Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007,2008 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/rbtree.h>
9#include <linux/mm.h>
10#include <linux/error-injection.h>
11#include "messages.h"
12#include "ctree.h"
13#include "disk-io.h"
14#include "transaction.h"
15#include "print-tree.h"
16#include "locking.h"
17#include "volumes.h"
18#include "qgroup.h"
19#include "tree-mod-log.h"
20#include "tree-checker.h"
21#include "fs.h"
22#include "accessors.h"
23#include "extent-tree.h"
24#include "relocation.h"
25#include "file-item.h"
26
27static struct kmem_cache *btrfs_path_cachep;
28
29static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
30 *root, struct btrfs_path *path, int level);
31static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
32 const struct btrfs_key *ins_key, struct btrfs_path *path,
33 int data_size, int extend);
34static int push_node_left(struct btrfs_trans_handle *trans,
35 struct extent_buffer *dst,
36 struct extent_buffer *src, int empty);
37static int balance_node_right(struct btrfs_trans_handle *trans,
38 struct extent_buffer *dst_buf,
39 struct extent_buffer *src_buf);
40
41static const struct btrfs_csums {
42 u16 size;
43 const char name[10];
44 const char driver[12];
45} btrfs_csums[] = {
46 [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
47 [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
48 [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
49 [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
50 .driver = "blake2b-256" },
51};
52
53/*
54 * The leaf data grows from end-to-front in the node. this returns the address
55 * of the start of the last item, which is the stop of the leaf data stack.
56 */
57static unsigned int leaf_data_end(const struct extent_buffer *leaf)
58{
59 u32 nr = btrfs_header_nritems(leaf);
60
61 if (nr == 0)
62 return BTRFS_LEAF_DATA_SIZE(leaf->fs_info);
63 return btrfs_item_offset(leaf, nr - 1);
64}
65
66/*
67 * Move data in a @leaf (using memmove, safe for overlapping ranges).
68 *
69 * @leaf: leaf that we're doing a memmove on
70 * @dst_offset: item data offset we're moving to
71 * @src_offset: item data offset were' moving from
72 * @len: length of the data we're moving
73 *
74 * Wrapper around memmove_extent_buffer() that takes into account the header on
75 * the leaf. The btrfs_item offset's start directly after the header, so we
76 * have to adjust any offsets to account for the header in the leaf. This
77 * handles that math to simplify the callers.
78 */
79static inline void memmove_leaf_data(const struct extent_buffer *leaf,
80 unsigned long dst_offset,
81 unsigned long src_offset,
82 unsigned long len)
83{
84 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, 0) + dst_offset,
85 btrfs_item_nr_offset(leaf, 0) + src_offset, len);
86}
87
88/*
89 * Copy item data from @src into @dst at the given @offset.
90 *
91 * @dst: destination leaf that we're copying into
92 * @src: source leaf that we're copying from
93 * @dst_offset: item data offset we're copying to
94 * @src_offset: item data offset were' copying from
95 * @len: length of the data we're copying
96 *
97 * Wrapper around copy_extent_buffer() that takes into account the header on
98 * the leaf. The btrfs_item offset's start directly after the header, so we
99 * have to adjust any offsets to account for the header in the leaf. This
100 * handles that math to simplify the callers.
101 */
102static inline void copy_leaf_data(const struct extent_buffer *dst,
103 const struct extent_buffer *src,
104 unsigned long dst_offset,
105 unsigned long src_offset, unsigned long len)
106{
107 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, 0) + dst_offset,
108 btrfs_item_nr_offset(src, 0) + src_offset, len);
109}
110
111/*
112 * Move items in a @leaf (using memmove).
113 *
114 * @dst: destination leaf for the items
115 * @dst_item: the item nr we're copying into
116 * @src_item: the item nr we're copying from
117 * @nr_items: the number of items to copy
118 *
119 * Wrapper around memmove_extent_buffer() that does the math to get the
120 * appropriate offsets into the leaf from the item numbers.
121 */
122static inline void memmove_leaf_items(const struct extent_buffer *leaf,
123 int dst_item, int src_item, int nr_items)
124{
125 memmove_extent_buffer(leaf, btrfs_item_nr_offset(leaf, dst_item),
126 btrfs_item_nr_offset(leaf, src_item),
127 nr_items * sizeof(struct btrfs_item));
128}
129
130/*
131 * Copy items from @src into @dst at the given @offset.
132 *
133 * @dst: destination leaf for the items
134 * @src: source leaf for the items
135 * @dst_item: the item nr we're copying into
136 * @src_item: the item nr we're copying from
137 * @nr_items: the number of items to copy
138 *
139 * Wrapper around copy_extent_buffer() that does the math to get the
140 * appropriate offsets into the leaf from the item numbers.
141 */
142static inline void copy_leaf_items(const struct extent_buffer *dst,
143 const struct extent_buffer *src,
144 int dst_item, int src_item, int nr_items)
145{
146 copy_extent_buffer(dst, src, btrfs_item_nr_offset(dst, dst_item),
147 btrfs_item_nr_offset(src, src_item),
148 nr_items * sizeof(struct btrfs_item));
149}
150
151/* This exists for btrfs-progs usages. */
152u16 btrfs_csum_type_size(u16 type)
153{
154 return btrfs_csums[type].size;
155}
156
157int btrfs_super_csum_size(const struct btrfs_super_block *s)
158{
159 u16 t = btrfs_super_csum_type(s);
160 /*
161 * csum type is validated at mount time
162 */
163 return btrfs_csum_type_size(t);
164}
165
166const char *btrfs_super_csum_name(u16 csum_type)
167{
168 /* csum type is validated at mount time */
169 return btrfs_csums[csum_type].name;
170}
171
172/*
173 * Return driver name if defined, otherwise the name that's also a valid driver
174 * name
175 */
176const char *btrfs_super_csum_driver(u16 csum_type)
177{
178 /* csum type is validated at mount time */
179 return btrfs_csums[csum_type].driver[0] ?
180 btrfs_csums[csum_type].driver :
181 btrfs_csums[csum_type].name;
182}
183
184size_t __attribute_const__ btrfs_get_num_csums(void)
185{
186 return ARRAY_SIZE(btrfs_csums);
187}
188
189struct btrfs_path *btrfs_alloc_path(void)
190{
191 might_sleep();
192
193 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
194}
195
196/* this also releases the path */
197void btrfs_free_path(struct btrfs_path *p)
198{
199 if (!p)
200 return;
201 btrfs_release_path(p);
202 kmem_cache_free(btrfs_path_cachep, p);
203}
204
205/*
206 * path release drops references on the extent buffers in the path
207 * and it drops any locks held by this path
208 *
209 * It is safe to call this on paths that no locks or extent buffers held.
210 */
211noinline void btrfs_release_path(struct btrfs_path *p)
212{
213 int i;
214
215 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
216 p->slots[i] = 0;
217 if (!p->nodes[i])
218 continue;
219 if (p->locks[i]) {
220 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
221 p->locks[i] = 0;
222 }
223 free_extent_buffer(p->nodes[i]);
224 p->nodes[i] = NULL;
225 }
226}
227
228/*
229 * We want the transaction abort to print stack trace only for errors where the
230 * cause could be a bug, eg. due to ENOSPC, and not for common errors that are
231 * caused by external factors.
232 */
233bool __cold abort_should_print_stack(int error)
234{
235 switch (error) {
236 case -EIO:
237 case -EROFS:
238 case -ENOMEM:
239 return false;
240 }
241 return true;
242}
243
244/*
245 * safely gets a reference on the root node of a tree. A lock
246 * is not taken, so a concurrent writer may put a different node
247 * at the root of the tree. See btrfs_lock_root_node for the
248 * looping required.
249 *
250 * The extent buffer returned by this has a reference taken, so
251 * it won't disappear. It may stop being the root of the tree
252 * at any time because there are no locks held.
253 */
254struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
255{
256 struct extent_buffer *eb;
257
258 while (1) {
259 rcu_read_lock();
260 eb = rcu_dereference(root->node);
261
262 /*
263 * RCU really hurts here, we could free up the root node because
264 * it was COWed but we may not get the new root node yet so do
265 * the inc_not_zero dance and if it doesn't work then
266 * synchronize_rcu and try again.
267 */
268 if (atomic_inc_not_zero(&eb->refs)) {
269 rcu_read_unlock();
270 break;
271 }
272 rcu_read_unlock();
273 synchronize_rcu();
274 }
275 return eb;
276}
277
278/*
279 * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
280 * just get put onto a simple dirty list. Transaction walks this list to make
281 * sure they get properly updated on disk.
282 */
283static void add_root_to_dirty_list(struct btrfs_root *root)
284{
285 struct btrfs_fs_info *fs_info = root->fs_info;
286
287 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
288 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
289 return;
290
291 spin_lock(&fs_info->trans_lock);
292 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
293 /* Want the extent tree to be the last on the list */
294 if (btrfs_root_id(root) == BTRFS_EXTENT_TREE_OBJECTID)
295 list_move_tail(&root->dirty_list,
296 &fs_info->dirty_cowonly_roots);
297 else
298 list_move(&root->dirty_list,
299 &fs_info->dirty_cowonly_roots);
300 }
301 spin_unlock(&fs_info->trans_lock);
302}
303
304/*
305 * used by snapshot creation to make a copy of a root for a tree with
306 * a given objectid. The buffer with the new root node is returned in
307 * cow_ret, and this func returns zero on success or a negative error code.
308 */
309int btrfs_copy_root(struct btrfs_trans_handle *trans,
310 struct btrfs_root *root,
311 struct extent_buffer *buf,
312 struct extent_buffer **cow_ret, u64 new_root_objectid)
313{
314 struct btrfs_fs_info *fs_info = root->fs_info;
315 struct extent_buffer *cow;
316 int ret = 0;
317 int level;
318 struct btrfs_disk_key disk_key;
319 u64 reloc_src_root = 0;
320
321 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
322 trans->transid != fs_info->running_transaction->transid);
323 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
324 trans->transid != btrfs_get_root_last_trans(root));
325
326 level = btrfs_header_level(buf);
327 if (level == 0)
328 btrfs_item_key(buf, &disk_key, 0);
329 else
330 btrfs_node_key(buf, &disk_key, 0);
331
332 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
333 reloc_src_root = btrfs_header_owner(buf);
334 cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
335 &disk_key, level, buf->start, 0,
336 reloc_src_root, BTRFS_NESTING_NEW_ROOT);
337 if (IS_ERR(cow))
338 return PTR_ERR(cow);
339
340 copy_extent_buffer_full(cow, buf);
341 btrfs_set_header_bytenr(cow, cow->start);
342 btrfs_set_header_generation(cow, trans->transid);
343 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
344 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
345 BTRFS_HEADER_FLAG_RELOC);
346 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
347 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
348 else
349 btrfs_set_header_owner(cow, new_root_objectid);
350
351 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
352
353 WARN_ON(btrfs_header_generation(buf) > trans->transid);
354 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
355 ret = btrfs_inc_ref(trans, root, cow, 1);
356 else
357 ret = btrfs_inc_ref(trans, root, cow, 0);
358 if (ret) {
359 btrfs_tree_unlock(cow);
360 free_extent_buffer(cow);
361 btrfs_abort_transaction(trans, ret);
362 return ret;
363 }
364
365 btrfs_mark_buffer_dirty(trans, cow);
366 *cow_ret = cow;
367 return 0;
368}
369
370/*
371 * check if the tree block can be shared by multiple trees
372 */
373bool btrfs_block_can_be_shared(struct btrfs_trans_handle *trans,
374 struct btrfs_root *root,
375 struct extent_buffer *buf)
376{
377 const u64 buf_gen = btrfs_header_generation(buf);
378
379 /*
380 * Tree blocks not in shareable trees and tree roots are never shared.
381 * If a block was allocated after the last snapshot and the block was
382 * not allocated by tree relocation, we know the block is not shared.
383 */
384
385 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
386 return false;
387
388 if (buf == root->node)
389 return false;
390
391 if (buf_gen > btrfs_root_last_snapshot(&root->root_item) &&
392 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC))
393 return false;
394
395 if (buf != root->commit_root)
396 return true;
397
398 /*
399 * An extent buffer that used to be the commit root may still be shared
400 * because the tree height may have increased and it became a child of a
401 * higher level root. This can happen when snapshotting a subvolume
402 * created in the current transaction.
403 */
404 if (buf_gen == trans->transid)
405 return true;
406
407 return false;
408}
409
410static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
411 struct btrfs_root *root,
412 struct extent_buffer *buf,
413 struct extent_buffer *cow,
414 int *last_ref)
415{
416 struct btrfs_fs_info *fs_info = root->fs_info;
417 u64 refs;
418 u64 owner;
419 u64 flags;
420 int ret;
421
422 /*
423 * Backrefs update rules:
424 *
425 * Always use full backrefs for extent pointers in tree block
426 * allocated by tree relocation.
427 *
428 * If a shared tree block is no longer referenced by its owner
429 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
430 * use full backrefs for extent pointers in tree block.
431 *
432 * If a tree block is been relocating
433 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
434 * use full backrefs for extent pointers in tree block.
435 * The reason for this is some operations (such as drop tree)
436 * are only allowed for blocks use full backrefs.
437 */
438
439 if (btrfs_block_can_be_shared(trans, root, buf)) {
440 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
441 btrfs_header_level(buf), 1,
442 &refs, &flags, NULL);
443 if (ret)
444 return ret;
445 if (unlikely(refs == 0)) {
446 btrfs_crit(fs_info,
447 "found 0 references for tree block at bytenr %llu level %d root %llu",
448 buf->start, btrfs_header_level(buf),
449 btrfs_root_id(root));
450 ret = -EUCLEAN;
451 btrfs_abort_transaction(trans, ret);
452 return ret;
453 }
454 } else {
455 refs = 1;
456 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID ||
457 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
458 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
459 else
460 flags = 0;
461 }
462
463 owner = btrfs_header_owner(buf);
464 if (unlikely(owner == BTRFS_TREE_RELOC_OBJECTID &&
465 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))) {
466 btrfs_crit(fs_info,
467"found tree block at bytenr %llu level %d root %llu refs %llu flags %llx without full backref flag set",
468 buf->start, btrfs_header_level(buf),
469 btrfs_root_id(root), refs, flags);
470 ret = -EUCLEAN;
471 btrfs_abort_transaction(trans, ret);
472 return ret;
473 }
474
475 if (refs > 1) {
476 if ((owner == btrfs_root_id(root) ||
477 btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID) &&
478 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
479 ret = btrfs_inc_ref(trans, root, buf, 1);
480 if (ret)
481 return ret;
482
483 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID) {
484 ret = btrfs_dec_ref(trans, root, buf, 0);
485 if (ret)
486 return ret;
487 ret = btrfs_inc_ref(trans, root, cow, 1);
488 if (ret)
489 return ret;
490 }
491 ret = btrfs_set_disk_extent_flags(trans, buf,
492 BTRFS_BLOCK_FLAG_FULL_BACKREF);
493 if (ret)
494 return ret;
495 } else {
496
497 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID)
498 ret = btrfs_inc_ref(trans, root, cow, 1);
499 else
500 ret = btrfs_inc_ref(trans, root, cow, 0);
501 if (ret)
502 return ret;
503 }
504 } else {
505 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
506 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID)
507 ret = btrfs_inc_ref(trans, root, cow, 1);
508 else
509 ret = btrfs_inc_ref(trans, root, cow, 0);
510 if (ret)
511 return ret;
512 ret = btrfs_dec_ref(trans, root, buf, 1);
513 if (ret)
514 return ret;
515 }
516 btrfs_clear_buffer_dirty(trans, buf);
517 *last_ref = 1;
518 }
519 return 0;
520}
521
522/*
523 * does the dirty work in cow of a single block. The parent block (if
524 * supplied) is updated to point to the new cow copy. The new buffer is marked
525 * dirty and returned locked. If you modify the block it needs to be marked
526 * dirty again.
527 *
528 * search_start -- an allocation hint for the new block
529 *
530 * empty_size -- a hint that you plan on doing more cow. This is the size in
531 * bytes the allocator should try to find free next to the block it returns.
532 * This is just a hint and may be ignored by the allocator.
533 */
534int btrfs_force_cow_block(struct btrfs_trans_handle *trans,
535 struct btrfs_root *root,
536 struct extent_buffer *buf,
537 struct extent_buffer *parent, int parent_slot,
538 struct extent_buffer **cow_ret,
539 u64 search_start, u64 empty_size,
540 enum btrfs_lock_nesting nest)
541{
542 struct btrfs_fs_info *fs_info = root->fs_info;
543 struct btrfs_disk_key disk_key;
544 struct extent_buffer *cow;
545 int level, ret;
546 int last_ref = 0;
547 int unlock_orig = 0;
548 u64 parent_start = 0;
549 u64 reloc_src_root = 0;
550
551 if (*cow_ret == buf)
552 unlock_orig = 1;
553
554 btrfs_assert_tree_write_locked(buf);
555
556 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
557 trans->transid != fs_info->running_transaction->transid);
558 WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
559 trans->transid != btrfs_get_root_last_trans(root));
560
561 level = btrfs_header_level(buf);
562
563 if (level == 0)
564 btrfs_item_key(buf, &disk_key, 0);
565 else
566 btrfs_node_key(buf, &disk_key, 0);
567
568 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID) {
569 if (parent)
570 parent_start = parent->start;
571 reloc_src_root = btrfs_header_owner(buf);
572 }
573 cow = btrfs_alloc_tree_block(trans, root, parent_start,
574 btrfs_root_id(root), &disk_key, level,
575 search_start, empty_size, reloc_src_root, nest);
576 if (IS_ERR(cow))
577 return PTR_ERR(cow);
578
579 /* cow is set to blocking by btrfs_init_new_buffer */
580
581 copy_extent_buffer_full(cow, buf);
582 btrfs_set_header_bytenr(cow, cow->start);
583 btrfs_set_header_generation(cow, trans->transid);
584 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
585 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
586 BTRFS_HEADER_FLAG_RELOC);
587 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID)
588 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
589 else
590 btrfs_set_header_owner(cow, btrfs_root_id(root));
591
592 write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
593
594 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
595 if (ret) {
596 btrfs_abort_transaction(trans, ret);
597 goto error_unlock_cow;
598 }
599
600 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
601 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
602 if (ret) {
603 btrfs_abort_transaction(trans, ret);
604 goto error_unlock_cow;
605 }
606 }
607
608 if (buf == root->node) {
609 WARN_ON(parent && parent != buf);
610 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID ||
611 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
612 parent_start = buf->start;
613
614 ret = btrfs_tree_mod_log_insert_root(root->node, cow, true);
615 if (ret < 0) {
616 btrfs_abort_transaction(trans, ret);
617 goto error_unlock_cow;
618 }
619 atomic_inc(&cow->refs);
620 rcu_assign_pointer(root->node, cow);
621
622 ret = btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
623 parent_start, last_ref);
624 free_extent_buffer(buf);
625 add_root_to_dirty_list(root);
626 if (ret < 0) {
627 btrfs_abort_transaction(trans, ret);
628 goto error_unlock_cow;
629 }
630 } else {
631 WARN_ON(trans->transid != btrfs_header_generation(parent));
632 ret = btrfs_tree_mod_log_insert_key(parent, parent_slot,
633 BTRFS_MOD_LOG_KEY_REPLACE);
634 if (ret) {
635 btrfs_abort_transaction(trans, ret);
636 goto error_unlock_cow;
637 }
638 btrfs_set_node_blockptr(parent, parent_slot,
639 cow->start);
640 btrfs_set_node_ptr_generation(parent, parent_slot,
641 trans->transid);
642 btrfs_mark_buffer_dirty(trans, parent);
643 if (last_ref) {
644 ret = btrfs_tree_mod_log_free_eb(buf);
645 if (ret) {
646 btrfs_abort_transaction(trans, ret);
647 goto error_unlock_cow;
648 }
649 }
650 ret = btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
651 parent_start, last_ref);
652 if (ret < 0) {
653 btrfs_abort_transaction(trans, ret);
654 goto error_unlock_cow;
655 }
656 }
657
658 trace_btrfs_cow_block(root, buf, cow);
659 if (unlock_orig)
660 btrfs_tree_unlock(buf);
661 free_extent_buffer_stale(buf);
662 btrfs_mark_buffer_dirty(trans, cow);
663 *cow_ret = cow;
664 return 0;
665
666error_unlock_cow:
667 btrfs_tree_unlock(cow);
668 free_extent_buffer(cow);
669 return ret;
670}
671
672static inline int should_cow_block(struct btrfs_trans_handle *trans,
673 struct btrfs_root *root,
674 struct extent_buffer *buf)
675{
676 if (btrfs_is_testing(root->fs_info))
677 return 0;
678
679 /* Ensure we can see the FORCE_COW bit */
680 smp_mb__before_atomic();
681
682 /*
683 * We do not need to cow a block if
684 * 1) this block is not created or changed in this transaction;
685 * 2) this block does not belong to TREE_RELOC tree;
686 * 3) the root is not forced COW.
687 *
688 * What is forced COW:
689 * when we create snapshot during committing the transaction,
690 * after we've finished copying src root, we must COW the shared
691 * block to ensure the metadata consistency.
692 */
693 if (btrfs_header_generation(buf) == trans->transid &&
694 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
695 !(btrfs_root_id(root) != BTRFS_TREE_RELOC_OBJECTID &&
696 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
697 !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
698 return 0;
699 return 1;
700}
701
702/*
703 * COWs a single block, see btrfs_force_cow_block() for the real work.
704 * This version of it has extra checks so that a block isn't COWed more than
705 * once per transaction, as long as it hasn't been written yet
706 */
707int btrfs_cow_block(struct btrfs_trans_handle *trans,
708 struct btrfs_root *root, struct extent_buffer *buf,
709 struct extent_buffer *parent, int parent_slot,
710 struct extent_buffer **cow_ret,
711 enum btrfs_lock_nesting nest)
712{
713 struct btrfs_fs_info *fs_info = root->fs_info;
714 u64 search_start;
715
716 if (unlikely(test_bit(BTRFS_ROOT_DELETING, &root->state))) {
717 btrfs_abort_transaction(trans, -EUCLEAN);
718 btrfs_crit(fs_info,
719 "attempt to COW block %llu on root %llu that is being deleted",
720 buf->start, btrfs_root_id(root));
721 return -EUCLEAN;
722 }
723
724 /*
725 * COWing must happen through a running transaction, which always
726 * matches the current fs generation (it's a transaction with a state
727 * less than TRANS_STATE_UNBLOCKED). If it doesn't, then turn the fs
728 * into error state to prevent the commit of any transaction.
729 */
730 if (unlikely(trans->transaction != fs_info->running_transaction ||
731 trans->transid != fs_info->generation)) {
732 btrfs_abort_transaction(trans, -EUCLEAN);
733 btrfs_crit(fs_info,
734"unexpected transaction when attempting to COW block %llu on root %llu, transaction %llu running transaction %llu fs generation %llu",
735 buf->start, btrfs_root_id(root), trans->transid,
736 fs_info->running_transaction->transid,
737 fs_info->generation);
738 return -EUCLEAN;
739 }
740
741 if (!should_cow_block(trans, root, buf)) {
742 *cow_ret = buf;
743 return 0;
744 }
745
746 search_start = round_down(buf->start, SZ_1G);
747
748 /*
749 * Before CoWing this block for later modification, check if it's
750 * the subtree root and do the delayed subtree trace if needed.
751 *
752 * Also We don't care about the error, as it's handled internally.
753 */
754 btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
755 return btrfs_force_cow_block(trans, root, buf, parent, parent_slot,
756 cow_ret, search_start, 0, nest);
757}
758ALLOW_ERROR_INJECTION(btrfs_cow_block, ERRNO);
759
760/*
761 * same as comp_keys only with two btrfs_key's
762 */
763int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
764{
765 if (k1->objectid > k2->objectid)
766 return 1;
767 if (k1->objectid < k2->objectid)
768 return -1;
769 if (k1->type > k2->type)
770 return 1;
771 if (k1->type < k2->type)
772 return -1;
773 if (k1->offset > k2->offset)
774 return 1;
775 if (k1->offset < k2->offset)
776 return -1;
777 return 0;
778}
779
780/*
781 * Search for a key in the given extent_buffer.
782 *
783 * The lower boundary for the search is specified by the slot number @first_slot.
784 * Use a value of 0 to search over the whole extent buffer. Works for both
785 * leaves and nodes.
786 *
787 * The slot in the extent buffer is returned via @slot. If the key exists in the
788 * extent buffer, then @slot will point to the slot where the key is, otherwise
789 * it points to the slot where you would insert the key.
790 *
791 * Slot may point to the total number of items (i.e. one position beyond the last
792 * key) if the key is bigger than the last key in the extent buffer.
793 */
794int btrfs_bin_search(struct extent_buffer *eb, int first_slot,
795 const struct btrfs_key *key, int *slot)
796{
797 unsigned long p;
798 int item_size;
799 /*
800 * Use unsigned types for the low and high slots, so that we get a more
801 * efficient division in the search loop below.
802 */
803 u32 low = first_slot;
804 u32 high = btrfs_header_nritems(eb);
805 int ret;
806 const int key_size = sizeof(struct btrfs_disk_key);
807
808 if (unlikely(low > high)) {
809 btrfs_err(eb->fs_info,
810 "%s: low (%u) > high (%u) eb %llu owner %llu level %d",
811 __func__, low, high, eb->start,
812 btrfs_header_owner(eb), btrfs_header_level(eb));
813 return -EINVAL;
814 }
815
816 if (btrfs_header_level(eb) == 0) {
817 p = offsetof(struct btrfs_leaf, items);
818 item_size = sizeof(struct btrfs_item);
819 } else {
820 p = offsetof(struct btrfs_node, ptrs);
821 item_size = sizeof(struct btrfs_key_ptr);
822 }
823
824 while (low < high) {
825 const int unit_size = eb->folio_size;
826 unsigned long oil;
827 unsigned long offset;
828 struct btrfs_disk_key *tmp;
829 struct btrfs_disk_key unaligned;
830 int mid;
831
832 mid = (low + high) / 2;
833 offset = p + mid * item_size;
834 oil = get_eb_offset_in_folio(eb, offset);
835
836 if (oil + key_size <= unit_size) {
837 const unsigned long idx = get_eb_folio_index(eb, offset);
838 char *kaddr = folio_address(eb->folios[idx]);
839
840 oil = get_eb_offset_in_folio(eb, offset);
841 tmp = (struct btrfs_disk_key *)(kaddr + oil);
842 } else {
843 read_extent_buffer(eb, &unaligned, offset, key_size);
844 tmp = &unaligned;
845 }
846
847 ret = btrfs_comp_keys(tmp, key);
848
849 if (ret < 0)
850 low = mid + 1;
851 else if (ret > 0)
852 high = mid;
853 else {
854 *slot = mid;
855 return 0;
856 }
857 }
858 *slot = low;
859 return 1;
860}
861
862static void root_add_used_bytes(struct btrfs_root *root)
863{
864 spin_lock(&root->accounting_lock);
865 btrfs_set_root_used(&root->root_item,
866 btrfs_root_used(&root->root_item) + root->fs_info->nodesize);
867 spin_unlock(&root->accounting_lock);
868}
869
870static void root_sub_used_bytes(struct btrfs_root *root)
871{
872 spin_lock(&root->accounting_lock);
873 btrfs_set_root_used(&root->root_item,
874 btrfs_root_used(&root->root_item) - root->fs_info->nodesize);
875 spin_unlock(&root->accounting_lock);
876}
877
878/* given a node and slot number, this reads the blocks it points to. The
879 * extent buffer is returned with a reference taken (but unlocked).
880 */
881struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
882 int slot)
883{
884 int level = btrfs_header_level(parent);
885 struct btrfs_tree_parent_check check = { 0 };
886 struct extent_buffer *eb;
887
888 if (slot < 0 || slot >= btrfs_header_nritems(parent))
889 return ERR_PTR(-ENOENT);
890
891 ASSERT(level);
892
893 check.level = level - 1;
894 check.transid = btrfs_node_ptr_generation(parent, slot);
895 check.owner_root = btrfs_header_owner(parent);
896 check.has_first_key = true;
897 btrfs_node_key_to_cpu(parent, &check.first_key, slot);
898
899 eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
900 &check);
901 if (IS_ERR(eb))
902 return eb;
903 if (!extent_buffer_uptodate(eb)) {
904 free_extent_buffer(eb);
905 return ERR_PTR(-EIO);
906 }
907
908 return eb;
909}
910
911/*
912 * node level balancing, used to make sure nodes are in proper order for
913 * item deletion. We balance from the top down, so we have to make sure
914 * that a deletion won't leave an node completely empty later on.
915 */
916static noinline int balance_level(struct btrfs_trans_handle *trans,
917 struct btrfs_root *root,
918 struct btrfs_path *path, int level)
919{
920 struct btrfs_fs_info *fs_info = root->fs_info;
921 struct extent_buffer *right = NULL;
922 struct extent_buffer *mid;
923 struct extent_buffer *left = NULL;
924 struct extent_buffer *parent = NULL;
925 int ret = 0;
926 int wret;
927 int pslot;
928 int orig_slot = path->slots[level];
929 u64 orig_ptr;
930
931 ASSERT(level > 0);
932
933 mid = path->nodes[level];
934
935 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK);
936 WARN_ON(btrfs_header_generation(mid) != trans->transid);
937
938 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
939
940 if (level < BTRFS_MAX_LEVEL - 1) {
941 parent = path->nodes[level + 1];
942 pslot = path->slots[level + 1];
943 }
944
945 /*
946 * deal with the case where there is only one pointer in the root
947 * by promoting the node below to a root
948 */
949 if (!parent) {
950 struct extent_buffer *child;
951
952 if (btrfs_header_nritems(mid) != 1)
953 return 0;
954
955 /* promote the child to a root */
956 child = btrfs_read_node_slot(mid, 0);
957 if (IS_ERR(child)) {
958 ret = PTR_ERR(child);
959 goto out;
960 }
961
962 btrfs_tree_lock(child);
963 ret = btrfs_cow_block(trans, root, child, mid, 0, &child,
964 BTRFS_NESTING_COW);
965 if (ret) {
966 btrfs_tree_unlock(child);
967 free_extent_buffer(child);
968 goto out;
969 }
970
971 ret = btrfs_tree_mod_log_insert_root(root->node, child, true);
972 if (ret < 0) {
973 btrfs_tree_unlock(child);
974 free_extent_buffer(child);
975 btrfs_abort_transaction(trans, ret);
976 goto out;
977 }
978 rcu_assign_pointer(root->node, child);
979
980 add_root_to_dirty_list(root);
981 btrfs_tree_unlock(child);
982
983 path->locks[level] = 0;
984 path->nodes[level] = NULL;
985 btrfs_clear_buffer_dirty(trans, mid);
986 btrfs_tree_unlock(mid);
987 /* once for the path */
988 free_extent_buffer(mid);
989
990 root_sub_used_bytes(root);
991 ret = btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
992 /* once for the root ptr */
993 free_extent_buffer_stale(mid);
994 if (ret < 0) {
995 btrfs_abort_transaction(trans, ret);
996 goto out;
997 }
998 return 0;
999 }
1000 if (btrfs_header_nritems(mid) >
1001 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1002 return 0;
1003
1004 if (pslot) {
1005 left = btrfs_read_node_slot(parent, pslot - 1);
1006 if (IS_ERR(left)) {
1007 ret = PTR_ERR(left);
1008 left = NULL;
1009 goto out;
1010 }
1011
1012 btrfs_tree_lock_nested(left, BTRFS_NESTING_LEFT);
1013 wret = btrfs_cow_block(trans, root, left,
1014 parent, pslot - 1, &left,
1015 BTRFS_NESTING_LEFT_COW);
1016 if (wret) {
1017 ret = wret;
1018 goto out;
1019 }
1020 }
1021
1022 if (pslot + 1 < btrfs_header_nritems(parent)) {
1023 right = btrfs_read_node_slot(parent, pslot + 1);
1024 if (IS_ERR(right)) {
1025 ret = PTR_ERR(right);
1026 right = NULL;
1027 goto out;
1028 }
1029
1030 btrfs_tree_lock_nested(right, BTRFS_NESTING_RIGHT);
1031 wret = btrfs_cow_block(trans, root, right,
1032 parent, pslot + 1, &right,
1033 BTRFS_NESTING_RIGHT_COW);
1034 if (wret) {
1035 ret = wret;
1036 goto out;
1037 }
1038 }
1039
1040 /* first, try to make some room in the middle buffer */
1041 if (left) {
1042 orig_slot += btrfs_header_nritems(left);
1043 wret = push_node_left(trans, left, mid, 1);
1044 if (wret < 0)
1045 ret = wret;
1046 }
1047
1048 /*
1049 * then try to empty the right most buffer into the middle
1050 */
1051 if (right) {
1052 wret = push_node_left(trans, mid, right, 1);
1053 if (wret < 0 && wret != -ENOSPC)
1054 ret = wret;
1055 if (btrfs_header_nritems(right) == 0) {
1056 btrfs_clear_buffer_dirty(trans, right);
1057 btrfs_tree_unlock(right);
1058 ret = btrfs_del_ptr(trans, root, path, level + 1, pslot + 1);
1059 if (ret < 0) {
1060 free_extent_buffer_stale(right);
1061 right = NULL;
1062 goto out;
1063 }
1064 root_sub_used_bytes(root);
1065 ret = btrfs_free_tree_block(trans, btrfs_root_id(root),
1066 right, 0, 1);
1067 free_extent_buffer_stale(right);
1068 right = NULL;
1069 if (ret < 0) {
1070 btrfs_abort_transaction(trans, ret);
1071 goto out;
1072 }
1073 } else {
1074 struct btrfs_disk_key right_key;
1075 btrfs_node_key(right, &right_key, 0);
1076 ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1077 BTRFS_MOD_LOG_KEY_REPLACE);
1078 if (ret < 0) {
1079 btrfs_abort_transaction(trans, ret);
1080 goto out;
1081 }
1082 btrfs_set_node_key(parent, &right_key, pslot + 1);
1083 btrfs_mark_buffer_dirty(trans, parent);
1084 }
1085 }
1086 if (btrfs_header_nritems(mid) == 1) {
1087 /*
1088 * we're not allowed to leave a node with one item in the
1089 * tree during a delete. A deletion from lower in the tree
1090 * could try to delete the only pointer in this node.
1091 * So, pull some keys from the left.
1092 * There has to be a left pointer at this point because
1093 * otherwise we would have pulled some pointers from the
1094 * right
1095 */
1096 if (unlikely(!left)) {
1097 btrfs_crit(fs_info,
1098"missing left child when middle child only has 1 item, parent bytenr %llu level %d mid bytenr %llu root %llu",
1099 parent->start, btrfs_header_level(parent),
1100 mid->start, btrfs_root_id(root));
1101 ret = -EUCLEAN;
1102 btrfs_abort_transaction(trans, ret);
1103 goto out;
1104 }
1105 wret = balance_node_right(trans, mid, left);
1106 if (wret < 0) {
1107 ret = wret;
1108 goto out;
1109 }
1110 if (wret == 1) {
1111 wret = push_node_left(trans, left, mid, 1);
1112 if (wret < 0)
1113 ret = wret;
1114 }
1115 BUG_ON(wret == 1);
1116 }
1117 if (btrfs_header_nritems(mid) == 0) {
1118 btrfs_clear_buffer_dirty(trans, mid);
1119 btrfs_tree_unlock(mid);
1120 ret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
1121 if (ret < 0) {
1122 free_extent_buffer_stale(mid);
1123 mid = NULL;
1124 goto out;
1125 }
1126 root_sub_used_bytes(root);
1127 ret = btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
1128 free_extent_buffer_stale(mid);
1129 mid = NULL;
1130 if (ret < 0) {
1131 btrfs_abort_transaction(trans, ret);
1132 goto out;
1133 }
1134 } else {
1135 /* update the parent key to reflect our changes */
1136 struct btrfs_disk_key mid_key;
1137 btrfs_node_key(mid, &mid_key, 0);
1138 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1139 BTRFS_MOD_LOG_KEY_REPLACE);
1140 if (ret < 0) {
1141 btrfs_abort_transaction(trans, ret);
1142 goto out;
1143 }
1144 btrfs_set_node_key(parent, &mid_key, pslot);
1145 btrfs_mark_buffer_dirty(trans, parent);
1146 }
1147
1148 /* update the path */
1149 if (left) {
1150 if (btrfs_header_nritems(left) > orig_slot) {
1151 atomic_inc(&left->refs);
1152 /* left was locked after cow */
1153 path->nodes[level] = left;
1154 path->slots[level + 1] -= 1;
1155 path->slots[level] = orig_slot;
1156 if (mid) {
1157 btrfs_tree_unlock(mid);
1158 free_extent_buffer(mid);
1159 }
1160 } else {
1161 orig_slot -= btrfs_header_nritems(left);
1162 path->slots[level] = orig_slot;
1163 }
1164 }
1165 /* double check we haven't messed things up */
1166 if (orig_ptr !=
1167 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1168 BUG();
1169out:
1170 if (right) {
1171 btrfs_tree_unlock(right);
1172 free_extent_buffer(right);
1173 }
1174 if (left) {
1175 if (path->nodes[level] != left)
1176 btrfs_tree_unlock(left);
1177 free_extent_buffer(left);
1178 }
1179 return ret;
1180}
1181
1182/* Node balancing for insertion. Here we only split or push nodes around
1183 * when they are completely full. This is also done top down, so we
1184 * have to be pessimistic.
1185 */
1186static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1187 struct btrfs_root *root,
1188 struct btrfs_path *path, int level)
1189{
1190 struct btrfs_fs_info *fs_info = root->fs_info;
1191 struct extent_buffer *right = NULL;
1192 struct extent_buffer *mid;
1193 struct extent_buffer *left = NULL;
1194 struct extent_buffer *parent = NULL;
1195 int ret = 0;
1196 int wret;
1197 int pslot;
1198 int orig_slot = path->slots[level];
1199
1200 if (level == 0)
1201 return 1;
1202
1203 mid = path->nodes[level];
1204 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1205
1206 if (level < BTRFS_MAX_LEVEL - 1) {
1207 parent = path->nodes[level + 1];
1208 pslot = path->slots[level + 1];
1209 }
1210
1211 if (!parent)
1212 return 1;
1213
1214 /* first, try to make some room in the middle buffer */
1215 if (pslot) {
1216 u32 left_nr;
1217
1218 left = btrfs_read_node_slot(parent, pslot - 1);
1219 if (IS_ERR(left))
1220 return PTR_ERR(left);
1221
1222 btrfs_tree_lock_nested(left, BTRFS_NESTING_LEFT);
1223
1224 left_nr = btrfs_header_nritems(left);
1225 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1226 wret = 1;
1227 } else {
1228 ret = btrfs_cow_block(trans, root, left, parent,
1229 pslot - 1, &left,
1230 BTRFS_NESTING_LEFT_COW);
1231 if (ret)
1232 wret = 1;
1233 else {
1234 wret = push_node_left(trans, left, mid, 0);
1235 }
1236 }
1237 if (wret < 0)
1238 ret = wret;
1239 if (wret == 0) {
1240 struct btrfs_disk_key disk_key;
1241 orig_slot += left_nr;
1242 btrfs_node_key(mid, &disk_key, 0);
1243 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1244 BTRFS_MOD_LOG_KEY_REPLACE);
1245 if (ret < 0) {
1246 btrfs_tree_unlock(left);
1247 free_extent_buffer(left);
1248 btrfs_abort_transaction(trans, ret);
1249 return ret;
1250 }
1251 btrfs_set_node_key(parent, &disk_key, pslot);
1252 btrfs_mark_buffer_dirty(trans, parent);
1253 if (btrfs_header_nritems(left) > orig_slot) {
1254 path->nodes[level] = left;
1255 path->slots[level + 1] -= 1;
1256 path->slots[level] = orig_slot;
1257 btrfs_tree_unlock(mid);
1258 free_extent_buffer(mid);
1259 } else {
1260 orig_slot -=
1261 btrfs_header_nritems(left);
1262 path->slots[level] = orig_slot;
1263 btrfs_tree_unlock(left);
1264 free_extent_buffer(left);
1265 }
1266 return 0;
1267 }
1268 btrfs_tree_unlock(left);
1269 free_extent_buffer(left);
1270 }
1271
1272 /*
1273 * then try to empty the right most buffer into the middle
1274 */
1275 if (pslot + 1 < btrfs_header_nritems(parent)) {
1276 u32 right_nr;
1277
1278 right = btrfs_read_node_slot(parent, pslot + 1);
1279 if (IS_ERR(right))
1280 return PTR_ERR(right);
1281
1282 btrfs_tree_lock_nested(right, BTRFS_NESTING_RIGHT);
1283
1284 right_nr = btrfs_header_nritems(right);
1285 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1286 wret = 1;
1287 } else {
1288 ret = btrfs_cow_block(trans, root, right,
1289 parent, pslot + 1,
1290 &right, BTRFS_NESTING_RIGHT_COW);
1291 if (ret)
1292 wret = 1;
1293 else {
1294 wret = balance_node_right(trans, right, mid);
1295 }
1296 }
1297 if (wret < 0)
1298 ret = wret;
1299 if (wret == 0) {
1300 struct btrfs_disk_key disk_key;
1301
1302 btrfs_node_key(right, &disk_key, 0);
1303 ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1304 BTRFS_MOD_LOG_KEY_REPLACE);
1305 if (ret < 0) {
1306 btrfs_tree_unlock(right);
1307 free_extent_buffer(right);
1308 btrfs_abort_transaction(trans, ret);
1309 return ret;
1310 }
1311 btrfs_set_node_key(parent, &disk_key, pslot + 1);
1312 btrfs_mark_buffer_dirty(trans, parent);
1313
1314 if (btrfs_header_nritems(mid) <= orig_slot) {
1315 path->nodes[level] = right;
1316 path->slots[level + 1] += 1;
1317 path->slots[level] = orig_slot -
1318 btrfs_header_nritems(mid);
1319 btrfs_tree_unlock(mid);
1320 free_extent_buffer(mid);
1321 } else {
1322 btrfs_tree_unlock(right);
1323 free_extent_buffer(right);
1324 }
1325 return 0;
1326 }
1327 btrfs_tree_unlock(right);
1328 free_extent_buffer(right);
1329 }
1330 return 1;
1331}
1332
1333/*
1334 * readahead one full node of leaves, finding things that are close
1335 * to the block in 'slot', and triggering ra on them.
1336 */
1337static void reada_for_search(struct btrfs_fs_info *fs_info,
1338 struct btrfs_path *path,
1339 int level, int slot, u64 objectid)
1340{
1341 struct extent_buffer *node;
1342 struct btrfs_disk_key disk_key;
1343 u32 nritems;
1344 u64 search;
1345 u64 target;
1346 u64 nread = 0;
1347 u64 nread_max;
1348 u32 nr;
1349 u32 blocksize;
1350 u32 nscan = 0;
1351
1352 if (level != 1 && path->reada != READA_FORWARD_ALWAYS)
1353 return;
1354
1355 if (!path->nodes[level])
1356 return;
1357
1358 node = path->nodes[level];
1359
1360 /*
1361 * Since the time between visiting leaves is much shorter than the time
1362 * between visiting nodes, limit read ahead of nodes to 1, to avoid too
1363 * much IO at once (possibly random).
1364 */
1365 if (path->reada == READA_FORWARD_ALWAYS) {
1366 if (level > 1)
1367 nread_max = node->fs_info->nodesize;
1368 else
1369 nread_max = SZ_128K;
1370 } else {
1371 nread_max = SZ_64K;
1372 }
1373
1374 search = btrfs_node_blockptr(node, slot);
1375 blocksize = fs_info->nodesize;
1376 if (path->reada != READA_FORWARD_ALWAYS) {
1377 struct extent_buffer *eb;
1378
1379 eb = find_extent_buffer(fs_info, search);
1380 if (eb) {
1381 free_extent_buffer(eb);
1382 return;
1383 }
1384 }
1385
1386 target = search;
1387
1388 nritems = btrfs_header_nritems(node);
1389 nr = slot;
1390
1391 while (1) {
1392 if (path->reada == READA_BACK) {
1393 if (nr == 0)
1394 break;
1395 nr--;
1396 } else if (path->reada == READA_FORWARD ||
1397 path->reada == READA_FORWARD_ALWAYS) {
1398 nr++;
1399 if (nr >= nritems)
1400 break;
1401 }
1402 if (path->reada == READA_BACK && objectid) {
1403 btrfs_node_key(node, &disk_key, nr);
1404 if (btrfs_disk_key_objectid(&disk_key) != objectid)
1405 break;
1406 }
1407 search = btrfs_node_blockptr(node, nr);
1408 if (path->reada == READA_FORWARD_ALWAYS ||
1409 (search <= target && target - search <= 65536) ||
1410 (search > target && search - target <= 65536)) {
1411 btrfs_readahead_node_child(node, nr);
1412 nread += blocksize;
1413 }
1414 nscan++;
1415 if (nread > nread_max || nscan > 32)
1416 break;
1417 }
1418}
1419
1420static noinline void reada_for_balance(struct btrfs_path *path, int level)
1421{
1422 struct extent_buffer *parent;
1423 int slot;
1424 int nritems;
1425
1426 parent = path->nodes[level + 1];
1427 if (!parent)
1428 return;
1429
1430 nritems = btrfs_header_nritems(parent);
1431 slot = path->slots[level + 1];
1432
1433 if (slot > 0)
1434 btrfs_readahead_node_child(parent, slot - 1);
1435 if (slot + 1 < nritems)
1436 btrfs_readahead_node_child(parent, slot + 1);
1437}
1438
1439
1440/*
1441 * when we walk down the tree, it is usually safe to unlock the higher layers
1442 * in the tree. The exceptions are when our path goes through slot 0, because
1443 * operations on the tree might require changing key pointers higher up in the
1444 * tree.
1445 *
1446 * callers might also have set path->keep_locks, which tells this code to keep
1447 * the lock if the path points to the last slot in the block. This is part of
1448 * walking through the tree, and selecting the next slot in the higher block.
1449 *
1450 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
1451 * if lowest_unlock is 1, level 0 won't be unlocked
1452 */
1453static noinline void unlock_up(struct btrfs_path *path, int level,
1454 int lowest_unlock, int min_write_lock_level,
1455 int *write_lock_level)
1456{
1457 int i;
1458 int skip_level = level;
1459 bool check_skip = true;
1460
1461 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1462 if (!path->nodes[i])
1463 break;
1464 if (!path->locks[i])
1465 break;
1466
1467 if (check_skip) {
1468 if (path->slots[i] == 0) {
1469 skip_level = i + 1;
1470 continue;
1471 }
1472
1473 if (path->keep_locks) {
1474 u32 nritems;
1475
1476 nritems = btrfs_header_nritems(path->nodes[i]);
1477 if (nritems < 1 || path->slots[i] >= nritems - 1) {
1478 skip_level = i + 1;
1479 continue;
1480 }
1481 }
1482 }
1483
1484 if (i >= lowest_unlock && i > skip_level) {
1485 check_skip = false;
1486 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
1487 path->locks[i] = 0;
1488 if (write_lock_level &&
1489 i > min_write_lock_level &&
1490 i <= *write_lock_level) {
1491 *write_lock_level = i - 1;
1492 }
1493 }
1494 }
1495}
1496
1497/*
1498 * Helper function for btrfs_search_slot() and other functions that do a search
1499 * on a btree. The goal is to find a tree block in the cache (the radix tree at
1500 * fs_info->buffer_radix), but if we can't find it, or it's not up to date, read
1501 * its pages from disk.
1502 *
1503 * Returns -EAGAIN, with the path unlocked, if the caller needs to repeat the
1504 * whole btree search, starting again from the current root node.
1505 */
1506static int
1507read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
1508 struct extent_buffer **eb_ret, int slot,
1509 const struct btrfs_key *key)
1510{
1511 struct btrfs_fs_info *fs_info = root->fs_info;
1512 struct btrfs_tree_parent_check check = { 0 };
1513 u64 blocknr;
1514 struct extent_buffer *tmp = NULL;
1515 int ret = 0;
1516 int parent_level;
1517 int err;
1518 bool read_tmp = false;
1519 bool tmp_locked = false;
1520 bool path_released = false;
1521
1522 blocknr = btrfs_node_blockptr(*eb_ret, slot);
1523 parent_level = btrfs_header_level(*eb_ret);
1524 btrfs_node_key_to_cpu(*eb_ret, &check.first_key, slot);
1525 check.has_first_key = true;
1526 check.level = parent_level - 1;
1527 check.transid = btrfs_node_ptr_generation(*eb_ret, slot);
1528 check.owner_root = btrfs_root_id(root);
1529
1530 /*
1531 * If we need to read an extent buffer from disk and we are holding locks
1532 * on upper level nodes, we unlock all the upper nodes before reading the
1533 * extent buffer, and then return -EAGAIN to the caller as it needs to
1534 * restart the search. We don't release the lock on the current level
1535 * because we need to walk this node to figure out which blocks to read.
1536 */
1537 tmp = find_extent_buffer(fs_info, blocknr);
1538 if (tmp) {
1539 if (p->reada == READA_FORWARD_ALWAYS)
1540 reada_for_search(fs_info, p, parent_level, slot, key->objectid);
1541
1542 /* first we do an atomic uptodate check */
1543 if (btrfs_buffer_uptodate(tmp, check.transid, 1) > 0) {
1544 /*
1545 * Do extra check for first_key, eb can be stale due to
1546 * being cached, read from scrub, or have multiple
1547 * parents (shared tree blocks).
1548 */
1549 if (btrfs_verify_level_key(tmp, &check)) {
1550 ret = -EUCLEAN;
1551 goto out;
1552 }
1553 *eb_ret = tmp;
1554 tmp = NULL;
1555 ret = 0;
1556 goto out;
1557 }
1558
1559 if (p->nowait) {
1560 ret = -EAGAIN;
1561 goto out;
1562 }
1563
1564 if (!p->skip_locking) {
1565 btrfs_unlock_up_safe(p, parent_level + 1);
1566 btrfs_maybe_reset_lockdep_class(root, tmp);
1567 tmp_locked = true;
1568 btrfs_tree_read_lock(tmp);
1569 btrfs_release_path(p);
1570 ret = -EAGAIN;
1571 path_released = true;
1572 }
1573
1574 /* Now we're allowed to do a blocking uptodate check. */
1575 err = btrfs_read_extent_buffer(tmp, &check);
1576 if (err) {
1577 ret = err;
1578 goto out;
1579 }
1580
1581 if (ret == 0) {
1582 ASSERT(!tmp_locked);
1583 *eb_ret = tmp;
1584 tmp = NULL;
1585 }
1586 goto out;
1587 } else if (p->nowait) {
1588 ret = -EAGAIN;
1589 goto out;
1590 }
1591
1592 if (!p->skip_locking) {
1593 btrfs_unlock_up_safe(p, parent_level + 1);
1594 ret = -EAGAIN;
1595 }
1596
1597 if (p->reada != READA_NONE)
1598 reada_for_search(fs_info, p, parent_level, slot, key->objectid);
1599
1600 tmp = btrfs_find_create_tree_block(fs_info, blocknr, check.owner_root, check.level);
1601 if (IS_ERR(tmp)) {
1602 ret = PTR_ERR(tmp);
1603 tmp = NULL;
1604 goto out;
1605 }
1606 read_tmp = true;
1607
1608 if (!p->skip_locking) {
1609 ASSERT(ret == -EAGAIN);
1610 btrfs_maybe_reset_lockdep_class(root, tmp);
1611 tmp_locked = true;
1612 btrfs_tree_read_lock(tmp);
1613 btrfs_release_path(p);
1614 path_released = true;
1615 }
1616
1617 /* Now we're allowed to do a blocking uptodate check. */
1618 err = btrfs_read_extent_buffer(tmp, &check);
1619 if (err) {
1620 ret = err;
1621 goto out;
1622 }
1623
1624 /*
1625 * If the read above didn't mark this buffer up to date,
1626 * it will never end up being up to date. Set ret to EIO now
1627 * and give up so that our caller doesn't loop forever
1628 * on our EAGAINs.
1629 */
1630 if (!extent_buffer_uptodate(tmp)) {
1631 ret = -EIO;
1632 goto out;
1633 }
1634
1635 if (ret == 0) {
1636 ASSERT(!tmp_locked);
1637 *eb_ret = tmp;
1638 tmp = NULL;
1639 }
1640out:
1641 if (tmp) {
1642 if (tmp_locked)
1643 btrfs_tree_read_unlock(tmp);
1644 if (read_tmp && ret && ret != -EAGAIN)
1645 free_extent_buffer_stale(tmp);
1646 else
1647 free_extent_buffer(tmp);
1648 }
1649 if (ret && !path_released)
1650 btrfs_release_path(p);
1651
1652 return ret;
1653}
1654
1655/*
1656 * helper function for btrfs_search_slot. This does all of the checks
1657 * for node-level blocks and does any balancing required based on
1658 * the ins_len.
1659 *
1660 * If no extra work was required, zero is returned. If we had to
1661 * drop the path, -EAGAIN is returned and btrfs_search_slot must
1662 * start over
1663 */
1664static int
1665setup_nodes_for_search(struct btrfs_trans_handle *trans,
1666 struct btrfs_root *root, struct btrfs_path *p,
1667 struct extent_buffer *b, int level, int ins_len,
1668 int *write_lock_level)
1669{
1670 struct btrfs_fs_info *fs_info = root->fs_info;
1671 int ret = 0;
1672
1673 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1674 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
1675
1676 if (*write_lock_level < level + 1) {
1677 *write_lock_level = level + 1;
1678 btrfs_release_path(p);
1679 return -EAGAIN;
1680 }
1681
1682 reada_for_balance(p, level);
1683 ret = split_node(trans, root, p, level);
1684
1685 b = p->nodes[level];
1686 } else if (ins_len < 0 && btrfs_header_nritems(b) <
1687 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
1688
1689 if (*write_lock_level < level + 1) {
1690 *write_lock_level = level + 1;
1691 btrfs_release_path(p);
1692 return -EAGAIN;
1693 }
1694
1695 reada_for_balance(p, level);
1696 ret = balance_level(trans, root, p, level);
1697 if (ret)
1698 return ret;
1699
1700 b = p->nodes[level];
1701 if (!b) {
1702 btrfs_release_path(p);
1703 return -EAGAIN;
1704 }
1705 BUG_ON(btrfs_header_nritems(b) == 1);
1706 }
1707 return ret;
1708}
1709
1710int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
1711 u64 iobjectid, u64 ioff, u8 key_type,
1712 struct btrfs_key *found_key)
1713{
1714 int ret;
1715 struct btrfs_key key;
1716 struct extent_buffer *eb;
1717
1718 ASSERT(path);
1719 ASSERT(found_key);
1720
1721 key.type = key_type;
1722 key.objectid = iobjectid;
1723 key.offset = ioff;
1724
1725 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1726 if (ret < 0)
1727 return ret;
1728
1729 eb = path->nodes[0];
1730 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1731 ret = btrfs_next_leaf(fs_root, path);
1732 if (ret)
1733 return ret;
1734 eb = path->nodes[0];
1735 }
1736
1737 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1738 if (found_key->type != key.type ||
1739 found_key->objectid != key.objectid)
1740 return 1;
1741
1742 return 0;
1743}
1744
1745static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
1746 struct btrfs_path *p,
1747 int write_lock_level)
1748{
1749 struct extent_buffer *b;
1750 int root_lock = 0;
1751 int level = 0;
1752
1753 if (p->search_commit_root) {
1754 b = root->commit_root;
1755 atomic_inc(&b->refs);
1756 level = btrfs_header_level(b);
1757 /*
1758 * Ensure that all callers have set skip_locking when
1759 * p->search_commit_root = 1.
1760 */
1761 ASSERT(p->skip_locking == 1);
1762
1763 goto out;
1764 }
1765
1766 if (p->skip_locking) {
1767 b = btrfs_root_node(root);
1768 level = btrfs_header_level(b);
1769 goto out;
1770 }
1771
1772 /* We try very hard to do read locks on the root */
1773 root_lock = BTRFS_READ_LOCK;
1774
1775 /*
1776 * If the level is set to maximum, we can skip trying to get the read
1777 * lock.
1778 */
1779 if (write_lock_level < BTRFS_MAX_LEVEL) {
1780 /*
1781 * We don't know the level of the root node until we actually
1782 * have it read locked
1783 */
1784 if (p->nowait) {
1785 b = btrfs_try_read_lock_root_node(root);
1786 if (IS_ERR(b))
1787 return b;
1788 } else {
1789 b = btrfs_read_lock_root_node(root);
1790 }
1791 level = btrfs_header_level(b);
1792 if (level > write_lock_level)
1793 goto out;
1794
1795 /* Whoops, must trade for write lock */
1796 btrfs_tree_read_unlock(b);
1797 free_extent_buffer(b);
1798 }
1799
1800 b = btrfs_lock_root_node(root);
1801 root_lock = BTRFS_WRITE_LOCK;
1802
1803 /* The level might have changed, check again */
1804 level = btrfs_header_level(b);
1805
1806out:
1807 /*
1808 * The root may have failed to write out at some point, and thus is no
1809 * longer valid, return an error in this case.
1810 */
1811 if (!extent_buffer_uptodate(b)) {
1812 if (root_lock)
1813 btrfs_tree_unlock_rw(b, root_lock);
1814 free_extent_buffer(b);
1815 return ERR_PTR(-EIO);
1816 }
1817
1818 p->nodes[level] = b;
1819 if (!p->skip_locking)
1820 p->locks[level] = root_lock;
1821 /*
1822 * Callers are responsible for dropping b's references.
1823 */
1824 return b;
1825}
1826
1827/*
1828 * Replace the extent buffer at the lowest level of the path with a cloned
1829 * version. The purpose is to be able to use it safely, after releasing the
1830 * commit root semaphore, even if relocation is happening in parallel, the
1831 * transaction used for relocation is committed and the extent buffer is
1832 * reallocated in the next transaction.
1833 *
1834 * This is used in a context where the caller does not prevent transaction
1835 * commits from happening, either by holding a transaction handle or holding
1836 * some lock, while it's doing searches through a commit root.
1837 * At the moment it's only used for send operations.
1838 */
1839static int finish_need_commit_sem_search(struct btrfs_path *path)
1840{
1841 const int i = path->lowest_level;
1842 const int slot = path->slots[i];
1843 struct extent_buffer *lowest = path->nodes[i];
1844 struct extent_buffer *clone;
1845
1846 ASSERT(path->need_commit_sem);
1847
1848 if (!lowest)
1849 return 0;
1850
1851 lockdep_assert_held_read(&lowest->fs_info->commit_root_sem);
1852
1853 clone = btrfs_clone_extent_buffer(lowest);
1854 if (!clone)
1855 return -ENOMEM;
1856
1857 btrfs_release_path(path);
1858 path->nodes[i] = clone;
1859 path->slots[i] = slot;
1860
1861 return 0;
1862}
1863
1864static inline int search_for_key_slot(struct extent_buffer *eb,
1865 int search_low_slot,
1866 const struct btrfs_key *key,
1867 int prev_cmp,
1868 int *slot)
1869{
1870 /*
1871 * If a previous call to btrfs_bin_search() on a parent node returned an
1872 * exact match (prev_cmp == 0), we can safely assume the target key will
1873 * always be at slot 0 on lower levels, since each key pointer
1874 * (struct btrfs_key_ptr) refers to the lowest key accessible from the
1875 * subtree it points to. Thus we can skip searching lower levels.
1876 */
1877 if (prev_cmp == 0) {
1878 *slot = 0;
1879 return 0;
1880 }
1881
1882 return btrfs_bin_search(eb, search_low_slot, key, slot);
1883}
1884
1885static int search_leaf(struct btrfs_trans_handle *trans,
1886 struct btrfs_root *root,
1887 const struct btrfs_key *key,
1888 struct btrfs_path *path,
1889 int ins_len,
1890 int prev_cmp)
1891{
1892 struct extent_buffer *leaf = path->nodes[0];
1893 int leaf_free_space = -1;
1894 int search_low_slot = 0;
1895 int ret;
1896 bool do_bin_search = true;
1897
1898 /*
1899 * If we are doing an insertion, the leaf has enough free space and the
1900 * destination slot for the key is not slot 0, then we can unlock our
1901 * write lock on the parent, and any other upper nodes, before doing the
1902 * binary search on the leaf (with search_for_key_slot()), allowing other
1903 * tasks to lock the parent and any other upper nodes.
1904 */
1905 if (ins_len > 0) {
1906 /*
1907 * Cache the leaf free space, since we will need it later and it
1908 * will not change until then.
1909 */
1910 leaf_free_space = btrfs_leaf_free_space(leaf);
1911
1912 /*
1913 * !path->locks[1] means we have a single node tree, the leaf is
1914 * the root of the tree.
1915 */
1916 if (path->locks[1] && leaf_free_space >= ins_len) {
1917 struct btrfs_disk_key first_key;
1918
1919 ASSERT(btrfs_header_nritems(leaf) > 0);
1920 btrfs_item_key(leaf, &first_key, 0);
1921
1922 /*
1923 * Doing the extra comparison with the first key is cheap,
1924 * taking into account that the first key is very likely
1925 * already in a cache line because it immediately follows
1926 * the extent buffer's header and we have recently accessed
1927 * the header's level field.
1928 */
1929 ret = btrfs_comp_keys(&first_key, key);
1930 if (ret < 0) {
1931 /*
1932 * The first key is smaller than the key we want
1933 * to insert, so we are safe to unlock all upper
1934 * nodes and we have to do the binary search.
1935 *
1936 * We do use btrfs_unlock_up_safe() and not
1937 * unlock_up() because the later does not unlock
1938 * nodes with a slot of 0 - we can safely unlock
1939 * any node even if its slot is 0 since in this
1940 * case the key does not end up at slot 0 of the
1941 * leaf and there's no need to split the leaf.
1942 */
1943 btrfs_unlock_up_safe(path, 1);
1944 search_low_slot = 1;
1945 } else {
1946 /*
1947 * The first key is >= then the key we want to
1948 * insert, so we can skip the binary search as
1949 * the target key will be at slot 0.
1950 *
1951 * We can not unlock upper nodes when the key is
1952 * less than the first key, because we will need
1953 * to update the key at slot 0 of the parent node
1954 * and possibly of other upper nodes too.
1955 * If the key matches the first key, then we can
1956 * unlock all the upper nodes, using
1957 * btrfs_unlock_up_safe() instead of unlock_up()
1958 * as stated above.
1959 */
1960 if (ret == 0)
1961 btrfs_unlock_up_safe(path, 1);
1962 /*
1963 * ret is already 0 or 1, matching the result of
1964 * a btrfs_bin_search() call, so there is no need
1965 * to adjust it.
1966 */
1967 do_bin_search = false;
1968 path->slots[0] = 0;
1969 }
1970 }
1971 }
1972
1973 if (do_bin_search) {
1974 ret = search_for_key_slot(leaf, search_low_slot, key,
1975 prev_cmp, &path->slots[0]);
1976 if (ret < 0)
1977 return ret;
1978 }
1979
1980 if (ins_len > 0) {
1981 /*
1982 * Item key already exists. In this case, if we are allowed to
1983 * insert the item (for example, in dir_item case, item key
1984 * collision is allowed), it will be merged with the original
1985 * item. Only the item size grows, no new btrfs item will be
1986 * added. If search_for_extension is not set, ins_len already
1987 * accounts the size btrfs_item, deduct it here so leaf space
1988 * check will be correct.
1989 */
1990 if (ret == 0 && !path->search_for_extension) {
1991 ASSERT(ins_len >= sizeof(struct btrfs_item));
1992 ins_len -= sizeof(struct btrfs_item);
1993 }
1994
1995 ASSERT(leaf_free_space >= 0);
1996
1997 if (leaf_free_space < ins_len) {
1998 int err;
1999
2000 err = split_leaf(trans, root, key, path, ins_len,
2001 (ret == 0));
2002 ASSERT(err <= 0);
2003 if (WARN_ON(err > 0))
2004 err = -EUCLEAN;
2005 if (err)
2006 ret = err;
2007 }
2008 }
2009
2010 return ret;
2011}
2012
2013/*
2014 * Look for a key in a tree and perform necessary modifications to preserve
2015 * tree invariants.
2016 *
2017 * @trans: Handle of transaction, used when modifying the tree
2018 * @p: Holds all btree nodes along the search path
2019 * @root: The root node of the tree
2020 * @key: The key we are looking for
2021 * @ins_len: Indicates purpose of search:
2022 * >0 for inserts it's size of item inserted (*)
2023 * <0 for deletions
2024 * 0 for plain searches, not modifying the tree
2025 *
2026 * (*) If size of item inserted doesn't include
2027 * sizeof(struct btrfs_item), then p->search_for_extension must
2028 * be set.
2029 * @cow: boolean should CoW operations be performed. Must always be 1
2030 * when modifying the tree.
2031 *
2032 * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2033 * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2034 *
2035 * If @key is found, 0 is returned and you can find the item in the leaf level
2036 * of the path (level 0)
2037 *
2038 * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2039 * points to the slot where it should be inserted
2040 *
2041 * If an error is encountered while searching the tree a negative error number
2042 * is returned
2043 */
2044int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2045 const struct btrfs_key *key, struct btrfs_path *p,
2046 int ins_len, int cow)
2047{
2048 struct btrfs_fs_info *fs_info;
2049 struct extent_buffer *b;
2050 int slot;
2051 int ret;
2052 int err;
2053 int level;
2054 int lowest_unlock = 1;
2055 /* everything at write_lock_level or lower must be write locked */
2056 int write_lock_level = 0;
2057 u8 lowest_level = 0;
2058 int min_write_lock_level;
2059 int prev_cmp;
2060
2061 if (!root)
2062 return -EINVAL;
2063
2064 fs_info = root->fs_info;
2065 might_sleep();
2066
2067 lowest_level = p->lowest_level;
2068 WARN_ON(lowest_level && ins_len > 0);
2069 WARN_ON(p->nodes[0] != NULL);
2070 BUG_ON(!cow && ins_len);
2071
2072 /*
2073 * For now only allow nowait for read only operations. There's no
2074 * strict reason why we can't, we just only need it for reads so it's
2075 * only implemented for reads.
2076 */
2077 ASSERT(!p->nowait || !cow);
2078
2079 if (ins_len < 0) {
2080 lowest_unlock = 2;
2081
2082 /* when we are removing items, we might have to go up to level
2083 * two as we update tree pointers Make sure we keep write
2084 * for those levels as well
2085 */
2086 write_lock_level = 2;
2087 } else if (ins_len > 0) {
2088 /*
2089 * for inserting items, make sure we have a write lock on
2090 * level 1 so we can update keys
2091 */
2092 write_lock_level = 1;
2093 }
2094
2095 if (!cow)
2096 write_lock_level = -1;
2097
2098 if (cow && (p->keep_locks || p->lowest_level))
2099 write_lock_level = BTRFS_MAX_LEVEL;
2100
2101 min_write_lock_level = write_lock_level;
2102
2103 if (p->need_commit_sem) {
2104 ASSERT(p->search_commit_root);
2105 if (p->nowait) {
2106 if (!down_read_trylock(&fs_info->commit_root_sem))
2107 return -EAGAIN;
2108 } else {
2109 down_read(&fs_info->commit_root_sem);
2110 }
2111 }
2112
2113again:
2114 prev_cmp = -1;
2115 b = btrfs_search_slot_get_root(root, p, write_lock_level);
2116 if (IS_ERR(b)) {
2117 ret = PTR_ERR(b);
2118 goto done;
2119 }
2120
2121 while (b) {
2122 int dec = 0;
2123
2124 level = btrfs_header_level(b);
2125
2126 if (cow) {
2127 bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2128
2129 /*
2130 * if we don't really need to cow this block
2131 * then we don't want to set the path blocking,
2132 * so we test it here
2133 */
2134 if (!should_cow_block(trans, root, b))
2135 goto cow_done;
2136
2137 /*
2138 * must have write locks on this node and the
2139 * parent
2140 */
2141 if (level > write_lock_level ||
2142 (level + 1 > write_lock_level &&
2143 level + 1 < BTRFS_MAX_LEVEL &&
2144 p->nodes[level + 1])) {
2145 write_lock_level = level + 1;
2146 btrfs_release_path(p);
2147 goto again;
2148 }
2149
2150 if (last_level)
2151 err = btrfs_cow_block(trans, root, b, NULL, 0,
2152 &b,
2153 BTRFS_NESTING_COW);
2154 else
2155 err = btrfs_cow_block(trans, root, b,
2156 p->nodes[level + 1],
2157 p->slots[level + 1], &b,
2158 BTRFS_NESTING_COW);
2159 if (err) {
2160 ret = err;
2161 goto done;
2162 }
2163 }
2164cow_done:
2165 p->nodes[level] = b;
2166
2167 /*
2168 * we have a lock on b and as long as we aren't changing
2169 * the tree, there is no way to for the items in b to change.
2170 * It is safe to drop the lock on our parent before we
2171 * go through the expensive btree search on b.
2172 *
2173 * If we're inserting or deleting (ins_len != 0), then we might
2174 * be changing slot zero, which may require changing the parent.
2175 * So, we can't drop the lock until after we know which slot
2176 * we're operating on.
2177 */
2178 if (!ins_len && !p->keep_locks) {
2179 int u = level + 1;
2180
2181 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2182 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2183 p->locks[u] = 0;
2184 }
2185 }
2186
2187 if (level == 0) {
2188 if (ins_len > 0)
2189 ASSERT(write_lock_level >= 1);
2190
2191 ret = search_leaf(trans, root, key, p, ins_len, prev_cmp);
2192 if (!p->search_for_split)
2193 unlock_up(p, level, lowest_unlock,
2194 min_write_lock_level, NULL);
2195 goto done;
2196 }
2197
2198 ret = search_for_key_slot(b, 0, key, prev_cmp, &slot);
2199 if (ret < 0)
2200 goto done;
2201 prev_cmp = ret;
2202
2203 if (ret && slot > 0) {
2204 dec = 1;
2205 slot--;
2206 }
2207 p->slots[level] = slot;
2208 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
2209 &write_lock_level);
2210 if (err == -EAGAIN)
2211 goto again;
2212 if (err) {
2213 ret = err;
2214 goto done;
2215 }
2216 b = p->nodes[level];
2217 slot = p->slots[level];
2218
2219 /*
2220 * Slot 0 is special, if we change the key we have to update
2221 * the parent pointer which means we must have a write lock on
2222 * the parent
2223 */
2224 if (slot == 0 && ins_len && write_lock_level < level + 1) {
2225 write_lock_level = level + 1;
2226 btrfs_release_path(p);
2227 goto again;
2228 }
2229
2230 unlock_up(p, level, lowest_unlock, min_write_lock_level,
2231 &write_lock_level);
2232
2233 if (level == lowest_level) {
2234 if (dec)
2235 p->slots[level]++;
2236 goto done;
2237 }
2238
2239 err = read_block_for_search(root, p, &b, slot, key);
2240 if (err == -EAGAIN && !p->nowait)
2241 goto again;
2242 if (err) {
2243 ret = err;
2244 goto done;
2245 }
2246
2247 if (!p->skip_locking) {
2248 level = btrfs_header_level(b);
2249
2250 btrfs_maybe_reset_lockdep_class(root, b);
2251
2252 if (level <= write_lock_level) {
2253 btrfs_tree_lock(b);
2254 p->locks[level] = BTRFS_WRITE_LOCK;
2255 } else {
2256 if (p->nowait) {
2257 if (!btrfs_try_tree_read_lock(b)) {
2258 free_extent_buffer(b);
2259 ret = -EAGAIN;
2260 goto done;
2261 }
2262 } else {
2263 btrfs_tree_read_lock(b);
2264 }
2265 p->locks[level] = BTRFS_READ_LOCK;
2266 }
2267 p->nodes[level] = b;
2268 }
2269 }
2270 ret = 1;
2271done:
2272 if (ret < 0 && !p->skip_release_on_error)
2273 btrfs_release_path(p);
2274
2275 if (p->need_commit_sem) {
2276 int ret2;
2277
2278 ret2 = finish_need_commit_sem_search(p);
2279 up_read(&fs_info->commit_root_sem);
2280 if (ret2)
2281 ret = ret2;
2282 }
2283
2284 return ret;
2285}
2286ALLOW_ERROR_INJECTION(btrfs_search_slot, ERRNO);
2287
2288/*
2289 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2290 * current state of the tree together with the operations recorded in the tree
2291 * modification log to search for the key in a previous version of this tree, as
2292 * denoted by the time_seq parameter.
2293 *
2294 * Naturally, there is no support for insert, delete or cow operations.
2295 *
2296 * The resulting path and return value will be set up as if we called
2297 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2298 */
2299int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2300 struct btrfs_path *p, u64 time_seq)
2301{
2302 struct btrfs_fs_info *fs_info = root->fs_info;
2303 struct extent_buffer *b;
2304 int slot;
2305 int ret;
2306 int err;
2307 int level;
2308 int lowest_unlock = 1;
2309 u8 lowest_level = 0;
2310
2311 lowest_level = p->lowest_level;
2312 WARN_ON(p->nodes[0] != NULL);
2313 ASSERT(!p->nowait);
2314
2315 if (p->search_commit_root) {
2316 BUG_ON(time_seq);
2317 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2318 }
2319
2320again:
2321 b = btrfs_get_old_root(root, time_seq);
2322 if (!b) {
2323 ret = -EIO;
2324 goto done;
2325 }
2326 level = btrfs_header_level(b);
2327 p->locks[level] = BTRFS_READ_LOCK;
2328
2329 while (b) {
2330 int dec = 0;
2331
2332 level = btrfs_header_level(b);
2333 p->nodes[level] = b;
2334
2335 /*
2336 * we have a lock on b and as long as we aren't changing
2337 * the tree, there is no way to for the items in b to change.
2338 * It is safe to drop the lock on our parent before we
2339 * go through the expensive btree search on b.
2340 */
2341 btrfs_unlock_up_safe(p, level + 1);
2342
2343 ret = btrfs_bin_search(b, 0, key, &slot);
2344 if (ret < 0)
2345 goto done;
2346
2347 if (level == 0) {
2348 p->slots[level] = slot;
2349 unlock_up(p, level, lowest_unlock, 0, NULL);
2350 goto done;
2351 }
2352
2353 if (ret && slot > 0) {
2354 dec = 1;
2355 slot--;
2356 }
2357 p->slots[level] = slot;
2358 unlock_up(p, level, lowest_unlock, 0, NULL);
2359
2360 if (level == lowest_level) {
2361 if (dec)
2362 p->slots[level]++;
2363 goto done;
2364 }
2365
2366 err = read_block_for_search(root, p, &b, slot, key);
2367 if (err == -EAGAIN && !p->nowait)
2368 goto again;
2369 if (err) {
2370 ret = err;
2371 goto done;
2372 }
2373
2374 level = btrfs_header_level(b);
2375 btrfs_tree_read_lock(b);
2376 b = btrfs_tree_mod_log_rewind(fs_info, b, time_seq);
2377 if (!b) {
2378 ret = -ENOMEM;
2379 goto done;
2380 }
2381 p->locks[level] = BTRFS_READ_LOCK;
2382 p->nodes[level] = b;
2383 }
2384 ret = 1;
2385done:
2386 if (ret < 0)
2387 btrfs_release_path(p);
2388
2389 return ret;
2390}
2391
2392/*
2393 * Search the tree again to find a leaf with smaller keys.
2394 * Returns 0 if it found something.
2395 * Returns 1 if there are no smaller keys.
2396 * Returns < 0 on error.
2397 *
2398 * This may release the path, and so you may lose any locks held at the
2399 * time you call it.
2400 */
2401static int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2402{
2403 struct btrfs_key key;
2404 struct btrfs_key orig_key;
2405 struct btrfs_disk_key found_key;
2406 int ret;
2407
2408 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
2409 orig_key = key;
2410
2411 if (key.offset > 0) {
2412 key.offset--;
2413 } else if (key.type > 0) {
2414 key.type--;
2415 key.offset = (u64)-1;
2416 } else if (key.objectid > 0) {
2417 key.objectid--;
2418 key.type = (u8)-1;
2419 key.offset = (u64)-1;
2420 } else {
2421 return 1;
2422 }
2423
2424 btrfs_release_path(path);
2425 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2426 if (ret <= 0)
2427 return ret;
2428
2429 /*
2430 * Previous key not found. Even if we were at slot 0 of the leaf we had
2431 * before releasing the path and calling btrfs_search_slot(), we now may
2432 * be in a slot pointing to the same original key - this can happen if
2433 * after we released the path, one of more items were moved from a
2434 * sibling leaf into the front of the leaf we had due to an insertion
2435 * (see push_leaf_right()).
2436 * If we hit this case and our slot is > 0 and just decrement the slot
2437 * so that the caller does not process the same key again, which may or
2438 * may not break the caller, depending on its logic.
2439 */
2440 if (path->slots[0] < btrfs_header_nritems(path->nodes[0])) {
2441 btrfs_item_key(path->nodes[0], &found_key, path->slots[0]);
2442 ret = btrfs_comp_keys(&found_key, &orig_key);
2443 if (ret == 0) {
2444 if (path->slots[0] > 0) {
2445 path->slots[0]--;
2446 return 0;
2447 }
2448 /*
2449 * At slot 0, same key as before, it means orig_key is
2450 * the lowest, leftmost, key in the tree. We're done.
2451 */
2452 return 1;
2453 }
2454 }
2455
2456 btrfs_item_key(path->nodes[0], &found_key, 0);
2457 ret = btrfs_comp_keys(&found_key, &key);
2458 /*
2459 * We might have had an item with the previous key in the tree right
2460 * before we released our path. And after we released our path, that
2461 * item might have been pushed to the first slot (0) of the leaf we
2462 * were holding due to a tree balance. Alternatively, an item with the
2463 * previous key can exist as the only element of a leaf (big fat item).
2464 * Therefore account for these 2 cases, so that our callers (like
2465 * btrfs_previous_item) don't miss an existing item with a key matching
2466 * the previous key we computed above.
2467 */
2468 if (ret <= 0)
2469 return 0;
2470 return 1;
2471}
2472
2473/*
2474 * helper to use instead of search slot if no exact match is needed but
2475 * instead the next or previous item should be returned.
2476 * When find_higher is true, the next higher item is returned, the next lower
2477 * otherwise.
2478 * When return_any and find_higher are both true, and no higher item is found,
2479 * return the next lower instead.
2480 * When return_any is true and find_higher is false, and no lower item is found,
2481 * return the next higher instead.
2482 * It returns 0 if any item is found, 1 if none is found (tree empty), and
2483 * < 0 on error
2484 */
2485int btrfs_search_slot_for_read(struct btrfs_root *root,
2486 const struct btrfs_key *key,
2487 struct btrfs_path *p, int find_higher,
2488 int return_any)
2489{
2490 int ret;
2491 struct extent_buffer *leaf;
2492
2493again:
2494 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
2495 if (ret <= 0)
2496 return ret;
2497 /*
2498 * a return value of 1 means the path is at the position where the
2499 * item should be inserted. Normally this is the next bigger item,
2500 * but in case the previous item is the last in a leaf, path points
2501 * to the first free slot in the previous leaf, i.e. at an invalid
2502 * item.
2503 */
2504 leaf = p->nodes[0];
2505
2506 if (find_higher) {
2507 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
2508 ret = btrfs_next_leaf(root, p);
2509 if (ret <= 0)
2510 return ret;
2511 if (!return_any)
2512 return 1;
2513 /*
2514 * no higher item found, return the next
2515 * lower instead
2516 */
2517 return_any = 0;
2518 find_higher = 0;
2519 btrfs_release_path(p);
2520 goto again;
2521 }
2522 } else {
2523 if (p->slots[0] == 0) {
2524 ret = btrfs_prev_leaf(root, p);
2525 if (ret < 0)
2526 return ret;
2527 if (!ret) {
2528 leaf = p->nodes[0];
2529 if (p->slots[0] == btrfs_header_nritems(leaf))
2530 p->slots[0]--;
2531 return 0;
2532 }
2533 if (!return_any)
2534 return 1;
2535 /*
2536 * no lower item found, return the next
2537 * higher instead
2538 */
2539 return_any = 0;
2540 find_higher = 1;
2541 btrfs_release_path(p);
2542 goto again;
2543 } else {
2544 --p->slots[0];
2545 }
2546 }
2547 return 0;
2548}
2549
2550/*
2551 * Execute search and call btrfs_previous_item to traverse backwards if the item
2552 * was not found.
2553 *
2554 * Return 0 if found, 1 if not found and < 0 if error.
2555 */
2556int btrfs_search_backwards(struct btrfs_root *root, struct btrfs_key *key,
2557 struct btrfs_path *path)
2558{
2559 int ret;
2560
2561 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
2562 if (ret > 0)
2563 ret = btrfs_previous_item(root, path, key->objectid, key->type);
2564
2565 if (ret == 0)
2566 btrfs_item_key_to_cpu(path->nodes[0], key, path->slots[0]);
2567
2568 return ret;
2569}
2570
2571/*
2572 * Search for a valid slot for the given path.
2573 *
2574 * @root: The root node of the tree.
2575 * @key: Will contain a valid item if found.
2576 * @path: The starting point to validate the slot.
2577 *
2578 * Return: 0 if the item is valid
2579 * 1 if not found
2580 * <0 if error.
2581 */
2582int btrfs_get_next_valid_item(struct btrfs_root *root, struct btrfs_key *key,
2583 struct btrfs_path *path)
2584{
2585 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2586 int ret;
2587
2588 ret = btrfs_next_leaf(root, path);
2589 if (ret)
2590 return ret;
2591 }
2592
2593 btrfs_item_key_to_cpu(path->nodes[0], key, path->slots[0]);
2594 return 0;
2595}
2596
2597/*
2598 * adjust the pointers going up the tree, starting at level
2599 * making sure the right key of each node is points to 'key'.
2600 * This is used after shifting pointers to the left, so it stops
2601 * fixing up pointers when a given leaf/node is not in slot 0 of the
2602 * higher levels
2603 *
2604 */
2605static void fixup_low_keys(struct btrfs_trans_handle *trans,
2606 const struct btrfs_path *path,
2607 const struct btrfs_disk_key *key, int level)
2608{
2609 int i;
2610 struct extent_buffer *t;
2611 int ret;
2612
2613 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2614 int tslot = path->slots[i];
2615
2616 if (!path->nodes[i])
2617 break;
2618 t = path->nodes[i];
2619 ret = btrfs_tree_mod_log_insert_key(t, tslot,
2620 BTRFS_MOD_LOG_KEY_REPLACE);
2621 BUG_ON(ret < 0);
2622 btrfs_set_node_key(t, key, tslot);
2623 btrfs_mark_buffer_dirty(trans, path->nodes[i]);
2624 if (tslot != 0)
2625 break;
2626 }
2627}
2628
2629/*
2630 * update item key.
2631 *
2632 * This function isn't completely safe. It's the caller's responsibility
2633 * that the new key won't break the order
2634 */
2635void btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
2636 const struct btrfs_path *path,
2637 const struct btrfs_key *new_key)
2638{
2639 struct btrfs_fs_info *fs_info = trans->fs_info;
2640 struct btrfs_disk_key disk_key;
2641 struct extent_buffer *eb;
2642 int slot;
2643
2644 eb = path->nodes[0];
2645 slot = path->slots[0];
2646 if (slot > 0) {
2647 btrfs_item_key(eb, &disk_key, slot - 1);
2648 if (unlikely(btrfs_comp_keys(&disk_key, new_key) >= 0)) {
2649 btrfs_print_leaf(eb);
2650 btrfs_crit(fs_info,
2651 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2652 slot, btrfs_disk_key_objectid(&disk_key),
2653 btrfs_disk_key_type(&disk_key),
2654 btrfs_disk_key_offset(&disk_key),
2655 new_key->objectid, new_key->type,
2656 new_key->offset);
2657 BUG();
2658 }
2659 }
2660 if (slot < btrfs_header_nritems(eb) - 1) {
2661 btrfs_item_key(eb, &disk_key, slot + 1);
2662 if (unlikely(btrfs_comp_keys(&disk_key, new_key) <= 0)) {
2663 btrfs_print_leaf(eb);
2664 btrfs_crit(fs_info,
2665 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2666 slot, btrfs_disk_key_objectid(&disk_key),
2667 btrfs_disk_key_type(&disk_key),
2668 btrfs_disk_key_offset(&disk_key),
2669 new_key->objectid, new_key->type,
2670 new_key->offset);
2671 BUG();
2672 }
2673 }
2674
2675 btrfs_cpu_key_to_disk(&disk_key, new_key);
2676 btrfs_set_item_key(eb, &disk_key, slot);
2677 btrfs_mark_buffer_dirty(trans, eb);
2678 if (slot == 0)
2679 fixup_low_keys(trans, path, &disk_key, 1);
2680}
2681
2682/*
2683 * Check key order of two sibling extent buffers.
2684 *
2685 * Return true if something is wrong.
2686 * Return false if everything is fine.
2687 *
2688 * Tree-checker only works inside one tree block, thus the following
2689 * corruption can not be detected by tree-checker:
2690 *
2691 * Leaf @left | Leaf @right
2692 * --------------------------------------------------------------
2693 * | 1 | 2 | 3 | 4 | 5 | f6 | | 7 | 8 |
2694 *
2695 * Key f6 in leaf @left itself is valid, but not valid when the next
2696 * key in leaf @right is 7.
2697 * This can only be checked at tree block merge time.
2698 * And since tree checker has ensured all key order in each tree block
2699 * is correct, we only need to bother the last key of @left and the first
2700 * key of @right.
2701 */
2702static bool check_sibling_keys(const struct extent_buffer *left,
2703 const struct extent_buffer *right)
2704{
2705 struct btrfs_key left_last;
2706 struct btrfs_key right_first;
2707 int level = btrfs_header_level(left);
2708 int nr_left = btrfs_header_nritems(left);
2709 int nr_right = btrfs_header_nritems(right);
2710
2711 /* No key to check in one of the tree blocks */
2712 if (!nr_left || !nr_right)
2713 return false;
2714
2715 if (level) {
2716 btrfs_node_key_to_cpu(left, &left_last, nr_left - 1);
2717 btrfs_node_key_to_cpu(right, &right_first, 0);
2718 } else {
2719 btrfs_item_key_to_cpu(left, &left_last, nr_left - 1);
2720 btrfs_item_key_to_cpu(right, &right_first, 0);
2721 }
2722
2723 if (unlikely(btrfs_comp_cpu_keys(&left_last, &right_first) >= 0)) {
2724 btrfs_crit(left->fs_info, "left extent buffer:");
2725 btrfs_print_tree(left, false);
2726 btrfs_crit(left->fs_info, "right extent buffer:");
2727 btrfs_print_tree(right, false);
2728 btrfs_crit(left->fs_info,
2729"bad key order, sibling blocks, left last (%llu %u %llu) right first (%llu %u %llu)",
2730 left_last.objectid, left_last.type,
2731 left_last.offset, right_first.objectid,
2732 right_first.type, right_first.offset);
2733 return true;
2734 }
2735 return false;
2736}
2737
2738/*
2739 * try to push data from one node into the next node left in the
2740 * tree.
2741 *
2742 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2743 * error, and > 0 if there was no room in the left hand block.
2744 */
2745static int push_node_left(struct btrfs_trans_handle *trans,
2746 struct extent_buffer *dst,
2747 struct extent_buffer *src, int empty)
2748{
2749 struct btrfs_fs_info *fs_info = trans->fs_info;
2750 int push_items = 0;
2751 int src_nritems;
2752 int dst_nritems;
2753 int ret = 0;
2754
2755 src_nritems = btrfs_header_nritems(src);
2756 dst_nritems = btrfs_header_nritems(dst);
2757 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2758 WARN_ON(btrfs_header_generation(src) != trans->transid);
2759 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2760
2761 if (!empty && src_nritems <= 8)
2762 return 1;
2763
2764 if (push_items <= 0)
2765 return 1;
2766
2767 if (empty) {
2768 push_items = min(src_nritems, push_items);
2769 if (push_items < src_nritems) {
2770 /* leave at least 8 pointers in the node if
2771 * we aren't going to empty it
2772 */
2773 if (src_nritems - push_items < 8) {
2774 if (push_items <= 8)
2775 return 1;
2776 push_items -= 8;
2777 }
2778 }
2779 } else
2780 push_items = min(src_nritems - 8, push_items);
2781
2782 /* dst is the left eb, src is the middle eb */
2783 if (check_sibling_keys(dst, src)) {
2784 ret = -EUCLEAN;
2785 btrfs_abort_transaction(trans, ret);
2786 return ret;
2787 }
2788 ret = btrfs_tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
2789 if (ret) {
2790 btrfs_abort_transaction(trans, ret);
2791 return ret;
2792 }
2793 copy_extent_buffer(dst, src,
2794 btrfs_node_key_ptr_offset(dst, dst_nritems),
2795 btrfs_node_key_ptr_offset(src, 0),
2796 push_items * sizeof(struct btrfs_key_ptr));
2797
2798 if (push_items < src_nritems) {
2799 /*
2800 * btrfs_tree_mod_log_eb_copy handles logging the move, so we
2801 * don't need to do an explicit tree mod log operation for it.
2802 */
2803 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(src, 0),
2804 btrfs_node_key_ptr_offset(src, push_items),
2805 (src_nritems - push_items) *
2806 sizeof(struct btrfs_key_ptr));
2807 }
2808 btrfs_set_header_nritems(src, src_nritems - push_items);
2809 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2810 btrfs_mark_buffer_dirty(trans, src);
2811 btrfs_mark_buffer_dirty(trans, dst);
2812
2813 return ret;
2814}
2815
2816/*
2817 * try to push data from one node into the next node right in the
2818 * tree.
2819 *
2820 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2821 * error, and > 0 if there was no room in the right hand block.
2822 *
2823 * this will only push up to 1/2 the contents of the left node over
2824 */
2825static int balance_node_right(struct btrfs_trans_handle *trans,
2826 struct extent_buffer *dst,
2827 struct extent_buffer *src)
2828{
2829 struct btrfs_fs_info *fs_info = trans->fs_info;
2830 int push_items = 0;
2831 int max_push;
2832 int src_nritems;
2833 int dst_nritems;
2834 int ret = 0;
2835
2836 WARN_ON(btrfs_header_generation(src) != trans->transid);
2837 WARN_ON(btrfs_header_generation(dst) != trans->transid);
2838
2839 src_nritems = btrfs_header_nritems(src);
2840 dst_nritems = btrfs_header_nritems(dst);
2841 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2842 if (push_items <= 0)
2843 return 1;
2844
2845 if (src_nritems < 4)
2846 return 1;
2847
2848 max_push = src_nritems / 2 + 1;
2849 /* don't try to empty the node */
2850 if (max_push >= src_nritems)
2851 return 1;
2852
2853 if (max_push < push_items)
2854 push_items = max_push;
2855
2856 /* dst is the right eb, src is the middle eb */
2857 if (check_sibling_keys(src, dst)) {
2858 ret = -EUCLEAN;
2859 btrfs_abort_transaction(trans, ret);
2860 return ret;
2861 }
2862
2863 /*
2864 * btrfs_tree_mod_log_eb_copy handles logging the move, so we don't
2865 * need to do an explicit tree mod log operation for it.
2866 */
2867 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(dst, push_items),
2868 btrfs_node_key_ptr_offset(dst, 0),
2869 (dst_nritems) *
2870 sizeof(struct btrfs_key_ptr));
2871
2872 ret = btrfs_tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
2873 push_items);
2874 if (ret) {
2875 btrfs_abort_transaction(trans, ret);
2876 return ret;
2877 }
2878 copy_extent_buffer(dst, src,
2879 btrfs_node_key_ptr_offset(dst, 0),
2880 btrfs_node_key_ptr_offset(src, src_nritems - push_items),
2881 push_items * sizeof(struct btrfs_key_ptr));
2882
2883 btrfs_set_header_nritems(src, src_nritems - push_items);
2884 btrfs_set_header_nritems(dst, dst_nritems + push_items);
2885
2886 btrfs_mark_buffer_dirty(trans, src);
2887 btrfs_mark_buffer_dirty(trans, dst);
2888
2889 return ret;
2890}
2891
2892/*
2893 * helper function to insert a new root level in the tree.
2894 * A new node is allocated, and a single item is inserted to
2895 * point to the existing root
2896 *
2897 * returns zero on success or < 0 on failure.
2898 */
2899static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2900 struct btrfs_root *root,
2901 struct btrfs_path *path, int level)
2902{
2903 u64 lower_gen;
2904 struct extent_buffer *lower;
2905 struct extent_buffer *c;
2906 struct extent_buffer *old;
2907 struct btrfs_disk_key lower_key;
2908 int ret;
2909
2910 BUG_ON(path->nodes[level]);
2911 BUG_ON(path->nodes[level-1] != root->node);
2912
2913 lower = path->nodes[level-1];
2914 if (level == 1)
2915 btrfs_item_key(lower, &lower_key, 0);
2916 else
2917 btrfs_node_key(lower, &lower_key, 0);
2918
2919 c = btrfs_alloc_tree_block(trans, root, 0, btrfs_root_id(root),
2920 &lower_key, level, root->node->start, 0,
2921 0, BTRFS_NESTING_NEW_ROOT);
2922 if (IS_ERR(c))
2923 return PTR_ERR(c);
2924
2925 root_add_used_bytes(root);
2926
2927 btrfs_set_header_nritems(c, 1);
2928 btrfs_set_node_key(c, &lower_key, 0);
2929 btrfs_set_node_blockptr(c, 0, lower->start);
2930 lower_gen = btrfs_header_generation(lower);
2931 WARN_ON(lower_gen != trans->transid);
2932
2933 btrfs_set_node_ptr_generation(c, 0, lower_gen);
2934
2935 btrfs_mark_buffer_dirty(trans, c);
2936
2937 old = root->node;
2938 ret = btrfs_tree_mod_log_insert_root(root->node, c, false);
2939 if (ret < 0) {
2940 int ret2;
2941
2942 ret2 = btrfs_free_tree_block(trans, btrfs_root_id(root), c, 0, 1);
2943 if (ret2 < 0)
2944 btrfs_abort_transaction(trans, ret2);
2945 btrfs_tree_unlock(c);
2946 free_extent_buffer(c);
2947 return ret;
2948 }
2949 rcu_assign_pointer(root->node, c);
2950
2951 /* the super has an extra ref to root->node */
2952 free_extent_buffer(old);
2953
2954 add_root_to_dirty_list(root);
2955 atomic_inc(&c->refs);
2956 path->nodes[level] = c;
2957 path->locks[level] = BTRFS_WRITE_LOCK;
2958 path->slots[level] = 0;
2959 return 0;
2960}
2961
2962/*
2963 * worker function to insert a single pointer in a node.
2964 * the node should have enough room for the pointer already
2965 *
2966 * slot and level indicate where you want the key to go, and
2967 * blocknr is the block the key points to.
2968 */
2969static int insert_ptr(struct btrfs_trans_handle *trans,
2970 const struct btrfs_path *path,
2971 const struct btrfs_disk_key *key, u64 bytenr,
2972 int slot, int level)
2973{
2974 struct extent_buffer *lower;
2975 int nritems;
2976 int ret;
2977
2978 BUG_ON(!path->nodes[level]);
2979 btrfs_assert_tree_write_locked(path->nodes[level]);
2980 lower = path->nodes[level];
2981 nritems = btrfs_header_nritems(lower);
2982 BUG_ON(slot > nritems);
2983 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
2984 if (slot != nritems) {
2985 if (level) {
2986 ret = btrfs_tree_mod_log_insert_move(lower, slot + 1,
2987 slot, nritems - slot);
2988 if (ret < 0) {
2989 btrfs_abort_transaction(trans, ret);
2990 return ret;
2991 }
2992 }
2993 memmove_extent_buffer(lower,
2994 btrfs_node_key_ptr_offset(lower, slot + 1),
2995 btrfs_node_key_ptr_offset(lower, slot),
2996 (nritems - slot) * sizeof(struct btrfs_key_ptr));
2997 }
2998 if (level) {
2999 ret = btrfs_tree_mod_log_insert_key(lower, slot,
3000 BTRFS_MOD_LOG_KEY_ADD);
3001 if (ret < 0) {
3002 btrfs_abort_transaction(trans, ret);
3003 return ret;
3004 }
3005 }
3006 btrfs_set_node_key(lower, key, slot);
3007 btrfs_set_node_blockptr(lower, slot, bytenr);
3008 WARN_ON(trans->transid == 0);
3009 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3010 btrfs_set_header_nritems(lower, nritems + 1);
3011 btrfs_mark_buffer_dirty(trans, lower);
3012
3013 return 0;
3014}
3015
3016/*
3017 * split the node at the specified level in path in two.
3018 * The path is corrected to point to the appropriate node after the split
3019 *
3020 * Before splitting this tries to make some room in the node by pushing
3021 * left and right, if either one works, it returns right away.
3022 *
3023 * returns 0 on success and < 0 on failure
3024 */
3025static noinline int split_node(struct btrfs_trans_handle *trans,
3026 struct btrfs_root *root,
3027 struct btrfs_path *path, int level)
3028{
3029 struct btrfs_fs_info *fs_info = root->fs_info;
3030 struct extent_buffer *c;
3031 struct extent_buffer *split;
3032 struct btrfs_disk_key disk_key;
3033 int mid;
3034 int ret;
3035 u32 c_nritems;
3036
3037 c = path->nodes[level];
3038 WARN_ON(btrfs_header_generation(c) != trans->transid);
3039 if (c == root->node) {
3040 /*
3041 * trying to split the root, lets make a new one
3042 *
3043 * tree mod log: We don't log_removal old root in
3044 * insert_new_root, because that root buffer will be kept as a
3045 * normal node. We are going to log removal of half of the
3046 * elements below with btrfs_tree_mod_log_eb_copy(). We're
3047 * holding a tree lock on the buffer, which is why we cannot
3048 * race with other tree_mod_log users.
3049 */
3050 ret = insert_new_root(trans, root, path, level + 1);
3051 if (ret)
3052 return ret;
3053 } else {
3054 ret = push_nodes_for_insert(trans, root, path, level);
3055 c = path->nodes[level];
3056 if (!ret && btrfs_header_nritems(c) <
3057 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3058 return 0;
3059 if (ret < 0)
3060 return ret;
3061 }
3062
3063 c_nritems = btrfs_header_nritems(c);
3064 mid = (c_nritems + 1) / 2;
3065 btrfs_node_key(c, &disk_key, mid);
3066
3067 split = btrfs_alloc_tree_block(trans, root, 0, btrfs_root_id(root),
3068 &disk_key, level, c->start, 0,
3069 0, BTRFS_NESTING_SPLIT);
3070 if (IS_ERR(split))
3071 return PTR_ERR(split);
3072
3073 root_add_used_bytes(root);
3074 ASSERT(btrfs_header_level(c) == level);
3075
3076 ret = btrfs_tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
3077 if (ret) {
3078 btrfs_tree_unlock(split);
3079 free_extent_buffer(split);
3080 btrfs_abort_transaction(trans, ret);
3081 return ret;
3082 }
3083 copy_extent_buffer(split, c,
3084 btrfs_node_key_ptr_offset(split, 0),
3085 btrfs_node_key_ptr_offset(c, mid),
3086 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3087 btrfs_set_header_nritems(split, c_nritems - mid);
3088 btrfs_set_header_nritems(c, mid);
3089
3090 btrfs_mark_buffer_dirty(trans, c);
3091 btrfs_mark_buffer_dirty(trans, split);
3092
3093 ret = insert_ptr(trans, path, &disk_key, split->start,
3094 path->slots[level + 1] + 1, level + 1);
3095 if (ret < 0) {
3096 btrfs_tree_unlock(split);
3097 free_extent_buffer(split);
3098 return ret;
3099 }
3100
3101 if (path->slots[level] >= mid) {
3102 path->slots[level] -= mid;
3103 btrfs_tree_unlock(c);
3104 free_extent_buffer(c);
3105 path->nodes[level] = split;
3106 path->slots[level + 1] += 1;
3107 } else {
3108 btrfs_tree_unlock(split);
3109 free_extent_buffer(split);
3110 }
3111 return 0;
3112}
3113
3114/*
3115 * how many bytes are required to store the items in a leaf. start
3116 * and nr indicate which items in the leaf to check. This totals up the
3117 * space used both by the item structs and the item data
3118 */
3119static int leaf_space_used(const struct extent_buffer *l, int start, int nr)
3120{
3121 int data_len;
3122 int nritems = btrfs_header_nritems(l);
3123 int end = min(nritems, start + nr) - 1;
3124
3125 if (!nr)
3126 return 0;
3127 data_len = btrfs_item_offset(l, start) + btrfs_item_size(l, start);
3128 data_len = data_len - btrfs_item_offset(l, end);
3129 data_len += sizeof(struct btrfs_item) * nr;
3130 WARN_ON(data_len < 0);
3131 return data_len;
3132}
3133
3134/*
3135 * The space between the end of the leaf items and
3136 * the start of the leaf data. IOW, how much room
3137 * the leaf has left for both items and data
3138 */
3139int btrfs_leaf_free_space(const struct extent_buffer *leaf)
3140{
3141 struct btrfs_fs_info *fs_info = leaf->fs_info;
3142 int nritems = btrfs_header_nritems(leaf);
3143 int ret;
3144
3145 ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3146 if (ret < 0) {
3147 btrfs_crit(fs_info,
3148 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3149 ret,
3150 (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3151 leaf_space_used(leaf, 0, nritems), nritems);
3152 }
3153 return ret;
3154}
3155
3156/*
3157 * min slot controls the lowest index we're willing to push to the
3158 * right. We'll push up to and including min_slot, but no lower
3159 */
3160static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
3161 struct btrfs_path *path,
3162 int data_size, int empty,
3163 struct extent_buffer *right,
3164 int free_space, u32 left_nritems,
3165 u32 min_slot)
3166{
3167 struct btrfs_fs_info *fs_info = right->fs_info;
3168 struct extent_buffer *left = path->nodes[0];
3169 struct extent_buffer *upper = path->nodes[1];
3170 struct btrfs_map_token token;
3171 struct btrfs_disk_key disk_key;
3172 int slot;
3173 u32 i;
3174 int push_space = 0;
3175 int push_items = 0;
3176 u32 nr;
3177 u32 right_nritems;
3178 u32 data_end;
3179 u32 this_item_size;
3180
3181 if (empty)
3182 nr = 0;
3183 else
3184 nr = max_t(u32, 1, min_slot);
3185
3186 if (path->slots[0] >= left_nritems)
3187 push_space += data_size;
3188
3189 slot = path->slots[1];
3190 i = left_nritems - 1;
3191 while (i >= nr) {
3192 if (!empty && push_items > 0) {
3193 if (path->slots[0] > i)
3194 break;
3195 if (path->slots[0] == i) {
3196 int space = btrfs_leaf_free_space(left);
3197
3198 if (space + push_space * 2 > free_space)
3199 break;
3200 }
3201 }
3202
3203 if (path->slots[0] == i)
3204 push_space += data_size;
3205
3206 this_item_size = btrfs_item_size(left, i);
3207 if (this_item_size + sizeof(struct btrfs_item) +
3208 push_space > free_space)
3209 break;
3210
3211 push_items++;
3212 push_space += this_item_size + sizeof(struct btrfs_item);
3213 if (i == 0)
3214 break;
3215 i--;
3216 }
3217
3218 if (push_items == 0)
3219 goto out_unlock;
3220
3221 WARN_ON(!empty && push_items == left_nritems);
3222
3223 /* push left to right */
3224 right_nritems = btrfs_header_nritems(right);
3225
3226 push_space = btrfs_item_data_end(left, left_nritems - push_items);
3227 push_space -= leaf_data_end(left);
3228
3229 /* make room in the right data area */
3230 data_end = leaf_data_end(right);
3231 memmove_leaf_data(right, data_end - push_space, data_end,
3232 BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3233
3234 /* copy from the left data area */
3235 copy_leaf_data(right, left, BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3236 leaf_data_end(left), push_space);
3237
3238 memmove_leaf_items(right, push_items, 0, right_nritems);
3239
3240 /* copy the items from left to right */
3241 copy_leaf_items(right, left, 0, left_nritems - push_items, push_items);
3242
3243 /* update the item pointers */
3244 btrfs_init_map_token(&token, right);
3245 right_nritems += push_items;
3246 btrfs_set_header_nritems(right, right_nritems);
3247 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3248 for (i = 0; i < right_nritems; i++) {
3249 push_space -= btrfs_token_item_size(&token, i);
3250 btrfs_set_token_item_offset(&token, i, push_space);
3251 }
3252
3253 left_nritems -= push_items;
3254 btrfs_set_header_nritems(left, left_nritems);
3255
3256 if (left_nritems)
3257 btrfs_mark_buffer_dirty(trans, left);
3258 else
3259 btrfs_clear_buffer_dirty(trans, left);
3260
3261 btrfs_mark_buffer_dirty(trans, right);
3262
3263 btrfs_item_key(right, &disk_key, 0);
3264 btrfs_set_node_key(upper, &disk_key, slot + 1);
3265 btrfs_mark_buffer_dirty(trans, upper);
3266
3267 /* then fixup the leaf pointer in the path */
3268 if (path->slots[0] >= left_nritems) {
3269 path->slots[0] -= left_nritems;
3270 if (btrfs_header_nritems(path->nodes[0]) == 0)
3271 btrfs_clear_buffer_dirty(trans, path->nodes[0]);
3272 btrfs_tree_unlock(path->nodes[0]);
3273 free_extent_buffer(path->nodes[0]);
3274 path->nodes[0] = right;
3275 path->slots[1] += 1;
3276 } else {
3277 btrfs_tree_unlock(right);
3278 free_extent_buffer(right);
3279 }
3280 return 0;
3281
3282out_unlock:
3283 btrfs_tree_unlock(right);
3284 free_extent_buffer(right);
3285 return 1;
3286}
3287
3288/*
3289 * push some data in the path leaf to the right, trying to free up at
3290 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3291 *
3292 * returns 1 if the push failed because the other node didn't have enough
3293 * room, 0 if everything worked out and < 0 if there were major errors.
3294 *
3295 * this will push starting from min_slot to the end of the leaf. It won't
3296 * push any slot lower than min_slot
3297 */
3298static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3299 *root, struct btrfs_path *path,
3300 int min_data_size, int data_size,
3301 int empty, u32 min_slot)
3302{
3303 struct extent_buffer *left = path->nodes[0];
3304 struct extent_buffer *right;
3305 struct extent_buffer *upper;
3306 int slot;
3307 int free_space;
3308 u32 left_nritems;
3309 int ret;
3310
3311 if (!path->nodes[1])
3312 return 1;
3313
3314 slot = path->slots[1];
3315 upper = path->nodes[1];
3316 if (slot >= btrfs_header_nritems(upper) - 1)
3317 return 1;
3318
3319 btrfs_assert_tree_write_locked(path->nodes[1]);
3320
3321 right = btrfs_read_node_slot(upper, slot + 1);
3322 if (IS_ERR(right))
3323 return PTR_ERR(right);
3324
3325 btrfs_tree_lock_nested(right, BTRFS_NESTING_RIGHT);
3326
3327 free_space = btrfs_leaf_free_space(right);
3328 if (free_space < data_size)
3329 goto out_unlock;
3330
3331 ret = btrfs_cow_block(trans, root, right, upper,
3332 slot + 1, &right, BTRFS_NESTING_RIGHT_COW);
3333 if (ret)
3334 goto out_unlock;
3335
3336 left_nritems = btrfs_header_nritems(left);
3337 if (left_nritems == 0)
3338 goto out_unlock;
3339
3340 if (check_sibling_keys(left, right)) {
3341 ret = -EUCLEAN;
3342 btrfs_abort_transaction(trans, ret);
3343 btrfs_tree_unlock(right);
3344 free_extent_buffer(right);
3345 return ret;
3346 }
3347 if (path->slots[0] == left_nritems && !empty) {
3348 /* Key greater than all keys in the leaf, right neighbor has
3349 * enough room for it and we're not emptying our leaf to delete
3350 * it, therefore use right neighbor to insert the new item and
3351 * no need to touch/dirty our left leaf. */
3352 btrfs_tree_unlock(left);
3353 free_extent_buffer(left);
3354 path->nodes[0] = right;
3355 path->slots[0] = 0;
3356 path->slots[1]++;
3357 return 0;
3358 }
3359
3360 return __push_leaf_right(trans, path, min_data_size, empty, right,
3361 free_space, left_nritems, min_slot);
3362out_unlock:
3363 btrfs_tree_unlock(right);
3364 free_extent_buffer(right);
3365 return 1;
3366}
3367
3368/*
3369 * push some data in the path leaf to the left, trying to free up at
3370 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3371 *
3372 * max_slot can put a limit on how far into the leaf we'll push items. The
3373 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3374 * items
3375 */
3376static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
3377 struct btrfs_path *path, int data_size,
3378 int empty, struct extent_buffer *left,
3379 int free_space, u32 right_nritems,
3380 u32 max_slot)
3381{
3382 struct btrfs_fs_info *fs_info = left->fs_info;
3383 struct btrfs_disk_key disk_key;
3384 struct extent_buffer *right = path->nodes[0];
3385 int i;
3386 int push_space = 0;
3387 int push_items = 0;
3388 u32 old_left_nritems;
3389 u32 nr;
3390 int ret = 0;
3391 u32 this_item_size;
3392 u32 old_left_item_size;
3393 struct btrfs_map_token token;
3394
3395 if (empty)
3396 nr = min(right_nritems, max_slot);
3397 else
3398 nr = min(right_nritems - 1, max_slot);
3399
3400 for (i = 0; i < nr; i++) {
3401 if (!empty && push_items > 0) {
3402 if (path->slots[0] < i)
3403 break;
3404 if (path->slots[0] == i) {
3405 int space = btrfs_leaf_free_space(right);
3406
3407 if (space + push_space * 2 > free_space)
3408 break;
3409 }
3410 }
3411
3412 if (path->slots[0] == i)
3413 push_space += data_size;
3414
3415 this_item_size = btrfs_item_size(right, i);
3416 if (this_item_size + sizeof(struct btrfs_item) + push_space >
3417 free_space)
3418 break;
3419
3420 push_items++;
3421 push_space += this_item_size + sizeof(struct btrfs_item);
3422 }
3423
3424 if (push_items == 0) {
3425 ret = 1;
3426 goto out;
3427 }
3428 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3429
3430 /* push data from right to left */
3431 copy_leaf_items(left, right, btrfs_header_nritems(left), 0, push_items);
3432
3433 push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3434 btrfs_item_offset(right, push_items - 1);
3435
3436 copy_leaf_data(left, right, leaf_data_end(left) - push_space,
3437 btrfs_item_offset(right, push_items - 1), push_space);
3438 old_left_nritems = btrfs_header_nritems(left);
3439 BUG_ON(old_left_nritems <= 0);
3440
3441 btrfs_init_map_token(&token, left);
3442 old_left_item_size = btrfs_item_offset(left, old_left_nritems - 1);
3443 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3444 u32 ioff;
3445
3446 ioff = btrfs_token_item_offset(&token, i);
3447 btrfs_set_token_item_offset(&token, i,
3448 ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
3449 }
3450 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3451
3452 /* fixup right node */
3453 if (push_items > right_nritems)
3454 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3455 right_nritems);
3456
3457 if (push_items < right_nritems) {
3458 push_space = btrfs_item_offset(right, push_items - 1) -
3459 leaf_data_end(right);
3460 memmove_leaf_data(right,
3461 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3462 leaf_data_end(right), push_space);
3463
3464 memmove_leaf_items(right, 0, push_items,
3465 btrfs_header_nritems(right) - push_items);
3466 }
3467
3468 btrfs_init_map_token(&token, right);
3469 right_nritems -= push_items;
3470 btrfs_set_header_nritems(right, right_nritems);
3471 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3472 for (i = 0; i < right_nritems; i++) {
3473 push_space = push_space - btrfs_token_item_size(&token, i);
3474 btrfs_set_token_item_offset(&token, i, push_space);
3475 }
3476
3477 btrfs_mark_buffer_dirty(trans, left);
3478 if (right_nritems)
3479 btrfs_mark_buffer_dirty(trans, right);
3480 else
3481 btrfs_clear_buffer_dirty(trans, right);
3482
3483 btrfs_item_key(right, &disk_key, 0);
3484 fixup_low_keys(trans, path, &disk_key, 1);
3485
3486 /* then fixup the leaf pointer in the path */
3487 if (path->slots[0] < push_items) {
3488 path->slots[0] += old_left_nritems;
3489 btrfs_tree_unlock(path->nodes[0]);
3490 free_extent_buffer(path->nodes[0]);
3491 path->nodes[0] = left;
3492 path->slots[1] -= 1;
3493 } else {
3494 btrfs_tree_unlock(left);
3495 free_extent_buffer(left);
3496 path->slots[0] -= push_items;
3497 }
3498 BUG_ON(path->slots[0] < 0);
3499 return ret;
3500out:
3501 btrfs_tree_unlock(left);
3502 free_extent_buffer(left);
3503 return ret;
3504}
3505
3506/*
3507 * push some data in the path leaf to the left, trying to free up at
3508 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3509 *
3510 * max_slot can put a limit on how far into the leaf we'll push items. The
3511 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
3512 * items
3513 */
3514static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3515 *root, struct btrfs_path *path, int min_data_size,
3516 int data_size, int empty, u32 max_slot)
3517{
3518 struct extent_buffer *right = path->nodes[0];
3519 struct extent_buffer *left;
3520 int slot;
3521 int free_space;
3522 u32 right_nritems;
3523 int ret = 0;
3524
3525 slot = path->slots[1];
3526 if (slot == 0)
3527 return 1;
3528 if (!path->nodes[1])
3529 return 1;
3530
3531 right_nritems = btrfs_header_nritems(right);
3532 if (right_nritems == 0)
3533 return 1;
3534
3535 btrfs_assert_tree_write_locked(path->nodes[1]);
3536
3537 left = btrfs_read_node_slot(path->nodes[1], slot - 1);
3538 if (IS_ERR(left))
3539 return PTR_ERR(left);
3540
3541 btrfs_tree_lock_nested(left, BTRFS_NESTING_LEFT);
3542
3543 free_space = btrfs_leaf_free_space(left);
3544 if (free_space < data_size) {
3545 ret = 1;
3546 goto out;
3547 }
3548
3549 ret = btrfs_cow_block(trans, root, left,
3550 path->nodes[1], slot - 1, &left,
3551 BTRFS_NESTING_LEFT_COW);
3552 if (ret) {
3553 /* we hit -ENOSPC, but it isn't fatal here */
3554 if (ret == -ENOSPC)
3555 ret = 1;
3556 goto out;
3557 }
3558
3559 if (check_sibling_keys(left, right)) {
3560 ret = -EUCLEAN;
3561 btrfs_abort_transaction(trans, ret);
3562 goto out;
3563 }
3564 return __push_leaf_left(trans, path, min_data_size, empty, left,
3565 free_space, right_nritems, max_slot);
3566out:
3567 btrfs_tree_unlock(left);
3568 free_extent_buffer(left);
3569 return ret;
3570}
3571
3572/*
3573 * split the path's leaf in two, making sure there is at least data_size
3574 * available for the resulting leaf level of the path.
3575 */
3576static noinline int copy_for_split(struct btrfs_trans_handle *trans,
3577 struct btrfs_path *path,
3578 struct extent_buffer *l,
3579 struct extent_buffer *right,
3580 int slot, int mid, int nritems)
3581{
3582 struct btrfs_fs_info *fs_info = trans->fs_info;
3583 int data_copy_size;
3584 int rt_data_off;
3585 int i;
3586 int ret;
3587 struct btrfs_disk_key disk_key;
3588 struct btrfs_map_token token;
3589
3590 nritems = nritems - mid;
3591 btrfs_set_header_nritems(right, nritems);
3592 data_copy_size = btrfs_item_data_end(l, mid) - leaf_data_end(l);
3593
3594 copy_leaf_items(right, l, 0, mid, nritems);
3595
3596 copy_leaf_data(right, l, BTRFS_LEAF_DATA_SIZE(fs_info) - data_copy_size,
3597 leaf_data_end(l), data_copy_size);
3598
3599 rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_data_end(l, mid);
3600
3601 btrfs_init_map_token(&token, right);
3602 for (i = 0; i < nritems; i++) {
3603 u32 ioff;
3604
3605 ioff = btrfs_token_item_offset(&token, i);
3606 btrfs_set_token_item_offset(&token, i, ioff + rt_data_off);
3607 }
3608
3609 btrfs_set_header_nritems(l, mid);
3610 btrfs_item_key(right, &disk_key, 0);
3611 ret = insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
3612 if (ret < 0)
3613 return ret;
3614
3615 btrfs_mark_buffer_dirty(trans, right);
3616 btrfs_mark_buffer_dirty(trans, l);
3617 BUG_ON(path->slots[0] != slot);
3618
3619 if (mid <= slot) {
3620 btrfs_tree_unlock(path->nodes[0]);
3621 free_extent_buffer(path->nodes[0]);
3622 path->nodes[0] = right;
3623 path->slots[0] -= mid;
3624 path->slots[1] += 1;
3625 } else {
3626 btrfs_tree_unlock(right);
3627 free_extent_buffer(right);
3628 }
3629
3630 BUG_ON(path->slots[0] < 0);
3631
3632 return 0;
3633}
3634
3635/*
3636 * double splits happen when we need to insert a big item in the middle
3637 * of a leaf. A double split can leave us with 3 mostly empty leaves:
3638 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
3639 * A B C
3640 *
3641 * We avoid this by trying to push the items on either side of our target
3642 * into the adjacent leaves. If all goes well we can avoid the double split
3643 * completely.
3644 */
3645static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
3646 struct btrfs_root *root,
3647 struct btrfs_path *path,
3648 int data_size)
3649{
3650 int ret;
3651 int progress = 0;
3652 int slot;
3653 u32 nritems;
3654 int space_needed = data_size;
3655
3656 slot = path->slots[0];
3657 if (slot < btrfs_header_nritems(path->nodes[0]))
3658 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3659
3660 /*
3661 * try to push all the items after our slot into the
3662 * right leaf
3663 */
3664 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
3665 if (ret < 0)
3666 return ret;
3667
3668 if (ret == 0)
3669 progress++;
3670
3671 nritems = btrfs_header_nritems(path->nodes[0]);
3672 /*
3673 * our goal is to get our slot at the start or end of a leaf. If
3674 * we've done so we're done
3675 */
3676 if (path->slots[0] == 0 || path->slots[0] == nritems)
3677 return 0;
3678
3679 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3680 return 0;
3681
3682 /* try to push all the items before our slot into the next leaf */
3683 slot = path->slots[0];
3684 space_needed = data_size;
3685 if (slot > 0)
3686 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3687 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
3688 if (ret < 0)
3689 return ret;
3690
3691 if (ret == 0)
3692 progress++;
3693
3694 if (progress)
3695 return 0;
3696 return 1;
3697}
3698
3699/*
3700 * split the path's leaf in two, making sure there is at least data_size
3701 * available for the resulting leaf level of the path.
3702 *
3703 * returns 0 if all went well and < 0 on failure.
3704 */
3705static noinline int split_leaf(struct btrfs_trans_handle *trans,
3706 struct btrfs_root *root,
3707 const struct btrfs_key *ins_key,
3708 struct btrfs_path *path, int data_size,
3709 int extend)
3710{
3711 struct btrfs_disk_key disk_key;
3712 struct extent_buffer *l;
3713 u32 nritems;
3714 int mid;
3715 int slot;
3716 struct extent_buffer *right;
3717 struct btrfs_fs_info *fs_info = root->fs_info;
3718 int ret = 0;
3719 int wret;
3720 int split;
3721 int num_doubles = 0;
3722 int tried_avoid_double = 0;
3723
3724 l = path->nodes[0];
3725 slot = path->slots[0];
3726 if (extend && data_size + btrfs_item_size(l, slot) +
3727 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
3728 return -EOVERFLOW;
3729
3730 /* first try to make some room by pushing left and right */
3731 if (data_size && path->nodes[1]) {
3732 int space_needed = data_size;
3733
3734 if (slot < btrfs_header_nritems(l))
3735 space_needed -= btrfs_leaf_free_space(l);
3736
3737 wret = push_leaf_right(trans, root, path, space_needed,
3738 space_needed, 0, 0);
3739 if (wret < 0)
3740 return wret;
3741 if (wret) {
3742 space_needed = data_size;
3743 if (slot > 0)
3744 space_needed -= btrfs_leaf_free_space(l);
3745 wret = push_leaf_left(trans, root, path, space_needed,
3746 space_needed, 0, (u32)-1);
3747 if (wret < 0)
3748 return wret;
3749 }
3750 l = path->nodes[0];
3751
3752 /* did the pushes work? */
3753 if (btrfs_leaf_free_space(l) >= data_size)
3754 return 0;
3755 }
3756
3757 if (!path->nodes[1]) {
3758 ret = insert_new_root(trans, root, path, 1);
3759 if (ret)
3760 return ret;
3761 }
3762again:
3763 split = 1;
3764 l = path->nodes[0];
3765 slot = path->slots[0];
3766 nritems = btrfs_header_nritems(l);
3767 mid = (nritems + 1) / 2;
3768
3769 if (mid <= slot) {
3770 if (nritems == 1 ||
3771 leaf_space_used(l, mid, nritems - mid) + data_size >
3772 BTRFS_LEAF_DATA_SIZE(fs_info)) {
3773 if (slot >= nritems) {
3774 split = 0;
3775 } else {
3776 mid = slot;
3777 if (mid != nritems &&
3778 leaf_space_used(l, mid, nritems - mid) +
3779 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3780 if (data_size && !tried_avoid_double)
3781 goto push_for_double;
3782 split = 2;
3783 }
3784 }
3785 }
3786 } else {
3787 if (leaf_space_used(l, 0, mid) + data_size >
3788 BTRFS_LEAF_DATA_SIZE(fs_info)) {
3789 if (!extend && data_size && slot == 0) {
3790 split = 0;
3791 } else if ((extend || !data_size) && slot == 0) {
3792 mid = 1;
3793 } else {
3794 mid = slot;
3795 if (mid != nritems &&
3796 leaf_space_used(l, mid, nritems - mid) +
3797 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3798 if (data_size && !tried_avoid_double)
3799 goto push_for_double;
3800 split = 2;
3801 }
3802 }
3803 }
3804 }
3805
3806 if (split == 0)
3807 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3808 else
3809 btrfs_item_key(l, &disk_key, mid);
3810
3811 /*
3812 * We have to about BTRFS_NESTING_NEW_ROOT here if we've done a double
3813 * split, because we're only allowed to have MAX_LOCKDEP_SUBCLASSES
3814 * subclasses, which is 8 at the time of this patch, and we've maxed it
3815 * out. In the future we could add a
3816 * BTRFS_NESTING_SPLIT_THE_SPLITTENING if we need to, but for now just
3817 * use BTRFS_NESTING_NEW_ROOT.
3818 */
3819 right = btrfs_alloc_tree_block(trans, root, 0, btrfs_root_id(root),
3820 &disk_key, 0, l->start, 0, 0,
3821 num_doubles ? BTRFS_NESTING_NEW_ROOT :
3822 BTRFS_NESTING_SPLIT);
3823 if (IS_ERR(right))
3824 return PTR_ERR(right);
3825
3826 root_add_used_bytes(root);
3827
3828 if (split == 0) {
3829 if (mid <= slot) {
3830 btrfs_set_header_nritems(right, 0);
3831 ret = insert_ptr(trans, path, &disk_key,
3832 right->start, path->slots[1] + 1, 1);
3833 if (ret < 0) {
3834 btrfs_tree_unlock(right);
3835 free_extent_buffer(right);
3836 return ret;
3837 }
3838 btrfs_tree_unlock(path->nodes[0]);
3839 free_extent_buffer(path->nodes[0]);
3840 path->nodes[0] = right;
3841 path->slots[0] = 0;
3842 path->slots[1] += 1;
3843 } else {
3844 btrfs_set_header_nritems(right, 0);
3845 ret = insert_ptr(trans, path, &disk_key,
3846 right->start, path->slots[1], 1);
3847 if (ret < 0) {
3848 btrfs_tree_unlock(right);
3849 free_extent_buffer(right);
3850 return ret;
3851 }
3852 btrfs_tree_unlock(path->nodes[0]);
3853 free_extent_buffer(path->nodes[0]);
3854 path->nodes[0] = right;
3855 path->slots[0] = 0;
3856 if (path->slots[1] == 0)
3857 fixup_low_keys(trans, path, &disk_key, 1);
3858 }
3859 /*
3860 * We create a new leaf 'right' for the required ins_len and
3861 * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
3862 * the content of ins_len to 'right'.
3863 */
3864 return ret;
3865 }
3866
3867 ret = copy_for_split(trans, path, l, right, slot, mid, nritems);
3868 if (ret < 0) {
3869 btrfs_tree_unlock(right);
3870 free_extent_buffer(right);
3871 return ret;
3872 }
3873
3874 if (split == 2) {
3875 BUG_ON(num_doubles != 0);
3876 num_doubles++;
3877 goto again;
3878 }
3879
3880 return 0;
3881
3882push_for_double:
3883 push_for_double_split(trans, root, path, data_size);
3884 tried_avoid_double = 1;
3885 if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3886 return 0;
3887 goto again;
3888}
3889
3890static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3891 struct btrfs_root *root,
3892 struct btrfs_path *path, int ins_len)
3893{
3894 struct btrfs_key key;
3895 struct extent_buffer *leaf;
3896 struct btrfs_file_extent_item *fi;
3897 u64 extent_len = 0;
3898 u32 item_size;
3899 int ret;
3900
3901 leaf = path->nodes[0];
3902 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3903
3904 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3905 key.type != BTRFS_EXTENT_CSUM_KEY);
3906
3907 if (btrfs_leaf_free_space(leaf) >= ins_len)
3908 return 0;
3909
3910 item_size = btrfs_item_size(leaf, path->slots[0]);
3911 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3912 fi = btrfs_item_ptr(leaf, path->slots[0],
3913 struct btrfs_file_extent_item);
3914 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3915 }
3916 btrfs_release_path(path);
3917
3918 path->keep_locks = 1;
3919 path->search_for_split = 1;
3920 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3921 path->search_for_split = 0;
3922 if (ret > 0)
3923 ret = -EAGAIN;
3924 if (ret < 0)
3925 goto err;
3926
3927 ret = -EAGAIN;
3928 leaf = path->nodes[0];
3929 /* if our item isn't there, return now */
3930 if (item_size != btrfs_item_size(leaf, path->slots[0]))
3931 goto err;
3932
3933 /* the leaf has changed, it now has room. return now */
3934 if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
3935 goto err;
3936
3937 if (key.type == BTRFS_EXTENT_DATA_KEY) {
3938 fi = btrfs_item_ptr(leaf, path->slots[0],
3939 struct btrfs_file_extent_item);
3940 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3941 goto err;
3942 }
3943
3944 ret = split_leaf(trans, root, &key, path, ins_len, 1);
3945 if (ret)
3946 goto err;
3947
3948 path->keep_locks = 0;
3949 btrfs_unlock_up_safe(path, 1);
3950 return 0;
3951err:
3952 path->keep_locks = 0;
3953 return ret;
3954}
3955
3956static noinline int split_item(struct btrfs_trans_handle *trans,
3957 struct btrfs_path *path,
3958 const struct btrfs_key *new_key,
3959 unsigned long split_offset)
3960{
3961 struct extent_buffer *leaf;
3962 int orig_slot, slot;
3963 char *buf;
3964 u32 nritems;
3965 u32 item_size;
3966 u32 orig_offset;
3967 struct btrfs_disk_key disk_key;
3968
3969 leaf = path->nodes[0];
3970 /*
3971 * Shouldn't happen because the caller must have previously called
3972 * setup_leaf_for_split() to make room for the new item in the leaf.
3973 */
3974 if (WARN_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item)))
3975 return -ENOSPC;
3976
3977 orig_slot = path->slots[0];
3978 orig_offset = btrfs_item_offset(leaf, path->slots[0]);
3979 item_size = btrfs_item_size(leaf, path->slots[0]);
3980
3981 buf = kmalloc(item_size, GFP_NOFS);
3982 if (!buf)
3983 return -ENOMEM;
3984
3985 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3986 path->slots[0]), item_size);
3987
3988 slot = path->slots[0] + 1;
3989 nritems = btrfs_header_nritems(leaf);
3990 if (slot != nritems) {
3991 /* shift the items */
3992 memmove_leaf_items(leaf, slot + 1, slot, nritems - slot);
3993 }
3994
3995 btrfs_cpu_key_to_disk(&disk_key, new_key);
3996 btrfs_set_item_key(leaf, &disk_key, slot);
3997
3998 btrfs_set_item_offset(leaf, slot, orig_offset);
3999 btrfs_set_item_size(leaf, slot, item_size - split_offset);
4000
4001 btrfs_set_item_offset(leaf, orig_slot,
4002 orig_offset + item_size - split_offset);
4003 btrfs_set_item_size(leaf, orig_slot, split_offset);
4004
4005 btrfs_set_header_nritems(leaf, nritems + 1);
4006
4007 /* write the data for the start of the original item */
4008 write_extent_buffer(leaf, buf,
4009 btrfs_item_ptr_offset(leaf, path->slots[0]),
4010 split_offset);
4011
4012 /* write the data for the new item */
4013 write_extent_buffer(leaf, buf + split_offset,
4014 btrfs_item_ptr_offset(leaf, slot),
4015 item_size - split_offset);
4016 btrfs_mark_buffer_dirty(trans, leaf);
4017
4018 BUG_ON(btrfs_leaf_free_space(leaf) < 0);
4019 kfree(buf);
4020 return 0;
4021}
4022
4023/*
4024 * This function splits a single item into two items,
4025 * giving 'new_key' to the new item and splitting the
4026 * old one at split_offset (from the start of the item).
4027 *
4028 * The path may be released by this operation. After
4029 * the split, the path is pointing to the old item. The
4030 * new item is going to be in the same node as the old one.
4031 *
4032 * Note, the item being split must be smaller enough to live alone on
4033 * a tree block with room for one extra struct btrfs_item
4034 *
4035 * This allows us to split the item in place, keeping a lock on the
4036 * leaf the entire time.
4037 */
4038int btrfs_split_item(struct btrfs_trans_handle *trans,
4039 struct btrfs_root *root,
4040 struct btrfs_path *path,
4041 const struct btrfs_key *new_key,
4042 unsigned long split_offset)
4043{
4044 int ret;
4045 ret = setup_leaf_for_split(trans, root, path,
4046 sizeof(struct btrfs_item));
4047 if (ret)
4048 return ret;
4049
4050 ret = split_item(trans, path, new_key, split_offset);
4051 return ret;
4052}
4053
4054/*
4055 * make the item pointed to by the path smaller. new_size indicates
4056 * how small to make it, and from_end tells us if we just chop bytes
4057 * off the end of the item or if we shift the item to chop bytes off
4058 * the front.
4059 */
4060void btrfs_truncate_item(struct btrfs_trans_handle *trans,
4061 const struct btrfs_path *path, u32 new_size, int from_end)
4062{
4063 int slot;
4064 struct extent_buffer *leaf;
4065 u32 nritems;
4066 unsigned int data_end;
4067 unsigned int old_data_start;
4068 unsigned int old_size;
4069 unsigned int size_diff;
4070 int i;
4071 struct btrfs_map_token token;
4072
4073 leaf = path->nodes[0];
4074 slot = path->slots[0];
4075
4076 old_size = btrfs_item_size(leaf, slot);
4077 if (old_size == new_size)
4078 return;
4079
4080 nritems = btrfs_header_nritems(leaf);
4081 data_end = leaf_data_end(leaf);
4082
4083 old_data_start = btrfs_item_offset(leaf, slot);
4084
4085 size_diff = old_size - new_size;
4086
4087 BUG_ON(slot < 0);
4088 BUG_ON(slot >= nritems);
4089
4090 /*
4091 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4092 */
4093 /* first correct the data pointers */
4094 btrfs_init_map_token(&token, leaf);
4095 for (i = slot; i < nritems; i++) {
4096 u32 ioff;
4097
4098 ioff = btrfs_token_item_offset(&token, i);
4099 btrfs_set_token_item_offset(&token, i, ioff + size_diff);
4100 }
4101
4102 /* shift the data */
4103 if (from_end) {
4104 memmove_leaf_data(leaf, data_end + size_diff, data_end,
4105 old_data_start + new_size - data_end);
4106 } else {
4107 struct btrfs_disk_key disk_key;
4108 u64 offset;
4109
4110 btrfs_item_key(leaf, &disk_key, slot);
4111
4112 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4113 unsigned long ptr;
4114 struct btrfs_file_extent_item *fi;
4115
4116 fi = btrfs_item_ptr(leaf, slot,
4117 struct btrfs_file_extent_item);
4118 fi = (struct btrfs_file_extent_item *)(
4119 (unsigned long)fi - size_diff);
4120
4121 if (btrfs_file_extent_type(leaf, fi) ==
4122 BTRFS_FILE_EXTENT_INLINE) {
4123 ptr = btrfs_item_ptr_offset(leaf, slot);
4124 memmove_extent_buffer(leaf, ptr,
4125 (unsigned long)fi,
4126 BTRFS_FILE_EXTENT_INLINE_DATA_START);
4127 }
4128 }
4129
4130 memmove_leaf_data(leaf, data_end + size_diff, data_end,
4131 old_data_start - data_end);
4132
4133 offset = btrfs_disk_key_offset(&disk_key);
4134 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4135 btrfs_set_item_key(leaf, &disk_key, slot);
4136 if (slot == 0)
4137 fixup_low_keys(trans, path, &disk_key, 1);
4138 }
4139
4140 btrfs_set_item_size(leaf, slot, new_size);
4141 btrfs_mark_buffer_dirty(trans, leaf);
4142
4143 if (btrfs_leaf_free_space(leaf) < 0) {
4144 btrfs_print_leaf(leaf);
4145 BUG();
4146 }
4147}
4148
4149/*
4150 * make the item pointed to by the path bigger, data_size is the added size.
4151 */
4152void btrfs_extend_item(struct btrfs_trans_handle *trans,
4153 const struct btrfs_path *path, u32 data_size)
4154{
4155 int slot;
4156 struct extent_buffer *leaf;
4157 u32 nritems;
4158 unsigned int data_end;
4159 unsigned int old_data;
4160 unsigned int old_size;
4161 int i;
4162 struct btrfs_map_token token;
4163
4164 leaf = path->nodes[0];
4165
4166 nritems = btrfs_header_nritems(leaf);
4167 data_end = leaf_data_end(leaf);
4168
4169 if (btrfs_leaf_free_space(leaf) < data_size) {
4170 btrfs_print_leaf(leaf);
4171 BUG();
4172 }
4173 slot = path->slots[0];
4174 old_data = btrfs_item_data_end(leaf, slot);
4175
4176 BUG_ON(slot < 0);
4177 if (slot >= nritems) {
4178 btrfs_print_leaf(leaf);
4179 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
4180 slot, nritems);
4181 BUG();
4182 }
4183
4184 /*
4185 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4186 */
4187 /* first correct the data pointers */
4188 btrfs_init_map_token(&token, leaf);
4189 for (i = slot; i < nritems; i++) {
4190 u32 ioff;
4191
4192 ioff = btrfs_token_item_offset(&token, i);
4193 btrfs_set_token_item_offset(&token, i, ioff - data_size);
4194 }
4195
4196 /* shift the data */
4197 memmove_leaf_data(leaf, data_end - data_size, data_end,
4198 old_data - data_end);
4199
4200 data_end = old_data;
4201 old_size = btrfs_item_size(leaf, slot);
4202 btrfs_set_item_size(leaf, slot, old_size + data_size);
4203 btrfs_mark_buffer_dirty(trans, leaf);
4204
4205 if (btrfs_leaf_free_space(leaf) < 0) {
4206 btrfs_print_leaf(leaf);
4207 BUG();
4208 }
4209}
4210
4211/*
4212 * Make space in the node before inserting one or more items.
4213 *
4214 * @trans: transaction handle
4215 * @root: root we are inserting items to
4216 * @path: points to the leaf/slot where we are going to insert new items
4217 * @batch: information about the batch of items to insert
4218 *
4219 * Main purpose is to save stack depth by doing the bulk of the work in a
4220 * function that doesn't call btrfs_search_slot
4221 */
4222static void setup_items_for_insert(struct btrfs_trans_handle *trans,
4223 struct btrfs_root *root, struct btrfs_path *path,
4224 const struct btrfs_item_batch *batch)
4225{
4226 struct btrfs_fs_info *fs_info = root->fs_info;
4227 int i;
4228 u32 nritems;
4229 unsigned int data_end;
4230 struct btrfs_disk_key disk_key;
4231 struct extent_buffer *leaf;
4232 int slot;
4233 struct btrfs_map_token token;
4234 u32 total_size;
4235
4236 /*
4237 * Before anything else, update keys in the parent and other ancestors
4238 * if needed, then release the write locks on them, so that other tasks
4239 * can use them while we modify the leaf.
4240 */
4241 if (path->slots[0] == 0) {
4242 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[0]);
4243 fixup_low_keys(trans, path, &disk_key, 1);
4244 }
4245 btrfs_unlock_up_safe(path, 1);
4246
4247 leaf = path->nodes[0];
4248 slot = path->slots[0];
4249
4250 nritems = btrfs_header_nritems(leaf);
4251 data_end = leaf_data_end(leaf);
4252 total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4253
4254 if (btrfs_leaf_free_space(leaf) < total_size) {
4255 btrfs_print_leaf(leaf);
4256 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4257 total_size, btrfs_leaf_free_space(leaf));
4258 BUG();
4259 }
4260
4261 btrfs_init_map_token(&token, leaf);
4262 if (slot != nritems) {
4263 unsigned int old_data = btrfs_item_data_end(leaf, slot);
4264
4265 if (old_data < data_end) {
4266 btrfs_print_leaf(leaf);
4267 btrfs_crit(fs_info,
4268 "item at slot %d with data offset %u beyond data end of leaf %u",
4269 slot, old_data, data_end);
4270 BUG();
4271 }
4272 /*
4273 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4274 */
4275 /* first correct the data pointers */
4276 for (i = slot; i < nritems; i++) {
4277 u32 ioff;
4278
4279 ioff = btrfs_token_item_offset(&token, i);
4280 btrfs_set_token_item_offset(&token, i,
4281 ioff - batch->total_data_size);
4282 }
4283 /* shift the items */
4284 memmove_leaf_items(leaf, slot + batch->nr, slot, nritems - slot);
4285
4286 /* shift the data */
4287 memmove_leaf_data(leaf, data_end - batch->total_data_size,
4288 data_end, old_data - data_end);
4289 data_end = old_data;
4290 }
4291
4292 /* setup the item for the new data */
4293 for (i = 0; i < batch->nr; i++) {
4294 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[i]);
4295 btrfs_set_item_key(leaf, &disk_key, slot + i);
4296 data_end -= batch->data_sizes[i];
4297 btrfs_set_token_item_offset(&token, slot + i, data_end);
4298 btrfs_set_token_item_size(&token, slot + i, batch->data_sizes[i]);
4299 }
4300
4301 btrfs_set_header_nritems(leaf, nritems + batch->nr);
4302 btrfs_mark_buffer_dirty(trans, leaf);
4303
4304 if (btrfs_leaf_free_space(leaf) < 0) {
4305 btrfs_print_leaf(leaf);
4306 BUG();
4307 }
4308}
4309
4310/*
4311 * Insert a new item into a leaf.
4312 *
4313 * @trans: Transaction handle.
4314 * @root: The root of the btree.
4315 * @path: A path pointing to the target leaf and slot.
4316 * @key: The key of the new item.
4317 * @data_size: The size of the data associated with the new key.
4318 */
4319void btrfs_setup_item_for_insert(struct btrfs_trans_handle *trans,
4320 struct btrfs_root *root,
4321 struct btrfs_path *path,
4322 const struct btrfs_key *key,
4323 u32 data_size)
4324{
4325 struct btrfs_item_batch batch;
4326
4327 batch.keys = key;
4328 batch.data_sizes = &data_size;
4329 batch.total_data_size = data_size;
4330 batch.nr = 1;
4331
4332 setup_items_for_insert(trans, root, path, &batch);
4333}
4334
4335/*
4336 * Given a key and some data, insert items into the tree.
4337 * This does all the path init required, making room in the tree if needed.
4338 *
4339 * Returns: 0 on success
4340 * -EEXIST if the first key already exists
4341 * < 0 on other errors
4342 */
4343int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4344 struct btrfs_root *root,
4345 struct btrfs_path *path,
4346 const struct btrfs_item_batch *batch)
4347{
4348 int ret = 0;
4349 int slot;
4350 u32 total_size;
4351
4352 total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4353 ret = btrfs_search_slot(trans, root, &batch->keys[0], path, total_size, 1);
4354 if (ret == 0)
4355 return -EEXIST;
4356 if (ret < 0)
4357 return ret;
4358
4359 slot = path->slots[0];
4360 BUG_ON(slot < 0);
4361
4362 setup_items_for_insert(trans, root, path, batch);
4363 return 0;
4364}
4365
4366/*
4367 * Given a key and some data, insert an item into the tree.
4368 * This does all the path init required, making room in the tree if needed.
4369 */
4370int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4371 const struct btrfs_key *cpu_key, void *data,
4372 u32 data_size)
4373{
4374 int ret = 0;
4375 struct btrfs_path *path;
4376 struct extent_buffer *leaf;
4377 unsigned long ptr;
4378
4379 path = btrfs_alloc_path();
4380 if (!path)
4381 return -ENOMEM;
4382 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4383 if (!ret) {
4384 leaf = path->nodes[0];
4385 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4386 write_extent_buffer(leaf, data, ptr, data_size);
4387 btrfs_mark_buffer_dirty(trans, leaf);
4388 }
4389 btrfs_free_path(path);
4390 return ret;
4391}
4392
4393/*
4394 * This function duplicates an item, giving 'new_key' to the new item.
4395 * It guarantees both items live in the same tree leaf and the new item is
4396 * contiguous with the original item.
4397 *
4398 * This allows us to split a file extent in place, keeping a lock on the leaf
4399 * the entire time.
4400 */
4401int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4402 struct btrfs_root *root,
4403 struct btrfs_path *path,
4404 const struct btrfs_key *new_key)
4405{
4406 struct extent_buffer *leaf;
4407 int ret;
4408 u32 item_size;
4409
4410 leaf = path->nodes[0];
4411 item_size = btrfs_item_size(leaf, path->slots[0]);
4412 ret = setup_leaf_for_split(trans, root, path,
4413 item_size + sizeof(struct btrfs_item));
4414 if (ret)
4415 return ret;
4416
4417 path->slots[0]++;
4418 btrfs_setup_item_for_insert(trans, root, path, new_key, item_size);
4419 leaf = path->nodes[0];
4420 memcpy_extent_buffer(leaf,
4421 btrfs_item_ptr_offset(leaf, path->slots[0]),
4422 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4423 item_size);
4424 return 0;
4425}
4426
4427/*
4428 * delete the pointer from a given node.
4429 *
4430 * the tree should have been previously balanced so the deletion does not
4431 * empty a node.
4432 *
4433 * This is exported for use inside btrfs-progs, don't un-export it.
4434 */
4435int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4436 struct btrfs_path *path, int level, int slot)
4437{
4438 struct extent_buffer *parent = path->nodes[level];
4439 u32 nritems;
4440 int ret;
4441
4442 nritems = btrfs_header_nritems(parent);
4443 if (slot != nritems - 1) {
4444 if (level) {
4445 ret = btrfs_tree_mod_log_insert_move(parent, slot,
4446 slot + 1, nritems - slot - 1);
4447 if (ret < 0) {
4448 btrfs_abort_transaction(trans, ret);
4449 return ret;
4450 }
4451 }
4452 memmove_extent_buffer(parent,
4453 btrfs_node_key_ptr_offset(parent, slot),
4454 btrfs_node_key_ptr_offset(parent, slot + 1),
4455 sizeof(struct btrfs_key_ptr) *
4456 (nritems - slot - 1));
4457 } else if (level) {
4458 ret = btrfs_tree_mod_log_insert_key(parent, slot,
4459 BTRFS_MOD_LOG_KEY_REMOVE);
4460 if (ret < 0) {
4461 btrfs_abort_transaction(trans, ret);
4462 return ret;
4463 }
4464 }
4465
4466 nritems--;
4467 btrfs_set_header_nritems(parent, nritems);
4468 if (nritems == 0 && parent == root->node) {
4469 BUG_ON(btrfs_header_level(root->node) != 1);
4470 /* just turn the root into a leaf and break */
4471 btrfs_set_header_level(root->node, 0);
4472 } else if (slot == 0) {
4473 struct btrfs_disk_key disk_key;
4474
4475 btrfs_node_key(parent, &disk_key, 0);
4476 fixup_low_keys(trans, path, &disk_key, level + 1);
4477 }
4478 btrfs_mark_buffer_dirty(trans, parent);
4479 return 0;
4480}
4481
4482/*
4483 * a helper function to delete the leaf pointed to by path->slots[1] and
4484 * path->nodes[1].
4485 *
4486 * This deletes the pointer in path->nodes[1] and frees the leaf
4487 * block extent. zero is returned if it all worked out, < 0 otherwise.
4488 *
4489 * The path must have already been setup for deleting the leaf, including
4490 * all the proper balancing. path->nodes[1] must be locked.
4491 */
4492static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
4493 struct btrfs_root *root,
4494 struct btrfs_path *path,
4495 struct extent_buffer *leaf)
4496{
4497 int ret;
4498
4499 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4500 ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
4501 if (ret < 0)
4502 return ret;
4503
4504 /*
4505 * btrfs_free_extent is expensive, we want to make sure we
4506 * aren't holding any locks when we call it
4507 */
4508 btrfs_unlock_up_safe(path, 0);
4509
4510 root_sub_used_bytes(root);
4511
4512 atomic_inc(&leaf->refs);
4513 ret = btrfs_free_tree_block(trans, btrfs_root_id(root), leaf, 0, 1);
4514 free_extent_buffer_stale(leaf);
4515 if (ret < 0)
4516 btrfs_abort_transaction(trans, ret);
4517
4518 return ret;
4519}
4520/*
4521 * delete the item at the leaf level in path. If that empties
4522 * the leaf, remove it from the tree
4523 */
4524int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4525 struct btrfs_path *path, int slot, int nr)
4526{
4527 struct btrfs_fs_info *fs_info = root->fs_info;
4528 struct extent_buffer *leaf;
4529 int ret = 0;
4530 int wret;
4531 u32 nritems;
4532
4533 leaf = path->nodes[0];
4534 nritems = btrfs_header_nritems(leaf);
4535
4536 if (slot + nr != nritems) {
4537 const u32 last_off = btrfs_item_offset(leaf, slot + nr - 1);
4538 const int data_end = leaf_data_end(leaf);
4539 struct btrfs_map_token token;
4540 u32 dsize = 0;
4541 int i;
4542
4543 for (i = 0; i < nr; i++)
4544 dsize += btrfs_item_size(leaf, slot + i);
4545
4546 memmove_leaf_data(leaf, data_end + dsize, data_end,
4547 last_off - data_end);
4548
4549 btrfs_init_map_token(&token, leaf);
4550 for (i = slot + nr; i < nritems; i++) {
4551 u32 ioff;
4552
4553 ioff = btrfs_token_item_offset(&token, i);
4554 btrfs_set_token_item_offset(&token, i, ioff + dsize);
4555 }
4556
4557 memmove_leaf_items(leaf, slot, slot + nr, nritems - slot - nr);
4558 }
4559 btrfs_set_header_nritems(leaf, nritems - nr);
4560 nritems -= nr;
4561
4562 /* delete the leaf if we've emptied it */
4563 if (nritems == 0) {
4564 if (leaf == root->node) {
4565 btrfs_set_header_level(leaf, 0);
4566 } else {
4567 btrfs_clear_buffer_dirty(trans, leaf);
4568 ret = btrfs_del_leaf(trans, root, path, leaf);
4569 if (ret < 0)
4570 return ret;
4571 }
4572 } else {
4573 int used = leaf_space_used(leaf, 0, nritems);
4574 if (slot == 0) {
4575 struct btrfs_disk_key disk_key;
4576
4577 btrfs_item_key(leaf, &disk_key, 0);
4578 fixup_low_keys(trans, path, &disk_key, 1);
4579 }
4580
4581 /*
4582 * Try to delete the leaf if it is mostly empty. We do this by
4583 * trying to move all its items into its left and right neighbours.
4584 * If we can't move all the items, then we don't delete it - it's
4585 * not ideal, but future insertions might fill the leaf with more
4586 * items, or items from other leaves might be moved later into our
4587 * leaf due to deletions on those leaves.
4588 */
4589 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4590 u32 min_push_space;
4591
4592 /* push_leaf_left fixes the path.
4593 * make sure the path still points to our leaf
4594 * for possible call to btrfs_del_ptr below
4595 */
4596 slot = path->slots[1];
4597 atomic_inc(&leaf->refs);
4598 /*
4599 * We want to be able to at least push one item to the
4600 * left neighbour leaf, and that's the first item.
4601 */
4602 min_push_space = sizeof(struct btrfs_item) +
4603 btrfs_item_size(leaf, 0);
4604 wret = push_leaf_left(trans, root, path, 0,
4605 min_push_space, 1, (u32)-1);
4606 if (wret < 0 && wret != -ENOSPC)
4607 ret = wret;
4608
4609 if (path->nodes[0] == leaf &&
4610 btrfs_header_nritems(leaf)) {
4611 /*
4612 * If we were not able to push all items from our
4613 * leaf to its left neighbour, then attempt to
4614 * either push all the remaining items to the
4615 * right neighbour or none. There's no advantage
4616 * in pushing only some items, instead of all, as
4617 * it's pointless to end up with a leaf having
4618 * too few items while the neighbours can be full
4619 * or nearly full.
4620 */
4621 nritems = btrfs_header_nritems(leaf);
4622 min_push_space = leaf_space_used(leaf, 0, nritems);
4623 wret = push_leaf_right(trans, root, path, 0,
4624 min_push_space, 1, 0);
4625 if (wret < 0 && wret != -ENOSPC)
4626 ret = wret;
4627 }
4628
4629 if (btrfs_header_nritems(leaf) == 0) {
4630 path->slots[1] = slot;
4631 ret = btrfs_del_leaf(trans, root, path, leaf);
4632 if (ret < 0)
4633 return ret;
4634 free_extent_buffer(leaf);
4635 ret = 0;
4636 } else {
4637 /* if we're still in the path, make sure
4638 * we're dirty. Otherwise, one of the
4639 * push_leaf functions must have already
4640 * dirtied this buffer
4641 */
4642 if (path->nodes[0] == leaf)
4643 btrfs_mark_buffer_dirty(trans, leaf);
4644 free_extent_buffer(leaf);
4645 }
4646 } else {
4647 btrfs_mark_buffer_dirty(trans, leaf);
4648 }
4649 }
4650 return ret;
4651}
4652
4653/*
4654 * A helper function to walk down the tree starting at min_key, and looking
4655 * for nodes or leaves that are have a minimum transaction id.
4656 * This is used by the btree defrag code, and tree logging
4657 *
4658 * This does not cow, but it does stuff the starting key it finds back
4659 * into min_key, so you can call btrfs_search_slot with cow=1 on the
4660 * key and get a writable path.
4661 *
4662 * This honors path->lowest_level to prevent descent past a given level
4663 * of the tree.
4664 *
4665 * min_trans indicates the oldest transaction that you are interested
4666 * in walking through. Any nodes or leaves older than min_trans are
4667 * skipped over (without reading them).
4668 *
4669 * returns zero if something useful was found, < 0 on error and 1 if there
4670 * was nothing in the tree that matched the search criteria.
4671 */
4672int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4673 struct btrfs_path *path,
4674 u64 min_trans)
4675{
4676 struct extent_buffer *cur;
4677 struct btrfs_key found_key;
4678 int slot;
4679 int sret;
4680 u32 nritems;
4681 int level;
4682 int ret = 1;
4683 int keep_locks = path->keep_locks;
4684
4685 ASSERT(!path->nowait);
4686 path->keep_locks = 1;
4687again:
4688 cur = btrfs_read_lock_root_node(root);
4689 level = btrfs_header_level(cur);
4690 WARN_ON(path->nodes[level]);
4691 path->nodes[level] = cur;
4692 path->locks[level] = BTRFS_READ_LOCK;
4693
4694 if (btrfs_header_generation(cur) < min_trans) {
4695 ret = 1;
4696 goto out;
4697 }
4698 while (1) {
4699 nritems = btrfs_header_nritems(cur);
4700 level = btrfs_header_level(cur);
4701 sret = btrfs_bin_search(cur, 0, min_key, &slot);
4702 if (sret < 0) {
4703 ret = sret;
4704 goto out;
4705 }
4706
4707 /* at the lowest level, we're done, setup the path and exit */
4708 if (level == path->lowest_level) {
4709 if (slot >= nritems)
4710 goto find_next_key;
4711 ret = 0;
4712 path->slots[level] = slot;
4713 btrfs_item_key_to_cpu(cur, &found_key, slot);
4714 goto out;
4715 }
4716 if (sret && slot > 0)
4717 slot--;
4718 /*
4719 * check this node pointer against the min_trans parameters.
4720 * If it is too old, skip to the next one.
4721 */
4722 while (slot < nritems) {
4723 u64 gen;
4724
4725 gen = btrfs_node_ptr_generation(cur, slot);
4726 if (gen < min_trans) {
4727 slot++;
4728 continue;
4729 }
4730 break;
4731 }
4732find_next_key:
4733 /*
4734 * we didn't find a candidate key in this node, walk forward
4735 * and find another one
4736 */
4737 if (slot >= nritems) {
4738 path->slots[level] = slot;
4739 sret = btrfs_find_next_key(root, path, min_key, level,
4740 min_trans);
4741 if (sret == 0) {
4742 btrfs_release_path(path);
4743 goto again;
4744 } else {
4745 goto out;
4746 }
4747 }
4748 /* save our key for returning back */
4749 btrfs_node_key_to_cpu(cur, &found_key, slot);
4750 path->slots[level] = slot;
4751 if (level == path->lowest_level) {
4752 ret = 0;
4753 goto out;
4754 }
4755 cur = btrfs_read_node_slot(cur, slot);
4756 if (IS_ERR(cur)) {
4757 ret = PTR_ERR(cur);
4758 goto out;
4759 }
4760
4761 btrfs_tree_read_lock(cur);
4762
4763 path->locks[level - 1] = BTRFS_READ_LOCK;
4764 path->nodes[level - 1] = cur;
4765 unlock_up(path, level, 1, 0, NULL);
4766 }
4767out:
4768 path->keep_locks = keep_locks;
4769 if (ret == 0) {
4770 btrfs_unlock_up_safe(path, path->lowest_level + 1);
4771 memcpy(min_key, &found_key, sizeof(found_key));
4772 }
4773 return ret;
4774}
4775
4776/*
4777 * this is similar to btrfs_next_leaf, but does not try to preserve
4778 * and fixup the path. It looks for and returns the next key in the
4779 * tree based on the current path and the min_trans parameters.
4780 *
4781 * 0 is returned if another key is found, < 0 if there are any errors
4782 * and 1 is returned if there are no higher keys in the tree
4783 *
4784 * path->keep_locks should be set to 1 on the search made before
4785 * calling this function.
4786 */
4787int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4788 struct btrfs_key *key, int level, u64 min_trans)
4789{
4790 int slot;
4791 struct extent_buffer *c;
4792
4793 WARN_ON(!path->keep_locks && !path->skip_locking);
4794 while (level < BTRFS_MAX_LEVEL) {
4795 if (!path->nodes[level])
4796 return 1;
4797
4798 slot = path->slots[level] + 1;
4799 c = path->nodes[level];
4800next:
4801 if (slot >= btrfs_header_nritems(c)) {
4802 int ret;
4803 int orig_lowest;
4804 struct btrfs_key cur_key;
4805 if (level + 1 >= BTRFS_MAX_LEVEL ||
4806 !path->nodes[level + 1])
4807 return 1;
4808
4809 if (path->locks[level + 1] || path->skip_locking) {
4810 level++;
4811 continue;
4812 }
4813
4814 slot = btrfs_header_nritems(c) - 1;
4815 if (level == 0)
4816 btrfs_item_key_to_cpu(c, &cur_key, slot);
4817 else
4818 btrfs_node_key_to_cpu(c, &cur_key, slot);
4819
4820 orig_lowest = path->lowest_level;
4821 btrfs_release_path(path);
4822 path->lowest_level = level;
4823 ret = btrfs_search_slot(NULL, root, &cur_key, path,
4824 0, 0);
4825 path->lowest_level = orig_lowest;
4826 if (ret < 0)
4827 return ret;
4828
4829 c = path->nodes[level];
4830 slot = path->slots[level];
4831 if (ret == 0)
4832 slot++;
4833 goto next;
4834 }
4835
4836 if (level == 0)
4837 btrfs_item_key_to_cpu(c, key, slot);
4838 else {
4839 u64 gen = btrfs_node_ptr_generation(c, slot);
4840
4841 if (gen < min_trans) {
4842 slot++;
4843 goto next;
4844 }
4845 btrfs_node_key_to_cpu(c, key, slot);
4846 }
4847 return 0;
4848 }
4849 return 1;
4850}
4851
4852int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
4853 u64 time_seq)
4854{
4855 int slot;
4856 int level;
4857 struct extent_buffer *c;
4858 struct extent_buffer *next;
4859 struct btrfs_fs_info *fs_info = root->fs_info;
4860 struct btrfs_key key;
4861 bool need_commit_sem = false;
4862 u32 nritems;
4863 int ret;
4864 int i;
4865
4866 /*
4867 * The nowait semantics are used only for write paths, where we don't
4868 * use the tree mod log and sequence numbers.
4869 */
4870 if (time_seq)
4871 ASSERT(!path->nowait);
4872
4873 nritems = btrfs_header_nritems(path->nodes[0]);
4874 if (nritems == 0)
4875 return 1;
4876
4877 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4878again:
4879 level = 1;
4880 next = NULL;
4881 btrfs_release_path(path);
4882
4883 path->keep_locks = 1;
4884
4885 if (time_seq) {
4886 ret = btrfs_search_old_slot(root, &key, path, time_seq);
4887 } else {
4888 if (path->need_commit_sem) {
4889 path->need_commit_sem = 0;
4890 need_commit_sem = true;
4891 if (path->nowait) {
4892 if (!down_read_trylock(&fs_info->commit_root_sem)) {
4893 ret = -EAGAIN;
4894 goto done;
4895 }
4896 } else {
4897 down_read(&fs_info->commit_root_sem);
4898 }
4899 }
4900 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4901 }
4902 path->keep_locks = 0;
4903
4904 if (ret < 0)
4905 goto done;
4906
4907 nritems = btrfs_header_nritems(path->nodes[0]);
4908 /*
4909 * by releasing the path above we dropped all our locks. A balance
4910 * could have added more items next to the key that used to be
4911 * at the very end of the block. So, check again here and
4912 * advance the path if there are now more items available.
4913 */
4914 if (nritems > 0 && path->slots[0] < nritems - 1) {
4915 if (ret == 0)
4916 path->slots[0]++;
4917 ret = 0;
4918 goto done;
4919 }
4920 /*
4921 * So the above check misses one case:
4922 * - after releasing the path above, someone has removed the item that
4923 * used to be at the very end of the block, and balance between leafs
4924 * gets another one with bigger key.offset to replace it.
4925 *
4926 * This one should be returned as well, or we can get leaf corruption
4927 * later(esp. in __btrfs_drop_extents()).
4928 *
4929 * And a bit more explanation about this check,
4930 * with ret > 0, the key isn't found, the path points to the slot
4931 * where it should be inserted, so the path->slots[0] item must be the
4932 * bigger one.
4933 */
4934 if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
4935 ret = 0;
4936 goto done;
4937 }
4938
4939 while (level < BTRFS_MAX_LEVEL) {
4940 if (!path->nodes[level]) {
4941 ret = 1;
4942 goto done;
4943 }
4944
4945 slot = path->slots[level] + 1;
4946 c = path->nodes[level];
4947 if (slot >= btrfs_header_nritems(c)) {
4948 level++;
4949 if (level == BTRFS_MAX_LEVEL) {
4950 ret = 1;
4951 goto done;
4952 }
4953 continue;
4954 }
4955
4956
4957 /*
4958 * Our current level is where we're going to start from, and to
4959 * make sure lockdep doesn't complain we need to drop our locks
4960 * and nodes from 0 to our current level.
4961 */
4962 for (i = 0; i < level; i++) {
4963 if (path->locks[level]) {
4964 btrfs_tree_read_unlock(path->nodes[i]);
4965 path->locks[i] = 0;
4966 }
4967 free_extent_buffer(path->nodes[i]);
4968 path->nodes[i] = NULL;
4969 }
4970
4971 next = c;
4972 ret = read_block_for_search(root, path, &next, slot, &key);
4973 if (ret == -EAGAIN && !path->nowait)
4974 goto again;
4975
4976 if (ret < 0) {
4977 btrfs_release_path(path);
4978 goto done;
4979 }
4980
4981 if (!path->skip_locking) {
4982 ret = btrfs_try_tree_read_lock(next);
4983 if (!ret && path->nowait) {
4984 ret = -EAGAIN;
4985 goto done;
4986 }
4987 if (!ret && time_seq) {
4988 /*
4989 * If we don't get the lock, we may be racing
4990 * with push_leaf_left, holding that lock while
4991 * itself waiting for the leaf we've currently
4992 * locked. To solve this situation, we give up
4993 * on our lock and cycle.
4994 */
4995 free_extent_buffer(next);
4996 btrfs_release_path(path);
4997 cond_resched();
4998 goto again;
4999 }
5000 if (!ret)
5001 btrfs_tree_read_lock(next);
5002 }
5003 break;
5004 }
5005 path->slots[level] = slot;
5006 while (1) {
5007 level--;
5008 path->nodes[level] = next;
5009 path->slots[level] = 0;
5010 if (!path->skip_locking)
5011 path->locks[level] = BTRFS_READ_LOCK;
5012 if (!level)
5013 break;
5014
5015 ret = read_block_for_search(root, path, &next, 0, &key);
5016 if (ret == -EAGAIN && !path->nowait)
5017 goto again;
5018
5019 if (ret < 0) {
5020 btrfs_release_path(path);
5021 goto done;
5022 }
5023
5024 if (!path->skip_locking) {
5025 if (path->nowait) {
5026 if (!btrfs_try_tree_read_lock(next)) {
5027 ret = -EAGAIN;
5028 goto done;
5029 }
5030 } else {
5031 btrfs_tree_read_lock(next);
5032 }
5033 }
5034 }
5035 ret = 0;
5036done:
5037 unlock_up(path, 0, 1, 0, NULL);
5038 if (need_commit_sem) {
5039 int ret2;
5040
5041 path->need_commit_sem = 1;
5042 ret2 = finish_need_commit_sem_search(path);
5043 up_read(&fs_info->commit_root_sem);
5044 if (ret2)
5045 ret = ret2;
5046 }
5047
5048 return ret;
5049}
5050
5051int btrfs_next_old_item(struct btrfs_root *root, struct btrfs_path *path, u64 time_seq)
5052{
5053 path->slots[0]++;
5054 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0]))
5055 return btrfs_next_old_leaf(root, path, time_seq);
5056 return 0;
5057}
5058
5059/*
5060 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5061 * searching until it gets past min_objectid or finds an item of 'type'
5062 *
5063 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5064 */
5065int btrfs_previous_item(struct btrfs_root *root,
5066 struct btrfs_path *path, u64 min_objectid,
5067 int type)
5068{
5069 struct btrfs_key found_key;
5070 struct extent_buffer *leaf;
5071 u32 nritems;
5072 int ret;
5073
5074 while (1) {
5075 if (path->slots[0] == 0) {
5076 ret = btrfs_prev_leaf(root, path);
5077 if (ret != 0)
5078 return ret;
5079 } else {
5080 path->slots[0]--;
5081 }
5082 leaf = path->nodes[0];
5083 nritems = btrfs_header_nritems(leaf);
5084 if (nritems == 0)
5085 return 1;
5086 if (path->slots[0] == nritems)
5087 path->slots[0]--;
5088
5089 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5090 if (found_key.objectid < min_objectid)
5091 break;
5092 if (found_key.type == type)
5093 return 0;
5094 if (found_key.objectid == min_objectid &&
5095 found_key.type < type)
5096 break;
5097 }
5098 return 1;
5099}
5100
5101/*
5102 * search in extent tree to find a previous Metadata/Data extent item with
5103 * min objecitd.
5104 *
5105 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5106 */
5107int btrfs_previous_extent_item(struct btrfs_root *root,
5108 struct btrfs_path *path, u64 min_objectid)
5109{
5110 struct btrfs_key found_key;
5111 struct extent_buffer *leaf;
5112 u32 nritems;
5113 int ret;
5114
5115 while (1) {
5116 if (path->slots[0] == 0) {
5117 ret = btrfs_prev_leaf(root, path);
5118 if (ret != 0)
5119 return ret;
5120 } else {
5121 path->slots[0]--;
5122 }
5123 leaf = path->nodes[0];
5124 nritems = btrfs_header_nritems(leaf);
5125 if (nritems == 0)
5126 return 1;
5127 if (path->slots[0] == nritems)
5128 path->slots[0]--;
5129
5130 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5131 if (found_key.objectid < min_objectid)
5132 break;
5133 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5134 found_key.type == BTRFS_METADATA_ITEM_KEY)
5135 return 0;
5136 if (found_key.objectid == min_objectid &&
5137 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5138 break;
5139 }
5140 return 1;
5141}
5142
5143int __init btrfs_ctree_init(void)
5144{
5145 btrfs_path_cachep = KMEM_CACHE(btrfs_path, 0);
5146 if (!btrfs_path_cachep)
5147 return -ENOMEM;
5148 return 0;
5149}
5150
5151void __cold btrfs_ctree_exit(void)
5152{
5153 kmem_cache_destroy(btrfs_path_cachep);
5154}
1/*
2 * Copyright (C) 2007,2008 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/sched.h>
20#include <linux/slab.h>
21#include <linux/rbtree.h>
22#include "ctree.h"
23#include "disk-io.h"
24#include "transaction.h"
25#include "print-tree.h"
26#include "locking.h"
27
28static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
29 *root, struct btrfs_path *path, int level);
30static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
31 *root, struct btrfs_key *ins_key,
32 struct btrfs_path *path, int data_size, int extend);
33static int push_node_left(struct btrfs_trans_handle *trans,
34 struct btrfs_root *root, struct extent_buffer *dst,
35 struct extent_buffer *src, int empty);
36static int balance_node_right(struct btrfs_trans_handle *trans,
37 struct btrfs_root *root,
38 struct extent_buffer *dst_buf,
39 struct extent_buffer *src_buf);
40static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
41 int level, int slot);
42static int tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
43 struct extent_buffer *eb);
44
45struct btrfs_path *btrfs_alloc_path(void)
46{
47 struct btrfs_path *path;
48 path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
49 return path;
50}
51
52/*
53 * set all locked nodes in the path to blocking locks. This should
54 * be done before scheduling
55 */
56noinline void btrfs_set_path_blocking(struct btrfs_path *p)
57{
58 int i;
59 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
60 if (!p->nodes[i] || !p->locks[i])
61 continue;
62 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
63 if (p->locks[i] == BTRFS_READ_LOCK)
64 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
65 else if (p->locks[i] == BTRFS_WRITE_LOCK)
66 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
67 }
68}
69
70/*
71 * reset all the locked nodes in the patch to spinning locks.
72 *
73 * held is used to keep lockdep happy, when lockdep is enabled
74 * we set held to a blocking lock before we go around and
75 * retake all the spinlocks in the path. You can safely use NULL
76 * for held
77 */
78noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
79 struct extent_buffer *held, int held_rw)
80{
81 int i;
82
83#ifdef CONFIG_DEBUG_LOCK_ALLOC
84 /* lockdep really cares that we take all of these spinlocks
85 * in the right order. If any of the locks in the path are not
86 * currently blocking, it is going to complain. So, make really
87 * really sure by forcing the path to blocking before we clear
88 * the path blocking.
89 */
90 if (held) {
91 btrfs_set_lock_blocking_rw(held, held_rw);
92 if (held_rw == BTRFS_WRITE_LOCK)
93 held_rw = BTRFS_WRITE_LOCK_BLOCKING;
94 else if (held_rw == BTRFS_READ_LOCK)
95 held_rw = BTRFS_READ_LOCK_BLOCKING;
96 }
97 btrfs_set_path_blocking(p);
98#endif
99
100 for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
101 if (p->nodes[i] && p->locks[i]) {
102 btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
103 if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
104 p->locks[i] = BTRFS_WRITE_LOCK;
105 else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
106 p->locks[i] = BTRFS_READ_LOCK;
107 }
108 }
109
110#ifdef CONFIG_DEBUG_LOCK_ALLOC
111 if (held)
112 btrfs_clear_lock_blocking_rw(held, held_rw);
113#endif
114}
115
116/* this also releases the path */
117void btrfs_free_path(struct btrfs_path *p)
118{
119 if (!p)
120 return;
121 btrfs_release_path(p);
122 kmem_cache_free(btrfs_path_cachep, p);
123}
124
125/*
126 * path release drops references on the extent buffers in the path
127 * and it drops any locks held by this path
128 *
129 * It is safe to call this on paths that no locks or extent buffers held.
130 */
131noinline void btrfs_release_path(struct btrfs_path *p)
132{
133 int i;
134
135 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
136 p->slots[i] = 0;
137 if (!p->nodes[i])
138 continue;
139 if (p->locks[i]) {
140 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
141 p->locks[i] = 0;
142 }
143 free_extent_buffer(p->nodes[i]);
144 p->nodes[i] = NULL;
145 }
146}
147
148/*
149 * safely gets a reference on the root node of a tree. A lock
150 * is not taken, so a concurrent writer may put a different node
151 * at the root of the tree. See btrfs_lock_root_node for the
152 * looping required.
153 *
154 * The extent buffer returned by this has a reference taken, so
155 * it won't disappear. It may stop being the root of the tree
156 * at any time because there are no locks held.
157 */
158struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
159{
160 struct extent_buffer *eb;
161
162 while (1) {
163 rcu_read_lock();
164 eb = rcu_dereference(root->node);
165
166 /*
167 * RCU really hurts here, we could free up the root node because
168 * it was cow'ed but we may not get the new root node yet so do
169 * the inc_not_zero dance and if it doesn't work then
170 * synchronize_rcu and try again.
171 */
172 if (atomic_inc_not_zero(&eb->refs)) {
173 rcu_read_unlock();
174 break;
175 }
176 rcu_read_unlock();
177 synchronize_rcu();
178 }
179 return eb;
180}
181
182/* loop around taking references on and locking the root node of the
183 * tree until you end up with a lock on the root. A locked buffer
184 * is returned, with a reference held.
185 */
186struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
187{
188 struct extent_buffer *eb;
189
190 while (1) {
191 eb = btrfs_root_node(root);
192 btrfs_tree_lock(eb);
193 if (eb == root->node)
194 break;
195 btrfs_tree_unlock(eb);
196 free_extent_buffer(eb);
197 }
198 return eb;
199}
200
201/* loop around taking references on and locking the root node of the
202 * tree until you end up with a lock on the root. A locked buffer
203 * is returned, with a reference held.
204 */
205static struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
206{
207 struct extent_buffer *eb;
208
209 while (1) {
210 eb = btrfs_root_node(root);
211 btrfs_tree_read_lock(eb);
212 if (eb == root->node)
213 break;
214 btrfs_tree_read_unlock(eb);
215 free_extent_buffer(eb);
216 }
217 return eb;
218}
219
220/* cowonly root (everything not a reference counted cow subvolume), just get
221 * put onto a simple dirty list. transaction.c walks this to make sure they
222 * get properly updated on disk.
223 */
224static void add_root_to_dirty_list(struct btrfs_root *root)
225{
226 spin_lock(&root->fs_info->trans_lock);
227 if (root->track_dirty && list_empty(&root->dirty_list)) {
228 list_add(&root->dirty_list,
229 &root->fs_info->dirty_cowonly_roots);
230 }
231 spin_unlock(&root->fs_info->trans_lock);
232}
233
234/*
235 * used by snapshot creation to make a copy of a root for a tree with
236 * a given objectid. The buffer with the new root node is returned in
237 * cow_ret, and this func returns zero on success or a negative error code.
238 */
239int btrfs_copy_root(struct btrfs_trans_handle *trans,
240 struct btrfs_root *root,
241 struct extent_buffer *buf,
242 struct extent_buffer **cow_ret, u64 new_root_objectid)
243{
244 struct extent_buffer *cow;
245 int ret = 0;
246 int level;
247 struct btrfs_disk_key disk_key;
248
249 WARN_ON(root->ref_cows && trans->transid !=
250 root->fs_info->running_transaction->transid);
251 WARN_ON(root->ref_cows && trans->transid != root->last_trans);
252
253 level = btrfs_header_level(buf);
254 if (level == 0)
255 btrfs_item_key(buf, &disk_key, 0);
256 else
257 btrfs_node_key(buf, &disk_key, 0);
258
259 cow = btrfs_alloc_free_block(trans, root, buf->len, 0,
260 new_root_objectid, &disk_key, level,
261 buf->start, 0);
262 if (IS_ERR(cow))
263 return PTR_ERR(cow);
264
265 copy_extent_buffer(cow, buf, 0, 0, cow->len);
266 btrfs_set_header_bytenr(cow, cow->start);
267 btrfs_set_header_generation(cow, trans->transid);
268 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
269 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
270 BTRFS_HEADER_FLAG_RELOC);
271 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
272 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
273 else
274 btrfs_set_header_owner(cow, new_root_objectid);
275
276 write_extent_buffer(cow, root->fs_info->fsid, btrfs_header_fsid(),
277 BTRFS_FSID_SIZE);
278
279 WARN_ON(btrfs_header_generation(buf) > trans->transid);
280 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
281 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
282 else
283 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
284
285 if (ret)
286 return ret;
287
288 btrfs_mark_buffer_dirty(cow);
289 *cow_ret = cow;
290 return 0;
291}
292
293enum mod_log_op {
294 MOD_LOG_KEY_REPLACE,
295 MOD_LOG_KEY_ADD,
296 MOD_LOG_KEY_REMOVE,
297 MOD_LOG_KEY_REMOVE_WHILE_FREEING,
298 MOD_LOG_KEY_REMOVE_WHILE_MOVING,
299 MOD_LOG_MOVE_KEYS,
300 MOD_LOG_ROOT_REPLACE,
301};
302
303struct tree_mod_move {
304 int dst_slot;
305 int nr_items;
306};
307
308struct tree_mod_root {
309 u64 logical;
310 u8 level;
311};
312
313struct tree_mod_elem {
314 struct rb_node node;
315 u64 index; /* shifted logical */
316 u64 seq;
317 enum mod_log_op op;
318
319 /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
320 int slot;
321
322 /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
323 u64 generation;
324
325 /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
326 struct btrfs_disk_key key;
327 u64 blockptr;
328
329 /* this is used for op == MOD_LOG_MOVE_KEYS */
330 struct tree_mod_move move;
331
332 /* this is used for op == MOD_LOG_ROOT_REPLACE */
333 struct tree_mod_root old_root;
334};
335
336static inline void tree_mod_log_read_lock(struct btrfs_fs_info *fs_info)
337{
338 read_lock(&fs_info->tree_mod_log_lock);
339}
340
341static inline void tree_mod_log_read_unlock(struct btrfs_fs_info *fs_info)
342{
343 read_unlock(&fs_info->tree_mod_log_lock);
344}
345
346static inline void tree_mod_log_write_lock(struct btrfs_fs_info *fs_info)
347{
348 write_lock(&fs_info->tree_mod_log_lock);
349}
350
351static inline void tree_mod_log_write_unlock(struct btrfs_fs_info *fs_info)
352{
353 write_unlock(&fs_info->tree_mod_log_lock);
354}
355
356/*
357 * Increment the upper half of tree_mod_seq, set lower half zero.
358 *
359 * Must be called with fs_info->tree_mod_seq_lock held.
360 */
361static inline u64 btrfs_inc_tree_mod_seq_major(struct btrfs_fs_info *fs_info)
362{
363 u64 seq = atomic64_read(&fs_info->tree_mod_seq);
364 seq &= 0xffffffff00000000ull;
365 seq += 1ull << 32;
366 atomic64_set(&fs_info->tree_mod_seq, seq);
367 return seq;
368}
369
370/*
371 * Increment the lower half of tree_mod_seq.
372 *
373 * Must be called with fs_info->tree_mod_seq_lock held. The way major numbers
374 * are generated should not technically require a spin lock here. (Rationale:
375 * incrementing the minor while incrementing the major seq number is between its
376 * atomic64_read and atomic64_set calls doesn't duplicate sequence numbers, it
377 * just returns a unique sequence number as usual.) We have decided to leave
378 * that requirement in here and rethink it once we notice it really imposes a
379 * problem on some workload.
380 */
381static inline u64 btrfs_inc_tree_mod_seq_minor(struct btrfs_fs_info *fs_info)
382{
383 return atomic64_inc_return(&fs_info->tree_mod_seq);
384}
385
386/*
387 * return the last minor in the previous major tree_mod_seq number
388 */
389u64 btrfs_tree_mod_seq_prev(u64 seq)
390{
391 return (seq & 0xffffffff00000000ull) - 1ull;
392}
393
394/*
395 * This adds a new blocker to the tree mod log's blocker list if the @elem
396 * passed does not already have a sequence number set. So when a caller expects
397 * to record tree modifications, it should ensure to set elem->seq to zero
398 * before calling btrfs_get_tree_mod_seq.
399 * Returns a fresh, unused tree log modification sequence number, even if no new
400 * blocker was added.
401 */
402u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
403 struct seq_list *elem)
404{
405 u64 seq;
406
407 tree_mod_log_write_lock(fs_info);
408 spin_lock(&fs_info->tree_mod_seq_lock);
409 if (!elem->seq) {
410 elem->seq = btrfs_inc_tree_mod_seq_major(fs_info);
411 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
412 }
413 seq = btrfs_inc_tree_mod_seq_minor(fs_info);
414 spin_unlock(&fs_info->tree_mod_seq_lock);
415 tree_mod_log_write_unlock(fs_info);
416
417 return seq;
418}
419
420void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
421 struct seq_list *elem)
422{
423 struct rb_root *tm_root;
424 struct rb_node *node;
425 struct rb_node *next;
426 struct seq_list *cur_elem;
427 struct tree_mod_elem *tm;
428 u64 min_seq = (u64)-1;
429 u64 seq_putting = elem->seq;
430
431 if (!seq_putting)
432 return;
433
434 spin_lock(&fs_info->tree_mod_seq_lock);
435 list_del(&elem->list);
436 elem->seq = 0;
437
438 list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
439 if (cur_elem->seq < min_seq) {
440 if (seq_putting > cur_elem->seq) {
441 /*
442 * blocker with lower sequence number exists, we
443 * cannot remove anything from the log
444 */
445 spin_unlock(&fs_info->tree_mod_seq_lock);
446 return;
447 }
448 min_seq = cur_elem->seq;
449 }
450 }
451 spin_unlock(&fs_info->tree_mod_seq_lock);
452
453 /*
454 * anything that's lower than the lowest existing (read: blocked)
455 * sequence number can be removed from the tree.
456 */
457 tree_mod_log_write_lock(fs_info);
458 tm_root = &fs_info->tree_mod_log;
459 for (node = rb_first(tm_root); node; node = next) {
460 next = rb_next(node);
461 tm = container_of(node, struct tree_mod_elem, node);
462 if (tm->seq > min_seq)
463 continue;
464 rb_erase(node, tm_root);
465 kfree(tm);
466 }
467 tree_mod_log_write_unlock(fs_info);
468}
469
470/*
471 * key order of the log:
472 * index -> sequence
473 *
474 * the index is the shifted logical of the *new* root node for root replace
475 * operations, or the shifted logical of the affected block for all other
476 * operations.
477 *
478 * Note: must be called with write lock (tree_mod_log_write_lock).
479 */
480static noinline int
481__tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
482{
483 struct rb_root *tm_root;
484 struct rb_node **new;
485 struct rb_node *parent = NULL;
486 struct tree_mod_elem *cur;
487
488 BUG_ON(!tm);
489
490 spin_lock(&fs_info->tree_mod_seq_lock);
491 tm->seq = btrfs_inc_tree_mod_seq_minor(fs_info);
492 spin_unlock(&fs_info->tree_mod_seq_lock);
493
494 tm_root = &fs_info->tree_mod_log;
495 new = &tm_root->rb_node;
496 while (*new) {
497 cur = container_of(*new, struct tree_mod_elem, node);
498 parent = *new;
499 if (cur->index < tm->index)
500 new = &((*new)->rb_left);
501 else if (cur->index > tm->index)
502 new = &((*new)->rb_right);
503 else if (cur->seq < tm->seq)
504 new = &((*new)->rb_left);
505 else if (cur->seq > tm->seq)
506 new = &((*new)->rb_right);
507 else
508 return -EEXIST;
509 }
510
511 rb_link_node(&tm->node, parent, new);
512 rb_insert_color(&tm->node, tm_root);
513 return 0;
514}
515
516/*
517 * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
518 * returns zero with the tree_mod_log_lock acquired. The caller must hold
519 * this until all tree mod log insertions are recorded in the rb tree and then
520 * call tree_mod_log_write_unlock() to release.
521 */
522static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
523 struct extent_buffer *eb) {
524 smp_mb();
525 if (list_empty(&(fs_info)->tree_mod_seq_list))
526 return 1;
527 if (eb && btrfs_header_level(eb) == 0)
528 return 1;
529
530 tree_mod_log_write_lock(fs_info);
531 if (list_empty(&(fs_info)->tree_mod_seq_list)) {
532 tree_mod_log_write_unlock(fs_info);
533 return 1;
534 }
535
536 return 0;
537}
538
539/* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
540static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
541 struct extent_buffer *eb)
542{
543 smp_mb();
544 if (list_empty(&(fs_info)->tree_mod_seq_list))
545 return 0;
546 if (eb && btrfs_header_level(eb) == 0)
547 return 0;
548
549 return 1;
550}
551
552static struct tree_mod_elem *
553alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
554 enum mod_log_op op, gfp_t flags)
555{
556 struct tree_mod_elem *tm;
557
558 tm = kzalloc(sizeof(*tm), flags);
559 if (!tm)
560 return NULL;
561
562 tm->index = eb->start >> PAGE_CACHE_SHIFT;
563 if (op != MOD_LOG_KEY_ADD) {
564 btrfs_node_key(eb, &tm->key, slot);
565 tm->blockptr = btrfs_node_blockptr(eb, slot);
566 }
567 tm->op = op;
568 tm->slot = slot;
569 tm->generation = btrfs_node_ptr_generation(eb, slot);
570 RB_CLEAR_NODE(&tm->node);
571
572 return tm;
573}
574
575static noinline int
576tree_mod_log_insert_key(struct btrfs_fs_info *fs_info,
577 struct extent_buffer *eb, int slot,
578 enum mod_log_op op, gfp_t flags)
579{
580 struct tree_mod_elem *tm;
581 int ret;
582
583 if (!tree_mod_need_log(fs_info, eb))
584 return 0;
585
586 tm = alloc_tree_mod_elem(eb, slot, op, flags);
587 if (!tm)
588 return -ENOMEM;
589
590 if (tree_mod_dont_log(fs_info, eb)) {
591 kfree(tm);
592 return 0;
593 }
594
595 ret = __tree_mod_log_insert(fs_info, tm);
596 tree_mod_log_write_unlock(fs_info);
597 if (ret)
598 kfree(tm);
599
600 return ret;
601}
602
603static noinline int
604tree_mod_log_insert_move(struct btrfs_fs_info *fs_info,
605 struct extent_buffer *eb, int dst_slot, int src_slot,
606 int nr_items, gfp_t flags)
607{
608 struct tree_mod_elem *tm = NULL;
609 struct tree_mod_elem **tm_list = NULL;
610 int ret = 0;
611 int i;
612 int locked = 0;
613
614 if (!tree_mod_need_log(fs_info, eb))
615 return 0;
616
617 tm_list = kzalloc(nr_items * sizeof(struct tree_mod_elem *), flags);
618 if (!tm_list)
619 return -ENOMEM;
620
621 tm = kzalloc(sizeof(*tm), flags);
622 if (!tm) {
623 ret = -ENOMEM;
624 goto free_tms;
625 }
626
627 tm->index = eb->start >> PAGE_CACHE_SHIFT;
628 tm->slot = src_slot;
629 tm->move.dst_slot = dst_slot;
630 tm->move.nr_items = nr_items;
631 tm->op = MOD_LOG_MOVE_KEYS;
632
633 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
634 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
635 MOD_LOG_KEY_REMOVE_WHILE_MOVING, flags);
636 if (!tm_list[i]) {
637 ret = -ENOMEM;
638 goto free_tms;
639 }
640 }
641
642 if (tree_mod_dont_log(fs_info, eb))
643 goto free_tms;
644 locked = 1;
645
646 /*
647 * When we override something during the move, we log these removals.
648 * This can only happen when we move towards the beginning of the
649 * buffer, i.e. dst_slot < src_slot.
650 */
651 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
652 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
653 if (ret)
654 goto free_tms;
655 }
656
657 ret = __tree_mod_log_insert(fs_info, tm);
658 if (ret)
659 goto free_tms;
660 tree_mod_log_write_unlock(fs_info);
661 kfree(tm_list);
662
663 return 0;
664free_tms:
665 for (i = 0; i < nr_items; i++) {
666 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
667 rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
668 kfree(tm_list[i]);
669 }
670 if (locked)
671 tree_mod_log_write_unlock(fs_info);
672 kfree(tm_list);
673 kfree(tm);
674
675 return ret;
676}
677
678static inline int
679__tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
680 struct tree_mod_elem **tm_list,
681 int nritems)
682{
683 int i, j;
684 int ret;
685
686 for (i = nritems - 1; i >= 0; i--) {
687 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
688 if (ret) {
689 for (j = nritems - 1; j > i; j--)
690 rb_erase(&tm_list[j]->node,
691 &fs_info->tree_mod_log);
692 return ret;
693 }
694 }
695
696 return 0;
697}
698
699static noinline int
700tree_mod_log_insert_root(struct btrfs_fs_info *fs_info,
701 struct extent_buffer *old_root,
702 struct extent_buffer *new_root, gfp_t flags,
703 int log_removal)
704{
705 struct tree_mod_elem *tm = NULL;
706 struct tree_mod_elem **tm_list = NULL;
707 int nritems = 0;
708 int ret = 0;
709 int i;
710
711 if (!tree_mod_need_log(fs_info, NULL))
712 return 0;
713
714 if (log_removal && btrfs_header_level(old_root) > 0) {
715 nritems = btrfs_header_nritems(old_root);
716 tm_list = kzalloc(nritems * sizeof(struct tree_mod_elem *),
717 flags);
718 if (!tm_list) {
719 ret = -ENOMEM;
720 goto free_tms;
721 }
722 for (i = 0; i < nritems; i++) {
723 tm_list[i] = alloc_tree_mod_elem(old_root, i,
724 MOD_LOG_KEY_REMOVE_WHILE_FREEING, flags);
725 if (!tm_list[i]) {
726 ret = -ENOMEM;
727 goto free_tms;
728 }
729 }
730 }
731
732 tm = kzalloc(sizeof(*tm), flags);
733 if (!tm) {
734 ret = -ENOMEM;
735 goto free_tms;
736 }
737
738 tm->index = new_root->start >> PAGE_CACHE_SHIFT;
739 tm->old_root.logical = old_root->start;
740 tm->old_root.level = btrfs_header_level(old_root);
741 tm->generation = btrfs_header_generation(old_root);
742 tm->op = MOD_LOG_ROOT_REPLACE;
743
744 if (tree_mod_dont_log(fs_info, NULL))
745 goto free_tms;
746
747 if (tm_list)
748 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
749 if (!ret)
750 ret = __tree_mod_log_insert(fs_info, tm);
751
752 tree_mod_log_write_unlock(fs_info);
753 if (ret)
754 goto free_tms;
755 kfree(tm_list);
756
757 return ret;
758
759free_tms:
760 if (tm_list) {
761 for (i = 0; i < nritems; i++)
762 kfree(tm_list[i]);
763 kfree(tm_list);
764 }
765 kfree(tm);
766
767 return ret;
768}
769
770static struct tree_mod_elem *
771__tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
772 int smallest)
773{
774 struct rb_root *tm_root;
775 struct rb_node *node;
776 struct tree_mod_elem *cur = NULL;
777 struct tree_mod_elem *found = NULL;
778 u64 index = start >> PAGE_CACHE_SHIFT;
779
780 tree_mod_log_read_lock(fs_info);
781 tm_root = &fs_info->tree_mod_log;
782 node = tm_root->rb_node;
783 while (node) {
784 cur = container_of(node, struct tree_mod_elem, node);
785 if (cur->index < index) {
786 node = node->rb_left;
787 } else if (cur->index > index) {
788 node = node->rb_right;
789 } else if (cur->seq < min_seq) {
790 node = node->rb_left;
791 } else if (!smallest) {
792 /* we want the node with the highest seq */
793 if (found)
794 BUG_ON(found->seq > cur->seq);
795 found = cur;
796 node = node->rb_left;
797 } else if (cur->seq > min_seq) {
798 /* we want the node with the smallest seq */
799 if (found)
800 BUG_ON(found->seq < cur->seq);
801 found = cur;
802 node = node->rb_right;
803 } else {
804 found = cur;
805 break;
806 }
807 }
808 tree_mod_log_read_unlock(fs_info);
809
810 return found;
811}
812
813/*
814 * this returns the element from the log with the smallest time sequence
815 * value that's in the log (the oldest log item). any element with a time
816 * sequence lower than min_seq will be ignored.
817 */
818static struct tree_mod_elem *
819tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
820 u64 min_seq)
821{
822 return __tree_mod_log_search(fs_info, start, min_seq, 1);
823}
824
825/*
826 * this returns the element from the log with the largest time sequence
827 * value that's in the log (the most recent log item). any element with
828 * a time sequence lower than min_seq will be ignored.
829 */
830static struct tree_mod_elem *
831tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
832{
833 return __tree_mod_log_search(fs_info, start, min_seq, 0);
834}
835
836static noinline int
837tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
838 struct extent_buffer *src, unsigned long dst_offset,
839 unsigned long src_offset, int nr_items)
840{
841 int ret = 0;
842 struct tree_mod_elem **tm_list = NULL;
843 struct tree_mod_elem **tm_list_add, **tm_list_rem;
844 int i;
845 int locked = 0;
846
847 if (!tree_mod_need_log(fs_info, NULL))
848 return 0;
849
850 if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
851 return 0;
852
853 tm_list = kzalloc(nr_items * 2 * sizeof(struct tree_mod_elem *),
854 GFP_NOFS);
855 if (!tm_list)
856 return -ENOMEM;
857
858 tm_list_add = tm_list;
859 tm_list_rem = tm_list + nr_items;
860 for (i = 0; i < nr_items; i++) {
861 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
862 MOD_LOG_KEY_REMOVE, GFP_NOFS);
863 if (!tm_list_rem[i]) {
864 ret = -ENOMEM;
865 goto free_tms;
866 }
867
868 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
869 MOD_LOG_KEY_ADD, GFP_NOFS);
870 if (!tm_list_add[i]) {
871 ret = -ENOMEM;
872 goto free_tms;
873 }
874 }
875
876 if (tree_mod_dont_log(fs_info, NULL))
877 goto free_tms;
878 locked = 1;
879
880 for (i = 0; i < nr_items; i++) {
881 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
882 if (ret)
883 goto free_tms;
884 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
885 if (ret)
886 goto free_tms;
887 }
888
889 tree_mod_log_write_unlock(fs_info);
890 kfree(tm_list);
891
892 return 0;
893
894free_tms:
895 for (i = 0; i < nr_items * 2; i++) {
896 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
897 rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
898 kfree(tm_list[i]);
899 }
900 if (locked)
901 tree_mod_log_write_unlock(fs_info);
902 kfree(tm_list);
903
904 return ret;
905}
906
907static inline void
908tree_mod_log_eb_move(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
909 int dst_offset, int src_offset, int nr_items)
910{
911 int ret;
912 ret = tree_mod_log_insert_move(fs_info, dst, dst_offset, src_offset,
913 nr_items, GFP_NOFS);
914 BUG_ON(ret < 0);
915}
916
917static noinline void
918tree_mod_log_set_node_key(struct btrfs_fs_info *fs_info,
919 struct extent_buffer *eb, int slot, int atomic)
920{
921 int ret;
922
923 ret = tree_mod_log_insert_key(fs_info, eb, slot,
924 MOD_LOG_KEY_REPLACE,
925 atomic ? GFP_ATOMIC : GFP_NOFS);
926 BUG_ON(ret < 0);
927}
928
929static noinline int
930tree_mod_log_free_eb(struct btrfs_fs_info *fs_info, struct extent_buffer *eb)
931{
932 struct tree_mod_elem **tm_list = NULL;
933 int nritems = 0;
934 int i;
935 int ret = 0;
936
937 if (btrfs_header_level(eb) == 0)
938 return 0;
939
940 if (!tree_mod_need_log(fs_info, NULL))
941 return 0;
942
943 nritems = btrfs_header_nritems(eb);
944 tm_list = kzalloc(nritems * sizeof(struct tree_mod_elem *),
945 GFP_NOFS);
946 if (!tm_list)
947 return -ENOMEM;
948
949 for (i = 0; i < nritems; i++) {
950 tm_list[i] = alloc_tree_mod_elem(eb, i,
951 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
952 if (!tm_list[i]) {
953 ret = -ENOMEM;
954 goto free_tms;
955 }
956 }
957
958 if (tree_mod_dont_log(fs_info, eb))
959 goto free_tms;
960
961 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
962 tree_mod_log_write_unlock(fs_info);
963 if (ret)
964 goto free_tms;
965 kfree(tm_list);
966
967 return 0;
968
969free_tms:
970 for (i = 0; i < nritems; i++)
971 kfree(tm_list[i]);
972 kfree(tm_list);
973
974 return ret;
975}
976
977static noinline void
978tree_mod_log_set_root_pointer(struct btrfs_root *root,
979 struct extent_buffer *new_root_node,
980 int log_removal)
981{
982 int ret;
983 ret = tree_mod_log_insert_root(root->fs_info, root->node,
984 new_root_node, GFP_NOFS, log_removal);
985 BUG_ON(ret < 0);
986}
987
988/*
989 * check if the tree block can be shared by multiple trees
990 */
991int btrfs_block_can_be_shared(struct btrfs_root *root,
992 struct extent_buffer *buf)
993{
994 /*
995 * Tree blocks not in refernece counted trees and tree roots
996 * are never shared. If a block was allocated after the last
997 * snapshot and the block was not allocated by tree relocation,
998 * we know the block is not shared.
999 */
1000 if (root->ref_cows &&
1001 buf != root->node && buf != root->commit_root &&
1002 (btrfs_header_generation(buf) <=
1003 btrfs_root_last_snapshot(&root->root_item) ||
1004 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
1005 return 1;
1006#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1007 if (root->ref_cows &&
1008 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1009 return 1;
1010#endif
1011 return 0;
1012}
1013
1014static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
1015 struct btrfs_root *root,
1016 struct extent_buffer *buf,
1017 struct extent_buffer *cow,
1018 int *last_ref)
1019{
1020 u64 refs;
1021 u64 owner;
1022 u64 flags;
1023 u64 new_flags = 0;
1024 int ret;
1025
1026 /*
1027 * Backrefs update rules:
1028 *
1029 * Always use full backrefs for extent pointers in tree block
1030 * allocated by tree relocation.
1031 *
1032 * If a shared tree block is no longer referenced by its owner
1033 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
1034 * use full backrefs for extent pointers in tree block.
1035 *
1036 * If a tree block is been relocating
1037 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
1038 * use full backrefs for extent pointers in tree block.
1039 * The reason for this is some operations (such as drop tree)
1040 * are only allowed for blocks use full backrefs.
1041 */
1042
1043 if (btrfs_block_can_be_shared(root, buf)) {
1044 ret = btrfs_lookup_extent_info(trans, root, buf->start,
1045 btrfs_header_level(buf), 1,
1046 &refs, &flags);
1047 if (ret)
1048 return ret;
1049 if (refs == 0) {
1050 ret = -EROFS;
1051 btrfs_std_error(root->fs_info, ret);
1052 return ret;
1053 }
1054 } else {
1055 refs = 1;
1056 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1057 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1058 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
1059 else
1060 flags = 0;
1061 }
1062
1063 owner = btrfs_header_owner(buf);
1064 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
1065 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
1066
1067 if (refs > 1) {
1068 if ((owner == root->root_key.objectid ||
1069 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
1070 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
1071 ret = btrfs_inc_ref(trans, root, buf, 1, 1);
1072 BUG_ON(ret); /* -ENOMEM */
1073
1074 if (root->root_key.objectid ==
1075 BTRFS_TREE_RELOC_OBJECTID) {
1076 ret = btrfs_dec_ref(trans, root, buf, 0, 1);
1077 BUG_ON(ret); /* -ENOMEM */
1078 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
1079 BUG_ON(ret); /* -ENOMEM */
1080 }
1081 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
1082 } else {
1083
1084 if (root->root_key.objectid ==
1085 BTRFS_TREE_RELOC_OBJECTID)
1086 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
1087 else
1088 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
1089 BUG_ON(ret); /* -ENOMEM */
1090 }
1091 if (new_flags != 0) {
1092 int level = btrfs_header_level(buf);
1093
1094 ret = btrfs_set_disk_extent_flags(trans, root,
1095 buf->start,
1096 buf->len,
1097 new_flags, level, 0);
1098 if (ret)
1099 return ret;
1100 }
1101 } else {
1102 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
1103 if (root->root_key.objectid ==
1104 BTRFS_TREE_RELOC_OBJECTID)
1105 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
1106 else
1107 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
1108 BUG_ON(ret); /* -ENOMEM */
1109 ret = btrfs_dec_ref(trans, root, buf, 1, 1);
1110 BUG_ON(ret); /* -ENOMEM */
1111 }
1112 clean_tree_block(trans, root, buf);
1113 *last_ref = 1;
1114 }
1115 return 0;
1116}
1117
1118/*
1119 * does the dirty work in cow of a single block. The parent block (if
1120 * supplied) is updated to point to the new cow copy. The new buffer is marked
1121 * dirty and returned locked. If you modify the block it needs to be marked
1122 * dirty again.
1123 *
1124 * search_start -- an allocation hint for the new block
1125 *
1126 * empty_size -- a hint that you plan on doing more cow. This is the size in
1127 * bytes the allocator should try to find free next to the block it returns.
1128 * This is just a hint and may be ignored by the allocator.
1129 */
1130static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1131 struct btrfs_root *root,
1132 struct extent_buffer *buf,
1133 struct extent_buffer *parent, int parent_slot,
1134 struct extent_buffer **cow_ret,
1135 u64 search_start, u64 empty_size)
1136{
1137 struct btrfs_disk_key disk_key;
1138 struct extent_buffer *cow;
1139 int level, ret;
1140 int last_ref = 0;
1141 int unlock_orig = 0;
1142 u64 parent_start;
1143
1144 if (*cow_ret == buf)
1145 unlock_orig = 1;
1146
1147 btrfs_assert_tree_locked(buf);
1148
1149 WARN_ON(root->ref_cows && trans->transid !=
1150 root->fs_info->running_transaction->transid);
1151 WARN_ON(root->ref_cows && trans->transid != root->last_trans);
1152
1153 level = btrfs_header_level(buf);
1154
1155 if (level == 0)
1156 btrfs_item_key(buf, &disk_key, 0);
1157 else
1158 btrfs_node_key(buf, &disk_key, 0);
1159
1160 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1161 if (parent)
1162 parent_start = parent->start;
1163 else
1164 parent_start = 0;
1165 } else
1166 parent_start = 0;
1167
1168 cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
1169 root->root_key.objectid, &disk_key,
1170 level, search_start, empty_size);
1171 if (IS_ERR(cow))
1172 return PTR_ERR(cow);
1173
1174 /* cow is set to blocking by btrfs_init_new_buffer */
1175
1176 copy_extent_buffer(cow, buf, 0, 0, cow->len);
1177 btrfs_set_header_bytenr(cow, cow->start);
1178 btrfs_set_header_generation(cow, trans->transid);
1179 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1180 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1181 BTRFS_HEADER_FLAG_RELOC);
1182 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1183 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1184 else
1185 btrfs_set_header_owner(cow, root->root_key.objectid);
1186
1187 write_extent_buffer(cow, root->fs_info->fsid, btrfs_header_fsid(),
1188 BTRFS_FSID_SIZE);
1189
1190 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1191 if (ret) {
1192 btrfs_abort_transaction(trans, root, ret);
1193 return ret;
1194 }
1195
1196 if (root->ref_cows) {
1197 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1198 if (ret)
1199 return ret;
1200 }
1201
1202 if (buf == root->node) {
1203 WARN_ON(parent && parent != buf);
1204 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1205 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1206 parent_start = buf->start;
1207 else
1208 parent_start = 0;
1209
1210 extent_buffer_get(cow);
1211 tree_mod_log_set_root_pointer(root, cow, 1);
1212 rcu_assign_pointer(root->node, cow);
1213
1214 btrfs_free_tree_block(trans, root, buf, parent_start,
1215 last_ref);
1216 free_extent_buffer(buf);
1217 add_root_to_dirty_list(root);
1218 } else {
1219 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1220 parent_start = parent->start;
1221 else
1222 parent_start = 0;
1223
1224 WARN_ON(trans->transid != btrfs_header_generation(parent));
1225 tree_mod_log_insert_key(root->fs_info, parent, parent_slot,
1226 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1227 btrfs_set_node_blockptr(parent, parent_slot,
1228 cow->start);
1229 btrfs_set_node_ptr_generation(parent, parent_slot,
1230 trans->transid);
1231 btrfs_mark_buffer_dirty(parent);
1232 if (last_ref) {
1233 ret = tree_mod_log_free_eb(root->fs_info, buf);
1234 if (ret) {
1235 btrfs_abort_transaction(trans, root, ret);
1236 return ret;
1237 }
1238 }
1239 btrfs_free_tree_block(trans, root, buf, parent_start,
1240 last_ref);
1241 }
1242 if (unlock_orig)
1243 btrfs_tree_unlock(buf);
1244 free_extent_buffer_stale(buf);
1245 btrfs_mark_buffer_dirty(cow);
1246 *cow_ret = cow;
1247 return 0;
1248}
1249
1250/*
1251 * returns the logical address of the oldest predecessor of the given root.
1252 * entries older than time_seq are ignored.
1253 */
1254static struct tree_mod_elem *
1255__tree_mod_log_oldest_root(struct btrfs_fs_info *fs_info,
1256 struct extent_buffer *eb_root, u64 time_seq)
1257{
1258 struct tree_mod_elem *tm;
1259 struct tree_mod_elem *found = NULL;
1260 u64 root_logical = eb_root->start;
1261 int looped = 0;
1262
1263 if (!time_seq)
1264 return NULL;
1265
1266 /*
1267 * the very last operation that's logged for a root is the replacement
1268 * operation (if it is replaced at all). this has the index of the *new*
1269 * root, making it the very first operation that's logged for this root.
1270 */
1271 while (1) {
1272 tm = tree_mod_log_search_oldest(fs_info, root_logical,
1273 time_seq);
1274 if (!looped && !tm)
1275 return NULL;
1276 /*
1277 * if there are no tree operation for the oldest root, we simply
1278 * return it. this should only happen if that (old) root is at
1279 * level 0.
1280 */
1281 if (!tm)
1282 break;
1283
1284 /*
1285 * if there's an operation that's not a root replacement, we
1286 * found the oldest version of our root. normally, we'll find a
1287 * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1288 */
1289 if (tm->op != MOD_LOG_ROOT_REPLACE)
1290 break;
1291
1292 found = tm;
1293 root_logical = tm->old_root.logical;
1294 looped = 1;
1295 }
1296
1297 /* if there's no old root to return, return what we found instead */
1298 if (!found)
1299 found = tm;
1300
1301 return found;
1302}
1303
1304/*
1305 * tm is a pointer to the first operation to rewind within eb. then, all
1306 * previous operations will be rewinded (until we reach something older than
1307 * time_seq).
1308 */
1309static void
1310__tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1311 u64 time_seq, struct tree_mod_elem *first_tm)
1312{
1313 u32 n;
1314 struct rb_node *next;
1315 struct tree_mod_elem *tm = first_tm;
1316 unsigned long o_dst;
1317 unsigned long o_src;
1318 unsigned long p_size = sizeof(struct btrfs_key_ptr);
1319
1320 n = btrfs_header_nritems(eb);
1321 tree_mod_log_read_lock(fs_info);
1322 while (tm && tm->seq >= time_seq) {
1323 /*
1324 * all the operations are recorded with the operator used for
1325 * the modification. as we're going backwards, we do the
1326 * opposite of each operation here.
1327 */
1328 switch (tm->op) {
1329 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1330 BUG_ON(tm->slot < n);
1331 /* Fallthrough */
1332 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1333 case MOD_LOG_KEY_REMOVE:
1334 btrfs_set_node_key(eb, &tm->key, tm->slot);
1335 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1336 btrfs_set_node_ptr_generation(eb, tm->slot,
1337 tm->generation);
1338 n++;
1339 break;
1340 case MOD_LOG_KEY_REPLACE:
1341 BUG_ON(tm->slot >= n);
1342 btrfs_set_node_key(eb, &tm->key, tm->slot);
1343 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1344 btrfs_set_node_ptr_generation(eb, tm->slot,
1345 tm->generation);
1346 break;
1347 case MOD_LOG_KEY_ADD:
1348 /* if a move operation is needed it's in the log */
1349 n--;
1350 break;
1351 case MOD_LOG_MOVE_KEYS:
1352 o_dst = btrfs_node_key_ptr_offset(tm->slot);
1353 o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1354 memmove_extent_buffer(eb, o_dst, o_src,
1355 tm->move.nr_items * p_size);
1356 break;
1357 case MOD_LOG_ROOT_REPLACE:
1358 /*
1359 * this operation is special. for roots, this must be
1360 * handled explicitly before rewinding.
1361 * for non-roots, this operation may exist if the node
1362 * was a root: root A -> child B; then A gets empty and
1363 * B is promoted to the new root. in the mod log, we'll
1364 * have a root-replace operation for B, a tree block
1365 * that is no root. we simply ignore that operation.
1366 */
1367 break;
1368 }
1369 next = rb_next(&tm->node);
1370 if (!next)
1371 break;
1372 tm = container_of(next, struct tree_mod_elem, node);
1373 if (tm->index != first_tm->index)
1374 break;
1375 }
1376 tree_mod_log_read_unlock(fs_info);
1377 btrfs_set_header_nritems(eb, n);
1378}
1379
1380/*
1381 * Called with eb read locked. If the buffer cannot be rewinded, the same buffer
1382 * is returned. If rewind operations happen, a fresh buffer is returned. The
1383 * returned buffer is always read-locked. If the returned buffer is not the
1384 * input buffer, the lock on the input buffer is released and the input buffer
1385 * is freed (its refcount is decremented).
1386 */
1387static struct extent_buffer *
1388tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1389 struct extent_buffer *eb, u64 time_seq)
1390{
1391 struct extent_buffer *eb_rewin;
1392 struct tree_mod_elem *tm;
1393
1394 if (!time_seq)
1395 return eb;
1396
1397 if (btrfs_header_level(eb) == 0)
1398 return eb;
1399
1400 tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1401 if (!tm)
1402 return eb;
1403
1404 btrfs_set_path_blocking(path);
1405 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1406
1407 if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1408 BUG_ON(tm->slot != 0);
1409 eb_rewin = alloc_dummy_extent_buffer(eb->start,
1410 fs_info->tree_root->nodesize);
1411 if (!eb_rewin) {
1412 btrfs_tree_read_unlock_blocking(eb);
1413 free_extent_buffer(eb);
1414 return NULL;
1415 }
1416 btrfs_set_header_bytenr(eb_rewin, eb->start);
1417 btrfs_set_header_backref_rev(eb_rewin,
1418 btrfs_header_backref_rev(eb));
1419 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1420 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1421 } else {
1422 eb_rewin = btrfs_clone_extent_buffer(eb);
1423 if (!eb_rewin) {
1424 btrfs_tree_read_unlock_blocking(eb);
1425 free_extent_buffer(eb);
1426 return NULL;
1427 }
1428 }
1429
1430 btrfs_clear_path_blocking(path, NULL, BTRFS_READ_LOCK);
1431 btrfs_tree_read_unlock_blocking(eb);
1432 free_extent_buffer(eb);
1433
1434 extent_buffer_get(eb_rewin);
1435 btrfs_tree_read_lock(eb_rewin);
1436 __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1437 WARN_ON(btrfs_header_nritems(eb_rewin) >
1438 BTRFS_NODEPTRS_PER_BLOCK(fs_info->tree_root));
1439
1440 return eb_rewin;
1441}
1442
1443/*
1444 * get_old_root() rewinds the state of @root's root node to the given @time_seq
1445 * value. If there are no changes, the current root->root_node is returned. If
1446 * anything changed in between, there's a fresh buffer allocated on which the
1447 * rewind operations are done. In any case, the returned buffer is read locked.
1448 * Returns NULL on error (with no locks held).
1449 */
1450static inline struct extent_buffer *
1451get_old_root(struct btrfs_root *root, u64 time_seq)
1452{
1453 struct tree_mod_elem *tm;
1454 struct extent_buffer *eb = NULL;
1455 struct extent_buffer *eb_root;
1456 struct extent_buffer *old;
1457 struct tree_mod_root *old_root = NULL;
1458 u64 old_generation = 0;
1459 u64 logical;
1460 u32 blocksize;
1461
1462 eb_root = btrfs_read_lock_root_node(root);
1463 tm = __tree_mod_log_oldest_root(root->fs_info, eb_root, time_seq);
1464 if (!tm)
1465 return eb_root;
1466
1467 if (tm->op == MOD_LOG_ROOT_REPLACE) {
1468 old_root = &tm->old_root;
1469 old_generation = tm->generation;
1470 logical = old_root->logical;
1471 } else {
1472 logical = eb_root->start;
1473 }
1474
1475 tm = tree_mod_log_search(root->fs_info, logical, time_seq);
1476 if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1477 btrfs_tree_read_unlock(eb_root);
1478 free_extent_buffer(eb_root);
1479 blocksize = btrfs_level_size(root, old_root->level);
1480 old = read_tree_block(root, logical, blocksize, 0);
1481 if (WARN_ON(!old || !extent_buffer_uptodate(old))) {
1482 free_extent_buffer(old);
1483 btrfs_warn(root->fs_info,
1484 "failed to read tree block %llu from get_old_root", logical);
1485 } else {
1486 eb = btrfs_clone_extent_buffer(old);
1487 free_extent_buffer(old);
1488 }
1489 } else if (old_root) {
1490 btrfs_tree_read_unlock(eb_root);
1491 free_extent_buffer(eb_root);
1492 eb = alloc_dummy_extent_buffer(logical, root->nodesize);
1493 } else {
1494 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1495 eb = btrfs_clone_extent_buffer(eb_root);
1496 btrfs_tree_read_unlock_blocking(eb_root);
1497 free_extent_buffer(eb_root);
1498 }
1499
1500 if (!eb)
1501 return NULL;
1502 extent_buffer_get(eb);
1503 btrfs_tree_read_lock(eb);
1504 if (old_root) {
1505 btrfs_set_header_bytenr(eb, eb->start);
1506 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1507 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1508 btrfs_set_header_level(eb, old_root->level);
1509 btrfs_set_header_generation(eb, old_generation);
1510 }
1511 if (tm)
1512 __tree_mod_log_rewind(root->fs_info, eb, time_seq, tm);
1513 else
1514 WARN_ON(btrfs_header_level(eb) != 0);
1515 WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(root));
1516
1517 return eb;
1518}
1519
1520int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1521{
1522 struct tree_mod_elem *tm;
1523 int level;
1524 struct extent_buffer *eb_root = btrfs_root_node(root);
1525
1526 tm = __tree_mod_log_oldest_root(root->fs_info, eb_root, time_seq);
1527 if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1528 level = tm->old_root.level;
1529 } else {
1530 level = btrfs_header_level(eb_root);
1531 }
1532 free_extent_buffer(eb_root);
1533
1534 return level;
1535}
1536
1537static inline int should_cow_block(struct btrfs_trans_handle *trans,
1538 struct btrfs_root *root,
1539 struct extent_buffer *buf)
1540{
1541 /* ensure we can see the force_cow */
1542 smp_rmb();
1543
1544 /*
1545 * We do not need to cow a block if
1546 * 1) this block is not created or changed in this transaction;
1547 * 2) this block does not belong to TREE_RELOC tree;
1548 * 3) the root is not forced COW.
1549 *
1550 * What is forced COW:
1551 * when we create snapshot during commiting the transaction,
1552 * after we've finished coping src root, we must COW the shared
1553 * block to ensure the metadata consistency.
1554 */
1555 if (btrfs_header_generation(buf) == trans->transid &&
1556 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1557 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1558 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1559 !root->force_cow)
1560 return 0;
1561 return 1;
1562}
1563
1564/*
1565 * cows a single block, see __btrfs_cow_block for the real work.
1566 * This version of it has extra checks so that a block isn't cow'd more than
1567 * once per transaction, as long as it hasn't been written yet
1568 */
1569noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1570 struct btrfs_root *root, struct extent_buffer *buf,
1571 struct extent_buffer *parent, int parent_slot,
1572 struct extent_buffer **cow_ret)
1573{
1574 u64 search_start;
1575 int ret;
1576
1577 if (trans->transaction != root->fs_info->running_transaction)
1578 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1579 trans->transid,
1580 root->fs_info->running_transaction->transid);
1581
1582 if (trans->transid != root->fs_info->generation)
1583 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1584 trans->transid, root->fs_info->generation);
1585
1586 if (!should_cow_block(trans, root, buf)) {
1587 *cow_ret = buf;
1588 return 0;
1589 }
1590
1591 search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
1592
1593 if (parent)
1594 btrfs_set_lock_blocking(parent);
1595 btrfs_set_lock_blocking(buf);
1596
1597 ret = __btrfs_cow_block(trans, root, buf, parent,
1598 parent_slot, cow_ret, search_start, 0);
1599
1600 trace_btrfs_cow_block(root, buf, *cow_ret);
1601
1602 return ret;
1603}
1604
1605/*
1606 * helper function for defrag to decide if two blocks pointed to by a
1607 * node are actually close by
1608 */
1609static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1610{
1611 if (blocknr < other && other - (blocknr + blocksize) < 32768)
1612 return 1;
1613 if (blocknr > other && blocknr - (other + blocksize) < 32768)
1614 return 1;
1615 return 0;
1616}
1617
1618/*
1619 * compare two keys in a memcmp fashion
1620 */
1621static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
1622{
1623 struct btrfs_key k1;
1624
1625 btrfs_disk_key_to_cpu(&k1, disk);
1626
1627 return btrfs_comp_cpu_keys(&k1, k2);
1628}
1629
1630/*
1631 * same as comp_keys only with two btrfs_key's
1632 */
1633int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
1634{
1635 if (k1->objectid > k2->objectid)
1636 return 1;
1637 if (k1->objectid < k2->objectid)
1638 return -1;
1639 if (k1->type > k2->type)
1640 return 1;
1641 if (k1->type < k2->type)
1642 return -1;
1643 if (k1->offset > k2->offset)
1644 return 1;
1645 if (k1->offset < k2->offset)
1646 return -1;
1647 return 0;
1648}
1649
1650/*
1651 * this is used by the defrag code to go through all the
1652 * leaves pointed to by a node and reallocate them so that
1653 * disk order is close to key order
1654 */
1655int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1656 struct btrfs_root *root, struct extent_buffer *parent,
1657 int start_slot, u64 *last_ret,
1658 struct btrfs_key *progress)
1659{
1660 struct extent_buffer *cur;
1661 u64 blocknr;
1662 u64 gen;
1663 u64 search_start = *last_ret;
1664 u64 last_block = 0;
1665 u64 other;
1666 u32 parent_nritems;
1667 int end_slot;
1668 int i;
1669 int err = 0;
1670 int parent_level;
1671 int uptodate;
1672 u32 blocksize;
1673 int progress_passed = 0;
1674 struct btrfs_disk_key disk_key;
1675
1676 parent_level = btrfs_header_level(parent);
1677
1678 WARN_ON(trans->transaction != root->fs_info->running_transaction);
1679 WARN_ON(trans->transid != root->fs_info->generation);
1680
1681 parent_nritems = btrfs_header_nritems(parent);
1682 blocksize = btrfs_level_size(root, parent_level - 1);
1683 end_slot = parent_nritems;
1684
1685 if (parent_nritems == 1)
1686 return 0;
1687
1688 btrfs_set_lock_blocking(parent);
1689
1690 for (i = start_slot; i < end_slot; i++) {
1691 int close = 1;
1692
1693 btrfs_node_key(parent, &disk_key, i);
1694 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1695 continue;
1696
1697 progress_passed = 1;
1698 blocknr = btrfs_node_blockptr(parent, i);
1699 gen = btrfs_node_ptr_generation(parent, i);
1700 if (last_block == 0)
1701 last_block = blocknr;
1702
1703 if (i > 0) {
1704 other = btrfs_node_blockptr(parent, i - 1);
1705 close = close_blocks(blocknr, other, blocksize);
1706 }
1707 if (!close && i < end_slot - 2) {
1708 other = btrfs_node_blockptr(parent, i + 1);
1709 close = close_blocks(blocknr, other, blocksize);
1710 }
1711 if (close) {
1712 last_block = blocknr;
1713 continue;
1714 }
1715
1716 cur = btrfs_find_tree_block(root, blocknr, blocksize);
1717 if (cur)
1718 uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1719 else
1720 uptodate = 0;
1721 if (!cur || !uptodate) {
1722 if (!cur) {
1723 cur = read_tree_block(root, blocknr,
1724 blocksize, gen);
1725 if (!cur || !extent_buffer_uptodate(cur)) {
1726 free_extent_buffer(cur);
1727 return -EIO;
1728 }
1729 } else if (!uptodate) {
1730 err = btrfs_read_buffer(cur, gen);
1731 if (err) {
1732 free_extent_buffer(cur);
1733 return err;
1734 }
1735 }
1736 }
1737 if (search_start == 0)
1738 search_start = last_block;
1739
1740 btrfs_tree_lock(cur);
1741 btrfs_set_lock_blocking(cur);
1742 err = __btrfs_cow_block(trans, root, cur, parent, i,
1743 &cur, search_start,
1744 min(16 * blocksize,
1745 (end_slot - i) * blocksize));
1746 if (err) {
1747 btrfs_tree_unlock(cur);
1748 free_extent_buffer(cur);
1749 break;
1750 }
1751 search_start = cur->start;
1752 last_block = cur->start;
1753 *last_ret = search_start;
1754 btrfs_tree_unlock(cur);
1755 free_extent_buffer(cur);
1756 }
1757 return err;
1758}
1759
1760/*
1761 * The leaf data grows from end-to-front in the node.
1762 * this returns the address of the start of the last item,
1763 * which is the stop of the leaf data stack
1764 */
1765static inline unsigned int leaf_data_end(struct btrfs_root *root,
1766 struct extent_buffer *leaf)
1767{
1768 u32 nr = btrfs_header_nritems(leaf);
1769 if (nr == 0)
1770 return BTRFS_LEAF_DATA_SIZE(root);
1771 return btrfs_item_offset_nr(leaf, nr - 1);
1772}
1773
1774
1775/*
1776 * search for key in the extent_buffer. The items start at offset p,
1777 * and they are item_size apart. There are 'max' items in p.
1778 *
1779 * the slot in the array is returned via slot, and it points to
1780 * the place where you would insert key if it is not found in
1781 * the array.
1782 *
1783 * slot may point to max if the key is bigger than all of the keys
1784 */
1785static noinline int generic_bin_search(struct extent_buffer *eb,
1786 unsigned long p,
1787 int item_size, struct btrfs_key *key,
1788 int max, int *slot)
1789{
1790 int low = 0;
1791 int high = max;
1792 int mid;
1793 int ret;
1794 struct btrfs_disk_key *tmp = NULL;
1795 struct btrfs_disk_key unaligned;
1796 unsigned long offset;
1797 char *kaddr = NULL;
1798 unsigned long map_start = 0;
1799 unsigned long map_len = 0;
1800 int err;
1801
1802 while (low < high) {
1803 mid = (low + high) / 2;
1804 offset = p + mid * item_size;
1805
1806 if (!kaddr || offset < map_start ||
1807 (offset + sizeof(struct btrfs_disk_key)) >
1808 map_start + map_len) {
1809
1810 err = map_private_extent_buffer(eb, offset,
1811 sizeof(struct btrfs_disk_key),
1812 &kaddr, &map_start, &map_len);
1813
1814 if (!err) {
1815 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1816 map_start);
1817 } else {
1818 read_extent_buffer(eb, &unaligned,
1819 offset, sizeof(unaligned));
1820 tmp = &unaligned;
1821 }
1822
1823 } else {
1824 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1825 map_start);
1826 }
1827 ret = comp_keys(tmp, key);
1828
1829 if (ret < 0)
1830 low = mid + 1;
1831 else if (ret > 0)
1832 high = mid;
1833 else {
1834 *slot = mid;
1835 return 0;
1836 }
1837 }
1838 *slot = low;
1839 return 1;
1840}
1841
1842/*
1843 * simple bin_search frontend that does the right thing for
1844 * leaves vs nodes
1845 */
1846static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1847 int level, int *slot)
1848{
1849 if (level == 0)
1850 return generic_bin_search(eb,
1851 offsetof(struct btrfs_leaf, items),
1852 sizeof(struct btrfs_item),
1853 key, btrfs_header_nritems(eb),
1854 slot);
1855 else
1856 return generic_bin_search(eb,
1857 offsetof(struct btrfs_node, ptrs),
1858 sizeof(struct btrfs_key_ptr),
1859 key, btrfs_header_nritems(eb),
1860 slot);
1861}
1862
1863int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1864 int level, int *slot)
1865{
1866 return bin_search(eb, key, level, slot);
1867}
1868
1869static void root_add_used(struct btrfs_root *root, u32 size)
1870{
1871 spin_lock(&root->accounting_lock);
1872 btrfs_set_root_used(&root->root_item,
1873 btrfs_root_used(&root->root_item) + size);
1874 spin_unlock(&root->accounting_lock);
1875}
1876
1877static void root_sub_used(struct btrfs_root *root, u32 size)
1878{
1879 spin_lock(&root->accounting_lock);
1880 btrfs_set_root_used(&root->root_item,
1881 btrfs_root_used(&root->root_item) - size);
1882 spin_unlock(&root->accounting_lock);
1883}
1884
1885/* given a node and slot number, this reads the blocks it points to. The
1886 * extent buffer is returned with a reference taken (but unlocked).
1887 * NULL is returned on error.
1888 */
1889static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
1890 struct extent_buffer *parent, int slot)
1891{
1892 int level = btrfs_header_level(parent);
1893 struct extent_buffer *eb;
1894
1895 if (slot < 0)
1896 return NULL;
1897 if (slot >= btrfs_header_nritems(parent))
1898 return NULL;
1899
1900 BUG_ON(level == 0);
1901
1902 eb = read_tree_block(root, btrfs_node_blockptr(parent, slot),
1903 btrfs_level_size(root, level - 1),
1904 btrfs_node_ptr_generation(parent, slot));
1905 if (eb && !extent_buffer_uptodate(eb)) {
1906 free_extent_buffer(eb);
1907 eb = NULL;
1908 }
1909
1910 return eb;
1911}
1912
1913/*
1914 * node level balancing, used to make sure nodes are in proper order for
1915 * item deletion. We balance from the top down, so we have to make sure
1916 * that a deletion won't leave an node completely empty later on.
1917 */
1918static noinline int balance_level(struct btrfs_trans_handle *trans,
1919 struct btrfs_root *root,
1920 struct btrfs_path *path, int level)
1921{
1922 struct extent_buffer *right = NULL;
1923 struct extent_buffer *mid;
1924 struct extent_buffer *left = NULL;
1925 struct extent_buffer *parent = NULL;
1926 int ret = 0;
1927 int wret;
1928 int pslot;
1929 int orig_slot = path->slots[level];
1930 u64 orig_ptr;
1931
1932 if (level == 0)
1933 return 0;
1934
1935 mid = path->nodes[level];
1936
1937 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1938 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1939 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1940
1941 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1942
1943 if (level < BTRFS_MAX_LEVEL - 1) {
1944 parent = path->nodes[level + 1];
1945 pslot = path->slots[level + 1];
1946 }
1947
1948 /*
1949 * deal with the case where there is only one pointer in the root
1950 * by promoting the node below to a root
1951 */
1952 if (!parent) {
1953 struct extent_buffer *child;
1954
1955 if (btrfs_header_nritems(mid) != 1)
1956 return 0;
1957
1958 /* promote the child to a root */
1959 child = read_node_slot(root, mid, 0);
1960 if (!child) {
1961 ret = -EROFS;
1962 btrfs_std_error(root->fs_info, ret);
1963 goto enospc;
1964 }
1965
1966 btrfs_tree_lock(child);
1967 btrfs_set_lock_blocking(child);
1968 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1969 if (ret) {
1970 btrfs_tree_unlock(child);
1971 free_extent_buffer(child);
1972 goto enospc;
1973 }
1974
1975 tree_mod_log_set_root_pointer(root, child, 1);
1976 rcu_assign_pointer(root->node, child);
1977
1978 add_root_to_dirty_list(root);
1979 btrfs_tree_unlock(child);
1980
1981 path->locks[level] = 0;
1982 path->nodes[level] = NULL;
1983 clean_tree_block(trans, root, mid);
1984 btrfs_tree_unlock(mid);
1985 /* once for the path */
1986 free_extent_buffer(mid);
1987
1988 root_sub_used(root, mid->len);
1989 btrfs_free_tree_block(trans, root, mid, 0, 1);
1990 /* once for the root ptr */
1991 free_extent_buffer_stale(mid);
1992 return 0;
1993 }
1994 if (btrfs_header_nritems(mid) >
1995 BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
1996 return 0;
1997
1998 left = read_node_slot(root, parent, pslot - 1);
1999 if (left) {
2000 btrfs_tree_lock(left);
2001 btrfs_set_lock_blocking(left);
2002 wret = btrfs_cow_block(trans, root, left,
2003 parent, pslot - 1, &left);
2004 if (wret) {
2005 ret = wret;
2006 goto enospc;
2007 }
2008 }
2009 right = read_node_slot(root, parent, pslot + 1);
2010 if (right) {
2011 btrfs_tree_lock(right);
2012 btrfs_set_lock_blocking(right);
2013 wret = btrfs_cow_block(trans, root, right,
2014 parent, pslot + 1, &right);
2015 if (wret) {
2016 ret = wret;
2017 goto enospc;
2018 }
2019 }
2020
2021 /* first, try to make some room in the middle buffer */
2022 if (left) {
2023 orig_slot += btrfs_header_nritems(left);
2024 wret = push_node_left(trans, root, left, mid, 1);
2025 if (wret < 0)
2026 ret = wret;
2027 }
2028
2029 /*
2030 * then try to empty the right most buffer into the middle
2031 */
2032 if (right) {
2033 wret = push_node_left(trans, root, mid, right, 1);
2034 if (wret < 0 && wret != -ENOSPC)
2035 ret = wret;
2036 if (btrfs_header_nritems(right) == 0) {
2037 clean_tree_block(trans, root, right);
2038 btrfs_tree_unlock(right);
2039 del_ptr(root, path, level + 1, pslot + 1);
2040 root_sub_used(root, right->len);
2041 btrfs_free_tree_block(trans, root, right, 0, 1);
2042 free_extent_buffer_stale(right);
2043 right = NULL;
2044 } else {
2045 struct btrfs_disk_key right_key;
2046 btrfs_node_key(right, &right_key, 0);
2047 tree_mod_log_set_node_key(root->fs_info, parent,
2048 pslot + 1, 0);
2049 btrfs_set_node_key(parent, &right_key, pslot + 1);
2050 btrfs_mark_buffer_dirty(parent);
2051 }
2052 }
2053 if (btrfs_header_nritems(mid) == 1) {
2054 /*
2055 * we're not allowed to leave a node with one item in the
2056 * tree during a delete. A deletion from lower in the tree
2057 * could try to delete the only pointer in this node.
2058 * So, pull some keys from the left.
2059 * There has to be a left pointer at this point because
2060 * otherwise we would have pulled some pointers from the
2061 * right
2062 */
2063 if (!left) {
2064 ret = -EROFS;
2065 btrfs_std_error(root->fs_info, ret);
2066 goto enospc;
2067 }
2068 wret = balance_node_right(trans, root, mid, left);
2069 if (wret < 0) {
2070 ret = wret;
2071 goto enospc;
2072 }
2073 if (wret == 1) {
2074 wret = push_node_left(trans, root, left, mid, 1);
2075 if (wret < 0)
2076 ret = wret;
2077 }
2078 BUG_ON(wret == 1);
2079 }
2080 if (btrfs_header_nritems(mid) == 0) {
2081 clean_tree_block(trans, root, mid);
2082 btrfs_tree_unlock(mid);
2083 del_ptr(root, path, level + 1, pslot);
2084 root_sub_used(root, mid->len);
2085 btrfs_free_tree_block(trans, root, mid, 0, 1);
2086 free_extent_buffer_stale(mid);
2087 mid = NULL;
2088 } else {
2089 /* update the parent key to reflect our changes */
2090 struct btrfs_disk_key mid_key;
2091 btrfs_node_key(mid, &mid_key, 0);
2092 tree_mod_log_set_node_key(root->fs_info, parent,
2093 pslot, 0);
2094 btrfs_set_node_key(parent, &mid_key, pslot);
2095 btrfs_mark_buffer_dirty(parent);
2096 }
2097
2098 /* update the path */
2099 if (left) {
2100 if (btrfs_header_nritems(left) > orig_slot) {
2101 extent_buffer_get(left);
2102 /* left was locked after cow */
2103 path->nodes[level] = left;
2104 path->slots[level + 1] -= 1;
2105 path->slots[level] = orig_slot;
2106 if (mid) {
2107 btrfs_tree_unlock(mid);
2108 free_extent_buffer(mid);
2109 }
2110 } else {
2111 orig_slot -= btrfs_header_nritems(left);
2112 path->slots[level] = orig_slot;
2113 }
2114 }
2115 /* double check we haven't messed things up */
2116 if (orig_ptr !=
2117 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2118 BUG();
2119enospc:
2120 if (right) {
2121 btrfs_tree_unlock(right);
2122 free_extent_buffer(right);
2123 }
2124 if (left) {
2125 if (path->nodes[level] != left)
2126 btrfs_tree_unlock(left);
2127 free_extent_buffer(left);
2128 }
2129 return ret;
2130}
2131
2132/* Node balancing for insertion. Here we only split or push nodes around
2133 * when they are completely full. This is also done top down, so we
2134 * have to be pessimistic.
2135 */
2136static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2137 struct btrfs_root *root,
2138 struct btrfs_path *path, int level)
2139{
2140 struct extent_buffer *right = NULL;
2141 struct extent_buffer *mid;
2142 struct extent_buffer *left = NULL;
2143 struct extent_buffer *parent = NULL;
2144 int ret = 0;
2145 int wret;
2146 int pslot;
2147 int orig_slot = path->slots[level];
2148
2149 if (level == 0)
2150 return 1;
2151
2152 mid = path->nodes[level];
2153 WARN_ON(btrfs_header_generation(mid) != trans->transid);
2154
2155 if (level < BTRFS_MAX_LEVEL - 1) {
2156 parent = path->nodes[level + 1];
2157 pslot = path->slots[level + 1];
2158 }
2159
2160 if (!parent)
2161 return 1;
2162
2163 left = read_node_slot(root, parent, pslot - 1);
2164
2165 /* first, try to make some room in the middle buffer */
2166 if (left) {
2167 u32 left_nr;
2168
2169 btrfs_tree_lock(left);
2170 btrfs_set_lock_blocking(left);
2171
2172 left_nr = btrfs_header_nritems(left);
2173 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
2174 wret = 1;
2175 } else {
2176 ret = btrfs_cow_block(trans, root, left, parent,
2177 pslot - 1, &left);
2178 if (ret)
2179 wret = 1;
2180 else {
2181 wret = push_node_left(trans, root,
2182 left, mid, 0);
2183 }
2184 }
2185 if (wret < 0)
2186 ret = wret;
2187 if (wret == 0) {
2188 struct btrfs_disk_key disk_key;
2189 orig_slot += left_nr;
2190 btrfs_node_key(mid, &disk_key, 0);
2191 tree_mod_log_set_node_key(root->fs_info, parent,
2192 pslot, 0);
2193 btrfs_set_node_key(parent, &disk_key, pslot);
2194 btrfs_mark_buffer_dirty(parent);
2195 if (btrfs_header_nritems(left) > orig_slot) {
2196 path->nodes[level] = left;
2197 path->slots[level + 1] -= 1;
2198 path->slots[level] = orig_slot;
2199 btrfs_tree_unlock(mid);
2200 free_extent_buffer(mid);
2201 } else {
2202 orig_slot -=
2203 btrfs_header_nritems(left);
2204 path->slots[level] = orig_slot;
2205 btrfs_tree_unlock(left);
2206 free_extent_buffer(left);
2207 }
2208 return 0;
2209 }
2210 btrfs_tree_unlock(left);
2211 free_extent_buffer(left);
2212 }
2213 right = read_node_slot(root, parent, pslot + 1);
2214
2215 /*
2216 * then try to empty the right most buffer into the middle
2217 */
2218 if (right) {
2219 u32 right_nr;
2220
2221 btrfs_tree_lock(right);
2222 btrfs_set_lock_blocking(right);
2223
2224 right_nr = btrfs_header_nritems(right);
2225 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
2226 wret = 1;
2227 } else {
2228 ret = btrfs_cow_block(trans, root, right,
2229 parent, pslot + 1,
2230 &right);
2231 if (ret)
2232 wret = 1;
2233 else {
2234 wret = balance_node_right(trans, root,
2235 right, mid);
2236 }
2237 }
2238 if (wret < 0)
2239 ret = wret;
2240 if (wret == 0) {
2241 struct btrfs_disk_key disk_key;
2242
2243 btrfs_node_key(right, &disk_key, 0);
2244 tree_mod_log_set_node_key(root->fs_info, parent,
2245 pslot + 1, 0);
2246 btrfs_set_node_key(parent, &disk_key, pslot + 1);
2247 btrfs_mark_buffer_dirty(parent);
2248
2249 if (btrfs_header_nritems(mid) <= orig_slot) {
2250 path->nodes[level] = right;
2251 path->slots[level + 1] += 1;
2252 path->slots[level] = orig_slot -
2253 btrfs_header_nritems(mid);
2254 btrfs_tree_unlock(mid);
2255 free_extent_buffer(mid);
2256 } else {
2257 btrfs_tree_unlock(right);
2258 free_extent_buffer(right);
2259 }
2260 return 0;
2261 }
2262 btrfs_tree_unlock(right);
2263 free_extent_buffer(right);
2264 }
2265 return 1;
2266}
2267
2268/*
2269 * readahead one full node of leaves, finding things that are close
2270 * to the block in 'slot', and triggering ra on them.
2271 */
2272static void reada_for_search(struct btrfs_root *root,
2273 struct btrfs_path *path,
2274 int level, int slot, u64 objectid)
2275{
2276 struct extent_buffer *node;
2277 struct btrfs_disk_key disk_key;
2278 u32 nritems;
2279 u64 search;
2280 u64 target;
2281 u64 nread = 0;
2282 u64 gen;
2283 int direction = path->reada;
2284 struct extent_buffer *eb;
2285 u32 nr;
2286 u32 blocksize;
2287 u32 nscan = 0;
2288
2289 if (level != 1)
2290 return;
2291
2292 if (!path->nodes[level])
2293 return;
2294
2295 node = path->nodes[level];
2296
2297 search = btrfs_node_blockptr(node, slot);
2298 blocksize = btrfs_level_size(root, level - 1);
2299 eb = btrfs_find_tree_block(root, search, blocksize);
2300 if (eb) {
2301 free_extent_buffer(eb);
2302 return;
2303 }
2304
2305 target = search;
2306
2307 nritems = btrfs_header_nritems(node);
2308 nr = slot;
2309
2310 while (1) {
2311 if (direction < 0) {
2312 if (nr == 0)
2313 break;
2314 nr--;
2315 } else if (direction > 0) {
2316 nr++;
2317 if (nr >= nritems)
2318 break;
2319 }
2320 if (path->reada < 0 && objectid) {
2321 btrfs_node_key(node, &disk_key, nr);
2322 if (btrfs_disk_key_objectid(&disk_key) != objectid)
2323 break;
2324 }
2325 search = btrfs_node_blockptr(node, nr);
2326 if ((search <= target && target - search <= 65536) ||
2327 (search > target && search - target <= 65536)) {
2328 gen = btrfs_node_ptr_generation(node, nr);
2329 readahead_tree_block(root, search, blocksize, gen);
2330 nread += blocksize;
2331 }
2332 nscan++;
2333 if ((nread > 65536 || nscan > 32))
2334 break;
2335 }
2336}
2337
2338static noinline void reada_for_balance(struct btrfs_root *root,
2339 struct btrfs_path *path, int level)
2340{
2341 int slot;
2342 int nritems;
2343 struct extent_buffer *parent;
2344 struct extent_buffer *eb;
2345 u64 gen;
2346 u64 block1 = 0;
2347 u64 block2 = 0;
2348 int blocksize;
2349
2350 parent = path->nodes[level + 1];
2351 if (!parent)
2352 return;
2353
2354 nritems = btrfs_header_nritems(parent);
2355 slot = path->slots[level + 1];
2356 blocksize = btrfs_level_size(root, level);
2357
2358 if (slot > 0) {
2359 block1 = btrfs_node_blockptr(parent, slot - 1);
2360 gen = btrfs_node_ptr_generation(parent, slot - 1);
2361 eb = btrfs_find_tree_block(root, block1, blocksize);
2362 /*
2363 * if we get -eagain from btrfs_buffer_uptodate, we
2364 * don't want to return eagain here. That will loop
2365 * forever
2366 */
2367 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2368 block1 = 0;
2369 free_extent_buffer(eb);
2370 }
2371 if (slot + 1 < nritems) {
2372 block2 = btrfs_node_blockptr(parent, slot + 1);
2373 gen = btrfs_node_ptr_generation(parent, slot + 1);
2374 eb = btrfs_find_tree_block(root, block2, blocksize);
2375 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2376 block2 = 0;
2377 free_extent_buffer(eb);
2378 }
2379
2380 if (block1)
2381 readahead_tree_block(root, block1, blocksize, 0);
2382 if (block2)
2383 readahead_tree_block(root, block2, blocksize, 0);
2384}
2385
2386
2387/*
2388 * when we walk down the tree, it is usually safe to unlock the higher layers
2389 * in the tree. The exceptions are when our path goes through slot 0, because
2390 * operations on the tree might require changing key pointers higher up in the
2391 * tree.
2392 *
2393 * callers might also have set path->keep_locks, which tells this code to keep
2394 * the lock if the path points to the last slot in the block. This is part of
2395 * walking through the tree, and selecting the next slot in the higher block.
2396 *
2397 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
2398 * if lowest_unlock is 1, level 0 won't be unlocked
2399 */
2400static noinline void unlock_up(struct btrfs_path *path, int level,
2401 int lowest_unlock, int min_write_lock_level,
2402 int *write_lock_level)
2403{
2404 int i;
2405 int skip_level = level;
2406 int no_skips = 0;
2407 struct extent_buffer *t;
2408
2409 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2410 if (!path->nodes[i])
2411 break;
2412 if (!path->locks[i])
2413 break;
2414 if (!no_skips && path->slots[i] == 0) {
2415 skip_level = i + 1;
2416 continue;
2417 }
2418 if (!no_skips && path->keep_locks) {
2419 u32 nritems;
2420 t = path->nodes[i];
2421 nritems = btrfs_header_nritems(t);
2422 if (nritems < 1 || path->slots[i] >= nritems - 1) {
2423 skip_level = i + 1;
2424 continue;
2425 }
2426 }
2427 if (skip_level < i && i >= lowest_unlock)
2428 no_skips = 1;
2429
2430 t = path->nodes[i];
2431 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2432 btrfs_tree_unlock_rw(t, path->locks[i]);
2433 path->locks[i] = 0;
2434 if (write_lock_level &&
2435 i > min_write_lock_level &&
2436 i <= *write_lock_level) {
2437 *write_lock_level = i - 1;
2438 }
2439 }
2440 }
2441}
2442
2443/*
2444 * This releases any locks held in the path starting at level and
2445 * going all the way up to the root.
2446 *
2447 * btrfs_search_slot will keep the lock held on higher nodes in a few
2448 * corner cases, such as COW of the block at slot zero in the node. This
2449 * ignores those rules, and it should only be called when there are no
2450 * more updates to be done higher up in the tree.
2451 */
2452noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2453{
2454 int i;
2455
2456 if (path->keep_locks)
2457 return;
2458
2459 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2460 if (!path->nodes[i])
2461 continue;
2462 if (!path->locks[i])
2463 continue;
2464 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2465 path->locks[i] = 0;
2466 }
2467}
2468
2469/*
2470 * helper function for btrfs_search_slot. The goal is to find a block
2471 * in cache without setting the path to blocking. If we find the block
2472 * we return zero and the path is unchanged.
2473 *
2474 * If we can't find the block, we set the path blocking and do some
2475 * reada. -EAGAIN is returned and the search must be repeated.
2476 */
2477static int
2478read_block_for_search(struct btrfs_trans_handle *trans,
2479 struct btrfs_root *root, struct btrfs_path *p,
2480 struct extent_buffer **eb_ret, int level, int slot,
2481 struct btrfs_key *key, u64 time_seq)
2482{
2483 u64 blocknr;
2484 u64 gen;
2485 u32 blocksize;
2486 struct extent_buffer *b = *eb_ret;
2487 struct extent_buffer *tmp;
2488 int ret;
2489
2490 blocknr = btrfs_node_blockptr(b, slot);
2491 gen = btrfs_node_ptr_generation(b, slot);
2492 blocksize = btrfs_level_size(root, level - 1);
2493
2494 tmp = btrfs_find_tree_block(root, blocknr, blocksize);
2495 if (tmp) {
2496 /* first we do an atomic uptodate check */
2497 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2498 *eb_ret = tmp;
2499 return 0;
2500 }
2501
2502 /* the pages were up to date, but we failed
2503 * the generation number check. Do a full
2504 * read for the generation number that is correct.
2505 * We must do this without dropping locks so
2506 * we can trust our generation number
2507 */
2508 btrfs_set_path_blocking(p);
2509
2510 /* now we're allowed to do a blocking uptodate check */
2511 ret = btrfs_read_buffer(tmp, gen);
2512 if (!ret) {
2513 *eb_ret = tmp;
2514 return 0;
2515 }
2516 free_extent_buffer(tmp);
2517 btrfs_release_path(p);
2518 return -EIO;
2519 }
2520
2521 /*
2522 * reduce lock contention at high levels
2523 * of the btree by dropping locks before
2524 * we read. Don't release the lock on the current
2525 * level because we need to walk this node to figure
2526 * out which blocks to read.
2527 */
2528 btrfs_unlock_up_safe(p, level + 1);
2529 btrfs_set_path_blocking(p);
2530
2531 free_extent_buffer(tmp);
2532 if (p->reada)
2533 reada_for_search(root, p, level, slot, key->objectid);
2534
2535 btrfs_release_path(p);
2536
2537 ret = -EAGAIN;
2538 tmp = read_tree_block(root, blocknr, blocksize, 0);
2539 if (tmp) {
2540 /*
2541 * If the read above didn't mark this buffer up to date,
2542 * it will never end up being up to date. Set ret to EIO now
2543 * and give up so that our caller doesn't loop forever
2544 * on our EAGAINs.
2545 */
2546 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2547 ret = -EIO;
2548 free_extent_buffer(tmp);
2549 }
2550 return ret;
2551}
2552
2553/*
2554 * helper function for btrfs_search_slot. This does all of the checks
2555 * for node-level blocks and does any balancing required based on
2556 * the ins_len.
2557 *
2558 * If no extra work was required, zero is returned. If we had to
2559 * drop the path, -EAGAIN is returned and btrfs_search_slot must
2560 * start over
2561 */
2562static int
2563setup_nodes_for_search(struct btrfs_trans_handle *trans,
2564 struct btrfs_root *root, struct btrfs_path *p,
2565 struct extent_buffer *b, int level, int ins_len,
2566 int *write_lock_level)
2567{
2568 int ret;
2569 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2570 BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
2571 int sret;
2572
2573 if (*write_lock_level < level + 1) {
2574 *write_lock_level = level + 1;
2575 btrfs_release_path(p);
2576 goto again;
2577 }
2578
2579 btrfs_set_path_blocking(p);
2580 reada_for_balance(root, p, level);
2581 sret = split_node(trans, root, p, level);
2582 btrfs_clear_path_blocking(p, NULL, 0);
2583
2584 BUG_ON(sret > 0);
2585 if (sret) {
2586 ret = sret;
2587 goto done;
2588 }
2589 b = p->nodes[level];
2590 } else if (ins_len < 0 && btrfs_header_nritems(b) <
2591 BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
2592 int sret;
2593
2594 if (*write_lock_level < level + 1) {
2595 *write_lock_level = level + 1;
2596 btrfs_release_path(p);
2597 goto again;
2598 }
2599
2600 btrfs_set_path_blocking(p);
2601 reada_for_balance(root, p, level);
2602 sret = balance_level(trans, root, p, level);
2603 btrfs_clear_path_blocking(p, NULL, 0);
2604
2605 if (sret) {
2606 ret = sret;
2607 goto done;
2608 }
2609 b = p->nodes[level];
2610 if (!b) {
2611 btrfs_release_path(p);
2612 goto again;
2613 }
2614 BUG_ON(btrfs_header_nritems(b) == 1);
2615 }
2616 return 0;
2617
2618again:
2619 ret = -EAGAIN;
2620done:
2621 return ret;
2622}
2623
2624static void key_search_validate(struct extent_buffer *b,
2625 struct btrfs_key *key,
2626 int level)
2627{
2628#ifdef CONFIG_BTRFS_ASSERT
2629 struct btrfs_disk_key disk_key;
2630
2631 btrfs_cpu_key_to_disk(&disk_key, key);
2632
2633 if (level == 0)
2634 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2635 offsetof(struct btrfs_leaf, items[0].key),
2636 sizeof(disk_key)));
2637 else
2638 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2639 offsetof(struct btrfs_node, ptrs[0].key),
2640 sizeof(disk_key)));
2641#endif
2642}
2643
2644static int key_search(struct extent_buffer *b, struct btrfs_key *key,
2645 int level, int *prev_cmp, int *slot)
2646{
2647 if (*prev_cmp != 0) {
2648 *prev_cmp = bin_search(b, key, level, slot);
2649 return *prev_cmp;
2650 }
2651
2652 key_search_validate(b, key, level);
2653 *slot = 0;
2654
2655 return 0;
2656}
2657
2658int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *found_path,
2659 u64 iobjectid, u64 ioff, u8 key_type,
2660 struct btrfs_key *found_key)
2661{
2662 int ret;
2663 struct btrfs_key key;
2664 struct extent_buffer *eb;
2665 struct btrfs_path *path;
2666
2667 key.type = key_type;
2668 key.objectid = iobjectid;
2669 key.offset = ioff;
2670
2671 if (found_path == NULL) {
2672 path = btrfs_alloc_path();
2673 if (!path)
2674 return -ENOMEM;
2675 } else
2676 path = found_path;
2677
2678 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2679 if ((ret < 0) || (found_key == NULL)) {
2680 if (path != found_path)
2681 btrfs_free_path(path);
2682 return ret;
2683 }
2684
2685 eb = path->nodes[0];
2686 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2687 ret = btrfs_next_leaf(fs_root, path);
2688 if (ret)
2689 return ret;
2690 eb = path->nodes[0];
2691 }
2692
2693 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2694 if (found_key->type != key.type ||
2695 found_key->objectid != key.objectid)
2696 return 1;
2697
2698 return 0;
2699}
2700
2701/*
2702 * look for key in the tree. path is filled in with nodes along the way
2703 * if key is found, we return zero and you can find the item in the leaf
2704 * level of the path (level 0)
2705 *
2706 * If the key isn't found, the path points to the slot where it should
2707 * be inserted, and 1 is returned. If there are other errors during the
2708 * search a negative error number is returned.
2709 *
2710 * if ins_len > 0, nodes and leaves will be split as we walk down the
2711 * tree. if ins_len < 0, nodes will be merged as we walk down the tree (if
2712 * possible)
2713 */
2714int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
2715 *root, struct btrfs_key *key, struct btrfs_path *p, int
2716 ins_len, int cow)
2717{
2718 struct extent_buffer *b;
2719 int slot;
2720 int ret;
2721 int err;
2722 int level;
2723 int lowest_unlock = 1;
2724 int root_lock;
2725 /* everything at write_lock_level or lower must be write locked */
2726 int write_lock_level = 0;
2727 u8 lowest_level = 0;
2728 int min_write_lock_level;
2729 int prev_cmp;
2730
2731 lowest_level = p->lowest_level;
2732 WARN_ON(lowest_level && ins_len > 0);
2733 WARN_ON(p->nodes[0] != NULL);
2734 BUG_ON(!cow && ins_len);
2735
2736 if (ins_len < 0) {
2737 lowest_unlock = 2;
2738
2739 /* when we are removing items, we might have to go up to level
2740 * two as we update tree pointers Make sure we keep write
2741 * for those levels as well
2742 */
2743 write_lock_level = 2;
2744 } else if (ins_len > 0) {
2745 /*
2746 * for inserting items, make sure we have a write lock on
2747 * level 1 so we can update keys
2748 */
2749 write_lock_level = 1;
2750 }
2751
2752 if (!cow)
2753 write_lock_level = -1;
2754
2755 if (cow && (p->keep_locks || p->lowest_level))
2756 write_lock_level = BTRFS_MAX_LEVEL;
2757
2758 min_write_lock_level = write_lock_level;
2759
2760again:
2761 prev_cmp = -1;
2762 /*
2763 * we try very hard to do read locks on the root
2764 */
2765 root_lock = BTRFS_READ_LOCK;
2766 level = 0;
2767 if (p->search_commit_root) {
2768 /*
2769 * the commit roots are read only
2770 * so we always do read locks
2771 */
2772 if (p->need_commit_sem)
2773 down_read(&root->fs_info->commit_root_sem);
2774 b = root->commit_root;
2775 extent_buffer_get(b);
2776 level = btrfs_header_level(b);
2777 if (p->need_commit_sem)
2778 up_read(&root->fs_info->commit_root_sem);
2779 if (!p->skip_locking)
2780 btrfs_tree_read_lock(b);
2781 } else {
2782 if (p->skip_locking) {
2783 b = btrfs_root_node(root);
2784 level = btrfs_header_level(b);
2785 } else {
2786 /* we don't know the level of the root node
2787 * until we actually have it read locked
2788 */
2789 b = btrfs_read_lock_root_node(root);
2790 level = btrfs_header_level(b);
2791 if (level <= write_lock_level) {
2792 /* whoops, must trade for write lock */
2793 btrfs_tree_read_unlock(b);
2794 free_extent_buffer(b);
2795 b = btrfs_lock_root_node(root);
2796 root_lock = BTRFS_WRITE_LOCK;
2797
2798 /* the level might have changed, check again */
2799 level = btrfs_header_level(b);
2800 }
2801 }
2802 }
2803 p->nodes[level] = b;
2804 if (!p->skip_locking)
2805 p->locks[level] = root_lock;
2806
2807 while (b) {
2808 level = btrfs_header_level(b);
2809
2810 /*
2811 * setup the path here so we can release it under lock
2812 * contention with the cow code
2813 */
2814 if (cow) {
2815 /*
2816 * if we don't really need to cow this block
2817 * then we don't want to set the path blocking,
2818 * so we test it here
2819 */
2820 if (!should_cow_block(trans, root, b))
2821 goto cow_done;
2822
2823 btrfs_set_path_blocking(p);
2824
2825 /*
2826 * must have write locks on this node and the
2827 * parent
2828 */
2829 if (level > write_lock_level ||
2830 (level + 1 > write_lock_level &&
2831 level + 1 < BTRFS_MAX_LEVEL &&
2832 p->nodes[level + 1])) {
2833 write_lock_level = level + 1;
2834 btrfs_release_path(p);
2835 goto again;
2836 }
2837
2838 err = btrfs_cow_block(trans, root, b,
2839 p->nodes[level + 1],
2840 p->slots[level + 1], &b);
2841 if (err) {
2842 ret = err;
2843 goto done;
2844 }
2845 }
2846cow_done:
2847 p->nodes[level] = b;
2848 btrfs_clear_path_blocking(p, NULL, 0);
2849
2850 /*
2851 * we have a lock on b and as long as we aren't changing
2852 * the tree, there is no way to for the items in b to change.
2853 * It is safe to drop the lock on our parent before we
2854 * go through the expensive btree search on b.
2855 *
2856 * If we're inserting or deleting (ins_len != 0), then we might
2857 * be changing slot zero, which may require changing the parent.
2858 * So, we can't drop the lock until after we know which slot
2859 * we're operating on.
2860 */
2861 if (!ins_len && !p->keep_locks) {
2862 int u = level + 1;
2863
2864 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2865 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2866 p->locks[u] = 0;
2867 }
2868 }
2869
2870 ret = key_search(b, key, level, &prev_cmp, &slot);
2871
2872 if (level != 0) {
2873 int dec = 0;
2874 if (ret && slot > 0) {
2875 dec = 1;
2876 slot -= 1;
2877 }
2878 p->slots[level] = slot;
2879 err = setup_nodes_for_search(trans, root, p, b, level,
2880 ins_len, &write_lock_level);
2881 if (err == -EAGAIN)
2882 goto again;
2883 if (err) {
2884 ret = err;
2885 goto done;
2886 }
2887 b = p->nodes[level];
2888 slot = p->slots[level];
2889
2890 /*
2891 * slot 0 is special, if we change the key
2892 * we have to update the parent pointer
2893 * which means we must have a write lock
2894 * on the parent
2895 */
2896 if (slot == 0 && ins_len &&
2897 write_lock_level < level + 1) {
2898 write_lock_level = level + 1;
2899 btrfs_release_path(p);
2900 goto again;
2901 }
2902
2903 unlock_up(p, level, lowest_unlock,
2904 min_write_lock_level, &write_lock_level);
2905
2906 if (level == lowest_level) {
2907 if (dec)
2908 p->slots[level]++;
2909 goto done;
2910 }
2911
2912 err = read_block_for_search(trans, root, p,
2913 &b, level, slot, key, 0);
2914 if (err == -EAGAIN)
2915 goto again;
2916 if (err) {
2917 ret = err;
2918 goto done;
2919 }
2920
2921 if (!p->skip_locking) {
2922 level = btrfs_header_level(b);
2923 if (level <= write_lock_level) {
2924 err = btrfs_try_tree_write_lock(b);
2925 if (!err) {
2926 btrfs_set_path_blocking(p);
2927 btrfs_tree_lock(b);
2928 btrfs_clear_path_blocking(p, b,
2929 BTRFS_WRITE_LOCK);
2930 }
2931 p->locks[level] = BTRFS_WRITE_LOCK;
2932 } else {
2933 err = btrfs_try_tree_read_lock(b);
2934 if (!err) {
2935 btrfs_set_path_blocking(p);
2936 btrfs_tree_read_lock(b);
2937 btrfs_clear_path_blocking(p, b,
2938 BTRFS_READ_LOCK);
2939 }
2940 p->locks[level] = BTRFS_READ_LOCK;
2941 }
2942 p->nodes[level] = b;
2943 }
2944 } else {
2945 p->slots[level] = slot;
2946 if (ins_len > 0 &&
2947 btrfs_leaf_free_space(root, b) < ins_len) {
2948 if (write_lock_level < 1) {
2949 write_lock_level = 1;
2950 btrfs_release_path(p);
2951 goto again;
2952 }
2953
2954 btrfs_set_path_blocking(p);
2955 err = split_leaf(trans, root, key,
2956 p, ins_len, ret == 0);
2957 btrfs_clear_path_blocking(p, NULL, 0);
2958
2959 BUG_ON(err > 0);
2960 if (err) {
2961 ret = err;
2962 goto done;
2963 }
2964 }
2965 if (!p->search_for_split)
2966 unlock_up(p, level, lowest_unlock,
2967 min_write_lock_level, &write_lock_level);
2968 goto done;
2969 }
2970 }
2971 ret = 1;
2972done:
2973 /*
2974 * we don't really know what they plan on doing with the path
2975 * from here on, so for now just mark it as blocking
2976 */
2977 if (!p->leave_spinning)
2978 btrfs_set_path_blocking(p);
2979 if (ret < 0)
2980 btrfs_release_path(p);
2981 return ret;
2982}
2983
2984/*
2985 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2986 * current state of the tree together with the operations recorded in the tree
2987 * modification log to search for the key in a previous version of this tree, as
2988 * denoted by the time_seq parameter.
2989 *
2990 * Naturally, there is no support for insert, delete or cow operations.
2991 *
2992 * The resulting path and return value will be set up as if we called
2993 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2994 */
2995int btrfs_search_old_slot(struct btrfs_root *root, struct btrfs_key *key,
2996 struct btrfs_path *p, u64 time_seq)
2997{
2998 struct extent_buffer *b;
2999 int slot;
3000 int ret;
3001 int err;
3002 int level;
3003 int lowest_unlock = 1;
3004 u8 lowest_level = 0;
3005 int prev_cmp = -1;
3006
3007 lowest_level = p->lowest_level;
3008 WARN_ON(p->nodes[0] != NULL);
3009
3010 if (p->search_commit_root) {
3011 BUG_ON(time_seq);
3012 return btrfs_search_slot(NULL, root, key, p, 0, 0);
3013 }
3014
3015again:
3016 b = get_old_root(root, time_seq);
3017 level = btrfs_header_level(b);
3018 p->locks[level] = BTRFS_READ_LOCK;
3019
3020 while (b) {
3021 level = btrfs_header_level(b);
3022 p->nodes[level] = b;
3023 btrfs_clear_path_blocking(p, NULL, 0);
3024
3025 /*
3026 * we have a lock on b and as long as we aren't changing
3027 * the tree, there is no way to for the items in b to change.
3028 * It is safe to drop the lock on our parent before we
3029 * go through the expensive btree search on b.
3030 */
3031 btrfs_unlock_up_safe(p, level + 1);
3032
3033 /*
3034 * Since we can unwind eb's we want to do a real search every
3035 * time.
3036 */
3037 prev_cmp = -1;
3038 ret = key_search(b, key, level, &prev_cmp, &slot);
3039
3040 if (level != 0) {
3041 int dec = 0;
3042 if (ret && slot > 0) {
3043 dec = 1;
3044 slot -= 1;
3045 }
3046 p->slots[level] = slot;
3047 unlock_up(p, level, lowest_unlock, 0, NULL);
3048
3049 if (level == lowest_level) {
3050 if (dec)
3051 p->slots[level]++;
3052 goto done;
3053 }
3054
3055 err = read_block_for_search(NULL, root, p, &b, level,
3056 slot, key, time_seq);
3057 if (err == -EAGAIN)
3058 goto again;
3059 if (err) {
3060 ret = err;
3061 goto done;
3062 }
3063
3064 level = btrfs_header_level(b);
3065 err = btrfs_try_tree_read_lock(b);
3066 if (!err) {
3067 btrfs_set_path_blocking(p);
3068 btrfs_tree_read_lock(b);
3069 btrfs_clear_path_blocking(p, b,
3070 BTRFS_READ_LOCK);
3071 }
3072 b = tree_mod_log_rewind(root->fs_info, p, b, time_seq);
3073 if (!b) {
3074 ret = -ENOMEM;
3075 goto done;
3076 }
3077 p->locks[level] = BTRFS_READ_LOCK;
3078 p->nodes[level] = b;
3079 } else {
3080 p->slots[level] = slot;
3081 unlock_up(p, level, lowest_unlock, 0, NULL);
3082 goto done;
3083 }
3084 }
3085 ret = 1;
3086done:
3087 if (!p->leave_spinning)
3088 btrfs_set_path_blocking(p);
3089 if (ret < 0)
3090 btrfs_release_path(p);
3091
3092 return ret;
3093}
3094
3095/*
3096 * helper to use instead of search slot if no exact match is needed but
3097 * instead the next or previous item should be returned.
3098 * When find_higher is true, the next higher item is returned, the next lower
3099 * otherwise.
3100 * When return_any and find_higher are both true, and no higher item is found,
3101 * return the next lower instead.
3102 * When return_any is true and find_higher is false, and no lower item is found,
3103 * return the next higher instead.
3104 * It returns 0 if any item is found, 1 if none is found (tree empty), and
3105 * < 0 on error
3106 */
3107int btrfs_search_slot_for_read(struct btrfs_root *root,
3108 struct btrfs_key *key, struct btrfs_path *p,
3109 int find_higher, int return_any)
3110{
3111 int ret;
3112 struct extent_buffer *leaf;
3113
3114again:
3115 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3116 if (ret <= 0)
3117 return ret;
3118 /*
3119 * a return value of 1 means the path is at the position where the
3120 * item should be inserted. Normally this is the next bigger item,
3121 * but in case the previous item is the last in a leaf, path points
3122 * to the first free slot in the previous leaf, i.e. at an invalid
3123 * item.
3124 */
3125 leaf = p->nodes[0];
3126
3127 if (find_higher) {
3128 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3129 ret = btrfs_next_leaf(root, p);
3130 if (ret <= 0)
3131 return ret;
3132 if (!return_any)
3133 return 1;
3134 /*
3135 * no higher item found, return the next
3136 * lower instead
3137 */
3138 return_any = 0;
3139 find_higher = 0;
3140 btrfs_release_path(p);
3141 goto again;
3142 }
3143 } else {
3144 if (p->slots[0] == 0) {
3145 ret = btrfs_prev_leaf(root, p);
3146 if (ret < 0)
3147 return ret;
3148 if (!ret) {
3149 leaf = p->nodes[0];
3150 if (p->slots[0] == btrfs_header_nritems(leaf))
3151 p->slots[0]--;
3152 return 0;
3153 }
3154 if (!return_any)
3155 return 1;
3156 /*
3157 * no lower item found, return the next
3158 * higher instead
3159 */
3160 return_any = 0;
3161 find_higher = 1;
3162 btrfs_release_path(p);
3163 goto again;
3164 } else {
3165 --p->slots[0];
3166 }
3167 }
3168 return 0;
3169}
3170
3171/*
3172 * adjust the pointers going up the tree, starting at level
3173 * making sure the right key of each node is points to 'key'.
3174 * This is used after shifting pointers to the left, so it stops
3175 * fixing up pointers when a given leaf/node is not in slot 0 of the
3176 * higher levels
3177 *
3178 */
3179static void fixup_low_keys(struct btrfs_root *root, struct btrfs_path *path,
3180 struct btrfs_disk_key *key, int level)
3181{
3182 int i;
3183 struct extent_buffer *t;
3184
3185 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3186 int tslot = path->slots[i];
3187 if (!path->nodes[i])
3188 break;
3189 t = path->nodes[i];
3190 tree_mod_log_set_node_key(root->fs_info, t, tslot, 1);
3191 btrfs_set_node_key(t, key, tslot);
3192 btrfs_mark_buffer_dirty(path->nodes[i]);
3193 if (tslot != 0)
3194 break;
3195 }
3196}
3197
3198/*
3199 * update item key.
3200 *
3201 * This function isn't completely safe. It's the caller's responsibility
3202 * that the new key won't break the order
3203 */
3204void btrfs_set_item_key_safe(struct btrfs_root *root, struct btrfs_path *path,
3205 struct btrfs_key *new_key)
3206{
3207 struct btrfs_disk_key disk_key;
3208 struct extent_buffer *eb;
3209 int slot;
3210
3211 eb = path->nodes[0];
3212 slot = path->slots[0];
3213 if (slot > 0) {
3214 btrfs_item_key(eb, &disk_key, slot - 1);
3215 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
3216 }
3217 if (slot < btrfs_header_nritems(eb) - 1) {
3218 btrfs_item_key(eb, &disk_key, slot + 1);
3219 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
3220 }
3221
3222 btrfs_cpu_key_to_disk(&disk_key, new_key);
3223 btrfs_set_item_key(eb, &disk_key, slot);
3224 btrfs_mark_buffer_dirty(eb);
3225 if (slot == 0)
3226 fixup_low_keys(root, path, &disk_key, 1);
3227}
3228
3229/*
3230 * try to push data from one node into the next node left in the
3231 * tree.
3232 *
3233 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3234 * error, and > 0 if there was no room in the left hand block.
3235 */
3236static int push_node_left(struct btrfs_trans_handle *trans,
3237 struct btrfs_root *root, struct extent_buffer *dst,
3238 struct extent_buffer *src, int empty)
3239{
3240 int push_items = 0;
3241 int src_nritems;
3242 int dst_nritems;
3243 int ret = 0;
3244
3245 src_nritems = btrfs_header_nritems(src);
3246 dst_nritems = btrfs_header_nritems(dst);
3247 push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
3248 WARN_ON(btrfs_header_generation(src) != trans->transid);
3249 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3250
3251 if (!empty && src_nritems <= 8)
3252 return 1;
3253
3254 if (push_items <= 0)
3255 return 1;
3256
3257 if (empty) {
3258 push_items = min(src_nritems, push_items);
3259 if (push_items < src_nritems) {
3260 /* leave at least 8 pointers in the node if
3261 * we aren't going to empty it
3262 */
3263 if (src_nritems - push_items < 8) {
3264 if (push_items <= 8)
3265 return 1;
3266 push_items -= 8;
3267 }
3268 }
3269 } else
3270 push_items = min(src_nritems - 8, push_items);
3271
3272 ret = tree_mod_log_eb_copy(root->fs_info, dst, src, dst_nritems, 0,
3273 push_items);
3274 if (ret) {
3275 btrfs_abort_transaction(trans, root, ret);
3276 return ret;
3277 }
3278 copy_extent_buffer(dst, src,
3279 btrfs_node_key_ptr_offset(dst_nritems),
3280 btrfs_node_key_ptr_offset(0),
3281 push_items * sizeof(struct btrfs_key_ptr));
3282
3283 if (push_items < src_nritems) {
3284 /*
3285 * don't call tree_mod_log_eb_move here, key removal was already
3286 * fully logged by tree_mod_log_eb_copy above.
3287 */
3288 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3289 btrfs_node_key_ptr_offset(push_items),
3290 (src_nritems - push_items) *
3291 sizeof(struct btrfs_key_ptr));
3292 }
3293 btrfs_set_header_nritems(src, src_nritems - push_items);
3294 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3295 btrfs_mark_buffer_dirty(src);
3296 btrfs_mark_buffer_dirty(dst);
3297
3298 return ret;
3299}
3300
3301/*
3302 * try to push data from one node into the next node right in the
3303 * tree.
3304 *
3305 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3306 * error, and > 0 if there was no room in the right hand block.
3307 *
3308 * this will only push up to 1/2 the contents of the left node over
3309 */
3310static int balance_node_right(struct btrfs_trans_handle *trans,
3311 struct btrfs_root *root,
3312 struct extent_buffer *dst,
3313 struct extent_buffer *src)
3314{
3315 int push_items = 0;
3316 int max_push;
3317 int src_nritems;
3318 int dst_nritems;
3319 int ret = 0;
3320
3321 WARN_ON(btrfs_header_generation(src) != trans->transid);
3322 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3323
3324 src_nritems = btrfs_header_nritems(src);
3325 dst_nritems = btrfs_header_nritems(dst);
3326 push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
3327 if (push_items <= 0)
3328 return 1;
3329
3330 if (src_nritems < 4)
3331 return 1;
3332
3333 max_push = src_nritems / 2 + 1;
3334 /* don't try to empty the node */
3335 if (max_push >= src_nritems)
3336 return 1;
3337
3338 if (max_push < push_items)
3339 push_items = max_push;
3340
3341 tree_mod_log_eb_move(root->fs_info, dst, push_items, 0, dst_nritems);
3342 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3343 btrfs_node_key_ptr_offset(0),
3344 (dst_nritems) *
3345 sizeof(struct btrfs_key_ptr));
3346
3347 ret = tree_mod_log_eb_copy(root->fs_info, dst, src, 0,
3348 src_nritems - push_items, push_items);
3349 if (ret) {
3350 btrfs_abort_transaction(trans, root, ret);
3351 return ret;
3352 }
3353 copy_extent_buffer(dst, src,
3354 btrfs_node_key_ptr_offset(0),
3355 btrfs_node_key_ptr_offset(src_nritems - push_items),
3356 push_items * sizeof(struct btrfs_key_ptr));
3357
3358 btrfs_set_header_nritems(src, src_nritems - push_items);
3359 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3360
3361 btrfs_mark_buffer_dirty(src);
3362 btrfs_mark_buffer_dirty(dst);
3363
3364 return ret;
3365}
3366
3367/*
3368 * helper function to insert a new root level in the tree.
3369 * A new node is allocated, and a single item is inserted to
3370 * point to the existing root
3371 *
3372 * returns zero on success or < 0 on failure.
3373 */
3374static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3375 struct btrfs_root *root,
3376 struct btrfs_path *path, int level)
3377{
3378 u64 lower_gen;
3379 struct extent_buffer *lower;
3380 struct extent_buffer *c;
3381 struct extent_buffer *old;
3382 struct btrfs_disk_key lower_key;
3383
3384 BUG_ON(path->nodes[level]);
3385 BUG_ON(path->nodes[level-1] != root->node);
3386
3387 lower = path->nodes[level-1];
3388 if (level == 1)
3389 btrfs_item_key(lower, &lower_key, 0);
3390 else
3391 btrfs_node_key(lower, &lower_key, 0);
3392
3393 c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
3394 root->root_key.objectid, &lower_key,
3395 level, root->node->start, 0);
3396 if (IS_ERR(c))
3397 return PTR_ERR(c);
3398
3399 root_add_used(root, root->nodesize);
3400
3401 memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
3402 btrfs_set_header_nritems(c, 1);
3403 btrfs_set_header_level(c, level);
3404 btrfs_set_header_bytenr(c, c->start);
3405 btrfs_set_header_generation(c, trans->transid);
3406 btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
3407 btrfs_set_header_owner(c, root->root_key.objectid);
3408
3409 write_extent_buffer(c, root->fs_info->fsid, btrfs_header_fsid(),
3410 BTRFS_FSID_SIZE);
3411
3412 write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
3413 btrfs_header_chunk_tree_uuid(c), BTRFS_UUID_SIZE);
3414
3415 btrfs_set_node_key(c, &lower_key, 0);
3416 btrfs_set_node_blockptr(c, 0, lower->start);
3417 lower_gen = btrfs_header_generation(lower);
3418 WARN_ON(lower_gen != trans->transid);
3419
3420 btrfs_set_node_ptr_generation(c, 0, lower_gen);
3421
3422 btrfs_mark_buffer_dirty(c);
3423
3424 old = root->node;
3425 tree_mod_log_set_root_pointer(root, c, 0);
3426 rcu_assign_pointer(root->node, c);
3427
3428 /* the super has an extra ref to root->node */
3429 free_extent_buffer(old);
3430
3431 add_root_to_dirty_list(root);
3432 extent_buffer_get(c);
3433 path->nodes[level] = c;
3434 path->locks[level] = BTRFS_WRITE_LOCK;
3435 path->slots[level] = 0;
3436 return 0;
3437}
3438
3439/*
3440 * worker function to insert a single pointer in a node.
3441 * the node should have enough room for the pointer already
3442 *
3443 * slot and level indicate where you want the key to go, and
3444 * blocknr is the block the key points to.
3445 */
3446static void insert_ptr(struct btrfs_trans_handle *trans,
3447 struct btrfs_root *root, struct btrfs_path *path,
3448 struct btrfs_disk_key *key, u64 bytenr,
3449 int slot, int level)
3450{
3451 struct extent_buffer *lower;
3452 int nritems;
3453 int ret;
3454
3455 BUG_ON(!path->nodes[level]);
3456 btrfs_assert_tree_locked(path->nodes[level]);
3457 lower = path->nodes[level];
3458 nritems = btrfs_header_nritems(lower);
3459 BUG_ON(slot > nritems);
3460 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(root));
3461 if (slot != nritems) {
3462 if (level)
3463 tree_mod_log_eb_move(root->fs_info, lower, slot + 1,
3464 slot, nritems - slot);
3465 memmove_extent_buffer(lower,
3466 btrfs_node_key_ptr_offset(slot + 1),
3467 btrfs_node_key_ptr_offset(slot),
3468 (nritems - slot) * sizeof(struct btrfs_key_ptr));
3469 }
3470 if (level) {
3471 ret = tree_mod_log_insert_key(root->fs_info, lower, slot,
3472 MOD_LOG_KEY_ADD, GFP_NOFS);
3473 BUG_ON(ret < 0);
3474 }
3475 btrfs_set_node_key(lower, key, slot);
3476 btrfs_set_node_blockptr(lower, slot, bytenr);
3477 WARN_ON(trans->transid == 0);
3478 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3479 btrfs_set_header_nritems(lower, nritems + 1);
3480 btrfs_mark_buffer_dirty(lower);
3481}
3482
3483/*
3484 * split the node at the specified level in path in two.
3485 * The path is corrected to point to the appropriate node after the split
3486 *
3487 * Before splitting this tries to make some room in the node by pushing
3488 * left and right, if either one works, it returns right away.
3489 *
3490 * returns 0 on success and < 0 on failure
3491 */
3492static noinline int split_node(struct btrfs_trans_handle *trans,
3493 struct btrfs_root *root,
3494 struct btrfs_path *path, int level)
3495{
3496 struct extent_buffer *c;
3497 struct extent_buffer *split;
3498 struct btrfs_disk_key disk_key;
3499 int mid;
3500 int ret;
3501 u32 c_nritems;
3502
3503 c = path->nodes[level];
3504 WARN_ON(btrfs_header_generation(c) != trans->transid);
3505 if (c == root->node) {
3506 /*
3507 * trying to split the root, lets make a new one
3508 *
3509 * tree mod log: We don't log_removal old root in
3510 * insert_new_root, because that root buffer will be kept as a
3511 * normal node. We are going to log removal of half of the
3512 * elements below with tree_mod_log_eb_copy. We're holding a
3513 * tree lock on the buffer, which is why we cannot race with
3514 * other tree_mod_log users.
3515 */
3516 ret = insert_new_root(trans, root, path, level + 1);
3517 if (ret)
3518 return ret;
3519 } else {
3520 ret = push_nodes_for_insert(trans, root, path, level);
3521 c = path->nodes[level];
3522 if (!ret && btrfs_header_nritems(c) <
3523 BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
3524 return 0;
3525 if (ret < 0)
3526 return ret;
3527 }
3528
3529 c_nritems = btrfs_header_nritems(c);
3530 mid = (c_nritems + 1) / 2;
3531 btrfs_node_key(c, &disk_key, mid);
3532
3533 split = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
3534 root->root_key.objectid,
3535 &disk_key, level, c->start, 0);
3536 if (IS_ERR(split))
3537 return PTR_ERR(split);
3538
3539 root_add_used(root, root->nodesize);
3540
3541 memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
3542 btrfs_set_header_level(split, btrfs_header_level(c));
3543 btrfs_set_header_bytenr(split, split->start);
3544 btrfs_set_header_generation(split, trans->transid);
3545 btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
3546 btrfs_set_header_owner(split, root->root_key.objectid);
3547 write_extent_buffer(split, root->fs_info->fsid,
3548 btrfs_header_fsid(), BTRFS_FSID_SIZE);
3549 write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
3550 btrfs_header_chunk_tree_uuid(split),
3551 BTRFS_UUID_SIZE);
3552
3553 ret = tree_mod_log_eb_copy(root->fs_info, split, c, 0,
3554 mid, c_nritems - mid);
3555 if (ret) {
3556 btrfs_abort_transaction(trans, root, ret);
3557 return ret;
3558 }
3559 copy_extent_buffer(split, c,
3560 btrfs_node_key_ptr_offset(0),
3561 btrfs_node_key_ptr_offset(mid),
3562 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3563 btrfs_set_header_nritems(split, c_nritems - mid);
3564 btrfs_set_header_nritems(c, mid);
3565 ret = 0;
3566
3567 btrfs_mark_buffer_dirty(c);
3568 btrfs_mark_buffer_dirty(split);
3569
3570 insert_ptr(trans, root, path, &disk_key, split->start,
3571 path->slots[level + 1] + 1, level + 1);
3572
3573 if (path->slots[level] >= mid) {
3574 path->slots[level] -= mid;
3575 btrfs_tree_unlock(c);
3576 free_extent_buffer(c);
3577 path->nodes[level] = split;
3578 path->slots[level + 1] += 1;
3579 } else {
3580 btrfs_tree_unlock(split);
3581 free_extent_buffer(split);
3582 }
3583 return ret;
3584}
3585
3586/*
3587 * how many bytes are required to store the items in a leaf. start
3588 * and nr indicate which items in the leaf to check. This totals up the
3589 * space used both by the item structs and the item data
3590 */
3591static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3592{
3593 struct btrfs_item *start_item;
3594 struct btrfs_item *end_item;
3595 struct btrfs_map_token token;
3596 int data_len;
3597 int nritems = btrfs_header_nritems(l);
3598 int end = min(nritems, start + nr) - 1;
3599
3600 if (!nr)
3601 return 0;
3602 btrfs_init_map_token(&token);
3603 start_item = btrfs_item_nr(start);
3604 end_item = btrfs_item_nr(end);
3605 data_len = btrfs_token_item_offset(l, start_item, &token) +
3606 btrfs_token_item_size(l, start_item, &token);
3607 data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3608 data_len += sizeof(struct btrfs_item) * nr;
3609 WARN_ON(data_len < 0);
3610 return data_len;
3611}
3612
3613/*
3614 * The space between the end of the leaf items and
3615 * the start of the leaf data. IOW, how much room
3616 * the leaf has left for both items and data
3617 */
3618noinline int btrfs_leaf_free_space(struct btrfs_root *root,
3619 struct extent_buffer *leaf)
3620{
3621 int nritems = btrfs_header_nritems(leaf);
3622 int ret;
3623 ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
3624 if (ret < 0) {
3625 btrfs_crit(root->fs_info,
3626 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3627 ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
3628 leaf_space_used(leaf, 0, nritems), nritems);
3629 }
3630 return ret;
3631}
3632
3633/*
3634 * min slot controls the lowest index we're willing to push to the
3635 * right. We'll push up to and including min_slot, but no lower
3636 */
3637static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
3638 struct btrfs_root *root,
3639 struct btrfs_path *path,
3640 int data_size, int empty,
3641 struct extent_buffer *right,
3642 int free_space, u32 left_nritems,
3643 u32 min_slot)
3644{
3645 struct extent_buffer *left = path->nodes[0];
3646 struct extent_buffer *upper = path->nodes[1];
3647 struct btrfs_map_token token;
3648 struct btrfs_disk_key disk_key;
3649 int slot;
3650 u32 i;
3651 int push_space = 0;
3652 int push_items = 0;
3653 struct btrfs_item *item;
3654 u32 nr;
3655 u32 right_nritems;
3656 u32 data_end;
3657 u32 this_item_size;
3658
3659 btrfs_init_map_token(&token);
3660
3661 if (empty)
3662 nr = 0;
3663 else
3664 nr = max_t(u32, 1, min_slot);
3665
3666 if (path->slots[0] >= left_nritems)
3667 push_space += data_size;
3668
3669 slot = path->slots[1];
3670 i = left_nritems - 1;
3671 while (i >= nr) {
3672 item = btrfs_item_nr(i);
3673
3674 if (!empty && push_items > 0) {
3675 if (path->slots[0] > i)
3676 break;
3677 if (path->slots[0] == i) {
3678 int space = btrfs_leaf_free_space(root, left);
3679 if (space + push_space * 2 > free_space)
3680 break;
3681 }
3682 }
3683
3684 if (path->slots[0] == i)
3685 push_space += data_size;
3686
3687 this_item_size = btrfs_item_size(left, item);
3688 if (this_item_size + sizeof(*item) + push_space > free_space)
3689 break;
3690
3691 push_items++;
3692 push_space += this_item_size + sizeof(*item);
3693 if (i == 0)
3694 break;
3695 i--;
3696 }
3697
3698 if (push_items == 0)
3699 goto out_unlock;
3700
3701 WARN_ON(!empty && push_items == left_nritems);
3702
3703 /* push left to right */
3704 right_nritems = btrfs_header_nritems(right);
3705
3706 push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3707 push_space -= leaf_data_end(root, left);
3708
3709 /* make room in the right data area */
3710 data_end = leaf_data_end(root, right);
3711 memmove_extent_buffer(right,
3712 btrfs_leaf_data(right) + data_end - push_space,
3713 btrfs_leaf_data(right) + data_end,
3714 BTRFS_LEAF_DATA_SIZE(root) - data_end);
3715
3716 /* copy from the left data area */
3717 copy_extent_buffer(right, left, btrfs_leaf_data(right) +
3718 BTRFS_LEAF_DATA_SIZE(root) - push_space,
3719 btrfs_leaf_data(left) + leaf_data_end(root, left),
3720 push_space);
3721
3722 memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3723 btrfs_item_nr_offset(0),
3724 right_nritems * sizeof(struct btrfs_item));
3725
3726 /* copy the items from left to right */
3727 copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3728 btrfs_item_nr_offset(left_nritems - push_items),
3729 push_items * sizeof(struct btrfs_item));
3730
3731 /* update the item pointers */
3732 right_nritems += push_items;
3733 btrfs_set_header_nritems(right, right_nritems);
3734 push_space = BTRFS_LEAF_DATA_SIZE(root);
3735 for (i = 0; i < right_nritems; i++) {
3736 item = btrfs_item_nr(i);
3737 push_space -= btrfs_token_item_size(right, item, &token);
3738 btrfs_set_token_item_offset(right, item, push_space, &token);
3739 }
3740
3741 left_nritems -= push_items;
3742 btrfs_set_header_nritems(left, left_nritems);
3743
3744 if (left_nritems)
3745 btrfs_mark_buffer_dirty(left);
3746 else
3747 clean_tree_block(trans, root, left);
3748
3749 btrfs_mark_buffer_dirty(right);
3750
3751 btrfs_item_key(right, &disk_key, 0);
3752 btrfs_set_node_key(upper, &disk_key, slot + 1);
3753 btrfs_mark_buffer_dirty(upper);
3754
3755 /* then fixup the leaf pointer in the path */
3756 if (path->slots[0] >= left_nritems) {
3757 path->slots[0] -= left_nritems;
3758 if (btrfs_header_nritems(path->nodes[0]) == 0)
3759 clean_tree_block(trans, root, path->nodes[0]);
3760 btrfs_tree_unlock(path->nodes[0]);
3761 free_extent_buffer(path->nodes[0]);
3762 path->nodes[0] = right;
3763 path->slots[1] += 1;
3764 } else {
3765 btrfs_tree_unlock(right);
3766 free_extent_buffer(right);
3767 }
3768 return 0;
3769
3770out_unlock:
3771 btrfs_tree_unlock(right);
3772 free_extent_buffer(right);
3773 return 1;
3774}
3775
3776/*
3777 * push some data in the path leaf to the right, trying to free up at
3778 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3779 *
3780 * returns 1 if the push failed because the other node didn't have enough
3781 * room, 0 if everything worked out and < 0 if there were major errors.
3782 *
3783 * this will push starting from min_slot to the end of the leaf. It won't
3784 * push any slot lower than min_slot
3785 */
3786static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3787 *root, struct btrfs_path *path,
3788 int min_data_size, int data_size,
3789 int empty, u32 min_slot)
3790{
3791 struct extent_buffer *left = path->nodes[0];
3792 struct extent_buffer *right;
3793 struct extent_buffer *upper;
3794 int slot;
3795 int free_space;
3796 u32 left_nritems;
3797 int ret;
3798
3799 if (!path->nodes[1])
3800 return 1;
3801
3802 slot = path->slots[1];
3803 upper = path->nodes[1];
3804 if (slot >= btrfs_header_nritems(upper) - 1)
3805 return 1;
3806
3807 btrfs_assert_tree_locked(path->nodes[1]);
3808
3809 right = read_node_slot(root, upper, slot + 1);
3810 if (right == NULL)
3811 return 1;
3812
3813 btrfs_tree_lock(right);
3814 btrfs_set_lock_blocking(right);
3815
3816 free_space = btrfs_leaf_free_space(root, right);
3817 if (free_space < data_size)
3818 goto out_unlock;
3819
3820 /* cow and double check */
3821 ret = btrfs_cow_block(trans, root, right, upper,
3822 slot + 1, &right);
3823 if (ret)
3824 goto out_unlock;
3825
3826 free_space = btrfs_leaf_free_space(root, right);
3827 if (free_space < data_size)
3828 goto out_unlock;
3829
3830 left_nritems = btrfs_header_nritems(left);
3831 if (left_nritems == 0)
3832 goto out_unlock;
3833
3834 if (path->slots[0] == left_nritems && !empty) {
3835 /* Key greater than all keys in the leaf, right neighbor has
3836 * enough room for it and we're not emptying our leaf to delete
3837 * it, therefore use right neighbor to insert the new item and
3838 * no need to touch/dirty our left leaft. */
3839 btrfs_tree_unlock(left);
3840 free_extent_buffer(left);
3841 path->nodes[0] = right;
3842 path->slots[0] = 0;
3843 path->slots[1]++;
3844 return 0;
3845 }
3846
3847 return __push_leaf_right(trans, root, path, min_data_size, empty,
3848 right, free_space, left_nritems, min_slot);
3849out_unlock:
3850 btrfs_tree_unlock(right);
3851 free_extent_buffer(right);
3852 return 1;
3853}
3854
3855/*
3856 * push some data in the path leaf to the left, trying to free up at
3857 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3858 *
3859 * max_slot can put a limit on how far into the leaf we'll push items. The
3860 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3861 * items
3862 */
3863static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
3864 struct btrfs_root *root,
3865 struct btrfs_path *path, int data_size,
3866 int empty, struct extent_buffer *left,
3867 int free_space, u32 right_nritems,
3868 u32 max_slot)
3869{
3870 struct btrfs_disk_key disk_key;
3871 struct extent_buffer *right = path->nodes[0];
3872 int i;
3873 int push_space = 0;
3874 int push_items = 0;
3875 struct btrfs_item *item;
3876 u32 old_left_nritems;
3877 u32 nr;
3878 int ret = 0;
3879 u32 this_item_size;
3880 u32 old_left_item_size;
3881 struct btrfs_map_token token;
3882
3883 btrfs_init_map_token(&token);
3884
3885 if (empty)
3886 nr = min(right_nritems, max_slot);
3887 else
3888 nr = min(right_nritems - 1, max_slot);
3889
3890 for (i = 0; i < nr; i++) {
3891 item = btrfs_item_nr(i);
3892
3893 if (!empty && push_items > 0) {
3894 if (path->slots[0] < i)
3895 break;
3896 if (path->slots[0] == i) {
3897 int space = btrfs_leaf_free_space(root, right);
3898 if (space + push_space * 2 > free_space)
3899 break;
3900 }
3901 }
3902
3903 if (path->slots[0] == i)
3904 push_space += data_size;
3905
3906 this_item_size = btrfs_item_size(right, item);
3907 if (this_item_size + sizeof(*item) + push_space > free_space)
3908 break;
3909
3910 push_items++;
3911 push_space += this_item_size + sizeof(*item);
3912 }
3913
3914 if (push_items == 0) {
3915 ret = 1;
3916 goto out;
3917 }
3918 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3919
3920 /* push data from right to left */
3921 copy_extent_buffer(left, right,
3922 btrfs_item_nr_offset(btrfs_header_nritems(left)),
3923 btrfs_item_nr_offset(0),
3924 push_items * sizeof(struct btrfs_item));
3925
3926 push_space = BTRFS_LEAF_DATA_SIZE(root) -
3927 btrfs_item_offset_nr(right, push_items - 1);
3928
3929 copy_extent_buffer(left, right, btrfs_leaf_data(left) +
3930 leaf_data_end(root, left) - push_space,
3931 btrfs_leaf_data(right) +
3932 btrfs_item_offset_nr(right, push_items - 1),
3933 push_space);
3934 old_left_nritems = btrfs_header_nritems(left);
3935 BUG_ON(old_left_nritems <= 0);
3936
3937 old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3938 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3939 u32 ioff;
3940
3941 item = btrfs_item_nr(i);
3942
3943 ioff = btrfs_token_item_offset(left, item, &token);
3944 btrfs_set_token_item_offset(left, item,
3945 ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size),
3946 &token);
3947 }
3948 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3949
3950 /* fixup right node */
3951 if (push_items > right_nritems)
3952 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3953 right_nritems);
3954
3955 if (push_items < right_nritems) {
3956 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3957 leaf_data_end(root, right);
3958 memmove_extent_buffer(right, btrfs_leaf_data(right) +
3959 BTRFS_LEAF_DATA_SIZE(root) - push_space,
3960 btrfs_leaf_data(right) +
3961 leaf_data_end(root, right), push_space);
3962
3963 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3964 btrfs_item_nr_offset(push_items),
3965 (btrfs_header_nritems(right) - push_items) *
3966 sizeof(struct btrfs_item));
3967 }
3968 right_nritems -= push_items;
3969 btrfs_set_header_nritems(right, right_nritems);
3970 push_space = BTRFS_LEAF_DATA_SIZE(root);
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 clean_tree_block(trans, root, right);
3984
3985 btrfs_item_key(right, &disk_key, 0);
3986 fixup_low_keys(root, 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 = read_node_slot(root, path->nodes[1], slot - 1);
4040 if (left == NULL)
4041 return 1;
4042
4043 btrfs_tree_lock(left);
4044 btrfs_set_lock_blocking(left);
4045
4046 free_space = btrfs_leaf_free_space(root, left);
4047 if (free_space < data_size) {
4048 ret = 1;
4049 goto out;
4050 }
4051
4052 /* cow and double check */
4053 ret = btrfs_cow_block(trans, root, left,
4054 path->nodes[1], slot - 1, &left);
4055 if (ret) {
4056 /* we hit -ENOSPC, but it isn't fatal here */
4057 if (ret == -ENOSPC)
4058 ret = 1;
4059 goto out;
4060 }
4061
4062 free_space = btrfs_leaf_free_space(root, left);
4063 if (free_space < data_size) {
4064 ret = 1;
4065 goto out;
4066 }
4067
4068 return __push_leaf_left(trans, root, path, min_data_size,
4069 empty, left, free_space, right_nritems,
4070 max_slot);
4071out:
4072 btrfs_tree_unlock(left);
4073 free_extent_buffer(left);
4074 return ret;
4075}
4076
4077/*
4078 * split the path's leaf in two, making sure there is at least data_size
4079 * available for the resulting leaf level of the path.
4080 */
4081static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4082 struct btrfs_root *root,
4083 struct btrfs_path *path,
4084 struct extent_buffer *l,
4085 struct extent_buffer *right,
4086 int slot, int mid, int nritems)
4087{
4088 int data_copy_size;
4089 int rt_data_off;
4090 int i;
4091 struct btrfs_disk_key disk_key;
4092 struct btrfs_map_token token;
4093
4094 btrfs_init_map_token(&token);
4095
4096 nritems = nritems - mid;
4097 btrfs_set_header_nritems(right, nritems);
4098 data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
4099
4100 copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4101 btrfs_item_nr_offset(mid),
4102 nritems * sizeof(struct btrfs_item));
4103
4104 copy_extent_buffer(right, l,
4105 btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
4106 data_copy_size, btrfs_leaf_data(l) +
4107 leaf_data_end(root, l), data_copy_size);
4108
4109 rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
4110 btrfs_item_end_nr(l, mid);
4111
4112 for (i = 0; i < nritems; i++) {
4113 struct btrfs_item *item = btrfs_item_nr(i);
4114 u32 ioff;
4115
4116 ioff = btrfs_token_item_offset(right, item, &token);
4117 btrfs_set_token_item_offset(right, item,
4118 ioff + rt_data_off, &token);
4119 }
4120
4121 btrfs_set_header_nritems(l, mid);
4122 btrfs_item_key(right, &disk_key, 0);
4123 insert_ptr(trans, root, path, &disk_key, right->start,
4124 path->slots[1] + 1, 1);
4125
4126 btrfs_mark_buffer_dirty(right);
4127 btrfs_mark_buffer_dirty(l);
4128 BUG_ON(path->slots[0] != slot);
4129
4130 if (mid <= slot) {
4131 btrfs_tree_unlock(path->nodes[0]);
4132 free_extent_buffer(path->nodes[0]);
4133 path->nodes[0] = right;
4134 path->slots[0] -= mid;
4135 path->slots[1] += 1;
4136 } else {
4137 btrfs_tree_unlock(right);
4138 free_extent_buffer(right);
4139 }
4140
4141 BUG_ON(path->slots[0] < 0);
4142}
4143
4144/*
4145 * double splits happen when we need to insert a big item in the middle
4146 * of a leaf. A double split can leave us with 3 mostly empty leaves:
4147 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4148 * A B C
4149 *
4150 * We avoid this by trying to push the items on either side of our target
4151 * into the adjacent leaves. If all goes well we can avoid the double split
4152 * completely.
4153 */
4154static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4155 struct btrfs_root *root,
4156 struct btrfs_path *path,
4157 int data_size)
4158{
4159 int ret;
4160 int progress = 0;
4161 int slot;
4162 u32 nritems;
4163 int space_needed = data_size;
4164
4165 slot = path->slots[0];
4166 if (slot < btrfs_header_nritems(path->nodes[0]))
4167 space_needed -= btrfs_leaf_free_space(root, path->nodes[0]);
4168
4169 /*
4170 * try to push all the items after our slot into the
4171 * right leaf
4172 */
4173 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4174 if (ret < 0)
4175 return ret;
4176
4177 if (ret == 0)
4178 progress++;
4179
4180 nritems = btrfs_header_nritems(path->nodes[0]);
4181 /*
4182 * our goal is to get our slot at the start or end of a leaf. If
4183 * we've done so we're done
4184 */
4185 if (path->slots[0] == 0 || path->slots[0] == nritems)
4186 return 0;
4187
4188 if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
4189 return 0;
4190
4191 /* try to push all the items before our slot into the next leaf */
4192 slot = path->slots[0];
4193 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4194 if (ret < 0)
4195 return ret;
4196
4197 if (ret == 0)
4198 progress++;
4199
4200 if (progress)
4201 return 0;
4202 return 1;
4203}
4204
4205/*
4206 * split the path's leaf in two, making sure there is at least data_size
4207 * available for the resulting leaf level of the path.
4208 *
4209 * returns 0 if all went well and < 0 on failure.
4210 */
4211static noinline int split_leaf(struct btrfs_trans_handle *trans,
4212 struct btrfs_root *root,
4213 struct btrfs_key *ins_key,
4214 struct btrfs_path *path, int data_size,
4215 int extend)
4216{
4217 struct btrfs_disk_key disk_key;
4218 struct extent_buffer *l;
4219 u32 nritems;
4220 int mid;
4221 int slot;
4222 struct extent_buffer *right;
4223 int ret = 0;
4224 int wret;
4225 int split;
4226 int num_doubles = 0;
4227 int tried_avoid_double = 0;
4228
4229 l = path->nodes[0];
4230 slot = path->slots[0];
4231 if (extend && data_size + btrfs_item_size_nr(l, slot) +
4232 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
4233 return -EOVERFLOW;
4234
4235 /* first try to make some room by pushing left and right */
4236 if (data_size && path->nodes[1]) {
4237 int space_needed = data_size;
4238
4239 if (slot < btrfs_header_nritems(l))
4240 space_needed -= btrfs_leaf_free_space(root, l);
4241
4242 wret = push_leaf_right(trans, root, path, space_needed,
4243 space_needed, 0, 0);
4244 if (wret < 0)
4245 return wret;
4246 if (wret) {
4247 wret = push_leaf_left(trans, root, path, space_needed,
4248 space_needed, 0, (u32)-1);
4249 if (wret < 0)
4250 return wret;
4251 }
4252 l = path->nodes[0];
4253
4254 /* did the pushes work? */
4255 if (btrfs_leaf_free_space(root, l) >= data_size)
4256 return 0;
4257 }
4258
4259 if (!path->nodes[1]) {
4260 ret = insert_new_root(trans, root, path, 1);
4261 if (ret)
4262 return ret;
4263 }
4264again:
4265 split = 1;
4266 l = path->nodes[0];
4267 slot = path->slots[0];
4268 nritems = btrfs_header_nritems(l);
4269 mid = (nritems + 1) / 2;
4270
4271 if (mid <= slot) {
4272 if (nritems == 1 ||
4273 leaf_space_used(l, mid, nritems - mid) + data_size >
4274 BTRFS_LEAF_DATA_SIZE(root)) {
4275 if (slot >= nritems) {
4276 split = 0;
4277 } else {
4278 mid = slot;
4279 if (mid != nritems &&
4280 leaf_space_used(l, mid, nritems - mid) +
4281 data_size > BTRFS_LEAF_DATA_SIZE(root)) {
4282 if (data_size && !tried_avoid_double)
4283 goto push_for_double;
4284 split = 2;
4285 }
4286 }
4287 }
4288 } else {
4289 if (leaf_space_used(l, 0, mid) + data_size >
4290 BTRFS_LEAF_DATA_SIZE(root)) {
4291 if (!extend && data_size && slot == 0) {
4292 split = 0;
4293 } else if ((extend || !data_size) && slot == 0) {
4294 mid = 1;
4295 } else {
4296 mid = slot;
4297 if (mid != nritems &&
4298 leaf_space_used(l, mid, nritems - mid) +
4299 data_size > BTRFS_LEAF_DATA_SIZE(root)) {
4300 if (data_size && !tried_avoid_double)
4301 goto push_for_double;
4302 split = 2;
4303 }
4304 }
4305 }
4306 }
4307
4308 if (split == 0)
4309 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4310 else
4311 btrfs_item_key(l, &disk_key, mid);
4312
4313 right = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
4314 root->root_key.objectid,
4315 &disk_key, 0, l->start, 0);
4316 if (IS_ERR(right))
4317 return PTR_ERR(right);
4318
4319 root_add_used(root, root->leafsize);
4320
4321 memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
4322 btrfs_set_header_bytenr(right, right->start);
4323 btrfs_set_header_generation(right, trans->transid);
4324 btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
4325 btrfs_set_header_owner(right, root->root_key.objectid);
4326 btrfs_set_header_level(right, 0);
4327 write_extent_buffer(right, root->fs_info->fsid,
4328 btrfs_header_fsid(), BTRFS_FSID_SIZE);
4329
4330 write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
4331 btrfs_header_chunk_tree_uuid(right),
4332 BTRFS_UUID_SIZE);
4333
4334 if (split == 0) {
4335 if (mid <= slot) {
4336 btrfs_set_header_nritems(right, 0);
4337 insert_ptr(trans, root, path, &disk_key, right->start,
4338 path->slots[1] + 1, 1);
4339 btrfs_tree_unlock(path->nodes[0]);
4340 free_extent_buffer(path->nodes[0]);
4341 path->nodes[0] = right;
4342 path->slots[0] = 0;
4343 path->slots[1] += 1;
4344 } else {
4345 btrfs_set_header_nritems(right, 0);
4346 insert_ptr(trans, root, path, &disk_key, right->start,
4347 path->slots[1], 1);
4348 btrfs_tree_unlock(path->nodes[0]);
4349 free_extent_buffer(path->nodes[0]);
4350 path->nodes[0] = right;
4351 path->slots[0] = 0;
4352 if (path->slots[1] == 0)
4353 fixup_low_keys(root, path, &disk_key, 1);
4354 }
4355 btrfs_mark_buffer_dirty(right);
4356 return ret;
4357 }
4358
4359 copy_for_split(trans, root, path, l, right, slot, mid, nritems);
4360
4361 if (split == 2) {
4362 BUG_ON(num_doubles != 0);
4363 num_doubles++;
4364 goto again;
4365 }
4366
4367 return 0;
4368
4369push_for_double:
4370 push_for_double_split(trans, root, path, data_size);
4371 tried_avoid_double = 1;
4372 if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
4373 return 0;
4374 goto again;
4375}
4376
4377static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4378 struct btrfs_root *root,
4379 struct btrfs_path *path, int ins_len)
4380{
4381 struct btrfs_key key;
4382 struct extent_buffer *leaf;
4383 struct btrfs_file_extent_item *fi;
4384 u64 extent_len = 0;
4385 u32 item_size;
4386 int ret;
4387
4388 leaf = path->nodes[0];
4389 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4390
4391 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4392 key.type != BTRFS_EXTENT_CSUM_KEY);
4393
4394 if (btrfs_leaf_free_space(root, leaf) >= ins_len)
4395 return 0;
4396
4397 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4398 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4399 fi = btrfs_item_ptr(leaf, path->slots[0],
4400 struct btrfs_file_extent_item);
4401 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4402 }
4403 btrfs_release_path(path);
4404
4405 path->keep_locks = 1;
4406 path->search_for_split = 1;
4407 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4408 path->search_for_split = 0;
4409 if (ret < 0)
4410 goto err;
4411
4412 ret = -EAGAIN;
4413 leaf = path->nodes[0];
4414 /* if our item isn't there or got smaller, return now */
4415 if (ret > 0 || 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(root, 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_trans_handle *trans,
4443 struct btrfs_root *root,
4444 struct btrfs_path *path,
4445 struct btrfs_key *new_key,
4446 unsigned long split_offset)
4447{
4448 struct extent_buffer *leaf;
4449 struct btrfs_item *item;
4450 struct btrfs_item *new_item;
4451 int slot;
4452 char *buf;
4453 u32 nritems;
4454 u32 item_size;
4455 u32 orig_offset;
4456 struct btrfs_disk_key disk_key;
4457
4458 leaf = path->nodes[0];
4459 BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
4460
4461 btrfs_set_path_blocking(path);
4462
4463 item = btrfs_item_nr(path->slots[0]);
4464 orig_offset = btrfs_item_offset(leaf, item);
4465 item_size = btrfs_item_size(leaf, item);
4466
4467 buf = kmalloc(item_size, GFP_NOFS);
4468 if (!buf)
4469 return -ENOMEM;
4470
4471 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4472 path->slots[0]), item_size);
4473
4474 slot = path->slots[0] + 1;
4475 nritems = btrfs_header_nritems(leaf);
4476 if (slot != nritems) {
4477 /* shift the items */
4478 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4479 btrfs_item_nr_offset(slot),
4480 (nritems - slot) * sizeof(struct btrfs_item));
4481 }
4482
4483 btrfs_cpu_key_to_disk(&disk_key, new_key);
4484 btrfs_set_item_key(leaf, &disk_key, slot);
4485
4486 new_item = btrfs_item_nr(slot);
4487
4488 btrfs_set_item_offset(leaf, new_item, orig_offset);
4489 btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4490
4491 btrfs_set_item_offset(leaf, item,
4492 orig_offset + item_size - split_offset);
4493 btrfs_set_item_size(leaf, item, split_offset);
4494
4495 btrfs_set_header_nritems(leaf, nritems + 1);
4496
4497 /* write the data for the start of the original item */
4498 write_extent_buffer(leaf, buf,
4499 btrfs_item_ptr_offset(leaf, path->slots[0]),
4500 split_offset);
4501
4502 /* write the data for the new item */
4503 write_extent_buffer(leaf, buf + split_offset,
4504 btrfs_item_ptr_offset(leaf, slot),
4505 item_size - split_offset);
4506 btrfs_mark_buffer_dirty(leaf);
4507
4508 BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
4509 kfree(buf);
4510 return 0;
4511}
4512
4513/*
4514 * This function splits a single item into two items,
4515 * giving 'new_key' to the new item and splitting the
4516 * old one at split_offset (from the start of the item).
4517 *
4518 * The path may be released by this operation. After
4519 * the split, the path is pointing to the old item. The
4520 * new item is going to be in the same node as the old one.
4521 *
4522 * Note, the item being split must be smaller enough to live alone on
4523 * a tree block with room for one extra struct btrfs_item
4524 *
4525 * This allows us to split the item in place, keeping a lock on the
4526 * leaf the entire time.
4527 */
4528int btrfs_split_item(struct btrfs_trans_handle *trans,
4529 struct btrfs_root *root,
4530 struct btrfs_path *path,
4531 struct btrfs_key *new_key,
4532 unsigned long split_offset)
4533{
4534 int ret;
4535 ret = setup_leaf_for_split(trans, root, path,
4536 sizeof(struct btrfs_item));
4537 if (ret)
4538 return ret;
4539
4540 ret = split_item(trans, root, path, new_key, split_offset);
4541 return ret;
4542}
4543
4544/*
4545 * This function duplicate a item, giving 'new_key' to the new item.
4546 * It guarantees both items live in the same tree leaf and the new item
4547 * is contiguous with the original item.
4548 *
4549 * This allows us to split file extent in place, keeping a lock on the
4550 * leaf the entire time.
4551 */
4552int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4553 struct btrfs_root *root,
4554 struct btrfs_path *path,
4555 struct btrfs_key *new_key)
4556{
4557 struct extent_buffer *leaf;
4558 int ret;
4559 u32 item_size;
4560
4561 leaf = path->nodes[0];
4562 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4563 ret = setup_leaf_for_split(trans, root, path,
4564 item_size + sizeof(struct btrfs_item));
4565 if (ret)
4566 return ret;
4567
4568 path->slots[0]++;
4569 setup_items_for_insert(root, path, new_key, &item_size,
4570 item_size, item_size +
4571 sizeof(struct btrfs_item), 1);
4572 leaf = path->nodes[0];
4573 memcpy_extent_buffer(leaf,
4574 btrfs_item_ptr_offset(leaf, path->slots[0]),
4575 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4576 item_size);
4577 return 0;
4578}
4579
4580/*
4581 * make the item pointed to by the path smaller. new_size indicates
4582 * how small to make it, and from_end tells us if we just chop bytes
4583 * off the end of the item or if we shift the item to chop bytes off
4584 * the front.
4585 */
4586void btrfs_truncate_item(struct btrfs_root *root, struct btrfs_path *path,
4587 u32 new_size, int from_end)
4588{
4589 int slot;
4590 struct extent_buffer *leaf;
4591 struct btrfs_item *item;
4592 u32 nritems;
4593 unsigned int data_end;
4594 unsigned int old_data_start;
4595 unsigned int old_size;
4596 unsigned int size_diff;
4597 int i;
4598 struct btrfs_map_token token;
4599
4600 btrfs_init_map_token(&token);
4601
4602 leaf = path->nodes[0];
4603 slot = path->slots[0];
4604
4605 old_size = btrfs_item_size_nr(leaf, slot);
4606 if (old_size == new_size)
4607 return;
4608
4609 nritems = btrfs_header_nritems(leaf);
4610 data_end = leaf_data_end(root, leaf);
4611
4612 old_data_start = btrfs_item_offset_nr(leaf, slot);
4613
4614 size_diff = old_size - new_size;
4615
4616 BUG_ON(slot < 0);
4617 BUG_ON(slot >= nritems);
4618
4619 /*
4620 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4621 */
4622 /* first correct the data pointers */
4623 for (i = slot; i < nritems; i++) {
4624 u32 ioff;
4625 item = btrfs_item_nr(i);
4626
4627 ioff = btrfs_token_item_offset(leaf, item, &token);
4628 btrfs_set_token_item_offset(leaf, item,
4629 ioff + size_diff, &token);
4630 }
4631
4632 /* shift the data */
4633 if (from_end) {
4634 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4635 data_end + size_diff, btrfs_leaf_data(leaf) +
4636 data_end, old_data_start + new_size - data_end);
4637 } else {
4638 struct btrfs_disk_key disk_key;
4639 u64 offset;
4640
4641 btrfs_item_key(leaf, &disk_key, slot);
4642
4643 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4644 unsigned long ptr;
4645 struct btrfs_file_extent_item *fi;
4646
4647 fi = btrfs_item_ptr(leaf, slot,
4648 struct btrfs_file_extent_item);
4649 fi = (struct btrfs_file_extent_item *)(
4650 (unsigned long)fi - size_diff);
4651
4652 if (btrfs_file_extent_type(leaf, fi) ==
4653 BTRFS_FILE_EXTENT_INLINE) {
4654 ptr = btrfs_item_ptr_offset(leaf, slot);
4655 memmove_extent_buffer(leaf, ptr,
4656 (unsigned long)fi,
4657 offsetof(struct btrfs_file_extent_item,
4658 disk_bytenr));
4659 }
4660 }
4661
4662 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4663 data_end + size_diff, btrfs_leaf_data(leaf) +
4664 data_end, old_data_start - data_end);
4665
4666 offset = btrfs_disk_key_offset(&disk_key);
4667 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4668 btrfs_set_item_key(leaf, &disk_key, slot);
4669 if (slot == 0)
4670 fixup_low_keys(root, path, &disk_key, 1);
4671 }
4672
4673 item = btrfs_item_nr(slot);
4674 btrfs_set_item_size(leaf, item, new_size);
4675 btrfs_mark_buffer_dirty(leaf);
4676
4677 if (btrfs_leaf_free_space(root, leaf) < 0) {
4678 btrfs_print_leaf(root, leaf);
4679 BUG();
4680 }
4681}
4682
4683/*
4684 * make the item pointed to by the path bigger, data_size is the added size.
4685 */
4686void btrfs_extend_item(struct btrfs_root *root, struct btrfs_path *path,
4687 u32 data_size)
4688{
4689 int slot;
4690 struct extent_buffer *leaf;
4691 struct btrfs_item *item;
4692 u32 nritems;
4693 unsigned int data_end;
4694 unsigned int old_data;
4695 unsigned int old_size;
4696 int i;
4697 struct btrfs_map_token token;
4698
4699 btrfs_init_map_token(&token);
4700
4701 leaf = path->nodes[0];
4702
4703 nritems = btrfs_header_nritems(leaf);
4704 data_end = leaf_data_end(root, leaf);
4705
4706 if (btrfs_leaf_free_space(root, leaf) < data_size) {
4707 btrfs_print_leaf(root, leaf);
4708 BUG();
4709 }
4710 slot = path->slots[0];
4711 old_data = btrfs_item_end_nr(leaf, slot);
4712
4713 BUG_ON(slot < 0);
4714 if (slot >= nritems) {
4715 btrfs_print_leaf(root, leaf);
4716 btrfs_crit(root->fs_info, "slot %d too large, nritems %d",
4717 slot, nritems);
4718 BUG_ON(1);
4719 }
4720
4721 /*
4722 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4723 */
4724 /* first correct the data pointers */
4725 for (i = slot; i < nritems; i++) {
4726 u32 ioff;
4727 item = btrfs_item_nr(i);
4728
4729 ioff = btrfs_token_item_offset(leaf, item, &token);
4730 btrfs_set_token_item_offset(leaf, item,
4731 ioff - data_size, &token);
4732 }
4733
4734 /* shift the data */
4735 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4736 data_end - data_size, btrfs_leaf_data(leaf) +
4737 data_end, old_data - data_end);
4738
4739 data_end = old_data;
4740 old_size = btrfs_item_size_nr(leaf, slot);
4741 item = btrfs_item_nr(slot);
4742 btrfs_set_item_size(leaf, item, old_size + data_size);
4743 btrfs_mark_buffer_dirty(leaf);
4744
4745 if (btrfs_leaf_free_space(root, leaf) < 0) {
4746 btrfs_print_leaf(root, leaf);
4747 BUG();
4748 }
4749}
4750
4751/*
4752 * this is a helper for btrfs_insert_empty_items, the main goal here is
4753 * to save stack depth by doing the bulk of the work in a function
4754 * that doesn't call btrfs_search_slot
4755 */
4756void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4757 struct btrfs_key *cpu_key, u32 *data_size,
4758 u32 total_data, u32 total_size, int nr)
4759{
4760 struct btrfs_item *item;
4761 int i;
4762 u32 nritems;
4763 unsigned int data_end;
4764 struct btrfs_disk_key disk_key;
4765 struct extent_buffer *leaf;
4766 int slot;
4767 struct btrfs_map_token token;
4768
4769 btrfs_init_map_token(&token);
4770
4771 leaf = path->nodes[0];
4772 slot = path->slots[0];
4773
4774 nritems = btrfs_header_nritems(leaf);
4775 data_end = leaf_data_end(root, leaf);
4776
4777 if (btrfs_leaf_free_space(root, leaf) < total_size) {
4778 btrfs_print_leaf(root, leaf);
4779 btrfs_crit(root->fs_info, "not enough freespace need %u have %d",
4780 total_size, btrfs_leaf_free_space(root, leaf));
4781 BUG();
4782 }
4783
4784 if (slot != nritems) {
4785 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4786
4787 if (old_data < data_end) {
4788 btrfs_print_leaf(root, leaf);
4789 btrfs_crit(root->fs_info, "slot %d old_data %d data_end %d",
4790 slot, old_data, data_end);
4791 BUG_ON(1);
4792 }
4793 /*
4794 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4795 */
4796 /* first correct the data pointers */
4797 for (i = slot; i < nritems; i++) {
4798 u32 ioff;
4799
4800 item = btrfs_item_nr( i);
4801 ioff = btrfs_token_item_offset(leaf, item, &token);
4802 btrfs_set_token_item_offset(leaf, item,
4803 ioff - total_data, &token);
4804 }
4805 /* shift the items */
4806 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4807 btrfs_item_nr_offset(slot),
4808 (nritems - slot) * sizeof(struct btrfs_item));
4809
4810 /* shift the data */
4811 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4812 data_end - total_data, btrfs_leaf_data(leaf) +
4813 data_end, old_data - data_end);
4814 data_end = old_data;
4815 }
4816
4817 /* setup the item for the new data */
4818 for (i = 0; i < nr; i++) {
4819 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4820 btrfs_set_item_key(leaf, &disk_key, slot + i);
4821 item = btrfs_item_nr(slot + i);
4822 btrfs_set_token_item_offset(leaf, item,
4823 data_end - data_size[i], &token);
4824 data_end -= data_size[i];
4825 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4826 }
4827
4828 btrfs_set_header_nritems(leaf, nritems + nr);
4829
4830 if (slot == 0) {
4831 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4832 fixup_low_keys(root, path, &disk_key, 1);
4833 }
4834 btrfs_unlock_up_safe(path, 1);
4835 btrfs_mark_buffer_dirty(leaf);
4836
4837 if (btrfs_leaf_free_space(root, leaf) < 0) {
4838 btrfs_print_leaf(root, leaf);
4839 BUG();
4840 }
4841}
4842
4843/*
4844 * Given a key and some data, insert items into the tree.
4845 * This does all the path init required, making room in the tree if needed.
4846 */
4847int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4848 struct btrfs_root *root,
4849 struct btrfs_path *path,
4850 struct btrfs_key *cpu_key, u32 *data_size,
4851 int nr)
4852{
4853 int ret = 0;
4854 int slot;
4855 int i;
4856 u32 total_size = 0;
4857 u32 total_data = 0;
4858
4859 for (i = 0; i < nr; i++)
4860 total_data += data_size[i];
4861
4862 total_size = total_data + (nr * sizeof(struct btrfs_item));
4863 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4864 if (ret == 0)
4865 return -EEXIST;
4866 if (ret < 0)
4867 return ret;
4868
4869 slot = path->slots[0];
4870 BUG_ON(slot < 0);
4871
4872 setup_items_for_insert(root, path, cpu_key, data_size,
4873 total_data, total_size, nr);
4874 return 0;
4875}
4876
4877/*
4878 * Given a key and some data, insert an item into the tree.
4879 * This does all the path init required, making room in the tree if needed.
4880 */
4881int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
4882 *root, struct btrfs_key *cpu_key, void *data, u32
4883 data_size)
4884{
4885 int ret = 0;
4886 struct btrfs_path *path;
4887 struct extent_buffer *leaf;
4888 unsigned long ptr;
4889
4890 path = btrfs_alloc_path();
4891 if (!path)
4892 return -ENOMEM;
4893 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4894 if (!ret) {
4895 leaf = path->nodes[0];
4896 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4897 write_extent_buffer(leaf, data, ptr, data_size);
4898 btrfs_mark_buffer_dirty(leaf);
4899 }
4900 btrfs_free_path(path);
4901 return ret;
4902}
4903
4904/*
4905 * delete the pointer from a given node.
4906 *
4907 * the tree should have been previously balanced so the deletion does not
4908 * empty a node.
4909 */
4910static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4911 int level, int slot)
4912{
4913 struct extent_buffer *parent = path->nodes[level];
4914 u32 nritems;
4915 int ret;
4916
4917 nritems = btrfs_header_nritems(parent);
4918 if (slot != nritems - 1) {
4919 if (level)
4920 tree_mod_log_eb_move(root->fs_info, parent, slot,
4921 slot + 1, nritems - slot - 1);
4922 memmove_extent_buffer(parent,
4923 btrfs_node_key_ptr_offset(slot),
4924 btrfs_node_key_ptr_offset(slot + 1),
4925 sizeof(struct btrfs_key_ptr) *
4926 (nritems - slot - 1));
4927 } else if (level) {
4928 ret = tree_mod_log_insert_key(root->fs_info, parent, slot,
4929 MOD_LOG_KEY_REMOVE, GFP_NOFS);
4930 BUG_ON(ret < 0);
4931 }
4932
4933 nritems--;
4934 btrfs_set_header_nritems(parent, nritems);
4935 if (nritems == 0 && parent == root->node) {
4936 BUG_ON(btrfs_header_level(root->node) != 1);
4937 /* just turn the root into a leaf and break */
4938 btrfs_set_header_level(root->node, 0);
4939 } else if (slot == 0) {
4940 struct btrfs_disk_key disk_key;
4941
4942 btrfs_node_key(parent, &disk_key, 0);
4943 fixup_low_keys(root, path, &disk_key, level + 1);
4944 }
4945 btrfs_mark_buffer_dirty(parent);
4946}
4947
4948/*
4949 * a helper function to delete the leaf pointed to by path->slots[1] and
4950 * path->nodes[1].
4951 *
4952 * This deletes the pointer in path->nodes[1] and frees the leaf
4953 * block extent. zero is returned if it all worked out, < 0 otherwise.
4954 *
4955 * The path must have already been setup for deleting the leaf, including
4956 * all the proper balancing. path->nodes[1] must be locked.
4957 */
4958static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4959 struct btrfs_root *root,
4960 struct btrfs_path *path,
4961 struct extent_buffer *leaf)
4962{
4963 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4964 del_ptr(root, path, 1, path->slots[1]);
4965
4966 /*
4967 * btrfs_free_extent is expensive, we want to make sure we
4968 * aren't holding any locks when we call it
4969 */
4970 btrfs_unlock_up_safe(path, 0);
4971
4972 root_sub_used(root, leaf->len);
4973
4974 extent_buffer_get(leaf);
4975 btrfs_free_tree_block(trans, root, leaf, 0, 1);
4976 free_extent_buffer_stale(leaf);
4977}
4978/*
4979 * delete the item at the leaf level in path. If that empties
4980 * the leaf, remove it from the tree
4981 */
4982int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4983 struct btrfs_path *path, int slot, int nr)
4984{
4985 struct extent_buffer *leaf;
4986 struct btrfs_item *item;
4987 int last_off;
4988 int dsize = 0;
4989 int ret = 0;
4990 int wret;
4991 int i;
4992 u32 nritems;
4993 struct btrfs_map_token token;
4994
4995 btrfs_init_map_token(&token);
4996
4997 leaf = path->nodes[0];
4998 last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4999
5000 for (i = 0; i < nr; i++)
5001 dsize += btrfs_item_size_nr(leaf, slot + i);
5002
5003 nritems = btrfs_header_nritems(leaf);
5004
5005 if (slot + nr != nritems) {
5006 int data_end = leaf_data_end(root, leaf);
5007
5008 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
5009 data_end + dsize,
5010 btrfs_leaf_data(leaf) + data_end,
5011 last_off - data_end);
5012
5013 for (i = slot + nr; i < nritems; i++) {
5014 u32 ioff;
5015
5016 item = btrfs_item_nr(i);
5017 ioff = btrfs_token_item_offset(leaf, item, &token);
5018 btrfs_set_token_item_offset(leaf, item,
5019 ioff + dsize, &token);
5020 }
5021
5022 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
5023 btrfs_item_nr_offset(slot + nr),
5024 sizeof(struct btrfs_item) *
5025 (nritems - slot - nr));
5026 }
5027 btrfs_set_header_nritems(leaf, nritems - nr);
5028 nritems -= nr;
5029
5030 /* delete the leaf if we've emptied it */
5031 if (nritems == 0) {
5032 if (leaf == root->node) {
5033 btrfs_set_header_level(leaf, 0);
5034 } else {
5035 btrfs_set_path_blocking(path);
5036 clean_tree_block(trans, root, leaf);
5037 btrfs_del_leaf(trans, root, path, leaf);
5038 }
5039 } else {
5040 int used = leaf_space_used(leaf, 0, nritems);
5041 if (slot == 0) {
5042 struct btrfs_disk_key disk_key;
5043
5044 btrfs_item_key(leaf, &disk_key, 0);
5045 fixup_low_keys(root, path, &disk_key, 1);
5046 }
5047
5048 /* delete the leaf if it is mostly empty */
5049 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
5050 /* push_leaf_left fixes the path.
5051 * make sure the path still points to our leaf
5052 * for possible call to del_ptr below
5053 */
5054 slot = path->slots[1];
5055 extent_buffer_get(leaf);
5056
5057 btrfs_set_path_blocking(path);
5058 wret = push_leaf_left(trans, root, path, 1, 1,
5059 1, (u32)-1);
5060 if (wret < 0 && wret != -ENOSPC)
5061 ret = wret;
5062
5063 if (path->nodes[0] == leaf &&
5064 btrfs_header_nritems(leaf)) {
5065 wret = push_leaf_right(trans, root, path, 1,
5066 1, 1, 0);
5067 if (wret < 0 && wret != -ENOSPC)
5068 ret = wret;
5069 }
5070
5071 if (btrfs_header_nritems(leaf) == 0) {
5072 path->slots[1] = slot;
5073 btrfs_del_leaf(trans, root, path, leaf);
5074 free_extent_buffer(leaf);
5075 ret = 0;
5076 } else {
5077 /* if we're still in the path, make sure
5078 * we're dirty. Otherwise, one of the
5079 * push_leaf functions must have already
5080 * dirtied this buffer
5081 */
5082 if (path->nodes[0] == leaf)
5083 btrfs_mark_buffer_dirty(leaf);
5084 free_extent_buffer(leaf);
5085 }
5086 } else {
5087 btrfs_mark_buffer_dirty(leaf);
5088 }
5089 }
5090 return ret;
5091}
5092
5093/*
5094 * search the tree again to find a leaf with lesser keys
5095 * returns 0 if it found something or 1 if there are no lesser leaves.
5096 * returns < 0 on io errors.
5097 *
5098 * This may release the path, and so you may lose any locks held at the
5099 * time you call it.
5100 */
5101int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5102{
5103 struct btrfs_key key;
5104 struct btrfs_disk_key found_key;
5105 int ret;
5106
5107 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5108
5109 if (key.offset > 0) {
5110 key.offset--;
5111 } else if (key.type > 0) {
5112 key.type--;
5113 key.offset = (u64)-1;
5114 } else if (key.objectid > 0) {
5115 key.objectid--;
5116 key.type = (u8)-1;
5117 key.offset = (u64)-1;
5118 } else {
5119 return 1;
5120 }
5121
5122 btrfs_release_path(path);
5123 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5124 if (ret < 0)
5125 return ret;
5126 btrfs_item_key(path->nodes[0], &found_key, 0);
5127 ret = comp_keys(&found_key, &key);
5128 if (ret < 0)
5129 return 0;
5130 return 1;
5131}
5132
5133/*
5134 * A helper function to walk down the tree starting at min_key, and looking
5135 * for nodes or leaves that are have a minimum transaction id.
5136 * This is used by the btree defrag code, and tree logging
5137 *
5138 * This does not cow, but it does stuff the starting key it finds back
5139 * into min_key, so you can call btrfs_search_slot with cow=1 on the
5140 * key and get a writable path.
5141 *
5142 * This does lock as it descends, and path->keep_locks should be set
5143 * to 1 by the caller.
5144 *
5145 * This honors path->lowest_level to prevent descent past a given level
5146 * of the tree.
5147 *
5148 * min_trans indicates the oldest transaction that you are interested
5149 * in walking through. Any nodes or leaves older than min_trans are
5150 * skipped over (without reading them).
5151 *
5152 * returns zero if something useful was found, < 0 on error and 1 if there
5153 * was nothing in the tree that matched the search criteria.
5154 */
5155int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5156 struct btrfs_path *path,
5157 u64 min_trans)
5158{
5159 struct extent_buffer *cur;
5160 struct btrfs_key found_key;
5161 int slot;
5162 int sret;
5163 u32 nritems;
5164 int level;
5165 int ret = 1;
5166
5167 WARN_ON(!path->keep_locks);
5168again:
5169 cur = btrfs_read_lock_root_node(root);
5170 level = btrfs_header_level(cur);
5171 WARN_ON(path->nodes[level]);
5172 path->nodes[level] = cur;
5173 path->locks[level] = BTRFS_READ_LOCK;
5174
5175 if (btrfs_header_generation(cur) < min_trans) {
5176 ret = 1;
5177 goto out;
5178 }
5179 while (1) {
5180 nritems = btrfs_header_nritems(cur);
5181 level = btrfs_header_level(cur);
5182 sret = bin_search(cur, min_key, level, &slot);
5183
5184 /* at the lowest level, we're done, setup the path and exit */
5185 if (level == path->lowest_level) {
5186 if (slot >= nritems)
5187 goto find_next_key;
5188 ret = 0;
5189 path->slots[level] = slot;
5190 btrfs_item_key_to_cpu(cur, &found_key, slot);
5191 goto out;
5192 }
5193 if (sret && slot > 0)
5194 slot--;
5195 /*
5196 * check this node pointer against the min_trans parameters.
5197 * If it is too old, old, skip to the next one.
5198 */
5199 while (slot < nritems) {
5200 u64 gen;
5201
5202 gen = btrfs_node_ptr_generation(cur, slot);
5203 if (gen < min_trans) {
5204 slot++;
5205 continue;
5206 }
5207 break;
5208 }
5209find_next_key:
5210 /*
5211 * we didn't find a candidate key in this node, walk forward
5212 * and find another one
5213 */
5214 if (slot >= nritems) {
5215 path->slots[level] = slot;
5216 btrfs_set_path_blocking(path);
5217 sret = btrfs_find_next_key(root, path, min_key, level,
5218 min_trans);
5219 if (sret == 0) {
5220 btrfs_release_path(path);
5221 goto again;
5222 } else {
5223 goto out;
5224 }
5225 }
5226 /* save our key for returning back */
5227 btrfs_node_key_to_cpu(cur, &found_key, slot);
5228 path->slots[level] = slot;
5229 if (level == path->lowest_level) {
5230 ret = 0;
5231 unlock_up(path, level, 1, 0, NULL);
5232 goto out;
5233 }
5234 btrfs_set_path_blocking(path);
5235 cur = read_node_slot(root, cur, slot);
5236 BUG_ON(!cur); /* -ENOMEM */
5237
5238 btrfs_tree_read_lock(cur);
5239
5240 path->locks[level - 1] = BTRFS_READ_LOCK;
5241 path->nodes[level - 1] = cur;
5242 unlock_up(path, level, 1, 0, NULL);
5243 btrfs_clear_path_blocking(path, NULL, 0);
5244 }
5245out:
5246 if (ret == 0)
5247 memcpy(min_key, &found_key, sizeof(found_key));
5248 btrfs_set_path_blocking(path);
5249 return ret;
5250}
5251
5252static void tree_move_down(struct btrfs_root *root,
5253 struct btrfs_path *path,
5254 int *level, int root_level)
5255{
5256 BUG_ON(*level == 0);
5257 path->nodes[*level - 1] = read_node_slot(root, path->nodes[*level],
5258 path->slots[*level]);
5259 path->slots[*level - 1] = 0;
5260 (*level)--;
5261}
5262
5263static int tree_move_next_or_upnext(struct btrfs_root *root,
5264 struct btrfs_path *path,
5265 int *level, int root_level)
5266{
5267 int ret = 0;
5268 int nritems;
5269 nritems = btrfs_header_nritems(path->nodes[*level]);
5270
5271 path->slots[*level]++;
5272
5273 while (path->slots[*level] >= nritems) {
5274 if (*level == root_level)
5275 return -1;
5276
5277 /* move upnext */
5278 path->slots[*level] = 0;
5279 free_extent_buffer(path->nodes[*level]);
5280 path->nodes[*level] = NULL;
5281 (*level)++;
5282 path->slots[*level]++;
5283
5284 nritems = btrfs_header_nritems(path->nodes[*level]);
5285 ret = 1;
5286 }
5287 return ret;
5288}
5289
5290/*
5291 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
5292 * or down.
5293 */
5294static int tree_advance(struct btrfs_root *root,
5295 struct btrfs_path *path,
5296 int *level, int root_level,
5297 int allow_down,
5298 struct btrfs_key *key)
5299{
5300 int ret;
5301
5302 if (*level == 0 || !allow_down) {
5303 ret = tree_move_next_or_upnext(root, path, level, root_level);
5304 } else {
5305 tree_move_down(root, path, level, root_level);
5306 ret = 0;
5307 }
5308 if (ret >= 0) {
5309 if (*level == 0)
5310 btrfs_item_key_to_cpu(path->nodes[*level], key,
5311 path->slots[*level]);
5312 else
5313 btrfs_node_key_to_cpu(path->nodes[*level], key,
5314 path->slots[*level]);
5315 }
5316 return ret;
5317}
5318
5319static int tree_compare_item(struct btrfs_root *left_root,
5320 struct btrfs_path *left_path,
5321 struct btrfs_path *right_path,
5322 char *tmp_buf)
5323{
5324 int cmp;
5325 int len1, len2;
5326 unsigned long off1, off2;
5327
5328 len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
5329 len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
5330 if (len1 != len2)
5331 return 1;
5332
5333 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
5334 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
5335 right_path->slots[0]);
5336
5337 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
5338
5339 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
5340 if (cmp)
5341 return 1;
5342 return 0;
5343}
5344
5345#define ADVANCE 1
5346#define ADVANCE_ONLY_NEXT -1
5347
5348/*
5349 * This function compares two trees and calls the provided callback for
5350 * every changed/new/deleted item it finds.
5351 * If shared tree blocks are encountered, whole subtrees are skipped, making
5352 * the compare pretty fast on snapshotted subvolumes.
5353 *
5354 * This currently works on commit roots only. As commit roots are read only,
5355 * we don't do any locking. The commit roots are protected with transactions.
5356 * Transactions are ended and rejoined when a commit is tried in between.
5357 *
5358 * This function checks for modifications done to the trees while comparing.
5359 * If it detects a change, it aborts immediately.
5360 */
5361int btrfs_compare_trees(struct btrfs_root *left_root,
5362 struct btrfs_root *right_root,
5363 btrfs_changed_cb_t changed_cb, void *ctx)
5364{
5365 int ret;
5366 int cmp;
5367 struct btrfs_path *left_path = NULL;
5368 struct btrfs_path *right_path = NULL;
5369 struct btrfs_key left_key;
5370 struct btrfs_key right_key;
5371 char *tmp_buf = NULL;
5372 int left_root_level;
5373 int right_root_level;
5374 int left_level;
5375 int right_level;
5376 int left_end_reached;
5377 int right_end_reached;
5378 int advance_left;
5379 int advance_right;
5380 u64 left_blockptr;
5381 u64 right_blockptr;
5382 u64 left_gen;
5383 u64 right_gen;
5384
5385 left_path = btrfs_alloc_path();
5386 if (!left_path) {
5387 ret = -ENOMEM;
5388 goto out;
5389 }
5390 right_path = btrfs_alloc_path();
5391 if (!right_path) {
5392 ret = -ENOMEM;
5393 goto out;
5394 }
5395
5396 tmp_buf = kmalloc(left_root->leafsize, GFP_NOFS);
5397 if (!tmp_buf) {
5398 ret = -ENOMEM;
5399 goto out;
5400 }
5401
5402 left_path->search_commit_root = 1;
5403 left_path->skip_locking = 1;
5404 right_path->search_commit_root = 1;
5405 right_path->skip_locking = 1;
5406
5407 /*
5408 * Strategy: Go to the first items of both trees. Then do
5409 *
5410 * If both trees are at level 0
5411 * Compare keys of current items
5412 * If left < right treat left item as new, advance left tree
5413 * and repeat
5414 * If left > right treat right item as deleted, advance right tree
5415 * and repeat
5416 * If left == right do deep compare of items, treat as changed if
5417 * needed, advance both trees and repeat
5418 * If both trees are at the same level but not at level 0
5419 * Compare keys of current nodes/leafs
5420 * If left < right advance left tree and repeat
5421 * If left > right advance right tree and repeat
5422 * If left == right compare blockptrs of the next nodes/leafs
5423 * If they match advance both trees but stay at the same level
5424 * and repeat
5425 * If they don't match advance both trees while allowing to go
5426 * deeper and repeat
5427 * If tree levels are different
5428 * Advance the tree that needs it and repeat
5429 *
5430 * Advancing a tree means:
5431 * If we are at level 0, try to go to the next slot. If that's not
5432 * possible, go one level up and repeat. Stop when we found a level
5433 * where we could go to the next slot. We may at this point be on a
5434 * node or a leaf.
5435 *
5436 * If we are not at level 0 and not on shared tree blocks, go one
5437 * level deeper.
5438 *
5439 * If we are not at level 0 and on shared tree blocks, go one slot to
5440 * the right if possible or go up and right.
5441 */
5442
5443 down_read(&left_root->fs_info->commit_root_sem);
5444 left_level = btrfs_header_level(left_root->commit_root);
5445 left_root_level = left_level;
5446 left_path->nodes[left_level] = left_root->commit_root;
5447 extent_buffer_get(left_path->nodes[left_level]);
5448
5449 right_level = btrfs_header_level(right_root->commit_root);
5450 right_root_level = right_level;
5451 right_path->nodes[right_level] = right_root->commit_root;
5452 extent_buffer_get(right_path->nodes[right_level]);
5453 up_read(&left_root->fs_info->commit_root_sem);
5454
5455 if (left_level == 0)
5456 btrfs_item_key_to_cpu(left_path->nodes[left_level],
5457 &left_key, left_path->slots[left_level]);
5458 else
5459 btrfs_node_key_to_cpu(left_path->nodes[left_level],
5460 &left_key, left_path->slots[left_level]);
5461 if (right_level == 0)
5462 btrfs_item_key_to_cpu(right_path->nodes[right_level],
5463 &right_key, right_path->slots[right_level]);
5464 else
5465 btrfs_node_key_to_cpu(right_path->nodes[right_level],
5466 &right_key, right_path->slots[right_level]);
5467
5468 left_end_reached = right_end_reached = 0;
5469 advance_left = advance_right = 0;
5470
5471 while (1) {
5472 if (advance_left && !left_end_reached) {
5473 ret = tree_advance(left_root, left_path, &left_level,
5474 left_root_level,
5475 advance_left != ADVANCE_ONLY_NEXT,
5476 &left_key);
5477 if (ret < 0)
5478 left_end_reached = ADVANCE;
5479 advance_left = 0;
5480 }
5481 if (advance_right && !right_end_reached) {
5482 ret = tree_advance(right_root, right_path, &right_level,
5483 right_root_level,
5484 advance_right != ADVANCE_ONLY_NEXT,
5485 &right_key);
5486 if (ret < 0)
5487 right_end_reached = ADVANCE;
5488 advance_right = 0;
5489 }
5490
5491 if (left_end_reached && right_end_reached) {
5492 ret = 0;
5493 goto out;
5494 } else if (left_end_reached) {
5495 if (right_level == 0) {
5496 ret = changed_cb(left_root, right_root,
5497 left_path, right_path,
5498 &right_key,
5499 BTRFS_COMPARE_TREE_DELETED,
5500 ctx);
5501 if (ret < 0)
5502 goto out;
5503 }
5504 advance_right = ADVANCE;
5505 continue;
5506 } else if (right_end_reached) {
5507 if (left_level == 0) {
5508 ret = changed_cb(left_root, right_root,
5509 left_path, right_path,
5510 &left_key,
5511 BTRFS_COMPARE_TREE_NEW,
5512 ctx);
5513 if (ret < 0)
5514 goto out;
5515 }
5516 advance_left = ADVANCE;
5517 continue;
5518 }
5519
5520 if (left_level == 0 && right_level == 0) {
5521 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5522 if (cmp < 0) {
5523 ret = changed_cb(left_root, right_root,
5524 left_path, right_path,
5525 &left_key,
5526 BTRFS_COMPARE_TREE_NEW,
5527 ctx);
5528 if (ret < 0)
5529 goto out;
5530 advance_left = ADVANCE;
5531 } else if (cmp > 0) {
5532 ret = changed_cb(left_root, right_root,
5533 left_path, right_path,
5534 &right_key,
5535 BTRFS_COMPARE_TREE_DELETED,
5536 ctx);
5537 if (ret < 0)
5538 goto out;
5539 advance_right = ADVANCE;
5540 } else {
5541 enum btrfs_compare_tree_result cmp;
5542
5543 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
5544 ret = tree_compare_item(left_root, left_path,
5545 right_path, tmp_buf);
5546 if (ret)
5547 cmp = BTRFS_COMPARE_TREE_CHANGED;
5548 else
5549 cmp = BTRFS_COMPARE_TREE_SAME;
5550 ret = changed_cb(left_root, right_root,
5551 left_path, right_path,
5552 &left_key, cmp, ctx);
5553 if (ret < 0)
5554 goto out;
5555 advance_left = ADVANCE;
5556 advance_right = ADVANCE;
5557 }
5558 } else if (left_level == right_level) {
5559 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5560 if (cmp < 0) {
5561 advance_left = ADVANCE;
5562 } else if (cmp > 0) {
5563 advance_right = ADVANCE;
5564 } else {
5565 left_blockptr = btrfs_node_blockptr(
5566 left_path->nodes[left_level],
5567 left_path->slots[left_level]);
5568 right_blockptr = btrfs_node_blockptr(
5569 right_path->nodes[right_level],
5570 right_path->slots[right_level]);
5571 left_gen = btrfs_node_ptr_generation(
5572 left_path->nodes[left_level],
5573 left_path->slots[left_level]);
5574 right_gen = btrfs_node_ptr_generation(
5575 right_path->nodes[right_level],
5576 right_path->slots[right_level]);
5577 if (left_blockptr == right_blockptr &&
5578 left_gen == right_gen) {
5579 /*
5580 * As we're on a shared block, don't
5581 * allow to go deeper.
5582 */
5583 advance_left = ADVANCE_ONLY_NEXT;
5584 advance_right = ADVANCE_ONLY_NEXT;
5585 } else {
5586 advance_left = ADVANCE;
5587 advance_right = ADVANCE;
5588 }
5589 }
5590 } else if (left_level < right_level) {
5591 advance_right = ADVANCE;
5592 } else {
5593 advance_left = ADVANCE;
5594 }
5595 }
5596
5597out:
5598 btrfs_free_path(left_path);
5599 btrfs_free_path(right_path);
5600 kfree(tmp_buf);
5601 return ret;
5602}
5603
5604/*
5605 * this is similar to btrfs_next_leaf, but does not try to preserve
5606 * and fixup the path. It looks for and returns the next key in the
5607 * tree based on the current path and the min_trans parameters.
5608 *
5609 * 0 is returned if another key is found, < 0 if there are any errors
5610 * and 1 is returned if there are no higher keys in the tree
5611 *
5612 * path->keep_locks should be set to 1 on the search made before
5613 * calling this function.
5614 */
5615int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5616 struct btrfs_key *key, int level, u64 min_trans)
5617{
5618 int slot;
5619 struct extent_buffer *c;
5620
5621 WARN_ON(!path->keep_locks);
5622 while (level < BTRFS_MAX_LEVEL) {
5623 if (!path->nodes[level])
5624 return 1;
5625
5626 slot = path->slots[level] + 1;
5627 c = path->nodes[level];
5628next:
5629 if (slot >= btrfs_header_nritems(c)) {
5630 int ret;
5631 int orig_lowest;
5632 struct btrfs_key cur_key;
5633 if (level + 1 >= BTRFS_MAX_LEVEL ||
5634 !path->nodes[level + 1])
5635 return 1;
5636
5637 if (path->locks[level + 1]) {
5638 level++;
5639 continue;
5640 }
5641
5642 slot = btrfs_header_nritems(c) - 1;
5643 if (level == 0)
5644 btrfs_item_key_to_cpu(c, &cur_key, slot);
5645 else
5646 btrfs_node_key_to_cpu(c, &cur_key, slot);
5647
5648 orig_lowest = path->lowest_level;
5649 btrfs_release_path(path);
5650 path->lowest_level = level;
5651 ret = btrfs_search_slot(NULL, root, &cur_key, path,
5652 0, 0);
5653 path->lowest_level = orig_lowest;
5654 if (ret < 0)
5655 return ret;
5656
5657 c = path->nodes[level];
5658 slot = path->slots[level];
5659 if (ret == 0)
5660 slot++;
5661 goto next;
5662 }
5663
5664 if (level == 0)
5665 btrfs_item_key_to_cpu(c, key, slot);
5666 else {
5667 u64 gen = btrfs_node_ptr_generation(c, slot);
5668
5669 if (gen < min_trans) {
5670 slot++;
5671 goto next;
5672 }
5673 btrfs_node_key_to_cpu(c, key, slot);
5674 }
5675 return 0;
5676 }
5677 return 1;
5678}
5679
5680/*
5681 * search the tree again to find a leaf with greater keys
5682 * returns 0 if it found something or 1 if there are no greater leaves.
5683 * returns < 0 on io errors.
5684 */
5685int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5686{
5687 return btrfs_next_old_leaf(root, path, 0);
5688}
5689
5690int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5691 u64 time_seq)
5692{
5693 int slot;
5694 int level;
5695 struct extent_buffer *c;
5696 struct extent_buffer *next;
5697 struct btrfs_key key;
5698 u32 nritems;
5699 int ret;
5700 int old_spinning = path->leave_spinning;
5701 int next_rw_lock = 0;
5702
5703 nritems = btrfs_header_nritems(path->nodes[0]);
5704 if (nritems == 0)
5705 return 1;
5706
5707 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5708again:
5709 level = 1;
5710 next = NULL;
5711 next_rw_lock = 0;
5712 btrfs_release_path(path);
5713
5714 path->keep_locks = 1;
5715 path->leave_spinning = 1;
5716
5717 if (time_seq)
5718 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5719 else
5720 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5721 path->keep_locks = 0;
5722
5723 if (ret < 0)
5724 return ret;
5725
5726 nritems = btrfs_header_nritems(path->nodes[0]);
5727 /*
5728 * by releasing the path above we dropped all our locks. A balance
5729 * could have added more items next to the key that used to be
5730 * at the very end of the block. So, check again here and
5731 * advance the path if there are now more items available.
5732 */
5733 if (nritems > 0 && path->slots[0] < nritems - 1) {
5734 if (ret == 0)
5735 path->slots[0]++;
5736 ret = 0;
5737 goto done;
5738 }
5739
5740 while (level < BTRFS_MAX_LEVEL) {
5741 if (!path->nodes[level]) {
5742 ret = 1;
5743 goto done;
5744 }
5745
5746 slot = path->slots[level] + 1;
5747 c = path->nodes[level];
5748 if (slot >= btrfs_header_nritems(c)) {
5749 level++;
5750 if (level == BTRFS_MAX_LEVEL) {
5751 ret = 1;
5752 goto done;
5753 }
5754 continue;
5755 }
5756
5757 if (next) {
5758 btrfs_tree_unlock_rw(next, next_rw_lock);
5759 free_extent_buffer(next);
5760 }
5761
5762 next = c;
5763 next_rw_lock = path->locks[level];
5764 ret = read_block_for_search(NULL, root, path, &next, level,
5765 slot, &key, 0);
5766 if (ret == -EAGAIN)
5767 goto again;
5768
5769 if (ret < 0) {
5770 btrfs_release_path(path);
5771 goto done;
5772 }
5773
5774 if (!path->skip_locking) {
5775 ret = btrfs_try_tree_read_lock(next);
5776 if (!ret && time_seq) {
5777 /*
5778 * If we don't get the lock, we may be racing
5779 * with push_leaf_left, holding that lock while
5780 * itself waiting for the leaf we've currently
5781 * locked. To solve this situation, we give up
5782 * on our lock and cycle.
5783 */
5784 free_extent_buffer(next);
5785 btrfs_release_path(path);
5786 cond_resched();
5787 goto again;
5788 }
5789 if (!ret) {
5790 btrfs_set_path_blocking(path);
5791 btrfs_tree_read_lock(next);
5792 btrfs_clear_path_blocking(path, next,
5793 BTRFS_READ_LOCK);
5794 }
5795 next_rw_lock = BTRFS_READ_LOCK;
5796 }
5797 break;
5798 }
5799 path->slots[level] = slot;
5800 while (1) {
5801 level--;
5802 c = path->nodes[level];
5803 if (path->locks[level])
5804 btrfs_tree_unlock_rw(c, path->locks[level]);
5805
5806 free_extent_buffer(c);
5807 path->nodes[level] = next;
5808 path->slots[level] = 0;
5809 if (!path->skip_locking)
5810 path->locks[level] = next_rw_lock;
5811 if (!level)
5812 break;
5813
5814 ret = read_block_for_search(NULL, root, path, &next, level,
5815 0, &key, 0);
5816 if (ret == -EAGAIN)
5817 goto again;
5818
5819 if (ret < 0) {
5820 btrfs_release_path(path);
5821 goto done;
5822 }
5823
5824 if (!path->skip_locking) {
5825 ret = btrfs_try_tree_read_lock(next);
5826 if (!ret) {
5827 btrfs_set_path_blocking(path);
5828 btrfs_tree_read_lock(next);
5829 btrfs_clear_path_blocking(path, next,
5830 BTRFS_READ_LOCK);
5831 }
5832 next_rw_lock = BTRFS_READ_LOCK;
5833 }
5834 }
5835 ret = 0;
5836done:
5837 unlock_up(path, 0, 1, 0, NULL);
5838 path->leave_spinning = old_spinning;
5839 if (!old_spinning)
5840 btrfs_set_path_blocking(path);
5841
5842 return ret;
5843}
5844
5845/*
5846 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5847 * searching until it gets past min_objectid or finds an item of 'type'
5848 *
5849 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5850 */
5851int btrfs_previous_item(struct btrfs_root *root,
5852 struct btrfs_path *path, u64 min_objectid,
5853 int type)
5854{
5855 struct btrfs_key found_key;
5856 struct extent_buffer *leaf;
5857 u32 nritems;
5858 int ret;
5859
5860 while (1) {
5861 if (path->slots[0] == 0) {
5862 btrfs_set_path_blocking(path);
5863 ret = btrfs_prev_leaf(root, path);
5864 if (ret != 0)
5865 return ret;
5866 } else {
5867 path->slots[0]--;
5868 }
5869 leaf = path->nodes[0];
5870 nritems = btrfs_header_nritems(leaf);
5871 if (nritems == 0)
5872 return 1;
5873 if (path->slots[0] == nritems)
5874 path->slots[0]--;
5875
5876 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5877 if (found_key.objectid < min_objectid)
5878 break;
5879 if (found_key.type == type)
5880 return 0;
5881 if (found_key.objectid == min_objectid &&
5882 found_key.type < type)
5883 break;
5884 }
5885 return 1;
5886}
5887
5888/*
5889 * search in extent tree to find a previous Metadata/Data extent item with
5890 * min objecitd.
5891 *
5892 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5893 */
5894int btrfs_previous_extent_item(struct btrfs_root *root,
5895 struct btrfs_path *path, u64 min_objectid)
5896{
5897 struct btrfs_key found_key;
5898 struct extent_buffer *leaf;
5899 u32 nritems;
5900 int ret;
5901
5902 while (1) {
5903 if (path->slots[0] == 0) {
5904 btrfs_set_path_blocking(path);
5905 ret = btrfs_prev_leaf(root, path);
5906 if (ret != 0)
5907 return ret;
5908 } else {
5909 path->slots[0]--;
5910 }
5911 leaf = path->nodes[0];
5912 nritems = btrfs_header_nritems(leaf);
5913 if (nritems == 0)
5914 return 1;
5915 if (path->slots[0] == nritems)
5916 path->slots[0]--;
5917
5918 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5919 if (found_key.objectid < min_objectid)
5920 break;
5921 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5922 found_key.type == BTRFS_METADATA_ITEM_KEY)
5923 return 0;
5924 if (found_key.objectid == min_objectid &&
5925 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5926 break;
5927 }
5928 return 1;
5929}