1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Copyright (C) 2009 Oracle. All rights reserved.
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public
10 * License version 2 as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License for more details.
18 #include <linux/sort.h>
19 #define MLOG_MASK_PREFIX ML_REFCOUNT
20 #include <cluster/masklog.h>
28 #include "buffer_head_io.h"
29 #include "blockcheck.h"
30 #include "refcounttree.h"
33 #include "extent_map.h"
35 static inline struct ocfs2_refcount_tree *
36 cache_info_to_refcount(struct ocfs2_caching_info *ci)
38 return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
41 static int ocfs2_validate_refcount_block(struct super_block *sb,
42 struct buffer_head *bh)
45 struct ocfs2_refcount_block *rb =
46 (struct ocfs2_refcount_block *)bh->b_data;
48 mlog(0, "Validating refcount block %llu\n",
49 (unsigned long long)bh->b_blocknr);
51 BUG_ON(!buffer_uptodate(bh));
54 * If the ecc fails, we return the error but otherwise
55 * leave the filesystem running. We know any error is
56 * local to this block.
58 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check);
60 mlog(ML_ERROR, "Checksum failed for refcount block %llu\n",
61 (unsigned long long)bh->b_blocknr);
66 if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) {
68 "Refcount block #%llu has bad signature %.*s",
69 (unsigned long long)bh->b_blocknr, 7,
74 if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
76 "Refcount block #%llu has an invalid rf_blkno "
78 (unsigned long long)bh->b_blocknr,
79 (unsigned long long)le64_to_cpu(rb->rf_blkno));
83 if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
85 "Refcount block #%llu has an invalid "
86 "rf_fs_generation of #%u",
87 (unsigned long long)bh->b_blocknr,
88 le32_to_cpu(rb->rf_fs_generation));
95 static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
97 struct buffer_head **bh)
100 struct buffer_head *tmp = *bh;
102 rc = ocfs2_read_block(ci, rb_blkno, &tmp,
103 ocfs2_validate_refcount_block);
105 /* If ocfs2_read_block() got us a new bh, pass it up. */
112 static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
114 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
119 static struct super_block *
120 ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
122 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
127 static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
129 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
131 spin_lock(&rf->rf_lock);
134 static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
136 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
138 spin_unlock(&rf->rf_lock);
141 static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
143 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
145 mutex_lock(&rf->rf_io_mutex);
148 static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
150 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
152 mutex_unlock(&rf->rf_io_mutex);
155 static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
156 .co_owner = ocfs2_refcount_cache_owner,
157 .co_get_super = ocfs2_refcount_cache_get_super,
158 .co_cache_lock = ocfs2_refcount_cache_lock,
159 .co_cache_unlock = ocfs2_refcount_cache_unlock,
160 .co_io_lock = ocfs2_refcount_cache_io_lock,
161 .co_io_unlock = ocfs2_refcount_cache_io_unlock,
164 static struct ocfs2_refcount_tree *
165 ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
167 struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
168 struct ocfs2_refcount_tree *tree = NULL;
171 tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
173 if (blkno < tree->rf_blkno)
175 else if (blkno > tree->rf_blkno)
184 /* osb_lock is already locked. */
185 static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
186 struct ocfs2_refcount_tree *new)
188 u64 rf_blkno = new->rf_blkno;
189 struct rb_node *parent = NULL;
190 struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
191 struct ocfs2_refcount_tree *tmp;
196 tmp = rb_entry(parent, struct ocfs2_refcount_tree,
199 if (rf_blkno < tmp->rf_blkno)
201 else if (rf_blkno > tmp->rf_blkno)
204 /* This should never happen! */
205 mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
206 (unsigned long long)rf_blkno);
211 rb_link_node(&new->rf_node, parent, p);
212 rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
215 static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
217 ocfs2_metadata_cache_exit(&tree->rf_ci);
218 ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
219 ocfs2_lock_res_free(&tree->rf_lockres);
224 ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
225 struct ocfs2_refcount_tree *tree)
227 rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
228 if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
229 osb->osb_ref_tree_lru = NULL;
232 static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
233 struct ocfs2_refcount_tree *tree)
235 spin_lock(&osb->osb_lock);
236 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
237 spin_unlock(&osb->osb_lock);
240 void ocfs2_kref_remove_refcount_tree(struct kref *kref)
242 struct ocfs2_refcount_tree *tree =
243 container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
245 ocfs2_free_refcount_tree(tree);
249 ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
251 kref_get(&tree->rf_getcnt);
255 ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
257 kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
260 static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
261 struct super_block *sb)
263 ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
264 mutex_init(&new->rf_io_mutex);
266 spin_lock_init(&new->rf_lock);
269 static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
270 struct ocfs2_refcount_tree *new,
271 u64 rf_blkno, u32 generation)
273 init_rwsem(&new->rf_sem);
274 ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
275 rf_blkno, generation);
278 static struct ocfs2_refcount_tree*
279 ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
281 struct ocfs2_refcount_tree *new;
283 new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
287 new->rf_blkno = rf_blkno;
288 kref_init(&new->rf_getcnt);
289 ocfs2_init_refcount_tree_ci(new, osb->sb);
294 static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
295 struct ocfs2_refcount_tree **ret_tree)
298 struct ocfs2_refcount_tree *tree, *new = NULL;
299 struct buffer_head *ref_root_bh = NULL;
300 struct ocfs2_refcount_block *ref_rb;
302 spin_lock(&osb->osb_lock);
303 if (osb->osb_ref_tree_lru &&
304 osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
305 tree = osb->osb_ref_tree_lru;
307 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
311 spin_unlock(&osb->osb_lock);
313 new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
320 * We need the generation to create the refcount tree lock and since
321 * it isn't changed during the tree modification, we are safe here to
322 * read without protection.
323 * We also have to purge the cache after we create the lock since the
324 * refcount block may have the stale data. It can only be trusted when
325 * we hold the refcount lock.
327 ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
330 ocfs2_metadata_cache_exit(&new->rf_ci);
335 ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
336 new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
337 ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
339 ocfs2_metadata_cache_purge(&new->rf_ci);
341 spin_lock(&osb->osb_lock);
342 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
346 ocfs2_insert_refcount_tree(osb, new);
354 osb->osb_ref_tree_lru = tree;
356 spin_unlock(&osb->osb_lock);
359 ocfs2_free_refcount_tree(new);
365 static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
368 struct buffer_head *di_bh = NULL;
369 struct ocfs2_dinode *di;
371 ret = ocfs2_read_inode_block(inode, &di_bh);
377 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
379 di = (struct ocfs2_dinode *)di_bh->b_data;
380 *ref_blkno = le64_to_cpu(di->i_refcount_loc);
386 static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
387 struct ocfs2_refcount_tree *tree, int rw)
391 ret = ocfs2_refcount_lock(tree, rw);
398 down_write(&tree->rf_sem);
400 down_read(&tree->rf_sem);
407 * Lock the refcount tree pointed by ref_blkno and return the tree.
408 * In most case, we lock the tree and read the refcount block.
409 * So read it here if the caller really needs it.
411 * If the tree has been re-created by other node, it will free the
412 * old one and re-create it.
414 int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
415 u64 ref_blkno, int rw,
416 struct ocfs2_refcount_tree **ret_tree,
417 struct buffer_head **ref_bh)
419 int ret, delete_tree = 0;
420 struct ocfs2_refcount_tree *tree = NULL;
421 struct buffer_head *ref_root_bh = NULL;
422 struct ocfs2_refcount_block *rb;
425 ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
431 ocfs2_refcount_tree_get(tree);
433 ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
436 ocfs2_refcount_tree_put(tree);
440 ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
444 ocfs2_unlock_refcount_tree(osb, tree, rw);
445 ocfs2_refcount_tree_put(tree);
449 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
451 * If the refcount block has been freed and re-created, we may need
452 * to recreate the refcount tree also.
454 * Here we just remove the tree from the rb-tree, and the last
455 * kref holder will unlock and delete this refcount_tree.
456 * Then we goto "again" and ocfs2_get_refcount_tree will create
457 * the new refcount tree for us.
459 if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
460 if (!tree->rf_removed) {
461 ocfs2_erase_refcount_tree_from_list(osb, tree);
462 tree->rf_removed = 1;
466 ocfs2_unlock_refcount_tree(osb, tree, rw);
468 * We get an extra reference when we create the refcount
469 * tree, so another put will destroy it.
472 ocfs2_refcount_tree_put(tree);
480 *ref_bh = ref_root_bh;
488 int ocfs2_lock_refcount_tree_by_inode(struct inode *inode, int rw,
489 struct ocfs2_refcount_tree **ret_tree,
490 struct buffer_head **ref_bh)
495 ret = ocfs2_get_refcount_block(inode, &ref_blkno);
501 return ocfs2_lock_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno,
502 rw, ret_tree, ref_bh);
505 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
506 struct ocfs2_refcount_tree *tree, int rw)
509 up_write(&tree->rf_sem);
511 up_read(&tree->rf_sem);
513 ocfs2_refcount_unlock(tree, rw);
514 ocfs2_refcount_tree_put(tree);
517 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
519 struct rb_node *node;
520 struct ocfs2_refcount_tree *tree;
521 struct rb_root *root = &osb->osb_rf_lock_tree;
523 while ((node = rb_last(root)) != NULL) {
524 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
526 mlog(0, "Purge tree %llu\n",
527 (unsigned long long) tree->rf_blkno);
529 rb_erase(&tree->rf_node, root);
530 ocfs2_free_refcount_tree(tree);
535 * Create a refcount tree for an inode.
536 * We take for granted that the inode is already locked.
538 static int ocfs2_create_refcount_tree(struct inode *inode,
539 struct buffer_head *di_bh)
542 handle_t *handle = NULL;
543 struct ocfs2_alloc_context *meta_ac = NULL;
544 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
545 struct ocfs2_inode_info *oi = OCFS2_I(inode);
546 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
547 struct buffer_head *new_bh = NULL;
548 struct ocfs2_refcount_block *rb;
549 struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
550 u16 suballoc_bit_start;
554 BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
556 mlog(0, "create tree for inode %lu\n", inode->i_ino);
558 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
564 handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
565 if (IS_ERR(handle)) {
566 ret = PTR_ERR(handle);
571 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
572 OCFS2_JOURNAL_ACCESS_WRITE);
578 ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1,
579 &suballoc_bit_start, &num_got,
586 new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
593 new_bh = sb_getblk(inode->i_sb, first_blkno);
594 ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
596 ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
597 OCFS2_JOURNAL_ACCESS_CREATE);
603 /* Initialize ocfs2_refcount_block. */
604 rb = (struct ocfs2_refcount_block *)new_bh->b_data;
605 memset(rb, 0, inode->i_sb->s_blocksize);
606 strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
607 rb->rf_suballoc_slot = cpu_to_le16(osb->slot_num);
608 rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
609 rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
610 rb->rf_blkno = cpu_to_le64(first_blkno);
611 rb->rf_count = cpu_to_le32(1);
612 rb->rf_records.rl_count =
613 cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
614 spin_lock(&osb->osb_lock);
615 rb->rf_generation = osb->s_next_generation++;
616 spin_unlock(&osb->osb_lock);
618 ocfs2_journal_dirty(handle, new_bh);
620 spin_lock(&oi->ip_lock);
621 oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
622 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
623 di->i_refcount_loc = cpu_to_le64(first_blkno);
624 spin_unlock(&oi->ip_lock);
626 mlog(0, "created tree for inode %lu, refblock %llu\n",
627 inode->i_ino, (unsigned long long)first_blkno);
629 ocfs2_journal_dirty(handle, di_bh);
632 * We have to init the tree lock here since it will use
633 * the generation number to create it.
635 new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
636 ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
637 new_tree->rf_generation);
639 spin_lock(&osb->osb_lock);
640 tree = ocfs2_find_refcount_tree(osb, first_blkno);
643 * We've just created a new refcount tree in this block. If
644 * we found a refcount tree on the ocfs2_super, it must be
645 * one we just deleted. We free the old tree before
646 * inserting the new tree.
648 BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
650 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
651 ocfs2_insert_refcount_tree(osb, new_tree);
652 spin_unlock(&osb->osb_lock);
655 ocfs2_refcount_tree_put(tree);
658 ocfs2_commit_trans(osb, handle);
662 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
668 ocfs2_free_alloc_context(meta_ac);
673 static int ocfs2_set_refcount_tree(struct inode *inode,
674 struct buffer_head *di_bh,
678 handle_t *handle = NULL;
679 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
680 struct ocfs2_inode_info *oi = OCFS2_I(inode);
681 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
682 struct buffer_head *ref_root_bh = NULL;
683 struct ocfs2_refcount_block *rb;
684 struct ocfs2_refcount_tree *ref_tree;
686 BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
688 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
689 &ref_tree, &ref_root_bh);
695 handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
696 if (IS_ERR(handle)) {
697 ret = PTR_ERR(handle);
702 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
703 OCFS2_JOURNAL_ACCESS_WRITE);
709 ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
710 OCFS2_JOURNAL_ACCESS_WRITE);
716 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
717 le32_add_cpu(&rb->rf_count, 1);
719 ocfs2_journal_dirty(handle, ref_root_bh);
721 spin_lock(&oi->ip_lock);
722 oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
723 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
724 di->i_refcount_loc = cpu_to_le64(refcount_loc);
725 spin_unlock(&oi->ip_lock);
726 ocfs2_journal_dirty(handle, di_bh);
729 ocfs2_commit_trans(osb, handle);
731 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
737 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
739 int ret, delete_tree = 0;
740 handle_t *handle = NULL;
741 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
742 struct ocfs2_inode_info *oi = OCFS2_I(inode);
743 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
744 struct ocfs2_refcount_block *rb;
745 struct inode *alloc_inode = NULL;
746 struct buffer_head *alloc_bh = NULL;
747 struct buffer_head *blk_bh = NULL;
748 struct ocfs2_refcount_tree *ref_tree;
749 int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
750 u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
753 if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
757 ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
763 rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
766 * If we are the last user, we need to free the block.
767 * So lock the allocator ahead.
769 if (le32_to_cpu(rb->rf_count) == 1) {
770 blk = le64_to_cpu(rb->rf_blkno);
771 bit = le16_to_cpu(rb->rf_suballoc_bit);
772 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
774 alloc_inode = ocfs2_get_system_file_inode(osb,
775 EXTENT_ALLOC_SYSTEM_INODE,
776 le16_to_cpu(rb->rf_suballoc_slot));
782 mutex_lock(&alloc_inode->i_mutex);
784 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
790 credits += OCFS2_SUBALLOC_FREE;
793 handle = ocfs2_start_trans(osb, credits);
794 if (IS_ERR(handle)) {
795 ret = PTR_ERR(handle);
800 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
801 OCFS2_JOURNAL_ACCESS_WRITE);
807 ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
808 OCFS2_JOURNAL_ACCESS_WRITE);
814 spin_lock(&oi->ip_lock);
815 oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
816 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
817 di->i_refcount_loc = 0;
818 spin_unlock(&oi->ip_lock);
819 ocfs2_journal_dirty(handle, di_bh);
821 le32_add_cpu(&rb->rf_count , -1);
822 ocfs2_journal_dirty(handle, blk_bh);
826 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
827 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
828 alloc_bh, bit, bg_blkno, 1);
834 ocfs2_commit_trans(osb, handle);
837 ocfs2_inode_unlock(alloc_inode, 1);
842 mutex_unlock(&alloc_inode->i_mutex);
846 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
848 ocfs2_refcount_tree_put(ref_tree);
854 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
855 struct buffer_head *ref_leaf_bh,
856 u64 cpos, unsigned int len,
857 struct ocfs2_refcount_rec *ret_rec,
861 struct ocfs2_refcount_block *rb =
862 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
863 struct ocfs2_refcount_rec *rec = NULL;
865 for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
866 rec = &rb->rf_records.rl_recs[i];
868 if (le64_to_cpu(rec->r_cpos) +
869 le32_to_cpu(rec->r_clusters) <= cpos)
871 else if (le64_to_cpu(rec->r_cpos) > cpos)
874 /* ok, cpos fail in this rec. Just return. */
881 /* We meet with a hole here, so fake the rec. */
882 ret_rec->r_cpos = cpu_to_le64(cpos);
883 ret_rec->r_refcount = 0;
884 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
885 le64_to_cpu(rec->r_cpos) < cpos + len)
886 ret_rec->r_clusters =
887 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
889 ret_rec->r_clusters = cpu_to_le32(len);
897 * Given a cpos and len, try to find the refcount record which contains cpos.
898 * 1. If cpos can be found in one refcount record, return the record.
899 * 2. If cpos can't be found, return a fake record which start from cpos
900 * and end at a small value between cpos+len and start of the next record.
901 * This fake record has r_refcount = 0.
903 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
904 struct buffer_head *ref_root_bh,
905 u64 cpos, unsigned int len,
906 struct ocfs2_refcount_rec *ret_rec,
908 struct buffer_head **ret_bh)
910 int ret = 0, i, found;
912 struct ocfs2_extent_list *el;
913 struct ocfs2_extent_rec *tmp, *rec = NULL;
914 struct ocfs2_extent_block *eb;
915 struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
916 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
917 struct ocfs2_refcount_block *rb =
918 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
920 if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
921 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
923 *ret_bh = ref_root_bh;
929 low_cpos = cpos & OCFS2_32BIT_POS_MASK;
931 if (el->l_tree_depth) {
932 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
938 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
941 if (el->l_tree_depth) {
943 "refcount tree %llu has non zero tree "
944 "depth in leaf btree tree block %llu\n",
945 (unsigned long long)ocfs2_metadata_cache_owner(ci),
946 (unsigned long long)eb_bh->b_blocknr);
953 for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
954 rec = &el->l_recs[i];
956 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
962 /* adjust len when we have ocfs2_extent_rec after it. */
963 if (found && i < le16_to_cpu(el->l_next_free_rec) - 1) {
964 tmp = &el->l_recs[i+1];
966 if (le32_to_cpu(tmp->e_cpos) < cpos + len)
967 len = le32_to_cpu(tmp->e_cpos) - cpos;
970 ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
977 ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
979 *ret_bh = ref_leaf_bh;
985 enum ocfs2_ref_rec_contig {
989 REF_CONTIG_LEFTRIGHT,
992 static enum ocfs2_ref_rec_contig
993 ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
996 if ((rb->rf_records.rl_recs[index].r_refcount ==
997 rb->rf_records.rl_recs[index + 1].r_refcount) &&
998 (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
999 le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1000 le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1001 return REF_CONTIG_RIGHT;
1003 return REF_CONTIG_NONE;
1006 static enum ocfs2_ref_rec_contig
1007 ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1010 enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1012 if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1013 ret = ocfs2_refcount_rec_adjacent(rb, index);
1016 enum ocfs2_ref_rec_contig tmp;
1018 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1020 if (tmp == REF_CONTIG_RIGHT) {
1021 if (ret == REF_CONTIG_RIGHT)
1022 ret = REF_CONTIG_LEFTRIGHT;
1024 ret = REF_CONTIG_LEFT;
1031 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1034 BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1035 rb->rf_records.rl_recs[index+1].r_refcount);
1037 le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1038 le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1040 if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1041 memmove(&rb->rf_records.rl_recs[index + 1],
1042 &rb->rf_records.rl_recs[index + 2],
1043 sizeof(struct ocfs2_refcount_rec) *
1044 (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1046 memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1047 0, sizeof(struct ocfs2_refcount_rec));
1048 le16_add_cpu(&rb->rf_records.rl_used, -1);
1052 * Merge the refcount rec if we are contiguous with the adjacent recs.
1054 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1057 enum ocfs2_ref_rec_contig contig =
1058 ocfs2_refcount_rec_contig(rb, index);
1060 if (contig == REF_CONTIG_NONE)
1063 if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1068 ocfs2_rotate_refcount_rec_left(rb, index);
1070 if (contig == REF_CONTIG_LEFTRIGHT)
1071 ocfs2_rotate_refcount_rec_left(rb, index);
1075 * Change the refcount indexed by "index" in ref_bh.
1076 * If refcount reaches 0, remove it.
1078 static int ocfs2_change_refcount_rec(handle_t *handle,
1079 struct ocfs2_caching_info *ci,
1080 struct buffer_head *ref_leaf_bh,
1081 int index, int change)
1084 struct ocfs2_refcount_block *rb =
1085 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1086 struct ocfs2_refcount_list *rl = &rb->rf_records;
1087 struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1089 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1090 OCFS2_JOURNAL_ACCESS_WRITE);
1096 mlog(0, "change index %d, old count %u, change %d\n", index,
1097 le32_to_cpu(rec->r_refcount), change);
1098 le32_add_cpu(&rec->r_refcount, change);
1100 if (!rec->r_refcount) {
1101 if (index != le16_to_cpu(rl->rl_used) - 1) {
1102 memmove(rec, rec + 1,
1103 (le16_to_cpu(rl->rl_used) - index - 1) *
1104 sizeof(struct ocfs2_refcount_rec));
1105 memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1106 0, sizeof(struct ocfs2_refcount_rec));
1109 le16_add_cpu(&rl->rl_used, -1);
1111 ocfs2_refcount_rec_merge(rb, index);
1113 ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1120 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1121 struct ocfs2_caching_info *ci,
1122 struct buffer_head *ref_root_bh,
1123 struct buffer_head **ref_leaf_bh,
1124 struct ocfs2_alloc_context *meta_ac)
1127 u16 suballoc_bit_start;
1130 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1131 struct buffer_head *new_bh = NULL;
1132 struct ocfs2_refcount_block *new_rb;
1133 struct ocfs2_refcount_block *root_rb =
1134 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1136 ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1137 OCFS2_JOURNAL_ACCESS_WRITE);
1143 ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1144 &suballoc_bit_start, &num_got,
1151 new_bh = sb_getblk(sb, blkno);
1152 if (new_bh == NULL) {
1157 ocfs2_set_new_buffer_uptodate(ci, new_bh);
1159 ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1160 OCFS2_JOURNAL_ACCESS_CREATE);
1167 * Initialize ocfs2_refcount_block.
1168 * It should contain the same information as the old root.
1169 * so just memcpy it and change the corresponding field.
1171 memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1173 new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1174 new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1175 new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1176 new_rb->rf_blkno = cpu_to_le64(blkno);
1177 new_rb->rf_cpos = cpu_to_le32(0);
1178 new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1179 new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1180 ocfs2_journal_dirty(handle, new_bh);
1182 /* Now change the root. */
1183 memset(&root_rb->rf_list, 0, sb->s_blocksize -
1184 offsetof(struct ocfs2_refcount_block, rf_list));
1185 root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1186 root_rb->rf_clusters = cpu_to_le32(1);
1187 root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1188 root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1189 root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1190 root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1192 ocfs2_journal_dirty(handle, ref_root_bh);
1194 mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno,
1195 le16_to_cpu(new_rb->rf_records.rl_used));
1197 *ref_leaf_bh = new_bh;
1204 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1205 struct ocfs2_refcount_rec *next)
1207 if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1208 ocfs2_get_ref_rec_low_cpos(next))
1214 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1216 const struct ocfs2_refcount_rec *l = a, *r = b;
1217 u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1218 u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1220 if (l_cpos > r_cpos)
1222 if (l_cpos < r_cpos)
1227 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1229 const struct ocfs2_refcount_rec *l = a, *r = b;
1230 u64 l_cpos = le64_to_cpu(l->r_cpos);
1231 u64 r_cpos = le64_to_cpu(r->r_cpos);
1233 if (l_cpos > r_cpos)
1235 if (l_cpos < r_cpos)
1240 static void swap_refcount_rec(void *a, void *b, int size)
1242 struct ocfs2_refcount_rec *l = a, *r = b, tmp;
1244 tmp = *(struct ocfs2_refcount_rec *)l;
1245 *(struct ocfs2_refcount_rec *)l =
1246 *(struct ocfs2_refcount_rec *)r;
1247 *(struct ocfs2_refcount_rec *)r = tmp;
1251 * The refcount cpos are ordered by their 64bit cpos,
1252 * But we will use the low 32 bit to be the e_cpos in the b-tree.
1253 * So we need to make sure that this pos isn't intersected with others.
1255 * Note: The refcount block is already sorted by their low 32 bit cpos,
1256 * So just try the middle pos first, and we will exit when we find
1257 * the good position.
1259 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1260 u32 *split_pos, int *split_index)
1262 int num_used = le16_to_cpu(rl->rl_used);
1263 int delta, middle = num_used / 2;
1265 for (delta = 0; delta < middle; delta++) {
1266 /* Let's check delta earlier than middle */
1267 if (ocfs2_refcount_rec_no_intersect(
1268 &rl->rl_recs[middle - delta - 1],
1269 &rl->rl_recs[middle - delta])) {
1270 *split_index = middle - delta;
1274 /* For even counts, don't walk off the end */
1275 if ((middle + delta + 1) == num_used)
1278 /* Now try delta past middle */
1279 if (ocfs2_refcount_rec_no_intersect(
1280 &rl->rl_recs[middle + delta],
1281 &rl->rl_recs[middle + delta + 1])) {
1282 *split_index = middle + delta + 1;
1287 if (delta >= middle)
1290 *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1294 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1295 struct buffer_head *new_bh,
1298 int split_index = 0, num_moved, ret;
1300 struct ocfs2_refcount_block *rb =
1301 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1302 struct ocfs2_refcount_list *rl = &rb->rf_records;
1303 struct ocfs2_refcount_block *new_rb =
1304 (struct ocfs2_refcount_block *)new_bh->b_data;
1305 struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1307 mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n",
1308 (unsigned long long)ref_leaf_bh->b_blocknr,
1309 le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used));
1312 * XXX: Improvement later.
1313 * If we know all the high 32 bit cpos is the same, no need to sort.
1315 * In order to make the whole process safe, we do:
1316 * 1. sort the entries by their low 32 bit cpos first so that we can
1317 * find the split cpos easily.
1318 * 2. call ocfs2_insert_extent to insert the new refcount block.
1319 * 3. move the refcount rec to the new block.
1320 * 4. sort the entries by their 64 bit cpos.
1321 * 5. dirty the new_rb and rb.
1323 sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1324 sizeof(struct ocfs2_refcount_rec),
1325 cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1327 ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1333 new_rb->rf_cpos = cpu_to_le32(cpos);
1335 /* move refcount records starting from split_index to the new block. */
1336 num_moved = le16_to_cpu(rl->rl_used) - split_index;
1337 memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1338 num_moved * sizeof(struct ocfs2_refcount_rec));
1340 /*ok, remove the entries we just moved over to the other block. */
1341 memset(&rl->rl_recs[split_index], 0,
1342 num_moved * sizeof(struct ocfs2_refcount_rec));
1344 /* change old and new rl_used accordingly. */
1345 le16_add_cpu(&rl->rl_used, -num_moved);
1346 new_rl->rl_used = cpu_to_le32(num_moved);
1348 sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1349 sizeof(struct ocfs2_refcount_rec),
1350 cmp_refcount_rec_by_cpos, swap_refcount_rec);
1352 sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1353 sizeof(struct ocfs2_refcount_rec),
1354 cmp_refcount_rec_by_cpos, swap_refcount_rec);
1360 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1361 struct ocfs2_caching_info *ci,
1362 struct buffer_head *ref_root_bh,
1363 struct buffer_head *ref_leaf_bh,
1364 struct ocfs2_alloc_context *meta_ac)
1367 u16 suballoc_bit_start;
1368 u32 num_got, new_cpos;
1370 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1371 struct ocfs2_refcount_block *root_rb =
1372 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1373 struct buffer_head *new_bh = NULL;
1374 struct ocfs2_refcount_block *new_rb;
1375 struct ocfs2_extent_tree ref_et;
1377 BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1379 ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1380 OCFS2_JOURNAL_ACCESS_WRITE);
1386 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1387 OCFS2_JOURNAL_ACCESS_WRITE);
1393 ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1394 &suballoc_bit_start, &num_got,
1401 new_bh = sb_getblk(sb, blkno);
1402 if (new_bh == NULL) {
1407 ocfs2_set_new_buffer_uptodate(ci, new_bh);
1409 ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1410 OCFS2_JOURNAL_ACCESS_CREATE);
1416 /* Initialize ocfs2_refcount_block. */
1417 new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1418 memset(new_rb, 0, sb->s_blocksize);
1419 strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1420 new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1421 new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1422 new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1423 new_rb->rf_blkno = cpu_to_le64(blkno);
1424 new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1425 new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1426 new_rb->rf_records.rl_count =
1427 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1428 new_rb->rf_generation = root_rb->rf_generation;
1430 ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1436 ocfs2_journal_dirty(handle, ref_leaf_bh);
1437 ocfs2_journal_dirty(handle, new_bh);
1439 ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1441 mlog(0, "insert new leaf block %llu at %u\n",
1442 (unsigned long long)new_bh->b_blocknr, new_cpos);
1444 /* Insert the new leaf block with the specific offset cpos. */
1445 ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1455 static int ocfs2_expand_refcount_tree(handle_t *handle,
1456 struct ocfs2_caching_info *ci,
1457 struct buffer_head *ref_root_bh,
1458 struct buffer_head *ref_leaf_bh,
1459 struct ocfs2_alloc_context *meta_ac)
1462 struct buffer_head *expand_bh = NULL;
1464 if (ref_root_bh == ref_leaf_bh) {
1466 * the old root bh hasn't been expanded to a b-tree,
1467 * so expand it first.
1469 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1470 &expand_bh, meta_ac);
1476 expand_bh = ref_leaf_bh;
1481 /* Now add a new refcount block into the tree.*/
1482 ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1483 expand_bh, meta_ac);
1492 * Adjust the extent rec in b-tree representing ref_leaf_bh.
1494 * Only called when we have inserted a new refcount rec at index 0
1495 * which means ocfs2_extent_rec.e_cpos may need some change.
1497 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1498 struct ocfs2_caching_info *ci,
1499 struct buffer_head *ref_root_bh,
1500 struct buffer_head *ref_leaf_bh,
1501 struct ocfs2_refcount_rec *rec)
1504 u32 new_cpos, old_cpos;
1505 struct ocfs2_path *path = NULL;
1506 struct ocfs2_extent_tree et;
1507 struct ocfs2_refcount_block *rb =
1508 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1509 struct ocfs2_extent_list *el;
1511 if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1514 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1515 old_cpos = le32_to_cpu(rb->rf_cpos);
1516 new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1517 if (old_cpos <= new_cpos)
1520 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1522 path = ocfs2_new_path_from_et(&et);
1529 ret = ocfs2_find_path(ci, path, old_cpos);
1536 * 2 more credits, one for the leaf refcount block, one for
1537 * the extent block contains the extent rec.
1539 ret = ocfs2_extend_trans(handle, handle->h_buffer_credits + 2);
1545 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1546 OCFS2_JOURNAL_ACCESS_WRITE);
1552 ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1553 OCFS2_JOURNAL_ACCESS_WRITE);
1559 /* change the leaf extent block first. */
1560 el = path_leaf_el(path);
1562 for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1563 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1566 BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1568 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1570 /* change the r_cpos in the leaf block. */
1571 rb->rf_cpos = cpu_to_le32(new_cpos);
1573 ocfs2_journal_dirty(handle, path_leaf_bh(path));
1574 ocfs2_journal_dirty(handle, ref_leaf_bh);
1577 ocfs2_free_path(path);
1581 static int ocfs2_insert_refcount_rec(handle_t *handle,
1582 struct ocfs2_caching_info *ci,
1583 struct buffer_head *ref_root_bh,
1584 struct buffer_head *ref_leaf_bh,
1585 struct ocfs2_refcount_rec *rec,
1587 struct ocfs2_alloc_context *meta_ac)
1590 struct ocfs2_refcount_block *rb =
1591 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1592 struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1593 struct buffer_head *new_bh = NULL;
1595 BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1597 if (rf_list->rl_used == rf_list->rl_count) {
1598 u64 cpos = le64_to_cpu(rec->r_cpos);
1599 u32 len = le32_to_cpu(rec->r_clusters);
1601 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1602 ref_leaf_bh, meta_ac);
1608 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1609 cpos, len, NULL, &index,
1616 ref_leaf_bh = new_bh;
1617 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1618 rf_list = &rb->rf_records;
1621 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1622 OCFS2_JOURNAL_ACCESS_WRITE);
1628 if (index < le16_to_cpu(rf_list->rl_used))
1629 memmove(&rf_list->rl_recs[index + 1],
1630 &rf_list->rl_recs[index],
1631 (le16_to_cpu(rf_list->rl_used) - index) *
1632 sizeof(struct ocfs2_refcount_rec));
1634 mlog(0, "insert refcount record start %llu, len %u, count %u "
1635 "to leaf block %llu at index %d\n",
1636 (unsigned long long)le64_to_cpu(rec->r_cpos),
1637 le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount),
1638 (unsigned long long)ref_leaf_bh->b_blocknr, index);
1640 rf_list->rl_recs[index] = *rec;
1642 le16_add_cpu(&rf_list->rl_used, 1);
1644 ocfs2_refcount_rec_merge(rb, index);
1646 ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1653 ret = ocfs2_adjust_refcount_rec(handle, ci,
1665 * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1666 * This is much simple than our b-tree code.
1667 * split_rec is the new refcount rec we want to insert.
1668 * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1669 * increase refcount or decrease a refcount to non-zero).
1670 * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1671 * rec( in case we decrease a refcount to zero).
1673 static int ocfs2_split_refcount_rec(handle_t *handle,
1674 struct ocfs2_caching_info *ci,
1675 struct buffer_head *ref_root_bh,
1676 struct buffer_head *ref_leaf_bh,
1677 struct ocfs2_refcount_rec *split_rec,
1679 struct ocfs2_alloc_context *meta_ac,
1680 struct ocfs2_cached_dealloc_ctxt *dealloc)
1684 struct ocfs2_refcount_block *rb =
1685 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1686 struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1687 struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1688 struct ocfs2_refcount_rec *tail_rec = NULL;
1689 struct buffer_head *new_bh = NULL;
1691 BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1693 mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n",
1694 le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters),
1695 le64_to_cpu(split_rec->r_cpos),
1696 le32_to_cpu(split_rec->r_clusters));
1699 * If we just need to split the header or tail clusters,
1700 * no more recs are needed, just split is OK.
1701 * Otherwise we at least need one new recs.
1703 if (!split_rec->r_refcount &&
1704 (split_rec->r_cpos == orig_rec->r_cpos ||
1705 le64_to_cpu(split_rec->r_cpos) +
1706 le32_to_cpu(split_rec->r_clusters) ==
1707 le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1713 * We need one more rec if we split in the middle and the new rec have
1714 * some refcount in it.
1716 if (split_rec->r_refcount &&
1717 (split_rec->r_cpos != orig_rec->r_cpos &&
1718 le64_to_cpu(split_rec->r_cpos) +
1719 le32_to_cpu(split_rec->r_clusters) !=
1720 le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1723 /* If the leaf block don't have enough record, expand it. */
1724 if (le16_to_cpu(rf_list->rl_used) + recs_need > rf_list->rl_count) {
1725 struct ocfs2_refcount_rec tmp_rec;
1726 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1727 len = le32_to_cpu(orig_rec->r_clusters);
1728 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1729 ref_leaf_bh, meta_ac);
1736 * We have to re-get it since now cpos may be moved to
1737 * another leaf block.
1739 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1740 cpos, len, &tmp_rec, &index,
1747 ref_leaf_bh = new_bh;
1748 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1749 rf_list = &rb->rf_records;
1750 orig_rec = &rf_list->rl_recs[index];
1753 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1754 OCFS2_JOURNAL_ACCESS_WRITE);
1761 * We have calculated out how many new records we need and store
1762 * in recs_need, so spare enough space first by moving the records
1763 * after "index" to the end.
1765 if (index != le16_to_cpu(rf_list->rl_used) - 1)
1766 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1767 &rf_list->rl_recs[index + 1],
1768 (le16_to_cpu(rf_list->rl_used) - index - 1) *
1769 sizeof(struct ocfs2_refcount_rec));
1771 len = (le64_to_cpu(orig_rec->r_cpos) +
1772 le32_to_cpu(orig_rec->r_clusters)) -
1773 (le64_to_cpu(split_rec->r_cpos) +
1774 le32_to_cpu(split_rec->r_clusters));
1777 * If we have "len", the we will split in the tail and move it
1778 * to the end of the space we have just spared.
1781 tail_rec = &rf_list->rl_recs[index + recs_need];
1783 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1784 le64_add_cpu(&tail_rec->r_cpos,
1785 le32_to_cpu(tail_rec->r_clusters) - len);
1786 tail_rec->r_clusters = le32_to_cpu(len);
1790 * If the split pos isn't the same as the original one, we need to
1791 * split in the head.
1793 * Note: We have the chance that split_rec.r_refcount = 0,
1794 * recs_need = 0 and len > 0, which means we just cut the head from
1795 * the orig_rec and in that case we have done some modification in
1796 * orig_rec above, so the check for r_cpos is faked.
1798 if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1799 len = le64_to_cpu(split_rec->r_cpos) -
1800 le64_to_cpu(orig_rec->r_cpos);
1801 orig_rec->r_clusters = cpu_to_le32(len);
1805 le16_add_cpu(&rf_list->rl_used, recs_need);
1807 if (split_rec->r_refcount) {
1808 rf_list->rl_recs[index] = *split_rec;
1809 mlog(0, "insert refcount record start %llu, len %u, count %u "
1810 "to leaf block %llu at index %d\n",
1811 (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1812 le32_to_cpu(split_rec->r_clusters),
1813 le32_to_cpu(split_rec->r_refcount),
1814 (unsigned long long)ref_leaf_bh->b_blocknr, index);
1816 ocfs2_refcount_rec_merge(rb, index);
1819 ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1828 static int __ocfs2_increase_refcount(handle_t *handle,
1829 struct ocfs2_caching_info *ci,
1830 struct buffer_head *ref_root_bh,
1832 struct ocfs2_alloc_context *meta_ac,
1833 struct ocfs2_cached_dealloc_ctxt *dealloc)
1836 struct buffer_head *ref_leaf_bh = NULL;
1837 struct ocfs2_refcount_rec rec;
1838 unsigned int set_len = 0;
1840 mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n",
1841 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1842 (unsigned long long)cpos, len);
1845 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1846 cpos, len, &rec, &index,
1853 set_len = le32_to_cpu(rec.r_clusters);
1856 * Here we may meet with 3 situations:
1858 * 1. If we find an already existing record, and the length
1859 * is the same, cool, we just need to increase the r_refcount
1861 * 2. If we find a hole, just insert it with r_refcount = 1.
1862 * 3. If we are in the middle of one extent record, split
1865 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
1867 mlog(0, "increase refcount rec, start %llu, len %u, "
1868 "count %u\n", (unsigned long long)cpos, set_len,
1869 le32_to_cpu(rec.r_refcount));
1870 ret = ocfs2_change_refcount_rec(handle, ci,
1871 ref_leaf_bh, index, 1);
1876 } else if (!rec.r_refcount) {
1877 rec.r_refcount = cpu_to_le32(1);
1879 mlog(0, "insert refcount rec, start %llu, len %u\n",
1880 (unsigned long long)le64_to_cpu(rec.r_cpos),
1882 ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
1884 &rec, index, meta_ac);
1890 set_len = min((u64)(cpos + len),
1891 le64_to_cpu(rec.r_cpos) + set_len) - cpos;
1892 rec.r_cpos = cpu_to_le64(cpos);
1893 rec.r_clusters = cpu_to_le32(set_len);
1894 le32_add_cpu(&rec.r_refcount, 1);
1896 mlog(0, "split refcount rec, start %llu, "
1897 "len %u, count %u\n",
1898 (unsigned long long)le64_to_cpu(rec.r_cpos),
1899 set_len, le32_to_cpu(rec.r_refcount));
1900 ret = ocfs2_split_refcount_rec(handle, ci,
1901 ref_root_bh, ref_leaf_bh,
1912 brelse(ref_leaf_bh);
1917 brelse(ref_leaf_bh);
1921 static int ocfs2_remove_refcount_extent(handle_t *handle,
1922 struct ocfs2_caching_info *ci,
1923 struct buffer_head *ref_root_bh,
1924 struct buffer_head *ref_leaf_bh,
1925 struct ocfs2_alloc_context *meta_ac,
1926 struct ocfs2_cached_dealloc_ctxt *dealloc)
1929 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1930 struct ocfs2_refcount_block *rb =
1931 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1932 struct ocfs2_extent_tree et;
1934 BUG_ON(rb->rf_records.rl_used);
1936 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1937 ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
1938 1, meta_ac, dealloc);
1944 ocfs2_remove_from_cache(ci, ref_leaf_bh);
1947 * add the freed block to the dealloc so that it will be freed
1948 * when we run dealloc.
1950 ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
1951 le16_to_cpu(rb->rf_suballoc_slot),
1952 le64_to_cpu(rb->rf_blkno),
1953 le16_to_cpu(rb->rf_suballoc_bit));
1959 ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1960 OCFS2_JOURNAL_ACCESS_WRITE);
1966 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1968 le32_add_cpu(&rb->rf_clusters, -1);
1971 * check whether we need to restore the root refcount block if
1972 * there is no leaf extent block at atll.
1974 if (!rb->rf_list.l_next_free_rec) {
1975 BUG_ON(rb->rf_clusters);
1977 mlog(0, "reset refcount tree root %llu to be a record block.\n",
1978 (unsigned long long)ref_root_bh->b_blocknr);
1983 memset(&rb->rf_records, 0, sb->s_blocksize -
1984 offsetof(struct ocfs2_refcount_block, rf_records));
1985 rb->rf_records.rl_count =
1986 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1989 ocfs2_journal_dirty(handle, ref_root_bh);
1995 static int ocfs2_decrease_refcount_rec(handle_t *handle,
1996 struct ocfs2_caching_info *ci,
1997 struct buffer_head *ref_root_bh,
1998 struct buffer_head *ref_leaf_bh,
1999 int index, u64 cpos, unsigned int len,
2000 struct ocfs2_alloc_context *meta_ac,
2001 struct ocfs2_cached_dealloc_ctxt *dealloc)
2004 struct ocfs2_refcount_block *rb =
2005 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2006 struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2008 BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2010 le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2012 if (cpos == le64_to_cpu(rec->r_cpos) &&
2013 len == le32_to_cpu(rec->r_clusters))
2014 ret = ocfs2_change_refcount_rec(handle, ci,
2015 ref_leaf_bh, index, -1);
2017 struct ocfs2_refcount_rec split = *rec;
2018 split.r_cpos = cpu_to_le64(cpos);
2019 split.r_clusters = cpu_to_le32(len);
2021 le32_add_cpu(&split.r_refcount, -1);
2023 mlog(0, "split refcount rec, start %llu, "
2024 "len %u, count %u, original start %llu, len %u\n",
2025 (unsigned long long)le64_to_cpu(split.r_cpos),
2026 len, le32_to_cpu(split.r_refcount),
2027 (unsigned long long)le64_to_cpu(rec->r_cpos),
2028 le32_to_cpu(rec->r_clusters));
2029 ret = ocfs2_split_refcount_rec(handle, ci,
2030 ref_root_bh, ref_leaf_bh,
2040 /* Remove the leaf refcount block if it contains no refcount record. */
2041 if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2042 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2043 ref_leaf_bh, meta_ac,
2053 static int __ocfs2_decrease_refcount(handle_t *handle,
2054 struct ocfs2_caching_info *ci,
2055 struct buffer_head *ref_root_bh,
2057 struct ocfs2_alloc_context *meta_ac,
2058 struct ocfs2_cached_dealloc_ctxt *dealloc)
2060 int ret = 0, index = 0;
2061 struct ocfs2_refcount_rec rec;
2062 unsigned int r_count = 0, r_len;
2063 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2064 struct buffer_head *ref_leaf_bh = NULL;
2066 mlog(0, "Tree owner %llu, decrease refcount start %llu, len %u\n",
2067 (unsigned long long)ocfs2_metadata_cache_owner(ci),
2068 (unsigned long long)cpos, len);
2071 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2072 cpos, len, &rec, &index,
2079 r_count = le32_to_cpu(rec.r_refcount);
2080 BUG_ON(r_count == 0);
2082 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2083 le32_to_cpu(rec.r_clusters)) - cpos;
2085 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2094 if (le32_to_cpu(rec.r_refcount) == 1) {
2095 ret = ocfs2_cache_cluster_dealloc(dealloc,
2096 ocfs2_clusters_to_blocks(sb, cpos),
2106 brelse(ref_leaf_bh);
2111 brelse(ref_leaf_bh);
2115 /* Caller must hold refcount tree lock. */
2116 int ocfs2_decrease_refcount(struct inode *inode,
2117 handle_t *handle, u32 cpos, u32 len,
2118 struct ocfs2_alloc_context *meta_ac,
2119 struct ocfs2_cached_dealloc_ctxt *dealloc)
2123 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2124 struct buffer_head *ref_root_bh = NULL;
2125 struct ocfs2_refcount_tree *tree;
2127 BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2129 ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2135 ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2141 ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2148 ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2149 cpos, len, meta_ac, dealloc);
2153 brelse(ref_root_bh);