 smp_mb();
471 if (list_empty(&(fs_info)->tree_mod_seq_list))
472 return 0;
473 if (eb && btrfs_header_level(eb) == 0)
474 return 0;
475
476 return 1;
477}
478
479static struct tree_mod_elem *
480alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
481 enum mod_log_op op, gfp_t flags)
482{
483 struct tree_mod_elem *tm;
484
485 tm = kzalloc(sizeof(*tm), flags);
486 if (!tm)
487 return NULL;
488
489 tm->logical = eb->start;
490 if (op != MOD_LOG_KEY_ADD) {
491 btrfs_node_key(eb, &tm->key, slot);
492 tm->blockptr = btrfs_node_blockptr(eb, slot);
493 }
494 tm->op = op;
495 tm->slot = slot;
496 tm->generation = btrfs_node_ptr_generation(eb, slot);
497 RB_CLEAR_NODE(&tm->node);
498
499 return tm;
500}
501
502static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
503 enum mod_log_op op, gfp_t flags)
504{
505 struct tree_mod_elem *tm;
506 int ret;
507
508 if (!tree_mod_need_log(eb->fs_info, eb))
509 return 0;
510
511 tm = alloc_tree_mod_elem(eb, slot, op, flags);
512 if (!tm)
513 return -ENOMEM;
514
515 if (tree_mod_dont_log(eb->fs_info, eb)) {
516 kfree(tm);
517 return 0;
518 }
519
520 ret = __tree_mod_log_insert(eb->fs_info, tm);
521 write_unlock(&eb->fs_info->tree_mod_log_lock);
522 if (ret)
523 kfree(tm);
524
525 return ret;
526}
527
528static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
529 int dst_slot, int src_slot, int nr_items)
530{
531 struct tree_mod_elem *tm = NULL;
532 struct tree_mod_elem **tm_list = NULL;
533 int ret = 0;
534 int i;
535 int locked = 0;
536
537 if (!tree_mod_need_log(eb->fs_info, eb))
538 return 0;
539
540 tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
541 if (!tm_list)
542 return -ENOMEM;
543
544 tm = kzalloc(sizeof(*tm), GFP_NOFS);
545 if (!tm) {
546 ret = -ENOMEM;
547 goto free_tms;
548 }
549
550 tm->logical = eb->start;
551 tm->slot = src_slot;
552 tm->move.dst_slot = dst_slot;
553 tm->move.nr_items = nr_items;
554 tm->op = MOD_LOG_MOVE_KEYS;
555
556 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
557 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
558 MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
559 if (!tm_list[i]) {
560 ret = -ENOMEM;
561 goto free_tms;
562 }
563 }
564
565 if (tree_mod_dont_log(eb->fs_info, eb))
566 goto free_tms;
567 locked = 1;
568
569 /*
570 * When we override something during the move, we log these removals.
571 * This can only happen when we move towards the beginning of the
572 * buffer, i.e. dst_slot < src_slot.
573 */
574 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
575 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
576 if (ret)
577 goto free_tms;
578 }
579
580 ret = __tree_mod_log_insert(eb->fs_info, tm);
581 if (ret)
582 goto free_tms;
583 write_unlock(&eb->fs_info->tree_mod_log_lock);
584 kfree(tm_list);
585
586 return 0;
587free_tms:
588 for (i = 0; i < nr_items; i++) {
589 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
590 rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
591 kfree(tm_list[i]);
592 }
593 if (locked)
594 write_unlock(&eb->fs_info->tree_mod_log_lock);
595 kfree(tm_list);
596 kfree(tm);
597
598 return ret;
599}
600
601static inline int
602__tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
603 struct tree_mod_elem **tm_list,
604 int nritems)
605{
606 int i, j;
607 int ret;
608
609 for (i = nritems - 1; i >= 0; i--) {
610 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
611 if (ret) {
612 for (j = nritems - 1; j > i; j--)
613 rb_erase(&tm_list[j]->node,
614 &fs_info->tree_mod_log);
615 return ret;
616 }
617 }
618
619 return 0;
620}
621
622static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
623 struct extent_buffer *new_root, int log_removal)
624{
625 struct btrfs_fs_info *fs_info = old_root->fs_info;
626 struct tree_mod_elem *tm = NULL;
627 struct tree_mod_elem **tm_list = NULL;
628 int nritems = 0;
629 int ret = 0;
630 int i;
631
632 if (!tree_mod_need_log(fs_info, NULL))
633 return 0;
634
635 if (log_removal && btrfs_header_level(old_root) > 0) {
636 nritems = btrfs_header_nritems(old_root);
637 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
638 GFP_NOFS);
639 if (!tm_list) {
640 ret = -ENOMEM;
641 goto free_tms;
642 }
643 for (i = 0; i < nritems; i++) {
644 tm_list[i] = alloc_tree_mod_elem(old_root, i,
645 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
646 if (!tm_list[i]) {
647 ret = -ENOMEM;
648 goto free_tms;
649 }
650 }
651 }
652
653 tm = kzalloc(sizeof(*tm), GFP_NOFS);
654 if (!tm) {
655 ret = -ENOMEM;
656 goto free_tms;
657 }
658
659 tm->logical = new_root->start;
660 tm->old_root.logical = old_root->start;
661 tm->old_root.level = btrfs_header_level(old_root);
662 tm->generation = btrfs_header_generation(old_root);
663 tm->op = MOD_LOG_ROOT_REPLACE;
664
665 if (tree_mod_dont_log(fs_info, NULL))
666 goto free_tms;
667
668 if (tm_list)
669 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
670 if (!ret)
671 ret = __tree_mod_log_insert(fs_info, tm);
672
673 write_unlock(&fs_info->tree_mod_log_lock);
674 if (ret)
675 goto free_tms;
676 kfree(tm_list);
677
678 return ret;
679
680free_tms:
681 if (tm_list) {
682 for (i = 0; i < nritems; i++)
683 kfree(tm_list[i]);
684 kfree(tm_list);
685 }
686 kfree(tm);
687
688 return ret;
689}
690
691static struct tree_mod_elem *
692__tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
693 int smallest)
694{
695 struct rb_root *tm_root;
696 struct rb_node *node;
697 struct tree_mod_elem *cur = NULL;
698 struct tree_mod_elem *found = NULL;
699
700 read_lock(&fs_info->tree_mod_log_lock);
701 tm_root = &fs_info->tree_mod_log;
702 node = tm_root->rb_node;
703 while (node) {
704 cur = rb_entry(node, struct tree_mod_elem, node);
705 if (cur->logical < start) {
706 node = node->rb_left;
707 } else if (cur->logical > start) {
708 node = node->rb_right;
709 } else if (cur->seq < min_seq) {
710 node = node->rb_left;
711 } else if (!smallest) {
712 /* we want the node with the highest seq */
713 if (found)
714 BUG_ON(found->seq > cur->seq);
715 found = cur;
716 node = node->rb_left;
717 } else if (cur->seq > min_seq) {
718 /* we want the node with the smallest seq */
719 if (found)
720 BUG_ON(found->seq < cur->seq);
721 found = cur;
722 node = node->rb_right;
723 } else {
724 found = cur;
725 break;
726 }
727 }
728 read_unlock(&fs_info->tree_mod_log_lock);
729
730 return found;
731}
732
733/*
734 * this returns the element from the log with the smallest time sequence
735 * value that's in the log (the oldest log item). any element with a time
736 * sequence lower than min_seq will be ignored.
737 */
738static struct tree_mod_elem *
739tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
740 u64 min_seq)
741{
742 return __tree_mod_log_search(fs_info, start, min_seq, 1);
743}
744
745/*
746 * this returns the element from the log with the largest time sequence
747 * value that's in the log (the most recent log item). any element with
748 * a time sequence lower than min_seq will be ignored.
749 */
750static struct tree_mod_elem *
751tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
752{
753 return __tree_mod_log_search(fs_info, start, min_seq, 0);
754}
755
756static noinline int
757tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
758 struct extent_buffer *src, unsigned long dst_offset,
759 unsigned long src_offset, int nr_items)
760{
761 int ret = 0;
762 struct tree_mod_elem **tm_list = NULL;
763 struct tree_mod_elem **tm_list_add, **tm_list_rem;
764 int i;
765 int locked = 0;
766
767 if (!tree_mod_need_log(fs_info, NULL))
768 return 0;
769
770 if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
771 return 0;
772
773 tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
774 GFP_NOFS);
775 if (!tm_list)
776 return -ENOMEM;
777
778 tm_list_add = tm_list;
779 tm_list_rem = tm_list + nr_items;
780 for (i = 0; i < nr_items; i++) {
781 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
782 MOD_LOG_KEY_REMOVE, GFP_NOFS);
783 if (!tm_list_rem[i]) {
784 ret = -ENOMEM;
785 goto free_tms;
786 }
787
788 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
789 MOD_LOG_KEY_ADD, GFP_NOFS);
790 if (!tm_list_add[i]) {
791 ret = -ENOMEM;
792 goto free_tms;
793 }
794 }
795
796 if (tree_mod_dont_log(fs_info, NULL))
797 goto free_tms;
798 locked = 1;
799
800 for (i = 0; i < nr_items; i++) {
801 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
802 if (ret)
803 goto free_tms;
804 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
805 if (ret)
806 goto free_tms;
807 }
808
809 write_unlock(&fs_info->tree_mod_log_lock);
810 kfree(tm_list);
811
812 return 0;
813
814free_tms:
815 for (i = 0; i < nr_items * 2; i++) {
816 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
817 rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
818 kfree(tm_list[i]);
819 }
820 if (locked)
821 write_unlock(&fs_info->tree_mod_log_lock);
822 kfree(tm_list);
823
824 return ret;
825}
826
827static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
828{
829 struct tree_mod_elem **tm_list = NULL;
830 int nritems = 0;
831 int i;
832 int ret = 0;
833
834 if (btrfs_header_level(eb) == 0)
835 return 0;
836
837 if (!tree_mod_need_log(eb->fs_info, NULL))
838 return 0;
839
840 nritems = btrfs_header_nritems(eb);
841 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
842 if (!tm_list)
843 return -ENOMEM;
844
845 for (i = 0; i < nritems; i++) {
846 tm_list[i] = alloc_tree_mod_elem(eb, i,
847 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
848 if (!tm_list[i]) {
849 ret = -ENOMEM;
850 goto free_tms;
851 }
852 }
853
854 if (tree_mod_dont_log(eb->fs_info, eb))
855 goto free_tms;
856
857 ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
858 write_unlock(&eb->fs_info->tree_mod_log_lock);
859 if (ret)
860 goto free_tms;
861 kfree(tm_list);
862
863 return 0;
864
865free_tms:
866 for (i = 0; i < nritems; i++)
867 kfree(tm_list[i]);
868 kfree(tm_list);
869
870 return ret;
871}
872
873/*
874 * check if the tree block can be shared by multiple trees
875 */
876int btrfs_block_can_be_shared(struct btrfs_root *root,
877 struct extent_buffer *buf)
878{
879 /*
880 * Tree blocks not in reference counted trees and tree roots
881 * are never shared. If a block was allocated after the last
882 * snapshot and the block was not allocated by tree relocation,
883 * we know the block is not shared.
884 */
885 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
886 buf != root->node && buf != root->commit_root &&
887 (btrfs_header_generation(buf) <=
888 btrfs_root_last_snapshot(&root->root_item) ||
889 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
890 return 1;
891#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
892 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
893 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
894 return 1;
895#endif
896 return 0;
897}
898
899static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
900 struct btrfs_root *root,
901 struct extent_buffer *buf,
902 struct extent_buffer *cow,
903 int *last_ref)
904{
905 struct btrfs_fs_info *fs_info = root->fs_info;
906 u64 refs;
907 u64 owner;
908 u64 flags;
909 u64 new_flags = 0;
910 int ret;
911
912 /*
913 * Backrefs update rules:
914 *
915 * Always use full backrefs for extent pointers in tree block
916 * allocated by tree relocation.
917 *
918 * If a shared tree block is no longer referenced by its owner
919 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
920 * use full backrefs for extent pointers in tree block.
921 *
922 * If a tree block is been relocating
923 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
924 * use full backrefs for extent pointers in tree block.
925 * The reason for this is some operations (such as drop tree)
926 * are only allowed for blocks use full backrefs.
927 */
928
929 if (btrfs_block_can_be_shared(root, buf)) {
930 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
931 btrfs_header_level(buf), 1,
932 &refs, &flags);
933 if (ret)
934 return ret;
935 if (refs == 0) {
936 ret = -EROFS;
937 btrfs_handle_fs_error(fs_info, ret, NULL);
938 return ret;
939 }
940 } else {
941 refs = 1;
942 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
943 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
944 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
945 else
946 flags = 0;
947 }
948
949 owner = btrfs_header_owner(buf);
950 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
951 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
952
953 if (refs > 1) {
954 if ((owner == root->root_key.objectid ||
955 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
956 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
957 ret = btrfs_inc_ref(trans, root, buf, 1);
958 if (ret)
959 return ret;
960
961 if (root->root_key.objectid ==
962 BTRFS_TREE_RELOC_OBJECTID) {
963 ret = btrfs_dec_ref(trans, root, buf, 0);
964 if (ret)
965 return ret;
966 ret = btrfs_inc_ref(trans, root, cow, 1);
967 if (ret)
968 return ret;
969 }
970 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
971 } else {
972
973 if (root->root_key.objectid ==
974 BTRFS_TREE_RELOC_OBJECTID)
975 ret = btrfs_inc_ref(trans, root, cow, 1);
976 else
977 ret = btrfs_inc_ref(trans, root, cow, 0);
978 if (ret)
979 return ret;
980 }
981 if (new_flags != 0) {
982 int level = btrfs_header_level(buf);
983
984 ret = btrfs_set_disk_extent_flags(trans, fs_info,
985 buf->start,
986 buf->len,
987 new_flags, level, 0);
988 if (ret)
989 return ret;
990 }
991 } else {
992 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
993 if (root->root_key.objectid ==
994 BTRFS_TREE_RELOC_OBJECTID)
995 ret = btrfs_inc_ref(trans, root, cow, 1);
996 else
997 ret = btrfs_inc_ref(trans, root, cow, 0);
998 if (ret)
999 return ret;
1000 ret = btrfs_dec_ref(trans, root, buf, 1);
1001 if (ret)
1002 return ret;
1003 }
1004 clean_tree_block(fs_info, buf);
1005 *last_ref = 1;
1006 }
1007 return 0;
1008}
1009
1010/*
1011 * does the dirty work in cow of a single block. The parent block (if
1012 * supplied) is updated to point to the new cow copy. The new buffer is marked
1013 * dirty and returned locked. If you modify the block it needs to be marked
1014 * dirty again.
1015 *
1016 * search_start -- an allocation hint for the new block
1017 *
1018 * empty_size -- a hint that you plan on doing more cow. This is the size in
1019 * bytes the allocator should try to find free next to the block it returns.
1020 * This is just a hint and may be ignored by the allocator.
1021 */
1022static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1023 struct btrfs_root *root,
1024 struct extent_buffer *buf,
1025 struct extent_buffer *parent, int parent_slot,
1026 struct extent_buffer **cow_ret,
1027 u64 search_start, u64 empty_size)
1028{
1029 struct btrfs_fs_info *fs_info = root->fs_info;
1030 struct btrfs_disk_key disk_key;
1031 struct extent_buffer *cow;
1032 int level, ret;
1033 int last_ref = 0;
1034 int unlock_orig = 0;
1035 u64 parent_start = 0;
1036
1037 if (*cow_ret == buf)
1038 unlock_orig = 1;
1039
1040 btrfs_assert_tree_locked(buf);
1041
1042 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1043 trans->transid != fs_info->running_transaction->transid);
1044 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1045 trans->transid != root->last_trans);
1046
1047 level = btrfs_header_level(buf);
1048
1049 if (level == 0)
1050 btrfs_item_key(buf, &disk_key, 0);
1051 else
1052 btrfs_node_key(buf, &disk_key, 0);
1053
1054 if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1055 parent_start = parent->start;
1056
1057 cow = btrfs_alloc_tree_block(trans, root, parent_start,
1058 root->root_key.objectid, &disk_key, level,
1059 search_start, empty_size);
1060 if (IS_ERR(cow))
1061 return PTR_ERR(cow);
1062
1063 /* cow is set to blocking by btrfs_init_new_buffer */
1064
1065 copy_extent_buffer_full(cow, buf);
1066 btrfs_set_header_bytenr(cow, cow->start);
1067 btrfs_set_header_generation(cow, trans->transid);
1068 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1069 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1070 BTRFS_HEADER_FLAG_RELOC);
1071 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1072 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1073 else
1074 btrfs_set_header_owner(cow, root->root_key.objectid);
1075
1076 write_extent_buffer_fsid(cow, fs_info->fsid);
1077
1078 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1079 if (ret) {
1080 btrfs_abort_transaction(trans, ret);
1081 return ret;
1082 }
1083
1084 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1085 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1086 if (ret) {
1087 btrfs_abort_transaction(trans, ret);
1088 return ret;
1089 }
1090 }
1091
1092 if (buf == root->node) {
1093 WARN_ON(parent && parent != buf);
1094 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1095 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1096 parent_start = buf->start;
1097
1098 extent_buffer_get(cow);
1099 ret = tree_mod_log_insert_root(root->node, cow, 1);
1100 BUG_ON(ret < 0);
1101 rcu_assign_pointer(root->node, cow);
1102
1103 btrfs_free_tree_block(trans, root, buf, parent_start,
1104 last_ref);
1105 free_extent_buffer(buf);
1106 add_root_to_dirty_list(root);
1107 } else {
1108 WARN_ON(trans->transid != btrfs_header_generation(parent));
1109 tree_mod_log_insert_key(parent, parent_slot,
1110 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1111 btrfs_set_node_blockptr(parent, parent_slot,
1112 cow->start);
1113 btrfs_set_node_ptr_generation(parent, parent_slot,
1114 trans->transid);
1115 btrfs_mark_buffer_dirty(parent);
1116 if (last_ref) {
1117 ret = tree_mod_log_free_eb(buf);
1118 if (ret) {
1119 btrfs_abort_transaction(trans, ret);
1120 return ret;
1121 }
1122 }
1123 btrfs_free_tree_block(trans, root, buf, parent_start,
1124 last_ref);
1125 }
1126 if (unlock_orig)
1127 btrfs_tree_unlock(buf);
1128 free_extent_buffer_stale(buf);
1129 btrfs_mark_buffer_dirty(cow);
1130 *cow_ret = cow;
1131 return 0;
1132}
1133
1134/*
1135 * returns the logical address of the oldest predecessor of the given root.
1136 * entries older than time_seq are ignored.
1137 */
1138static struct tree_mod_elem *__tree_mod_log_oldest_root(
1139 struct extent_buffer *eb_root, u64 time_seq)
1140{
1141 struct tree_mod_elem *tm;
1142 struct tree_mod_elem *found = NULL;
1143 u64 root_logical = eb_root->start;
1144 int looped = 0;
1145
1146 if (!time_seq)
1147 return NULL;
1148
1149 /*
1150 * the very last operation that's logged for a root is the
1151 * replacement operation (if it is replaced at all). this has
1152 * the logical address of the *new* root, making it the very
1153 * first operation that's logged for this root.
1154 */
1155 while (1) {
1156 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1157 time_seq);
1158 if (!looped && !tm)
1159 return NULL;
1160 /*
1161 * if there are no tree operation for the oldest root, we simply
1162 * return it. this should only happen if that (old) root is at
1163 * level 0.
1164 */
1165 if (!tm)
1166 break;
1167
1168 /*
1169 * if there's an operation that's not a root replacement, we
1170 * found the oldest version of our root. normally, we'll find a
1171 * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1172 */
1173 if (tm->op != MOD_LOG_ROOT_REPLACE)
1174 break;
1175
1176 found = tm;
1177 root_logical = tm->old_root.logical;
1178 looped = 1;
1179 }
1180
1181 /* if there's no old root to return, return what we found instead */
1182 if (!found)
1183 found = tm;
1184
1185 return found;
1186}
1187
1188/*
1189 * tm is a pointer to the first operation to rewind within eb. then, all
1190 * previous operations will be rewound (until we reach something older than
1191 * time_seq).
1192 */
1193static void
1194__tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1195 u64 time_seq, struct tree_mod_elem *first_tm)
1196{
1197 u32 n;
1198 struct rb_node *next;
1199 struct tree_mod_elem *tm = first_tm;
1200 unsigned long o_dst;
1201 unsigned long o_src;
1202 unsigned long p_size = sizeof(struct btrfs_key_ptr);
1203
1204 n = btrfs_header_nritems(eb);
1205 read_lock(&fs_info->tree_mod_log_lock);
1206 while (tm && tm->seq >= time_seq) {
1207 /*
1208 * all the operations are recorded with the operator used for
1209 * the modification. as we're going backwards, we do the
1210 * opposite of each operation here.
1211 */
1212 switch (tm->op) {
1213 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1214 BUG_ON(tm->slot < n);
1215 /* Fallthrough */
1216 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1217 case MOD_LOG_KEY_REMOVE:
1218 btrfs_set_node_key(eb, &tm->key, tm->slot);
1219 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1220 btrfs_set_node_ptr_generation(eb, tm->slot,
1221 tm->generation);
1222 n++;
1223 break;
1224 case MOD_LOG_KEY_REPLACE:
1225 BUG_ON(tm->slot >= n);
1226 btrfs_set_node_key(eb, &tm->key, tm->slot);
1227 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1228 btrfs_set_node_ptr_generation(eb, tm->slot,
1229 tm->generation);
1230 break;
1231 case MOD_LOG_KEY_ADD:
1232 /* if a move operation is needed it's in the log */
1233 n--;
1234 break;
1235 case MOD_LOG_MOVE_KEYS:
1236 o_dst = btrfs_node_key_ptr_offset(tm->slot);
1237 o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1238 memmove_extent_buffer(eb, o_dst, o_src,
1239 tm->move.nr_items * p_size);
1240 break;
1241 case MOD_LOG_ROOT_REPLACE:
1242 /*
1243 * this operation is special. for roots, this must be
1244 * handled explicitly before rewinding.
1245 * for non-roots, this operation may exist if the node
1246 * was a root: root A -> child B; then A gets empty and
1247 * B is promoted to the new root. in the mod log, we'll
1248 * have a root-replace operation for B, a tree block
1249 * that is no root. we simply ignore that operation.
1250 */
1251 break;
1252 }
1253 next = rb_next(&tm->node);
1254 if (!next)
1255 break;
1256 tm = rb_entry(next, struct tree_mod_elem, node);
1257 if (tm->logical != first_tm->logical)
1258 break;
1259 }
1260 read_unlock(&fs_info->tree_mod_log_lock);
1261 btrfs_set_header_nritems(eb, n);
1262}
1263
1264/*
1265 * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1266 * is returned. If rewind operations happen, a fresh buffer is returned. The
1267 * returned buffer is always read-locked. If the returned buffer is not the
1268 * input buffer, the lock on the input buffer is released and the input buffer
1269 * is freed (its refcount is decremented).
1270 */
1271static struct extent_buffer *
1272tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1273 struct extent_buffer *eb, u64 time_seq)
1274{
1275 struct extent_buffer *eb_rewin;
1276 struct tree_mod_elem *tm;
1277
1278 if (!time_seq)
1279 return eb;
1280
1281 if (btrfs_header_level(eb) == 0)
1282 return eb;
1283
1284 tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1285 if (!tm)
1286 return eb;
1287
1288 btrfs_set_path_blocking(path);
1289 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1290
1291 if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1292 BUG_ON(tm->slot != 0);
1293 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1294 if (!eb_rewin) {
1295 btrfs_tree_read_unlock_blocking(eb);
1296 free_extent_buffer(eb);
1297 return NULL;
1298 }
1299 btrfs_set_header_bytenr(eb_rewin, eb->start);
1300 btrfs_set_header_backref_rev(eb_rewin,
1301 btrfs_header_backref_rev(eb));
1302 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1303 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1304 } else {
1305 eb_rewin = btrfs_clone_extent_buffer(eb);
1306 if (!eb_rewin) {
1307 btrfs_tree_read_unlock_blocking(eb);
1308 free_extent_buffer(eb);
1309 return NULL;
1310 }
1311 }
1312
1313 btrfs_clear_path_blocking(path, NULL, BTRFS_READ_LOCK);
1314 btrfs_tree_read_unlock_blocking(eb);
1315 free_extent_buffer(eb);
1316
1317 extent_buffer_get(eb_rewin);
1318 btrfs_tree_read_lock(eb_rewin);
1319 __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1320 WARN_ON(btrfs_header_nritems(eb_rewin) >
1321 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1322
1323 return eb_rewin;
1324}
1325
1326/*
1327 * get_old_root() rewinds the state of @root's root node to the given @time_seq
1328 * value. If there are no changes, the current root->root_node is returned. If
1329 * anything changed in between, there's a fresh buffer allocated on which the
1330 * rewind operations are done. In any case, the returned buffer is read locked.
1331 * Returns NULL on error (with no locks held).
1332 */
1333static inline struct extent_buffer *
1334get_old_root(struct btrfs_root *root, u64 time_seq)
1335{
1336 struct btrfs_fs_info *fs_info = root->fs_info;
1337 struct tree_mod_elem *tm;
1338 struct extent_buffer *eb = NULL;
1339 struct extent_buffer *eb_root;
1340 struct extent_buffer *old;
1341 struct tree_mod_root *old_root = NULL;
1342 u64 old_generation = 0;
1343 u64 logical;
1344 int level;
1345
1346 eb_root = btrfs_read_lock_root_node(root);
1347 tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1348 if (!tm)
1349 return eb_root;
1350
1351 if (tm->op == MOD_LOG_ROOT_REPLACE) {
1352 old_root = &tm->old_root;
1353 old_generation = tm->generation;
1354 logical = old_root->logical;
1355 level = old_root->level;
1356 } else {
1357 logical = eb_root->start;
1358 level = btrfs_header_level(eb_root);
1359 }
1360
1361 tm = tree_mod_log_search(fs_info, logical, time_seq);
1362 if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1363 btrfs_tree_read_unlock(eb_root);
1364 free_extent_buffer(eb_root);
1365 old = read_tree_block(fs_info, logical, 0, level, NULL);
1366 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1367 if (!IS_ERR(old))
1368 free_extent_buffer(old);
1369 btrfs_warn(fs_info,
1370 "failed to read tree block %llu from get_old_root",
1371 logical);
1372 } else {
1373 eb = btrfs_clone_extent_buffer(old);
1374 free_extent_buffer(old);
1375 }
1376 } else if (old_root) {
1377 btrfs_tree_read_unlock(eb_root);
1378 free_extent_buffer(eb_root);
1379 eb = alloc_dummy_extent_buffer(fs_info, logical);
1380 } else {
1381 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1382 eb = btrfs_clone_extent_buffer(eb_root);
1383 btrfs_tree_read_unlock_blocking(eb_root);
1384 free_extent_buffer(eb_root);
1385 }
1386
1387 if (!eb)
1388 return NULL;
1389 extent_buffer_get(eb);
1390 btrfs_tree_read_lock(eb);
1391 if (old_root) {
1392 btrfs_set_header_bytenr(eb, eb->start);
1393 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1394 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1395 btrfs_set_header_level(eb, old_root->level);
1396 btrfs_set_header_generation(eb, old_generation);
1397 }
1398 if (tm)
1399 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1400 else
1401 WARN_ON(btrfs_header_level(eb) != 0);
1402 WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1403
1404 return eb;
1405}
1406
1407int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1408{
1409 struct tree_mod_elem *tm;
1410 int level;
1411 struct extent_buffer *eb_root = btrfs_root_node(root);
1412
1413 tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1414 if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1415 level = tm->old_root.level;
1416 } else {
1417 level = btrfs_header_level(eb_root);
1418 }
1419 free_extent_buffer(eb_root);
1420
1421 return level;
1422}
1423
1424static inline int should_cow_block(struct btrfs_trans_handle *trans,
1425 struct btrfs_root *root,
1426 struct extent_buffer *buf)
1427{
1428 if (btrfs_is_testing(root->fs_info))
1429 return 0;
1430
1431 /* Ensure we can see the FORCE_COW bit */
1432 smp_mb__before_atomic();
1433
1434 /*
1435 * We do not need to cow a block if
1436 * 1) this block is not created or changed in this transaction;
1437 * 2) this block does not belong to TREE_RELOC tree;
1438 * 3) the root is not forced COW.
1439 *
1440 * What is forced COW:
1441 * when we create snapshot during committing the transaction,
1442 * after we've finished coping src root, we must COW the shared
1443 * block to ensure the metadata consistency.
1444 */
1445 if (btrfs_header_generation(buf) == trans->transid &&
1446 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1447 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1448 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1449 !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1450 return 0;
1451 return 1;
1452}
1453
1454/*
1455 * cows a single block, see __btrfs_cow_block for the real work.
1456 * This version of it has extra checks so that a block isn't COWed more than
1457 * once per transaction, as long as it hasn't been written yet
1458 */
1459noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1460 struct btrfs_root *root, struct extent_buffer *buf,
1461 struct extent_buffer *parent, int parent_slot,
1462 struct extent_buffer **cow_ret)
1463{
1464 struct btrfs_fs_info *fs_info = root->fs_info;
1465 u64 search_start;
1466 int ret;
1467
1468 if (trans->transaction != fs_info->running_transaction)
1469 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1470 trans->transid,
1471 fs_info->running_transaction->transid);
1472
1473 if (trans->transid != fs_info->generation)
1474 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1475 trans->transid, fs_info->generation);
1476
1477 if (!should_cow_block(trans, root, buf)) {
1478 trans->dirty = true;
1479 *cow_ret = buf;
1480 return 0;
1481 }
1482
1483 search_start = buf->start & ~((u64)SZ_1G - 1);
1484
1485 if (parent)
1486 btrfs_set_lock_blocking(parent);
1487 btrfs_set_lock_blocking(buf);
1488
1489 ret = __btrfs_cow_block(trans, root, buf, parent,
1490 parent_slot, cow_ret, search_start, 0);
1491
1492 trace_btrfs_cow_block(root, buf, *cow_ret);
1493
1494 return ret;
1495}
1496
1497/*
1498 * helper function for defrag to decide if two blocks pointed to by a
1499 * node are actually close by
1500 */
1501static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1502{
1503 if (blocknr < other && other - (blocknr + blocksize) < 32768)
1504 return 1;
1505 if (blocknr > other && blocknr - (other + blocksize) < 32768)
1506 return 1;
1507 return 0;
1508}
1509
1510/*
1511 * compare two keys in a memcmp fashion
1512 */
1513static int comp_keys(const struct btrfs_disk_key *disk,
1514 const struct btrfs_key *k2)
1515{
1516 struct btrfs_key k1;
1517
1518 btrfs_disk_key_to_cpu(&k1, disk);
1519
1520 return btrfs_comp_cpu_keys(&k1, k2);
1521}
1522
1523/*
1524 * same as comp_keys only with two btrfs_key's
1525 */
1526int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1527{
1528 if (k1->objectid > k2->objectid)
1529 return 1;
1530 if (k1->objectid < k2->objectid)
1531 return -1;
1532 if (k1->type > k2->type)
1533 return 1;
1534 if (k1->type < k2->type)
1535 return -1;
1536 if (k1->offset > k2->offset)
1537 return 1;
1538 if (k1->offset < k2->offset)
1539 return -1;
1540 return 0;
1541}
1542
1543/*
1544 * this is used by the defrag code to go through all the
1545 * leaves pointed to by a node and reallocate them so that
1546 * disk order is close to key order
1547 */
1548int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1549 struct btrfs_root *root, struct extent_buffer *parent,
1550 int start_slot, u64 *last_ret,
1551 struct btrfs_key *progress)
1552{
1553 struct btrfs_fs_info *fs_info = root->fs_info;
1554 struct extent_buffer *cur;
1555 u64 blocknr;
1556 u64 gen;
1557 u64 search_start = *last_ret;
1558 u64 last_block = 0;
1559 u64 other;
1560 u32 parent_nritems;
1561 int end_slot;
1562 int i;
1563 int err = 0;
1564 int parent_level;
1565 int uptodate;
1566 u32 blocksize;
1567 int progress_passed = 0;
1568 struct btrfs_disk_key disk_key;
1569
1570 parent_level = btrfs_header_level(parent);
1571
1572 WARN_ON(trans->transaction != fs_info->running_transaction);
1573 WARN_ON(trans->transid != fs_info->generation);
1574
1575 parent_nritems = btrfs_header_nritems(parent);
1576 blocksize = fs_info->nodesize;
1577 end_slot = parent_nritems - 1;
1578
1579 if (parent_nritems <= 1)
1580 return 0;
1581
1582 btrfs_set_lock_blocking(parent);
1583
1584 for (i = start_slot; i <= end_slot; i++) {
1585 struct btrfs_key first_key;
1586 int close = 1;
1587
1588 btrfs_node_key(parent, &disk_key, i);
1589 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1590 continue;
1591
1592 progress_passed = 1;
1593 blocknr = btrfs_node_blockptr(parent, i);
1594 gen = btrfs_node_ptr_generation(parent, i);
1595 btrfs_node_key_to_cpu(parent, &first_key, i);
1596 if (last_block == 0)
1597 last_block = blocknr;
1598
1599 if (i > 0) {
1600 other = btrfs_node_blockptr(parent, i - 1);
1601 close = close_blocks(blocknr, other, blocksize);
1602 }
1603 if (!close && i < end_slot) {
1604 other = btrfs_node_blockptr(parent, i + 1);
1605 close = close_blocks(blocknr, other, blocksize);
1606 }
1607 if (close) {
1608 last_block = blocknr;
1609 continue;
1610 }
1611
1612 cur = find_extent_buffer(fs_info, blocknr);
1613 if (cur)
1614 uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1615 else
1616 uptodate = 0;
1617 if (!cur || !uptodate) {
1618 if (!cur) {
1619 cur = read_tree_block(fs_info, blocknr, gen,
1620 parent_level - 1,
1621 &first_key);
1622 if (IS_ERR(cur)) {
1623 return PTR_ERR(cur);
1624 } else if (!extent_buffer_uptodate(cur)) {
1625 free_extent_buffer(cur);
1626 return -EIO;
1627 }
1628 } else if (!uptodate) {
1629 err = btrfs_read_buffer(cur, gen,
1630 parent_level - 1,&first_key);
1631 if (err) {
1632 free_extent_buffer(cur);
1633 return err;
1634 }
1635 }
1636 }
1637 if (search_start == 0)
1638 search_start = last_block;
1639
1640 btrfs_tree_lock(cur);
1641 btrfs_set_lock_blocking(cur);
1642 err = __btrfs_cow_block(trans, root, cur, parent, i,
1643 &cur, search_start,
1644 min(16 * blocksize,
1645 (end_slot - i) * blocksize));
1646 if (err) {
1647 btrfs_tree_unlock(cur);
1648 free_extent_buffer(cur);
1649 break;
1650 }
1651 search_start = cur->start;
1652 last_block = cur->start;
1653 *last_ret = search_start;
1654 btrfs_tree_unlock(cur);
1655 free_extent_buffer(cur);
1656 }
1657 return err;
1658}
1659
1660/*
1661 * search for key in the extent_buffer. The items start at offset p,
1662 * and they are item_size apart. There are 'max' items in p.
1663 *
1664 * the slot in the array is returned via slot, and it points to
1665 * the place where you would insert key if it is not found in
1666 * the array.
1667 *
1668 * slot may point to max if the key is bigger than all of the keys
1669 */
1670static noinline int generic_bin_search(struct extent_buffer *eb,
1671 unsigned long p, int item_size,
1672 const struct btrfs_key *key,
1673 int max, int *slot)
1674{
1675 int low = 0;
1676 int high = max;
1677 int mid;
1678 int ret;
1679 struct btrfs_disk_key *tmp = NULL;
1680 struct btrfs_disk_key unaligned;
1681 unsigned long offset;
1682 char *kaddr = NULL;
1683 unsigned long map_start = 0;
1684 unsigned long map_len = 0;
1685 int err;
1686
1687 if (low > high) {
1688 btrfs_err(eb->fs_info,
1689 "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1690 __func__, low, high, eb->start,
1691 btrfs_header_owner(eb), btrfs_header_level(eb));
1692 return -EINVAL;
1693 }
1694
1695 while (low < high) {
1696 mid = (low + high) / 2;
1697 offset = p + mid * item_size;
1698
1699 if (!kaddr || offset < map_start ||
1700 (offset + sizeof(struct btrfs_disk_key)) >
1701 map_start + map_len) {
1702
1703 err = map_private_extent_buffer(eb, offset,
1704 sizeof(struct btrfs_disk_key),
1705 &kaddr, &map_start, &map_len);
1706
1707 if (!err) {
1708 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1709 map_start);
1710 } else if (err == 1) {
1711 read_extent_buffer(eb, &unaligned,
1712 offset, sizeof(unaligned));
1713 tmp = &unaligned;
1714 } else {
1715 return err;
1716 }
1717
1718 } else {
1719 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1720 map_start);
1721 }
1722 ret = comp_keys(tmp, key);
1723
1724 if (ret < 0)
1725 low = mid + 1;
1726 else if (ret > 0)
1727 high = mid;
1728 else {
1729 *slot = mid;
1730 return 0;
1731 }
1732 }
1733 *slot = low;
1734 return 1;
1735}
1736
1737/*
1738 * simple bin_search frontend that does the right thing for
1739 * leaves vs nodes
1740 */
1741int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1742 int level, int *slot)
1743{
1744 if (level == 0)
1745 return generic_bin_search(eb,
1746 offsetof(struct btrfs_leaf, items),
1747 sizeof(struct btrfs_item),
1748 key, btrfs_header_nritems(eb),
1749 slot);
1750 else
1751 return generic_bin_search(eb,
1752 offsetof(struct btrfs_node, ptrs),
1753 sizeof(struct btrfs_key_ptr),
1754 key, btrfs_header_nritems(eb),
1755 slot);
1756}
1757
1758static void root_add_used(struct btrfs_root *root, u32 size)
1759{
1760 spin_lock(&root->accounting_lock);
1761 btrfs_set_root_used(&root->root_item,
1762 btrfs_root_used(&root->root_item) + size);
1763 spin_unlock(&root->accounting_lock);
1764}
1765
1766static void root_sub_used(struct btrfs_root *root, u32 size)
1767{
1768 spin_lock(&root->accounting_lock);
1769 btrfs_set_root_used(&root->root_item,
1770 btrfs_root_used(&root->root_item) - size);
1771 spin_unlock(&root->accounting_lock);
1772}
1773
1774/* given a node and slot number, this reads the blocks it points to. The
1775 * extent buffer is returned with a reference taken (but unlocked).
1776 */
1777static noinline struct extent_buffer *
1778read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1779 int slot)
1780{
1781 int level = btrfs_header_level(parent);
1782 struct extent_buffer *eb;
1783 struct btrfs_key first_key;
1784
1785 if (slot < 0 || slot >= btrfs_header_nritems(parent))
1786 return ERR_PTR(-ENOENT);
1787
1788 BUG_ON(level == 0);
1789
1790 btrfs_node_key_to_cpu(parent, &first_key, slot);
1791 eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1792 btrfs_node_ptr_generation(parent, slot),
1793 level - 1, &first_key);
1794 if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1795 free_extent_buffer(eb);
1796 eb = ERR_PTR(-EIO);
1797 }
1798
1799 return eb;
1800}
1801
1802/*
1803 * node level balancing, used to make sure nodes are in proper order for
1804 * item deletion. We balance from the top down, so we have to make sure
1805 * that a deletion won't leave an node completely empty later on.
1806 */
1807static noinline int balance_level(struct btrfs_trans_handle *trans,
1808 struct btrfs_root *root,
1809 struct btrfs_path *path, int level)
1810{
1811 struct btrfs_fs_info *fs_info = root->fs_info;
1812 struct extent_buffer *right = NULL;
1813 struct extent_buffer *mid;
1814 struct extent_buffer *left = NULL;
1815 struct extent_buffer *parent = NULL;
1816 int ret = 0;
1817 int wret;
1818 int pslot;
1819 int orig_slot = path->slots[level];
1820 u64 orig_ptr;
1821
1822 if (level == 0)
1823 return 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 = read_node_slot(fs_info, 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(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 clean_tree_block(fs_info, 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 = read_node_slot(fs_info, 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(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 = read_node_slot(fs_info, 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(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, fs_info, 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, fs_info, mid, right, 1);
1932 if (wret < 0 && wret != -ENOSPC)
1933 ret = wret;
1934 if (btrfs_header_nritems(right) == 0) {
1935 clean_tree_block(fs_info, 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, fs_info, mid, left);
1968 if (wret < 0) {
1969 ret = wret;
1970 goto enospc;
1971 }
1972 if (wret == 1) {
1973 wret = push_node_left(trans, fs_info, 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 clean_tree_block(fs_info, 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 extent_buffer_get(left);
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 = read_node_slot(fs_info, 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(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, fs_info,
2085 left, mid, 0);
2086 }
2087 }
2088 if (wret < 0)
2089 ret = wret;
2090 if (wret == 0) {
2091 struct btrfs_disk_key disk_key;
2092 orig_slot += left_nr;
2093 btrfs_node_key(mid, &disk_key, 0);
2094 ret = tree_mod_log_insert_key(parent, pslot,
2095 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2096 BUG_ON(ret < 0);
2097 btrfs_set_node_key(parent, &disk_key, pslot);
2098 btrfs_mark_buffer_dirty(parent);
2099 if (btrfs_header_nritems(left) > orig_slot) {
2100 path->nodes[level] = left;
2101 path->slots[level + 1] -= 1;
2102 path->slots[level] = orig_slot;
2103 btrfs_tree_unlock(mid);
2104 free_extent_buffer(mid);
2105 } else {
2106 orig_slot -=
2107 btrfs_header_nritems(left);
2108 path->slots[level] = orig_slot;
2109 btrfs_tree_unlock(left);
2110 free_extent_buffer(left);
2111 }
2112 return 0;
2113 }
2114 btrfs_tree_unlock(left);
2115 free_extent_buffer(left);
2116 }
2117 right = read_node_slot(fs_info, parent, pslot + 1);
2118 if (IS_ERR(right))
2119 right = NULL;
2120
2121 /*
2122 * then try to empty the right most buffer into the middle
2123 */
2124 if (right) {
2125 u32 right_nr;
2126
2127 btrfs_tree_lock(right);
2128 btrfs_set_lock_blocking(right);
2129
2130 right_nr = btrfs_header_nritems(right);
2131 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2132 wret = 1;
2133 } else {
2134 ret = btrfs_cow_block(trans, root, right,
2135 parent, pslot + 1,
2136 &right);
2137 if (ret)
2138 wret = 1;
2139 else {
2140 wret = balance_node_right(trans, fs_info,
2141 right, mid);
2142 }
2143 }
2144 if (wret < 0)
2145 ret = wret;
2146 if (wret == 0) {
2147 struct btrfs_disk_key disk_key;
2148
2149 btrfs_node_key(right, &disk_key, 0);
2150 ret = tree_mod_log_insert_key(parent, pslot + 1,
2151 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2152 BUG_ON(ret < 0);
2153 btrfs_set_node_key(parent, &disk_key, pslot + 1);
2154 btrfs_mark_buffer_dirty(parent);
2155
2156 if (btrfs_header_nritems(mid) <= orig_slot) {
2157 path->nodes[level] = right;
2158 path->slots[level + 1] += 1;
2159 path->slots[level] = orig_slot -
2160 btrfs_header_nritems(mid);
2161 btrfs_tree_unlock(mid);
2162 free_extent_buffer(mid);
2163 } else {
2164 btrfs_tree_unlock(right);
2165 free_extent_buffer(right);
2166 }
2167 return 0;
2168 }
2169 btrfs_tree_unlock(right);
2170 free_extent_buffer(right);
2171 }
2172 return 1;
2173}
2174
2175/*
2176 * readahead one full node of leaves, finding things that are close
2177 * to the block in 'slot', and triggering ra on them.
2178 */
2179static void reada_for_search(struct btrfs_fs_info *fs_info,
2180 struct btrfs_path *path,
2181 int level, int slot, u64 objectid)
2182{
2183 struct extent_buffer *node;
2184 struct btrfs_disk_key disk_key;
2185 u32 nritems;
2186 u64 search;
2187 u64 target;
2188 u64 nread = 0;
2189 struct extent_buffer *eb;
2190 u32 nr;
2191 u32 blocksize;
2192 u32 nscan = 0;
2193
2194 if (level != 1)
2195 return;
2196
2197 if (!path->nodes[level])
2198 return;
2199
2200 node = path->nodes[level];
2201
2202 search = btrfs_node_blockptr(node, slot);
2203 blocksize = fs_info->nodesize;
2204 eb = find_extent_buffer(fs_info, search);
2205 if (eb) {
2206 free_extent_buffer(eb);
2207 return;
2208 }
2209
2210 target = search;
2211
2212 nritems = btrfs_header_nritems(node);
2213 nr = slot;
2214
2215 while (1) {
2216 if (path->reada == READA_BACK) {
2217 if (nr == 0)
2218 break;
2219 nr--;
2220 } else if (path->reada == READA_FORWARD) {
2221 nr++;
2222 if (nr >= nritems)
2223 break;
2224 }
2225 if (path->reada == READA_BACK && objectid) {
2226 btrfs_node_key(node, &disk_key, nr);
2227 if (btrfs_disk_key_objectid(&disk_key) != objectid)
2228 break;
2229 }
2230 search = btrfs_node_blockptr(node, nr);
2231 if ((search <= target && target - search <= 65536) ||
2232 (search > target && search - target <= 65536)) {
2233 readahead_tree_block(fs_info, search);
2234 nread += blocksize;
2235 }
2236 nscan++;
2237 if ((nread > 65536 || nscan > 32))
2238 break;
2239 }
2240}
2241
2242static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2243 struct btrfs_path *path, int level)
2244{
2245 int slot;
2246 int nritems;
2247 struct extent_buffer *parent;
2248 struct extent_buffer *eb;
2249 u64 gen;
2250 u64 block1 = 0;
2251 u64 block2 = 0;
2252
2253 parent = path->nodes[level + 1];
2254 if (!parent)
2255 return;
2256
2257 nritems = btrfs_header_nritems(parent);
2258 slot = path->slots[level + 1];
2259
2260 if (slot > 0) {
2261 block1 = btrfs_node_blockptr(parent, slot - 1);
2262 gen = btrfs_node_ptr_generation(parent, slot - 1);
2263 eb = find_extent_buffer(fs_info, block1);
2264 /*
2265 * if we get -eagain from btrfs_buffer_uptodate, we
2266 * don't want to return eagain here. That will loop
2267 * forever
2268 */
2269 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2270 block1 = 0;
2271 free_extent_buffer(eb);
2272 }
2273 if (slot + 1 < nritems) {
2274 block2 = btrfs_node_blockptr(parent, slot + 1);
2275 gen = btrfs_node_ptr_generation(parent, slot + 1);
2276 eb = find_extent_buffer(fs_info, block2);
2277 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2278 block2 = 0;
2279 free_extent_buffer(eb);
2280 }
2281
2282 if (block1)
2283 readahead_tree_block(fs_info, block1);
2284 if (block2)
2285 readahead_tree_block(fs_info, block2);
2286}
2287
2288
2289/*
2290 * when we walk down the tree, it is usually safe to unlock the higher layers
2291 * in the tree. The exceptions are when our path goes through slot 0, because
2292 * operations on the tree might require changing key pointers higher up in the
2293 * tree.
2294 *
2295 * callers might also have set path->keep_locks, which tells this code to keep
2296 * the lock if the path points to the last slot in the block. This is part of
2297 * walking through the tree, and selecting the next slot in the higher block.
2298 *
2299 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
2300 * if lowest_unlock is 1, level 0 won't be unlocked
2301 */
2302static noinline void unlock_up(struct btrfs_path *path, int level,
2303 int lowest_unlock, int min_write_lock_level,
2304 int *write_lock_level)
2305{
2306 int i;
2307 int skip_level = level;
2308 int no_skips = 0;
2309 struct extent_buffer *t;
2310
2311 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2312 if (!path->nodes[i])
2313 break;
2314 if (!path->locks[i])
2315 break;
2316 if (!no_skips && path->slots[i] == 0) {
2317 skip_level = i + 1;
2318 continue;
2319 }
2320 if (!no_skips && path->keep_locks) {
2321 u32 nritems;
2322 t = path->nodes[i];
2323 nritems = btrfs_header_nritems(t);
2324 if (nritems < 1 || path->slots[i] >= nritems - 1) {
2325 skip_level = i + 1;
2326 continue;
2327 }
2328 }
2329 if (skip_level < i && i >= lowest_unlock)
2330 no_skips = 1;
2331
2332 t = path->nodes[i];
2333 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2334 btrfs_tree_unlock_rw(t, path->locks[i]);
2335 path->locks[i] = 0;
2336 if (write_lock_level &&
2337 i > min_write_lock_level &&
2338 i <= *write_lock_level) {
2339 *write_lock_level = i - 1;
2340 }
2341 }
2342 }
2343}
2344
2345/*
2346 * This releases any locks held in the path starting at level and
2347 * going all the way up to the root.
2348 *
2349 * btrfs_search_slot will keep the lock held on higher nodes in a few
2350 * corner cases, such as COW of the block at slot zero in the node. This
2351 * ignores those rules, and it should only be called when there are no
2352 * more updates to be done higher up in the tree.
2353 */
2354noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2355{
2356 int i;
2357
2358 if (path->keep_locks)
2359 return;
2360
2361 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2362 if (!path->nodes[i])
2363 continue;
2364 if (!path->locks[i])
2365 continue;
2366 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2367 path->locks[i] = 0;
2368 }
2369}
2370
2371/*
2372 * helper function for btrfs_search_slot. The goal is to find a block
2373 * in cache without setting the path to blocking. If we find the block
2374 * we return zero and the path is unchanged.
2375 *
2376 * If we can't find the block, we set the path blocking and do some
2377 * reada. -EAGAIN is returned and the search must be repeated.
2378 */
2379static int
2380read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2381 struct extent_buffer **eb_ret, int level, int slot,
2382 const struct btrfs_key *key)
2383{
2384 struct btrfs_fs_info *fs_info = root->fs_info;
2385 u64 blocknr;
2386 u64 gen;
2387 struct extent_buffer *b = *eb_ret;
2388 struct extent_buffer *tmp;
2389 struct btrfs_key first_key;
2390 int ret;
2391 int parent_level;
2392
2393 blocknr = btrfs_node_blockptr(b, slot);
2394 gen = btrfs_node_ptr_generation(b, slot);
2395 parent_level = btrfs_header_level(b);
2396 btrfs_node_key_to_cpu(b, &first_key, slot);
2397
2398 tmp = find_extent_buffer(fs_info, blocknr);
2399 if (tmp) {
2400 /* first we do an atomic uptodate check */
2401 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2402 *eb_ret = tmp;
2403 return 0;
2404 }
2405
2406 /* the pages were up to date, but we failed
2407 * the generation number check. Do a full
2408 * read for the generation number that is correct.
2409 * We must do this without dropping locks so
2410 * we can trust our generation number
2411 */
2412 btrfs_set_path_blocking(p);
2413
2414 /* now we're allowed to do a blocking uptodate check */
2415 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2416 if (!ret) {
2417 *eb_ret = tmp;
2418 return 0;
2419 }
2420 free_extent_buffer(tmp);
2421 btrfs_release_path(p);
2422 return -EIO;
2423 }
2424
2425 /*
2426 * reduce lock contention at high levels
2427 * of the btree by dropping locks before
2428 * we read. Don't release the lock on the current
2429 * level because we need to walk this node to figure
2430 * out which blocks to read.
2431 */
2432 btrfs_unlock_up_safe(p, level + 1);
2433 btrfs_set_path_blocking(p);
2434
2435 free_extent_buffer(tmp);
2436 if (p->reada != READA_NONE)
2437 reada_for_search(fs_info, p, level, slot, key->objectid);
2438
2439 ret = -EAGAIN;
2440 tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2441 &first_key);
2442 if (!IS_ERR(tmp)) {
2443 /*
2444 * If the read above didn't mark this buffer up to date,
2445 * it will never end up being up to date. Set ret to EIO now
2446 * and give up so that our caller doesn't loop forever
2447 * on our EAGAINs.
2448 */
2449 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2450 ret = -EIO;
2451 free_extent_buffer(tmp);
2452 } else {
2453 ret = PTR_ERR(tmp);
2454 }
2455
2456 btrfs_release_path(p);
2457 return ret;
2458}
2459
2460/*
2461 * helper function for btrfs_search_slot. This does all of the checks
2462 * for node-level blocks and does any balancing required based on
2463 * the ins_len.
2464 *
2465 * If no extra work was required, zero is returned. If we had to
2466 * drop the path, -EAGAIN is returned and btrfs_search_slot must
2467 * start over
2468 */
2469static int
2470setup_nodes_for_search(struct btrfs_trans_handle *trans,
2471 struct btrfs_root *root, struct btrfs_path *p,
2472 struct extent_buffer *b, int level, int ins_len,
2473 int *write_lock_level)
2474{
2475 struct btrfs_fs_info *fs_info = root->fs_info;
2476 int ret;
2477
2478 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2479 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
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 = split_node(trans, root, p, level);
2491 btrfs_clear_path_blocking(p, NULL, 0);
2492
2493 BUG_ON(sret > 0);
2494 if (sret) {
2495 ret = sret;
2496 goto done;
2497 }
2498 b = p->nodes[level];
2499 } else if (ins_len < 0 && btrfs_header_nritems(b) <
2500 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2501 int sret;
2502
2503 if (*write_lock_level < level + 1) {
2504 *write_lock_level = level + 1;
2505 btrfs_release_path(p);
2506 goto again;
2507 }
2508
2509 btrfs_set_path_blocking(p);
2510 reada_for_balance(fs_info, p, level);
2511 sret = balance_level(trans, root, p, level);
2512 btrfs_clear_path_blocking(p, NULL, 0);
2513
2514 if (sret) {
2515 ret = sret;
2516 goto done;
2517 }
2518 b = p->nodes[level];
2519 if (!b) {
2520 btrfs_release_path(p);
2521 goto again;
2522 }
2523 BUG_ON(btrfs_header_nritems(b) == 1);
2524 }
2525 return 0;
2526
2527again:
2528 ret = -EAGAIN;
2529done:
2530 return ret;
2531}
2532
2533static void key_search_validate(struct extent_buffer *b,
2534 const struct btrfs_key *key,
2535 int level)
2536{
2537#ifdef CONFIG_BTRFS_ASSERT
2538 struct btrfs_disk_key disk_key;
2539
2540 btrfs_cpu_key_to_disk(&disk_key, key);
2541
2542 if (level == 0)
2543 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2544 offsetof(struct btrfs_leaf, items[0].key),
2545 sizeof(disk_key)));
2546 else
2547 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2548 offsetof(struct btrfs_node, ptrs[0].key),
2549 sizeof(disk_key)));
2550#endif
2551}
2552
2553static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2554 int level, int *prev_cmp, int *slot)
2555{
2556 if (*prev_cmp != 0) {
2557 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2558 return *prev_cmp;
2559 }
2560
2561 key_search_validate(b, key, level);
2562 *slot = 0;
2563
2564 return 0;
2565}
2566
2567int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2568 u64 iobjectid, u64 ioff, u8 key_type,
2569 struct btrfs_key *found_key)
2570{
2571 int ret;
2572 struct btrfs_key key;
2573 struct extent_buffer *eb;
2574
2575 ASSERT(path);
2576 ASSERT(found_key);
2577
2578 key.type = key_type;
2579 key.objectid = iobjectid;
2580 key.offset = ioff;
2581
2582 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2583 if (ret < 0)
2584 return ret;
2585
2586 eb = path->nodes[0];
2587 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2588 ret = btrfs_next_leaf(fs_root, path);
2589 if (ret)
2590 return ret;
2591 eb = path->nodes[0];
2592 }
2593
2594 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2595 if (found_key->type != key.type ||
2596 found_key->objectid != key.objectid)
2597 return 1;
2598
2599 return 0;
2600}
2601
2602/*
2603 * btrfs_search_slot - look for a key in a tree and perform necessary
2604 * modifications to preserve tree invariants.
2605 *
2606 * @trans: Handle of transaction, used when modifying the tree
2607 * @p: Holds all btree nodes along the search path
2608 * @root: The root node of the tree
2609 * @key: The key we are looking for
2610 * @ins_len: Indicates purpose of search, for inserts it is 1, for
2611 * deletions it's -1. 0 for plain searches
2612 * @cow: boolean should CoW operations be performed. Must always be 1
2613 * when modifying the tree.
2614 *
2615 * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2616 * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2617 *
2618 * If @key is found, 0 is returned and you can find the item in the leaf level
2619 * of the path (level 0)
2620 *
2621 * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2622 * points to the slot where it should be inserted
2623 *
2624 * If an error is encountered while searching the tree a negative error number
2625 * is returned
2626 */
2627int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2628 const struct btrfs_key *key, struct btrfs_path *p,
2629 int ins_len, int cow)
2630{
2631 struct btrfs_fs_info *fs_info = root->fs_info;
2632 struct extent_buffer *b;
2633 int slot;
2634 int ret;
2635 int err;
2636 int level;
2637 int lowest_unlock = 1;
2638 int root_lock;
2639 /* everything at write_lock_level or lower must be write locked */
2640 int write_lock_level = 0;
2641 u8 lowest_level = 0;
2642 int min_write_lock_level;
2643 int prev_cmp;
2644
2645 lowest_level = p->lowest_level;
2646 WARN_ON(lowest_level && ins_len > 0);
2647 WARN_ON(p->nodes[0] != NULL);
2648 BUG_ON(!cow && ins_len);
2649
2650 if (ins_len < 0) {
2651 lowest_unlock = 2;
2652
2653 /* when we are removing items, we might have to go up to level
2654 * two as we update tree pointers Make sure we keep write
2655 * for those levels as well
2656 */
2657 write_lock_level = 2;
2658 } else if (ins_len > 0) {
2659 /*
2660 * for inserting items, make sure we have a write lock on
2661 * level 1 so we can update keys
2662 */
2663 write_lock_level = 1;
2664 }
2665
2666 if (!cow)
2667 write_lock_level = -1;
2668
2669 if (cow && (p->keep_locks || p->lowest_level))
2670 write_lock_level = BTRFS_MAX_LEVEL;
2671
2672 min_write_lock_level = write_lock_level;
2673
2674again:
2675 prev_cmp = -1;
2676 /*
2677 * we try very hard to do read locks on the root
2678 */
2679 root_lock = BTRFS_READ_LOCK;
2680 level = 0;
2681 if (p->search_commit_root) {
2682 /*
2683 * the commit roots are read only
2684 * so we always do read locks
2685 */
2686 if (p->need_commit_sem)
2687 down_read(&fs_info->commit_root_sem);
2688 b = root->commit_root;
2689 extent_buffer_get(b);
2690 level = btrfs_header_level(b);
2691 if (p->need_commit_sem)
2692 up_read(&fs_info->commit_root_sem);
2693 if (!p->skip_locking)
2694 btrfs_tree_read_lock(b);
2695 } else {
2696 if (p->skip_locking) {
2697 b = btrfs_root_node(root);
2698 level = btrfs_header_level(b);
2699 } else {
2700 /* we don't know the level of the root node
2701 * until we actually have it read locked
2702 */
2703 b = btrfs_read_lock_root_node(root);
2704 level = btrfs_header_level(b);
2705 if (level <= write_lock_level) {
2706 /* whoops, must trade for write lock */
2707 btrfs_tree_read_unlock(b);
2708 free_extent_buffer(b);
2709 b = btrfs_lock_root_node(root);
2710 root_lock = BTRFS_WRITE_LOCK;
2711
2712 /* the level might have changed, check again */
2713 level = btrfs_header_level(b);
2714 }
2715 }
2716 }
2717 p->nodes[level] = b;
2718 if (!p->skip_locking)
2719 p->locks[level] = root_lock;
2720
2721 while (b) {
2722 level = btrfs_header_level(b);
2723
2724 /*
2725 * setup the path here so we can release it under lock
2726 * contention with the cow code
2727 */
2728 if (cow) {
2729 bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2730
2731 /*
2732 * if we don't really need to cow this block
2733 * then we don't want to set the path blocking,
2734 * so we test it here
2735 */
2736 if (!should_cow_block(trans, root, b)) {
2737 trans->dirty = true;
2738 goto cow_done;
2739 }
2740
2741 /*
2742 * must have write locks on this node and the
2743 * parent
2744 */
2745 if (level > write_lock_level ||
2746 (level + 1 > write_lock_level &&
2747 level + 1 < BTRFS_MAX_LEVEL &&
2748 p->nodes[level + 1])) {
2749 write_lock_level = level + 1;
2750 btrfs_release_path(p);
2751 goto again;
2752 }
2753
2754 btrfs_set_path_blocking(p);
2755 if (last_level)
2756 err = btrfs_cow_block(trans, root, b, NULL, 0,
2757 &b);
2758 else
2759 err = btrfs_cow_block(trans, root, b,
2760 p->nodes[level + 1],
2761 p->slots[level + 1], &b);
2762 if (err) {
2763 ret = err;
2764 goto done;
2765 }
2766 }
2767cow_done:
2768 p->nodes[level] = b;
2769 btrfs_clear_path_blocking(p, NULL, 0);
2770
2771 /*
2772 * we have a lock on b and as long as we aren't changing
2773 * the tree, there is no way to for the items in b to change.
2774 * It is safe to drop the lock on our parent before we
2775 * go through the expensive btree search on b.
2776 *
2777 * If we're inserting or deleting (ins_len != 0), then we might
2778 * be changing slot zero, which may require changing the parent.
2779 * So, we can't drop the lock until after we know which slot
2780 * we're operating on.
2781 */
2782 if (!ins_len && !p->keep_locks) {
2783 int u = level + 1;
2784
2785 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2786 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2787 p->locks[u] = 0;
2788 }
2789 }
2790
2791 ret = key_search(b, key, level, &prev_cmp, &slot);
2792 if (ret < 0)
2793 goto done;
2794
2795 if (level != 0) {
2796 int dec = 0;
2797 if (ret && slot > 0) {
2798 dec = 1;
2799 slot -= 1;
2800 }
2801 p->slots[level] = slot;
2802 err = setup_nodes_for_search(trans, root, p, b, level,
2803 ins_len, &write_lock_level);
2804 if (err == -EAGAIN)
2805 goto again;
2806 if (err) {
2807 ret = err;
2808 goto done;
2809 }
2810 b = p->nodes[level];
2811 slot = p->slots[level];
2812
2813 /*
2814 * slot 0 is special, if we change the key
2815 * we have to update the parent pointer
2816 * which means we must have a write lock
2817 * on the parent
2818 */
2819 if (slot == 0 && ins_len &&
2820 write_lock_level < level + 1) {
2821 write_lock_level = level + 1;
2822 btrfs_release_path(p);
2823 goto again;
2824 }
2825
2826 unlock_up(p, level, lowest_unlock,
2827 min_write_lock_level, &write_lock_level);
2828
2829 if (level == lowest_level) {
2830 if (dec)
2831 p->slots[level]++;
2832 goto done;
2833 }
2834
2835 err = read_block_for_search(root, p, &b, level,
2836 slot, key);
2837 if (err == -EAGAIN)
2838 goto again;
2839 if (err) {
2840 ret = err;
2841 goto done;
2842 }
2843
2844 if (!p->skip_locking) {
2845 level = btrfs_header_level(b);
2846 if (level <= write_lock_level) {
2847 err = btrfs_try_tree_write_lock(b);
2848 if (!err) {
2849 btrfs_set_path_blocking(p);
2850 btrfs_tree_lock(b);
2851 btrfs_clear_path_blocking(p, b,
2852 BTRFS_WRITE_LOCK);
2853 }
2854 p->locks[level] = BTRFS_WRITE_LOCK;
2855 } else {
2856 err = btrfs_tree_read_lock_atomic(b);
2857 if (!err) {
2858 btrfs_set_path_blocking(p);
2859 btrfs_tree_read_lock(b);
2860 btrfs_clear_path_blocking(p, b,
2861 BTRFS_READ_LOCK);
2862 }
2863 p->locks[level] = BTRFS_READ_LOCK;
2864 }
2865 p->nodes[level] = b;
2866 }
2867 } else {
2868 p->slots[level] = slot;
2869 if (ins_len > 0 &&
2870 btrfs_leaf_free_space(fs_info, b) < ins_len) {
2871 if (write_lock_level < 1) {
2872 write_lock_level = 1;
2873 btrfs_release_path(p);
2874 goto again;
2875 }
2876
2877 btrfs_set_path_blocking(p);
2878 err = split_leaf(trans, root, key,
2879 p, ins_len, ret == 0);
2880 btrfs_clear_path_blocking(p, NULL, 0);
2881
2882 BUG_ON(err > 0);
2883 if (err) {
2884 ret = err;
2885 goto done;
2886 }
2887 }
2888 if (!p->search_for_split)
2889 unlock_up(p, level, lowest_unlock,
2890 min_write_lock_level, &write_lock_level);
2891 goto done;
2892 }
2893 }
2894 ret = 1;
2895done:
2896 /*
2897 * we don't really know what they plan on doing with the path
2898 * from here on, so for now just mark it as blocking
2899 */
2900 if (!p->leave_spinning)
2901 btrfs_set_path_blocking(p);
2902 if (ret < 0 && !p->skip_release_on_error)
2903 btrfs_release_path(p);
2904 return ret;
2905}
2906
2907/*
2908 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2909 * current state of the tree together with the operations recorded in the tree
2910 * modification log to search for the key in a previous version of this tree, as
2911 * denoted by the time_seq parameter.
2912 *
2913 * Naturally, there is no support for insert, delete or cow operations.
2914 *
2915 * The resulting path and return value will be set up as if we called
2916 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2917 */
2918int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2919 struct btrfs_path *p, u64 time_seq)
2920{
2921 struct btrfs_fs_info *fs_info = root->fs_info;
2922 struct extent_buffer *b;
2923 int slot;
2924 int ret;
2925 int err;
2926 int level;
2927 int lowest_unlock = 1;
2928 u8 lowest_level = 0;
2929 int prev_cmp = -1;
2930
2931 lowest_level = p->lowest_level;
2932 WARN_ON(p->nodes[0] != NULL);
2933
2934 if (p->search_commit_root) {
2935 BUG_ON(time_seq);
2936 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2937 }
2938
2939again:
2940 b = get_old_root(root, time_seq);
2941 level = btrfs_header_level(b);
2942 p->locks[level] = BTRFS_READ_LOCK;
2943
2944 while (b) {
2945 level = btrfs_header_level(b);
2946 p->nodes[level] = b;
2947 btrfs_clear_path_blocking(p, NULL, 0);
2948
2949 /*
2950 * we have a lock on b and as long as we aren't changing
2951 * the tree, there is no way to for the items in b to change.
2952 * It is safe to drop the lock on our parent before we
2953 * go through the expensive btree search on b.
2954 */
2955 btrfs_unlock_up_safe(p, level + 1);
2956
2957 /*
2958 * Since we can unwind ebs we want to do a real search every
2959 * time.
2960 */
2961 prev_cmp = -1;
2962 ret = key_search(b, key, level, &prev_cmp, &slot);
2963
2964 if (level != 0) {
2965 int dec = 0;
2966 if (ret && slot > 0) {
2967 dec = 1;
2968 slot -= 1;
2969 }
2970 p->slots[level] = slot;
2971 unlock_up(p, level, lowest_unlock, 0, NULL);
2972
2973 if (level == lowest_level) {
2974 if (dec)
2975 p->slots[level]++;
2976 goto done;
2977 }
2978
2979 err = read_block_for_search(root, p, &b, level,
2980 slot, key);
2981 if (err == -EAGAIN)
2982 goto again;
2983 if (err) {
2984 ret = err;
2985 goto done;
2986 }
2987
2988 level = btrfs_header_level(b);
2989 err = btrfs_tree_read_lock_atomic(b);
2990 if (!err) {
2991 btrfs_set_path_blocking(p);
2992 btrfs_tree_read_lock(b);
2993 btrfs_clear_path_blocking(p, b,
2994 BTRFS_READ_LOCK);
2995 }
2996 b = tree_mod_log_rewind(fs_info, p, b, time_seq);
2997 if (!b) {
2998 ret = -ENOMEM;
2999 goto done;
3000 }
3001 p->locks[level] = BTRFS_READ_LOCK;
3002 p->nodes[level] = b;
3003 } else {
3004 p->slots[level] = slot;
3005 unlock_up(p, level, lowest_unlock, 0, NULL);
3006 goto done;
3007 }
3008 }
3009 ret = 1;
3010done:
3011 if (!p->leave_spinning)
3012 btrfs_set_path_blocking(p);
3013 if (ret < 0)
3014 btrfs_release_path(p);
3015
3016 return ret;
3017}
3018
3019/*
3020 * helper to use instead of search slot if no exact match is needed but
3021 * instead the next or previous item should be returned.
3022 * When find_higher is true, the next higher item is returned, the next lower
3023 * otherwise.
3024 * When return_any and find_higher are both true, and no higher item is found,
3025 * return the next lower instead.
3026 * When return_any is true and find_higher is false, and no lower item is found,
3027 * return the next higher instead.
3028 * It returns 0 if any item is found, 1 if none is found (tree empty), and
3029 * < 0 on error
3030 */
3031int btrfs_search_slot_for_read(struct btrfs_root *root,
3032 const struct btrfs_key *key,
3033 struct btrfs_path *p, int find_higher,
3034 int return_any)
3035{
3036 int ret;
3037 struct extent_buffer *leaf;
3038
3039again:
3040 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3041 if (ret <= 0)
3042 return ret;
3043 /*
3044 * a return value of 1 means the path is at the position where the
3045 * item should be inserted. Normally this is the next bigger item,
3046 * but in case the previous item is the last in a leaf, path points
3047 * to the first free slot in the previous leaf, i.e. at an invalid
3048 * item.
3049 */
3050 leaf = p->nodes[0];
3051
3052 if (find_higher) {
3053 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3054 ret = btrfs_next_leaf(root, p);
3055 if (ret <= 0)
3056 return ret;
3057 if (!return_any)
3058 return 1;
3059 /*
3060 * no higher item found, return the next
3061 * lower instead
3062 */
3063 return_any = 0;
3064 find_higher = 0;
3065 btrfs_release_path(p);
3066 goto again;
3067 }
3068 } else {
3069 if (p->slots[0] == 0) {
3070 ret = btrfs_prev_leaf(root, p);
3071 if (ret < 0)
3072 return ret;
3073 if (!ret) {
3074 leaf = p->nodes[0];
3075 if (p->slots[0] == btrfs_header_nritems(leaf))
3076 p->slots[0]--;
3077 return 0;
3078 }
3079 if (!return_any)
3080 return 1;
3081 /*
3082 * no lower item found, return the next
3083 * higher instead
3084 */
3085 return_any = 0;
3086 find_higher = 1;
3087 btrfs_release_path(p);
3088 goto again;
3089 } else {
3090 --p->slots[0];
3091 }
3092 }
3093 return 0;
3094}
3095
3096/*
3097 * adjust the pointers going up the tree, starting at level
3098 * making sure the right key of each node is points to 'key'.
3099 * This is used after shifting pointers to the left, so it stops
3100 * fixing up pointers when a given leaf/node is not in slot 0 of the
3101 * higher levels
3102 *
3103 */
3104static void fixup_low_keys(struct btrfs_fs_info *fs_info,
3105 struct btrfs_path *path,
3106 struct btrfs_disk_key *key, int level)
3107{
3108 int i;
3109 struct extent_buffer *t;
3110 int ret;
3111
3112 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3113 int tslot = path->slots[i];
3114
3115 if (!path->nodes[i])
3116 break;
3117 t = path->nodes[i];
3118 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3119 GFP_ATOMIC);
3120 BUG_ON(ret < 0);
3121 btrfs_set_node_key(t, key, tslot);
3122 btrfs_mark_buffer_dirty(path->nodes[i]);
3123 if (tslot != 0)
3124 break;
3125 }
3126}
3127
3128/*
3129 * update item key.
3130 *
3131 * This function isn't completely safe. It's the caller's responsibility
3132 * that the new key won't break the order
3133 */
3134void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3135 struct btrfs_path *path,
3136 const struct btrfs_key *new_key)
3137{
3138 struct btrfs_disk_key disk_key;
3139 struct extent_buffer *eb;
3140 int slot;
3141
3142 eb = path->nodes[0];
3143 slot = path->slots[0];
3144 if (slot > 0) {
3145 btrfs_item_key(eb, &disk_key, slot - 1);
3146 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
3147 }
3148 if (slot < btrfs_header_nritems(eb) - 1) {
3149 btrfs_item_key(eb, &disk_key, slot + 1);
3150 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
3151 }
3152
3153 btrfs_cpu_key_to_disk(&disk_key, new_key);
3154 btrfs_set_item_key(eb, &disk_key, slot);
3155 btrfs_mark_buffer_dirty(eb);
3156 if (slot == 0)
3157 fixup_low_keys(fs_info, path, &disk_key, 1);
3158}
3159
3160/*
3161 * try to push data from one node into the next node left in the
3162 * tree.
3163 *
3164 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3165 * error, and > 0 if there was no room in the left hand block.
3166 */
3167static int push_node_left(struct btrfs_trans_handle *trans,
3168 struct btrfs_fs_info *fs_info,
3169 struct extent_buffer *dst,
3170 struct extent_buffer *src, int empty)
3171{
3172 int push_items = 0;
3173 int src_nritems;
3174 int dst_nritems;
3175 int ret = 0;
3176
3177 src_nritems = btrfs_header_nritems(src);
3178 dst_nritems = btrfs_header_nritems(dst);
3179 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3180 WARN_ON(btrfs_header_generation(src) != trans->transid);
3181 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3182
3183 if (!empty && src_nritems <= 8)
3184 return 1;
3185
3186 if (push_items <= 0)
3187 return 1;
3188
3189 if (empty) {
3190 push_items = min(src_nritems, push_items);
3191 if (push_items < src_nritems) {
3192 /* leave at least 8 pointers in the node if
3193 * we aren't going to empty it
3194 */
3195 if (src_nritems - push_items < 8) {
3196 if (push_items <= 8)
3197 return 1;
3198 push_items -= 8;
3199 }
3200 }
3201 } else
3202 push_items = min(src_nritems - 8, push_items);
3203
3204 ret = tree_mod_log_eb_copy(fs_info, dst, src, dst_nritems, 0,
3205 push_items);
3206 if (ret) {
3207 btrfs_abort_transaction(trans, ret);
3208 return ret;
3209 }
3210 copy_extent_buffer(dst, src,
3211 btrfs_node_key_ptr_offset(dst_nritems),
3212 btrfs_node_key_ptr_offset(0),
3213 push_items * sizeof(struct btrfs_key_ptr));
3214
3215 if (push_items < src_nritems) {
3216 /*
3217 * Don't call tree_mod_log_insert_move here, key removal was
3218 * already fully logged by tree_mod_log_eb_copy above.
3219 */
3220 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3221 btrfs_node_key_ptr_offset(push_items),
3222 (src_nritems - push_items) *
3223 sizeof(struct btrfs_key_ptr));
3224 }
3225 btrfs_set_header_nritems(src, src_nritems - push_items);
3226 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3227 btrfs_mark_buffer_dirty(src);
3228 btrfs_mark_buffer_dirty(dst);
3229
3230 return ret;
3231}
3232
3233/*
3234 * try to push data from one node into the next node right in the
3235 * tree.
3236 *
3237 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3238 * error, and > 0 if there was no room in the right hand block.
3239 *
3240 * this will only push up to 1/2 the contents of the left node over
3241 */
3242static int balance_node_right(struct btrfs_trans_handle *trans,
3243 struct btrfs_fs_info *fs_info,
3244 struct extent_buffer *dst,
3245 struct extent_buffer *src)
3246{
3247 int push_items = 0;
3248 int max_push;
3249 int src_nritems;
3250 int dst_nritems;
3251 int ret = 0;
3252
3253 WARN_ON(btrfs_header_generation(src) != trans->transid);
3254 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3255
3256 src_nritems = btrfs_header_nritems(src);
3257 dst_nritems = btrfs_header_nritems(dst);
3258 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3259 if (push_items <= 0)
3260 return 1;
3261
3262 if (src_nritems < 4)
3263 return 1;
3264
3265 max_push = src_nritems / 2 + 1;
3266 /* don't try to empty the node */
3267 if (max_push >= src_nritems)
3268 return 1;
3269
3270 if (max_push < push_items)
3271 push_items = max_push;
3272
3273 ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3274 BUG_ON(ret < 0);
3275 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3276 btrfs_node_key_ptr_offset(0),
3277 (dst_nritems) *
3278 sizeof(struct btrfs_key_ptr));
3279
3280 ret = tree_mod_log_eb_copy(fs_info, dst, src, 0,
3281 src_nritems - push_items, push_items);
3282 if (ret) {
3283 btrfs_abort_transaction(trans, ret);
3284 return ret;
3285 }
3286 copy_extent_buffer(dst, src,
3287 btrfs_node_key_ptr_offset(0),
3288 btrfs_node_key_ptr_offset(src_nritems - push_items),
3289 push_items * sizeof(struct btrfs_key_ptr));
3290
3291 btrfs_set_header_nritems(src, src_nritems - push_items);
3292 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3293
3294 btrfs_mark_buffer_dirty(src);
3295 btrfs_mark_buffer_dirty(dst);
3296
3297 return ret;
3298}
3299
3300/*
3301 * helper function to insert a new root level in the tree.
3302 * A new node is allocated, and a single item is inserted to
3303 * point to the existing root
3304 *
3305 * returns zero on success or < 0 on failure.
3306 */
3307static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3308 struct btrfs_root *root,
3309 struct btrfs_path *path, int level)
3310{
3311 struct btrfs_fs_info *fs_info = root->fs_info;
3312 u64 lower_gen;
3313 struct extent_buffer *lower;
3314 struct extent_buffer *c;
3315 struct extent_buffer *old;
3316 struct btrfs_disk_key lower_key;
3317 int ret;
3318
3319 BUG_ON(path->nodes[level]);
3320 BUG_ON(path->nodes[level-1] != root->node);
3321
3322 lower = path->nodes[level-1];
3323 if (level == 1)
3324 btrfs_item_key(lower, &lower_key, 0);
3325 else
3326 btrfs_node_key(lower, &lower_key, 0);
3327
3328 c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3329 &lower_key, level, root->node->start, 0);
3330 if (IS_ERR(c))
3331 return PTR_ERR(c);
3332
3333 root_add_used(root, fs_info->nodesize);
3334
3335 memzero_extent_buffer(c, 0, sizeof(struct btrfs_header));
3336 btrfs_set_header_nritems(c, 1);
3337 btrfs_set_header_level(c, level);
3338 btrfs_set_header_bytenr(c, c->start);
3339 btrfs_set_header_generation(c, trans->transid);
3340 btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
3341 btrfs_set_header_owner(c, root->root_key.objectid);
3342
3343 write_extent_buffer_fsid(c, fs_info->fsid);
3344 write_extent_buffer_chunk_tree_uuid(c, fs_info->chunk_tree_uuid);
3345
3346 btrfs_set_node_key(c, &lower_key, 0);
3347 btrfs_set_node_blockptr(c, 0, lower->start);
3348 lower_gen = btrfs_header_generation(lower);
3349 WARN_ON(lower_gen != trans->transid);
3350
3351 btrfs_set_node_ptr_generation(c, 0, lower_gen);
3352
3353 btrfs_mark_buffer_dirty(c);
3354
3355 old = root->node;
3356 ret = tree_mod_log_insert_root(root->node, c, 0);
3357 BUG_ON(ret < 0);
3358 rcu_assign_pointer(root->node, c);
3359
3360 /* the super has an extra ref to root->node */
3361 free_extent_buffer(old);
3362
3363 add_root_to_dirty_list(root);
3364 extent_buffer_get(c);
3365 path->nodes[level] = c;
3366 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3367 path->slots[level] = 0;
3368 return 0;
3369}
3370
3371/*
3372 * worker function to insert a single pointer in a node.
3373 * the node should have enough room for the pointer already
3374 *
3375 * slot and level indicate where you want the key to go, and
3376 * blocknr is the block the key points to.
3377 */
3378static void insert_ptr(struct btrfs_trans_handle *trans,
3379 struct btrfs_fs_info *fs_info, struct btrfs_path *path,
3380 struct btrfs_disk_key *key, u64 bytenr,
3381 int slot, int level)
3382{
3383 struct extent_buffer *lower;
3384 int nritems;
3385 int ret;
3386
3387 BUG_ON(!path->nodes[level]);
3388 btrfs_assert_tree_locked(path->nodes[level]);
3389 lower = path->nodes[level];
3390 nritems = btrfs_header_nritems(lower);
3391 BUG_ON(slot > nritems);
3392 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(fs_info));
3393 if (slot != nritems) {
3394 if (level) {
3395 ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3396 nritems - slot);
3397 BUG_ON(ret < 0);
3398 }
3399 memmove_extent_buffer(lower,
3400 btrfs_node_key_ptr_offset(slot + 1),
3401 btrfs_node_key_ptr_offset(slot),
3402 (nritems - slot) * sizeof(struct btrfs_key_ptr));
3403 }
3404 if (level) {
3405 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3406 GFP_NOFS);
3407 BUG_ON(ret < 0);
3408 }
3409 btrfs_set_node_key(lower, key, slot);
3410 btrfs_set_node_blockptr(lower, slot, bytenr);
3411 WARN_ON(trans->transid == 0);
3412 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3413 btrfs_set_header_nritems(lower, nritems + 1);
3414 btrfs_mark_buffer_dirty(lower);
3415}
3416
3417/*
3418 * split the node at the specified level in path in two.
3419 * The path is corrected to point to the appropriate node after the split
3420 *
3421 * Before splitting this tries to make some room in the node by pushing
3422 * left and right, if either one works, it returns right away.
3423 *
3424 * returns 0 on success and < 0 on failure
3425 */
3426static noinline int split_node(struct btrfs_trans_handle *trans,
3427 struct btrfs_root *root,
3428 struct btrfs_path *path, int level)
3429{
3430 struct btrfs_fs_info *fs_info = root->fs_info;
3431 struct extent_buffer *c;
3432 struct extent_buffer *split;
3433 struct btrfs_disk_key disk_key;
3434 int mid;
3435 int ret;
3436 u32 c_nritems;
3437
3438 c = path->nodes[level];
3439 WARN_ON(btrfs_header_generation(c) != trans->transid);
3440 if (c == root->node) {
3441 /*
3442 * trying to split the root, lets make a new one
3443 *
3444 * tree mod log: We don't log_removal old root in
3445 * insert_new_root, because that root buffer will be kept as a
3446 * normal node. We are going to log removal of half of the
3447 * elements below with tree_mod_log_eb_copy. We're holding a
3448 * tree lock on the buffer, which is why we cannot race with
3449 * other tree_mod_log users.
3450 */
3451 ret = insert_new_root(trans, root, path, level + 1);
3452 if (ret)
3453 return ret;
3454 } else {
3455 ret = push_nodes_for_insert(trans, root, path, level);
3456 c = path->nodes[level];
3457 if (!ret && btrfs_header_nritems(c) <
3458 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3459 return 0;
3460 if (ret < 0)
3461 return ret;
3462 }
3463
3464 c_nritems = btrfs_header_nritems(c);
3465 mid = (c_nritems + 1) / 2;
3466 btrfs_node_key(c, &disk_key, mid);
3467
3468 split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3469 &disk_key, level, c->start, 0);
3470 if (IS_ERR(split))
3471 return PTR_ERR(split);
3472
3473 root_add_used(root, fs_info->nodesize);
3474
3475 memzero_extent_buffer(split, 0, sizeof(struct btrfs_header));
3476 btrfs_set_header_level(split, btrfs_header_level(c));
3477 btrfs_set_header_bytenr(split, split->start);
3478 btrfs_set_header_generation(split, trans->transid);
3479 btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
3480 btrfs_set_header_owner(split, root->root_key.objectid);
3481 write_extent_buffer_fsid(split, fs_info->fsid);
3482 write_extent_buffer_chunk_tree_uuid(split, fs_info->chunk_tree_uuid);
3483
3484 ret = tree_mod_log_eb_copy(fs_info, split, c, 0, mid, c_nritems - mid);
3485 if (ret) {
3486 btrfs_abort_transaction(trans, ret);
3487 return ret;
3488 }
3489 copy_extent_buffer(split, c,
3490 btrfs_node_key_ptr_offset(0),
3491 btrfs_node_key_ptr_offset(mid),
3492 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3493 btrfs_set_header_nritems(split, c_nritems - mid);
3494 btrfs_set_header_nritems(c, mid);
3495 ret = 0;
3496
3497 btrfs_mark_buffer_dirty(c);
3498 btrfs_mark_buffer_dirty(split);
3499
3500 insert_ptr(trans, fs_info, path, &disk_key, split->start,
3501 path->slots[level + 1] + 1, level + 1);
3502
3503 if (path->slots[level] >= mid) {
3504 path->slots[level] -= mid;
3505 btrfs_tree_unlock(c);
3506 free_extent_buffer(c);
3507 path->nodes[level] = split;
3508 path->slots[level + 1] += 1;
3509 } else {
3510 btrfs_tree_unlock(split);
3511 free_extent_buffer(split);
3512 }
3513 return ret;
3514}
3515
3516/*
3517 * how many bytes are required to store the items in a leaf. start
3518 * and nr indicate which items in the leaf to check. This totals up the
3519 * space used both by the item structs and the item data
3520 */
3521static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3522{
3523 struct btrfs_item *start_item;
3524 struct btrfs_item *end_item;
3525 struct btrfs_map_token token;
3526 int data_len;
3527 int nritems = btrfs_header_nritems(l);
3528 int end = min(nritems, start + nr) - 1;
3529
3530 if (!nr)
3531 return 0;
3532 btrfs_init_map_token(&token);
3533 start_item = btrfs_item_nr(start);
3534 end_item = btrfs_item_nr(end);
3535 data_len = btrfs_token_item_offset(l, start_item, &token) +
3536 btrfs_token_item_size(l, start_item, &token);
3537 data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3538 data_len += sizeof(struct btrfs_item) * nr;
3539 WARN_ON(data_len < 0);
3540 return data_len;
3541}
3542
3543/*
3544 * The space between the end of the leaf items and
3545 * the start of the leaf data. IOW, how much room
3546 * the leaf has left for both items and data
3547 */
3548noinline int btrfs_leaf_free_space(struct btrfs_fs_info *fs_info,
3549 struct extent_buffer *leaf)
3550{
3551 int nritems = btrfs_header_nritems(leaf);
3552 int ret;
3553
3554 ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3555 if (ret < 0) {
3556 btrfs_crit(fs_info,
3557 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3558 ret,
3559 (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3560 leaf_space_used(leaf, 0, nritems), nritems);
3561 }
3562 return ret;
3563}
3564
3565/*
3566 * min slot controls the lowest index we're willing to push to the
3567 * right. We'll push up to and including min_slot, but no lower
3568 */
3569static noinline int __push_leaf_right(struct btrfs_fs_info *fs_info,
3570 struct btrfs_path *path,
3571 int data_size, int empty,
3572 struct extent_buffer *right,
3573 int free_space, u32 left_nritems,
3574 u32 min_slot)
3575{
3576 struct extent_buffer *left = path->nodes[0];
3577 struct extent_buffer *upper = path->nodes[1];
3578 struct btrfs_map_token token;
3579 struct btrfs_disk_key disk_key;
3580 int slot;
3581 u32 i;
3582 int push_space = 0;
3583 int push_items = 0;
3584 struct btrfs_item *item;
3585 u32 nr;
3586 u32 right_nritems;
3587 u32 data_end;
3588 u32 this_item_size;
3589
3590 btrfs_init_map_token(&token);
3591
3592 if (empty)
3593 nr = 0;
3594 else
3595 nr = max_t(u32, 1, min_slot);
3596
3597 if (path->slots[0] >= left_nritems)
3598 push_space += data_size;
3599
3600 slot = path->slots[1];
3601 i = left_nritems - 1;
3602 while (i >= nr) {
3603 item = btrfs_item_nr(i);
3604
3605 if (!empty && push_items > 0) {
3606 if (path->slots[0] > i)
3607 break;
3608 if (path->slots[0] == i) {
3609 int space = btrfs_leaf_free_space(fs_info, left);
3610 if (space + push_space * 2 > free_space)
3611 break;
3612 }
3613 }
3614
3615 if (path->slots[0] == i)
3616 push_space += data_size;
3617
3618 this_item_size = btrfs_item_size(left, item);
3619 if (this_item_size + sizeof(*item) + push_space > free_space)
3620 break;
3621
3622 push_items++;
3623 push_space += this_item_size + sizeof(*item);
3624 if (i == 0)
3625 break;
3626 i--;
3627 }
3628
3629 if (push_items == 0)
3630 goto out_unlock;
3631
3632 WARN_ON(!empty && push_items == left_nritems);
3633
3634 /* push left to right */
3635 right_nritems = btrfs_header_nritems(right);
3636
3637 push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3638 push_space -= leaf_data_end(fs_info, left);
3639
3640 /* make room in the right data area */
3641 data_end = leaf_data_end(fs_info, right);
3642 memmove_extent_buffer(right,
3643 BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
3644 BTRFS_LEAF_DATA_OFFSET + data_end,
3645 BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3646
3647 /* copy from the left data area */
3648 copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
3649 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3650 BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, left),
3651 push_space);
3652
3653 memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3654 btrfs_item_nr_offset(0),
3655 right_nritems * sizeof(struct btrfs_item));
3656
3657 /* copy the items from left to right */
3658 copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3659 btrfs_item_nr_offset(left_nritems - push_items),
3660 push_items * sizeof(struct btrfs_item));
3661
3662 /* update the item pointers */
3663 right_nritems += push_items;
3664 btrfs_set_header_nritems(right, right_nritems);
3665 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3666 for (i = 0; i < right_nritems; i++) {
3667 item = btrfs_item_nr(i);
3668 push_space -= btrfs_token_item_size(right, item, &token);
3669 btrfs_set_token_item_offset(right, item, push_space, &token);
3670 }
3671
3672 left_nritems -= push_items;
3673 btrfs_set_header_nritems(left, left_nritems);
3674
3675 if (left_nritems)
3676 btrfs_mark_buffer_dirty(left);
3677 else
3678 clean_tree_block(fs_info, left);
3679
3680 btrfs_mark_buffer_dirty(right);
3681
3682 btrfs_item_key(right, &disk_key, 0);
3683 btrfs_set_node_key(upper, &disk_key, slot + 1);
3684 btrfs_mark_buffer_dirty(upper);
3685
3686 /* then fixup the leaf pointer in the path */
3687 if (path->slots[0] >= left_nritems) {
3688 path->slots[0] -= left_nritems;
3689 if (btrfs_header_nritems(path->nodes[0]) == 0)
3690 clean_tree_block(fs_info, path->nodes[0]);
3691 btrfs_tree_unlock(path->nodes[0]);
3692 free_extent_buffer(path->nodes[0]);
3693 path->nodes[0] = right;
3694 path->slots[1] += 1;
3695 } else {
3696 btrfs_tree_unlock(right);
3697 free_extent_buffer(right);
3698 }
3699 return 0;
3700
3701out_unlock:
3702 btrfs_tree_unlock(right);
3703 free_extent_buffer(right);
3704 return 1;
3705}
3706
3707/*
3708 * push some data in the path leaf to the right, trying to free up at
3709 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3710 *
3711 * returns 1 if the push failed because the other node didn't have enough
3712 * room, 0 if everything worked out and < 0 if there were major errors.
3713 *
3714 * this will push starting from min_slot to the end of the leaf. It won't
3715 * push any slot lower than min_slot
3716 */
3717static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3718 *root, struct btrfs_path *path,
3719 int min_data_size, int data_size,
3720 int empty, u32 min_slot)
3721{
3722 struct btrfs_fs_info *fs_info = root->fs_info;
3723 struct extent_buffer *left = path->nodes[0];
3724 struct extent_buffer *right;
3725 struct extent_buffer *upper;
3726 int slot;
3727 int free_space;
3728 u32 left_nritems;
3729 int ret;
3730
3731 if (!path->nodes[1])
3732 return 1;
3733
3734 slot = path->slots[1];
3735 upper = path->nodes[1];
3736 if (slot >= btrfs_header_nritems(upper) - 1)
3737 return 1;
3738
3739 btrfs_assert_tree_locked(path->nodes[1]);
3740
3741 right = read_node_slot(fs_info, upper, slot + 1);
3742 /*
3743 * slot + 1 is not valid or we fail to read the right node,
3744 * no big deal, just return.
3745 */
3746 if (IS_ERR(right))
3747 return 1;
3748
3749 btrfs_tree_lock(right);
3750 btrfs_set_lock_blocking(right);
3751
3752 free_space = btrfs_leaf_free_space(fs_info, right);
3753 if (free_space < data_size)
3754 goto out_unlock;
3755
3756 /* cow and double check */
3757 ret = btrfs_cow_block(trans, root, right, upper,
3758 slot + 1, &right);
3759 if (ret)
3760 goto out_unlock;
3761
3762 free_space = btrfs_leaf_free_space(fs_info, right);
3763 if (free_space < data_size)
3764 goto out_unlock;
3765
3766 left_nritems = btrfs_header_nritems(left);
3767 if (left_nritems == 0)
3768 goto out_unlock;
3769
3770 if (path->slots[0] == left_nritems && !empty) {
3771 /* Key greater than all keys in the leaf, right neighbor has
3772 * enough room for it and we're not emptying our leaf to delete
3773 * it, therefore use right neighbor to insert the new item and
3774 * no need to touch/dirty our left leaft. */
3775 btrfs_tree_unlock(left);
3776 free_extent_buffer(left);
3777 path->nodes[0] = right;
3778 path->slots[0] = 0;
3779 path->slots[1]++;
3780 return 0;
3781 }
3782
3783 return __push_leaf_right(fs_info, path, min_data_size, empty,
3784 right, free_space, left_nritems, min_slot);
3785out_unlock:
3786 btrfs_tree_unlock(right);
3787 free_extent_buffer(right);
3788 return 1;
3789}
3790
3791/*
3792 * push some data in the path leaf to the left, trying to free up at
3793 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3794 *
3795 * max_slot can put a limit on how far into the leaf we'll push items. The
3796 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3797 * items
3798 */
3799static noinline int __push_leaf_left(struct btrfs_fs_info *fs_info,
3800 struct btrfs_path *path, int data_size,
3801 int empty, struct extent_buffer *left,
3802 int free_space, u32 right_nritems,
3803 u32 max_slot)
3804{
3805 struct btrfs_disk_key disk_key;
3806 struct extent_buffer *right = path->nodes[0];
3807 int i;
3808 int push_space = 0;
3809 int push_items = 0;
3810 struct btrfs_item *item;
3811 u32 old_left_nritems;
3812 u32 nr;
3813 int ret = 0;
3814 u32 this_item_size;
3815 u32 old_left_item_size;
3816 struct btrfs_map_token token;
3817
3818 btrfs_init_map_token(&token);
3819
3820 if (empty)
3821 nr = min(right_nritems, max_slot);
3822 else
3823 nr = min(right_nritems - 1, max_slot);
3824
3825 for (i = 0; i < nr; i++) {
3826 item = btrfs_item_nr(i);
3827
3828 if (!empty && push_items > 0) {
3829 if (path->slots[0] < i)
3830 break;
3831 if (path->slots[0] == i) {
3832 int space = btrfs_leaf_free_space(fs_info, right);
3833 if (space + push_space * 2 > free_space)
3834 break;
3835 }
3836 }
3837
3838 if (path->slots[0] == i)
3839 push_space += data_size;
3840
3841 this_item_size = btrfs_item_size(right, item);
3842 if (this_item_size + sizeof(*item) + push_space > free_space)
3843 break;
3844
3845 push_items++;
3846 push_space += this_item_size + sizeof(*item);
3847 }
3848
3849 if (push_items == 0) {
3850 ret = 1;
3851 goto out;
3852 }
3853 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3854
3855 /* push data from right to left */
3856 copy_extent_buffer(left, right,
3857 btrfs_item_nr_offset(btrfs_header_nritems(left)),
3858 btrfs_item_nr_offset(0),
3859 push_items * sizeof(struct btrfs_item));
3860
3861 push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3862 btrfs_item_offset_nr(right, push_items - 1);
3863
3864 copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
3865 leaf_data_end(fs_info, left) - push_space,
3866 BTRFS_LEAF_DATA_OFFSET +
3867 btrfs_item_offset_nr(right, push_items - 1),
3868 push_space);
3869 old_left_nritems = btrfs_header_nritems(left);
3870 BUG_ON(old_left_nritems <= 0);
3871
3872 old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3873 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3874 u32 ioff;
3875
3876 item = btrfs_item_nr(i);
3877
3878 ioff = btrfs_token_item_offset(left, item, &token);
3879 btrfs_set_token_item_offset(left, item,
3880 ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size),
3881 &token);
3882 }
3883 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3884
3885 /* fixup right node */
3886 if (push_items > right_nritems)
3887 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3888 right_nritems);
3889
3890 if (push_items < right_nritems) {
3891 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3892 leaf_data_end(fs_info, right);
3893 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
3894 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3895 BTRFS_LEAF_DATA_OFFSET +
3896 leaf_data_end(fs_info, right), push_space);
3897
3898 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3899 btrfs_item_nr_offset(push_items),
3900 (btrfs_header_nritems(right) - push_items) *
3901 sizeof(struct btrfs_item));
3902 }
3903 right_nritems -= push_items;
3904 btrfs_set_header_nritems(right, right_nritems);
3905 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3906 for (i = 0; i < right_nritems; i++) {
3907 item = btrfs_item_nr(i);
3908
3909 push_space = push_space - btrfs_token_item_size(right,
3910 item, &token);
3911 btrfs_set_token_item_offset(right, item, push_space, &token);
3912 }
3913
3914 btrfs_mark_buffer_dirty(left);
3915 if (right_nritems)
3916 btrfs_mark_buffer_dirty(right);
3917 else
3918 clean_tree_block(fs_info, right);
3919
3920 btrfs_item_key(right, &disk_key, 0);
3921 fixup_low_keys(fs_info, path, &disk_key, 1);
3922
3923 /* then fixup the leaf pointer in the path */
3924 if (path->slots[0] < push_items) {
3925 path->slots[0] += old_left_nritems;
3926 btrfs_tree_unlock(path->nodes[0]);
3927 free_extent_buffer(path->nodes[0]);
3928 path->nodes[0] = left;
3929 path->slots[1] -= 1;
3930 } else {
3931 btrfs_tree_unlock(left);
3932 free_extent_buffer(left);
3933 path->slots[0] -= push_items;
3934 }
3935 BUG_ON(path->slots[0] < 0);
3936 return ret;
3937out:
3938 btrfs_tree_unlock(left);
3939 free_extent_buffer(left);
3940 return ret;
3941}
3942
3943/*
3944 * push some data in the path leaf to the left, trying to free up at
3945 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3946 *
3947 * max_slot can put a limit on how far into the leaf we'll push items. The
3948 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
3949 * items
3950 */
3951static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3952 *root, struct btrfs_path *path, int min_data_size,
3953 int data_size, int empty, u32 max_slot)
3954{
3955 struct btrfs_fs_info *fs_info = root->fs_info;
3956 struct extent_buffer *right = path->nodes[0];
3957 struct extent_buffer *left;
3958 int slot;
3959 int free_space;
3960 u32 right_nritems;
3961 int ret = 0;
3962
3963 slot = path->slots[1];
3964 if (slot == 0)
3965 return 1;
3966 if (!path->nodes[1])
3967 return 1;
3968
3969 right_nritems = btrfs_header_nritems(right);
3970 if (right_nritems == 0)
3971 return 1;
3972
3973 btrfs_assert_tree_locked(path->nodes[1]);
3974
3975 left = read_node_slot(fs_info, path->nodes[1], slot - 1);
3976 /*
3977 * slot - 1 is not valid or we fail to read the left node,
3978 * no big deal, just return.
3979 */
3980 if (IS_ERR(left))
3981 return 1;
3982
3983 btrfs_tree_lock(left);
3984 btrfs_set_lock_blocking(left);
3985
3986 free_space = btrfs_leaf_free_space(fs_info, left);
3987 if (free_space < data_size) {
3988 ret = 1;
3989 goto out;
3990 }
3991
3992 /* cow and double check */
3993 ret = btrfs_cow_block(trans, root, left,
3994 path->nodes[1], slot - 1, &left);
3995 if (ret) {
3996 /* we hit -ENOSPC, but it isn't fatal here */
3997 if (ret == -ENOSPC)
3998 ret = 1;
3999 goto out;
4000 }
4001
4002 free_space = btrfs_leaf_free_space(fs_info, left);
4003 if (free_space < data_size) {
4004 ret = 1;
4005 goto out;
4006 }
4007
4008 return __push_leaf_left(fs_info, path, min_data_size,
4009 empty, left, free_space, right_nritems,
4010 max_slot);
4011out:
4012 btrfs_tree_unlock(left);
4013 free_extent_buffer(left);
4014 return ret;
4015}
4016
4017/*
4018 * split the path's leaf in two, making sure there is at least data_size
4019 * available for the resulting leaf level of the path.
4020 */
4021static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4022 struct btrfs_fs_info *fs_info,
4023 struct btrfs_path *path,
4024 struct extent_buffer *l,
4025 struct extent_buffer *right,
4026 int slot, int mid, int nritems)
4027{
4028 int data_copy_size;
4029 int rt_data_off;
4030 int i;
4031 struct btrfs_disk_key disk_key;
4032 struct btrfs_map_token token;
4033
4034 btrfs_init_map_token(&token);
4035
4036 nritems = nritems - mid;
4037 btrfs_set_header_nritems(right, nritems);
4038 data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(fs_info, l);
4039
4040 copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4041 btrfs_item_nr_offset(mid),
4042 nritems * sizeof(struct btrfs_item));
4043
4044 copy_extent_buffer(right, l,
4045 BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
4046 data_copy_size, BTRFS_LEAF_DATA_OFFSET +
4047 leaf_data_end(fs_info, l), data_copy_size);
4048
4049 rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4050
4051 for (i = 0; i < nritems; i++) {
4052 struct btrfs_item *item = btrfs_item_nr(i);
4053 u32 ioff;
4054
4055 ioff = btrfs_token_item_offset(right, item, &token);
4056 btrfs_set_token_item_offset(right, item,
4057 ioff + rt_data_off, &token);
4058 }
4059
4060 btrfs_set_header_nritems(l, mid);
4061 btrfs_item_key(right, &disk_key, 0);
4062 insert_ptr(trans, fs_info, path, &disk_key, right->start,
4063 path->slots[1] + 1, 1);
4064
4065 btrfs_mark_buffer_dirty(right);
4066 btrfs_mark_buffer_dirty(l);
4067 BUG_ON(path->slots[0] != slot);
4068
4069 if (mid <= slot) {
4070 btrfs_tree_unlock(path->nodes[0]);
4071 free_extent_buffer(path->nodes[0]);
4072 path->nodes[0] = right;
4073 path->slots[0] -= mid;
4074 path->slots[1] += 1;
4075 } else {
4076 btrfs_tree_unlock(right);
4077 free_extent_buffer(right);
4078 }
4079
4080 BUG_ON(path->slots[0] < 0);
4081}
4082
4083/*
4084 * double splits happen when we need to insert a big item in the middle
4085 * of a leaf. A double split can leave us with 3 mostly empty leaves:
4086 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4087 * A B C
4088 *
4089 * We avoid this by trying to push the items on either side of our target
4090 * into the adjacent leaves. If all goes well we can avoid the double split
4091 * completely.
4092 */
4093static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4094 struct btrfs_root *root,
4095 struct btrfs_path *path,
4096 int data_size)
4097{
4098 struct btrfs_fs_info *fs_info = root->fs_info;
4099 int ret;
4100 int progress = 0;
4101 int slot;
4102 u32 nritems;
4103 int space_needed = data_size;
4104
4105 slot = path->slots[0];
4106 if (slot < btrfs_header_nritems(path->nodes[0]))
4107 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4108
4109 /*
4110 * try to push all the items after our slot into the
4111 * right leaf
4112 */
4113 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4114 if (ret < 0)
4115 return ret;
4116
4117 if (ret == 0)
4118 progress++;
4119
4120 nritems = btrfs_header_nritems(path->nodes[0]);
4121 /*
4122 * our goal is to get our slot at the start or end of a leaf. If
4123 * we've done so we're done
4124 */
4125 if (path->slots[0] == 0 || path->slots[0] == nritems)
4126 return 0;
4127
4128 if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4129 return 0;
4130
4131 /* try to push all the items before our slot into the next leaf */
4132 slot = path->slots[0];
4133 space_needed = data_size;
4134 if (slot > 0)
4135 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4136 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4137 if (ret < 0)
4138 return ret;
4139
4140 if (ret == 0)
4141 progress++;
4142
4143 if (progress)
4144 return 0;
4145 return 1;
4146}
4147
4148/*
4149 * split the path's leaf in two, making sure there is at least data_size
4150 * available for the resulting leaf level of the path.
4151 *
4152 * returns 0 if all went well and < 0 on failure.
4153 */
4154static noinline int split_leaf(struct btrfs_trans_handle *trans,
4155 struct btrfs_root *root,
4156 const struct btrfs_key *ins_key,
4157 struct btrfs_path *path, int data_size,
4158 int extend)
4159{
4160 struct btrfs_disk_key disk_key;
4161 struct extent_buffer *l;
4162 u32 nritems;
4163 int mid;
4164 int slot;
4165 struct extent_buffer *right;
4166 struct btrfs_fs_info *fs_info = root->fs_info;
4167 int ret = 0;
4168 int wret;
4169 int split;
4170 int num_doubles = 0;
4171 int tried_avoid_double = 0;
4172
4173 l = path->nodes[0];
4174 slot = path->slots[0];
4175 if (extend && data_size + btrfs_item_size_nr(l, slot) +
4176 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4177 return -EOVERFLOW;
4178
4179 /* first try to make some room by pushing left and right */
4180 if (data_size && path->nodes[1]) {
4181 int space_needed = data_size;
4182
4183 if (slot < btrfs_header_nritems(l))
4184 space_needed -= btrfs_leaf_free_space(fs_info, l);
4185
4186 wret = push_leaf_right(trans, root, path, space_needed,
4187 space_needed, 0, 0);
4188 if (wret < 0)
4189 return wret;
4190 if (wret) {
4191 space_needed = data_size;
4192 if (slot > 0)
4193 space_needed -= btrfs_leaf_free_space(fs_info,
4194 l);
4195 wret = push_leaf_left(trans, root, path, space_needed,
4196 space_needed, 0, (u32)-1);
4197 if (wret < 0)
4198 return wret;
4199 }
4200 l = path->nodes[0];
4201
4202 /* did the pushes work? */
4203 if (btrfs_leaf_free_space(fs_info, l) >= data_size)
4204 return 0;
4205 }
4206
4207 if (!path->nodes[1]) {
4208 ret = insert_new_root(trans, root, path, 1);
4209 if (ret)
4210 return ret;
4211 }
4212again:
4213 split = 1;
4214 l = path->nodes[0];
4215 slot = path->slots[0];
4216 nritems = btrfs_header_nritems(l);
4217 mid = (nritems + 1) / 2;
4218
4219 if (mid <= slot) {
4220 if (nritems == 1 ||
4221 leaf_space_used(l, mid, nritems - mid) + data_size >
4222 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4223 if (slot >= nritems) {
4224 split = 0;
4225 } else {
4226 mid = slot;
4227 if (mid != nritems &&
4228 leaf_space_used(l, mid, nritems - mid) +
4229 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4230 if (data_size && !tried_avoid_double)
4231 goto push_for_double;
4232 split = 2;
4233 }
4234 }
4235 }
4236 } else {
4237 if (leaf_space_used(l, 0, mid) + data_size >
4238 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4239 if (!extend && data_size && slot == 0) {
4240 split = 0;
4241 } else if ((extend || !data_size) && slot == 0) {
4242 mid = 1;
4243 } else {
4244 mid = slot;
4245 if (mid != nritems &&
4246 leaf_space_used(l, mid, nritems - mid) +
4247 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4248 if (data_size && !tried_avoid_double)
4249 goto push_for_double;
4250 split = 2;
4251 }
4252 }
4253 }
4254 }
4255
4256 if (split == 0)
4257 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4258 else
4259 btrfs_item_key(l, &disk_key, mid);
4260
4261 right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
4262 &disk_key, 0, l->start, 0);
4263 if (IS_ERR(right))
4264 return PTR_ERR(right);
4265
4266 root_add_used(root, fs_info->nodesize);
4267
4268 memzero_extent_buffer(right, 0, sizeof(struct btrfs_header));
4269 btrfs_set_header_bytenr(right, right->start);
4270 btrfs_set_header_generation(right, trans->transid);
4271 btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
4272 btrfs_set_header_owner(right, root->root_key.objectid);
4273 btrfs_set_header_level(right, 0);
4274 write_extent_buffer_fsid(right, fs_info->fsid);
4275 write_extent_buffer_chunk_tree_uuid(right, fs_info->chunk_tree_uuid);
4276
4277 if (split == 0) {
4278 if (mid <= slot) {
4279 btrfs_set_header_nritems(right, 0);
4280 insert_ptr(trans, fs_info, path, &disk_key,
4281 right->start, path->slots[1] + 1, 1);
4282 btrfs_tree_unlock(path->nodes[0]);
4283 free_extent_buffer(path->nodes[0]);
4284 path->nodes[0] = right;
4285 path->slots[0] = 0;
4286 path->slots[1] += 1;
4287 } else {
4288 btrfs_set_header_nritems(right, 0);
4289 insert_ptr(trans, fs_info, path, &disk_key,
4290 right->start, path->slots[1], 1);
4291 btrfs_tree_unlock(path->nodes[0]);
4292 free_extent_buffer(path->nodes[0]);
4293 path->nodes[0] = right;
4294 path->slots[0] = 0;
4295 if (path->slots[1] == 0)
4296 fixup_low_keys(fs_info, path, &disk_key, 1);
4297 }
4298 /*
4299 * We create a new leaf 'right' for the required ins_len and
4300 * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4301 * the content of ins_len to 'right'.
4302 */
4303 return ret;
4304 }
4305
4306 copy_for_split(trans, fs_info, path, l, right, slot, mid, nritems);
4307
4308 if (split == 2) {
4309 BUG_ON(num_doubles != 0);
4310 num_doubles++;
4311 goto again;
4312 }
4313
4314 return 0;
4315
4316push_for_double:
4317 push_for_double_split(trans, root, path, data_size);
4318 tried_avoid_double = 1;
4319 if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4320 return 0;
4321 goto again;
4322}
4323
4324static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4325 struct btrfs_root *root,
4326 struct btrfs_path *path, int ins_len)
4327{
4328 struct btrfs_fs_info *fs_info = root->fs_info;
4329 struct btrfs_key key;
4330 struct extent_buffer *leaf;
4331 struct btrfs_file_extent_item *fi;
4332 u64 extent_len = 0;
4333 u32 item_size;
4334 int ret;
4335
4336 leaf = path->nodes[0];
4337 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4338
4339 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4340 key.type != BTRFS_EXTENT_CSUM_KEY);
4341
4342 if (btrfs_leaf_free_space(fs_info, leaf) >= ins_len)
4343 return 0;
4344
4345 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4346 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4347 fi = btrfs_item_ptr(leaf, path->slots[0],
4348 struct btrfs_file_extent_item);
4349 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4350 }
4351 btrfs_release_path(path);
4352
4353 path->keep_locks = 1;
4354 path->search_for_split = 1;
4355 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4356 path->search_for_split = 0;
4357 if (ret > 0)
4358 ret = -EAGAIN;
4359 if (ret < 0)
4360 goto err;
4361
4362 ret = -EAGAIN;
4363 leaf = path->nodes[0];
4364 /* if our item isn't there, return now */
4365 if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4366 goto err;
4367
4368 /* the leaf has changed, it now has room. return now */
4369 if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= ins_len)
4370 goto err;
4371
4372 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4373 fi = btrfs_item_ptr(leaf, path->slots[0],
4374 struct btrfs_file_extent_item);
4375 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4376 goto err;
4377 }
4378
4379 btrfs_set_path_blocking(path);
4380 ret = split_leaf(trans, root, &key, path, ins_len, 1);
4381 if (ret)
4382 goto err;
4383
4384 path->keep_locks = 0;
4385 btrfs_unlock_up_safe(path, 1);
4386 return 0;
4387err:
4388 path->keep_locks = 0;
4389 return ret;
4390}
4391
4392static noinline int split_item(struct btrfs_fs_info *fs_info,
4393 struct btrfs_path *path,
4394 const struct btrfs_key *new_key,
4395 unsigned long split_offset)
4396{
4397 struct extent_buffer *leaf;
4398 struct btrfs_item *item;
4399 struct btrfs_item *new_item;
4400 int slot;
4401 char *buf;
4402 u32 nritems;
4403 u32 item_size;
4404 u32 orig_offset;
4405 struct btrfs_disk_key disk_key;
4406
4407 leaf = path->nodes[0];
4408 BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < sizeof(struct btrfs_item));
4409
4410 btrfs_set_path_blocking(path);
4411
4412 item = btrfs_item_nr(path->slots[0]);
4413 orig_offset = btrfs_item_offset(leaf, item);
4414 item_size = btrfs_item_size(leaf, item);
4415
4416 buf = kmalloc(item_size, GFP_NOFS);
4417 if (!buf)
4418 return -ENOMEM;
4419
4420 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4421 path->slots[0]), item_size);
4422
4423 slot = path->slots[0] + 1;
4424 nritems = btrfs_header_nritems(leaf);
4425 if (slot != nritems) {
4426 /* shift the items */
4427 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4428 btrfs_item_nr_offset(slot),
4429 (nritems - slot) * sizeof(struct btrfs_item));
4430 }
4431
4432 btrfs_cpu_key_to_disk(&disk_key, new_key);
4433 btrfs_set_item_key(leaf, &disk_key, slot);
4434
4435 new_item = btrfs_item_nr(slot);
4436
4437 btrfs_set_item_offset(leaf, new_item, orig_offset);
4438 btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4439
4440 btrfs_set_item_offset(leaf, item,
4441 orig_offset + item_size - split_offset);
4442 btrfs_set_item_size(leaf, item, split_offset);
4443
4444 btrfs_set_header_nritems(leaf, nritems + 1);
4445
4446 /* write the data for the start of the original item */
4447 write_extent_buffer(leaf, buf,
4448 btrfs_item_ptr_offset(leaf, path->slots[0]),
4449 split_offset);
4450
4451 /* write the data for the new item */
4452 write_extent_buffer(leaf, buf + split_offset,
4453 btrfs_item_ptr_offset(leaf, slot),
4454 item_size - split_offset);
4455 btrfs_mark_buffer_dirty(leaf);
4456
4457 BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < 0);
4458 kfree(buf);
4459 return 0;
4460}
4461
4462/*
4463 * This function splits a single item into two items,
4464 * giving 'new_key' to the new item and splitting the
4465 * old one at split_offset (from the start of the item).
4466 *
4467 * The path may be released by this operation. After
4468 * the split, the path is pointing to the old item. The
4469 * new item is going to be in the same node as the old one.
4470 *
4471 * Note, the item being split must be smaller enough to live alone on
4472 * a tree block with room for one extra struct btrfs_item
4473 *
4474 * This allows us to split the item in place, keeping a lock on the
4475 * leaf the entire time.
4476 */
4477int btrfs_split_item(struct btrfs_trans_handle *trans,
4478 struct btrfs_root *root,
4479 struct btrfs_path *path,
4480 const struct btrfs_key *new_key,
4481 unsigned long split_offset)
4482{
4483 int ret;
4484 ret = setup_leaf_for_split(trans, root, path,
4485 sizeof(struct btrfs_item));
4486 if (ret)
4487 return ret;
4488
4489 ret = split_item(root->fs_info, path, new_key, split_offset);
4490 return ret;
4491}
4492
4493/*
4494 * This function duplicate a item, giving 'new_key' to the new item.
4495 * It guarantees both items live in the same tree leaf and the new item
4496 * is contiguous with the original item.
4497 *
4498 * This allows us to split file extent in place, keeping a lock on the
4499 * leaf the entire time.
4500 */
4501int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4502 struct btrfs_root *root,
4503 struct btrfs_path *path,
4504 const struct btrfs_key *new_key)
4505{
4506 struct extent_buffer *leaf;
4507 int ret;
4508 u32 item_size;
4509
4510 leaf = path->nodes[0];
4511 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4512 ret = setup_leaf_for_split(trans, root, path,
4513 item_size + sizeof(struct btrfs_item));
4514 if (ret)
4515 return ret;
4516
4517 path->slots[0]++;
4518 setup_items_for_insert(root, path, new_key, &item_size,
4519 item_size, item_size +
4520 sizeof(struct btrfs_item), 1);
4521 leaf = path->nodes[0];
4522 memcpy_extent_buffer(leaf,
4523 btrfs_item_ptr_offset(leaf, path->slots[0]),
4524 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4525 item_size);
4526 return 0;
4527}
4528
4529/*
4530 * make the item pointed to by the path smaller. new_size indicates
4531 * how small to make it, and from_end tells us if we just chop bytes
4532 * off the end of the item or if we shift the item to chop bytes off
4533 * the front.
4534 */
4535void btrfs_truncate_item(struct btrfs_fs_info *fs_info,
4536 struct btrfs_path *path, u32 new_size, int from_end)
4537{
4538 int slot;
4539 struct extent_buffer *leaf;
4540 struct btrfs_item *item;
4541 u32 nritems;
4542 unsigned int data_end;
4543 unsigned int old_data_start;
4544 unsigned int old_size;
4545 unsigned int size_diff;
4546 int i;
4547 struct btrfs_map_token token;
4548
4549 btrfs_init_map_token(&token);
4550
4551 leaf = path->nodes[0];
4552 slot = path->slots[0];
4553
4554 old_size = btrfs_item_size_nr(leaf, slot);
4555 if (old_size == new_size)
4556 return;
4557
4558 nritems = btrfs_header_nritems(leaf);
4559 data_end = leaf_data_end(fs_info, leaf);
4560
4561 old_data_start = btrfs_item_offset_nr(leaf, slot);
4562
4563 size_diff = old_size - new_size;
4564
4565 BUG_ON(slot < 0);
4566 BUG_ON(slot >= nritems);
4567
4568 /*
4569 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4570 */
4571 /* first correct the data pointers */
4572 for (i = slot; i < nritems; i++) {
4573 u32 ioff;
4574 item = btrfs_item_nr(i);
4575
4576 ioff = btrfs_token_item_offset(leaf, item, &token);
4577 btrfs_set_token_item_offset(leaf, item,
4578 ioff + size_diff, &token);
4579 }
4580
4581 /* shift the data */
4582 if (from_end) {
4583 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4584 data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4585 data_end, old_data_start + new_size - data_end);
4586 } else {
4587 struct btrfs_disk_key disk_key;
4588 u64 offset;
4589
4590 btrfs_item_key(leaf, &disk_key, slot);
4591
4592 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4593 unsigned long ptr;
4594 struct btrfs_file_extent_item *fi;
4595
4596 fi = btrfs_item_ptr(leaf, slot,
4597 struct btrfs_file_extent_item);
4598 fi = (struct btrfs_file_extent_item *)(
4599 (unsigned long)fi - size_diff);
4600
4601 if (btrfs_file_extent_type(leaf, fi) ==
4602 BTRFS_FILE_EXTENT_INLINE) {
4603 ptr = btrfs_item_ptr_offset(leaf, slot);
4604 memmove_extent_buffer(leaf, ptr,
4605 (unsigned long)fi,
4606 BTRFS_FILE_EXTENT_INLINE_DATA_START);
4607 }
4608 }
4609
4610 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4611 data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4612 data_end, old_data_start - data_end);
4613
4614 offset = btrfs_disk_key_offset(&disk_key);
4615 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4616 btrfs_set_item_key(leaf, &disk_key, slot);
4617 if (slot == 0)
4618 fixup_low_keys(fs_info, path, &disk_key, 1);
4619 }
4620
4621 item = btrfs_item_nr(slot);
4622 btrfs_set_item_size(leaf, item, new_size);
4623 btrfs_mark_buffer_dirty(leaf);
4624
4625 if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4626 btrfs_print_leaf(leaf);
4627 BUG();
4628 }
4629}
4630
4631/*
4632 * make the item pointed to by the path bigger, data_size is the added size.
4633 */
4634void btrfs_extend_item(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
4635 u32 data_size)
4636{
4637 int slot;
4638 struct extent_buffer *leaf;
4639 struct btrfs_item *item;
4640 u32 nritems;
4641 unsigned int data_end;
4642 unsigned int old_data;
4643 unsigned int old_size;
4644 int i;
4645 struct btrfs_map_token token;
4646
4647 btrfs_init_map_token(&token);
4648
4649 leaf = path->nodes[0];
4650
4651 nritems = btrfs_header_nritems(leaf);
4652 data_end = leaf_data_end(fs_info, leaf);
4653
4654 if (btrfs_leaf_free_space(fs_info, leaf) < data_size) {
4655 btrfs_print_leaf(leaf);
4656 BUG();
4657 }
4658 slot = path->slots[0];
4659 old_data = btrfs_item_end_nr(leaf, slot);
4660
4661 BUG_ON(slot < 0);
4662 if (slot >= nritems) {
4663 btrfs_print_leaf(leaf);
4664 btrfs_crit(fs_info, "slot %d too large, nritems %d",
4665 slot, nritems);
4666 BUG_ON(1);
4667 }
4668
4669 /*
4670 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4671 */
4672 /* first correct the data pointers */
4673 for (i = slot; i < nritems; i++) {
4674 u32 ioff;
4675 item = btrfs_item_nr(i);
4676
4677 ioff = btrfs_token_item_offset(leaf, item, &token);
4678 btrfs_set_token_item_offset(leaf, item,
4679 ioff - data_size, &token);
4680 }
4681
4682 /* shift the data */
4683 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4684 data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
4685 data_end, old_data - data_end);
4686
4687 data_end = old_data;
4688 old_size = btrfs_item_size_nr(leaf, slot);
4689 item = btrfs_item_nr(slot);
4690 btrfs_set_item_size(leaf, item, old_size + data_size);
4691 btrfs_mark_buffer_dirty(leaf);
4692
4693 if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4694 btrfs_print_leaf(leaf);
4695 BUG();
4696 }
4697}
4698
4699/*
4700 * this is a helper for btrfs_insert_empty_items, the main goal here is
4701 * to save stack depth by doing the bulk of the work in a function
4702 * that doesn't call btrfs_search_slot
4703 */
4704void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4705 const struct btrfs_key *cpu_key, u32 *data_size,
4706 u32 total_data, u32 total_size, int nr)
4707{
4708 struct btrfs_fs_info *fs_info = root->fs_info;
4709 struct btrfs_item *item;
4710 int i;
4711 u32 nritems;
4712 unsigned int data_end;
4713 struct btrfs_disk_key disk_key;
4714 struct extent_buffer *leaf;
4715 int slot;
4716 struct btrfs_map_token token;
4717
4718 if (path->slots[0] == 0) {
4719 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4720 fixup_low_keys(fs_info, path, &disk_key, 1);
4721 }
4722 btrfs_unlock_up_safe(path, 1);
4723
4724 btrfs_init_map_token(&token);
4725
4726 leaf = path->nodes[0];
4727 slot = path->slots[0];
4728
4729 nritems = btrfs_header_nritems(leaf);
4730 data_end = leaf_data_end(fs_info, leaf);
4731
4732 if (btrfs_leaf_free_space(fs_info, leaf) < total_size) {
4733 btrfs_print_leaf(leaf);
4734 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4735 total_size, btrfs_leaf_free_space(fs_info, leaf));
4736 BUG();
4737 }
4738
4739 if (slot != nritems) {
4740 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4741
4742 if (old_data < data_end) {
4743 btrfs_print_leaf(leaf);
4744 btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4745 slot, old_data, data_end);
4746 BUG_ON(1);
4747 }
4748 /*
4749 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4750 */
4751 /* first correct the data pointers */
4752 for (i = slot; i < nritems; i++) {
4753 u32 ioff;
4754
4755 item = btrfs_item_nr(i);
4756 ioff = btrfs_token_item_offset(leaf, item, &token);
4757 btrfs_set_token_item_offset(leaf, item,
4758 ioff - total_data, &token);
4759 }
4760 /* shift the items */
4761 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4762 btrfs_item_nr_offset(slot),
4763 (nritems - slot) * sizeof(struct btrfs_item));
4764
4765 /* shift the data */
4766 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4767 data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
4768 data_end, old_data - data_end);
4769 data_end = old_data;
4770 }
4771
4772 /* setup the item for the new data */
4773 for (i = 0; i < nr; i++) {
4774 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4775 btrfs_set_item_key(leaf, &disk_key, slot + i);
4776 item = btrfs_item_nr(slot + i);
4777 btrfs_set_token_item_offset(leaf, item,
4778 data_end - data_size[i], &token);
4779 data_end -= data_size[i];
4780 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4781 }
4782
4783 btrfs_set_header_nritems(leaf, nritems + nr);
4784 btrfs_mark_buffer_dirty(leaf);
4785
4786 if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4787 btrfs_print_leaf(leaf);
4788 BUG();
4789 }
4790}
4791
4792/*
4793 * Given a key and some data, insert items into the tree.
4794 * This does all the path init required, making room in the tree if needed.
4795 */
4796int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4797 struct btrfs_root *root,
4798 struct btrfs_path *path,
4799 const struct btrfs_key *cpu_key, u32 *data_size,
4800 int nr)
4801{
4802 int ret = 0;
4803 int slot;
4804 int i;
4805 u32 total_size = 0;
4806 u32 total_data = 0;
4807
4808 for (i = 0; i < nr; i++)
4809 total_data += data_size[i];
4810
4811 total_size = total_data + (nr * sizeof(struct btrfs_item));
4812 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4813 if (ret == 0)
4814 return -EEXIST;
4815 if (ret < 0)
4816 return ret;
4817
4818 slot = path->slots[0];
4819 BUG_ON(slot < 0);
4820
4821 setup_items_for_insert(root, path, cpu_key, data_size,
4822 total_data, total_size, nr);
4823 return 0;
4824}
4825
4826/*
4827 * Given a key and some data, insert an item into the tree.
4828 * This does all the path init required, making room in the tree if needed.
4829 */
4830int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4831 const struct btrfs_key *cpu_key, void *data,
4832 u32 data_size)
4833{
4834 int ret = 0;
4835 struct btrfs_path *path;
4836 struct extent_buffer *leaf;
4837 unsigned long ptr;
4838
4839 path = btrfs_alloc_path();
4840 if (!path)
4841 return -ENOMEM;
4842 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4843 if (!ret) {
4844 leaf = path->nodes[0];
4845 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4846 write_extent_buffer(leaf, data, ptr, data_size);
4847 btrfs_mark_buffer_dirty(leaf);
4848 }
4849 btrfs_free_path(path);
4850 return ret;
4851}
4852
4853/*
4854 * delete the pointer from a given node.
4855 *
4856 * the tree should have been previously balanced so the deletion does not
4857 * empty a node.
4858 */
4859static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4860 int level, int slot)
4861{
4862 struct btrfs_fs_info *fs_info = root->fs_info;
4863 struct extent_buffer *parent = path->nodes[level];
4864 u32 nritems;
4865 int ret;
4866
4867 nritems = btrfs_header_nritems(parent);
4868 if (slot != nritems - 1) {
4869 if (level) {
4870 ret = tree_mod_log_insert_move(parent, slot, slot + 1,
4871 nritems - slot - 1);
4872 BUG_ON(ret < 0);
4873 }
4874 memmove_extent_buffer(parent,
4875 btrfs_node_key_ptr_offset(slot),
4876 btrfs_node_key_ptr_offset(slot + 1),
4877 sizeof(struct btrfs_key_ptr) *
4878 (nritems - slot - 1));
4879 } else if (level) {
4880 ret = tree_mod_log_insert_key(parent, slot, MOD_LOG_KEY_REMOVE,
4881 GFP_NOFS);
4882 BUG_ON(ret < 0);
4883 }
4884
4885 nritems--;
4886 btrfs_set_header_nritems(parent, nritems);
4887 if (nritems == 0 && parent == root->node) {
4888 BUG_ON(btrfs_header_level(root->node) != 1);
4889 /* just turn the root into a leaf and break */
4890 btrfs_set_header_level(root->node, 0);
4891 } else if (slot == 0) {
4892 struct btrfs_disk_key disk_key;
4893
4894 btrfs_node_key(parent, &disk_key, 0);
4895 fixup_low_keys(fs_info, path, &disk_key, level + 1);
4896 }
4897 btrfs_mark_buffer_dirty(parent);
4898}
4899
4900/*
4901 * a helper function to delete the leaf pointed to by path->slots[1] and
4902 * path->nodes[1].
4903 *
4904 * This deletes the pointer in path->nodes[1] and frees the leaf
4905 * block extent. zero is returned if it all worked out, < 0 otherwise.
4906 *
4907 * The path must have already been setup for deleting the leaf, including
4908 * all the proper balancing. path->nodes[1] must be locked.
4909 */
4910static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4911 struct btrfs_root *root,
4912 struct btrfs_path *path,
4913 struct extent_buffer *leaf)
4914{
4915 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4916 del_ptr(root, path, 1, path->slots[1]);
4917
4918 /*
4919 * btrfs_free_extent is expensive, we want to make sure we
4920 * aren't holding any locks when we call it
4921 */
4922 btrfs_unlock_up_safe(path, 0);
4923
4924 root_sub_used(root, leaf->len);
4925
4926 extent_buffer_get(leaf);
4927 btrfs_free_tree_block(trans, root, leaf, 0, 1);
4928 free_extent_buffer_stale(leaf);
4929}
4930/*
4931 * delete the item at the leaf level in path. If that empties
4932 * the leaf, remove it from the tree
4933 */
4934int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4935 struct btrfs_path *path, int slot, int nr)
4936{
4937 struct btrfs_fs_info *fs_info = root->fs_info;
4938 struct extent_buffer *leaf;
4939 struct btrfs_item *item;
4940 u32 last_off;
4941 u32 dsize = 0;
4942 int ret = 0;
4943 int wret;
4944 int i;
4945 u32 nritems;
4946 struct btrfs_map_token token;
4947
4948 btrfs_init_map_token(&token);
4949
4950 leaf = path->nodes[0];
4951 last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4952
4953 for (i = 0; i < nr; i++)
4954 dsize += btrfs_item_size_nr(leaf, slot + i);
4955
4956 nritems = btrfs_header_nritems(leaf);
4957
4958 if (slot + nr != nritems) {
4959 int data_end = leaf_data_end(fs_info, leaf);
4960
4961 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4962 data_end + dsize,
4963 BTRFS_LEAF_DATA_OFFSET + data_end,
4964 last_off - data_end);
4965
4966 for (i = slot + nr; i < nritems; i++) {
4967 u32 ioff;
4968
4969 item = btrfs_item_nr(i);
4970 ioff = btrfs_token_item_offset(leaf, item, &token);
4971 btrfs_set_token_item_offset(leaf, item,
4972 ioff + dsize, &token);
4973 }
4974
4975 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4976 btrfs_item_nr_offset(slot + nr),
4977 sizeof(struct btrfs_item) *
4978 (nritems - slot - nr));
4979 }
4980 btrfs_set_header_nritems(leaf, nritems - nr);
4981 nritems -= nr;
4982
4983 /* delete the leaf if we've emptied it */
4984 if (nritems == 0) {
4985 if (leaf == root->node) {
4986 btrfs_set_header_level(leaf, 0);
4987 } else {
4988 btrfs_set_path_blocking(path);
4989 clean_tree_block(fs_info, leaf);
4990 btrfs_del_leaf(trans, root, path, leaf);
4991 }
4992 } else {
4993 int used = leaf_space_used(leaf, 0, nritems);
4994 if (slot == 0) {
4995 struct btrfs_disk_key disk_key;
4996
4997 btrfs_item_key(leaf, &disk_key, 0);
4998 fixup_low_keys(fs_info, path, &disk_key, 1);
4999 }
5000
5001 /* delete the leaf if it is mostly empty */
5002 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
5003 /* push_leaf_left fixes the path.
5004 * make sure the path still points to our leaf
5005 * for possible call to del_ptr below
5006 */
5007 slot = path->slots[1];
5008 extent_buffer_get(leaf);
5009
5010 btrfs_set_path_blocking(path);
5011 wret = push_leaf_left(trans, root, path, 1, 1,
5012 1, (u32)-1);
5013 if (wret < 0 && wret != -ENOSPC)
5014 ret = wret;
5015
5016 if (path->nodes[0] == leaf &&
5017 btrfs_header_nritems(leaf)) {
5018 wret = push_leaf_right(trans, root, path, 1,
5019 1, 1, 0);
5020 if (wret < 0 && wret != -ENOSPC)
5021 ret = wret;
5022 }
5023
5024 if (btrfs_header_nritems(leaf) == 0) {
5025 path->slots[1] = slot;
5026 btrfs_del_leaf(trans, root, path, leaf);
5027 free_extent_buffer(leaf);
5028 ret = 0;
5029 } else {
5030 /* if we're still in the path, make sure
5031 * we're dirty. Otherwise, one of the
5032 * push_leaf functions must have already
5033 * dirtied this buffer
5034 */
5035 if (path->nodes[0] == leaf)
5036 btrfs_mark_buffer_dirty(leaf);
5037 free_extent_buffer(leaf);
5038 }
5039 } else {
5040 btrfs_mark_buffer_dirty(leaf);
5041 }
5042 }
5043 return ret;
5044}
5045
5046/*
5047 * search the tree again to find a leaf with lesser keys
5048 * returns 0 if it found something or 1 if there are no lesser leaves.
5049 * returns < 0 on io errors.
5050 *
5051 * This may release the path, and so you may lose any locks held at the
5052 * time you call it.
5053 */
5054int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5055{
5056 struct btrfs_key key;
5057 struct btrfs_disk_key found_key;
5058 int ret;
5059
5060 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5061
5062 if (key.offset > 0) {
5063 key.offset--;
5064 } else if (key.type > 0) {
5065 key.type--;
5066 key.offset = (u64)-1;
5067 } else if (key.objectid > 0) {
5068 key.objectid--;
5069 key.type = (u8)-1;
5070 key.offset = (u64)-1;
5071 } else {
5072 return 1;
5073 }
5074
5075 btrfs_release_path(path);
5076 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5077 if (ret < 0)
5078 return ret;
5079 btrfs_item_key(path->nodes[0], &found_key, 0);
5080 ret = comp_keys(&found_key, &key);
5081 /*
5082 * We might have had an item with the previous key in the tree right
5083 * before we released our path. And after we released our path, that
5084 * item might have been pushed to the first slot (0) of the leaf we
5085 * were holding due to a tree balance. Alternatively, an item with the
5086 * previous key can exist as the only element of a leaf (big fat item).
5087 * Therefore account for these 2 cases, so that our callers (like
5088 * btrfs_previous_item) don't miss an existing item with a key matching
5089 * the previous key we computed above.
5090 */
5091 if (ret <= 0)
5092 return 0;
5093 return 1;
5094}
5095
5096/*
5097 * A helper function to walk down the tree starting at min_key, and looking
5098 * for nodes or leaves that are have a minimum transaction id.
5099 * This is used by the btree defrag code, and tree logging
5100 *
5101 * This does not cow, but it does stuff the starting key it finds back
5102 * into min_key, so you can call btrfs_search_slot with cow=1 on the
5103 * key and get a writable path.
5104 *
5105 * This honors path->lowest_level to prevent descent past a given level
5106 * of the tree.
5107 *
5108 * min_trans indicates the oldest transaction that you are interested
5109 * in walking through. Any nodes or leaves older than min_trans are
5110 * skipped over (without reading them).
5111 *
5112 * returns zero if something useful was found, < 0 on error and 1 if there
5113 * was nothing in the tree that matched the search criteria.
5114 */
5115int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5116 struct btrfs_path *path,
5117 u64 min_trans)
5118{
5119 struct btrfs_fs_info *fs_info = root->fs_info;
5120 struct extent_buffer *cur;
5121 struct btrfs_key found_key;
5122 int slot;
5123 int sret;
5124 u32 nritems;
5125 int level;
5126 int ret = 1;
5127 int keep_locks = path->keep_locks;
5128
5129 path->keep_locks = 1;
5130again:
5131 cur = btrfs_read_lock_root_node(root);
5132 level = btrfs_header_level(cur);
5133 WARN_ON(path->nodes[level]);
5134 path->nodes[level] = cur;
5135 path->locks[level] = BTRFS_READ_LOCK;
5136
5137 if (btrfs_header_generation(cur) < min_trans) {
5138 ret = 1;
5139 goto out;
5140 }
5141 while (1) {
5142 nritems = btrfs_header_nritems(cur);
5143 level = btrfs_header_level(cur);
5144 sret = btrfs_bin_search(cur, min_key, level, &slot);
5145
5146 /* at the lowest level, we're done, setup the path and exit */
5147 if (level == path->lowest_level) {
5148 if (slot >= nritems)
5149 goto find_next_key;
5150 ret = 0;
5151 path->slots[level] = slot;
5152 btrfs_item_key_to_cpu(cur, &found_key, slot);
5153 goto out;
5154 }
5155 if (sret && slot > 0)
5156 slot--;
5157 /*
5158 * check this node pointer against the min_trans parameters.
5159 * If it is too old, old, skip to the next one.
5160 */
5161 while (slot < nritems) {
5162 u64 gen;
5163
5164 gen = btrfs_node_ptr_generation(cur, slot);
5165 if (gen < min_trans) {
5166 slot++;
5167 continue;
5168 }
5169 break;
5170 }
5171find_next_key:
5172 /*
5173 * we didn't find a candidate key in this node, walk forward
5174 * and find another one
5175 */
5176 if (slot >= nritems) {
5177 path->slots[level] = slot;
5178 btrfs_set_path_blocking(path);
5179 sret = btrfs_find_next_key(root, path, min_key, level,
5180 min_trans);
5181 if (sret == 0) {
5182 btrfs_release_path(path);
5183 goto again;
5184 } else {
5185 goto out;
5186 }
5187 }
5188 /* save our key for returning back */
5189 btrfs_node_key_to_cpu(cur, &found_key, slot);
5190 path->slots[level] = slot;
5191 if (level == path->lowest_level) {
5192 ret = 0;
5193 goto out;
5194 }
5195 btrfs_set_path_blocking(path);
5196 cur = read_node_slot(fs_info, cur, slot);
5197 if (IS_ERR(cur)) {
5198 ret = PTR_ERR(cur);
5199 goto out;
5200 }
5201
5202 btrfs_tree_read_lock(cur);
5203
5204 path->locks[level - 1] = BTRFS_READ_LOCK;
5205 path->nodes[level - 1] = cur;
5206 unlock_up(path, level, 1, 0, NULL);
5207 btrfs_clear_path_blocking(path, NULL, 0);
5208 }
5209out:
5210 path->keep_locks = keep_locks;
5211 if (ret == 0) {
5212 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5213 btrfs_set_path_blocking(path);
5214 memcpy(min_key, &found_key, sizeof(found_key));
5215 }
5216 return ret;
5217}
5218
5219static int tree_move_down(struct btrfs_fs_info *fs_info,
5220 struct btrfs_path *path,
5221 int *level)
5222{
5223 struct extent_buffer *eb;
5224
5225 BUG_ON(*level == 0);
5226 eb = read_node_slot(fs_info, path->nodes[*level], path->slots[*level]);
5227 if (IS_ERR(eb))
5228 return PTR_ERR(eb);
5229
5230 path->nodes[*level - 1] = eb;
5231 path->slots[*level - 1] = 0;
5232 (*level)--;
5233 return 0;
5234}
5235
5236static int tree_move_next_or_upnext(struct btrfs_path *path,
5237 int *level, int root_level)
5238{
5239 int ret = 0;
5240 int nritems;
5241 nritems = btrfs_header_nritems(path->nodes[*level]);
5242
5243 path->slots[*level]++;
5244
5245 while (path->slots[*level] >= nritems) {
5246 if (*level == root_level)
5247 return -1;
5248
5249 /* move upnext */
5250 path->slots[*level] = 0;
5251 free_extent_buffer(path->nodes[*level]);
5252 path->nodes[*level] = NULL;
5253 (*level)++;
5254 path->slots[*level]++;
5255
5256 nritems = btrfs_header_nritems(path->nodes[*level]);
5257 ret = 1;
5258 }
5259 return ret;
5260}
5261
5262/*
5263 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
5264 * or down.
5265 */
5266static int tree_advance(struct btrfs_fs_info *fs_info,
5267 struct btrfs_path *path,
5268 int *level, int root_level,
5269 int allow_down,
5270 struct btrfs_key *key)
5271{
5272 int ret;
5273
5274 if (*level == 0 || !allow_down) {
5275 ret = tree_move_next_or_upnext(path, level, root_level);
5276 } else {
5277 ret = tree_move_down(fs_info, path, level);
5278 }
5279 if (ret >= 0) {
5280 if (*level == 0)
5281 btrfs_item_key_to_cpu(path->nodes[*level], key,
5282 path->slots[*level]);
5283 else
5284 btrfs_node_key_to_cpu(path->nodes[*level], key,
5285 path->slots[*level]);
5286 }
5287 return ret;
5288}
5289
5290static int tree_compare_item(struct btrfs_path *left_path,
5291 struct btrfs_path *right_path,
5292 char *tmp_buf)
5293{
5294 int cmp;
5295 int len1, len2;
5296 unsigned long off1, off2;
5297
5298 len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
5299 len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
5300 if (len1 != len2)
5301 return 1;
5302
5303 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
5304 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
5305 right_path->slots[0]);
5306
5307 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
5308
5309 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
5310 if (cmp)
5311 return 1;
5312 return 0;
5313}
5314
5315#define ADVANCE 1
5316#define ADVANCE_ONLY_NEXT -1
5317
5318/*
5319 * This function compares two trees and calls the provided callback for
5320 * every changed/new/deleted item it finds.
5321 * If shared tree blocks are encountered, whole subtrees are skipped, making
5322 * the compare pretty fast on snapshotted subvolumes.
5323 *
5324 * This currently works on commit roots only. As commit roots are read only,
5325 * we don't do any locking. The commit roots are protected with transactions.
5326 * Transactions are ended and rejoined when a commit is tried in between.
5327 *
5328 * This function checks for modifications done to the trees while comparing.
5329 * If it detects a change, it aborts immediately.
5330 */
5331int btrfs_compare_trees(struct btrfs_root *left_root,
5332 struct btrfs_root *right_root,
5333 btrfs_changed_cb_t changed_cb, void *ctx)
5334{
5335 struct btrfs_fs_info *fs_info = left_root->fs_info;
5336 int ret;
5337 int cmp;
5338 struct btrfs_path *left_path = NULL;
5339 struct btrfs_path *right_path = NULL;
5340 struct btrfs_key left_key;
5341 struct btrfs_key right_key;
5342 char *tmp_buf = NULL;
5343 int left_root_level;
5344 int right_root_level;
5345 int left_level;
5346 int right_level;
5347 int left_end_reached;
5348 int right_end_reached;
5349 int advance_left;
5350 int advance_right;
5351 u64 left_blockptr;
5352 u64 right_blockptr;
5353 u64 left_gen;
5354 u64 right_gen;
5355
5356 left_path = btrfs_alloc_path();
5357 if (!left_path) {
5358 ret = -ENOMEM;
5359 goto out;
5360 }
5361 right_path = btrfs_alloc_path();
5362 if (!right_path) {
5363 ret = -ENOMEM;
5364 goto out;
5365 }
5366
5367 tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
5368 if (!tmp_buf) {
5369 ret = -ENOMEM;
5370 goto out;
5371 }
5372
5373 left_path->search_commit_root = 1;
5374 left_path->skip_locking = 1;
5375 right_path->search_commit_root = 1;
5376 right_path->skip_locking = 1;
5377
5378 /*
5379 * Strategy: Go to the first items of both trees. Then do
5380 *
5381 * If both trees are at level 0
5382 * Compare keys of current items
5383 * If left < right treat left item as new, advance left tree
5384 * and repeat
5385 * If left > right treat right item as deleted, advance right tree
5386 * and repeat
5387 * If left == right do deep compare of items, treat as changed if
5388 * needed, advance both trees and repeat
5389 * If both trees are at the same level but not at level 0
5390 * Compare keys of current nodes/leafs
5391 * If left < right advance left tree and repeat
5392 * If left > right advance right tree and repeat
5393 * If left == right compare blockptrs of the next nodes/leafs
5394 * If they match advance both trees but stay at the same level
5395 * and repeat
5396 * If they don't match advance both trees while allowing to go
5397 * deeper and repeat
5398 * If tree levels are different
5399 * Advance the tree that needs it and repeat
5400 *
5401 * Advancing a tree means:
5402 * If we are at level 0, try to go to the next slot. If that's not
5403 * possible, go one level up and repeat. Stop when we found a level
5404 * where we could go to the next slot. We may at this point be on a
5405 * node or a leaf.
5406 *
5407 * If we are not at level 0 and not on shared tree blocks, go one
5408 * level deeper.
5409 *
5410 * If we are not at level 0 and on shared tree blocks, go one slot to
5411 * the right if possible or go up and right.
5412 */
5413
5414 down_read(&fs_info->commit_root_sem);
5415 left_level = btrfs_header_level(left_root->commit_root);
5416 left_root_level = left_level;
5417 left_path->nodes[left_level] =
5418 btrfs_clone_extent_buffer(left_root->commit_root);
5419 if (!left_path->nodes[left_level]) {
5420 up_read(&fs_info->commit_root_sem);
5421 ret = -ENOMEM;
5422 goto out;
5423 }
5424 extent_buffer_get(left_path->nodes[left_level]);
5425
5426 right_level = btrfs_header_level(right_root->commit_root);
5427 right_root_level = right_level;
5428 right_path->nodes[right_level] =
5429 btrfs_clone_extent_buffer(right_root->commit_root);
5430 if (!right_path->nodes[right_level]) {
5431 up_read(&fs_info->commit_root_sem);
5432 ret = -ENOMEM;
5433 goto out;
5434 }
5435 extent_buffer_get(right_path->nodes[right_level]);
5436 up_read(&fs_info->commit_root_sem);
5437
5438 if (left_level == 0)
5439 btrfs_item_key_to_cpu(left_path->nodes[left_level],
5440 &left_key, left_path->slots[left_level]);
5441 else
5442 btrfs_node_key_to_cpu(left_path->nodes[left_level],
5443 &left_key, left_path->slots[left_level]);
5444 if (right_level == 0)
5445 btrfs_item_key_to_cpu(right_path->nodes[right_level],
5446 &right_key, right_path->slots[right_level]);
5447 else
5448 btrfs_node_key_to_cpu(right_path->nodes[right_level],
5449 &right_key, right_path->slots[right_level]);
5450
5451 left_end_reached = right_end_reached = 0;
5452 advance_left = advance_right = 0;
5453
5454 while (1) {
5455 if (advance_left && !left_end_reached) {
5456 ret = tree_advance(fs_info, left_path, &left_level,
5457 left_root_level,
5458 advance_left != ADVANCE_ONLY_NEXT,
5459 &left_key);
5460 if (ret == -1)
5461 left_end_reached = ADVANCE;
5462 else if (ret < 0)
5463 goto out;
5464 advance_left = 0;
5465 }
5466 if (advance_right && !right_end_reached) {
5467 ret = tree_advance(fs_info, right_path, &right_level,
5468 right_root_level,
5469 advance_right != ADVANCE_ONLY_NEXT,
5470 &right_key);
5471 if (ret == -1)
5472 right_end_reached = ADVANCE;
5473 else if (ret < 0)
5474 goto out;
5475 advance_right = 0;
5476 }
5477
5478 if (left_end_reached && right_end_reached) {
5479 ret = 0;
5480 goto out;
5481 } else if (left_end_reached) {
5482 if (right_level == 0) {
5483 ret = changed_cb(left_path, right_path,
5484 &right_key,
5485 BTRFS_COMPARE_TREE_DELETED,
5486 ctx);
5487 if (ret < 0)
5488 goto out;
5489 }
5490 advance_right = ADVANCE;
5491 continue;
5492 } else if (right_end_reached) {
5493 if (left_level == 0) {
5494 ret = changed_cb(left_path, right_path,
5495 &left_key,
5496 BTRFS_COMPARE_TREE_NEW,
5497 ctx);
5498 if (ret < 0)
5499 goto out;
5500 }
5501 advance_left = ADVANCE;
5502 continue;
5503 }
5504
5505 if (left_level == 0 && right_level == 0) {
5506 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5507 if (cmp < 0) {
5508 ret = changed_cb(left_path, right_path,
5509 &left_key,
5510 BTRFS_COMPARE_TREE_NEW,
5511 ctx);
5512 if (ret < 0)
5513 goto out;
5514 advance_left = ADVANCE;
5515 } else if (cmp > 0) {
5516 ret = changed_cb(left_path, right_path,
5517 &right_key,
5518 BTRFS_COMPARE_TREE_DELETED,
5519 ctx);
5520 if (ret < 0)
5521 goto out;
5522 advance_right = ADVANCE;
5523 } else {
5524 enum btrfs_compare_tree_result result;
5525
5526 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
5527 ret = tree_compare_item(left_path, right_path,
5528 tmp_buf);
5529 if (ret)
5530 result = BTRFS_COMPARE_TREE_CHANGED;
5531 else
5532 result = BTRFS_COMPARE_TREE_SAME;
5533 ret = changed_cb(left_path, right_path,
5534 &left_key, result, ctx);
5535 if (ret < 0)
5536 goto out;
5537 advance_left = ADVANCE;
5538 advance_right = ADVANCE;
5539 }
5540 } else if (left_level == right_level) {
5541 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5542 if (cmp < 0) {
5543 advance_left = ADVANCE;
5544 } else if (cmp > 0) {
5545 advance_right = ADVANCE;
5546 } else {
5547 left_blockptr = btrfs_node_blockptr(
5548 left_path->nodes[left_level],
5549 left_path->slots[left_level]);
5550 right_blockptr = btrfs_node_blockptr(
5551 right_path->nodes[right_level],
5552 right_path->slots[right_level]);
5553 left_gen = btrfs_node_ptr_generation(
5554 left_path->nodes[left_level],
5555 left_path->slots[left_level]);
5556 right_gen = btrfs_node_ptr_generation(
5557 right_path->nodes[right_level],
5558 right_path->slots[right_level]);
5559 if (left_blockptr == right_blockptr &&
5560 left_gen == right_gen) {
5561 /*
5562 * As we're on a shared block, don't
5563 * allow to go deeper.
5564 */
5565 advance_left = ADVANCE_ONLY_NEXT;
5566 advance_right = ADVANCE_ONLY_NEXT;
5567 } else {
5568 advance_left = ADVANCE;
5569 advance_right = ADVANCE;
5570 }
5571 }
5572 } else if (left_level < right_level) {
5573 advance_right = ADVANCE;
5574 } else {
5575 advance_left = ADVANCE;
5576 }
5577 }
5578
5579out:
5580 btrfs_free_path(left_path);
5581 btrfs_free_path(right_path);
5582 kvfree(tmp_buf);
5583 return ret;
5584}
5585
5586/*
5587 * this is similar to btrfs_next_leaf, but does not try to preserve
5588 * and fixup the path. It looks for and returns the next key in the
5589 * tree based on the current path and the min_trans parameters.
5590 *
5591 * 0 is returned if another key is found, < 0 if there are any errors
5592 * and 1 is returned if there are no higher keys in the tree
5593 *
5594 * path->keep_locks should be set to 1 on the search made before
5595 * calling this function.
5596 */
5597int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5598 struct btrfs_key *key, int level, u64 min_trans)
5599{
5600 int slot;
5601 struct extent_buffer *c;
5602
5603 WARN_ON(!path->keep_locks);
5604 while (level < BTRFS_MAX_LEVEL) {
5605 if (!path->nodes[level])
5606 return 1;
5607
5608 slot = path->slots[level] + 1;
5609 c = path->nodes[level];
5610next:
5611 if (slot >= btrfs_header_nritems(c)) {
5612 int ret;
5613 int orig_lowest;
5614 struct btrfs_key cur_key;
5615 if (level + 1 >= BTRFS_MAX_LEVEL ||
5616 !path->nodes[level + 1])
5617 return 1;
5618
5619 if (path->locks[level + 1]) {
5620 level++;
5621 continue;
5622 }
5623
5624 slot = btrfs_header_nritems(c) - 1;
5625 if (level == 0)
5626 btrfs_item_key_to_cpu(c, &cur_key, slot);
5627 else
5628 btrfs_node_key_to_cpu(c, &cur_key, slot);
5629
5630 orig_lowest = path->lowest_level;
5631 btrfs_release_path(path);
5632 path->lowest_level = level;
5633 ret = btrfs_search_slot(NULL, root, &cur_key, path,
5634 0, 0);
5635 path->lowest_level = orig_lowest;
5636 if (ret < 0)
5637 return ret;
5638
5639 c = path->nodes[level];
5640 slot = path->slots[level];
5641 if (ret == 0)
5642 slot++;
5643 goto next;
5644 }
5645
5646 if (level == 0)
5647 btrfs_item_key_to_cpu(c, key, slot);
5648 else {
5649 u64 gen = btrfs_node_ptr_generation(c, slot);
5650
5651 if (gen < min_trans) {
5652 slot++;
5653 goto next;
5654 }
5655 btrfs_node_key_to_cpu(c, key, slot);
5656 }
5657 return 0;
5658 }
5659 return 1;
5660}
5661
5662/*
5663 * search the tree again to find a leaf with greater keys
5664 * returns 0 if it found something or 1 if there are no greater leaves.
5665 * returns < 0 on io errors.
5666 */
5667int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5668{
5669 return btrfs_next_old_leaf(root, path, 0);
5670}
5671
5672int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5673 u64 time_seq)
5674{
5675 int slot;
5676 int level;
5677 struct extent_buffer *c;
5678 struct extent_buffer *next;
5679 struct btrfs_key key;
5680 u32 nritems;
5681 int ret;
5682 int old_spinning = path->leave_spinning;
5683 int next_rw_lock = 0;
5684
5685 nritems = btrfs_header_nritems(path->nodes[0]);
5686 if (nritems == 0)
5687 return 1;
5688
5689 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5690again:
5691 level = 1;
5692 next = NULL;
5693 next_rw_lock = 0;
5694 btrfs_release_path(path);
5695
5696 path->keep_locks = 1;
5697 path->leave_spinning = 1;
5698
5699 if (time_seq)
5700 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5701 else
5702 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5703 path->keep_locks = 0;
5704
5705 if (ret < 0)
5706 return ret;
5707
5708 nritems = btrfs_header_nritems(path->nodes[0]);
5709 /*
5710 * by releasing the path above we dropped all our locks. A balance
5711 * could have added more items next to the key that used to be
5712 * at the very end of the block. So, check again here and
5713 * advance the path if there are now more items available.
5714 */
5715 if (nritems > 0 && path->slots[0] < nritems - 1) {
5716 if (ret == 0)
5717 path->slots[0]++;
5718 ret = 0;
5719 goto done;
5720 }
5721 /*
5722 * So the above check misses one case:
5723 * - after releasing the path above, someone has removed the item that
5724 * used to be at the very end of the block, and balance between leafs
5725 * gets another one with bigger key.offset to replace it.
5726 *
5727 * This one should be returned as well, or we can get leaf corruption
5728 * later(esp. in __btrfs_drop_extents()).
5729 *
5730 * And a bit more explanation about this check,
5731 * with ret > 0, the key isn't found, the path points to the slot
5732 * where it should be inserted, so the path->slots[0] item must be the
5733 * bigger one.
5734 */
5735 if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5736 ret = 0;
5737 goto done;
5738 }
5739
5740 while (level < BTRFS_MAX_LEVEL) {
5741 if (!path->nodes[level]) {
5742 ret = 1;
5743 goto done;
5744 }
5745
5746 slot = path->slots[level] + 1;
5747 c = path->nodes[level];
5748 if (slot >= btrfs_header_nritems(c)) {
5749 level++;
5750 if (level == BTRFS_MAX_LEVEL) {
5751 ret = 1;
5752 goto done;
5753 }
5754 continue;
5755 }
5756
5757 if (next) {
5758 btrfs_tree_unlock_rw(next, next_rw_lock);
5759 free_extent_buffer(next);
5760 }
5761
5762 next = c;
5763 next_rw_lock = path->locks[level];
5764 ret = read_block_for_search(root, path, &next, level,
5765 slot, &key);
5766 if (ret == -EAGAIN)
5767 goto again;
5768
5769 if (ret < 0) {
5770 btrfs_release_path(path);
5771 goto done;
5772 }
5773
5774 if (!path->skip_locking) {
5775 ret = btrfs_try_tree_read_lock(next);
5776 if (!ret && time_seq) {
5777 /*
5778 * If we don't get the lock, we may be racing
5779 * with push_leaf_left, holding that lock while
5780 * itself waiting for the leaf we've currently
5781 * locked. To solve this situation, we give up
5782 * on our lock and cycle.
5783 */
5784 free_extent_buffer(next);
5785 btrfs_release_path(path);
5786 cond_resched();
5787 goto again;
5788 }
5789 if (!ret) {
5790 btrfs_set_path_blocking(path);
5791 btrfs_tree_read_lock(next);
5792 btrfs_clear_path_blocking(path, next,
5793 BTRFS_READ_LOCK);
5794 }
5795 next_rw_lock = BTRFS_READ_LOCK;
5796 }
5797 break;
5798 }
5799 path->slots[level] = slot;
5800 while (1) {
5801 level--;
5802 c = path->nodes[level];
5803 if (path->locks[level])
5804 btrfs_tree_unlock_rw(c, path->locks[level]);
5805
5806 free_extent_buffer(c);
5807 path->nodes[level] = next;
5808 path->slots[level] = 0;
5809 if (!path->skip_locking)
5810 path->locks[level] = next_rw_lock;
5811 if (!level)
5812 break;
5813
5814 ret = read_block_for_search(root, path, &next, level,
5815 0, &key);
5816 if (ret == -EAGAIN)
5817 goto again;
5818
5819 if (ret < 0) {
5820 btrfs_release_path(path);
5821 goto done;
5822 }
5823
5824 if (!path->skip_locking) {
5825 ret = btrfs_try_tree_read_lock(next);
5826 if (!ret) {
5827 btrfs_set_path_blocking(path);
5828 btrfs_tree_read_lock(next);
5829 btrfs_clear_path_blocking(path, next,
5830 BTRFS_READ_LOCK);
5831 }
5832 next_rw_lock = BTRFS_READ_LOCK;
5833 }
5834 }
5835 ret = 0;
5836done:
5837 unlock_up(path, 0, 1, 0, NULL);
5838 path->leave_spinning = old_spinning;
5839 if (!old_spinning)
5840 btrfs_set_path_blocking(path);
5841
5842 return ret;
5843}
5844
5845/*
5846 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5847 * searching until it gets past min_objectid or finds an item of 'type'
5848 *
5849 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5850 */
5851int btrfs_previous_item(struct btrfs_root *root,
5852 struct btrfs_path *path, u64 min_objectid,
5853 int type)
5854{
5855 struct btrfs_key found_key;
5856 struct extent_buffer *leaf;
5857 u32 nritems;
5858 int ret;
5859
5860 while (1) {
5861 if (path->slots[0] == 0) {
5862 btrfs_set_path_blocking(path);
5863 ret = btrfs_prev_leaf(root, path);
5864 if (ret != 0)
5865 return ret;
5866 } else {
5867 path->slots[0]--;
5868 }
5869 leaf = path->nodes[0];
5870 nritems = btrfs_header_nritems(leaf);
5871 if (nritems == 0)
5872 return 1;
5873 if (path->slots[0] == nritems)
5874 path->slots[0]--;
5875
5876 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5877 if (found_key.objectid < min_objectid)
5878 break;
5879 if (found_key.type == type)
5880 return 0;
5881 if (found_key.objectid == min_objectid &&
5882 found_key.type < type)
5883 break;
5884 }
5885 return 1;
5886}
5887
5888/*
5889 * search in extent tree to find a previous Metadata/Data extent item with
5890 * min objecitd.
5891 *
5892 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5893 */
5894int btrfs_previous_extent_item(struct btrfs_root *root,
5895 struct btrfs_path *path, u64 min_objectid)
5896{
5897 struct btrfs_key found_key;
5898 struct extent_buffer *leaf;
5899 u32 nritems;
5900 int ret;
5901
5902 while (1) {
5903 if (path->slots[0] == 0) {
5904 btrfs_set_path_blocking(path);
5905 ret = btrfs_prev_leaf(root, path);
5906 if (ret != 0)
5907 return ret;
5908 } else {
5909 path->slots[0]--;
5910 }
5911 leaf = path->nodes[0];
5912 nritems = btrfs_header_nritems(leaf);
5913 if (nritems == 0)
5914 return 1;
5915 if (path->slots[0] == nritems)
5916 path->slots[0]--;
5917
5918 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5919 if (found_key.objectid < min_objectid)
5920 break;
5921 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5922 found_key.type == BTRFS_METADATA_ITEM_KEY)
5923 return 0;
5924 if (found_key.objectid == min_objectid &&
5925 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5926 break;
5927 }
5928 return 1;
5929}