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"
38 #include <linux/bio.h>
39 #include <linux/blkdev.h>
40 #include <linux/gfp.h>
41 #include <linux/slab.h>
42 #include <linux/writeback.h>
43 #include <linux/pagevec.h>
44 #include <linux/swap.h>
46 struct ocfs2_cow_context {
50 struct ocfs2_extent_tree data_et;
51 struct ocfs2_refcount_tree *ref_tree;
52 struct buffer_head *ref_root_bh;
53 struct ocfs2_alloc_context *meta_ac;
54 struct ocfs2_alloc_context *data_ac;
55 struct ocfs2_cached_dealloc_ctxt dealloc;
57 struct ocfs2_post_refcount *post_refcount;
59 int (*get_clusters)(struct ocfs2_cow_context *context,
60 u32 v_cluster, u32 *p_cluster,
62 unsigned int *extent_flags);
63 int (*cow_duplicate_clusters)(handle_t *handle,
64 struct ocfs2_cow_context *context,
65 u32 cpos, u32 old_cluster,
66 u32 new_cluster, u32 new_len);
69 static inline struct ocfs2_refcount_tree *
70 cache_info_to_refcount(struct ocfs2_caching_info *ci)
72 return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
75 static int ocfs2_validate_refcount_block(struct super_block *sb,
76 struct buffer_head *bh)
79 struct ocfs2_refcount_block *rb =
80 (struct ocfs2_refcount_block *)bh->b_data;
82 mlog(0, "Validating refcount block %llu\n",
83 (unsigned long long)bh->b_blocknr);
85 BUG_ON(!buffer_uptodate(bh));
88 * If the ecc fails, we return the error but otherwise
89 * leave the filesystem running. We know any error is
90 * local to this block.
92 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check);
94 mlog(ML_ERROR, "Checksum failed for refcount block %llu\n",
95 (unsigned long long)bh->b_blocknr);
100 if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) {
102 "Refcount block #%llu has bad signature %.*s",
103 (unsigned long long)bh->b_blocknr, 7,
108 if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
110 "Refcount block #%llu has an invalid rf_blkno "
112 (unsigned long long)bh->b_blocknr,
113 (unsigned long long)le64_to_cpu(rb->rf_blkno));
117 if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
119 "Refcount block #%llu has an invalid "
120 "rf_fs_generation of #%u",
121 (unsigned long long)bh->b_blocknr,
122 le32_to_cpu(rb->rf_fs_generation));
129 static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
131 struct buffer_head **bh)
134 struct buffer_head *tmp = *bh;
136 rc = ocfs2_read_block(ci, rb_blkno, &tmp,
137 ocfs2_validate_refcount_block);
139 /* If ocfs2_read_block() got us a new bh, pass it up. */
146 static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
148 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
153 static struct super_block *
154 ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
156 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
161 static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
163 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
165 spin_lock(&rf->rf_lock);
168 static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
170 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
172 spin_unlock(&rf->rf_lock);
175 static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
177 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
179 mutex_lock(&rf->rf_io_mutex);
182 static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
184 struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
186 mutex_unlock(&rf->rf_io_mutex);
189 static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
190 .co_owner = ocfs2_refcount_cache_owner,
191 .co_get_super = ocfs2_refcount_cache_get_super,
192 .co_cache_lock = ocfs2_refcount_cache_lock,
193 .co_cache_unlock = ocfs2_refcount_cache_unlock,
194 .co_io_lock = ocfs2_refcount_cache_io_lock,
195 .co_io_unlock = ocfs2_refcount_cache_io_unlock,
198 static struct ocfs2_refcount_tree *
199 ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
201 struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
202 struct ocfs2_refcount_tree *tree = NULL;
205 tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
207 if (blkno < tree->rf_blkno)
209 else if (blkno > tree->rf_blkno)
218 /* osb_lock is already locked. */
219 static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
220 struct ocfs2_refcount_tree *new)
222 u64 rf_blkno = new->rf_blkno;
223 struct rb_node *parent = NULL;
224 struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
225 struct ocfs2_refcount_tree *tmp;
230 tmp = rb_entry(parent, struct ocfs2_refcount_tree,
233 if (rf_blkno < tmp->rf_blkno)
235 else if (rf_blkno > tmp->rf_blkno)
238 /* This should never happen! */
239 mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
240 (unsigned long long)rf_blkno);
245 rb_link_node(&new->rf_node, parent, p);
246 rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
249 static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
251 ocfs2_metadata_cache_exit(&tree->rf_ci);
252 ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
253 ocfs2_lock_res_free(&tree->rf_lockres);
258 ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
259 struct ocfs2_refcount_tree *tree)
261 rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
262 if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
263 osb->osb_ref_tree_lru = NULL;
266 static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
267 struct ocfs2_refcount_tree *tree)
269 spin_lock(&osb->osb_lock);
270 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
271 spin_unlock(&osb->osb_lock);
274 void ocfs2_kref_remove_refcount_tree(struct kref *kref)
276 struct ocfs2_refcount_tree *tree =
277 container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
279 ocfs2_free_refcount_tree(tree);
283 ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
285 kref_get(&tree->rf_getcnt);
289 ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
291 kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
294 static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
295 struct super_block *sb)
297 ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
298 mutex_init(&new->rf_io_mutex);
300 spin_lock_init(&new->rf_lock);
303 static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
304 struct ocfs2_refcount_tree *new,
305 u64 rf_blkno, u32 generation)
307 init_rwsem(&new->rf_sem);
308 ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
309 rf_blkno, generation);
312 static struct ocfs2_refcount_tree*
313 ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
315 struct ocfs2_refcount_tree *new;
317 new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
321 new->rf_blkno = rf_blkno;
322 kref_init(&new->rf_getcnt);
323 ocfs2_init_refcount_tree_ci(new, osb->sb);
328 static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
329 struct ocfs2_refcount_tree **ret_tree)
332 struct ocfs2_refcount_tree *tree, *new = NULL;
333 struct buffer_head *ref_root_bh = NULL;
334 struct ocfs2_refcount_block *ref_rb;
336 spin_lock(&osb->osb_lock);
337 if (osb->osb_ref_tree_lru &&
338 osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
339 tree = osb->osb_ref_tree_lru;
341 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
345 spin_unlock(&osb->osb_lock);
347 new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
354 * We need the generation to create the refcount tree lock and since
355 * it isn't changed during the tree modification, we are safe here to
356 * read without protection.
357 * We also have to purge the cache after we create the lock since the
358 * refcount block may have the stale data. It can only be trusted when
359 * we hold the refcount lock.
361 ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
364 ocfs2_metadata_cache_exit(&new->rf_ci);
369 ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
370 new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
371 ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
373 ocfs2_metadata_cache_purge(&new->rf_ci);
375 spin_lock(&osb->osb_lock);
376 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
380 ocfs2_insert_refcount_tree(osb, new);
388 osb->osb_ref_tree_lru = tree;
390 spin_unlock(&osb->osb_lock);
393 ocfs2_free_refcount_tree(new);
399 static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
402 struct buffer_head *di_bh = NULL;
403 struct ocfs2_dinode *di;
405 ret = ocfs2_read_inode_block(inode, &di_bh);
411 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
413 di = (struct ocfs2_dinode *)di_bh->b_data;
414 *ref_blkno = le64_to_cpu(di->i_refcount_loc);
420 static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
421 struct ocfs2_refcount_tree *tree, int rw)
425 ret = ocfs2_refcount_lock(tree, rw);
432 down_write(&tree->rf_sem);
434 down_read(&tree->rf_sem);
441 * Lock the refcount tree pointed by ref_blkno and return the tree.
442 * In most case, we lock the tree and read the refcount block.
443 * So read it here if the caller really needs it.
445 * If the tree has been re-created by other node, it will free the
446 * old one and re-create it.
448 int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
449 u64 ref_blkno, int rw,
450 struct ocfs2_refcount_tree **ret_tree,
451 struct buffer_head **ref_bh)
453 int ret, delete_tree = 0;
454 struct ocfs2_refcount_tree *tree = NULL;
455 struct buffer_head *ref_root_bh = NULL;
456 struct ocfs2_refcount_block *rb;
459 ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
465 ocfs2_refcount_tree_get(tree);
467 ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
470 ocfs2_refcount_tree_put(tree);
474 ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
478 ocfs2_unlock_refcount_tree(osb, tree, rw);
479 ocfs2_refcount_tree_put(tree);
483 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
485 * If the refcount block has been freed and re-created, we may need
486 * to recreate the refcount tree also.
488 * Here we just remove the tree from the rb-tree, and the last
489 * kref holder will unlock and delete this refcount_tree.
490 * Then we goto "again" and ocfs2_get_refcount_tree will create
491 * the new refcount tree for us.
493 if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
494 if (!tree->rf_removed) {
495 ocfs2_erase_refcount_tree_from_list(osb, tree);
496 tree->rf_removed = 1;
500 ocfs2_unlock_refcount_tree(osb, tree, rw);
502 * We get an extra reference when we create the refcount
503 * tree, so another put will destroy it.
506 ocfs2_refcount_tree_put(tree);
514 *ref_bh = ref_root_bh;
522 int ocfs2_lock_refcount_tree_by_inode(struct inode *inode, int rw,
523 struct ocfs2_refcount_tree **ret_tree,
524 struct buffer_head **ref_bh)
529 ret = ocfs2_get_refcount_block(inode, &ref_blkno);
535 return ocfs2_lock_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno,
536 rw, ret_tree, ref_bh);
539 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
540 struct ocfs2_refcount_tree *tree, int rw)
543 up_write(&tree->rf_sem);
545 up_read(&tree->rf_sem);
547 ocfs2_refcount_unlock(tree, rw);
548 ocfs2_refcount_tree_put(tree);
551 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
553 struct rb_node *node;
554 struct ocfs2_refcount_tree *tree;
555 struct rb_root *root = &osb->osb_rf_lock_tree;
557 while ((node = rb_last(root)) != NULL) {
558 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
560 mlog(0, "Purge tree %llu\n",
561 (unsigned long long) tree->rf_blkno);
563 rb_erase(&tree->rf_node, root);
564 ocfs2_free_refcount_tree(tree);
569 * Create a refcount tree for an inode.
570 * We take for granted that the inode is already locked.
572 static int ocfs2_create_refcount_tree(struct inode *inode,
573 struct buffer_head *di_bh)
576 handle_t *handle = NULL;
577 struct ocfs2_alloc_context *meta_ac = NULL;
578 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
579 struct ocfs2_inode_info *oi = OCFS2_I(inode);
580 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
581 struct buffer_head *new_bh = NULL;
582 struct ocfs2_refcount_block *rb;
583 struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
584 u16 suballoc_bit_start;
588 BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
590 mlog(0, "create tree for inode %lu\n", inode->i_ino);
592 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
598 handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
599 if (IS_ERR(handle)) {
600 ret = PTR_ERR(handle);
605 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
606 OCFS2_JOURNAL_ACCESS_WRITE);
612 ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1,
613 &suballoc_bit_start, &num_got,
620 new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
627 new_bh = sb_getblk(inode->i_sb, first_blkno);
628 ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
630 ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
631 OCFS2_JOURNAL_ACCESS_CREATE);
637 /* Initialize ocfs2_refcount_block. */
638 rb = (struct ocfs2_refcount_block *)new_bh->b_data;
639 memset(rb, 0, inode->i_sb->s_blocksize);
640 strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
641 rb->rf_suballoc_slot = cpu_to_le16(osb->slot_num);
642 rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
643 rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
644 rb->rf_blkno = cpu_to_le64(first_blkno);
645 rb->rf_count = cpu_to_le32(1);
646 rb->rf_records.rl_count =
647 cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
648 spin_lock(&osb->osb_lock);
649 rb->rf_generation = osb->s_next_generation++;
650 spin_unlock(&osb->osb_lock);
652 ocfs2_journal_dirty(handle, new_bh);
654 spin_lock(&oi->ip_lock);
655 oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
656 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
657 di->i_refcount_loc = cpu_to_le64(first_blkno);
658 spin_unlock(&oi->ip_lock);
660 mlog(0, "created tree for inode %lu, refblock %llu\n",
661 inode->i_ino, (unsigned long long)first_blkno);
663 ocfs2_journal_dirty(handle, di_bh);
666 * We have to init the tree lock here since it will use
667 * the generation number to create it.
669 new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
670 ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
671 new_tree->rf_generation);
673 spin_lock(&osb->osb_lock);
674 tree = ocfs2_find_refcount_tree(osb, first_blkno);
677 * We've just created a new refcount tree in this block. If
678 * we found a refcount tree on the ocfs2_super, it must be
679 * one we just deleted. We free the old tree before
680 * inserting the new tree.
682 BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
684 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
685 ocfs2_insert_refcount_tree(osb, new_tree);
686 spin_unlock(&osb->osb_lock);
689 ocfs2_refcount_tree_put(tree);
692 ocfs2_commit_trans(osb, handle);
696 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
702 ocfs2_free_alloc_context(meta_ac);
707 static int ocfs2_set_refcount_tree(struct inode *inode,
708 struct buffer_head *di_bh,
712 handle_t *handle = NULL;
713 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
714 struct ocfs2_inode_info *oi = OCFS2_I(inode);
715 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
716 struct buffer_head *ref_root_bh = NULL;
717 struct ocfs2_refcount_block *rb;
718 struct ocfs2_refcount_tree *ref_tree;
720 BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
722 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
723 &ref_tree, &ref_root_bh);
729 handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
730 if (IS_ERR(handle)) {
731 ret = PTR_ERR(handle);
736 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
737 OCFS2_JOURNAL_ACCESS_WRITE);
743 ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
744 OCFS2_JOURNAL_ACCESS_WRITE);
750 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
751 le32_add_cpu(&rb->rf_count, 1);
753 ocfs2_journal_dirty(handle, ref_root_bh);
755 spin_lock(&oi->ip_lock);
756 oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
757 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
758 di->i_refcount_loc = cpu_to_le64(refcount_loc);
759 spin_unlock(&oi->ip_lock);
760 ocfs2_journal_dirty(handle, di_bh);
763 ocfs2_commit_trans(osb, handle);
765 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
771 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
773 int ret, delete_tree = 0;
774 handle_t *handle = NULL;
775 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
776 struct ocfs2_inode_info *oi = OCFS2_I(inode);
777 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
778 struct ocfs2_refcount_block *rb;
779 struct inode *alloc_inode = NULL;
780 struct buffer_head *alloc_bh = NULL;
781 struct buffer_head *blk_bh = NULL;
782 struct ocfs2_refcount_tree *ref_tree;
783 int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
784 u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
787 if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
791 ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
797 rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
800 * If we are the last user, we need to free the block.
801 * So lock the allocator ahead.
803 if (le32_to_cpu(rb->rf_count) == 1) {
804 blk = le64_to_cpu(rb->rf_blkno);
805 bit = le16_to_cpu(rb->rf_suballoc_bit);
806 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
808 alloc_inode = ocfs2_get_system_file_inode(osb,
809 EXTENT_ALLOC_SYSTEM_INODE,
810 le16_to_cpu(rb->rf_suballoc_slot));
816 mutex_lock(&alloc_inode->i_mutex);
818 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
824 credits += OCFS2_SUBALLOC_FREE;
827 handle = ocfs2_start_trans(osb, credits);
828 if (IS_ERR(handle)) {
829 ret = PTR_ERR(handle);
834 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
835 OCFS2_JOURNAL_ACCESS_WRITE);
841 ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
842 OCFS2_JOURNAL_ACCESS_WRITE);
848 spin_lock(&oi->ip_lock);
849 oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
850 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
851 di->i_refcount_loc = 0;
852 spin_unlock(&oi->ip_lock);
853 ocfs2_journal_dirty(handle, di_bh);
855 le32_add_cpu(&rb->rf_count , -1);
856 ocfs2_journal_dirty(handle, blk_bh);
860 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
861 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
862 alloc_bh, bit, bg_blkno, 1);
868 ocfs2_commit_trans(osb, handle);
871 ocfs2_inode_unlock(alloc_inode, 1);
876 mutex_unlock(&alloc_inode->i_mutex);
880 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
882 ocfs2_refcount_tree_put(ref_tree);
888 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
889 struct buffer_head *ref_leaf_bh,
890 u64 cpos, unsigned int len,
891 struct ocfs2_refcount_rec *ret_rec,
895 struct ocfs2_refcount_block *rb =
896 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
897 struct ocfs2_refcount_rec *rec = NULL;
899 for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
900 rec = &rb->rf_records.rl_recs[i];
902 if (le64_to_cpu(rec->r_cpos) +
903 le32_to_cpu(rec->r_clusters) <= cpos)
905 else if (le64_to_cpu(rec->r_cpos) > cpos)
908 /* ok, cpos fail in this rec. Just return. */
915 /* We meet with a hole here, so fake the rec. */
916 ret_rec->r_cpos = cpu_to_le64(cpos);
917 ret_rec->r_refcount = 0;
918 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
919 le64_to_cpu(rec->r_cpos) < cpos + len)
920 ret_rec->r_clusters =
921 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
923 ret_rec->r_clusters = cpu_to_le32(len);
931 * Try to remove refcount tree. The mechanism is:
932 * 1) Check whether i_clusters == 0, if no, exit.
933 * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
934 * 3) Check whether we have inline xattr stored outside, if yes, exit.
935 * 4) Remove the tree.
937 int ocfs2_try_remove_refcount_tree(struct inode *inode,
938 struct buffer_head *di_bh)
941 struct ocfs2_inode_info *oi = OCFS2_I(inode);
942 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
944 down_write(&oi->ip_xattr_sem);
945 down_write(&oi->ip_alloc_sem);
950 if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
953 if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
954 ocfs2_has_inline_xattr_value_outside(inode, di))
957 ret = ocfs2_remove_refcount_tree(inode, di_bh);
961 up_write(&oi->ip_alloc_sem);
962 up_write(&oi->ip_xattr_sem);
967 * Given a cpos and len, try to find the refcount record which contains cpos.
968 * 1. If cpos can be found in one refcount record, return the record.
969 * 2. If cpos can't be found, return a fake record which start from cpos
970 * and end at a small value between cpos+len and start of the next record.
971 * This fake record has r_refcount = 0.
973 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
974 struct buffer_head *ref_root_bh,
975 u64 cpos, unsigned int len,
976 struct ocfs2_refcount_rec *ret_rec,
978 struct buffer_head **ret_bh)
980 int ret = 0, i, found;
982 struct ocfs2_extent_list *el;
983 struct ocfs2_extent_rec *tmp, *rec = NULL;
984 struct ocfs2_extent_block *eb;
985 struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
986 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
987 struct ocfs2_refcount_block *rb =
988 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
990 if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
991 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
993 *ret_bh = ref_root_bh;
999 low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1001 if (el->l_tree_depth) {
1002 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1008 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1011 if (el->l_tree_depth) {
1013 "refcount tree %llu has non zero tree "
1014 "depth in leaf btree tree block %llu\n",
1015 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1016 (unsigned long long)eb_bh->b_blocknr);
1023 for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1024 rec = &el->l_recs[i];
1026 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1032 /* adjust len when we have ocfs2_extent_rec after it. */
1033 if (found && i < le16_to_cpu(el->l_next_free_rec) - 1) {
1034 tmp = &el->l_recs[i+1];
1036 if (le32_to_cpu(tmp->e_cpos) < cpos + len)
1037 len = le32_to_cpu(tmp->e_cpos) - cpos;
1040 ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1047 ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1049 *ret_bh = ref_leaf_bh;
1055 enum ocfs2_ref_rec_contig {
1056 REF_CONTIG_NONE = 0,
1059 REF_CONTIG_LEFTRIGHT,
1062 static enum ocfs2_ref_rec_contig
1063 ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1066 if ((rb->rf_records.rl_recs[index].r_refcount ==
1067 rb->rf_records.rl_recs[index + 1].r_refcount) &&
1068 (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1069 le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1070 le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1071 return REF_CONTIG_RIGHT;
1073 return REF_CONTIG_NONE;
1076 static enum ocfs2_ref_rec_contig
1077 ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1080 enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1082 if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1083 ret = ocfs2_refcount_rec_adjacent(rb, index);
1086 enum ocfs2_ref_rec_contig tmp;
1088 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1090 if (tmp == REF_CONTIG_RIGHT) {
1091 if (ret == REF_CONTIG_RIGHT)
1092 ret = REF_CONTIG_LEFTRIGHT;
1094 ret = REF_CONTIG_LEFT;
1101 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1104 BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1105 rb->rf_records.rl_recs[index+1].r_refcount);
1107 le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1108 le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1110 if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1111 memmove(&rb->rf_records.rl_recs[index + 1],
1112 &rb->rf_records.rl_recs[index + 2],
1113 sizeof(struct ocfs2_refcount_rec) *
1114 (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1116 memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1117 0, sizeof(struct ocfs2_refcount_rec));
1118 le16_add_cpu(&rb->rf_records.rl_used, -1);
1122 * Merge the refcount rec if we are contiguous with the adjacent recs.
1124 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1127 enum ocfs2_ref_rec_contig contig =
1128 ocfs2_refcount_rec_contig(rb, index);
1130 if (contig == REF_CONTIG_NONE)
1133 if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1138 ocfs2_rotate_refcount_rec_left(rb, index);
1140 if (contig == REF_CONTIG_LEFTRIGHT)
1141 ocfs2_rotate_refcount_rec_left(rb, index);
1145 * Change the refcount indexed by "index" in ref_bh.
1146 * If refcount reaches 0, remove it.
1148 static int ocfs2_change_refcount_rec(handle_t *handle,
1149 struct ocfs2_caching_info *ci,
1150 struct buffer_head *ref_leaf_bh,
1151 int index, int merge, int change)
1154 struct ocfs2_refcount_block *rb =
1155 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1156 struct ocfs2_refcount_list *rl = &rb->rf_records;
1157 struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1159 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1160 OCFS2_JOURNAL_ACCESS_WRITE);
1166 mlog(0, "change index %d, old count %u, change %d\n", index,
1167 le32_to_cpu(rec->r_refcount), change);
1168 le32_add_cpu(&rec->r_refcount, change);
1170 if (!rec->r_refcount) {
1171 if (index != le16_to_cpu(rl->rl_used) - 1) {
1172 memmove(rec, rec + 1,
1173 (le16_to_cpu(rl->rl_used) - index - 1) *
1174 sizeof(struct ocfs2_refcount_rec));
1175 memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1176 0, sizeof(struct ocfs2_refcount_rec));
1179 le16_add_cpu(&rl->rl_used, -1);
1181 ocfs2_refcount_rec_merge(rb, index);
1183 ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1190 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1191 struct ocfs2_caching_info *ci,
1192 struct buffer_head *ref_root_bh,
1193 struct buffer_head **ref_leaf_bh,
1194 struct ocfs2_alloc_context *meta_ac)
1197 u16 suballoc_bit_start;
1200 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1201 struct buffer_head *new_bh = NULL;
1202 struct ocfs2_refcount_block *new_rb;
1203 struct ocfs2_refcount_block *root_rb =
1204 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1206 ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1207 OCFS2_JOURNAL_ACCESS_WRITE);
1213 ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1214 &suballoc_bit_start, &num_got,
1221 new_bh = sb_getblk(sb, blkno);
1222 if (new_bh == NULL) {
1227 ocfs2_set_new_buffer_uptodate(ci, new_bh);
1229 ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1230 OCFS2_JOURNAL_ACCESS_CREATE);
1237 * Initialize ocfs2_refcount_block.
1238 * It should contain the same information as the old root.
1239 * so just memcpy it and change the corresponding field.
1241 memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1243 new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1244 new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1245 new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1246 new_rb->rf_blkno = cpu_to_le64(blkno);
1247 new_rb->rf_cpos = cpu_to_le32(0);
1248 new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1249 new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1250 ocfs2_journal_dirty(handle, new_bh);
1252 /* Now change the root. */
1253 memset(&root_rb->rf_list, 0, sb->s_blocksize -
1254 offsetof(struct ocfs2_refcount_block, rf_list));
1255 root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1256 root_rb->rf_clusters = cpu_to_le32(1);
1257 root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1258 root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1259 root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1260 root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1262 ocfs2_journal_dirty(handle, ref_root_bh);
1264 mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno,
1265 le16_to_cpu(new_rb->rf_records.rl_used));
1267 *ref_leaf_bh = new_bh;
1274 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1275 struct ocfs2_refcount_rec *next)
1277 if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1278 ocfs2_get_ref_rec_low_cpos(next))
1284 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1286 const struct ocfs2_refcount_rec *l = a, *r = b;
1287 u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1288 u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1290 if (l_cpos > r_cpos)
1292 if (l_cpos < r_cpos)
1297 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1299 const struct ocfs2_refcount_rec *l = a, *r = b;
1300 u64 l_cpos = le64_to_cpu(l->r_cpos);
1301 u64 r_cpos = le64_to_cpu(r->r_cpos);
1303 if (l_cpos > r_cpos)
1305 if (l_cpos < r_cpos)
1310 static void swap_refcount_rec(void *a, void *b, int size)
1312 struct ocfs2_refcount_rec *l = a, *r = b, tmp;
1314 tmp = *(struct ocfs2_refcount_rec *)l;
1315 *(struct ocfs2_refcount_rec *)l =
1316 *(struct ocfs2_refcount_rec *)r;
1317 *(struct ocfs2_refcount_rec *)r = tmp;
1321 * The refcount cpos are ordered by their 64bit cpos,
1322 * But we will use the low 32 bit to be the e_cpos in the b-tree.
1323 * So we need to make sure that this pos isn't intersected with others.
1325 * Note: The refcount block is already sorted by their low 32 bit cpos,
1326 * So just try the middle pos first, and we will exit when we find
1327 * the good position.
1329 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1330 u32 *split_pos, int *split_index)
1332 int num_used = le16_to_cpu(rl->rl_used);
1333 int delta, middle = num_used / 2;
1335 for (delta = 0; delta < middle; delta++) {
1336 /* Let's check delta earlier than middle */
1337 if (ocfs2_refcount_rec_no_intersect(
1338 &rl->rl_recs[middle - delta - 1],
1339 &rl->rl_recs[middle - delta])) {
1340 *split_index = middle - delta;
1344 /* For even counts, don't walk off the end */
1345 if ((middle + delta + 1) == num_used)
1348 /* Now try delta past middle */
1349 if (ocfs2_refcount_rec_no_intersect(
1350 &rl->rl_recs[middle + delta],
1351 &rl->rl_recs[middle + delta + 1])) {
1352 *split_index = middle + delta + 1;
1357 if (delta >= middle)
1360 *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1364 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1365 struct buffer_head *new_bh,
1368 int split_index = 0, num_moved, ret;
1370 struct ocfs2_refcount_block *rb =
1371 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1372 struct ocfs2_refcount_list *rl = &rb->rf_records;
1373 struct ocfs2_refcount_block *new_rb =
1374 (struct ocfs2_refcount_block *)new_bh->b_data;
1375 struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1377 mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n",
1378 (unsigned long long)ref_leaf_bh->b_blocknr,
1379 le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used));
1382 * XXX: Improvement later.
1383 * If we know all the high 32 bit cpos is the same, no need to sort.
1385 * In order to make the whole process safe, we do:
1386 * 1. sort the entries by their low 32 bit cpos first so that we can
1387 * find the split cpos easily.
1388 * 2. call ocfs2_insert_extent to insert the new refcount block.
1389 * 3. move the refcount rec to the new block.
1390 * 4. sort the entries by their 64 bit cpos.
1391 * 5. dirty the new_rb and rb.
1393 sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1394 sizeof(struct ocfs2_refcount_rec),
1395 cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1397 ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1403 new_rb->rf_cpos = cpu_to_le32(cpos);
1405 /* move refcount records starting from split_index to the new block. */
1406 num_moved = le16_to_cpu(rl->rl_used) - split_index;
1407 memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1408 num_moved * sizeof(struct ocfs2_refcount_rec));
1410 /*ok, remove the entries we just moved over to the other block. */
1411 memset(&rl->rl_recs[split_index], 0,
1412 num_moved * sizeof(struct ocfs2_refcount_rec));
1414 /* change old and new rl_used accordingly. */
1415 le16_add_cpu(&rl->rl_used, -num_moved);
1416 new_rl->rl_used = cpu_to_le32(num_moved);
1418 sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1419 sizeof(struct ocfs2_refcount_rec),
1420 cmp_refcount_rec_by_cpos, swap_refcount_rec);
1422 sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1423 sizeof(struct ocfs2_refcount_rec),
1424 cmp_refcount_rec_by_cpos, swap_refcount_rec);
1430 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1431 struct ocfs2_caching_info *ci,
1432 struct buffer_head *ref_root_bh,
1433 struct buffer_head *ref_leaf_bh,
1434 struct ocfs2_alloc_context *meta_ac)
1437 u16 suballoc_bit_start;
1438 u32 num_got, new_cpos;
1440 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1441 struct ocfs2_refcount_block *root_rb =
1442 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1443 struct buffer_head *new_bh = NULL;
1444 struct ocfs2_refcount_block *new_rb;
1445 struct ocfs2_extent_tree ref_et;
1447 BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1449 ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1450 OCFS2_JOURNAL_ACCESS_WRITE);
1456 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1457 OCFS2_JOURNAL_ACCESS_WRITE);
1463 ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1464 &suballoc_bit_start, &num_got,
1471 new_bh = sb_getblk(sb, blkno);
1472 if (new_bh == NULL) {
1477 ocfs2_set_new_buffer_uptodate(ci, new_bh);
1479 ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1480 OCFS2_JOURNAL_ACCESS_CREATE);
1486 /* Initialize ocfs2_refcount_block. */
1487 new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1488 memset(new_rb, 0, sb->s_blocksize);
1489 strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1490 new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1491 new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1492 new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1493 new_rb->rf_blkno = cpu_to_le64(blkno);
1494 new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1495 new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1496 new_rb->rf_records.rl_count =
1497 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1498 new_rb->rf_generation = root_rb->rf_generation;
1500 ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1506 ocfs2_journal_dirty(handle, ref_leaf_bh);
1507 ocfs2_journal_dirty(handle, new_bh);
1509 ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1511 mlog(0, "insert new leaf block %llu at %u\n",
1512 (unsigned long long)new_bh->b_blocknr, new_cpos);
1514 /* Insert the new leaf block with the specific offset cpos. */
1515 ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1525 static int ocfs2_expand_refcount_tree(handle_t *handle,
1526 struct ocfs2_caching_info *ci,
1527 struct buffer_head *ref_root_bh,
1528 struct buffer_head *ref_leaf_bh,
1529 struct ocfs2_alloc_context *meta_ac)
1532 struct buffer_head *expand_bh = NULL;
1534 if (ref_root_bh == ref_leaf_bh) {
1536 * the old root bh hasn't been expanded to a b-tree,
1537 * so expand it first.
1539 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1540 &expand_bh, meta_ac);
1546 expand_bh = ref_leaf_bh;
1551 /* Now add a new refcount block into the tree.*/
1552 ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1553 expand_bh, meta_ac);
1562 * Adjust the extent rec in b-tree representing ref_leaf_bh.
1564 * Only called when we have inserted a new refcount rec at index 0
1565 * which means ocfs2_extent_rec.e_cpos may need some change.
1567 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1568 struct ocfs2_caching_info *ci,
1569 struct buffer_head *ref_root_bh,
1570 struct buffer_head *ref_leaf_bh,
1571 struct ocfs2_refcount_rec *rec)
1574 u32 new_cpos, old_cpos;
1575 struct ocfs2_path *path = NULL;
1576 struct ocfs2_extent_tree et;
1577 struct ocfs2_refcount_block *rb =
1578 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1579 struct ocfs2_extent_list *el;
1581 if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1584 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1585 old_cpos = le32_to_cpu(rb->rf_cpos);
1586 new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1587 if (old_cpos <= new_cpos)
1590 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1592 path = ocfs2_new_path_from_et(&et);
1599 ret = ocfs2_find_path(ci, path, old_cpos);
1606 * 2 more credits, one for the leaf refcount block, one for
1607 * the extent block contains the extent rec.
1609 ret = ocfs2_extend_trans(handle, handle->h_buffer_credits + 2);
1615 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1616 OCFS2_JOURNAL_ACCESS_WRITE);
1622 ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1623 OCFS2_JOURNAL_ACCESS_WRITE);
1629 /* change the leaf extent block first. */
1630 el = path_leaf_el(path);
1632 for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1633 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1636 BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1638 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1640 /* change the r_cpos in the leaf block. */
1641 rb->rf_cpos = cpu_to_le32(new_cpos);
1643 ocfs2_journal_dirty(handle, path_leaf_bh(path));
1644 ocfs2_journal_dirty(handle, ref_leaf_bh);
1647 ocfs2_free_path(path);
1651 static int ocfs2_insert_refcount_rec(handle_t *handle,
1652 struct ocfs2_caching_info *ci,
1653 struct buffer_head *ref_root_bh,
1654 struct buffer_head *ref_leaf_bh,
1655 struct ocfs2_refcount_rec *rec,
1656 int index, int merge,
1657 struct ocfs2_alloc_context *meta_ac)
1660 struct ocfs2_refcount_block *rb =
1661 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1662 struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1663 struct buffer_head *new_bh = NULL;
1665 BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1667 if (rf_list->rl_used == rf_list->rl_count) {
1668 u64 cpos = le64_to_cpu(rec->r_cpos);
1669 u32 len = le32_to_cpu(rec->r_clusters);
1671 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1672 ref_leaf_bh, meta_ac);
1678 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1679 cpos, len, NULL, &index,
1686 ref_leaf_bh = new_bh;
1687 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1688 rf_list = &rb->rf_records;
1691 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1692 OCFS2_JOURNAL_ACCESS_WRITE);
1698 if (index < le16_to_cpu(rf_list->rl_used))
1699 memmove(&rf_list->rl_recs[index + 1],
1700 &rf_list->rl_recs[index],
1701 (le16_to_cpu(rf_list->rl_used) - index) *
1702 sizeof(struct ocfs2_refcount_rec));
1704 mlog(0, "insert refcount record start %llu, len %u, count %u "
1705 "to leaf block %llu at index %d\n",
1706 (unsigned long long)le64_to_cpu(rec->r_cpos),
1707 le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount),
1708 (unsigned long long)ref_leaf_bh->b_blocknr, index);
1710 rf_list->rl_recs[index] = *rec;
1712 le16_add_cpu(&rf_list->rl_used, 1);
1715 ocfs2_refcount_rec_merge(rb, index);
1717 ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1724 ret = ocfs2_adjust_refcount_rec(handle, ci,
1736 * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1737 * This is much simple than our b-tree code.
1738 * split_rec is the new refcount rec we want to insert.
1739 * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1740 * increase refcount or decrease a refcount to non-zero).
1741 * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1742 * rec( in case we decrease a refcount to zero).
1744 static int ocfs2_split_refcount_rec(handle_t *handle,
1745 struct ocfs2_caching_info *ci,
1746 struct buffer_head *ref_root_bh,
1747 struct buffer_head *ref_leaf_bh,
1748 struct ocfs2_refcount_rec *split_rec,
1749 int index, int merge,
1750 struct ocfs2_alloc_context *meta_ac,
1751 struct ocfs2_cached_dealloc_ctxt *dealloc)
1755 struct ocfs2_refcount_block *rb =
1756 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1757 struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1758 struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1759 struct ocfs2_refcount_rec *tail_rec = NULL;
1760 struct buffer_head *new_bh = NULL;
1762 BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1764 mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n",
1765 le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters),
1766 le64_to_cpu(split_rec->r_cpos),
1767 le32_to_cpu(split_rec->r_clusters));
1770 * If we just need to split the header or tail clusters,
1771 * no more recs are needed, just split is OK.
1772 * Otherwise we at least need one new recs.
1774 if (!split_rec->r_refcount &&
1775 (split_rec->r_cpos == orig_rec->r_cpos ||
1776 le64_to_cpu(split_rec->r_cpos) +
1777 le32_to_cpu(split_rec->r_clusters) ==
1778 le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1784 * We need one more rec if we split in the middle and the new rec have
1785 * some refcount in it.
1787 if (split_rec->r_refcount &&
1788 (split_rec->r_cpos != orig_rec->r_cpos &&
1789 le64_to_cpu(split_rec->r_cpos) +
1790 le32_to_cpu(split_rec->r_clusters) !=
1791 le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1794 /* If the leaf block don't have enough record, expand it. */
1795 if (le16_to_cpu(rf_list->rl_used) + recs_need > rf_list->rl_count) {
1796 struct ocfs2_refcount_rec tmp_rec;
1797 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1798 len = le32_to_cpu(orig_rec->r_clusters);
1799 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1800 ref_leaf_bh, meta_ac);
1807 * We have to re-get it since now cpos may be moved to
1808 * another leaf block.
1810 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1811 cpos, len, &tmp_rec, &index,
1818 ref_leaf_bh = new_bh;
1819 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1820 rf_list = &rb->rf_records;
1821 orig_rec = &rf_list->rl_recs[index];
1824 ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1825 OCFS2_JOURNAL_ACCESS_WRITE);
1832 * We have calculated out how many new records we need and store
1833 * in recs_need, so spare enough space first by moving the records
1834 * after "index" to the end.
1836 if (index != le16_to_cpu(rf_list->rl_used) - 1)
1837 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1838 &rf_list->rl_recs[index + 1],
1839 (le16_to_cpu(rf_list->rl_used) - index - 1) *
1840 sizeof(struct ocfs2_refcount_rec));
1842 len = (le64_to_cpu(orig_rec->r_cpos) +
1843 le32_to_cpu(orig_rec->r_clusters)) -
1844 (le64_to_cpu(split_rec->r_cpos) +
1845 le32_to_cpu(split_rec->r_clusters));
1848 * If we have "len", the we will split in the tail and move it
1849 * to the end of the space we have just spared.
1852 tail_rec = &rf_list->rl_recs[index + recs_need];
1854 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1855 le64_add_cpu(&tail_rec->r_cpos,
1856 le32_to_cpu(tail_rec->r_clusters) - len);
1857 tail_rec->r_clusters = le32_to_cpu(len);
1861 * If the split pos isn't the same as the original one, we need to
1862 * split in the head.
1864 * Note: We have the chance that split_rec.r_refcount = 0,
1865 * recs_need = 0 and len > 0, which means we just cut the head from
1866 * the orig_rec and in that case we have done some modification in
1867 * orig_rec above, so the check for r_cpos is faked.
1869 if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1870 len = le64_to_cpu(split_rec->r_cpos) -
1871 le64_to_cpu(orig_rec->r_cpos);
1872 orig_rec->r_clusters = cpu_to_le32(len);
1876 le16_add_cpu(&rf_list->rl_used, recs_need);
1878 if (split_rec->r_refcount) {
1879 rf_list->rl_recs[index] = *split_rec;
1880 mlog(0, "insert refcount record start %llu, len %u, count %u "
1881 "to leaf block %llu at index %d\n",
1882 (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1883 le32_to_cpu(split_rec->r_clusters),
1884 le32_to_cpu(split_rec->r_refcount),
1885 (unsigned long long)ref_leaf_bh->b_blocknr, index);
1888 ocfs2_refcount_rec_merge(rb, index);
1891 ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1900 static int __ocfs2_increase_refcount(handle_t *handle,
1901 struct ocfs2_caching_info *ci,
1902 struct buffer_head *ref_root_bh,
1903 u64 cpos, u32 len, int merge,
1904 struct ocfs2_alloc_context *meta_ac,
1905 struct ocfs2_cached_dealloc_ctxt *dealloc)
1908 struct buffer_head *ref_leaf_bh = NULL;
1909 struct ocfs2_refcount_rec rec;
1910 unsigned int set_len = 0;
1912 mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n",
1913 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1914 (unsigned long long)cpos, len);
1917 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1918 cpos, len, &rec, &index,
1925 set_len = le32_to_cpu(rec.r_clusters);
1928 * Here we may meet with 3 situations:
1930 * 1. If we find an already existing record, and the length
1931 * is the same, cool, we just need to increase the r_refcount
1933 * 2. If we find a hole, just insert it with r_refcount = 1.
1934 * 3. If we are in the middle of one extent record, split
1937 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
1939 mlog(0, "increase refcount rec, start %llu, len %u, "
1940 "count %u\n", (unsigned long long)cpos, set_len,
1941 le32_to_cpu(rec.r_refcount));
1942 ret = ocfs2_change_refcount_rec(handle, ci,
1949 } else if (!rec.r_refcount) {
1950 rec.r_refcount = cpu_to_le32(1);
1952 mlog(0, "insert refcount rec, start %llu, len %u\n",
1953 (unsigned long long)le64_to_cpu(rec.r_cpos),
1955 ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
1964 set_len = min((u64)(cpos + len),
1965 le64_to_cpu(rec.r_cpos) + set_len) - cpos;
1966 rec.r_cpos = cpu_to_le64(cpos);
1967 rec.r_clusters = cpu_to_le32(set_len);
1968 le32_add_cpu(&rec.r_refcount, 1);
1970 mlog(0, "split refcount rec, start %llu, "
1971 "len %u, count %u\n",
1972 (unsigned long long)le64_to_cpu(rec.r_cpos),
1973 set_len, le32_to_cpu(rec.r_refcount));
1974 ret = ocfs2_split_refcount_rec(handle, ci,
1975 ref_root_bh, ref_leaf_bh,
1986 brelse(ref_leaf_bh);
1991 brelse(ref_leaf_bh);
1995 static int ocfs2_remove_refcount_extent(handle_t *handle,
1996 struct ocfs2_caching_info *ci,
1997 struct buffer_head *ref_root_bh,
1998 struct buffer_head *ref_leaf_bh,
1999 struct ocfs2_alloc_context *meta_ac,
2000 struct ocfs2_cached_dealloc_ctxt *dealloc)
2003 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2004 struct ocfs2_refcount_block *rb =
2005 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2006 struct ocfs2_extent_tree et;
2008 BUG_ON(rb->rf_records.rl_used);
2010 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2011 ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
2012 1, meta_ac, dealloc);
2018 ocfs2_remove_from_cache(ci, ref_leaf_bh);
2021 * add the freed block to the dealloc so that it will be freed
2022 * when we run dealloc.
2024 ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
2025 le16_to_cpu(rb->rf_suballoc_slot),
2026 le64_to_cpu(rb->rf_blkno),
2027 le16_to_cpu(rb->rf_suballoc_bit));
2033 ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
2034 OCFS2_JOURNAL_ACCESS_WRITE);
2040 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2042 le32_add_cpu(&rb->rf_clusters, -1);
2045 * check whether we need to restore the root refcount block if
2046 * there is no leaf extent block at atll.
2048 if (!rb->rf_list.l_next_free_rec) {
2049 BUG_ON(rb->rf_clusters);
2051 mlog(0, "reset refcount tree root %llu to be a record block.\n",
2052 (unsigned long long)ref_root_bh->b_blocknr);
2057 memset(&rb->rf_records, 0, sb->s_blocksize -
2058 offsetof(struct ocfs2_refcount_block, rf_records));
2059 rb->rf_records.rl_count =
2060 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
2063 ocfs2_journal_dirty(handle, ref_root_bh);
2069 int ocfs2_increase_refcount(handle_t *handle,
2070 struct ocfs2_caching_info *ci,
2071 struct buffer_head *ref_root_bh,
2073 struct ocfs2_alloc_context *meta_ac,
2074 struct ocfs2_cached_dealloc_ctxt *dealloc)
2076 return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
2081 static int ocfs2_decrease_refcount_rec(handle_t *handle,
2082 struct ocfs2_caching_info *ci,
2083 struct buffer_head *ref_root_bh,
2084 struct buffer_head *ref_leaf_bh,
2085 int index, u64 cpos, unsigned int len,
2086 struct ocfs2_alloc_context *meta_ac,
2087 struct ocfs2_cached_dealloc_ctxt *dealloc)
2090 struct ocfs2_refcount_block *rb =
2091 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2092 struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2094 BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2096 le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2098 if (cpos == le64_to_cpu(rec->r_cpos) &&
2099 len == le32_to_cpu(rec->r_clusters))
2100 ret = ocfs2_change_refcount_rec(handle, ci,
2101 ref_leaf_bh, index, 1, -1);
2103 struct ocfs2_refcount_rec split = *rec;
2104 split.r_cpos = cpu_to_le64(cpos);
2105 split.r_clusters = cpu_to_le32(len);
2107 le32_add_cpu(&split.r_refcount, -1);
2109 mlog(0, "split refcount rec, start %llu, "
2110 "len %u, count %u, original start %llu, len %u\n",
2111 (unsigned long long)le64_to_cpu(split.r_cpos),
2112 len, le32_to_cpu(split.r_refcount),
2113 (unsigned long long)le64_to_cpu(rec->r_cpos),
2114 le32_to_cpu(rec->r_clusters));
2115 ret = ocfs2_split_refcount_rec(handle, ci,
2116 ref_root_bh, ref_leaf_bh,
2126 /* Remove the leaf refcount block if it contains no refcount record. */
2127 if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2128 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2129 ref_leaf_bh, meta_ac,
2139 static int __ocfs2_decrease_refcount(handle_t *handle,
2140 struct ocfs2_caching_info *ci,
2141 struct buffer_head *ref_root_bh,
2143 struct ocfs2_alloc_context *meta_ac,
2144 struct ocfs2_cached_dealloc_ctxt *dealloc,
2147 int ret = 0, index = 0;
2148 struct ocfs2_refcount_rec rec;
2149 unsigned int r_count = 0, r_len;
2150 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2151 struct buffer_head *ref_leaf_bh = NULL;
2153 mlog(0, "Tree owner %llu, decrease refcount start %llu, "
2154 "len %u, delete %u\n",
2155 (unsigned long long)ocfs2_metadata_cache_owner(ci),
2156 (unsigned long long)cpos, len, delete);
2159 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2160 cpos, len, &rec, &index,
2167 r_count = le32_to_cpu(rec.r_refcount);
2168 BUG_ON(r_count == 0);
2170 BUG_ON(r_count > 1);
2172 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2173 le32_to_cpu(rec.r_clusters)) - cpos;
2175 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2184 if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
2185 ret = ocfs2_cache_cluster_dealloc(dealloc,
2186 ocfs2_clusters_to_blocks(sb, cpos),
2196 brelse(ref_leaf_bh);
2201 brelse(ref_leaf_bh);
2205 /* Caller must hold refcount tree lock. */
2206 int ocfs2_decrease_refcount(struct inode *inode,
2207 handle_t *handle, u32 cpos, u32 len,
2208 struct ocfs2_alloc_context *meta_ac,
2209 struct ocfs2_cached_dealloc_ctxt *dealloc,
2214 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2215 struct buffer_head *ref_root_bh = NULL;
2216 struct ocfs2_refcount_tree *tree;
2218 BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2220 ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2226 ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2232 ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2239 ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2240 cpos, len, meta_ac, dealloc, delete);
2244 brelse(ref_root_bh);
2249 * Mark the already-existing extent at cpos as refcounted for len clusters.
2250 * This adds the refcount extent flag.
2252 * If the existing extent is larger than the request, initiate a
2253 * split. An attempt will be made at merging with adjacent extents.
2255 * The caller is responsible for passing down meta_ac if we'll need it.
2257 static int ocfs2_mark_extent_refcounted(struct inode *inode,
2258 struct ocfs2_extent_tree *et,
2259 handle_t *handle, u32 cpos,
2261 struct ocfs2_alloc_context *meta_ac,
2262 struct ocfs2_cached_dealloc_ctxt *dealloc)
2266 mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n",
2267 inode->i_ino, cpos, len, phys);
2269 if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2270 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2271 "tree, but the feature bit is not set in the "
2272 "super block.", inode->i_ino);
2277 ret = ocfs2_change_extent_flag(handle, et, cpos,
2278 len, phys, meta_ac, dealloc,
2279 OCFS2_EXT_REFCOUNTED, 0);
2288 * Given some contiguous physical clusters, calculate what we need
2289 * for modifying their refcount.
2291 static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2292 struct ocfs2_caching_info *ci,
2293 struct buffer_head *ref_root_bh,
2299 int ret = 0, index, ref_blocks = 0, recs_add = 0;
2300 u64 cpos = start_cpos;
2301 struct ocfs2_refcount_block *rb;
2302 struct ocfs2_refcount_rec rec;
2303 struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2306 mlog(0, "start_cpos %llu, clusters %u\n",
2307 (unsigned long long)start_cpos, clusters);
2309 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2310 cpos, clusters, &rec,
2311 &index, &ref_leaf_bh);
2317 if (ref_leaf_bh != prev_bh) {
2319 * Now we encounter a new leaf block, so calculate
2320 * whether we need to extend the old leaf.
2323 rb = (struct ocfs2_refcount_block *)
2326 if (le64_to_cpu(rb->rf_records.rl_used) +
2328 le16_to_cpu(rb->rf_records.rl_count))
2335 prev_bh = ref_leaf_bh;
2339 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2341 mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu,"
2342 "rec->r_clusters %u, rec->r_refcount %u, index %d\n",
2343 recs_add, (unsigned long long)cpos, clusters,
2344 (unsigned long long)le64_to_cpu(rec.r_cpos),
2345 le32_to_cpu(rec.r_clusters),
2346 le32_to_cpu(rec.r_refcount), index);
2348 len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2349 le32_to_cpu(rec.r_clusters)) - cpos;
2351 * If the refcount rec already exist, cool. We just need
2352 * to check whether there is a split. Otherwise we just need
2353 * to increase the refcount.
2354 * If we will insert one, increases recs_add.
2356 * We record all the records which will be inserted to the
2357 * same refcount block, so that we can tell exactly whether
2358 * we need a new refcount block or not.
2360 if (rec.r_refcount) {
2361 /* Check whether we need a split at the beginning. */
2362 if (cpos == start_cpos &&
2363 cpos != le64_to_cpu(rec.r_cpos))
2366 /* Check whether we need a split in the end. */
2367 if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2368 le32_to_cpu(rec.r_clusters))
2373 brelse(ref_leaf_bh);
2380 rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2382 if (le64_to_cpu(rb->rf_records.rl_used) + recs_add >
2383 le16_to_cpu(rb->rf_records.rl_count))
2392 mlog(0, "we need ref_blocks %d\n", ref_blocks);
2393 *meta_add += ref_blocks;
2394 *credits += ref_blocks;
2397 * So we may need ref_blocks to insert into the tree.
2398 * That also means we need to change the b-tree and add that number
2399 * of records since we never merge them.
2400 * We need one more block for expansion since the new created leaf
2401 * block is also full and needs split.
2403 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2404 if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2405 struct ocfs2_extent_tree et;
2407 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2408 *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2409 *credits += ocfs2_calc_extend_credits(sb,
2413 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2418 brelse(ref_leaf_bh);
2424 * For refcount tree, we will decrease some contiguous clusters
2425 * refcount count, so just go through it to see how many blocks
2426 * we gonna touch and whether we need to create new blocks.
2428 * Normally the refcount blocks store these refcount should be
2429 * continguous also, so that we can get the number easily.
2430 * As for meta_ac, we will at most add split 2 refcount record and
2431 * 2 more refcount block, so just check it in a rough way.
2433 * Caller must hold refcount tree lock.
2435 int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
2436 struct buffer_head *di_bh,
2440 struct ocfs2_alloc_context **meta_ac)
2442 int ret, ref_blocks = 0;
2443 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2444 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2445 struct buffer_head *ref_root_bh = NULL;
2446 struct ocfs2_refcount_tree *tree;
2447 u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2449 if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2450 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2451 "tree, but the feature bit is not set in the "
2452 "super block.", inode->i_ino);
2457 BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2459 ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
2460 le64_to_cpu(di->i_refcount_loc), &tree);
2466 ret = ocfs2_read_refcount_block(&tree->rf_ci,
2467 le64_to_cpu(di->i_refcount_loc),
2474 ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2477 start_cpos, clusters,
2478 &ref_blocks, credits);
2484 mlog(0, "reserve new metadata %d, credits = %d\n",
2485 ref_blocks, *credits);
2488 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
2489 ref_blocks, meta_ac);
2495 brelse(ref_root_bh);
2499 #define MAX_CONTIG_BYTES 1048576
2501 static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb)
2503 return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES);
2506 static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb)
2508 return ~(ocfs2_cow_contig_clusters(sb) - 1);
2512 * Given an extent that starts at 'start' and an I/O that starts at 'cpos',
2513 * find an offset (start + (n * contig_clusters)) that is closest to cpos
2514 * while still being less than or equal to it.
2516 * The goal is to break the extent at a multiple of contig_clusters.
2518 static inline unsigned int ocfs2_cow_align_start(struct super_block *sb,
2522 BUG_ON(start > cpos);
2524 return start + ((cpos - start) & ocfs2_cow_contig_mask(sb));
2528 * Given a cluster count of len, pad it out so that it is a multiple
2529 * of contig_clusters.
2531 static inline unsigned int ocfs2_cow_align_length(struct super_block *sb,
2534 unsigned int padded =
2535 (len + (ocfs2_cow_contig_clusters(sb) - 1)) &
2536 ocfs2_cow_contig_mask(sb);
2546 * Calculate out the start and number of virtual clusters we need to to CoW.
2548 * cpos is vitual start cluster position we want to do CoW in a
2549 * file and write_len is the cluster length.
2550 * max_cpos is the place where we want to stop CoW intentionally.
2552 * Normal we will start CoW from the beginning of extent record cotaining cpos.
2553 * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we
2554 * get good I/O from the resulting extent tree.
2556 static int ocfs2_refcount_cal_cow_clusters(struct inode *inode,
2557 struct ocfs2_extent_list *el,
2565 int tree_height = le16_to_cpu(el->l_tree_depth), i;
2566 struct buffer_head *eb_bh = NULL;
2567 struct ocfs2_extent_block *eb = NULL;
2568 struct ocfs2_extent_rec *rec;
2569 unsigned int want_clusters, rec_end = 0;
2570 int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb);
2573 BUG_ON(cpos + write_len > max_cpos);
2575 if (tree_height > 0) {
2576 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh);
2582 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2585 if (el->l_tree_depth) {
2586 ocfs2_error(inode->i_sb,
2587 "Inode %lu has non zero tree depth in "
2588 "leaf block %llu\n", inode->i_ino,
2589 (unsigned long long)eb_bh->b_blocknr);
2596 for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
2597 rec = &el->l_recs[i];
2599 if (ocfs2_is_empty_extent(rec)) {
2600 mlog_bug_on_msg(i != 0, "Inode %lu has empty record in "
2601 "index %d\n", inode->i_ino, i);
2605 if (le32_to_cpu(rec->e_cpos) +
2606 le16_to_cpu(rec->e_leaf_clusters) <= cpos)
2609 if (*cow_len == 0) {
2611 * We should find a refcounted record in the
2614 BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED));
2615 *cow_start = le32_to_cpu(rec->e_cpos);
2619 * If we encounter a hole, a non-refcounted record or
2620 * pass the max_cpos, stop the search.
2622 if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) ||
2623 (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) ||
2624 (max_cpos <= le32_to_cpu(rec->e_cpos)))
2627 leaf_clusters = le16_to_cpu(rec->e_leaf_clusters);
2628 rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters;
2629 if (rec_end > max_cpos) {
2631 leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos);
2635 * How many clusters do we actually need from
2636 * this extent? First we see how many we actually
2637 * need to complete the write. If that's smaller
2638 * than contig_clusters, we try for contig_clusters.
2641 want_clusters = write_len;
2643 want_clusters = (cpos + write_len) -
2644 (*cow_start + *cow_len);
2645 if (want_clusters < contig_clusters)
2646 want_clusters = contig_clusters;
2649 * If the write does not cover the whole extent, we
2650 * need to calculate how we're going to split the extent.
2651 * We try to do it on contig_clusters boundaries.
2653 * Any extent smaller than contig_clusters will be
2654 * CoWed in its entirety.
2656 if (leaf_clusters <= contig_clusters)
2657 *cow_len += leaf_clusters;
2658 else if (*cow_len || (*cow_start == cpos)) {
2660 * This extent needs to be CoW'd from its
2661 * beginning, so all we have to do is compute
2662 * how many clusters to grab. We align
2663 * want_clusters to the edge of contig_clusters
2664 * to get better I/O.
2666 want_clusters = ocfs2_cow_align_length(inode->i_sb,
2669 if (leaf_clusters < want_clusters)
2670 *cow_len += leaf_clusters;
2672 *cow_len += want_clusters;
2673 } else if ((*cow_start + contig_clusters) >=
2674 (cpos + write_len)) {
2676 * Breaking off contig_clusters at the front
2677 * of the extent will cover our write. That's
2680 *cow_len = contig_clusters;
2681 } else if ((rec_end - cpos) <= contig_clusters) {
2683 * Breaking off contig_clusters at the tail of
2684 * this extent will cover cpos.
2686 *cow_start = rec_end - contig_clusters;
2687 *cow_len = contig_clusters;
2688 } else if ((rec_end - cpos) <= want_clusters) {
2690 * While we can't fit the entire write in this
2691 * extent, we know that the write goes from cpos
2692 * to the end of the extent. Break that off.
2693 * We try to break it at some multiple of
2694 * contig_clusters from the front of the extent.
2695 * Failing that (ie, cpos is within
2696 * contig_clusters of the front), we'll CoW the
2699 *cow_start = ocfs2_cow_align_start(inode->i_sb,
2701 *cow_len = rec_end - *cow_start;
2704 * Ok, the entire write lives in the middle of
2705 * this extent. Let's try to slice the extent up
2706 * nicely. Optimally, our CoW region starts at
2707 * m*contig_clusters from the beginning of the
2708 * extent and goes for n*contig_clusters,
2709 * covering the entire write.
2711 *cow_start = ocfs2_cow_align_start(inode->i_sb,
2714 want_clusters = (cpos + write_len) - *cow_start;
2715 want_clusters = ocfs2_cow_align_length(inode->i_sb,
2717 if (*cow_start + want_clusters <= rec_end)
2718 *cow_len = want_clusters;
2720 *cow_len = rec_end - *cow_start;
2723 /* Have we covered our entire write yet? */
2724 if ((*cow_start + *cow_len) >= (cpos + write_len))
2728 * If we reach the end of the extent block and don't get enough
2729 * clusters, continue with the next extent block if possible.
2731 if (i + 1 == le16_to_cpu(el->l_next_free_rec) &&
2732 eb && eb->h_next_leaf_blk) {
2736 ret = ocfs2_read_extent_block(INODE_CACHE(inode),
2737 le64_to_cpu(eb->h_next_leaf_blk),
2744 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2756 * Prepare meta_ac, data_ac and calculate credits when we want to add some
2757 * num_clusters in data_tree "et" and change the refcount for the old
2758 * clusters(starting form p_cluster) in the refcount tree.
2761 * 1. since we may split the old tree, so we at most will need num_clusters + 2
2762 * more new leaf records.
2763 * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so
2764 * just give data_ac = NULL.
2766 static int ocfs2_lock_refcount_allocators(struct super_block *sb,
2767 u32 p_cluster, u32 num_clusters,
2768 struct ocfs2_extent_tree *et,
2769 struct ocfs2_caching_info *ref_ci,
2770 struct buffer_head *ref_root_bh,
2771 struct ocfs2_alloc_context **meta_ac,
2772 struct ocfs2_alloc_context **data_ac,
2775 int ret = 0, meta_add = 0;
2776 int num_free_extents = ocfs2_num_free_extents(OCFS2_SB(sb), et);
2778 if (num_free_extents < 0) {
2779 ret = num_free_extents;
2784 if (num_free_extents < num_clusters + 2)
2786 ocfs2_extend_meta_needed(et->et_root_el);
2788 *credits += ocfs2_calc_extend_credits(sb, et->et_root_el,
2791 ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh,
2792 p_cluster, num_clusters,
2793 &meta_add, credits);
2799 mlog(0, "reserve new metadata %d, clusters %u, credits = %d\n",
2800 meta_add, num_clusters, *credits);
2801 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add,
2809 ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters,
2818 ocfs2_free_alloc_context(*meta_ac);
2826 static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh)
2828 BUG_ON(buffer_dirty(bh));
2830 clear_buffer_mapped(bh);
2835 static int ocfs2_duplicate_clusters_by_page(handle_t *handle,
2836 struct ocfs2_cow_context *context,
2837 u32 cpos, u32 old_cluster,
2838 u32 new_cluster, u32 new_len)
2840 int ret = 0, partial;
2841 struct ocfs2_caching_info *ci = context->data_et.et_ci;
2842 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2843 u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2846 unsigned int from, to;
2847 loff_t offset, end, map_end;
2848 struct address_space *mapping = context->inode->i_mapping;
2850 mlog(0, "old_cluster %u, new %u, len %u at offset %u\n", old_cluster,
2851 new_cluster, new_len, cpos);
2853 offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
2854 end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2856 while (offset < end) {
2857 page_index = offset >> PAGE_CACHE_SHIFT;
2858 map_end = (page_index + 1) << PAGE_CACHE_SHIFT;
2862 /* from, to is the offset within the page. */
2863 from = offset & (PAGE_CACHE_SIZE - 1);
2864 to = PAGE_CACHE_SIZE;
2865 if (map_end & (PAGE_CACHE_SIZE - 1))
2866 to = map_end & (PAGE_CACHE_SIZE - 1);
2868 page = grab_cache_page(mapping, page_index);
2870 /* This page can't be dirtied before we CoW it out. */
2871 BUG_ON(PageDirty(page));
2873 if (!PageUptodate(page)) {
2874 ret = block_read_full_page(page, ocfs2_get_block);
2882 if (page_has_buffers(page)) {
2883 ret = walk_page_buffers(handle, page_buffers(page),
2885 ocfs2_clear_cow_buffer);
2892 ocfs2_map_and_dirty_page(context->inode,
2894 page, 0, &new_block);
2895 mark_page_accessed(page);
2898 page_cache_release(page);
2908 static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
2909 struct ocfs2_cow_context *context,
2910 u32 cpos, u32 old_cluster,
2911 u32 new_cluster, u32 new_len)
2914 struct super_block *sb = context->inode->i_sb;
2915 struct ocfs2_caching_info *ci = context->data_et.et_ci;
2916 int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
2917 u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
2918 u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2919 struct ocfs2_super *osb = OCFS2_SB(sb);
2920 struct buffer_head *old_bh = NULL;
2921 struct buffer_head *new_bh = NULL;
2923 mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster,
2924 new_cluster, new_len);
2926 for (i = 0; i < blocks; i++, old_block++, new_block++) {
2927 new_bh = sb_getblk(osb->sb, new_block);
2928 if (new_bh == NULL) {
2934 ocfs2_set_new_buffer_uptodate(ci, new_bh);
2936 ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
2942 ret = ocfs2_journal_access(handle, ci, new_bh,
2943 OCFS2_JOURNAL_ACCESS_CREATE);
2949 memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
2950 ret = ocfs2_journal_dirty(handle, new_bh);
2967 static int ocfs2_clear_ext_refcount(handle_t *handle,
2968 struct ocfs2_extent_tree *et,
2969 u32 cpos, u32 p_cluster, u32 len,
2970 unsigned int ext_flags,
2971 struct ocfs2_alloc_context *meta_ac,
2972 struct ocfs2_cached_dealloc_ctxt *dealloc)
2975 struct ocfs2_extent_rec replace_rec;
2976 struct ocfs2_path *path = NULL;
2977 struct ocfs2_extent_list *el;
2978 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2979 u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
2981 mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n",
2982 (unsigned long long)ino, cpos, len, p_cluster, ext_flags);
2984 memset(&replace_rec, 0, sizeof(replace_rec));
2985 replace_rec.e_cpos = cpu_to_le32(cpos);
2986 replace_rec.e_leaf_clusters = cpu_to_le16(len);
2987 replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
2989 replace_rec.e_flags = ext_flags;
2990 replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
2992 path = ocfs2_new_path_from_et(et);
2999 ret = ocfs2_find_path(et->et_ci, path, cpos);
3005 el = path_leaf_el(path);
3007 index = ocfs2_search_extent_list(el, cpos);
3008 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
3010 "Inode %llu has an extent at cpos %u which can no "
3011 "longer be found.\n",
3012 (unsigned long long)ino, cpos);
3017 ret = ocfs2_split_extent(handle, et, path, index,
3018 &replace_rec, meta_ac, dealloc);
3023 ocfs2_free_path(path);
3027 static int ocfs2_replace_clusters(handle_t *handle,
3028 struct ocfs2_cow_context *context,
3031 unsigned int ext_flags)
3034 struct ocfs2_caching_info *ci = context->data_et.et_ci;
3035 u64 ino = ocfs2_metadata_cache_owner(ci);
3037 mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n",
3038 (unsigned long long)ino, cpos, old, new, len, ext_flags);
3040 /*If the old clusters is unwritten, no need to duplicate. */
3041 if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3042 ret = context->cow_duplicate_clusters(handle, context, cpos,
3050 ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
3051 cpos, new, len, ext_flags,
3052 context->meta_ac, &context->dealloc);
3059 static int ocfs2_cow_sync_writeback(struct super_block *sb,
3060 struct ocfs2_cow_context *context,
3061 u32 cpos, u32 num_clusters)
3064 loff_t offset, end, map_end;
3068 if (ocfs2_should_order_data(context->inode))
3071 offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
3072 end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits);
3074 ret = filemap_fdatawrite_range(context->inode->i_mapping,
3081 while (offset < end) {
3082 page_index = offset >> PAGE_CACHE_SHIFT;
3083 map_end = (page_index + 1) << PAGE_CACHE_SHIFT;
3087 page = grab_cache_page(context->inode->i_mapping, page_index);
3090 wait_on_page_writeback(page);
3091 if (PageError(page)) {
3095 mark_page_accessed(page);
3098 page_cache_release(page);
3108 static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3109 u32 v_cluster, u32 *p_cluster,
3111 unsigned int *extent_flags)
3113 return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3114 num_clusters, extent_flags);
3117 static int ocfs2_make_clusters_writable(struct super_block *sb,
3118 struct ocfs2_cow_context *context,
3119 u32 cpos, u32 p_cluster,
3120 u32 num_clusters, unsigned int e_flags)
3122 int ret, delete, index, credits = 0;
3123 u32 new_bit, new_len;
3124 unsigned int set_len;
3125 struct ocfs2_super *osb = OCFS2_SB(sb);
3127 struct buffer_head *ref_leaf_bh = NULL;
3128 struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
3129 struct ocfs2_refcount_rec rec;
3131 mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n",
3132 cpos, p_cluster, num_clusters, e_flags);
3134 ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3137 context->ref_root_bh,
3139 &context->data_ac, &credits);
3145 if (context->post_refcount)
3146 credits += context->post_refcount->credits;
3148 credits += context->extra_credits;
3149 handle = ocfs2_start_trans(osb, credits);
3150 if (IS_ERR(handle)) {
3151 ret = PTR_ERR(handle);
3156 while (num_clusters) {
3157 ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
3158 p_cluster, num_clusters,
3159 &rec, &index, &ref_leaf_bh);
3165 BUG_ON(!rec.r_refcount);
3166 set_len = min((u64)p_cluster + num_clusters,
3167 le64_to_cpu(rec.r_cpos) +
3168 le32_to_cpu(rec.r_clusters)) - p_cluster;
3171 * There are many different situation here.
3172 * 1. If refcount == 1, remove the flag and don't COW.
3173 * 2. If refcount > 1, allocate clusters.
3174 * Here we may not allocate r_len once at a time, so continue
3175 * until we reach num_clusters.
3177 if (le32_to_cpu(rec.r_refcount) == 1) {
3179 ret = ocfs2_clear_ext_refcount(handle,
3192 ret = __ocfs2_claim_clusters(osb, handle,
3195 &new_bit, &new_len);
3201 ret = ocfs2_replace_clusters(handle, context,
3202 cpos, p_cluster, new_bit,
3211 ret = __ocfs2_decrease_refcount(handle, ref_ci,
3212 context->ref_root_bh,
3215 &context->dealloc, delete);
3222 p_cluster += set_len;
3223 num_clusters -= set_len;
3224 brelse(ref_leaf_bh);
3228 /* handle any post_cow action. */
3229 if (context->post_refcount && context->post_refcount->func) {
3230 ret = context->post_refcount->func(context->inode, handle,
3231 context->post_refcount->para);
3239 * Here we should write the new page out first if we are
3240 * in write-back mode.
3242 if (context->get_clusters == ocfs2_di_get_clusters) {
3243 ret = ocfs2_cow_sync_writeback(sb, context, cpos, num_clusters);
3249 ocfs2_commit_trans(osb, handle);
3252 if (context->data_ac) {
3253 ocfs2_free_alloc_context(context->data_ac);
3254 context->data_ac = NULL;
3256 if (context->meta_ac) {
3257 ocfs2_free_alloc_context(context->meta_ac);
3258 context->meta_ac = NULL;
3260 brelse(ref_leaf_bh);
3265 static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
3268 struct inode *inode = context->inode;
3269 u32 cow_start = context->cow_start, cow_len = context->cow_len;
3270 u32 p_cluster, num_clusters;
3271 unsigned int ext_flags;
3272 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3274 if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
3275 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
3276 "tree, but the feature bit is not set in the "
3277 "super block.", inode->i_ino);
3281 ocfs2_init_dealloc_ctxt(&context->dealloc);
3284 ret = context->get_clusters(context, cow_start, &p_cluster,
3285 &num_clusters, &ext_flags);
3291 BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
3293 if (cow_len < num_clusters)
3294 num_clusters = cow_len;
3296 ret = ocfs2_make_clusters_writable(inode->i_sb, context,
3297 cow_start, p_cluster,
3298 num_clusters, ext_flags);
3304 cow_len -= num_clusters;
3305 cow_start += num_clusters;
3308 if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
3309 ocfs2_schedule_truncate_log_flush(osb, 1);
3310 ocfs2_run_deallocs(osb, &context->dealloc);
3317 * Starting at cpos, try to CoW write_len clusters. Don't CoW
3318 * past max_cpos. This will stop when it runs into a hole or an
3319 * unrefcounted extent.
3321 static int ocfs2_refcount_cow_hunk(struct inode *inode,
3322 struct buffer_head *di_bh,
3323 u32 cpos, u32 write_len, u32 max_cpos)
3326 u32 cow_start = 0, cow_len = 0;
3327 struct ocfs2_inode_info *oi = OCFS2_I(inode);
3328 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3329 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3330 struct buffer_head *ref_root_bh = NULL;
3331 struct ocfs2_refcount_tree *ref_tree;
3332 struct ocfs2_cow_context *context = NULL;
3334 BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3336 ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
3337 cpos, write_len, max_cpos,
3338 &cow_start, &cow_len);
3344 mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, "
3345 "cow_len %u\n", inode->i_ino,
3346 cpos, write_len, cow_start, cow_len);
3348 BUG_ON(cow_len == 0);
3350 context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3357 ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3358 1, &ref_tree, &ref_root_bh);
3364 context->inode = inode;
3365 context->cow_start = cow_start;
3366 context->cow_len = cow_len;
3367 context->ref_tree = ref_tree;
3368 context->ref_root_bh = ref_root_bh;
3369 context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3370 context->get_clusters = ocfs2_di_get_clusters;
3372 ocfs2_init_dinode_extent_tree(&context->data_et,
3373 INODE_CACHE(inode), di_bh);
3375 ret = ocfs2_replace_cow(context);
3380 * truncate the extent map here since no matter whether we meet with
3381 * any error during the action, we shouldn't trust cached extent map
3384 ocfs2_extent_map_trunc(inode, cow_start);
3386 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3387 brelse(ref_root_bh);
3394 * CoW any and all clusters between cpos and cpos+write_len.
3395 * Don't CoW past max_cpos. If this returns successfully, all
3396 * clusters between cpos and cpos+write_len are safe to modify.
3398 int ocfs2_refcount_cow(struct inode *inode,
3399 struct buffer_head *di_bh,
3400 u32 cpos, u32 write_len, u32 max_cpos)
3403 u32 p_cluster, num_clusters;
3404 unsigned int ext_flags;
3407 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3408 &num_clusters, &ext_flags);
3414 if (write_len < num_clusters)
3415 num_clusters = write_len;
3417 if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3418 ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
3419 num_clusters, max_cpos);
3426 write_len -= num_clusters;
3427 cpos += num_clusters;
3433 static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3434 u32 v_cluster, u32 *p_cluster,
3436 unsigned int *extent_flags)
3438 struct inode *inode = context->inode;
3439 struct ocfs2_xattr_value_root *xv = context->cow_object;
3441 return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3442 num_clusters, &xv->xr_list,
3447 * Given a xattr value root, calculate the most meta/credits we need for
3448 * refcount tree change if we truncate it to 0.
3450 int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3451 struct ocfs2_caching_info *ref_ci,
3452 struct buffer_head *ref_root_bh,
3453 struct ocfs2_xattr_value_root *xv,
3454 int *meta_add, int *credits)
3456 int ret = 0, index, ref_blocks = 0;
3457 u32 p_cluster, num_clusters;
3458 u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3459 struct ocfs2_refcount_block *rb;
3460 struct ocfs2_refcount_rec rec;
3461 struct buffer_head *ref_leaf_bh = NULL;
3463 while (cpos < clusters) {
3464 ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3465 &num_clusters, &xv->xr_list,
3472 cpos += num_clusters;
3474 while (num_clusters) {
3475 ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3476 p_cluster, num_clusters,
3484 BUG_ON(!rec.r_refcount);
3486 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3489 * We really don't know whether the other clusters is in
3490 * this refcount block or not, so just take the worst
3491 * case that all the clusters are in this block and each
3492 * one will split a refcount rec, so totally we need
3493 * clusters * 2 new refcount rec.
3495 if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3496 le16_to_cpu(rb->rf_records.rl_count))
3500 brelse(ref_leaf_bh);
3503 if (num_clusters <= le32_to_cpu(rec.r_clusters))
3506 num_clusters -= le32_to_cpu(rec.r_clusters);
3507 p_cluster += num_clusters;
3511 *meta_add += ref_blocks;
3515 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3516 if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3517 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3519 struct ocfs2_extent_tree et;
3521 ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3522 *credits += ocfs2_calc_extend_credits(inode->i_sb,
3528 brelse(ref_leaf_bh);
3535 int ocfs2_refcount_cow_xattr(struct inode *inode,
3536 struct ocfs2_dinode *di,
3537 struct ocfs2_xattr_value_buf *vb,
3538 struct ocfs2_refcount_tree *ref_tree,
3539 struct buffer_head *ref_root_bh,
3540 u32 cpos, u32 write_len,
3541 struct ocfs2_post_refcount *post)
3544 struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3545 struct ocfs2_inode_info *oi = OCFS2_I(inode);
3546 struct ocfs2_cow_context *context = NULL;
3547 u32 cow_start, cow_len;
3549 BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3551 ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3552 cpos, write_len, UINT_MAX,
3553 &cow_start, &cow_len);
3559 BUG_ON(cow_len == 0);
3561 context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3568 context->inode = inode;
3569 context->cow_start = cow_start;
3570 context->cow_len = cow_len;
3571 context->ref_tree = ref_tree;
3572 context->ref_root_bh = ref_root_bh;;
3573 context->cow_object = xv;
3575 context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3576 /* We need the extra credits for duplicate_clusters by jbd. */
3577 context->extra_credits =
3578 ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3579 context->get_clusters = ocfs2_xattr_value_get_clusters;
3580 context->post_refcount = post;
3582 ocfs2_init_xattr_value_extent_tree(&context->data_et,
3583 INODE_CACHE(inode), vb);
3585 ret = ocfs2_replace_cow(context);
3595 * Insert a new extent into refcount tree and mark a extent rec
3596 * as refcounted in the dinode tree.
3598 int ocfs2_add_refcount_flag(struct inode *inode,
3599 struct ocfs2_extent_tree *data_et,
3600 struct ocfs2_caching_info *ref_ci,
3601 struct buffer_head *ref_root_bh,
3602 u32 cpos, u32 p_cluster, u32 num_clusters,
3603 struct ocfs2_cached_dealloc_ctxt *dealloc,
3604 struct ocfs2_post_refcount *post)
3608 int credits = 1, ref_blocks = 0;
3609 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3610 struct ocfs2_alloc_context *meta_ac = NULL;
3612 ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3613 ref_ci, ref_root_bh,
3614 p_cluster, num_clusters,
3615 &ref_blocks, &credits);
3621 mlog(0, "reserve new metadata %d, credits = %d\n",
3622 ref_blocks, credits);
3625 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
3626 ref_blocks, &meta_ac);
3634 credits += post->credits;
3636 handle = ocfs2_start_trans(osb, credits);
3637 if (IS_ERR(handle)) {
3638 ret = PTR_ERR(handle);
3643 ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3644 cpos, num_clusters, p_cluster,
3651 ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3652 p_cluster, num_clusters, 0,
3659 if (post && post->func) {
3660 ret = post->func(inode, handle, post->para);
3666 ocfs2_commit_trans(osb, handle);
3669 ocfs2_free_alloc_context(meta_ac);
3673 static int ocfs2_change_ctime(struct inode *inode,
3674 struct buffer_head *di_bh)
3678 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3680 handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3681 OCFS2_INODE_UPDATE_CREDITS);
3682 if (IS_ERR(handle)) {
3683 ret = PTR_ERR(handle);
3688 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3689 OCFS2_JOURNAL_ACCESS_WRITE);
3695 inode->i_ctime = CURRENT_TIME;
3696 di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
3697 di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
3699 ocfs2_journal_dirty(handle, di_bh);
3702 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3707 static int ocfs2_attach_refcount_tree(struct inode *inode,
3708 struct buffer_head *di_bh)
3710 int ret, data_changed = 0;
3711 struct buffer_head *ref_root_bh = NULL;
3712 struct ocfs2_inode_info *oi = OCFS2_I(inode);
3713 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3714 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3715 struct ocfs2_refcount_tree *ref_tree;
3716 unsigned int ext_flags;
3718 u32 cpos, num_clusters, clusters, p_cluster;
3719 struct ocfs2_cached_dealloc_ctxt dealloc;
3720 struct ocfs2_extent_tree di_et;
3722 ocfs2_init_dealloc_ctxt(&dealloc);
3724 if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) {
3725 ret = ocfs2_create_refcount_tree(inode, di_bh);
3732 BUG_ON(!di->i_refcount_loc);
3733 ret = ocfs2_lock_refcount_tree(osb,
3734 le64_to_cpu(di->i_refcount_loc), 1,
3735 &ref_tree, &ref_root_bh);
3741 ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3743 size = i_size_read(inode);
3744 clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3747 while (cpos < clusters) {
3748 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3749 &num_clusters, &ext_flags);
3751 if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3752 ret = ocfs2_add_refcount_flag(inode, &di_et,
3755 p_cluster, num_clusters,
3764 cpos += num_clusters;
3767 if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
3768 ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
3779 ret = ocfs2_change_ctime(inode, di_bh);
3785 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3786 brelse(ref_root_bh);
3788 if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3789 ocfs2_schedule_truncate_log_flush(osb, 1);
3790 ocfs2_run_deallocs(osb, &dealloc);
3794 * Empty the extent map so that we may get the right extent
3795 * record from the disk.
3797 ocfs2_extent_map_trunc(inode, 0);
3802 static int ocfs2_add_refcounted_extent(struct inode *inode,
3803 struct ocfs2_extent_tree *et,
3804 struct ocfs2_caching_info *ref_ci,
3805 struct buffer_head *ref_root_bh,
3806 u32 cpos, u32 p_cluster, u32 num_clusters,
3807 unsigned int ext_flags,
3808 struct ocfs2_cached_dealloc_ctxt *dealloc)
3813 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3814 struct ocfs2_alloc_context *meta_ac = NULL;
3816 ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3817 p_cluster, num_clusters,
3819 ref_root_bh, &meta_ac,
3826 handle = ocfs2_start_trans(osb, credits);
3827 if (IS_ERR(handle)) {
3828 ret = PTR_ERR(handle);
3833 ret = ocfs2_insert_extent(handle, et, cpos,
3834 cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
3836 num_clusters, ext_flags, meta_ac);
3842 ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3843 p_cluster, num_clusters,
3849 ocfs2_commit_trans(osb, handle);
3852 ocfs2_free_alloc_context(meta_ac);
3856 static int ocfs2_duplicate_extent_list(struct inode *s_inode,
3857 struct inode *t_inode,
3858 struct buffer_head *t_bh,
3859 struct ocfs2_caching_info *ref_ci,
3860 struct buffer_head *ref_root_bh,
3861 struct ocfs2_cached_dealloc_ctxt *dealloc)
3864 u32 p_cluster, num_clusters, clusters, cpos;
3866 unsigned int ext_flags;
3867 struct ocfs2_extent_tree et;
3869 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
3871 size = i_size_read(s_inode);
3872 clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
3875 while (cpos < clusters) {
3876 ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
3877 &num_clusters, &ext_flags);
3880 ret = ocfs2_add_refcounted_extent(t_inode, &et,
3881 ref_ci, ref_root_bh,
3892 cpos += num_clusters;
3900 * change the new file's attributes to the src.
3902 * reflink creates a snapshot of a file, that means the attributes
3903 * must be identical except for three exceptions - nlink, ino, and ctime.
3905 static int ocfs2_complete_reflink(struct inode *s_inode,
3906 struct buffer_head *s_bh,
3907 struct inode *t_inode,
3908 struct buffer_head *t_bh,
3913 struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
3914 struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
3915 loff_t size = i_size_read(s_inode);
3917 handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
3918 OCFS2_INODE_UPDATE_CREDITS);
3919 if (IS_ERR(handle)) {
3920 ret = PTR_ERR(handle);
3925 ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
3926 OCFS2_JOURNAL_ACCESS_WRITE);
3932 spin_lock(&OCFS2_I(t_inode)->ip_lock);
3933 OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
3934 OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
3935 OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
3936 spin_unlock(&OCFS2_I(t_inode)->ip_lock);
3937 i_size_write(t_inode, size);
3939 di->i_xattr_inline_size = s_di->i_xattr_inline_size;
3940 di->i_clusters = s_di->i_clusters;
3941 di->i_size = s_di->i_size;
3942 di->i_dyn_features = s_di->i_dyn_features;
3943 di->i_attr = s_di->i_attr;
3946 di->i_uid = s_di->i_uid;
3947 di->i_gid = s_di->i_gid;
3948 di->i_mode = s_di->i_mode;
3952 * we want mtime to appear identical to the source and
3955 t_inode->i_ctime = CURRENT_TIME;
3957 di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec);
3958 di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec);
3960 t_inode->i_mtime = s_inode->i_mtime;
3961 di->i_mtime = s_di->i_mtime;
3962 di->i_mtime_nsec = s_di->i_mtime_nsec;
3965 ocfs2_journal_dirty(handle, t_bh);
3968 ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
3972 static int ocfs2_create_reflink_node(struct inode *s_inode,
3973 struct buffer_head *s_bh,
3974 struct inode *t_inode,
3975 struct buffer_head *t_bh,
3979 struct buffer_head *ref_root_bh = NULL;
3980 struct ocfs2_cached_dealloc_ctxt dealloc;
3981 struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
3982 struct ocfs2_refcount_block *rb;
3983 struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
3984 struct ocfs2_refcount_tree *ref_tree;
3986 ocfs2_init_dealloc_ctxt(&dealloc);
3988 ret = ocfs2_set_refcount_tree(t_inode, t_bh,
3989 le64_to_cpu(di->i_refcount_loc));
3995 ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3996 1, &ref_tree, &ref_root_bh);
4001 rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
4003 ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4004 &ref_tree->rf_ci, ref_root_bh,
4008 goto out_unlock_refcount;
4011 ret = ocfs2_complete_reflink(s_inode, s_bh, t_inode, t_bh, preserve);
4015 out_unlock_refcount:
4016 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4017 brelse(ref_root_bh);
4019 if (ocfs2_dealloc_has_cluster(&dealloc)) {
4020 ocfs2_schedule_truncate_log_flush(osb, 1);
4021 ocfs2_run_deallocs(osb, &dealloc);
4027 static int __ocfs2_reflink(struct dentry *old_dentry,
4028 struct buffer_head *old_bh,
4029 struct inode *new_inode,
4033 struct inode *inode = old_dentry->d_inode;
4034 struct buffer_head *new_bh = NULL;
4036 ret = filemap_fdatawrite(inode->i_mapping);
4042 ret = ocfs2_attach_refcount_tree(inode, old_bh);
4048 mutex_lock(&new_inode->i_mutex);
4049 ret = ocfs2_inode_lock(new_inode, &new_bh, 1);
4055 ret = ocfs2_create_reflink_node(inode, old_bh,
4056 new_inode, new_bh, preserve);
4062 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
4063 ret = ocfs2_reflink_xattrs(inode, old_bh,
4070 ocfs2_inode_unlock(new_inode, 1);
4073 mutex_unlock(&new_inode->i_mutex);
4076 ret = filemap_fdatawait(inode->i_mapping);
4083 static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
4084 struct dentry *new_dentry, bool preserve)
4087 struct inode *inode = old_dentry->d_inode;
4088 struct buffer_head *old_bh = NULL;
4089 struct inode *new_orphan_inode = NULL;
4091 if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4094 error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
4101 error = ocfs2_inode_lock(inode, &old_bh, 1);
4107 down_write(&OCFS2_I(inode)->ip_xattr_sem);
4108 down_write(&OCFS2_I(inode)->ip_alloc_sem);
4109 error = __ocfs2_reflink(old_dentry, old_bh,
4110 new_orphan_inode, preserve);
4111 up_write(&OCFS2_I(inode)->ip_alloc_sem);
4112 up_write(&OCFS2_I(inode)->ip_xattr_sem);
4114 ocfs2_inode_unlock(inode, 1);
4122 /* If the security isn't preserved, we need to re-initialize them. */
4124 error = ocfs2_init_security_and_acl(dir, new_orphan_inode);
4130 error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
4136 if (new_orphan_inode) {
4138 * We need to open_unlock the inode no matter whether we
4139 * succeed or not, so that other nodes can delete it later.
4141 ocfs2_open_unlock(new_orphan_inode);
4143 iput(new_orphan_inode);