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