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 <linux/vmalloc.h>
23#include "ctree.h"
24#include "disk-io.h"
25#include "transaction.h"
26#include "print-tree.h"
27#include "locking.h"
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
32 *root, struct btrfs_key *ins_key,
33 struct btrfs_path *path, int data_size, int extend);
34static int push_node_left(struct btrfs_trans_handle *trans,
35 struct btrfs_fs_info *fs_info,
36 struct extent_buffer *dst,
37 struct extent_buffer *src, int empty);
38static int balance_node_right(struct btrfs_trans_handle *trans,
39 struct btrfs_fs_info *fs_info,
40 struct extent_buffer *dst_buf,
41 struct extent_buffer *src_buf);
42static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
43 int level, int slot);
44static int tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
45 struct extent_buffer *eb);
46
47struct btrfs_path *btrfs_alloc_path(void)
48{
49 return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
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 if (held) {
84 btrfs_set_lock_blocking_rw(held, held_rw);
85 if (held_rw == BTRFS_WRITE_LOCK)
86 held_rw = BTRFS_WRITE_LOCK_BLOCKING;
87 else if (held_rw == BTRFS_READ_LOCK)
88 held_rw = BTRFS_READ_LOCK_BLOCKING;
89 }
90 btrfs_set_path_blocking(p);
91
92 for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
93 if (p->nodes[i] && p->locks[i]) {
94 btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
95 if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
96 p->locks[i] = BTRFS_WRITE_LOCK;
97 else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
98 p->locks[i] = BTRFS_READ_LOCK;
99 }
100 }
101
102 if (held)
103 btrfs_clear_lock_blocking_rw(held, held_rw);
104}
105
106/* this also releases the path */
107void btrfs_free_path(struct btrfs_path *p)
108{
109 if (!p)
110 return;
111 btrfs_release_path(p);
112 kmem_cache_free(btrfs_path_cachep, p);
113}
114
115/*
116 * path release drops references on the extent buffers in the path
117 * and it drops any locks held by this path
118 *
119 * It is safe to call this on paths that no locks or extent buffers held.
120 */
121noinline void btrfs_release_path(struct btrfs_path *p)
122{
123 int i;
124
125 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
126 p->slots[i] = 0;
127 if (!p->nodes[i])
128 continue;
129 if (p->locks[i]) {
130 btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
131 p->locks[i] = 0;
132 }
133 free_extent_buffer(p->nodes[i]);
134 p->nodes[i] = NULL;
135 }
136}
137
138/*
139 * safely gets a reference on the root node of a tree. A lock
140 * is not taken, so a concurrent writer may put a different node
141 * at the root of the tree. See btrfs_lock_root_node for the
142 * looping required.
143 *
144 * The extent buffer returned by this has a reference taken, so
145 * it won't disappear. It may stop being the root of the tree
146 * at any time because there are no locks held.
147 */
148struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
149{
150 struct extent_buffer *eb;
151
152 while (1) {
153 rcu_read_lock();
154 eb = rcu_dereference(root->node);
155
156 /*
157 * RCU really hurts here, we could free up the root node because
158 * it was COWed but we may not get the new root node yet so do
159 * the inc_not_zero dance and if it doesn't work then
160 * synchronize_rcu and try again.
161 */
162 if (atomic_inc_not_zero(&eb->refs)) {
163 rcu_read_unlock();
164 break;
165 }
166 rcu_read_unlock();
167 synchronize_rcu();
168 }
169 return eb;
170}
171
172/* loop around taking references on and locking the root node of the
173 * tree until you end up with a lock on the root. A locked buffer
174 * is returned, with a reference held.
175 */
176struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
177{
178 struct extent_buffer *eb;
179
180 while (1) {
181 eb = btrfs_root_node(root);
182 btrfs_tree_lock(eb);
183 if (eb == root->node)
184 break;
185 btrfs_tree_unlock(eb);
186 free_extent_buffer(eb);
187 }
188 return eb;
189}
190
191/* loop around taking references on and locking the root node of the
192 * tree until you end up with a lock on the root. A locked buffer
193 * is returned, with a reference held.
194 */
195static struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
196{
197 struct extent_buffer *eb;
198
199 while (1) {
200 eb = btrfs_root_node(root);
201 btrfs_tree_read_lock(eb);
202 if (eb == root->node)
203 break;
204 btrfs_tree_read_unlock(eb);
205 free_extent_buffer(eb);
206 }
207 return eb;
208}
209
210/* cowonly root (everything not a reference counted cow subvolume), just get
211 * put onto a simple dirty list. transaction.c walks this to make sure they
212 * get properly updated on disk.
213 */
214static void add_root_to_dirty_list(struct btrfs_root *root)
215{
216 struct btrfs_fs_info *fs_info = root->fs_info;
217
218 if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
219 !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
220 return;
221
222 spin_lock(&fs_info->trans_lock);
223 if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
224 /* Want the extent tree to be the last on the list */
225 if (root->objectid == BTRFS_EXTENT_TREE_OBJECTID)
226 list_move_tail(&root->dirty_list,
227 &fs_info->dirty_cowonly_roots);
228 else
229 list_move(&root->dirty_list,
230 &fs_info->dirty_cowonly_roots);
231 }
232 spin_unlock(&fs_info->trans_lock);
233}
234
235/*
236 * used by snapshot creation to make a copy of a root for a tree with
237 * a given objectid. The buffer with the new root node is returned in
238 * cow_ret, and this func returns zero on success or a negative error code.
239 */
240int btrfs_copy_root(struct btrfs_trans_handle *trans,
241 struct btrfs_root *root,
242 struct extent_buffer *buf,
243 struct extent_buffer **cow_ret, u64 new_root_objectid)
244{
245 struct btrfs_fs_info *fs_info = root->fs_info;
246 struct extent_buffer *cow;
247 int ret = 0;
248 int level;
249 struct btrfs_disk_key disk_key;
250
251 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
252 trans->transid != fs_info->running_transaction->transid);
253 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
254 trans->transid != root->last_trans);
255
256 level = btrfs_header_level(buf);
257 if (level == 0)
258 btrfs_item_key(buf, &disk_key, 0);
259 else
260 btrfs_node_key(buf, &disk_key, 0);
261
262 cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
263 &disk_key, level, buf->start, 0);
264 if (IS_ERR(cow))
265 return PTR_ERR(cow);
266
267 copy_extent_buffer_full(cow, buf);
268 btrfs_set_header_bytenr(cow, cow->start);
269 btrfs_set_header_generation(cow, trans->transid);
270 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
271 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
272 BTRFS_HEADER_FLAG_RELOC);
273 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
274 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
275 else
276 btrfs_set_header_owner(cow, new_root_objectid);
277
278 write_extent_buffer_fsid(cow, fs_info->fsid);
279
280 WARN_ON(btrfs_header_generation(buf) > trans->transid);
281 if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
282 ret = btrfs_inc_ref(trans, root, cow, 1);
283 else
284 ret = btrfs_inc_ref(trans, root, cow, 0);
285
286 if (ret)
287 return ret;
288
289 btrfs_mark_buffer_dirty(cow);
290 *cow_ret = cow;
291 return 0;
292}
293
294enum mod_log_op {
295 MOD_LOG_KEY_REPLACE,
296 MOD_LOG_KEY_ADD,
297 MOD_LOG_KEY_REMOVE,
298 MOD_LOG_KEY_REMOVE_WHILE_FREEING,
299 MOD_LOG_KEY_REMOVE_WHILE_MOVING,
300 MOD_LOG_MOVE_KEYS,
301 MOD_LOG_ROOT_REPLACE,
302};
303
304struct tree_mod_move {
305 int dst_slot;
306 int nr_items;
307};
308
309struct tree_mod_root {
310 u64 logical;
311 u8 level;
312};
313
314struct tree_mod_elem {
315 struct rb_node node;
316 u64 logical;
317 u64 seq;
318 enum mod_log_op op;
319
320 /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
321 int slot;
322
323 /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
324 u64 generation;
325
326 /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
327 struct btrfs_disk_key key;
328 u64 blockptr;
329
330 /* this is used for op == MOD_LOG_MOVE_KEYS */
331 struct tree_mod_move move;
332
333 /* this is used for op == MOD_LOG_ROOT_REPLACE */
334 struct tree_mod_root old_root;
335};
336
337static inline void tree_mod_log_read_lock(struct btrfs_fs_info *fs_info)
338{
339 read_lock(&fs_info->tree_mod_log_lock);
340}
341
342static inline void tree_mod_log_read_unlock(struct btrfs_fs_info *fs_info)
343{
344 read_unlock(&fs_info->tree_mod_log_lock);
345}
346
347static inline void tree_mod_log_write_lock(struct btrfs_fs_info *fs_info)
348{
349 write_lock(&fs_info->tree_mod_log_lock);
350}
351
352static inline void tree_mod_log_write_unlock(struct btrfs_fs_info *fs_info)
353{
354 write_unlock(&fs_info->tree_mod_log_lock);
355}
356
357/*
358 * Pull a new tree mod seq number for our operation.
359 */
360static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
361{
362 return atomic64_inc_return(&fs_info->tree_mod_seq);
363}
364
365/*
366 * This adds a new blocker to the tree mod log's blocker list if the @elem
367 * passed does not already have a sequence number set. So when a caller expects
368 * to record tree modifications, it should ensure to set elem->seq to zero
369 * before calling btrfs_get_tree_mod_seq.
370 * Returns a fresh, unused tree log modification sequence number, even if no new
371 * blocker was added.
372 */
373u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
374 struct seq_list *elem)
375{
376 tree_mod_log_write_lock(fs_info);
377 spin_lock(&fs_info->tree_mod_seq_lock);
378 if (!elem->seq) {
379 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
380 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
381 }
382 spin_unlock(&fs_info->tree_mod_seq_lock);
383 tree_mod_log_write_unlock(fs_info);
384
385 return elem->seq;
386}
387
388void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
389 struct seq_list *elem)
390{
391 struct rb_root *tm_root;
392 struct rb_node *node;
393 struct rb_node *next;
394 struct seq_list *cur_elem;
395 struct tree_mod_elem *tm;
396 u64 min_seq = (u64)-1;
397 u64 seq_putting = elem->seq;
398
399 if (!seq_putting)
400 return;
401
402 spin_lock(&fs_info->tree_mod_seq_lock);
403 list_del(&elem->list);
404 elem->seq = 0;
405
406 list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
407 if (cur_elem->seq < min_seq) {
408 if (seq_putting > cur_elem->seq) {
409 /*
410 * blocker with lower sequence number exists, we
411 * cannot remove anything from the log
412 */
413 spin_unlock(&fs_info->tree_mod_seq_lock);
414 return;
415 }
416 min_seq = cur_elem->seq;
417 }
418 }
419 spin_unlock(&fs_info->tree_mod_seq_lock);
420
421 /*
422 * anything that's lower than the lowest existing (read: blocked)
423 * sequence number can be removed from the tree.
424 */
425 tree_mod_log_write_lock(fs_info);
426 tm_root = &fs_info->tree_mod_log;
427 for (node = rb_first(tm_root); node; node = next) {
428 next = rb_next(node);
429 tm = container_of(node, struct tree_mod_elem, node);
430 if (tm->seq > min_seq)
431 continue;
432 rb_erase(node, tm_root);
433 kfree(tm);
434 }
435 tree_mod_log_write_unlock(fs_info);
436}
437
438/*
439 * key order of the log:
440 * node/leaf start address -> sequence
441 *
442 * The 'start address' is the logical address of the *new* root node
443 * for root replace operations, or the logical address of the affected
444 * block for all other operations.
445 *
446 * Note: must be called with write lock (tree_mod_log_write_lock).
447 */
448static noinline int
449__tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
450{
451 struct rb_root *tm_root;
452 struct rb_node **new;
453 struct rb_node *parent = NULL;
454 struct tree_mod_elem *cur;
455
456 BUG_ON(!tm);
457
458 tm->seq = btrfs_inc_tree_mod_seq(fs_info);
459
460 tm_root = &fs_info->tree_mod_log;
461 new = &tm_root->rb_node;
462 while (*new) {
463 cur = container_of(*new, struct tree_mod_elem, node);
464 parent = *new;
465 if (cur->logical < tm->logical)
466 new = &((*new)->rb_left);
467 else if (cur->logical > tm->logical)
468 new = &((*new)->rb_right);
469 else if (cur->seq < tm->seq)
470 new = &((*new)->rb_left);
471 else if (cur->seq > tm->seq)
472 new = &((*new)->rb_right);
473 else
474 return -EEXIST;
475 }
476
477 rb_link_node(&tm->node, parent, new);
478 rb_insert_color(&tm->node, tm_root);
479 return 0;
480}
481
482/*
483 * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
484 * returns zero with the tree_mod_log_lock acquired. The caller must hold
485 * this until all tree mod log insertions are recorded in the rb tree and then
486 * call tree_mod_log_write_unlock() to release.
487 */
488static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
489 struct extent_buffer *eb) {
490 smp_mb();
491 if (list_empty(&(fs_info)->tree_mod_seq_list))
492 return 1;
493 if (eb && btrfs_header_level(eb) == 0)
494 return 1;
495
496 tree_mod_log_write_lock(fs_info);
497 if (list_empty(&(fs_info)->tree_mod_seq_list)) {
498 tree_mod_log_write_unlock(fs_info);
499 return 1;
500 }
501
502 return 0;
503}
504
505/* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
506static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
507 struct extent_buffer *eb)
508{
509 smp_mb();
510 if (list_empty(&(fs_info)->tree_mod_seq_list))
511 return 0;
512 if (eb && btrfs_header_level(eb) == 0)
513 return 0;
514
515 return 1;
516}
517
518static struct tree_mod_elem *
519alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
520 enum mod_log_op op, gfp_t flags)
521{
522 struct tree_mod_elem *tm;
523
524 tm = kzalloc(sizeof(*tm), flags);
525 if (!tm)
526 return NULL;
527
528 tm->logical = eb->start;
529 if (op != MOD_LOG_KEY_ADD) {
530 btrfs_node_key(eb, &tm->key, slot);
531 tm->blockptr = btrfs_node_blockptr(eb, slot);
532 }
533 tm->op = op;
534 tm->slot = slot;
535 tm->generation = btrfs_node_ptr_generation(eb, slot);
536 RB_CLEAR_NODE(&tm->node);
537
538 return tm;
539}
540
541static noinline int
542tree_mod_log_insert_key(struct btrfs_fs_info *fs_info,
543 struct extent_buffer *eb, int slot,
544 enum mod_log_op op, gfp_t flags)
545{
546 struct tree_mod_elem *tm;
547 int ret;
548
549 if (!tree_mod_need_log(fs_info, eb))
550 return 0;
551
552 tm = alloc_tree_mod_elem(eb, slot, op, flags);
553 if (!tm)
554 return -ENOMEM;
555
556 if (tree_mod_dont_log(fs_info, eb)) {
557 kfree(tm);
558 return 0;
559 }
560
561 ret = __tree_mod_log_insert(fs_info, tm);
562 tree_mod_log_write_unlock(fs_info);
563 if (ret)
564 kfree(tm);
565
566 return ret;
567}
568
569static noinline int
570tree_mod_log_insert_move(struct btrfs_fs_info *fs_info,
571 struct extent_buffer *eb, int dst_slot, int src_slot,
572 int nr_items, gfp_t flags)
573{
574 struct tree_mod_elem *tm = NULL;
575 struct tree_mod_elem **tm_list = NULL;
576 int ret = 0;
577 int i;
578 int locked = 0;
579
580 if (!tree_mod_need_log(fs_info, eb))
581 return 0;
582
583 tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), flags);
584 if (!tm_list)
585 return -ENOMEM;
586
587 tm = kzalloc(sizeof(*tm), flags);
588 if (!tm) {
589 ret = -ENOMEM;
590 goto free_tms;
591 }
592
593 tm->logical = eb->start;
594 tm->slot = src_slot;
595 tm->move.dst_slot = dst_slot;
596 tm->move.nr_items = nr_items;
597 tm->op = MOD_LOG_MOVE_KEYS;
598
599 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
600 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
601 MOD_LOG_KEY_REMOVE_WHILE_MOVING, flags);
602 if (!tm_list[i]) {
603 ret = -ENOMEM;
604 goto free_tms;
605 }
606 }
607
608 if (tree_mod_dont_log(fs_info, eb))
609 goto free_tms;
610 locked = 1;
611
612 /*
613 * When we override something during the move, we log these removals.
614 * This can only happen when we move towards the beginning of the
615 * buffer, i.e. dst_slot < src_slot.
616 */
617 for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
618 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
619 if (ret)
620 goto free_tms;
621 }
622
623 ret = __tree_mod_log_insert(fs_info, tm);
624 if (ret)
625 goto free_tms;
626 tree_mod_log_write_unlock(fs_info);
627 kfree(tm_list);
628
629 return 0;
630free_tms:
631 for (i = 0; i < nr_items; i++) {
632 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
633 rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
634 kfree(tm_list[i]);
635 }
636 if (locked)
637 tree_mod_log_write_unlock(fs_info);
638 kfree(tm_list);
639 kfree(tm);
640
641 return ret;
642}
643
644static inline int
645__tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
646 struct tree_mod_elem **tm_list,
647 int nritems)
648{
649 int i, j;
650 int ret;
651
652 for (i = nritems - 1; i >= 0; i--) {
653 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
654 if (ret) {
655 for (j = nritems - 1; j > i; j--)
656 rb_erase(&tm_list[j]->node,
657 &fs_info->tree_mod_log);
658 return ret;
659 }
660 }
661
662 return 0;
663}
664
665static noinline int
666tree_mod_log_insert_root(struct btrfs_fs_info *fs_info,
667 struct extent_buffer *old_root,
668 struct extent_buffer *new_root, gfp_t flags,
669 int log_removal)
670{
671 struct tree_mod_elem *tm = NULL;
672 struct tree_mod_elem **tm_list = NULL;
673 int nritems = 0;
674 int ret = 0;
675 int i;
676
677 if (!tree_mod_need_log(fs_info, NULL))
678 return 0;
679
680 if (log_removal && btrfs_header_level(old_root) > 0) {
681 nritems = btrfs_header_nritems(old_root);
682 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
683 flags);
684 if (!tm_list) {
685 ret = -ENOMEM;
686 goto free_tms;
687 }
688 for (i = 0; i < nritems; i++) {
689 tm_list[i] = alloc_tree_mod_elem(old_root, i,
690 MOD_LOG_KEY_REMOVE_WHILE_FREEING, flags);
691 if (!tm_list[i]) {
692 ret = -ENOMEM;
693 goto free_tms;
694 }
695 }
696 }
697
698 tm = kzalloc(sizeof(*tm), flags);
699 if (!tm) {
700 ret = -ENOMEM;
701 goto free_tms;
702 }
703
704 tm->logical = new_root->start;
705 tm->old_root.logical = old_root->start;
706 tm->old_root.level = btrfs_header_level(old_root);
707 tm->generation = btrfs_header_generation(old_root);
708 tm->op = MOD_LOG_ROOT_REPLACE;
709
710 if (tree_mod_dont_log(fs_info, NULL))
711 goto free_tms;
712
713 if (tm_list)
714 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
715 if (!ret)
716 ret = __tree_mod_log_insert(fs_info, tm);
717
718 tree_mod_log_write_unlock(fs_info);
719 if (ret)
720 goto free_tms;
721 kfree(tm_list);
722
723 return ret;
724
725free_tms:
726 if (tm_list) {
727 for (i = 0; i < nritems; i++)
728 kfree(tm_list[i]);
729 kfree(tm_list);
730 }
731 kfree(tm);
732
733 return ret;
734}
735
736static struct tree_mod_elem *
737__tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
738 int smallest)
739{
740 struct rb_root *tm_root;
741 struct rb_node *node;
742 struct tree_mod_elem *cur = NULL;
743 struct tree_mod_elem *found = NULL;
744
745 tree_mod_log_read_lock(fs_info);
746 tm_root = &fs_info->tree_mod_log;
747 node = tm_root->rb_node;
748 while (node) {
749 cur = container_of(node, struct tree_mod_elem, node);
750 if (cur->logical < start) {
751 node = node->rb_left;
752 } else if (cur->logical > start) {
753 node = node->rb_right;
754 } else if (cur->seq < min_seq) {
755 node = node->rb_left;
756 } else if (!smallest) {
757 /* we want the node with the highest seq */
758 if (found)
759 BUG_ON(found->seq > cur->seq);
760 found = cur;
761 node = node->rb_left;
762 } else if (cur->seq > min_seq) {
763 /* we want the node with the smallest seq */
764 if (found)
765 BUG_ON(found->seq < cur->seq);
766 found = cur;
767 node = node->rb_right;
768 } else {
769 found = cur;
770 break;
771 }
772 }
773 tree_mod_log_read_unlock(fs_info);
774
775 return found;
776}
777
778/*
779 * this returns the element from the log with the smallest time sequence
780 * value that's in the log (the oldest log item). any element with a time
781 * sequence lower than min_seq will be ignored.
782 */
783static struct tree_mod_elem *
784tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
785 u64 min_seq)
786{
787 return __tree_mod_log_search(fs_info, start, min_seq, 1);
788}
789
790/*
791 * this returns the element from the log with the largest time sequence
792 * value that's in the log (the most recent log item). any element with
793 * a time sequence lower than min_seq will be ignored.
794 */
795static struct tree_mod_elem *
796tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
797{
798 return __tree_mod_log_search(fs_info, start, min_seq, 0);
799}
800
801static noinline int
802tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
803 struct extent_buffer *src, unsigned long dst_offset,
804 unsigned long src_offset, int nr_items)
805{
806 int ret = 0;
807 struct tree_mod_elem **tm_list = NULL;
808 struct tree_mod_elem **tm_list_add, **tm_list_rem;
809 int i;
810 int locked = 0;
811
812 if (!tree_mod_need_log(fs_info, NULL))
813 return 0;
814
815 if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
816 return 0;
817
818 tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
819 GFP_NOFS);
820 if (!tm_list)
821 return -ENOMEM;
822
823 tm_list_add = tm_list;
824 tm_list_rem = tm_list + nr_items;
825 for (i = 0; i < nr_items; i++) {
826 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
827 MOD_LOG_KEY_REMOVE, GFP_NOFS);
828 if (!tm_list_rem[i]) {
829 ret = -ENOMEM;
830 goto free_tms;
831 }
832
833 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
834 MOD_LOG_KEY_ADD, GFP_NOFS);
835 if (!tm_list_add[i]) {
836 ret = -ENOMEM;
837 goto free_tms;
838 }
839 }
840
841 if (tree_mod_dont_log(fs_info, NULL))
842 goto free_tms;
843 locked = 1;
844
845 for (i = 0; i < nr_items; i++) {
846 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
847 if (ret)
848 goto free_tms;
849 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
850 if (ret)
851 goto free_tms;
852 }
853
854 tree_mod_log_write_unlock(fs_info);
855 kfree(tm_list);
856
857 return 0;
858
859free_tms:
860 for (i = 0; i < nr_items * 2; i++) {
861 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
862 rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
863 kfree(tm_list[i]);
864 }
865 if (locked)
866 tree_mod_log_write_unlock(fs_info);
867 kfree(tm_list);
868
869 return ret;
870}
871
872static inline void
873tree_mod_log_eb_move(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
874 int dst_offset, int src_offset, int nr_items)
875{
876 int ret;
877 ret = tree_mod_log_insert_move(fs_info, dst, dst_offset, src_offset,
878 nr_items, GFP_NOFS);
879 BUG_ON(ret < 0);
880}
881
882static noinline void
883tree_mod_log_set_node_key(struct btrfs_fs_info *fs_info,
884 struct extent_buffer *eb, int slot, int atomic)
885{
886 int ret;
887
888 ret = tree_mod_log_insert_key(fs_info, eb, slot,
889 MOD_LOG_KEY_REPLACE,
890 atomic ? GFP_ATOMIC : GFP_NOFS);
891 BUG_ON(ret < 0);
892}
893
894static noinline int
895tree_mod_log_free_eb(struct btrfs_fs_info *fs_info, struct extent_buffer *eb)
896{
897 struct tree_mod_elem **tm_list = NULL;
898 int nritems = 0;
899 int i;
900 int ret = 0;
901
902 if (btrfs_header_level(eb) == 0)
903 return 0;
904
905 if (!tree_mod_need_log(fs_info, NULL))
906 return 0;
907
908 nritems = btrfs_header_nritems(eb);
909 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
910 if (!tm_list)
911 return -ENOMEM;
912
913 for (i = 0; i < nritems; i++) {
914 tm_list[i] = alloc_tree_mod_elem(eb, i,
915 MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
916 if (!tm_list[i]) {
917 ret = -ENOMEM;
918 goto free_tms;
919 }
920 }
921
922 if (tree_mod_dont_log(fs_info, eb))
923 goto free_tms;
924
925 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
926 tree_mod_log_write_unlock(fs_info);
927 if (ret)
928 goto free_tms;
929 kfree(tm_list);
930
931 return 0;
932
933free_tms:
934 for (i = 0; i < nritems; i++)
935 kfree(tm_list[i]);
936 kfree(tm_list);
937
938 return ret;
939}
940
941static noinline void
942tree_mod_log_set_root_pointer(struct btrfs_root *root,
943 struct extent_buffer *new_root_node,
944 int log_removal)
945{
946 int ret;
947 ret = tree_mod_log_insert_root(root->fs_info, root->node,
948 new_root_node, GFP_NOFS, log_removal);
949 BUG_ON(ret < 0);
950}
951
952/*
953 * check if the tree block can be shared by multiple trees
954 */
955int btrfs_block_can_be_shared(struct btrfs_root *root,
956 struct extent_buffer *buf)
957{
958 /*
959 * Tree blocks not in reference counted trees and tree roots
960 * are never shared. If a block was allocated after the last
961 * snapshot and the block was not allocated by tree relocation,
962 * we know the block is not shared.
963 */
964 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
965 buf != root->node && buf != root->commit_root &&
966 (btrfs_header_generation(buf) <=
967 btrfs_root_last_snapshot(&root->root_item) ||
968 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
969 return 1;
970#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
971 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
972 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
973 return 1;
974#endif
975 return 0;
976}
977
978static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
979 struct btrfs_root *root,
980 struct extent_buffer *buf,
981 struct extent_buffer *cow,
982 int *last_ref)
983{
984 struct btrfs_fs_info *fs_info = root->fs_info;
985 u64 refs;
986 u64 owner;
987 u64 flags;
988 u64 new_flags = 0;
989 int ret;
990
991 /*
992 * Backrefs update rules:
993 *
994 * Always use full backrefs for extent pointers in tree block
995 * allocated by tree relocation.
996 *
997 * If a shared tree block is no longer referenced by its owner
998 * tree (btrfs_header_owner(buf) == root->root_key.objectid),
999 * use full backrefs for extent pointers in tree block.
1000 *
1001 * If a tree block is been relocating
1002 * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
1003 * use full backrefs for extent pointers in tree block.
1004 * The reason for this is some operations (such as drop tree)
1005 * are only allowed for blocks use full backrefs.
1006 */
1007
1008 if (btrfs_block_can_be_shared(root, buf)) {
1009 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
1010 btrfs_header_level(buf), 1,
1011 &refs, &flags);
1012 if (ret)
1013 return ret;
1014 if (refs == 0) {
1015 ret = -EROFS;
1016 btrfs_handle_fs_error(fs_info, ret, NULL);
1017 return ret;
1018 }
1019 } else {
1020 refs = 1;
1021 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1022 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1023 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
1024 else
1025 flags = 0;
1026 }
1027
1028 owner = btrfs_header_owner(buf);
1029 BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
1030 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
1031
1032 if (refs > 1) {
1033 if ((owner == root->root_key.objectid ||
1034 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
1035 !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
1036 ret = btrfs_inc_ref(trans, root, buf, 1);
1037 BUG_ON(ret); /* -ENOMEM */
1038
1039 if (root->root_key.objectid ==
1040 BTRFS_TREE_RELOC_OBJECTID) {
1041 ret = btrfs_dec_ref(trans, root, buf, 0);
1042 BUG_ON(ret); /* -ENOMEM */
1043 ret = btrfs_inc_ref(trans, root, cow, 1);
1044 BUG_ON(ret); /* -ENOMEM */
1045 }
1046 new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
1047 } else {
1048
1049 if (root->root_key.objectid ==
1050 BTRFS_TREE_RELOC_OBJECTID)
1051 ret = btrfs_inc_ref(trans, root, cow, 1);
1052 else
1053 ret = btrfs_inc_ref(trans, root, cow, 0);
1054 BUG_ON(ret); /* -ENOMEM */
1055 }
1056 if (new_flags != 0) {
1057 int level = btrfs_header_level(buf);
1058
1059 ret = btrfs_set_disk_extent_flags(trans, fs_info,
1060 buf->start,
1061 buf->len,
1062 new_flags, level, 0);
1063 if (ret)
1064 return ret;
1065 }
1066 } else {
1067 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
1068 if (root->root_key.objectid ==
1069 BTRFS_TREE_RELOC_OBJECTID)
1070 ret = btrfs_inc_ref(trans, root, cow, 1);
1071 else
1072 ret = btrfs_inc_ref(trans, root, cow, 0);
1073 BUG_ON(ret); /* -ENOMEM */
1074 ret = btrfs_dec_ref(trans, root, buf, 1);
1075 BUG_ON(ret); /* -ENOMEM */
1076 }
1077 clean_tree_block(trans, fs_info, buf);
1078 *last_ref = 1;
1079 }
1080 return 0;
1081}
1082
1083/*
1084 * does the dirty work in cow of a single block. The parent block (if
1085 * supplied) is updated to point to the new cow copy. The new buffer is marked
1086 * dirty and returned locked. If you modify the block it needs to be marked
1087 * dirty again.
1088 *
1089 * search_start -- an allocation hint for the new block
1090 *
1091 * empty_size -- a hint that you plan on doing more cow. This is the size in
1092 * bytes the allocator should try to find free next to the block it returns.
1093 * This is just a hint and may be ignored by the allocator.
1094 */
1095static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1096 struct btrfs_root *root,
1097 struct extent_buffer *buf,
1098 struct extent_buffer *parent, int parent_slot,
1099 struct extent_buffer **cow_ret,
1100 u64 search_start, u64 empty_size)
1101{
1102 struct btrfs_fs_info *fs_info = root->fs_info;
1103 struct btrfs_disk_key disk_key;
1104 struct extent_buffer *cow;
1105 int level, ret;
1106 int last_ref = 0;
1107 int unlock_orig = 0;
1108 u64 parent_start = 0;
1109
1110 if (*cow_ret == buf)
1111 unlock_orig = 1;
1112
1113 btrfs_assert_tree_locked(buf);
1114
1115 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1116 trans->transid != fs_info->running_transaction->transid);
1117 WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1118 trans->transid != root->last_trans);
1119
1120 level = btrfs_header_level(buf);
1121
1122 if (level == 0)
1123 btrfs_item_key(buf, &disk_key, 0);
1124 else
1125 btrfs_node_key(buf, &disk_key, 0);
1126
1127 if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1128 parent_start = parent->start;
1129
1130 cow = btrfs_alloc_tree_block(trans, root, parent_start,
1131 root->root_key.objectid, &disk_key, level,
1132 search_start, empty_size);
1133 if (IS_ERR(cow))
1134 return PTR_ERR(cow);
1135
1136 /* cow is set to blocking by btrfs_init_new_buffer */
1137
1138 copy_extent_buffer_full(cow, buf);
1139 btrfs_set_header_bytenr(cow, cow->start);
1140 btrfs_set_header_generation(cow, trans->transid);
1141 btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1142 btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1143 BTRFS_HEADER_FLAG_RELOC);
1144 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1145 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1146 else
1147 btrfs_set_header_owner(cow, root->root_key.objectid);
1148
1149 write_extent_buffer_fsid(cow, fs_info->fsid);
1150
1151 ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1152 if (ret) {
1153 btrfs_abort_transaction(trans, ret);
1154 return ret;
1155 }
1156
1157 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1158 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1159 if (ret) {
1160 btrfs_abort_transaction(trans, ret);
1161 return ret;
1162 }
1163 }
1164
1165 if (buf == root->node) {
1166 WARN_ON(parent && parent != buf);
1167 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1168 btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1169 parent_start = buf->start;
1170
1171 extent_buffer_get(cow);
1172 tree_mod_log_set_root_pointer(root, cow, 1);
1173 rcu_assign_pointer(root->node, cow);
1174
1175 btrfs_free_tree_block(trans, root, buf, parent_start,
1176 last_ref);
1177 free_extent_buffer(buf);
1178 add_root_to_dirty_list(root);
1179 } else {
1180 WARN_ON(trans->transid != btrfs_header_generation(parent));
1181 tree_mod_log_insert_key(fs_info, parent, parent_slot,
1182 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1183 btrfs_set_node_blockptr(parent, parent_slot,
1184 cow->start);
1185 btrfs_set_node_ptr_generation(parent, parent_slot,
1186 trans->transid);
1187 btrfs_mark_buffer_dirty(parent);
1188 if (last_ref) {
1189 ret = tree_mod_log_free_eb(fs_info, buf);
1190 if (ret) {
1191 btrfs_abort_transaction(trans, ret);
1192 return ret;
1193 }
1194 }
1195 btrfs_free_tree_block(trans, root, buf, parent_start,
1196 last_ref);
1197 }
1198 if (unlock_orig)
1199 btrfs_tree_unlock(buf);
1200 free_extent_buffer_stale(buf);
1201 btrfs_mark_buffer_dirty(cow);
1202 *cow_ret = cow;
1203 return 0;
1204}
1205
1206/*
1207 * returns the logical address of the oldest predecessor of the given root.
1208 * entries older than time_seq are ignored.
1209 */
1210static struct tree_mod_elem *
1211__tree_mod_log_oldest_root(struct btrfs_fs_info *fs_info,
1212 struct extent_buffer *eb_root, u64 time_seq)
1213{
1214 struct tree_mod_elem *tm;
1215 struct tree_mod_elem *found = NULL;
1216 u64 root_logical = eb_root->start;
1217 int looped = 0;
1218
1219 if (!time_seq)
1220 return NULL;
1221
1222 /*
1223 * the very last operation that's logged for a root is the
1224 * replacement operation (if it is replaced at all). this has
1225 * the logical address of the *new* root, making it the very
1226 * first operation that's logged for this root.
1227 */
1228 while (1) {
1229 tm = tree_mod_log_search_oldest(fs_info, root_logical,
1230 time_seq);
1231 if (!looped && !tm)
1232 return NULL;
1233 /*
1234 * if there are no tree operation for the oldest root, we simply
1235 * return it. this should only happen if that (old) root is at
1236 * level 0.
1237 */
1238 if (!tm)
1239 break;
1240
1241 /*
1242 * if there's an operation that's not a root replacement, we
1243 * found the oldest version of our root. normally, we'll find a
1244 * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1245 */
1246 if (tm->op != MOD_LOG_ROOT_REPLACE)
1247 break;
1248
1249 found = tm;
1250 root_logical = tm->old_root.logical;
1251 looped = 1;
1252 }
1253
1254 /* if there's no old root to return, return what we found instead */
1255 if (!found)
1256 found = tm;
1257
1258 return found;
1259}
1260
1261/*
1262 * tm is a pointer to the first operation to rewind within eb. then, all
1263 * previous operations will be rewound (until we reach something older than
1264 * time_seq).
1265 */
1266static void
1267__tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1268 u64 time_seq, struct tree_mod_elem *first_tm)
1269{
1270 u32 n;
1271 struct rb_node *next;
1272 struct tree_mod_elem *tm = first_tm;
1273 unsigned long o_dst;
1274 unsigned long o_src;
1275 unsigned long p_size = sizeof(struct btrfs_key_ptr);
1276
1277 n = btrfs_header_nritems(eb);
1278 tree_mod_log_read_lock(fs_info);
1279 while (tm && tm->seq >= time_seq) {
1280 /*
1281 * all the operations are recorded with the operator used for
1282 * the modification. as we're going backwards, we do the
1283 * opposite of each operation here.
1284 */
1285 switch (tm->op) {
1286 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1287 BUG_ON(tm->slot < n);
1288 /* Fallthrough */
1289 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1290 case MOD_LOG_KEY_REMOVE:
1291 btrfs_set_node_key(eb, &tm->key, tm->slot);
1292 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1293 btrfs_set_node_ptr_generation(eb, tm->slot,
1294 tm->generation);
1295 n++;
1296 break;
1297 case MOD_LOG_KEY_REPLACE:
1298 BUG_ON(tm->slot >= n);
1299 btrfs_set_node_key(eb, &tm->key, tm->slot);
1300 btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1301 btrfs_set_node_ptr_generation(eb, tm->slot,
1302 tm->generation);
1303 break;
1304 case MOD_LOG_KEY_ADD:
1305 /* if a move operation is needed it's in the log */
1306 n--;
1307 break;
1308 case MOD_LOG_MOVE_KEYS:
1309 o_dst = btrfs_node_key_ptr_offset(tm->slot);
1310 o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1311 memmove_extent_buffer(eb, o_dst, o_src,
1312 tm->move.nr_items * p_size);
1313 break;
1314 case MOD_LOG_ROOT_REPLACE:
1315 /*
1316 * this operation is special. for roots, this must be
1317 * handled explicitly before rewinding.
1318 * for non-roots, this operation may exist if the node
1319 * was a root: root A -> child B; then A gets empty and
1320 * B is promoted to the new root. in the mod log, we'll
1321 * have a root-replace operation for B, a tree block
1322 * that is no root. we simply ignore that operation.
1323 */
1324 break;
1325 }
1326 next = rb_next(&tm->node);
1327 if (!next)
1328 break;
1329 tm = container_of(next, struct tree_mod_elem, node);
1330 if (tm->logical != first_tm->logical)
1331 break;
1332 }
1333 tree_mod_log_read_unlock(fs_info);
1334 btrfs_set_header_nritems(eb, n);
1335}
1336
1337/*
1338 * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1339 * is returned. If rewind operations happen, a fresh buffer is returned. The
1340 * returned buffer is always read-locked. If the returned buffer is not the
1341 * input buffer, the lock on the input buffer is released and the input buffer
1342 * is freed (its refcount is decremented).
1343 */
1344static struct extent_buffer *
1345tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1346 struct extent_buffer *eb, u64 time_seq)
1347{
1348 struct extent_buffer *eb_rewin;
1349 struct tree_mod_elem *tm;
1350
1351 if (!time_seq)
1352 return eb;
1353
1354 if (btrfs_header_level(eb) == 0)
1355 return eb;
1356
1357 tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1358 if (!tm)
1359 return eb;
1360
1361 btrfs_set_path_blocking(path);
1362 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1363
1364 if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1365 BUG_ON(tm->slot != 0);
1366 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1367 if (!eb_rewin) {
1368 btrfs_tree_read_unlock_blocking(eb);
1369 free_extent_buffer(eb);
1370 return NULL;
1371 }
1372 btrfs_set_header_bytenr(eb_rewin, eb->start);
1373 btrfs_set_header_backref_rev(eb_rewin,
1374 btrfs_header_backref_rev(eb));
1375 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1376 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1377 } else {
1378 eb_rewin = btrfs_clone_extent_buffer(eb);
1379 if (!eb_rewin) {
1380 btrfs_tree_read_unlock_blocking(eb);
1381 free_extent_buffer(eb);
1382 return NULL;
1383 }
1384 }
1385
1386 btrfs_clear_path_blocking(path, NULL, BTRFS_READ_LOCK);
1387 btrfs_tree_read_unlock_blocking(eb);
1388 free_extent_buffer(eb);
1389
1390 extent_buffer_get(eb_rewin);
1391 btrfs_tree_read_lock(eb_rewin);
1392 __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1393 WARN_ON(btrfs_header_nritems(eb_rewin) >
1394 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1395
1396 return eb_rewin;
1397}
1398
1399/*
1400 * get_old_root() rewinds the state of @root's root node to the given @time_seq
1401 * value. If there are no changes, the current root->root_node is returned. If
1402 * anything changed in between, there's a fresh buffer allocated on which the
1403 * rewind operations are done. In any case, the returned buffer is read locked.
1404 * Returns NULL on error (with no locks held).
1405 */
1406static inline struct extent_buffer *
1407get_old_root(struct btrfs_root *root, u64 time_seq)
1408{
1409 struct btrfs_fs_info *fs_info = root->fs_info;
1410 struct tree_mod_elem *tm;
1411 struct extent_buffer *eb = NULL;
1412 struct extent_buffer *eb_root;
1413 struct extent_buffer *old;
1414 struct tree_mod_root *old_root = NULL;
1415 u64 old_generation = 0;
1416 u64 logical;
1417
1418 eb_root = btrfs_read_lock_root_node(root);
1419 tm = __tree_mod_log_oldest_root(fs_info, eb_root, time_seq);
1420 if (!tm)
1421 return eb_root;
1422
1423 if (tm->op == MOD_LOG_ROOT_REPLACE) {
1424 old_root = &tm->old_root;
1425 old_generation = tm->generation;
1426 logical = old_root->logical;
1427 } else {
1428 logical = eb_root->start;
1429 }
1430
1431 tm = tree_mod_log_search(fs_info, logical, time_seq);
1432 if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1433 btrfs_tree_read_unlock(eb_root);
1434 free_extent_buffer(eb_root);
1435 old = read_tree_block(fs_info, logical, 0);
1436 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1437 if (!IS_ERR(old))
1438 free_extent_buffer(old);
1439 btrfs_warn(fs_info,
1440 "failed to read tree block %llu from get_old_root",
1441 logical);
1442 } else {
1443 eb = btrfs_clone_extent_buffer(old);
1444 free_extent_buffer(old);
1445 }
1446 } else if (old_root) {
1447 btrfs_tree_read_unlock(eb_root);
1448 free_extent_buffer(eb_root);
1449 eb = alloc_dummy_extent_buffer(fs_info, logical);
1450 } else {
1451 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1452 eb = btrfs_clone_extent_buffer(eb_root);
1453 btrfs_tree_read_unlock_blocking(eb_root);
1454 free_extent_buffer(eb_root);
1455 }
1456
1457 if (!eb)
1458 return NULL;
1459 extent_buffer_get(eb);
1460 btrfs_tree_read_lock(eb);
1461 if (old_root) {
1462 btrfs_set_header_bytenr(eb, eb->start);
1463 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1464 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1465 btrfs_set_header_level(eb, old_root->level);
1466 btrfs_set_header_generation(eb, old_generation);
1467 }
1468 if (tm)
1469 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1470 else
1471 WARN_ON(btrfs_header_level(eb) != 0);
1472 WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1473
1474 return eb;
1475}
1476
1477int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1478{
1479 struct tree_mod_elem *tm;
1480 int level;
1481 struct extent_buffer *eb_root = btrfs_root_node(root);
1482
1483 tm = __tree_mod_log_oldest_root(root->fs_info, eb_root, time_seq);
1484 if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1485 level = tm->old_root.level;
1486 } else {
1487 level = btrfs_header_level(eb_root);
1488 }
1489 free_extent_buffer(eb_root);
1490
1491 return level;
1492}
1493
1494static inline int should_cow_block(struct btrfs_trans_handle *trans,
1495 struct btrfs_root *root,
1496 struct extent_buffer *buf)
1497{
1498 if (btrfs_is_testing(root->fs_info))
1499 return 0;
1500
1501 /* ensure we can see the force_cow */
1502 smp_rmb();
1503
1504 /*
1505 * We do not need to cow a block if
1506 * 1) this block is not created or changed in this transaction;
1507 * 2) this block does not belong to TREE_RELOC tree;
1508 * 3) the root is not forced COW.
1509 *
1510 * What is forced COW:
1511 * when we create snapshot during committing the transaction,
1512 * after we've finished coping src root, we must COW the shared
1513 * block to ensure the metadata consistency.
1514 */
1515 if (btrfs_header_generation(buf) == trans->transid &&
1516 !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1517 !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1518 btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1519 !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1520 return 0;
1521 return 1;
1522}
1523
1524/*
1525 * cows a single block, see __btrfs_cow_block for the real work.
1526 * This version of it has extra checks so that a block isn't COWed more than
1527 * once per transaction, as long as it hasn't been written yet
1528 */
1529noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1530 struct btrfs_root *root, struct extent_buffer *buf,
1531 struct extent_buffer *parent, int parent_slot,
1532 struct extent_buffer **cow_ret)
1533{
1534 struct btrfs_fs_info *fs_info = root->fs_info;
1535 u64 search_start;
1536 int ret;
1537
1538 if (trans->transaction != fs_info->running_transaction)
1539 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1540 trans->transid,
1541 fs_info->running_transaction->transid);
1542
1543 if (trans->transid != fs_info->generation)
1544 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1545 trans->transid, fs_info->generation);
1546
1547 if (!should_cow_block(trans, root, buf)) {
1548 trans->dirty = true;
1549 *cow_ret = buf;
1550 return 0;
1551 }
1552
1553 search_start = buf->start & ~((u64)SZ_1G - 1);
1554
1555 if (parent)
1556 btrfs_set_lock_blocking(parent);
1557 btrfs_set_lock_blocking(buf);
1558
1559 ret = __btrfs_cow_block(trans, root, buf, parent,
1560 parent_slot, cow_ret, search_start, 0);
1561
1562 trace_btrfs_cow_block(root, buf, *cow_ret);
1563
1564 return ret;
1565}
1566
1567/*
1568 * helper function for defrag to decide if two blocks pointed to by a
1569 * node are actually close by
1570 */
1571static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1572{
1573 if (blocknr < other && other - (blocknr + blocksize) < 32768)
1574 return 1;
1575 if (blocknr > other && blocknr - (other + blocksize) < 32768)
1576 return 1;
1577 return 0;
1578}
1579
1580/*
1581 * compare two keys in a memcmp fashion
1582 */
1583static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
1584{
1585 struct btrfs_key k1;
1586
1587 btrfs_disk_key_to_cpu(&k1, disk);
1588
1589 return btrfs_comp_cpu_keys(&k1, k2);
1590}
1591
1592/*
1593 * same as comp_keys only with two btrfs_key's
1594 */
1595int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
1596{
1597 if (k1->objectid > k2->objectid)
1598 return 1;
1599 if (k1->objectid < k2->objectid)
1600 return -1;
1601 if (k1->type > k2->type)
1602 return 1;
1603 if (k1->type < k2->type)
1604 return -1;
1605 if (k1->offset > k2->offset)
1606 return 1;
1607 if (k1->offset < k2->offset)
1608 return -1;
1609 return 0;
1610}
1611
1612/*
1613 * this is used by the defrag code to go through all the
1614 * leaves pointed to by a node and reallocate them so that
1615 * disk order is close to key order
1616 */
1617int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1618 struct btrfs_root *root, struct extent_buffer *parent,
1619 int start_slot, u64 *last_ret,
1620 struct btrfs_key *progress)
1621{
1622 struct btrfs_fs_info *fs_info = root->fs_info;
1623 struct extent_buffer *cur;
1624 u64 blocknr;
1625 u64 gen;
1626 u64 search_start = *last_ret;
1627 u64 last_block = 0;
1628 u64 other;
1629 u32 parent_nritems;
1630 int end_slot;
1631 int i;
1632 int err = 0;
1633 int parent_level;
1634 int uptodate;
1635 u32 blocksize;
1636 int progress_passed = 0;
1637 struct btrfs_disk_key disk_key;
1638
1639 parent_level = btrfs_header_level(parent);
1640
1641 WARN_ON(trans->transaction != fs_info->running_transaction);
1642 WARN_ON(trans->transid != fs_info->generation);
1643
1644 parent_nritems = btrfs_header_nritems(parent);
1645 blocksize = fs_info->nodesize;
1646 end_slot = parent_nritems - 1;
1647
1648 if (parent_nritems <= 1)
1649 return 0;
1650
1651 btrfs_set_lock_blocking(parent);
1652
1653 for (i = start_slot; i <= end_slot; i++) {
1654 int close = 1;
1655
1656 btrfs_node_key(parent, &disk_key, i);
1657 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1658 continue;
1659
1660 progress_passed = 1;
1661 blocknr = btrfs_node_blockptr(parent, i);
1662 gen = btrfs_node_ptr_generation(parent, i);
1663 if (last_block == 0)
1664 last_block = blocknr;
1665
1666 if (i > 0) {
1667 other = btrfs_node_blockptr(parent, i - 1);
1668 close = close_blocks(blocknr, other, blocksize);
1669 }
1670 if (!close && i < end_slot) {
1671 other = btrfs_node_blockptr(parent, i + 1);
1672 close = close_blocks(blocknr, other, blocksize);
1673 }
1674 if (close) {
1675 last_block = blocknr;
1676 continue;
1677 }
1678
1679 cur = find_extent_buffer(fs_info, blocknr);
1680 if (cur)
1681 uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1682 else
1683 uptodate = 0;
1684 if (!cur || !uptodate) {
1685 if (!cur) {
1686 cur = read_tree_block(fs_info, blocknr, gen);
1687 if (IS_ERR(cur)) {
1688 return PTR_ERR(cur);
1689 } else if (!extent_buffer_uptodate(cur)) {
1690 free_extent_buffer(cur);
1691 return -EIO;
1692 }
1693 } else if (!uptodate) {
1694 err = btrfs_read_buffer(cur, gen);
1695 if (err) {
1696 free_extent_buffer(cur);
1697 return err;
1698 }
1699 }
1700 }
1701 if (search_start == 0)
1702 search_start = last_block;
1703
1704 btrfs_tree_lock(cur);
1705 btrfs_set_lock_blocking(cur);
1706 err = __btrfs_cow_block(trans, root, cur, parent, i,
1707 &cur, search_start,
1708 min(16 * blocksize,
1709 (end_slot - i) * blocksize));
1710 if (err) {
1711 btrfs_tree_unlock(cur);
1712 free_extent_buffer(cur);
1713 break;
1714 }
1715 search_start = cur->start;
1716 last_block = cur->start;
1717 *last_ret = search_start;
1718 btrfs_tree_unlock(cur);
1719 free_extent_buffer(cur);
1720 }
1721 return err;
1722}
1723
1724/*
1725 * search for key in the extent_buffer. The items start at offset p,
1726 * and they are item_size apart. There are 'max' items in p.
1727 *
1728 * the slot in the array is returned via slot, and it points to
1729 * the place where you would insert key if it is not found in
1730 * the array.
1731 *
1732 * slot may point to max if the key is bigger than all of the keys
1733 */
1734static noinline int generic_bin_search(struct extent_buffer *eb,
1735 unsigned long p,
1736 int item_size, struct btrfs_key *key,
1737 int max, int *slot)
1738{
1739 int low = 0;
1740 int high = max;
1741 int mid;
1742 int ret;
1743 struct btrfs_disk_key *tmp = NULL;
1744 struct btrfs_disk_key unaligned;
1745 unsigned long offset;
1746 char *kaddr = NULL;
1747 unsigned long map_start = 0;
1748 unsigned long map_len = 0;
1749 int err;
1750
1751 if (low > high) {
1752 btrfs_err(eb->fs_info,
1753 "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1754 __func__, low, high, eb->start,
1755 btrfs_header_owner(eb), btrfs_header_level(eb));
1756 return -EINVAL;
1757 }
1758
1759 while (low < high) {
1760 mid = (low + high) / 2;
1761 offset = p + mid * item_size;
1762
1763 if (!kaddr || offset < map_start ||
1764 (offset + sizeof(struct btrfs_disk_key)) >
1765 map_start + map_len) {
1766
1767 err = map_private_extent_buffer(eb, offset,
1768 sizeof(struct btrfs_disk_key),
1769 &kaddr, &map_start, &map_len);
1770
1771 if (!err) {
1772 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1773 map_start);
1774 } else if (err == 1) {
1775 read_extent_buffer(eb, &unaligned,
1776 offset, sizeof(unaligned));
1777 tmp = &unaligned;
1778 } else {
1779 return err;
1780 }
1781
1782 } else {
1783 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1784 map_start);
1785 }
1786 ret = comp_keys(tmp, key);
1787
1788 if (ret < 0)
1789 low = mid + 1;
1790 else if (ret > 0)
1791 high = mid;
1792 else {
1793 *slot = mid;
1794 return 0;
1795 }
1796 }
1797 *slot = low;
1798 return 1;
1799}
1800
1801/*
1802 * simple bin_search frontend that does the right thing for
1803 * leaves vs nodes
1804 */
1805static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1806 int level, int *slot)
1807{
1808 if (level == 0)
1809 return generic_bin_search(eb,
1810 offsetof(struct btrfs_leaf, items),
1811 sizeof(struct btrfs_item),
1812 key, btrfs_header_nritems(eb),
1813 slot);
1814 else
1815 return generic_bin_search(eb,
1816 offsetof(struct btrfs_node, ptrs),
1817 sizeof(struct btrfs_key_ptr),
1818 key, btrfs_header_nritems(eb),
1819 slot);
1820}
1821
1822int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1823 int level, int *slot)
1824{
1825 return bin_search(eb, key, level, slot);
1826}
1827
1828static void root_add_used(struct btrfs_root *root, u32 size)
1829{
1830 spin_lock(&root->accounting_lock);
1831 btrfs_set_root_used(&root->root_item,
1832 btrfs_root_used(&root->root_item) + size);
1833 spin_unlock(&root->accounting_lock);
1834}
1835
1836static void root_sub_used(struct btrfs_root *root, u32 size)
1837{
1838 spin_lock(&root->accounting_lock);
1839 btrfs_set_root_used(&root->root_item,
1840 btrfs_root_used(&root->root_item) - size);
1841 spin_unlock(&root->accounting_lock);
1842}
1843
1844/* given a node and slot number, this reads the blocks it points to. The
1845 * extent buffer is returned with a reference taken (but unlocked).
1846 */
1847static noinline struct extent_buffer *
1848read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1849 int slot)
1850{
1851 int level = btrfs_header_level(parent);
1852 struct extent_buffer *eb;
1853
1854 if (slot < 0 || slot >= btrfs_header_nritems(parent))
1855 return ERR_PTR(-ENOENT);
1856
1857 BUG_ON(level == 0);
1858
1859 eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1860 btrfs_node_ptr_generation(parent, slot));
1861 if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1862 free_extent_buffer(eb);
1863 eb = ERR_PTR(-EIO);
1864 }
1865
1866 return eb;
1867}
1868
1869/*
1870 * node level balancing, used to make sure nodes are in proper order for
1871 * item deletion. We balance from the top down, so we have to make sure
1872 * that a deletion won't leave an node completely empty later on.
1873 */
1874static noinline int balance_level(struct btrfs_trans_handle *trans,
1875 struct btrfs_root *root,
1876 struct btrfs_path *path, int level)
1877{
1878 struct btrfs_fs_info *fs_info = root->fs_info;
1879 struct extent_buffer *right = NULL;
1880 struct extent_buffer *mid;
1881 struct extent_buffer *left = NULL;
1882 struct extent_buffer *parent = NULL;
1883 int ret = 0;
1884 int wret;
1885 int pslot;
1886 int orig_slot = path->slots[level];
1887 u64 orig_ptr;
1888
1889 if (level == 0)
1890 return 0;
1891
1892 mid = path->nodes[level];
1893
1894 WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1895 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1896 WARN_ON(btrfs_header_generation(mid) != trans->transid);
1897
1898 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1899
1900 if (level < BTRFS_MAX_LEVEL - 1) {
1901 parent = path->nodes[level + 1];
1902 pslot = path->slots[level + 1];
1903 }
1904
1905 /*
1906 * deal with the case where there is only one pointer in the root
1907 * by promoting the node below to a root
1908 */
1909 if (!parent) {
1910 struct extent_buffer *child;
1911
1912 if (btrfs_header_nritems(mid) != 1)
1913 return 0;
1914
1915 /* promote the child to a root */
1916 child = read_node_slot(fs_info, mid, 0);
1917 if (IS_ERR(child)) {
1918 ret = PTR_ERR(child);
1919 btrfs_handle_fs_error(fs_info, ret, NULL);
1920 goto enospc;
1921 }
1922
1923 btrfs_tree_lock(child);
1924 btrfs_set_lock_blocking(child);
1925 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1926 if (ret) {
1927 btrfs_tree_unlock(child);
1928 free_extent_buffer(child);
1929 goto enospc;
1930 }
1931
1932 tree_mod_log_set_root_pointer(root, child, 1);
1933 rcu_assign_pointer(root->node, child);
1934
1935 add_root_to_dirty_list(root);
1936 btrfs_tree_unlock(child);
1937
1938 path->locks[level] = 0;
1939 path->nodes[level] = NULL;
1940 clean_tree_block(trans, fs_info, mid);
1941 btrfs_tree_unlock(mid);
1942 /* once for the path */
1943 free_extent_buffer(mid);
1944
1945 root_sub_used(root, mid->len);
1946 btrfs_free_tree_block(trans, root, mid, 0, 1);
1947 /* once for the root ptr */
1948 free_extent_buffer_stale(mid);
1949 return 0;
1950 }
1951 if (btrfs_header_nritems(mid) >
1952 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1953 return 0;
1954
1955 left = read_node_slot(fs_info, parent, pslot - 1);
1956 if (IS_ERR(left))
1957 left = NULL;
1958
1959 if (left) {
1960 btrfs_tree_lock(left);
1961 btrfs_set_lock_blocking(left);
1962 wret = btrfs_cow_block(trans, root, left,
1963 parent, pslot - 1, &left);
1964 if (wret) {
1965 ret = wret;
1966 goto enospc;
1967 }
1968 }
1969
1970 right = read_node_slot(fs_info, parent, pslot + 1);
1971 if (IS_ERR(right))
1972 right = NULL;
1973
1974 if (right) {
1975 btrfs_tree_lock(right);
1976 btrfs_set_lock_blocking(right);
1977 wret = btrfs_cow_block(trans, root, right,
1978 parent, pslot + 1, &right);
1979 if (wret) {
1980 ret = wret;
1981 goto enospc;
1982 }
1983 }
1984
1985 /* first, try to make some room in the middle buffer */
1986 if (left) {
1987 orig_slot += btrfs_header_nritems(left);
1988 wret = push_node_left(trans, fs_info, left, mid, 1);
1989 if (wret < 0)
1990 ret = wret;
1991 }
1992
1993 /*
1994 * then try to empty the right most buffer into the middle
1995 */
1996 if (right) {
1997 wret = push_node_left(trans, fs_info, mid, right, 1);
1998 if (wret < 0 && wret != -ENOSPC)
1999 ret = wret;
2000 if (btrfs_header_nritems(right) == 0) {
2001 clean_tree_block(trans, fs_info, right);
2002 btrfs_tree_unlock(right);
2003 del_ptr(root, path, level + 1, pslot + 1);
2004 root_sub_used(root, right->len);
2005 btrfs_free_tree_block(trans, root, right, 0, 1);
2006 free_extent_buffer_stale(right);
2007 right = NULL;
2008 } else {
2009 struct btrfs_disk_key right_key;
2010 btrfs_node_key(right, &right_key, 0);
2011 tree_mod_log_set_node_key(fs_info, parent,
2012 pslot + 1, 0);
2013 btrfs_set_node_key(parent, &right_key, pslot + 1);
2014 btrfs_mark_buffer_dirty(parent);
2015 }
2016 }
2017 if (btrfs_header_nritems(mid) == 1) {
2018 /*
2019 * we're not allowed to leave a node with one item in the
2020 * tree during a delete. A deletion from lower in the tree
2021 * could try to delete the only pointer in this node.
2022 * So, pull some keys from the left.
2023 * There has to be a left pointer at this point because
2024 * otherwise we would have pulled some pointers from the
2025 * right
2026 */
2027 if (!left) {
2028 ret = -EROFS;
2029 btrfs_handle_fs_error(fs_info, ret, NULL);
2030 goto enospc;
2031 }
2032 wret = balance_node_right(trans, fs_info, mid, left);
2033 if (wret < 0) {
2034 ret = wret;
2035 goto enospc;
2036 }
2037 if (wret == 1) {
2038 wret = push_node_left(trans, fs_info, left, mid, 1);
2039 if (wret < 0)
2040 ret = wret;
2041 }
2042 BUG_ON(wret == 1);
2043 }
2044 if (btrfs_header_nritems(mid) == 0) {
2045 clean_tree_block(trans, fs_info, mid);
2046 btrfs_tree_unlock(mid);
2047 del_ptr(root, path, level + 1, pslot);
2048 root_sub_used(root, mid->len);
2049 btrfs_free_tree_block(trans, root, mid, 0, 1);
2050 free_extent_buffer_stale(mid);
2051 mid = NULL;
2052 } else {
2053 /* update the parent key to reflect our changes */
2054 struct btrfs_disk_key mid_key;
2055 btrfs_node_key(mid, &mid_key, 0);
2056 tree_mod_log_set_node_key(fs_info, parent, pslot, 0);
2057 btrfs_set_node_key(parent, &mid_key, pslot);
2058 btrfs_mark_buffer_dirty(parent);
2059 }
2060
2061 /* update the path */
2062 if (left) {
2063 if (btrfs_header_nritems(left) > orig_slot) {
2064 extent_buffer_get(left);
2065 /* left was locked after cow */
2066 path->nodes[level] = left;
2067 path->slots[level + 1] -= 1;
2068 path->slots[level] = orig_slot;
2069 if (mid) {
2070 btrfs_tree_unlock(mid);
2071 free_extent_buffer(mid);
2072 }
2073 } else {
2074 orig_slot -= btrfs_header_nritems(left);
2075 path->slots[level] = orig_slot;
2076 }
2077 }
2078 /* double check we haven't messed things up */
2079 if (orig_ptr !=
2080 btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2081 BUG();
2082enospc:
2083 if (right) {
2084 btrfs_tree_unlock(right);
2085 free_extent_buffer(right);
2086 }
2087 if (left) {
2088 if (path->nodes[level] != left)
2089 btrfs_tree_unlock(left);
2090 free_extent_buffer(left);
2091 }
2092 return ret;
2093}
2094
2095/* Node balancing for insertion. Here we only split or push nodes around
2096 * when they are completely full. This is also done top down, so we
2097 * have to be pessimistic.
2098 */
2099static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2100 struct btrfs_root *root,
2101 struct btrfs_path *path, int level)
2102{
2103 struct btrfs_fs_info *fs_info = root->fs_info;
2104 struct extent_buffer *right = NULL;
2105 struct extent_buffer *mid;
2106 struct extent_buffer *left = NULL;
2107 struct extent_buffer *parent = NULL;
2108 int ret = 0;
2109 int wret;
2110 int pslot;
2111 int orig_slot = path->slots[level];
2112
2113 if (level == 0)
2114 return 1;
2115
2116 mid = path->nodes[level];
2117 WARN_ON(btrfs_header_generation(mid) != trans->transid);
2118
2119 if (level < BTRFS_MAX_LEVEL - 1) {
2120 parent = path->nodes[level + 1];
2121 pslot = path->slots[level + 1];
2122 }
2123
2124 if (!parent)
2125 return 1;
2126
2127 left = read_node_slot(fs_info, parent, pslot - 1);
2128 if (IS_ERR(left))
2129 left = NULL;
2130
2131 /* first, try to make some room in the middle buffer */
2132 if (left) {
2133 u32 left_nr;
2134
2135 btrfs_tree_lock(left);
2136 btrfs_set_lock_blocking(left);
2137
2138 left_nr = btrfs_header_nritems(left);
2139 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2140 wret = 1;
2141 } else {
2142 ret = btrfs_cow_block(trans, root, left, parent,
2143 pslot - 1, &left);
2144 if (ret)
2145 wret = 1;
2146 else {
2147 wret = push_node_left(trans, fs_info,
2148 left, mid, 0);
2149 }
2150 }
2151 if (wret < 0)
2152 ret = wret;
2153 if (wret == 0) {
2154 struct btrfs_disk_key disk_key;
2155 orig_slot += left_nr;
2156 btrfs_node_key(mid, &disk_key, 0);
2157 tree_mod_log_set_node_key(fs_info, parent, pslot, 0);
2158 btrfs_set_node_key(parent, &disk_key, pslot);
2159 btrfs_mark_buffer_dirty(parent);
2160 if (btrfs_header_nritems(left) > orig_slot) {
2161 path->nodes[level] = left;
2162 path->slots[level + 1] -= 1;
2163 path->slots[level] = orig_slot;
2164 btrfs_tree_unlock(mid);
2165 free_extent_buffer(mid);
2166 } else {
2167 orig_slot -=
2168 btrfs_header_nritems(left);
2169 path->slots[level] = orig_slot;
2170 btrfs_tree_unlock(left);
2171 free_extent_buffer(left);
2172 }
2173 return 0;
2174 }
2175 btrfs_tree_unlock(left);
2176 free_extent_buffer(left);
2177 }
2178 right = read_node_slot(fs_info, parent, pslot + 1);
2179 if (IS_ERR(right))
2180 right = NULL;
2181
2182 /*
2183 * then try to empty the right most buffer into the middle
2184 */
2185 if (right) {
2186 u32 right_nr;
2187
2188 btrfs_tree_lock(right);
2189 btrfs_set_lock_blocking(right);
2190
2191 right_nr = btrfs_header_nritems(right);
2192 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2193 wret = 1;
2194 } else {
2195 ret = btrfs_cow_block(trans, root, right,
2196 parent, pslot + 1,
2197 &right);
2198 if (ret)
2199 wret = 1;
2200 else {
2201 wret = balance_node_right(trans, fs_info,
2202 right, mid);
2203 }
2204 }
2205 if (wret < 0)
2206 ret = wret;
2207 if (wret == 0) {
2208 struct btrfs_disk_key disk_key;
2209
2210 btrfs_node_key(right, &disk_key, 0);
2211 tree_mod_log_set_node_key(fs_info, parent,
2212 pslot + 1, 0);
2213 btrfs_set_node_key(parent, &disk_key, pslot + 1);
2214 btrfs_mark_buffer_dirty(parent);
2215
2216 if (btrfs_header_nritems(mid) <= orig_slot) {
2217 path->nodes[level] = right;
2218 path->slots[level + 1] += 1;
2219 path->slots[level] = orig_slot -
2220 btrfs_header_nritems(mid);
2221 btrfs_tree_unlock(mid);
2222 free_extent_buffer(mid);
2223 } else {
2224 btrfs_tree_unlock(right);
2225 free_extent_buffer(right);
2226 }
2227 return 0;
2228 }
2229 btrfs_tree_unlock(right);
2230 free_extent_buffer(right);
2231 }
2232 return 1;
2233}
2234
2235/*
2236 * readahead one full node of leaves, finding things that are close
2237 * to the block in 'slot', and triggering ra on them.
2238 */
2239static void reada_for_search(struct btrfs_fs_info *fs_info,
2240 struct btrfs_path *path,
2241 int level, int slot, u64 objectid)
2242{
2243 struct extent_buffer *node;
2244 struct btrfs_disk_key disk_key;
2245 u32 nritems;
2246 u64 search;
2247 u64 target;
2248 u64 nread = 0;
2249 struct extent_buffer *eb;
2250 u32 nr;
2251 u32 blocksize;
2252 u32 nscan = 0;
2253
2254 if (level != 1)
2255 return;
2256
2257 if (!path->nodes[level])
2258 return;
2259
2260 node = path->nodes[level];
2261
2262 search = btrfs_node_blockptr(node, slot);
2263 blocksize = fs_info->nodesize;
2264 eb = find_extent_buffer(fs_info, search);
2265 if (eb) {
2266 free_extent_buffer(eb);
2267 return;
2268 }
2269
2270 target = search;
2271
2272 nritems = btrfs_header_nritems(node);
2273 nr = slot;
2274
2275 while (1) {
2276 if (path->reada == READA_BACK) {
2277 if (nr == 0)
2278 break;
2279 nr--;
2280 } else if (path->reada == READA_FORWARD) {
2281 nr++;
2282 if (nr >= nritems)
2283 break;
2284 }
2285 if (path->reada == READA_BACK && objectid) {
2286 btrfs_node_key(node, &disk_key, nr);
2287 if (btrfs_disk_key_objectid(&disk_key) != objectid)
2288 break;
2289 }
2290 search = btrfs_node_blockptr(node, nr);
2291 if ((search <= target && target - search <= 65536) ||
2292 (search > target && search - target <= 65536)) {
2293 readahead_tree_block(fs_info, search);
2294 nread += blocksize;
2295 }
2296 nscan++;
2297 if ((nread > 65536 || nscan > 32))
2298 break;
2299 }
2300}
2301
2302static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2303 struct btrfs_path *path, int level)
2304{
2305 int slot;
2306 int nritems;
2307 struct extent_buffer *parent;
2308 struct extent_buffer *eb;
2309 u64 gen;
2310 u64 block1 = 0;
2311 u64 block2 = 0;
2312
2313 parent = path->nodes[level + 1];
2314 if (!parent)
2315 return;
2316
2317 nritems = btrfs_header_nritems(parent);
2318 slot = path->slots[level + 1];
2319
2320 if (slot > 0) {
2321 block1 = btrfs_node_blockptr(parent, slot - 1);
2322 gen = btrfs_node_ptr_generation(parent, slot - 1);
2323 eb = find_extent_buffer(fs_info, block1);
2324 /*
2325 * if we get -eagain from btrfs_buffer_uptodate, we
2326 * don't want to return eagain here. That will loop
2327 * forever
2328 */
2329 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2330 block1 = 0;
2331 free_extent_buffer(eb);
2332 }
2333 if (slot + 1 < nritems) {
2334 block2 = btrfs_node_blockptr(parent, slot + 1);
2335 gen = btrfs_node_ptr_generation(parent, slot + 1);
2336 eb = find_extent_buffer(fs_info, block2);
2337 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2338 block2 = 0;
2339 free_extent_buffer(eb);
2340 }
2341
2342 if (block1)
2343 readahead_tree_block(fs_info, block1);
2344 if (block2)
2345 readahead_tree_block(fs_info, block2);
2346}
2347
2348
2349/*
2350 * when we walk down the tree, it is usually safe to unlock the higher layers
2351 * in the tree. The exceptions are when our path goes through slot 0, because
2352 * operations on the tree might require changing key pointers higher up in the
2353 * tree.
2354 *
2355 * callers might also have set path->keep_locks, which tells this code to keep
2356 * the lock if the path points to the last slot in the block. This is part of
2357 * walking through the tree, and selecting the next slot in the higher block.
2358 *
2359 * lowest_unlock sets the lowest level in the tree we're allowed to unlock. so
2360 * if lowest_unlock is 1, level 0 won't be unlocked
2361 */
2362static noinline void unlock_up(struct btrfs_path *path, int level,
2363 int lowest_unlock, int min_write_lock_level,
2364 int *write_lock_level)
2365{
2366 int i;
2367 int skip_level = level;
2368 int no_skips = 0;
2369 struct extent_buffer *t;
2370
2371 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2372 if (!path->nodes[i])
2373 break;
2374 if (!path->locks[i])
2375 break;
2376 if (!no_skips && path->slots[i] == 0) {
2377 skip_level = i + 1;
2378 continue;
2379 }
2380 if (!no_skips && path->keep_locks) {
2381 u32 nritems;
2382 t = path->nodes[i];
2383 nritems = btrfs_header_nritems(t);
2384 if (nritems < 1 || path->slots[i] >= nritems - 1) {
2385 skip_level = i + 1;
2386 continue;
2387 }
2388 }
2389 if (skip_level < i && i >= lowest_unlock)
2390 no_skips = 1;
2391
2392 t = path->nodes[i];
2393 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2394 btrfs_tree_unlock_rw(t, path->locks[i]);
2395 path->locks[i] = 0;
2396 if (write_lock_level &&
2397 i > min_write_lock_level &&
2398 i <= *write_lock_level) {
2399 *write_lock_level = i - 1;
2400 }
2401 }
2402 }
2403}
2404
2405/*
2406 * This releases any locks held in the path starting at level and
2407 * going all the way up to the root.
2408 *
2409 * btrfs_search_slot will keep the lock held on higher nodes in a few
2410 * corner cases, such as COW of the block at slot zero in the node. This
2411 * ignores those rules, and it should only be called when there are no
2412 * more updates to be done higher up in the tree.
2413 */
2414noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2415{
2416 int i;
2417
2418 if (path->keep_locks)
2419 return;
2420
2421 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2422 if (!path->nodes[i])
2423 continue;
2424 if (!path->locks[i])
2425 continue;
2426 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2427 path->locks[i] = 0;
2428 }
2429}
2430
2431/*
2432 * helper function for btrfs_search_slot. The goal is to find a block
2433 * in cache without setting the path to blocking. If we find the block
2434 * we return zero and the path is unchanged.
2435 *
2436 * If we can't find the block, we set the path blocking and do some
2437 * reada. -EAGAIN is returned and the search must be repeated.
2438 */
2439static int
2440read_block_for_search(struct btrfs_trans_handle *trans,
2441 struct btrfs_root *root, struct btrfs_path *p,
2442 struct extent_buffer **eb_ret, int level, int slot,
2443 struct btrfs_key *key, u64 time_seq)
2444{
2445 struct btrfs_fs_info *fs_info = root->fs_info;
2446 u64 blocknr;
2447 u64 gen;
2448 struct extent_buffer *b = *eb_ret;
2449 struct extent_buffer *tmp;
2450 int ret;
2451
2452 blocknr = btrfs_node_blockptr(b, slot);
2453 gen = btrfs_node_ptr_generation(b, slot);
2454
2455 tmp = find_extent_buffer(fs_info, blocknr);
2456 if (tmp) {
2457 /* first we do an atomic uptodate check */
2458 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2459 *eb_ret = tmp;
2460 return 0;
2461 }
2462
2463 /* the pages were up to date, but we failed
2464 * the generation number check. Do a full
2465 * read for the generation number that is correct.
2466 * We must do this without dropping locks so
2467 * we can trust our generation number
2468 */
2469 btrfs_set_path_blocking(p);
2470
2471 /* now we're allowed to do a blocking uptodate check */
2472 ret = btrfs_read_buffer(tmp, gen);
2473 if (!ret) {
2474 *eb_ret = tmp;
2475 return 0;
2476 }
2477 free_extent_buffer(tmp);
2478 btrfs_release_path(p);
2479 return -EIO;
2480 }
2481
2482 /*
2483 * reduce lock contention at high levels
2484 * of the btree by dropping locks before
2485 * we read. Don't release the lock on the current
2486 * level because we need to walk this node to figure
2487 * out which blocks to read.
2488 */
2489 btrfs_unlock_up_safe(p, level + 1);
2490 btrfs_set_path_blocking(p);
2491
2492 free_extent_buffer(tmp);
2493 if (p->reada != READA_NONE)
2494 reada_for_search(fs_info, p, level, slot, key->objectid);
2495
2496 btrfs_release_path(p);
2497
2498 ret = -EAGAIN;
2499 tmp = read_tree_block(fs_info, blocknr, 0);
2500 if (!IS_ERR(tmp)) {
2501 /*
2502 * If the read above didn't mark this buffer up to date,
2503 * it will never end up being up to date. Set ret to EIO now
2504 * and give up so that our caller doesn't loop forever
2505 * on our EAGAINs.
2506 */
2507 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2508 ret = -EIO;
2509 free_extent_buffer(tmp);
2510 } else {
2511 ret = PTR_ERR(tmp);
2512 }
2513 return ret;
2514}
2515
2516/*
2517 * helper function for btrfs_search_slot. This does all of the checks
2518 * for node-level blocks and does any balancing required based on
2519 * the ins_len.
2520 *
2521 * If no extra work was required, zero is returned. If we had to
2522 * drop the path, -EAGAIN is returned and btrfs_search_slot must
2523 * start over
2524 */
2525static int
2526setup_nodes_for_search(struct btrfs_trans_handle *trans,
2527 struct btrfs_root *root, struct btrfs_path *p,
2528 struct extent_buffer *b, int level, int ins_len,
2529 int *write_lock_level)
2530{
2531 struct btrfs_fs_info *fs_info = root->fs_info;
2532 int ret;
2533
2534 if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2535 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2536 int sret;
2537
2538 if (*write_lock_level < level + 1) {
2539 *write_lock_level = level + 1;
2540 btrfs_release_path(p);
2541 goto again;
2542 }
2543
2544 btrfs_set_path_blocking(p);
2545 reada_for_balance(fs_info, p, level);
2546 sret = split_node(trans, root, p, level);
2547 btrfs_clear_path_blocking(p, NULL, 0);
2548
2549 BUG_ON(sret > 0);
2550 if (sret) {
2551 ret = sret;
2552 goto done;
2553 }
2554 b = p->nodes[level];
2555 } else if (ins_len < 0 && btrfs_header_nritems(b) <
2556 BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2557 int sret;
2558
2559 if (*write_lock_level < level + 1) {
2560 *write_lock_level = level + 1;
2561 btrfs_release_path(p);
2562 goto again;
2563 }
2564
2565 btrfs_set_path_blocking(p);
2566 reada_for_balance(fs_info, p, level);
2567 sret = balance_level(trans, root, p, level);
2568 btrfs_clear_path_blocking(p, NULL, 0);
2569
2570 if (sret) {
2571 ret = sret;
2572 goto done;
2573 }
2574 b = p->nodes[level];
2575 if (!b) {
2576 btrfs_release_path(p);
2577 goto again;
2578 }
2579 BUG_ON(btrfs_header_nritems(b) == 1);
2580 }
2581 return 0;
2582
2583again:
2584 ret = -EAGAIN;
2585done:
2586 return ret;
2587}
2588
2589static void key_search_validate(struct extent_buffer *b,
2590 struct btrfs_key *key,
2591 int level)
2592{
2593#ifdef CONFIG_BTRFS_ASSERT
2594 struct btrfs_disk_key disk_key;
2595
2596 btrfs_cpu_key_to_disk(&disk_key, key);
2597
2598 if (level == 0)
2599 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2600 offsetof(struct btrfs_leaf, items[0].key),
2601 sizeof(disk_key)));
2602 else
2603 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2604 offsetof(struct btrfs_node, ptrs[0].key),
2605 sizeof(disk_key)));
2606#endif
2607}
2608
2609static int key_search(struct extent_buffer *b, struct btrfs_key *key,
2610 int level, int *prev_cmp, int *slot)
2611{
2612 if (*prev_cmp != 0) {
2613 *prev_cmp = bin_search(b, key, level, slot);
2614 return *prev_cmp;
2615 }
2616
2617 key_search_validate(b, key, level);
2618 *slot = 0;
2619
2620 return 0;
2621}
2622
2623int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2624 u64 iobjectid, u64 ioff, u8 key_type,
2625 struct btrfs_key *found_key)
2626{
2627 int ret;
2628 struct btrfs_key key;
2629 struct extent_buffer *eb;
2630
2631 ASSERT(path);
2632 ASSERT(found_key);
2633
2634 key.type = key_type;
2635 key.objectid = iobjectid;
2636 key.offset = ioff;
2637
2638 ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2639 if (ret < 0)
2640 return ret;
2641
2642 eb = path->nodes[0];
2643 if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2644 ret = btrfs_next_leaf(fs_root, path);
2645 if (ret)
2646 return ret;
2647 eb = path->nodes[0];
2648 }
2649
2650 btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2651 if (found_key->type != key.type ||
2652 found_key->objectid != key.objectid)
2653 return 1;
2654
2655 return 0;
2656}
2657
2658/*
2659 * look for key in the tree. path is filled in with nodes along the way
2660 * if key is found, we return zero and you can find the item in the leaf
2661 * level of the path (level 0)
2662 *
2663 * If the key isn't found, the path points to the slot where it should
2664 * be inserted, and 1 is returned. If there are other errors during the
2665 * search a negative error number is returned.
2666 *
2667 * if ins_len > 0, nodes and leaves will be split as we walk down the
2668 * tree. if ins_len < 0, nodes will be merged as we walk down the tree (if
2669 * possible)
2670 */
2671int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
2672 *root, struct btrfs_key *key, struct btrfs_path *p, int
2673 ins_len, int cow)
2674{
2675 struct btrfs_fs_info *fs_info = root->fs_info;
2676 struct extent_buffer *b;
2677 int slot;
2678 int ret;
2679 int err;
2680 int level;
2681 int lowest_unlock = 1;
2682 int root_lock;
2683 /* everything at write_lock_level or lower must be write locked */
2684 int write_lock_level = 0;
2685 u8 lowest_level = 0;
2686 int min_write_lock_level;
2687 int prev_cmp;
2688
2689 lowest_level = p->lowest_level;
2690 WARN_ON(lowest_level && ins_len > 0);
2691 WARN_ON(p->nodes[0] != NULL);
2692 BUG_ON(!cow && ins_len);
2693
2694 if (ins_len < 0) {
2695 lowest_unlock = 2;
2696
2697 /* when we are removing items, we might have to go up to level
2698 * two as we update tree pointers Make sure we keep write
2699 * for those levels as well
2700 */
2701 write_lock_level = 2;
2702 } else if (ins_len > 0) {
2703 /*
2704 * for inserting items, make sure we have a write lock on
2705 * level 1 so we can update keys
2706 */
2707 write_lock_level = 1;
2708 }
2709
2710 if (!cow)
2711 write_lock_level = -1;
2712
2713 if (cow && (p->keep_locks || p->lowest_level))
2714 write_lock_level = BTRFS_MAX_LEVEL;
2715
2716 min_write_lock_level = write_lock_level;
2717
2718again:
2719 prev_cmp = -1;
2720 /*
2721 * we try very hard to do read locks on the root
2722 */
2723 root_lock = BTRFS_READ_LOCK;
2724 level = 0;
2725 if (p->search_commit_root) {
2726 /*
2727 * the commit roots are read only
2728 * so we always do read locks
2729 */
2730 if (p->need_commit_sem)
2731 down_read(&fs_info->commit_root_sem);
2732 b = root->commit_root;
2733 extent_buffer_get(b);
2734 level = btrfs_header_level(b);
2735 if (p->need_commit_sem)
2736 up_read(&fs_info->commit_root_sem);
2737 if (!p->skip_locking)
2738 btrfs_tree_read_lock(b);
2739 } else {
2740 if (p->skip_locking) {
2741 b = btrfs_root_node(root);
2742 level = btrfs_header_level(b);
2743 } else {
2744 /* we don't know the level of the root node
2745 * until we actually have it read locked
2746 */
2747 b = btrfs_read_lock_root_node(root);
2748 level = btrfs_header_level(b);
2749 if (level <= write_lock_level) {
2750 /* whoops, must trade for write lock */
2751 btrfs_tree_read_unlock(b);
2752 free_extent_buffer(b);
2753 b = btrfs_lock_root_node(root);
2754 root_lock = BTRFS_WRITE_LOCK;
2755
2756 /* the level might have changed, check again */
2757 level = btrfs_header_level(b);
2758 }
2759 }
2760 }
2761 p->nodes[level] = b;
2762 if (!p->skip_locking)
2763 p->locks[level] = root_lock;
2764
2765 while (b) {
2766 level = btrfs_header_level(b);
2767
2768 /*
2769 * setup the path here so we can release it under lock
2770 * contention with the cow code
2771 */
2772 if (cow) {
2773 /*
2774 * if we don't really need to cow this block
2775 * then we don't want to set the path blocking,
2776 * so we test it here
2777 */
2778 if (!should_cow_block(trans, root, b)) {
2779 trans->dirty = true;
2780 goto cow_done;
2781 }
2782
2783 /*
2784 * must have write locks on this node and the
2785 * parent
2786 */
2787 if (level > write_lock_level ||
2788 (level + 1 > write_lock_level &&
2789 level + 1 < BTRFS_MAX_LEVEL &&
2790 p->nodes[level + 1])) {
2791 write_lock_level = level + 1;
2792 btrfs_release_path(p);
2793 goto again;
2794 }
2795
2796 btrfs_set_path_blocking(p);
2797 err = btrfs_cow_block(trans, root, b,
2798 p->nodes[level + 1],
2799 p->slots[level + 1], &b);
2800 if (err) {
2801 ret = err;
2802 goto done;
2803 }
2804 }
2805cow_done:
2806 p->nodes[level] = b;
2807 btrfs_clear_path_blocking(p, NULL, 0);
2808
2809 /*
2810 * we have a lock on b and as long as we aren't changing
2811 * the tree, there is no way to for the items in b to change.
2812 * It is safe to drop the lock on our parent before we
2813 * go through the expensive btree search on b.
2814 *
2815 * If we're inserting or deleting (ins_len != 0), then we might
2816 * be changing slot zero, which may require changing the parent.
2817 * So, we can't drop the lock until after we know which slot
2818 * we're operating on.
2819 */
2820 if (!ins_len && !p->keep_locks) {
2821 int u = level + 1;
2822
2823 if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2824 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2825 p->locks[u] = 0;
2826 }
2827 }
2828
2829 ret = key_search(b, key, level, &prev_cmp, &slot);
2830 if (ret < 0)
2831 goto done;
2832
2833 if (level != 0) {
2834 int dec = 0;
2835 if (ret && slot > 0) {
2836 dec = 1;
2837 slot -= 1;
2838 }
2839 p->slots[level] = slot;
2840 err = setup_nodes_for_search(trans, root, p, b, level,
2841 ins_len, &write_lock_level);
2842 if (err == -EAGAIN)
2843 goto again;
2844 if (err) {
2845 ret = err;
2846 goto done;
2847 }
2848 b = p->nodes[level];
2849 slot = p->slots[level];
2850
2851 /*
2852 * slot 0 is special, if we change the key
2853 * we have to update the parent pointer
2854 * which means we must have a write lock
2855 * on the parent
2856 */
2857 if (slot == 0 && ins_len &&
2858 write_lock_level < level + 1) {
2859 write_lock_level = level + 1;
2860 btrfs_release_path(p);
2861 goto again;
2862 }
2863
2864 unlock_up(p, level, lowest_unlock,
2865 min_write_lock_level, &write_lock_level);
2866
2867 if (level == lowest_level) {
2868 if (dec)
2869 p->slots[level]++;
2870 goto done;
2871 }
2872
2873 err = read_block_for_search(trans, root, p,
2874 &b, level, slot, key, 0);
2875 if (err == -EAGAIN)
2876 goto again;
2877 if (err) {
2878 ret = err;
2879 goto done;
2880 }
2881
2882 if (!p->skip_locking) {
2883 level = btrfs_header_level(b);
2884 if (level <= write_lock_level) {
2885 err = btrfs_try_tree_write_lock(b);
2886 if (!err) {
2887 btrfs_set_path_blocking(p);
2888 btrfs_tree_lock(b);
2889 btrfs_clear_path_blocking(p, b,
2890 BTRFS_WRITE_LOCK);
2891 }
2892 p->locks[level] = BTRFS_WRITE_LOCK;
2893 } else {
2894 err = btrfs_tree_read_lock_atomic(b);
2895 if (!err) {
2896 btrfs_set_path_blocking(p);
2897 btrfs_tree_read_lock(b);
2898 btrfs_clear_path_blocking(p, b,
2899 BTRFS_READ_LOCK);
2900 }
2901 p->locks[level] = BTRFS_READ_LOCK;
2902 }
2903 p->nodes[level] = b;
2904 }
2905 } else {
2906 p->slots[level] = slot;
2907 if (ins_len > 0 &&
2908 btrfs_leaf_free_space(fs_info, b) < ins_len) {
2909 if (write_lock_level < 1) {
2910 write_lock_level = 1;
2911 btrfs_release_path(p);
2912 goto again;
2913 }
2914
2915 btrfs_set_path_blocking(p);
2916 err = split_leaf(trans, root, key,
2917 p, ins_len, ret == 0);
2918 btrfs_clear_path_blocking(p, NULL, 0);
2919
2920 BUG_ON(err > 0);
2921 if (err) {
2922 ret = err;
2923 goto done;
2924 }
2925 }
2926 if (!p->search_for_split)
2927 unlock_up(p, level, lowest_unlock,
2928 min_write_lock_level, &write_lock_level);
2929 goto done;
2930 }
2931 }
2932 ret = 1;
2933done:
2934 /*
2935 * we don't really know what they plan on doing with the path
2936 * from here on, so for now just mark it as blocking
2937 */
2938 if (!p->leave_spinning)
2939 btrfs_set_path_blocking(p);
2940 if (ret < 0 && !p->skip_release_on_error)
2941 btrfs_release_path(p);
2942 return ret;
2943}
2944
2945/*
2946 * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2947 * current state of the tree together with the operations recorded in the tree
2948 * modification log to search for the key in a previous version of this tree, as
2949 * denoted by the time_seq parameter.
2950 *
2951 * Naturally, there is no support for insert, delete or cow operations.
2952 *
2953 * The resulting path and return value will be set up as if we called
2954 * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2955 */
2956int btrfs_search_old_slot(struct btrfs_root *root, struct btrfs_key *key,
2957 struct btrfs_path *p, u64 time_seq)
2958{
2959 struct btrfs_fs_info *fs_info = root->fs_info;
2960 struct extent_buffer *b;
2961 int slot;
2962 int ret;
2963 int err;
2964 int level;
2965 int lowest_unlock = 1;
2966 u8 lowest_level = 0;
2967 int prev_cmp = -1;
2968
2969 lowest_level = p->lowest_level;
2970 WARN_ON(p->nodes[0] != NULL);
2971
2972 if (p->search_commit_root) {
2973 BUG_ON(time_seq);
2974 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2975 }
2976
2977again:
2978 b = get_old_root(root, time_seq);
2979 level = btrfs_header_level(b);
2980 p->locks[level] = BTRFS_READ_LOCK;
2981
2982 while (b) {
2983 level = btrfs_header_level(b);
2984 p->nodes[level] = b;
2985 btrfs_clear_path_blocking(p, NULL, 0);
2986
2987 /*
2988 * we have a lock on b and as long as we aren't changing
2989 * the tree, there is no way to for the items in b to change.
2990 * It is safe to drop the lock on our parent before we
2991 * go through the expensive btree search on b.
2992 */
2993 btrfs_unlock_up_safe(p, level + 1);
2994
2995 /*
2996 * Since we can unwind ebs we want to do a real search every
2997 * time.
2998 */
2999 prev_cmp = -1;
3000 ret = key_search(b, key, level, &prev_cmp, &slot);
3001
3002 if (level != 0) {
3003 int dec = 0;
3004 if (ret && slot > 0) {
3005 dec = 1;
3006 slot -= 1;
3007 }
3008 p->slots[level] = slot;
3009 unlock_up(p, level, lowest_unlock, 0, NULL);
3010
3011 if (level == lowest_level) {
3012 if (dec)
3013 p->slots[level]++;
3014 goto done;
3015 }
3016
3017 err = read_block_for_search(NULL, root, p, &b, level,
3018 slot, key, time_seq);
3019 if (err == -EAGAIN)
3020 goto again;
3021 if (err) {
3022 ret = err;
3023 goto done;
3024 }
3025
3026 level = btrfs_header_level(b);
3027 err = btrfs_tree_read_lock_atomic(b);
3028 if (!err) {
3029 btrfs_set_path_blocking(p);
3030 btrfs_tree_read_lock(b);
3031 btrfs_clear_path_blocking(p, b,
3032 BTRFS_READ_LOCK);
3033 }
3034 b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3035 if (!b) {
3036 ret = -ENOMEM;
3037 goto done;
3038 }
3039 p->locks[level] = BTRFS_READ_LOCK;
3040 p->nodes[level] = b;
3041 } else {
3042 p->slots[level] = slot;
3043 unlock_up(p, level, lowest_unlock, 0, NULL);
3044 goto done;
3045 }
3046 }
3047 ret = 1;
3048done:
3049 if (!p->leave_spinning)
3050 btrfs_set_path_blocking(p);
3051 if (ret < 0)
3052 btrfs_release_path(p);
3053
3054 return ret;
3055}
3056
3057/*
3058 * helper to use instead of search slot if no exact match is needed but
3059 * instead the next or previous item should be returned.
3060 * When find_higher is true, the next higher item is returned, the next lower
3061 * otherwise.
3062 * When return_any and find_higher are both true, and no higher item is found,
3063 * return the next lower instead.
3064 * When return_any is true and find_higher is false, and no lower item is found,
3065 * return the next higher instead.
3066 * It returns 0 if any item is found, 1 if none is found (tree empty), and
3067 * < 0 on error
3068 */
3069int btrfs_search_slot_for_read(struct btrfs_root *root,
3070 struct btrfs_key *key, struct btrfs_path *p,
3071 int find_higher, int return_any)
3072{
3073 int ret;
3074 struct extent_buffer *leaf;
3075
3076again:
3077 ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3078 if (ret <= 0)
3079 return ret;
3080 /*
3081 * a return value of 1 means the path is at the position where the
3082 * item should be inserted. Normally this is the next bigger item,
3083 * but in case the previous item is the last in a leaf, path points
3084 * to the first free slot in the previous leaf, i.e. at an invalid
3085 * item.
3086 */
3087 leaf = p->nodes[0];
3088
3089 if (find_higher) {
3090 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3091 ret = btrfs_next_leaf(root, p);
3092 if (ret <= 0)
3093 return ret;
3094 if (!return_any)
3095 return 1;
3096 /*
3097 * no higher item found, return the next
3098 * lower instead
3099 */
3100 return_any = 0;
3101 find_higher = 0;
3102 btrfs_release_path(p);
3103 goto again;
3104 }
3105 } else {
3106 if (p->slots[0] == 0) {
3107 ret = btrfs_prev_leaf(root, p);
3108 if (ret < 0)
3109 return ret;
3110 if (!ret) {
3111 leaf = p->nodes[0];
3112 if (p->slots[0] == btrfs_header_nritems(leaf))
3113 p->slots[0]--;
3114 return 0;
3115 }
3116 if (!return_any)
3117 return 1;
3118 /*
3119 * no lower item found, return the next
3120 * higher instead
3121 */
3122 return_any = 0;
3123 find_higher = 1;
3124 btrfs_release_path(p);
3125 goto again;
3126 } else {
3127 --p->slots[0];
3128 }
3129 }
3130 return 0;
3131}
3132
3133/*
3134 * adjust the pointers going up the tree, starting at level
3135 * making sure the right key of each node is points to 'key'.
3136 * This is used after shifting pointers to the left, so it stops
3137 * fixing up pointers when a given leaf/node is not in slot 0 of the
3138 * higher levels
3139 *
3140 */
3141static void fixup_low_keys(struct btrfs_fs_info *fs_info,
3142 struct btrfs_path *path,
3143 struct btrfs_disk_key *key, int level)
3144{
3145 int i;
3146 struct extent_buffer *t;
3147
3148 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3149 int tslot = path->slots[i];
3150 if (!path->nodes[i])
3151 break;
3152 t = path->nodes[i];
3153 tree_mod_log_set_node_key(fs_info, t, tslot, 1);
3154 btrfs_set_node_key(t, key, tslot);
3155 btrfs_mark_buffer_dirty(path->nodes[i]);
3156 if (tslot != 0)
3157 break;
3158 }
3159}
3160
3161/*
3162 * update item key.
3163 *
3164 * This function isn't completely safe. It's the caller's responsibility
3165 * that the new key won't break the order
3166 */
3167void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3168 struct btrfs_path *path,
3169 struct btrfs_key *new_key)
3170{
3171 struct btrfs_disk_key disk_key;
3172 struct extent_buffer *eb;
3173 int slot;
3174
3175 eb = path->nodes[0];
3176 slot = path->slots[0];
3177 if (slot > 0) {
3178 btrfs_item_key(eb, &disk_key, slot - 1);
3179 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
3180 }
3181 if (slot < btrfs_header_nritems(eb) - 1) {
3182 btrfs_item_key(eb, &disk_key, slot + 1);
3183 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
3184 }
3185
3186 btrfs_cpu_key_to_disk(&disk_key, new_key);
3187 btrfs_set_item_key(eb, &disk_key, slot);
3188 btrfs_mark_buffer_dirty(eb);
3189 if (slot == 0)
3190 fixup_low_keys(fs_info, path, &disk_key, 1);
3191}
3192
3193/*
3194 * try to push data from one node into the next node left in the
3195 * tree.
3196 *
3197 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3198 * error, and > 0 if there was no room in the left hand block.
3199 */
3200static int push_node_left(struct btrfs_trans_handle *trans,
3201 struct btrfs_fs_info *fs_info,
3202 struct extent_buffer *dst,
3203 struct extent_buffer *src, int empty)
3204{
3205 int push_items = 0;
3206 int src_nritems;
3207 int dst_nritems;
3208 int ret = 0;
3209
3210 src_nritems = btrfs_header_nritems(src);
3211 dst_nritems = btrfs_header_nritems(dst);
3212 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3213 WARN_ON(btrfs_header_generation(src) != trans->transid);
3214 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3215
3216 if (!empty && src_nritems <= 8)
3217 return 1;
3218
3219 if (push_items <= 0)
3220 return 1;
3221
3222 if (empty) {
3223 push_items = min(src_nritems, push_items);
3224 if (push_items < src_nritems) {
3225 /* leave at least 8 pointers in the node if
3226 * we aren't going to empty it
3227 */
3228 if (src_nritems - push_items < 8) {
3229 if (push_items <= 8)
3230 return 1;
3231 push_items -= 8;
3232 }
3233 }
3234 } else
3235 push_items = min(src_nritems - 8, push_items);
3236
3237 ret = tree_mod_log_eb_copy(fs_info, dst, src, dst_nritems, 0,
3238 push_items);
3239 if (ret) {
3240 btrfs_abort_transaction(trans, ret);
3241 return ret;
3242 }
3243 copy_extent_buffer(dst, src,
3244 btrfs_node_key_ptr_offset(dst_nritems),
3245 btrfs_node_key_ptr_offset(0),
3246 push_items * sizeof(struct btrfs_key_ptr));
3247
3248 if (push_items < src_nritems) {
3249 /*
3250 * don't call tree_mod_log_eb_move here, key removal was already
3251 * fully logged by tree_mod_log_eb_copy above.
3252 */
3253 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3254 btrfs_node_key_ptr_offset(push_items),
3255 (src_nritems - push_items) *
3256 sizeof(struct btrfs_key_ptr));
3257 }
3258 btrfs_set_header_nritems(src, src_nritems - push_items);
3259 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3260 btrfs_mark_buffer_dirty(src);
3261 btrfs_mark_buffer_dirty(dst);
3262
3263 return ret;
3264}
3265
3266/*
3267 * try to push data from one node into the next node right in the
3268 * tree.
3269 *
3270 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3271 * error, and > 0 if there was no room in the right hand block.
3272 *
3273 * this will only push up to 1/2 the contents of the left node over
3274 */
3275static int balance_node_right(struct btrfs_trans_handle *trans,
3276 struct btrfs_fs_info *fs_info,
3277 struct extent_buffer *dst,
3278 struct extent_buffer *src)
3279{
3280 int push_items = 0;
3281 int max_push;
3282 int src_nritems;
3283 int dst_nritems;
3284 int ret = 0;
3285
3286 WARN_ON(btrfs_header_generation(src) != trans->transid);
3287 WARN_ON(btrfs_header_generation(dst) != trans->transid);
3288
3289 src_nritems = btrfs_header_nritems(src);
3290 dst_nritems = btrfs_header_nritems(dst);
3291 push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3292 if (push_items <= 0)
3293 return 1;
3294
3295 if (src_nritems < 4)
3296 return 1;
3297
3298 max_push = src_nritems / 2 + 1;
3299 /* don't try to empty the node */
3300 if (max_push >= src_nritems)
3301 return 1;
3302
3303 if (max_push < push_items)
3304 push_items = max_push;
3305
3306 tree_mod_log_eb_move(fs_info, dst, push_items, 0, dst_nritems);
3307 memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3308 btrfs_node_key_ptr_offset(0),
3309 (dst_nritems) *
3310 sizeof(struct btrfs_key_ptr));
3311
3312 ret = tree_mod_log_eb_copy(fs_info, dst, src, 0,
3313 src_nritems - push_items, push_items);
3314 if (ret) {
3315 btrfs_abort_transaction(trans, ret);
3316 return ret;
3317 }
3318 copy_extent_buffer(dst, src,
3319 btrfs_node_key_ptr_offset(0),
3320 btrfs_node_key_ptr_offset(src_nritems - push_items),
3321 push_items * sizeof(struct btrfs_key_ptr));
3322
3323 btrfs_set_header_nritems(src, src_nritems - push_items);
3324 btrfs_set_header_nritems(dst, dst_nritems + push_items);
3325
3326 btrfs_mark_buffer_dirty(src);
3327 btrfs_mark_buffer_dirty(dst);
3328
3329 return ret;
3330}
3331
3332/*
3333 * helper function to insert a new root level in the tree.
3334 * A new node is allocated, and a single item is inserted to
3335 * point to the existing root
3336 *
3337 * returns zero on success or < 0 on failure.
3338 */
3339static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3340 struct btrfs_root *root,
3341 struct btrfs_path *path, int level)
3342{
3343 struct btrfs_fs_info *fs_info = root->fs_info;
3344 u64 lower_gen;
3345 struct extent_buffer *lower;
3346 struct extent_buffer *c;
3347 struct extent_buffer *old;
3348 struct btrfs_disk_key lower_key;
3349
3350 BUG_ON(path->nodes[level]);
3351 BUG_ON(path->nodes[level-1] != root->node);
3352
3353 lower = path->nodes[level-1];
3354 if (level == 1)
3355 btrfs_item_key(lower, &lower_key, 0);
3356 else
3357 btrfs_node_key(lower, &lower_key, 0);
3358
3359 c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3360 &lower_key, level, root->node->start, 0);
3361 if (IS_ERR(c))
3362 return PTR_ERR(c);
3363
3364 root_add_used(root, fs_info->nodesize);
3365
3366 memzero_extent_buffer(c, 0, sizeof(struct btrfs_header));
3367 btrfs_set_header_nritems(c, 1);
3368 btrfs_set_header_level(c, level);
3369 btrfs_set_header_bytenr(c, c->start);
3370 btrfs_set_header_generation(c, trans->transid);
3371 btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
3372 btrfs_set_header_owner(c, root->root_key.objectid);
3373
3374 write_extent_buffer_fsid(c, fs_info->fsid);
3375 write_extent_buffer_chunk_tree_uuid(c, fs_info->chunk_tree_uuid);
3376
3377 btrfs_set_node_key(c, &lower_key, 0);
3378 btrfs_set_node_blockptr(c, 0, lower->start);
3379 lower_gen = btrfs_header_generation(lower);
3380 WARN_ON(lower_gen != trans->transid);
3381
3382 btrfs_set_node_ptr_generation(c, 0, lower_gen);
3383
3384 btrfs_mark_buffer_dirty(c);
3385
3386 old = root->node;
3387 tree_mod_log_set_root_pointer(root, c, 0);
3388 rcu_assign_pointer(root->node, c);
3389
3390 /* the super has an extra ref to root->node */
3391 free_extent_buffer(old);
3392
3393 add_root_to_dirty_list(root);
3394 extent_buffer_get(c);
3395 path->nodes[level] = c;
3396 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3397 path->slots[level] = 0;
3398 return 0;
3399}
3400
3401/*
3402 * worker function to insert a single pointer in a node.
3403 * the node should have enough room for the pointer already
3404 *
3405 * slot and level indicate where you want the key to go, and
3406 * blocknr is the block the key points to.
3407 */
3408static void insert_ptr(struct btrfs_trans_handle *trans,
3409 struct btrfs_fs_info *fs_info, struct btrfs_path *path,
3410 struct btrfs_disk_key *key, u64 bytenr,
3411 int slot, int level)
3412{
3413 struct extent_buffer *lower;
3414 int nritems;
3415 int ret;
3416
3417 BUG_ON(!path->nodes[level]);
3418 btrfs_assert_tree_locked(path->nodes[level]);
3419 lower = path->nodes[level];
3420 nritems = btrfs_header_nritems(lower);
3421 BUG_ON(slot > nritems);
3422 BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(fs_info));
3423 if (slot != nritems) {
3424 if (level)
3425 tree_mod_log_eb_move(fs_info, lower, slot + 1,
3426 slot, nritems - slot);
3427 memmove_extent_buffer(lower,
3428 btrfs_node_key_ptr_offset(slot + 1),
3429 btrfs_node_key_ptr_offset(slot),
3430 (nritems - slot) * sizeof(struct btrfs_key_ptr));
3431 }
3432 if (level) {
3433 ret = tree_mod_log_insert_key(fs_info, lower, slot,
3434 MOD_LOG_KEY_ADD, GFP_NOFS);
3435 BUG_ON(ret < 0);
3436 }
3437 btrfs_set_node_key(lower, key, slot);
3438 btrfs_set_node_blockptr(lower, slot, bytenr);
3439 WARN_ON(trans->transid == 0);
3440 btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3441 btrfs_set_header_nritems(lower, nritems + 1);
3442 btrfs_mark_buffer_dirty(lower);
3443}
3444
3445/*
3446 * split the node at the specified level in path in two.
3447 * The path is corrected to point to the appropriate node after the split
3448 *
3449 * Before splitting this tries to make some room in the node by pushing
3450 * left and right, if either one works, it returns right away.
3451 *
3452 * returns 0 on success and < 0 on failure
3453 */
3454static noinline int split_node(struct btrfs_trans_handle *trans,
3455 struct btrfs_root *root,
3456 struct btrfs_path *path, int level)
3457{
3458 struct btrfs_fs_info *fs_info = root->fs_info;
3459 struct extent_buffer *c;
3460 struct extent_buffer *split;
3461 struct btrfs_disk_key disk_key;
3462 int mid;
3463 int ret;
3464 u32 c_nritems;
3465
3466 c = path->nodes[level];
3467 WARN_ON(btrfs_header_generation(c) != trans->transid);
3468 if (c == root->node) {
3469 /*
3470 * trying to split the root, lets make a new one
3471 *
3472 * tree mod log: We don't log_removal old root in
3473 * insert_new_root, because that root buffer will be kept as a
3474 * normal node. We are going to log removal of half of the
3475 * elements below with tree_mod_log_eb_copy. We're holding a
3476 * tree lock on the buffer, which is why we cannot race with
3477 * other tree_mod_log users.
3478 */
3479 ret = insert_new_root(trans, root, path, level + 1);
3480 if (ret)
3481 return ret;
3482 } else {
3483 ret = push_nodes_for_insert(trans, root, path, level);
3484 c = path->nodes[level];
3485 if (!ret && btrfs_header_nritems(c) <
3486 BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3487 return 0;
3488 if (ret < 0)
3489 return ret;
3490 }
3491
3492 c_nritems = btrfs_header_nritems(c);
3493 mid = (c_nritems + 1) / 2;
3494 btrfs_node_key(c, &disk_key, mid);
3495
3496 split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3497 &disk_key, level, c->start, 0);
3498 if (IS_ERR(split))
3499 return PTR_ERR(split);
3500
3501 root_add_used(root, fs_info->nodesize);
3502
3503 memzero_extent_buffer(split, 0, sizeof(struct btrfs_header));
3504 btrfs_set_header_level(split, btrfs_header_level(c));
3505 btrfs_set_header_bytenr(split, split->start);
3506 btrfs_set_header_generation(split, trans->transid);
3507 btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
3508 btrfs_set_header_owner(split, root->root_key.objectid);
3509 write_extent_buffer_fsid(split, fs_info->fsid);
3510 write_extent_buffer_chunk_tree_uuid(split, fs_info->chunk_tree_uuid);
3511
3512 ret = tree_mod_log_eb_copy(fs_info, split, c, 0, mid, c_nritems - mid);
3513 if (ret) {
3514 btrfs_abort_transaction(trans, ret);
3515 return ret;
3516 }
3517 copy_extent_buffer(split, c,
3518 btrfs_node_key_ptr_offset(0),
3519 btrfs_node_key_ptr_offset(mid),
3520 (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3521 btrfs_set_header_nritems(split, c_nritems - mid);
3522 btrfs_set_header_nritems(c, mid);
3523 ret = 0;
3524
3525 btrfs_mark_buffer_dirty(c);
3526 btrfs_mark_buffer_dirty(split);
3527
3528 insert_ptr(trans, fs_info, path, &disk_key, split->start,
3529 path->slots[level + 1] + 1, level + 1);
3530
3531 if (path->slots[level] >= mid) {
3532 path->slots[level] -= mid;
3533 btrfs_tree_unlock(c);
3534 free_extent_buffer(c);
3535 path->nodes[level] = split;
3536 path->slots[level + 1] += 1;
3537 } else {
3538 btrfs_tree_unlock(split);
3539 free_extent_buffer(split);
3540 }
3541 return ret;
3542}
3543
3544/*
3545 * how many bytes are required to store the items in a leaf. start
3546 * and nr indicate which items in the leaf to check. This totals up the
3547 * space used both by the item structs and the item data
3548 */
3549static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3550{
3551 struct btrfs_item *start_item;
3552 struct btrfs_item *end_item;
3553 struct btrfs_map_token token;
3554 int data_len;
3555 int nritems = btrfs_header_nritems(l);
3556 int end = min(nritems, start + nr) - 1;
3557
3558 if (!nr)
3559 return 0;
3560 btrfs_init_map_token(&token);
3561 start_item = btrfs_item_nr(start);
3562 end_item = btrfs_item_nr(end);
3563 data_len = btrfs_token_item_offset(l, start_item, &token) +
3564 btrfs_token_item_size(l, start_item, &token);
3565 data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3566 data_len += sizeof(struct btrfs_item) * nr;
3567 WARN_ON(data_len < 0);
3568 return data_len;
3569}
3570
3571/*
3572 * The space between the end of the leaf items and
3573 * the start of the leaf data. IOW, how much room
3574 * the leaf has left for both items and data
3575 */
3576noinline int btrfs_leaf_free_space(struct btrfs_fs_info *fs_info,
3577 struct extent_buffer *leaf)
3578{
3579 int nritems = btrfs_header_nritems(leaf);
3580 int ret;
3581
3582 ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3583 if (ret < 0) {
3584 btrfs_crit(fs_info,
3585 "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3586 ret,
3587 (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3588 leaf_space_used(leaf, 0, nritems), nritems);
3589 }
3590 return ret;
3591}
3592
3593/*
3594 * min slot controls the lowest index we're willing to push to the
3595 * right. We'll push up to and including min_slot, but no lower
3596 */
3597static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
3598 struct btrfs_fs_info *fs_info,
3599 struct btrfs_path *path,
3600 int data_size, int empty,
3601 struct extent_buffer *right,
3602 int free_space, u32 left_nritems,
3603 u32 min_slot)
3604{
3605 struct extent_buffer *left = path->nodes[0];
3606 struct extent_buffer *upper = path->nodes[1];
3607 struct btrfs_map_token token;
3608 struct btrfs_disk_key disk_key;
3609 int slot;
3610 u32 i;
3611 int push_space = 0;
3612 int push_items = 0;
3613 struct btrfs_item *item;
3614 u32 nr;
3615 u32 right_nritems;
3616 u32 data_end;
3617 u32 this_item_size;
3618
3619 btrfs_init_map_token(&token);
3620
3621 if (empty)
3622 nr = 0;
3623 else
3624 nr = max_t(u32, 1, min_slot);
3625
3626 if (path->slots[0] >= left_nritems)
3627 push_space += data_size;
3628
3629 slot = path->slots[1];
3630 i = left_nritems - 1;
3631 while (i >= nr) {
3632 item = btrfs_item_nr(i);
3633
3634 if (!empty && push_items > 0) {
3635 if (path->slots[0] > i)
3636 break;
3637 if (path->slots[0] == i) {
3638 int space = btrfs_leaf_free_space(fs_info, left);
3639 if (space + push_space * 2 > free_space)
3640 break;
3641 }
3642 }
3643
3644 if (path->slots[0] == i)
3645 push_space += data_size;
3646
3647 this_item_size = btrfs_item_size(left, item);
3648 if (this_item_size + sizeof(*item) + push_space > free_space)
3649 break;
3650
3651 push_items++;
3652 push_space += this_item_size + sizeof(*item);
3653 if (i == 0)
3654 break;
3655 i--;
3656 }
3657
3658 if (push_items == 0)
3659 goto out_unlock;
3660
3661 WARN_ON(!empty && push_items == left_nritems);
3662
3663 /* push left to right */
3664 right_nritems = btrfs_header_nritems(right);
3665
3666 push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3667 push_space -= leaf_data_end(fs_info, left);
3668
3669 /* make room in the right data area */
3670 data_end = leaf_data_end(fs_info, right);
3671 memmove_extent_buffer(right,
3672 btrfs_leaf_data(right) + data_end - push_space,
3673 btrfs_leaf_data(right) + data_end,
3674 BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3675
3676 /* copy from the left data area */
3677 copy_extent_buffer(right, left, btrfs_leaf_data(right) +
3678 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3679 btrfs_leaf_data(left) + leaf_data_end(fs_info, left),
3680 push_space);
3681
3682 memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3683 btrfs_item_nr_offset(0),
3684 right_nritems * sizeof(struct btrfs_item));
3685
3686 /* copy the items from left to right */
3687 copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3688 btrfs_item_nr_offset(left_nritems - push_items),
3689 push_items * sizeof(struct btrfs_item));
3690
3691 /* update the item pointers */
3692 right_nritems += push_items;
3693 btrfs_set_header_nritems(right, right_nritems);
3694 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3695 for (i = 0; i < right_nritems; i++) {
3696 item = btrfs_item_nr(i);
3697 push_space -= btrfs_token_item_size(right, item, &token);
3698 btrfs_set_token_item_offset(right, item, push_space, &token);
3699 }
3700
3701 left_nritems -= push_items;
3702 btrfs_set_header_nritems(left, left_nritems);
3703
3704 if (left_nritems)
3705 btrfs_mark_buffer_dirty(left);
3706 else
3707 clean_tree_block(trans, fs_info, left);
3708
3709 btrfs_mark_buffer_dirty(right);
3710
3711 btrfs_item_key(right, &disk_key, 0);
3712 btrfs_set_node_key(upper, &disk_key, slot + 1);
3713 btrfs_mark_buffer_dirty(upper);
3714
3715 /* then fixup the leaf pointer in the path */
3716 if (path->slots[0] >= left_nritems) {
3717 path->slots[0] -= left_nritems;
3718 if (btrfs_header_nritems(path->nodes[0]) == 0)
3719 clean_tree_block(trans, fs_info, path->nodes[0]);
3720 btrfs_tree_unlock(path->nodes[0]);
3721 free_extent_buffer(path->nodes[0]);
3722 path->nodes[0] = right;
3723 path->slots[1] += 1;
3724 } else {
3725 btrfs_tree_unlock(right);
3726 free_extent_buffer(right);
3727 }
3728 return 0;
3729
3730out_unlock:
3731 btrfs_tree_unlock(right);
3732 free_extent_buffer(right);
3733 return 1;
3734}
3735
3736/*
3737 * push some data in the path leaf to the right, trying to free up at
3738 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3739 *
3740 * returns 1 if the push failed because the other node didn't have enough
3741 * room, 0 if everything worked out and < 0 if there were major errors.
3742 *
3743 * this will push starting from min_slot to the end of the leaf. It won't
3744 * push any slot lower than min_slot
3745 */
3746static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3747 *root, struct btrfs_path *path,
3748 int min_data_size, int data_size,
3749 int empty, u32 min_slot)
3750{
3751 struct btrfs_fs_info *fs_info = root->fs_info;
3752 struct extent_buffer *left = path->nodes[0];
3753 struct extent_buffer *right;
3754 struct extent_buffer *upper;
3755 int slot;
3756 int free_space;
3757 u32 left_nritems;
3758 int ret;
3759
3760 if (!path->nodes[1])
3761 return 1;
3762
3763 slot = path->slots[1];
3764 upper = path->nodes[1];
3765 if (slot >= btrfs_header_nritems(upper) - 1)
3766 return 1;
3767
3768 btrfs_assert_tree_locked(path->nodes[1]);
3769
3770 right = read_node_slot(fs_info, upper, slot + 1);
3771 /*
3772 * slot + 1 is not valid or we fail to read the right node,
3773 * no big deal, just return.
3774 */
3775 if (IS_ERR(right))
3776 return 1;
3777
3778 btrfs_tree_lock(right);
3779 btrfs_set_lock_blocking(right);
3780
3781 free_space = btrfs_leaf_free_space(fs_info, right);
3782 if (free_space < data_size)
3783 goto out_unlock;
3784
3785 /* cow and double check */
3786 ret = btrfs_cow_block(trans, root, right, upper,
3787 slot + 1, &right);
3788 if (ret)
3789 goto out_unlock;
3790
3791 free_space = btrfs_leaf_free_space(fs_info, right);
3792 if (free_space < data_size)
3793 goto out_unlock;
3794
3795 left_nritems = btrfs_header_nritems(left);
3796 if (left_nritems == 0)
3797 goto out_unlock;
3798
3799 if (path->slots[0] == left_nritems && !empty) {
3800 /* Key greater than all keys in the leaf, right neighbor has
3801 * enough room for it and we're not emptying our leaf to delete
3802 * it, therefore use right neighbor to insert the new item and
3803 * no need to touch/dirty our left leaft. */
3804 btrfs_tree_unlock(left);
3805 free_extent_buffer(left);
3806 path->nodes[0] = right;
3807 path->slots[0] = 0;
3808 path->slots[1]++;
3809 return 0;
3810 }
3811
3812 return __push_leaf_right(trans, fs_info, path, min_data_size, empty,
3813 right, free_space, left_nritems, min_slot);
3814out_unlock:
3815 btrfs_tree_unlock(right);
3816 free_extent_buffer(right);
3817 return 1;
3818}
3819
3820/*
3821 * push some data in the path leaf to the left, trying to free up at
3822 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3823 *
3824 * max_slot can put a limit on how far into the leaf we'll push items. The
3825 * item at 'max_slot' won't be touched. Use (u32)-1 to make us do all the
3826 * items
3827 */
3828static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
3829 struct btrfs_fs_info *fs_info,
3830 struct btrfs_path *path, int data_size,
3831 int empty, struct extent_buffer *left,
3832 int free_space, u32 right_nritems,
3833 u32 max_slot)
3834{
3835 struct btrfs_disk_key disk_key;
3836 struct extent_buffer *right = path->nodes[0];
3837 int i;
3838 int push_space = 0;
3839 int push_items = 0;
3840 struct btrfs_item *item;
3841 u32 old_left_nritems;
3842 u32 nr;
3843 int ret = 0;
3844 u32 this_item_size;
3845 u32 old_left_item_size;
3846 struct btrfs_map_token token;
3847
3848 btrfs_init_map_token(&token);
3849
3850 if (empty)
3851 nr = min(right_nritems, max_slot);
3852 else
3853 nr = min(right_nritems - 1, max_slot);
3854
3855 for (i = 0; i < nr; i++) {
3856 item = btrfs_item_nr(i);
3857
3858 if (!empty && push_items > 0) {
3859 if (path->slots[0] < i)
3860 break;
3861 if (path->slots[0] == i) {
3862 int space = btrfs_leaf_free_space(fs_info, right);
3863 if (space + push_space * 2 > free_space)
3864 break;
3865 }
3866 }
3867
3868 if (path->slots[0] == i)
3869 push_space += data_size;
3870
3871 this_item_size = btrfs_item_size(right, item);
3872 if (this_item_size + sizeof(*item) + push_space > free_space)
3873 break;
3874
3875 push_items++;
3876 push_space += this_item_size + sizeof(*item);
3877 }
3878
3879 if (push_items == 0) {
3880 ret = 1;
3881 goto out;
3882 }
3883 WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3884
3885 /* push data from right to left */
3886 copy_extent_buffer(left, right,
3887 btrfs_item_nr_offset(btrfs_header_nritems(left)),
3888 btrfs_item_nr_offset(0),
3889 push_items * sizeof(struct btrfs_item));
3890
3891 push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3892 btrfs_item_offset_nr(right, push_items - 1);
3893
3894 copy_extent_buffer(left, right, btrfs_leaf_data(left) +
3895 leaf_data_end(fs_info, left) - push_space,
3896 btrfs_leaf_data(right) +
3897 btrfs_item_offset_nr(right, push_items - 1),
3898 push_space);
3899 old_left_nritems = btrfs_header_nritems(left);
3900 BUG_ON(old_left_nritems <= 0);
3901
3902 old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3903 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3904 u32 ioff;
3905
3906 item = btrfs_item_nr(i);
3907
3908 ioff = btrfs_token_item_offset(left, item, &token);
3909 btrfs_set_token_item_offset(left, item,
3910 ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size),
3911 &token);
3912 }
3913 btrfs_set_header_nritems(left, old_left_nritems + push_items);
3914
3915 /* fixup right node */
3916 if (push_items > right_nritems)
3917 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3918 right_nritems);
3919
3920 if (push_items < right_nritems) {
3921 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3922 leaf_data_end(fs_info, right);
3923 memmove_extent_buffer(right, btrfs_leaf_data(right) +
3924 BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3925 btrfs_leaf_data(right) +
3926 leaf_data_end(fs_info, right), push_space);
3927
3928 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3929 btrfs_item_nr_offset(push_items),
3930 (btrfs_header_nritems(right) - push_items) *
3931 sizeof(struct btrfs_item));
3932 }
3933 right_nritems -= push_items;
3934 btrfs_set_header_nritems(right, right_nritems);
3935 push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3936 for (i = 0; i < right_nritems; i++) {
3937 item = btrfs_item_nr(i);
3938
3939 push_space = push_space - btrfs_token_item_size(right,
3940 item, &token);
3941 btrfs_set_token_item_offset(right, item, push_space, &token);
3942 }
3943
3944 btrfs_mark_buffer_dirty(left);
3945 if (right_nritems)
3946 btrfs_mark_buffer_dirty(right);
3947 else
3948 clean_tree_block(trans, fs_info, right);
3949
3950 btrfs_item_key(right, &disk_key, 0);
3951 fixup_low_keys(fs_info, path, &disk_key, 1);
3952
3953 /* then fixup the leaf pointer in the path */
3954 if (path->slots[0] < push_items) {
3955 path->slots[0] += old_left_nritems;
3956 btrfs_tree_unlock(path->nodes[0]);
3957 free_extent_buffer(path->nodes[0]);
3958 path->nodes[0] = left;
3959 path->slots[1] -= 1;
3960 } else {
3961 btrfs_tree_unlock(left);
3962 free_extent_buffer(left);
3963 path->slots[0] -= push_items;
3964 }
3965 BUG_ON(path->slots[0] < 0);
3966 return ret;
3967out:
3968 btrfs_tree_unlock(left);
3969 free_extent_buffer(left);
3970 return ret;
3971}
3972
3973/*
3974 * push some data in the path leaf to the left, trying to free up at
3975 * least data_size bytes. returns zero if the push worked, nonzero otherwise
3976 *
3977 * max_slot can put a limit on how far into the leaf we'll push items. The
3978 * item at 'max_slot' won't be touched. Use (u32)-1 to make us push all the
3979 * items
3980 */
3981static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3982 *root, struct btrfs_path *path, int min_data_size,
3983 int data_size, int empty, u32 max_slot)
3984{
3985 struct btrfs_fs_info *fs_info = root->fs_info;
3986 struct extent_buffer *right = path->nodes[0];
3987 struct extent_buffer *left;
3988 int slot;
3989 int free_space;
3990 u32 right_nritems;
3991 int ret = 0;
3992
3993 slot = path->slots[1];
3994 if (slot == 0)
3995 return 1;
3996 if (!path->nodes[1])
3997 return 1;
3998
3999 right_nritems = btrfs_header_nritems(right);
4000 if (right_nritems == 0)
4001 return 1;
4002
4003 btrfs_assert_tree_locked(path->nodes[1]);
4004
4005 left = read_node_slot(fs_info, path->nodes[1], slot - 1);
4006 /*
4007 * slot - 1 is not valid or we fail to read the left node,
4008 * no big deal, just return.
4009 */
4010 if (IS_ERR(left))
4011 return 1;
4012
4013 btrfs_tree_lock(left);
4014 btrfs_set_lock_blocking(left);
4015
4016 free_space = btrfs_leaf_free_space(fs_info, left);
4017 if (free_space < data_size) {
4018 ret = 1;
4019 goto out;
4020 }
4021
4022 /* cow and double check */
4023 ret = btrfs_cow_block(trans, root, left,
4024 path->nodes[1], slot - 1, &left);
4025 if (ret) {
4026 /* we hit -ENOSPC, but it isn't fatal here */
4027 if (ret == -ENOSPC)
4028 ret = 1;
4029 goto out;
4030 }
4031
4032 free_space = btrfs_leaf_free_space(fs_info, left);
4033 if (free_space < data_size) {
4034 ret = 1;
4035 goto out;
4036 }
4037
4038 return __push_leaf_left(trans, fs_info, path, min_data_size,
4039 empty, left, free_space, right_nritems,
4040 max_slot);
4041out:
4042 btrfs_tree_unlock(left);
4043 free_extent_buffer(left);
4044 return ret;
4045}
4046
4047/*
4048 * split the path's leaf in two, making sure there is at least data_size
4049 * available for the resulting leaf level of the path.
4050 */
4051static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4052 struct btrfs_fs_info *fs_info,
4053 struct btrfs_path *path,
4054 struct extent_buffer *l,
4055 struct extent_buffer *right,
4056 int slot, int mid, int nritems)
4057{
4058 int data_copy_size;
4059 int rt_data_off;
4060 int i;
4061 struct btrfs_disk_key disk_key;
4062 struct btrfs_map_token token;
4063
4064 btrfs_init_map_token(&token);
4065
4066 nritems = nritems - mid;
4067 btrfs_set_header_nritems(right, nritems);
4068 data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(fs_info, l);
4069
4070 copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4071 btrfs_item_nr_offset(mid),
4072 nritems * sizeof(struct btrfs_item));
4073
4074 copy_extent_buffer(right, l,
4075 btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(fs_info) -
4076 data_copy_size, btrfs_leaf_data(l) +
4077 leaf_data_end(fs_info, l), data_copy_size);
4078
4079 rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4080
4081 for (i = 0; i < nritems; i++) {
4082 struct btrfs_item *item = btrfs_item_nr(i);
4083 u32 ioff;
4084
4085 ioff = btrfs_token_item_offset(right, item, &token);
4086 btrfs_set_token_item_offset(right, item,
4087 ioff + rt_data_off, &token);
4088 }
4089
4090 btrfs_set_header_nritems(l, mid);
4091 btrfs_item_key(right, &disk_key, 0);
4092 insert_ptr(trans, fs_info, path, &disk_key, right->start,
4093 path->slots[1] + 1, 1);
4094
4095 btrfs_mark_buffer_dirty(right);
4096 btrfs_mark_buffer_dirty(l);
4097 BUG_ON(path->slots[0] != slot);
4098
4099 if (mid <= slot) {
4100 btrfs_tree_unlock(path->nodes[0]);
4101 free_extent_buffer(path->nodes[0]);
4102 path->nodes[0] = right;
4103 path->slots[0] -= mid;
4104 path->slots[1] += 1;
4105 } else {
4106 btrfs_tree_unlock(right);
4107 free_extent_buffer(right);
4108 }
4109
4110 BUG_ON(path->slots[0] < 0);
4111}
4112
4113/*
4114 * double splits happen when we need to insert a big item in the middle
4115 * of a leaf. A double split can leave us with 3 mostly empty leaves:
4116 * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4117 * A B C
4118 *
4119 * We avoid this by trying to push the items on either side of our target
4120 * into the adjacent leaves. If all goes well we can avoid the double split
4121 * completely.
4122 */
4123static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4124 struct btrfs_root *root,
4125 struct btrfs_path *path,
4126 int data_size)
4127{
4128 struct btrfs_fs_info *fs_info = root->fs_info;
4129 int ret;
4130 int progress = 0;
4131 int slot;
4132 u32 nritems;
4133 int space_needed = data_size;
4134
4135 slot = path->slots[0];
4136 if (slot < btrfs_header_nritems(path->nodes[0]))
4137 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4138
4139 /*
4140 * try to push all the items after our slot into the
4141 * right leaf
4142 */
4143 ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4144 if (ret < 0)
4145 return ret;
4146
4147 if (ret == 0)
4148 progress++;
4149
4150 nritems = btrfs_header_nritems(path->nodes[0]);
4151 /*
4152 * our goal is to get our slot at the start or end of a leaf. If
4153 * we've done so we're done
4154 */
4155 if (path->slots[0] == 0 || path->slots[0] == nritems)
4156 return 0;
4157
4158 if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4159 return 0;
4160
4161 /* try to push all the items before our slot into the next leaf */
4162 slot = path->slots[0];
4163 ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4164 if (ret < 0)
4165 return ret;
4166
4167 if (ret == 0)
4168 progress++;
4169
4170 if (progress)
4171 return 0;
4172 return 1;
4173}
4174
4175/*
4176 * split the path's leaf in two, making sure there is at least data_size
4177 * available for the resulting leaf level of the path.
4178 *
4179 * returns 0 if all went well and < 0 on failure.
4180 */
4181static noinline int split_leaf(struct btrfs_trans_handle *trans,
4182 struct btrfs_root *root,
4183 struct btrfs_key *ins_key,
4184 struct btrfs_path *path, int data_size,
4185 int extend)
4186{
4187 struct btrfs_disk_key disk_key;
4188 struct extent_buffer *l;
4189 u32 nritems;
4190 int mid;
4191 int slot;
4192 struct extent_buffer *right;
4193 struct btrfs_fs_info *fs_info = root->fs_info;
4194 int ret = 0;
4195 int wret;
4196 int split;
4197 int num_doubles = 0;
4198 int tried_avoid_double = 0;
4199
4200 l = path->nodes[0];
4201 slot = path->slots[0];
4202 if (extend && data_size + btrfs_item_size_nr(l, slot) +
4203 sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4204 return -EOVERFLOW;
4205
4206 /* first try to make some room by pushing left and right */
4207 if (data_size && path->nodes[1]) {
4208 int space_needed = data_size;
4209
4210 if (slot < btrfs_header_nritems(l))
4211 space_needed -= btrfs_leaf_free_space(fs_info, l);
4212
4213 wret = push_leaf_right(trans, root, path, space_needed,
4214 space_needed, 0, 0);
4215 if (wret < 0)
4216 return wret;
4217 if (wret) {
4218 wret = push_leaf_left(trans, root, path, space_needed,
4219 space_needed, 0, (u32)-1);
4220 if (wret < 0)
4221 return wret;
4222 }
4223 l = path->nodes[0];
4224
4225 /* did the pushes work? */
4226 if (btrfs_leaf_free_space(fs_info, l) >= data_size)
4227 return 0;
4228 }
4229
4230 if (!path->nodes[1]) {
4231 ret = insert_new_root(trans, root, path, 1);
4232 if (ret)
4233 return ret;
4234 }
4235again:
4236 split = 1;
4237 l = path->nodes[0];
4238 slot = path->slots[0];
4239 nritems = btrfs_header_nritems(l);
4240 mid = (nritems + 1) / 2;
4241
4242 if (mid <= slot) {
4243 if (nritems == 1 ||
4244 leaf_space_used(l, mid, nritems - mid) + data_size >
4245 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4246 if (slot >= nritems) {
4247 split = 0;
4248 } else {
4249 mid = slot;
4250 if (mid != nritems &&
4251 leaf_space_used(l, mid, nritems - mid) +
4252 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4253 if (data_size && !tried_avoid_double)
4254 goto push_for_double;
4255 split = 2;
4256 }
4257 }
4258 }
4259 } else {
4260 if (leaf_space_used(l, 0, mid) + data_size >
4261 BTRFS_LEAF_DATA_SIZE(fs_info)) {
4262 if (!extend && data_size && slot == 0) {
4263 split = 0;
4264 } else if ((extend || !data_size) && slot == 0) {
4265 mid = 1;
4266 } else {
4267 mid = slot;
4268 if (mid != nritems &&
4269 leaf_space_used(l, mid, nritems - mid) +
4270 data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4271 if (data_size && !tried_avoid_double)
4272 goto push_for_double;
4273 split = 2;
4274 }
4275 }
4276 }
4277 }
4278
4279 if (split == 0)
4280 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4281 else
4282 btrfs_item_key(l, &disk_key, mid);
4283
4284 right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
4285 &disk_key, 0, l->start, 0);
4286 if (IS_ERR(right))
4287 return PTR_ERR(right);
4288
4289 root_add_used(root, fs_info->nodesize);
4290
4291 memzero_extent_buffer(right, 0, sizeof(struct btrfs_header));
4292 btrfs_set_header_bytenr(right, right->start);
4293 btrfs_set_header_generation(right, trans->transid);
4294 btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
4295 btrfs_set_header_owner(right, root->root_key.objectid);
4296 btrfs_set_header_level(right, 0);
4297 write_extent_buffer_fsid(right, fs_info->fsid);
4298 write_extent_buffer_chunk_tree_uuid(right, fs_info->chunk_tree_uuid);
4299
4300 if (split == 0) {
4301 if (mid <= slot) {
4302 btrfs_set_header_nritems(right, 0);
4303 insert_ptr(trans, fs_info, path, &disk_key,
4304 right->start, path->slots[1] + 1, 1);
4305 btrfs_tree_unlock(path->nodes[0]);
4306 free_extent_buffer(path->nodes[0]);
4307 path->nodes[0] = right;
4308 path->slots[0] = 0;
4309 path->slots[1] += 1;
4310 } else {
4311 btrfs_set_header_nritems(right, 0);
4312 insert_ptr(trans, fs_info, path, &disk_key,
4313 right->start, path->slots[1], 1);
4314 btrfs_tree_unlock(path->nodes[0]);
4315 free_extent_buffer(path->nodes[0]);
4316 path->nodes[0] = right;
4317 path->slots[0] = 0;
4318 if (path->slots[1] == 0)
4319 fixup_low_keys(fs_info, path, &disk_key, 1);
4320 }
4321 /*
4322 * We create a new leaf 'right' for the required ins_len and
4323 * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4324 * the content of ins_len to 'right'.
4325 */
4326 return ret;
4327 }
4328
4329 copy_for_split(trans, fs_info, path, l, right, slot, mid, nritems);
4330
4331 if (split == 2) {
4332 BUG_ON(num_doubles != 0);
4333 num_doubles++;
4334 goto again;
4335 }
4336
4337 return 0;
4338
4339push_for_double:
4340 push_for_double_split(trans, root, path, data_size);
4341 tried_avoid_double = 1;
4342 if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4343 return 0;
4344 goto again;
4345}
4346
4347static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4348 struct btrfs_root *root,
4349 struct btrfs_path *path, int ins_len)
4350{
4351 struct btrfs_fs_info *fs_info = root->fs_info;
4352 struct btrfs_key key;
4353 struct extent_buffer *leaf;
4354 struct btrfs_file_extent_item *fi;
4355 u64 extent_len = 0;
4356 u32 item_size;
4357 int ret;
4358
4359 leaf = path->nodes[0];
4360 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4361
4362 BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4363 key.type != BTRFS_EXTENT_CSUM_KEY);
4364
4365 if (btrfs_leaf_free_space(fs_info, leaf) >= ins_len)
4366 return 0;
4367
4368 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4369 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4370 fi = btrfs_item_ptr(leaf, path->slots[0],
4371 struct btrfs_file_extent_item);
4372 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4373 }
4374 btrfs_release_path(path);
4375
4376 path->keep_locks = 1;
4377 path->search_for_split = 1;
4378 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4379 path->search_for_split = 0;
4380 if (ret > 0)
4381 ret = -EAGAIN;
4382 if (ret < 0)
4383 goto err;
4384
4385 ret = -EAGAIN;
4386 leaf = path->nodes[0];
4387 /* if our item isn't there, return now */
4388 if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4389 goto err;
4390
4391 /* the leaf has changed, it now has room. return now */
4392 if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= ins_len)
4393 goto err;
4394
4395 if (key.type == BTRFS_EXTENT_DATA_KEY) {
4396 fi = btrfs_item_ptr(leaf, path->slots[0],
4397 struct btrfs_file_extent_item);
4398 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4399 goto err;
4400 }
4401
4402 btrfs_set_path_blocking(path);
4403 ret = split_leaf(trans, root, &key, path, ins_len, 1);
4404 if (ret)
4405 goto err;
4406
4407 path->keep_locks = 0;
4408 btrfs_unlock_up_safe(path, 1);
4409 return 0;
4410err:
4411 path->keep_locks = 0;
4412 return ret;
4413}
4414
4415static noinline int split_item(struct btrfs_trans_handle *trans,
4416 struct btrfs_fs_info *fs_info,
4417 struct btrfs_path *path,
4418 struct btrfs_key *new_key,
4419 unsigned long split_offset)
4420{
4421 struct extent_buffer *leaf;
4422 struct btrfs_item *item;
4423 struct btrfs_item *new_item;
4424 int slot;
4425 char *buf;
4426 u32 nritems;
4427 u32 item_size;
4428 u32 orig_offset;
4429 struct btrfs_disk_key disk_key;
4430
4431 leaf = path->nodes[0];
4432 BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < sizeof(struct btrfs_item));
4433
4434 btrfs_set_path_blocking(path);
4435
4436 item = btrfs_item_nr(path->slots[0]);
4437 orig_offset = btrfs_item_offset(leaf, item);
4438 item_size = btrfs_item_size(leaf, item);
4439
4440 buf = kmalloc(item_size, GFP_NOFS);
4441 if (!buf)
4442 return -ENOMEM;
4443
4444 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4445 path->slots[0]), item_size);
4446
4447 slot = path->slots[0] + 1;
4448 nritems = btrfs_header_nritems(leaf);
4449 if (slot != nritems) {
4450 /* shift the items */
4451 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4452 btrfs_item_nr_offset(slot),
4453 (nritems - slot) * sizeof(struct btrfs_item));
4454 }
4455
4456 btrfs_cpu_key_to_disk(&disk_key, new_key);
4457 btrfs_set_item_key(leaf, &disk_key, slot);
4458
4459 new_item = btrfs_item_nr(slot);
4460
4461 btrfs_set_item_offset(leaf, new_item, orig_offset);
4462 btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4463
4464 btrfs_set_item_offset(leaf, item,
4465 orig_offset + item_size - split_offset);
4466 btrfs_set_item_size(leaf, item, split_offset);
4467
4468 btrfs_set_header_nritems(leaf, nritems + 1);
4469
4470 /* write the data for the start of the original item */
4471 write_extent_buffer(leaf, buf,
4472 btrfs_item_ptr_offset(leaf, path->slots[0]),
4473 split_offset);
4474
4475 /* write the data for the new item */
4476 write_extent_buffer(leaf, buf + split_offset,
4477 btrfs_item_ptr_offset(leaf, slot),
4478 item_size - split_offset);
4479 btrfs_mark_buffer_dirty(leaf);
4480
4481 BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < 0);
4482 kfree(buf);
4483 return 0;
4484}
4485
4486/*
4487 * This function splits a single item into two items,
4488 * giving 'new_key' to the new item and splitting the
4489 * old one at split_offset (from the start of the item).
4490 *
4491 * The path may be released by this operation. After
4492 * the split, the path is pointing to the old item. The
4493 * new item is going to be in the same node as the old one.
4494 *
4495 * Note, the item being split must be smaller enough to live alone on
4496 * a tree block with room for one extra struct btrfs_item
4497 *
4498 * This allows us to split the item in place, keeping a lock on the
4499 * leaf the entire time.
4500 */
4501int btrfs_split_item(struct btrfs_trans_handle *trans,
4502 struct btrfs_root *root,
4503 struct btrfs_path *path,
4504 struct btrfs_key *new_key,
4505 unsigned long split_offset)
4506{
4507 int ret;
4508 ret = setup_leaf_for_split(trans, root, path,
4509 sizeof(struct btrfs_item));
4510 if (ret)
4511 return ret;
4512
4513 ret = split_item(trans, root->fs_info, path, new_key, split_offset);
4514 return ret;
4515}
4516
4517/*
4518 * This function duplicate a item, giving 'new_key' to the new item.
4519 * It guarantees both items live in the same tree leaf and the new item
4520 * is contiguous with the original item.
4521 *
4522 * This allows us to split file extent in place, keeping a lock on the
4523 * leaf the entire time.
4524 */
4525int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4526 struct btrfs_root *root,
4527 struct btrfs_path *path,
4528 struct btrfs_key *new_key)
4529{
4530 struct extent_buffer *leaf;
4531 int ret;
4532 u32 item_size;
4533
4534 leaf = path->nodes[0];
4535 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4536 ret = setup_leaf_for_split(trans, root, path,
4537 item_size + sizeof(struct btrfs_item));
4538 if (ret)
4539 return ret;
4540
4541 path->slots[0]++;
4542 setup_items_for_insert(root, path, new_key, &item_size,
4543 item_size, item_size +
4544 sizeof(struct btrfs_item), 1);
4545 leaf = path->nodes[0];
4546 memcpy_extent_buffer(leaf,
4547 btrfs_item_ptr_offset(leaf, path->slots[0]),
4548 btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4549 item_size);
4550 return 0;
4551}
4552
4553/*
4554 * make the item pointed to by the path smaller. new_size indicates
4555 * how small to make it, and from_end tells us if we just chop bytes
4556 * off the end of the item or if we shift the item to chop bytes off
4557 * the front.
4558 */
4559void btrfs_truncate_item(struct btrfs_fs_info *fs_info,
4560 struct btrfs_path *path, u32 new_size, int from_end)
4561{
4562 int slot;
4563 struct extent_buffer *leaf;
4564 struct btrfs_item *item;
4565 u32 nritems;
4566 unsigned int data_end;
4567 unsigned int old_data_start;
4568 unsigned int old_size;
4569 unsigned int size_diff;
4570 int i;
4571 struct btrfs_map_token token;
4572
4573 btrfs_init_map_token(&token);
4574
4575 leaf = path->nodes[0];
4576 slot = path->slots[0];
4577
4578 old_size = btrfs_item_size_nr(leaf, slot);
4579 if (old_size == new_size)
4580 return;
4581
4582 nritems = btrfs_header_nritems(leaf);
4583 data_end = leaf_data_end(fs_info, leaf);
4584
4585 old_data_start = btrfs_item_offset_nr(leaf, slot);
4586
4587 size_diff = old_size - new_size;
4588
4589 BUG_ON(slot < 0);
4590 BUG_ON(slot >= nritems);
4591
4592 /*
4593 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4594 */
4595 /* first correct the data pointers */
4596 for (i = slot; i < nritems; i++) {
4597 u32 ioff;
4598 item = btrfs_item_nr(i);
4599
4600 ioff = btrfs_token_item_offset(leaf, item, &token);
4601 btrfs_set_token_item_offset(leaf, item,
4602 ioff + size_diff, &token);
4603 }
4604
4605 /* shift the data */
4606 if (from_end) {
4607 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4608 data_end + size_diff, btrfs_leaf_data(leaf) +
4609 data_end, old_data_start + new_size - data_end);
4610 } else {
4611 struct btrfs_disk_key disk_key;
4612 u64 offset;
4613
4614 btrfs_item_key(leaf, &disk_key, slot);
4615
4616 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4617 unsigned long ptr;
4618 struct btrfs_file_extent_item *fi;
4619
4620 fi = btrfs_item_ptr(leaf, slot,
4621 struct btrfs_file_extent_item);
4622 fi = (struct btrfs_file_extent_item *)(
4623 (unsigned long)fi - size_diff);
4624
4625 if (btrfs_file_extent_type(leaf, fi) ==
4626 BTRFS_FILE_EXTENT_INLINE) {
4627 ptr = btrfs_item_ptr_offset(leaf, slot);
4628 memmove_extent_buffer(leaf, ptr,
4629 (unsigned long)fi,
4630 BTRFS_FILE_EXTENT_INLINE_DATA_START);
4631 }
4632 }
4633
4634 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4635 data_end + size_diff, btrfs_leaf_data(leaf) +
4636 data_end, old_data_start - data_end);
4637
4638 offset = btrfs_disk_key_offset(&disk_key);
4639 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4640 btrfs_set_item_key(leaf, &disk_key, slot);
4641 if (slot == 0)
4642 fixup_low_keys(fs_info, path, &disk_key, 1);
4643 }
4644
4645 item = btrfs_item_nr(slot);
4646 btrfs_set_item_size(leaf, item, new_size);
4647 btrfs_mark_buffer_dirty(leaf);
4648
4649 if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4650 btrfs_print_leaf(fs_info, leaf);
4651 BUG();
4652 }
4653}
4654
4655/*
4656 * make the item pointed to by the path bigger, data_size is the added size.
4657 */
4658void btrfs_extend_item(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
4659 u32 data_size)
4660{
4661 int slot;
4662 struct extent_buffer *leaf;
4663 struct btrfs_item *item;
4664 u32 nritems;
4665 unsigned int data_end;
4666 unsigned int old_data;
4667 unsigned int old_size;
4668 int i;
4669 struct btrfs_map_token token;
4670
4671 btrfs_init_map_token(&token);
4672
4673 leaf = path->nodes[0];
4674
4675 nritems = btrfs_header_nritems(leaf);
4676 data_end = leaf_data_end(fs_info, leaf);
4677
4678 if (btrfs_leaf_free_space(fs_info, leaf) < data_size) {
4679 btrfs_print_leaf(fs_info, leaf);
4680 BUG();
4681 }
4682 slot = path->slots[0];
4683 old_data = btrfs_item_end_nr(leaf, slot);
4684
4685 BUG_ON(slot < 0);
4686 if (slot >= nritems) {
4687 btrfs_print_leaf(fs_info, leaf);
4688 btrfs_crit(fs_info, "slot %d too large, nritems %d",
4689 slot, nritems);
4690 BUG_ON(1);
4691 }
4692
4693 /*
4694 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4695 */
4696 /* first correct the data pointers */
4697 for (i = slot; i < nritems; i++) {
4698 u32 ioff;
4699 item = btrfs_item_nr(i);
4700
4701 ioff = btrfs_token_item_offset(leaf, item, &token);
4702 btrfs_set_token_item_offset(leaf, item,
4703 ioff - data_size, &token);
4704 }
4705
4706 /* shift the data */
4707 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4708 data_end - data_size, btrfs_leaf_data(leaf) +
4709 data_end, old_data - data_end);
4710
4711 data_end = old_data;
4712 old_size = btrfs_item_size_nr(leaf, slot);
4713 item = btrfs_item_nr(slot);
4714 btrfs_set_item_size(leaf, item, old_size + data_size);
4715 btrfs_mark_buffer_dirty(leaf);
4716
4717 if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4718 btrfs_print_leaf(fs_info, leaf);
4719 BUG();
4720 }
4721}
4722
4723/*
4724 * this is a helper for btrfs_insert_empty_items, the main goal here is
4725 * to save stack depth by doing the bulk of the work in a function
4726 * that doesn't call btrfs_search_slot
4727 */
4728void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4729 struct btrfs_key *cpu_key, u32 *data_size,
4730 u32 total_data, u32 total_size, int nr)
4731{
4732 struct btrfs_fs_info *fs_info = root->fs_info;
4733 struct btrfs_item *item;
4734 int i;
4735 u32 nritems;
4736 unsigned int data_end;
4737 struct btrfs_disk_key disk_key;
4738 struct extent_buffer *leaf;
4739 int slot;
4740 struct btrfs_map_token token;
4741
4742 if (path->slots[0] == 0) {
4743 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4744 fixup_low_keys(fs_info, path, &disk_key, 1);
4745 }
4746 btrfs_unlock_up_safe(path, 1);
4747
4748 btrfs_init_map_token(&token);
4749
4750 leaf = path->nodes[0];
4751 slot = path->slots[0];
4752
4753 nritems = btrfs_header_nritems(leaf);
4754 data_end = leaf_data_end(fs_info, leaf);
4755
4756 if (btrfs_leaf_free_space(fs_info, leaf) < total_size) {
4757 btrfs_print_leaf(fs_info, leaf);
4758 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4759 total_size, btrfs_leaf_free_space(fs_info, leaf));
4760 BUG();
4761 }
4762
4763 if (slot != nritems) {
4764 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4765
4766 if (old_data < data_end) {
4767 btrfs_print_leaf(fs_info, leaf);
4768 btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4769 slot, old_data, data_end);
4770 BUG_ON(1);
4771 }
4772 /*
4773 * item0..itemN ... dataN.offset..dataN.size .. data0.size
4774 */
4775 /* first correct the data pointers */
4776 for (i = slot; i < nritems; i++) {
4777 u32 ioff;
4778
4779 item = btrfs_item_nr(i);
4780 ioff = btrfs_token_item_offset(leaf, item, &token);
4781 btrfs_set_token_item_offset(leaf, item,
4782 ioff - total_data, &token);
4783 }
4784 /* shift the items */
4785 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4786 btrfs_item_nr_offset(slot),
4787 (nritems - slot) * sizeof(struct btrfs_item));
4788
4789 /* shift the data */
4790 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4791 data_end - total_data, btrfs_leaf_data(leaf) +
4792 data_end, old_data - data_end);
4793 data_end = old_data;
4794 }
4795
4796 /* setup the item for the new data */
4797 for (i = 0; i < nr; i++) {
4798 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4799 btrfs_set_item_key(leaf, &disk_key, slot + i);
4800 item = btrfs_item_nr(slot + i);
4801 btrfs_set_token_item_offset(leaf, item,
4802 data_end - data_size[i], &token);
4803 data_end -= data_size[i];
4804 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4805 }
4806
4807 btrfs_set_header_nritems(leaf, nritems + nr);
4808 btrfs_mark_buffer_dirty(leaf);
4809
4810 if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4811 btrfs_print_leaf(fs_info, leaf);
4812 BUG();
4813 }
4814}
4815
4816/*
4817 * Given a key and some data, insert items into the tree.
4818 * This does all the path init required, making room in the tree if needed.
4819 */
4820int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4821 struct btrfs_root *root,
4822 struct btrfs_path *path,
4823 struct btrfs_key *cpu_key, u32 *data_size,
4824 int nr)
4825{
4826 int ret = 0;
4827 int slot;
4828 int i;
4829 u32 total_size = 0;
4830 u32 total_data = 0;
4831
4832 for (i = 0; i < nr; i++)
4833 total_data += data_size[i];
4834
4835 total_size = total_data + (nr * sizeof(struct btrfs_item));
4836 ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4837 if (ret == 0)
4838 return -EEXIST;
4839 if (ret < 0)
4840 return ret;
4841
4842 slot = path->slots[0];
4843 BUG_ON(slot < 0);
4844
4845 setup_items_for_insert(root, path, cpu_key, data_size,
4846 total_data, total_size, nr);
4847 return 0;
4848}
4849
4850/*
4851 * Given a key and some data, insert an item into the tree.
4852 * This does all the path init required, making room in the tree if needed.
4853 */
4854int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
4855 *root, struct btrfs_key *cpu_key, void *data, u32
4856 data_size)
4857{
4858 int ret = 0;
4859 struct btrfs_path *path;
4860 struct extent_buffer *leaf;
4861 unsigned long ptr;
4862
4863 path = btrfs_alloc_path();
4864 if (!path)
4865 return -ENOMEM;
4866 ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4867 if (!ret) {
4868 leaf = path->nodes[0];
4869 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4870 write_extent_buffer(leaf, data, ptr, data_size);
4871 btrfs_mark_buffer_dirty(leaf);
4872 }
4873 btrfs_free_path(path);
4874 return ret;
4875}
4876
4877/*
4878 * delete the pointer from a given node.
4879 *
4880 * the tree should have been previously balanced so the deletion does not
4881 * empty a node.
4882 */
4883static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4884 int level, int slot)
4885{
4886 struct btrfs_fs_info *fs_info = root->fs_info;
4887 struct extent_buffer *parent = path->nodes[level];
4888 u32 nritems;
4889 int ret;
4890
4891 nritems = btrfs_header_nritems(parent);
4892 if (slot != nritems - 1) {
4893 if (level)
4894 tree_mod_log_eb_move(fs_info, parent, slot,
4895 slot + 1, nritems - slot - 1);
4896 memmove_extent_buffer(parent,
4897 btrfs_node_key_ptr_offset(slot),
4898 btrfs_node_key_ptr_offset(slot + 1),
4899 sizeof(struct btrfs_key_ptr) *
4900 (nritems - slot - 1));
4901 } else if (level) {
4902 ret = tree_mod_log_insert_key(fs_info, parent, slot,
4903 MOD_LOG_KEY_REMOVE, GFP_NOFS);
4904 BUG_ON(ret < 0);
4905 }
4906
4907 nritems--;
4908 btrfs_set_header_nritems(parent, nritems);
4909 if (nritems == 0 && parent == root->node) {
4910 BUG_ON(btrfs_header_level(root->node) != 1);
4911 /* just turn the root into a leaf and break */
4912 btrfs_set_header_level(root->node, 0);
4913 } else if (slot == 0) {
4914 struct btrfs_disk_key disk_key;
4915
4916 btrfs_node_key(parent, &disk_key, 0);
4917 fixup_low_keys(fs_info, path, &disk_key, level + 1);
4918 }
4919 btrfs_mark_buffer_dirty(parent);
4920}
4921
4922/*
4923 * a helper function to delete the leaf pointed to by path->slots[1] and
4924 * path->nodes[1].
4925 *
4926 * This deletes the pointer in path->nodes[1] and frees the leaf
4927 * block extent. zero is returned if it all worked out, < 0 otherwise.
4928 *
4929 * The path must have already been setup for deleting the leaf, including
4930 * all the proper balancing. path->nodes[1] must be locked.
4931 */
4932static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4933 struct btrfs_root *root,
4934 struct btrfs_path *path,
4935 struct extent_buffer *leaf)
4936{
4937 WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4938 del_ptr(root, path, 1, path->slots[1]);
4939
4940 /*
4941 * btrfs_free_extent is expensive, we want to make sure we
4942 * aren't holding any locks when we call it
4943 */
4944 btrfs_unlock_up_safe(path, 0);
4945
4946 root_sub_used(root, leaf->len);
4947
4948 extent_buffer_get(leaf);
4949 btrfs_free_tree_block(trans, root, leaf, 0, 1);
4950 free_extent_buffer_stale(leaf);
4951}
4952/*
4953 * delete the item at the leaf level in path. If that empties
4954 * the leaf, remove it from the tree
4955 */
4956int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4957 struct btrfs_path *path, int slot, int nr)
4958{
4959 struct btrfs_fs_info *fs_info = root->fs_info;
4960 struct extent_buffer *leaf;
4961 struct btrfs_item *item;
4962 u32 last_off;
4963 u32 dsize = 0;
4964 int ret = 0;
4965 int wret;
4966 int i;
4967 u32 nritems;
4968 struct btrfs_map_token token;
4969
4970 btrfs_init_map_token(&token);
4971
4972 leaf = path->nodes[0];
4973 last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4974
4975 for (i = 0; i < nr; i++)
4976 dsize += btrfs_item_size_nr(leaf, slot + i);
4977
4978 nritems = btrfs_header_nritems(leaf);
4979
4980 if (slot + nr != nritems) {
4981 int data_end = leaf_data_end(fs_info, leaf);
4982
4983 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4984 data_end + dsize,
4985 btrfs_leaf_data(leaf) + data_end,
4986 last_off - data_end);
4987
4988 for (i = slot + nr; i < nritems; i++) {
4989 u32 ioff;
4990
4991 item = btrfs_item_nr(i);
4992 ioff = btrfs_token_item_offset(leaf, item, &token);
4993 btrfs_set_token_item_offset(leaf, item,
4994 ioff + dsize, &token);
4995 }
4996
4997 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4998 btrfs_item_nr_offset(slot + nr),
4999 sizeof(struct btrfs_item) *
5000 (nritems - slot - nr));
5001 }
5002 btrfs_set_header_nritems(leaf, nritems - nr);
5003 nritems -= nr;
5004
5005 /* delete the leaf if we've emptied it */
5006 if (nritems == 0) {
5007 if (leaf == root->node) {
5008 btrfs_set_header_level(leaf, 0);
5009 } else {
5010 btrfs_set_path_blocking(path);
5011 clean_tree_block(trans, fs_info, leaf);
5012 btrfs_del_leaf(trans, root, path, leaf);
5013 }
5014 } else {
5015 int used = leaf_space_used(leaf, 0, nritems);
5016 if (slot == 0) {
5017 struct btrfs_disk_key disk_key;
5018
5019 btrfs_item_key(leaf, &disk_key, 0);
5020 fixup_low_keys(fs_info, path, &disk_key, 1);
5021 }
5022
5023 /* delete the leaf if it is mostly empty */
5024 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
5025 /* push_leaf_left fixes the path.
5026 * make sure the path still points to our leaf
5027 * for possible call to del_ptr below
5028 */
5029 slot = path->slots[1];
5030 extent_buffer_get(leaf);
5031
5032 btrfs_set_path_blocking(path);
5033 wret = push_leaf_left(trans, root, path, 1, 1,
5034 1, (u32)-1);
5035 if (wret < 0 && wret != -ENOSPC)
5036 ret = wret;
5037
5038 if (path->nodes[0] == leaf &&
5039 btrfs_header_nritems(leaf)) {
5040 wret = push_leaf_right(trans, root, path, 1,
5041 1, 1, 0);
5042 if (wret < 0 && wret != -ENOSPC)
5043 ret = wret;
5044 }
5045
5046 if (btrfs_header_nritems(leaf) == 0) {
5047 path->slots[1] = slot;
5048 btrfs_del_leaf(trans, root, path, leaf);
5049 free_extent_buffer(leaf);
5050 ret = 0;
5051 } else {
5052 /* if we're still in the path, make sure
5053 * we're dirty. Otherwise, one of the
5054 * push_leaf functions must have already
5055 * dirtied this buffer
5056 */
5057 if (path->nodes[0] == leaf)
5058 btrfs_mark_buffer_dirty(leaf);
5059 free_extent_buffer(leaf);
5060 }
5061 } else {
5062 btrfs_mark_buffer_dirty(leaf);
5063 }
5064 }
5065 return ret;
5066}
5067
5068/*
5069 * search the tree again to find a leaf with lesser keys
5070 * returns 0 if it found something or 1 if there are no lesser leaves.
5071 * returns < 0 on io errors.
5072 *
5073 * This may release the path, and so you may lose any locks held at the
5074 * time you call it.
5075 */
5076int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5077{
5078 struct btrfs_key key;
5079 struct btrfs_disk_key found_key;
5080 int ret;
5081
5082 btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5083
5084 if (key.offset > 0) {
5085 key.offset--;
5086 } else if (key.type > 0) {
5087 key.type--;
5088 key.offset = (u64)-1;
5089 } else if (key.objectid > 0) {
5090 key.objectid--;
5091 key.type = (u8)-1;
5092 key.offset = (u64)-1;
5093 } else {
5094 return 1;
5095 }
5096
5097 btrfs_release_path(path);
5098 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5099 if (ret < 0)
5100 return ret;
5101 btrfs_item_key(path->nodes[0], &found_key, 0);
5102 ret = comp_keys(&found_key, &key);
5103 /*
5104 * We might have had an item with the previous key in the tree right
5105 * before we released our path. And after we released our path, that
5106 * item might have been pushed to the first slot (0) of the leaf we
5107 * were holding due to a tree balance. Alternatively, an item with the
5108 * previous key can exist as the only element of a leaf (big fat item).
5109 * Therefore account for these 2 cases, so that our callers (like
5110 * btrfs_previous_item) don't miss an existing item with a key matching
5111 * the previous key we computed above.
5112 */
5113 if (ret <= 0)
5114 return 0;
5115 return 1;
5116}
5117
5118/*
5119 * A helper function to walk down the tree starting at min_key, and looking
5120 * for nodes or leaves that are have a minimum transaction id.
5121 * This is used by the btree defrag code, and tree logging
5122 *
5123 * This does not cow, but it does stuff the starting key it finds back
5124 * into min_key, so you can call btrfs_search_slot with cow=1 on the
5125 * key and get a writable path.
5126 *
5127 * This does lock as it descends, and path->keep_locks should be set
5128 * to 1 by the caller.
5129 *
5130 * This honors path->lowest_level to prevent descent past a given level
5131 * of the tree.
5132 *
5133 * min_trans indicates the oldest transaction that you are interested
5134 * in walking through. Any nodes or leaves older than min_trans are
5135 * skipped over (without reading them).
5136 *
5137 * returns zero if something useful was found, < 0 on error and 1 if there
5138 * was nothing in the tree that matched the search criteria.
5139 */
5140int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5141 struct btrfs_path *path,
5142 u64 min_trans)
5143{
5144 struct btrfs_fs_info *fs_info = root->fs_info;
5145 struct extent_buffer *cur;
5146 struct btrfs_key found_key;
5147 int slot;
5148 int sret;
5149 u32 nritems;
5150 int level;
5151 int ret = 1;
5152 int keep_locks = path->keep_locks;
5153
5154 path->keep_locks = 1;
5155again:
5156 cur = btrfs_read_lock_root_node(root);
5157 level = btrfs_header_level(cur);
5158 WARN_ON(path->nodes[level]);
5159 path->nodes[level] = cur;
5160 path->locks[level] = BTRFS_READ_LOCK;
5161
5162 if (btrfs_header_generation(cur) < min_trans) {
5163 ret = 1;
5164 goto out;
5165 }
5166 while (1) {
5167 nritems = btrfs_header_nritems(cur);
5168 level = btrfs_header_level(cur);
5169 sret = bin_search(cur, min_key, level, &slot);
5170
5171 /* at the lowest level, we're done, setup the path and exit */
5172 if (level == path->lowest_level) {
5173 if (slot >= nritems)
5174 goto find_next_key;
5175 ret = 0;
5176 path->slots[level] = slot;
5177 btrfs_item_key_to_cpu(cur, &found_key, slot);
5178 goto out;
5179 }
5180 if (sret && slot > 0)
5181 slot--;
5182 /*
5183 * check this node pointer against the min_trans parameters.
5184 * If it is too old, old, skip to the next one.
5185 */
5186 while (slot < nritems) {
5187 u64 gen;
5188
5189 gen = btrfs_node_ptr_generation(cur, slot);
5190 if (gen < min_trans) {
5191 slot++;
5192 continue;
5193 }
5194 break;
5195 }
5196find_next_key:
5197 /*
5198 * we didn't find a candidate key in this node, walk forward
5199 * and find another one
5200 */
5201 if (slot >= nritems) {
5202 path->slots[level] = slot;
5203 btrfs_set_path_blocking(path);
5204 sret = btrfs_find_next_key(root, path, min_key, level,
5205 min_trans);
5206 if (sret == 0) {
5207 btrfs_release_path(path);
5208 goto again;
5209 } else {
5210 goto out;
5211 }
5212 }
5213 /* save our key for returning back */
5214 btrfs_node_key_to_cpu(cur, &found_key, slot);
5215 path->slots[level] = slot;
5216 if (level == path->lowest_level) {
5217 ret = 0;
5218 goto out;
5219 }
5220 btrfs_set_path_blocking(path);
5221 cur = read_node_slot(fs_info, cur, slot);
5222 if (IS_ERR(cur)) {
5223 ret = PTR_ERR(cur);
5224 goto out;
5225 }
5226
5227 btrfs_tree_read_lock(cur);
5228
5229 path->locks[level - 1] = BTRFS_READ_LOCK;
5230 path->nodes[level - 1] = cur;
5231 unlock_up(path, level, 1, 0, NULL);
5232 btrfs_clear_path_blocking(path, NULL, 0);
5233 }
5234out:
5235 path->keep_locks = keep_locks;
5236 if (ret == 0) {
5237 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5238 btrfs_set_path_blocking(path);
5239 memcpy(min_key, &found_key, sizeof(found_key));
5240 }
5241 return ret;
5242}
5243
5244static int tree_move_down(struct btrfs_fs_info *fs_info,
5245 struct btrfs_path *path,
5246 int *level, int root_level)
5247{
5248 struct extent_buffer *eb;
5249
5250 BUG_ON(*level == 0);
5251 eb = read_node_slot(fs_info, path->nodes[*level], path->slots[*level]);
5252 if (IS_ERR(eb))
5253 return PTR_ERR(eb);
5254
5255 path->nodes[*level - 1] = eb;
5256 path->slots[*level - 1] = 0;
5257 (*level)--;
5258 return 0;
5259}
5260
5261static int tree_move_next_or_upnext(struct btrfs_fs_info *fs_info,
5262 struct btrfs_path *path,
5263 int *level, int root_level)
5264{
5265 int ret = 0;
5266 int nritems;
5267 nritems = btrfs_header_nritems(path->nodes[*level]);
5268
5269 path->slots[*level]++;
5270
5271 while (path->slots[*level] >= nritems) {
5272 if (*level == root_level)
5273 return -1;
5274
5275 /* move upnext */
5276 path->slots[*level] = 0;
5277 free_extent_buffer(path->nodes[*level]);
5278 path->nodes[*level] = NULL;
5279 (*level)++;
5280 path->slots[*level]++;
5281
5282 nritems = btrfs_header_nritems(path->nodes[*level]);
5283 ret = 1;
5284 }
5285 return ret;
5286}
5287
5288/*
5289 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
5290 * or down.
5291 */
5292static int tree_advance(struct btrfs_fs_info *fs_info,
5293 struct btrfs_path *path,
5294 int *level, int root_level,
5295 int allow_down,
5296 struct btrfs_key *key)
5297{
5298 int ret;
5299
5300 if (*level == 0 || !allow_down) {
5301 ret = tree_move_next_or_upnext(fs_info, path, level,
5302 root_level);
5303 } else {
5304 ret = tree_move_down(fs_info, path, level, root_level);
5305 }
5306 if (ret >= 0) {
5307 if (*level == 0)
5308 btrfs_item_key_to_cpu(path->nodes[*level], key,
5309 path->slots[*level]);
5310 else
5311 btrfs_node_key_to_cpu(path->nodes[*level], key,
5312 path->slots[*level]);
5313 }
5314 return ret;
5315}
5316
5317static int tree_compare_item(struct btrfs_path *left_path,
5318 struct btrfs_path *right_path,
5319 char *tmp_buf)
5320{
5321 int cmp;
5322 int len1, len2;
5323 unsigned long off1, off2;
5324
5325 len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
5326 len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
5327 if (len1 != len2)
5328 return 1;
5329
5330 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
5331 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
5332 right_path->slots[0]);
5333
5334 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
5335
5336 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
5337 if (cmp)
5338 return 1;
5339 return 0;
5340}
5341
5342#define ADVANCE 1
5343#define ADVANCE_ONLY_NEXT -1
5344
5345/*
5346 * This function compares two trees and calls the provided callback for
5347 * every changed/new/deleted item it finds.
5348 * If shared tree blocks are encountered, whole subtrees are skipped, making
5349 * the compare pretty fast on snapshotted subvolumes.
5350 *
5351 * This currently works on commit roots only. As commit roots are read only,
5352 * we don't do any locking. The commit roots are protected with transactions.
5353 * Transactions are ended and rejoined when a commit is tried in between.
5354 *
5355 * This function checks for modifications done to the trees while comparing.
5356 * If it detects a change, it aborts immediately.
5357 */
5358int btrfs_compare_trees(struct btrfs_root *left_root,
5359 struct btrfs_root *right_root,
5360 btrfs_changed_cb_t changed_cb, void *ctx)
5361{
5362 struct btrfs_fs_info *fs_info = left_root->fs_info;
5363 int ret;
5364 int cmp;
5365 struct btrfs_path *left_path = NULL;
5366 struct btrfs_path *right_path = NULL;
5367 struct btrfs_key left_key;
5368 struct btrfs_key right_key;
5369 char *tmp_buf = NULL;
5370 int left_root_level;
5371 int right_root_level;
5372 int left_level;
5373 int right_level;
5374 int left_end_reached;
5375 int right_end_reached;
5376 int advance_left;
5377 int advance_right;
5378 u64 left_blockptr;
5379 u64 right_blockptr;
5380 u64 left_gen;
5381 u64 right_gen;
5382
5383 left_path = btrfs_alloc_path();
5384 if (!left_path) {
5385 ret = -ENOMEM;
5386 goto out;
5387 }
5388 right_path = btrfs_alloc_path();
5389 if (!right_path) {
5390 ret = -ENOMEM;
5391 goto out;
5392 }
5393
5394 tmp_buf = kmalloc(fs_info->nodesize, GFP_KERNEL | __GFP_NOWARN);
5395 if (!tmp_buf) {
5396 tmp_buf = vmalloc(fs_info->nodesize);
5397 if (!tmp_buf) {
5398 ret = -ENOMEM;
5399 goto out;
5400 }
5401 }
5402
5403 left_path->search_commit_root = 1;
5404 left_path->skip_locking = 1;
5405 right_path->search_commit_root = 1;
5406 right_path->skip_locking = 1;
5407
5408 /*
5409 * Strategy: Go to the first items of both trees. Then do
5410 *
5411 * If both trees are at level 0
5412 * Compare keys of current items
5413 * If left < right treat left item as new, advance left tree
5414 * and repeat
5415 * If left > right treat right item as deleted, advance right tree
5416 * and repeat
5417 * If left == right do deep compare of items, treat as changed if
5418 * needed, advance both trees and repeat
5419 * If both trees are at the same level but not at level 0
5420 * Compare keys of current nodes/leafs
5421 * If left < right advance left tree and repeat
5422 * If left > right advance right tree and repeat
5423 * If left == right compare blockptrs of the next nodes/leafs
5424 * If they match advance both trees but stay at the same level
5425 * and repeat
5426 * If they don't match advance both trees while allowing to go
5427 * deeper and repeat
5428 * If tree levels are different
5429 * Advance the tree that needs it and repeat
5430 *
5431 * Advancing a tree means:
5432 * If we are at level 0, try to go to the next slot. If that's not
5433 * possible, go one level up and repeat. Stop when we found a level
5434 * where we could go to the next slot. We may at this point be on a
5435 * node or a leaf.
5436 *
5437 * If we are not at level 0 and not on shared tree blocks, go one
5438 * level deeper.
5439 *
5440 * If we are not at level 0 and on shared tree blocks, go one slot to
5441 * the right if possible or go up and right.
5442 */
5443
5444 down_read(&fs_info->commit_root_sem);
5445 left_level = btrfs_header_level(left_root->commit_root);
5446 left_root_level = left_level;
5447 left_path->nodes[left_level] = left_root->commit_root;
5448 extent_buffer_get(left_path->nodes[left_level]);
5449
5450 right_level = btrfs_header_level(right_root->commit_root);
5451 right_root_level = right_level;
5452 right_path->nodes[right_level] = right_root->commit_root;
5453 extent_buffer_get(right_path->nodes[right_level]);
5454 up_read(&fs_info->commit_root_sem);
5455
5456 if (left_level == 0)
5457 btrfs_item_key_to_cpu(left_path->nodes[left_level],
5458 &left_key, left_path->slots[left_level]);
5459 else
5460 btrfs_node_key_to_cpu(left_path->nodes[left_level],
5461 &left_key, left_path->slots[left_level]);
5462 if (right_level == 0)
5463 btrfs_item_key_to_cpu(right_path->nodes[right_level],
5464 &right_key, right_path->slots[right_level]);
5465 else
5466 btrfs_node_key_to_cpu(right_path->nodes[right_level],
5467 &right_key, right_path->slots[right_level]);
5468
5469 left_end_reached = right_end_reached = 0;
5470 advance_left = advance_right = 0;
5471
5472 while (1) {
5473 if (advance_left && !left_end_reached) {
5474 ret = tree_advance(fs_info, left_path, &left_level,
5475 left_root_level,
5476 advance_left != ADVANCE_ONLY_NEXT,
5477 &left_key);
5478 if (ret == -1)
5479 left_end_reached = ADVANCE;
5480 else if (ret < 0)
5481 goto out;
5482 advance_left = 0;
5483 }
5484 if (advance_right && !right_end_reached) {
5485 ret = tree_advance(fs_info, right_path, &right_level,
5486 right_root_level,
5487 advance_right != ADVANCE_ONLY_NEXT,
5488 &right_key);
5489 if (ret == -1)
5490 right_end_reached = ADVANCE;
5491 else if (ret < 0)
5492 goto out;
5493 advance_right = 0;
5494 }
5495
5496 if (left_end_reached && right_end_reached) {
5497 ret = 0;
5498 goto out;
5499 } else if (left_end_reached) {
5500 if (right_level == 0) {
5501 ret = changed_cb(left_root, right_root,
5502 left_path, right_path,
5503 &right_key,
5504 BTRFS_COMPARE_TREE_DELETED,
5505 ctx);
5506 if (ret < 0)
5507 goto out;
5508 }
5509 advance_right = ADVANCE;
5510 continue;
5511 } else if (right_end_reached) {
5512 if (left_level == 0) {
5513 ret = changed_cb(left_root, right_root,
5514 left_path, right_path,
5515 &left_key,
5516 BTRFS_COMPARE_TREE_NEW,
5517 ctx);
5518 if (ret < 0)
5519 goto out;
5520 }
5521 advance_left = ADVANCE;
5522 continue;
5523 }
5524
5525 if (left_level == 0 && right_level == 0) {
5526 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5527 if (cmp < 0) {
5528 ret = changed_cb(left_root, right_root,
5529 left_path, right_path,
5530 &left_key,
5531 BTRFS_COMPARE_TREE_NEW,
5532 ctx);
5533 if (ret < 0)
5534 goto out;
5535 advance_left = ADVANCE;
5536 } else if (cmp > 0) {
5537 ret = changed_cb(left_root, right_root,
5538 left_path, right_path,
5539 &right_key,
5540 BTRFS_COMPARE_TREE_DELETED,
5541 ctx);
5542 if (ret < 0)
5543 goto out;
5544 advance_right = ADVANCE;
5545 } else {
5546 enum btrfs_compare_tree_result result;
5547
5548 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
5549 ret = tree_compare_item(left_path, right_path,
5550 tmp_buf);
5551 if (ret)
5552 result = BTRFS_COMPARE_TREE_CHANGED;
5553 else
5554 result = BTRFS_COMPARE_TREE_SAME;
5555 ret = changed_cb(left_root, right_root,
5556 left_path, right_path,
5557 &left_key, result, ctx);
5558 if (ret < 0)
5559 goto out;
5560 advance_left = ADVANCE;
5561 advance_right = ADVANCE;
5562 }
5563 } else if (left_level == right_level) {
5564 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5565 if (cmp < 0) {
5566 advance_left = ADVANCE;
5567 } else if (cmp > 0) {
5568 advance_right = ADVANCE;
5569 } else {
5570 left_blockptr = btrfs_node_blockptr(
5571 left_path->nodes[left_level],
5572 left_path->slots[left_level]);
5573 right_blockptr = btrfs_node_blockptr(
5574 right_path->nodes[right_level],
5575 right_path->slots[right_level]);
5576 left_gen = btrfs_node_ptr_generation(
5577 left_path->nodes[left_level],
5578 left_path->slots[left_level]);
5579 right_gen = btrfs_node_ptr_generation(
5580 right_path->nodes[right_level],
5581 right_path->slots[right_level]);
5582 if (left_blockptr == right_blockptr &&
5583 left_gen == right_gen) {
5584 /*
5585 * As we're on a shared block, don't
5586 * allow to go deeper.
5587 */
5588 advance_left = ADVANCE_ONLY_NEXT;
5589 advance_right = ADVANCE_ONLY_NEXT;
5590 } else {
5591 advance_left = ADVANCE;
5592 advance_right = ADVANCE;
5593 }
5594 }
5595 } else if (left_level < right_level) {
5596 advance_right = ADVANCE;
5597 } else {
5598 advance_left = ADVANCE;
5599 }
5600 }
5601
5602out:
5603 btrfs_free_path(left_path);
5604 btrfs_free_path(right_path);
5605 kvfree(tmp_buf);
5606 return ret;
5607}
5608
5609/*
5610 * this is similar to btrfs_next_leaf, but does not try to preserve
5611 * and fixup the path. It looks for and returns the next key in the
5612 * tree based on the current path and the min_trans parameters.
5613 *
5614 * 0 is returned if another key is found, < 0 if there are any errors
5615 * and 1 is returned if there are no higher keys in the tree
5616 *
5617 * path->keep_locks should be set to 1 on the search made before
5618 * calling this function.
5619 */
5620int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5621 struct btrfs_key *key, int level, u64 min_trans)
5622{
5623 int slot;
5624 struct extent_buffer *c;
5625
5626 WARN_ON(!path->keep_locks);
5627 while (level < BTRFS_MAX_LEVEL) {
5628 if (!path->nodes[level])
5629 return 1;
5630
5631 slot = path->slots[level] + 1;
5632 c = path->nodes[level];
5633next:
5634 if (slot >= btrfs_header_nritems(c)) {
5635 int ret;
5636 int orig_lowest;
5637 struct btrfs_key cur_key;
5638 if (level + 1 >= BTRFS_MAX_LEVEL ||
5639 !path->nodes[level + 1])
5640 return 1;
5641
5642 if (path->locks[level + 1]) {
5643 level++;
5644 continue;
5645 }
5646
5647 slot = btrfs_header_nritems(c) - 1;
5648 if (level == 0)
5649 btrfs_item_key_to_cpu(c, &cur_key, slot);
5650 else
5651 btrfs_node_key_to_cpu(c, &cur_key, slot);
5652
5653 orig_lowest = path->lowest_level;
5654 btrfs_release_path(path);
5655 path->lowest_level = level;
5656 ret = btrfs_search_slot(NULL, root, &cur_key, path,
5657 0, 0);
5658 path->lowest_level = orig_lowest;
5659 if (ret < 0)
5660 return ret;
5661
5662 c = path->nodes[level];
5663 slot = path->slots[level];
5664 if (ret == 0)
5665 slot++;
5666 goto next;
5667 }
5668
5669 if (level == 0)
5670 btrfs_item_key_to_cpu(c, key, slot);
5671 else {
5672 u64 gen = btrfs_node_ptr_generation(c, slot);
5673
5674 if (gen < min_trans) {
5675 slot++;
5676 goto next;
5677 }
5678 btrfs_node_key_to_cpu(c, key, slot);
5679 }
5680 return 0;
5681 }
5682 return 1;
5683}
5684
5685/*
5686 * search the tree again to find a leaf with greater keys
5687 * returns 0 if it found something or 1 if there are no greater leaves.
5688 * returns < 0 on io errors.
5689 */
5690int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5691{
5692 return btrfs_next_old_leaf(root, path, 0);
5693}
5694
5695int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5696 u64 time_seq)
5697{
5698 int slot;
5699 int level;
5700 struct extent_buffer *c;
5701 struct extent_buffer *next;
5702 struct btrfs_key key;
5703 u32 nritems;
5704 int ret;
5705 int old_spinning = path->leave_spinning;
5706 int next_rw_lock = 0;
5707
5708 nritems = btrfs_header_nritems(path->nodes[0]);
5709 if (nritems == 0)
5710 return 1;
5711
5712 btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5713again:
5714 level = 1;
5715 next = NULL;
5716 next_rw_lock = 0;
5717 btrfs_release_path(path);
5718
5719 path->keep_locks = 1;
5720 path->leave_spinning = 1;
5721
5722 if (time_seq)
5723 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5724 else
5725 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5726 path->keep_locks = 0;
5727
5728 if (ret < 0)
5729 return ret;
5730
5731 nritems = btrfs_header_nritems(path->nodes[0]);
5732 /*
5733 * by releasing the path above we dropped all our locks. A balance
5734 * could have added more items next to the key that used to be
5735 * at the very end of the block. So, check again here and
5736 * advance the path if there are now more items available.
5737 */
5738 if (nritems > 0 && path->slots[0] < nritems - 1) {
5739 if (ret == 0)
5740 path->slots[0]++;
5741 ret = 0;
5742 goto done;
5743 }
5744 /*
5745 * So the above check misses one case:
5746 * - after releasing the path above, someone has removed the item that
5747 * used to be at the very end of the block, and balance between leafs
5748 * gets another one with bigger key.offset to replace it.
5749 *
5750 * This one should be returned as well, or we can get leaf corruption
5751 * later(esp. in __btrfs_drop_extents()).
5752 *
5753 * And a bit more explanation about this check,
5754 * with ret > 0, the key isn't found, the path points to the slot
5755 * where it should be inserted, so the path->slots[0] item must be the
5756 * bigger one.
5757 */
5758 if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5759 ret = 0;
5760 goto done;
5761 }
5762
5763 while (level < BTRFS_MAX_LEVEL) {
5764 if (!path->nodes[level]) {
5765 ret = 1;
5766 goto done;
5767 }
5768
5769 slot = path->slots[level] + 1;
5770 c = path->nodes[level];
5771 if (slot >= btrfs_header_nritems(c)) {
5772 level++;
5773 if (level == BTRFS_MAX_LEVEL) {
5774 ret = 1;
5775 goto done;
5776 }
5777 continue;
5778 }
5779
5780 if (next) {
5781 btrfs_tree_unlock_rw(next, next_rw_lock);
5782 free_extent_buffer(next);
5783 }
5784
5785 next = c;
5786 next_rw_lock = path->locks[level];
5787 ret = read_block_for_search(NULL, root, path, &next, level,
5788 slot, &key, 0);
5789 if (ret == -EAGAIN)
5790 goto again;
5791
5792 if (ret < 0) {
5793 btrfs_release_path(path);
5794 goto done;
5795 }
5796
5797 if (!path->skip_locking) {
5798 ret = btrfs_try_tree_read_lock(next);
5799 if (!ret && time_seq) {
5800 /*
5801 * If we don't get the lock, we may be racing
5802 * with push_leaf_left, holding that lock while
5803 * itself waiting for the leaf we've currently
5804 * locked. To solve this situation, we give up
5805 * on our lock and cycle.
5806 */
5807 free_extent_buffer(next);
5808 btrfs_release_path(path);
5809 cond_resched();
5810 goto again;
5811 }
5812 if (!ret) {
5813 btrfs_set_path_blocking(path);
5814 btrfs_tree_read_lock(next);
5815 btrfs_clear_path_blocking(path, next,
5816 BTRFS_READ_LOCK);
5817 }
5818 next_rw_lock = BTRFS_READ_LOCK;
5819 }
5820 break;
5821 }
5822 path->slots[level] = slot;
5823 while (1) {
5824 level--;
5825 c = path->nodes[level];
5826 if (path->locks[level])
5827 btrfs_tree_unlock_rw(c, path->locks[level]);
5828
5829 free_extent_buffer(c);
5830 path->nodes[level] = next;
5831 path->slots[level] = 0;
5832 if (!path->skip_locking)
5833 path->locks[level] = next_rw_lock;
5834 if (!level)
5835 break;
5836
5837 ret = read_block_for_search(NULL, root, path, &next, level,
5838 0, &key, 0);
5839 if (ret == -EAGAIN)
5840 goto again;
5841
5842 if (ret < 0) {
5843 btrfs_release_path(path);
5844 goto done;
5845 }
5846
5847 if (!path->skip_locking) {
5848 ret = btrfs_try_tree_read_lock(next);
5849 if (!ret) {
5850 btrfs_set_path_blocking(path);
5851 btrfs_tree_read_lock(next);
5852 btrfs_clear_path_blocking(path, next,
5853 BTRFS_READ_LOCK);
5854 }
5855 next_rw_lock = BTRFS_READ_LOCK;
5856 }
5857 }
5858 ret = 0;
5859done:
5860 unlock_up(path, 0, 1, 0, NULL);
5861 path->leave_spinning = old_spinning;
5862 if (!old_spinning)
5863 btrfs_set_path_blocking(path);
5864
5865 return ret;
5866}
5867
5868/*
5869 * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5870 * searching until it gets past min_objectid or finds an item of 'type'
5871 *
5872 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5873 */
5874int btrfs_previous_item(struct btrfs_root *root,
5875 struct btrfs_path *path, u64 min_objectid,
5876 int type)
5877{
5878 struct btrfs_key found_key;
5879 struct extent_buffer *leaf;
5880 u32 nritems;
5881 int ret;
5882
5883 while (1) {
5884 if (path->slots[0] == 0) {
5885 btrfs_set_path_blocking(path);
5886 ret = btrfs_prev_leaf(root, path);
5887 if (ret != 0)
5888 return ret;
5889 } else {
5890 path->slots[0]--;
5891 }
5892 leaf = path->nodes[0];
5893 nritems = btrfs_header_nritems(leaf);
5894 if (nritems == 0)
5895 return 1;
5896 if (path->slots[0] == nritems)
5897 path->slots[0]--;
5898
5899 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5900 if (found_key.objectid < min_objectid)
5901 break;
5902 if (found_key.type == type)
5903 return 0;
5904 if (found_key.objectid == min_objectid &&
5905 found_key.type < type)
5906 break;
5907 }
5908 return 1;
5909}
5910
5911/*
5912 * search in extent tree to find a previous Metadata/Data extent item with
5913 * min objecitd.
5914 *
5915 * returns 0 if something is found, 1 if nothing was found and < 0 on error
5916 */
5917int btrfs_previous_extent_item(struct btrfs_root *root,
5918 struct btrfs_path *path, u64 min_objectid)
5919{
5920 struct btrfs_key found_key;
5921 struct extent_buffer *leaf;
5922 u32 nritems;
5923 int ret;
5924
5925 while (1) {
5926 if (path->slots[0] == 0) {
5927 btrfs_set_path_blocking(path);
5928 ret = btrfs_prev_leaf(root, path);
5929 if (ret != 0)
5930 return ret;
5931 } else {
5932 path->slots[0]--;
5933 }
5934 leaf = path->nodes[0];
5935 nritems = btrfs_header_nritems(leaf);
5936 if (nritems == 0)
5937 return 1;
5938 if (path->slots[0] == nritems)
5939 path->slots[0]--;
5940
5941 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5942 if (found_key.objectid < min_objectid)
5943 break;
5944 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5945 found_key.type == BTRFS_METADATA_ITEM_KEY)
5946 return 0;
5947 if (found_key.objectid == min_objectid &&
5948 found_key.type < BTRFS_EXTENT_ITEM_KEY)
5949 break;
5950 }
5951 return 1;
5952}