1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Extent allocs and frees
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
32 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
33 #include <cluster/masklog.h>
40 #include "extent_map.h"
43 #include "localalloc.h"
50 #include "buffer_head_io.h"
52 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
);
53 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
54 struct ocfs2_extent_block
*eb
);
57 * Structures which describe a path through a btree, and functions to
60 * The idea here is to be as generic as possible with the tree
63 struct ocfs2_path_item
{
64 struct buffer_head
*bh
;
65 struct ocfs2_extent_list
*el
;
68 #define OCFS2_MAX_PATH_DEPTH 5
72 struct ocfs2_path_item p_node
[OCFS2_MAX_PATH_DEPTH
];
75 #define path_root_bh(_path) ((_path)->p_node[0].bh)
76 #define path_root_el(_path) ((_path)->p_node[0].el)
77 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
78 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
79 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
82 * Reset the actual path elements so that we can re-use the structure
83 * to build another path. Generally, this involves freeing the buffer
86 static void ocfs2_reinit_path(struct ocfs2_path
*path
, int keep_root
)
88 int i
, start
= 0, depth
= 0;
89 struct ocfs2_path_item
*node
;
94 for(i
= start
; i
< path_num_items(path
); i
++) {
95 node
= &path
->p_node
[i
];
103 * Tree depth may change during truncate, or insert. If we're
104 * keeping the root extent list, then make sure that our path
105 * structure reflects the proper depth.
108 depth
= le16_to_cpu(path_root_el(path
)->l_tree_depth
);
110 path
->p_tree_depth
= depth
;
113 static void ocfs2_free_path(struct ocfs2_path
*path
)
116 ocfs2_reinit_path(path
, 0);
122 * All the elements of src into dest. After this call, src could be freed
123 * without affecting dest.
125 * Both paths should have the same root. Any non-root elements of dest
128 static void ocfs2_cp_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
132 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
133 BUG_ON(path_root_el(dest
) != path_root_el(src
));
135 ocfs2_reinit_path(dest
, 1);
137 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
138 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
139 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
141 if (dest
->p_node
[i
].bh
)
142 get_bh(dest
->p_node
[i
].bh
);
147 * Make the *dest path the same as src and re-initialize src path to
150 static void ocfs2_mv_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
154 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
156 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
157 brelse(dest
->p_node
[i
].bh
);
159 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
160 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
162 src
->p_node
[i
].bh
= NULL
;
163 src
->p_node
[i
].el
= NULL
;
168 * Insert an extent block at given index.
170 * This will not take an additional reference on eb_bh.
172 static inline void ocfs2_path_insert_eb(struct ocfs2_path
*path
, int index
,
173 struct buffer_head
*eb_bh
)
175 struct ocfs2_extent_block
*eb
= (struct ocfs2_extent_block
*)eb_bh
->b_data
;
178 * Right now, no root bh is an extent block, so this helps
179 * catch code errors with dinode trees. The assertion can be
180 * safely removed if we ever need to insert extent block
181 * structures at the root.
185 path
->p_node
[index
].bh
= eb_bh
;
186 path
->p_node
[index
].el
= &eb
->h_list
;
189 static struct ocfs2_path
*ocfs2_new_path(struct buffer_head
*root_bh
,
190 struct ocfs2_extent_list
*root_el
)
192 struct ocfs2_path
*path
;
194 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) >= OCFS2_MAX_PATH_DEPTH
);
196 path
= kzalloc(sizeof(*path
), GFP_NOFS
);
198 path
->p_tree_depth
= le16_to_cpu(root_el
->l_tree_depth
);
200 path_root_bh(path
) = root_bh
;
201 path_root_el(path
) = root_el
;
208 * Allocate and initialize a new path based on a disk inode tree.
210 static struct ocfs2_path
*ocfs2_new_inode_path(struct buffer_head
*di_bh
)
212 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
213 struct ocfs2_extent_list
*el
= &di
->id2
.i_list
;
215 return ocfs2_new_path(di_bh
, el
);
219 * Convenience function to journal all components in a path.
221 static int ocfs2_journal_access_path(struct inode
*inode
, handle_t
*handle
,
222 struct ocfs2_path
*path
)
229 for(i
= 0; i
< path_num_items(path
); i
++) {
230 ret
= ocfs2_journal_access(handle
, inode
, path
->p_node
[i
].bh
,
231 OCFS2_JOURNAL_ACCESS_WRITE
);
243 * Return the index of the extent record which contains cluster #v_cluster.
244 * -1 is returned if it was not found.
246 * Should work fine on interior and exterior nodes.
248 int ocfs2_search_extent_list(struct ocfs2_extent_list
*el
, u32 v_cluster
)
252 struct ocfs2_extent_rec
*rec
;
253 u32 rec_end
, rec_start
, clusters
;
255 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
256 rec
= &el
->l_recs
[i
];
258 rec_start
= le32_to_cpu(rec
->e_cpos
);
259 clusters
= ocfs2_rec_clusters(el
, rec
);
261 rec_end
= rec_start
+ clusters
;
263 if (v_cluster
>= rec_start
&& v_cluster
< rec_end
) {
272 enum ocfs2_contig_type
{
281 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
282 * ocfs2_extent_contig only work properly against leaf nodes!
284 static int ocfs2_block_extent_contig(struct super_block
*sb
,
285 struct ocfs2_extent_rec
*ext
,
288 u64 blk_end
= le64_to_cpu(ext
->e_blkno
);
290 blk_end
+= ocfs2_clusters_to_blocks(sb
,
291 le16_to_cpu(ext
->e_leaf_clusters
));
293 return blkno
== blk_end
;
296 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec
*left
,
297 struct ocfs2_extent_rec
*right
)
301 left_range
= le32_to_cpu(left
->e_cpos
) +
302 le16_to_cpu(left
->e_leaf_clusters
);
304 return (left_range
== le32_to_cpu(right
->e_cpos
));
307 static enum ocfs2_contig_type
308 ocfs2_extent_contig(struct inode
*inode
,
309 struct ocfs2_extent_rec
*ext
,
310 struct ocfs2_extent_rec
*insert_rec
)
312 u64 blkno
= le64_to_cpu(insert_rec
->e_blkno
);
315 * Refuse to coalesce extent records with different flag
316 * fields - we don't want to mix unwritten extents with user
319 if (ext
->e_flags
!= insert_rec
->e_flags
)
322 if (ocfs2_extents_adjacent(ext
, insert_rec
) &&
323 ocfs2_block_extent_contig(inode
->i_sb
, ext
, blkno
))
326 blkno
= le64_to_cpu(ext
->e_blkno
);
327 if (ocfs2_extents_adjacent(insert_rec
, ext
) &&
328 ocfs2_block_extent_contig(inode
->i_sb
, insert_rec
, blkno
))
335 * NOTE: We can have pretty much any combination of contiguousness and
338 * The usefulness of APPEND_TAIL is more in that it lets us know that
339 * we'll have to update the path to that leaf.
341 enum ocfs2_append_type
{
346 enum ocfs2_split_type
{
352 struct ocfs2_insert_type
{
353 enum ocfs2_split_type ins_split
;
354 enum ocfs2_append_type ins_appending
;
355 enum ocfs2_contig_type ins_contig
;
356 int ins_contig_index
;
360 struct ocfs2_merge_ctxt
{
361 enum ocfs2_contig_type c_contig_type
;
362 int c_has_empty_extent
;
363 int c_split_covers_rec
;
367 * How many free extents have we got before we need more meta data?
369 int ocfs2_num_free_extents(struct ocfs2_super
*osb
,
371 struct buffer_head
*bh
)
374 struct ocfs2_extent_list
*el
;
375 struct ocfs2_extent_block
*eb
;
376 struct buffer_head
*eb_bh
= NULL
;
377 struct ocfs2_dinode
*fe
= (struct ocfs2_dinode
*)bh
->b_data
;
381 if (!OCFS2_IS_VALID_DINODE(fe
)) {
382 OCFS2_RO_ON_INVALID_DINODE(inode
->i_sb
, fe
);
387 if (fe
->i_last_eb_blk
) {
388 retval
= ocfs2_read_block(osb
, le64_to_cpu(fe
->i_last_eb_blk
),
389 &eb_bh
, OCFS2_BH_CACHED
, inode
);
394 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
397 el
= &fe
->id2
.i_list
;
399 BUG_ON(el
->l_tree_depth
!= 0);
401 retval
= le16_to_cpu(el
->l_count
) - le16_to_cpu(el
->l_next_free_rec
);
410 /* expects array to already be allocated
412 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
415 static int ocfs2_create_new_meta_bhs(struct ocfs2_super
*osb
,
419 struct ocfs2_alloc_context
*meta_ac
,
420 struct buffer_head
*bhs
[])
422 int count
, status
, i
;
423 u16 suballoc_bit_start
;
426 struct ocfs2_extent_block
*eb
;
431 while (count
< wanted
) {
432 status
= ocfs2_claim_metadata(osb
,
444 for(i
= count
; i
< (num_got
+ count
); i
++) {
445 bhs
[i
] = sb_getblk(osb
->sb
, first_blkno
);
446 if (bhs
[i
] == NULL
) {
451 ocfs2_set_new_buffer_uptodate(inode
, bhs
[i
]);
453 status
= ocfs2_journal_access(handle
, inode
, bhs
[i
],
454 OCFS2_JOURNAL_ACCESS_CREATE
);
460 memset(bhs
[i
]->b_data
, 0, osb
->sb
->s_blocksize
);
461 eb
= (struct ocfs2_extent_block
*) bhs
[i
]->b_data
;
462 /* Ok, setup the minimal stuff here. */
463 strcpy(eb
->h_signature
, OCFS2_EXTENT_BLOCK_SIGNATURE
);
464 eb
->h_blkno
= cpu_to_le64(first_blkno
);
465 eb
->h_fs_generation
= cpu_to_le32(osb
->fs_generation
);
466 eb
->h_suballoc_slot
= cpu_to_le16(osb
->slot_num
);
467 eb
->h_suballoc_bit
= cpu_to_le16(suballoc_bit_start
);
469 cpu_to_le16(ocfs2_extent_recs_per_eb(osb
->sb
));
471 suballoc_bit_start
++;
474 /* We'll also be dirtied by the caller, so
475 * this isn't absolutely necessary. */
476 status
= ocfs2_journal_dirty(handle
, bhs
[i
]);
489 for(i
= 0; i
< wanted
; i
++) {
500 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
502 * Returns the sum of the rightmost extent rec logical offset and
505 * ocfs2_add_branch() uses this to determine what logical cluster
506 * value should be populated into the leftmost new branch records.
508 * ocfs2_shift_tree_depth() uses this to determine the # clusters
509 * value for the new topmost tree record.
511 static inline u32
ocfs2_sum_rightmost_rec(struct ocfs2_extent_list
*el
)
515 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
517 return le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
518 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
522 * Add an entire tree branch to our inode. eb_bh is the extent block
523 * to start at, if we don't want to start the branch at the dinode
526 * last_eb_bh is required as we have to update it's next_leaf pointer
527 * for the new last extent block.
529 * the new branch will be 'empty' in the sense that every block will
530 * contain a single record with cluster count == 0.
532 static int ocfs2_add_branch(struct ocfs2_super
*osb
,
535 struct buffer_head
*fe_bh
,
536 struct buffer_head
*eb_bh
,
537 struct buffer_head
**last_eb_bh
,
538 struct ocfs2_alloc_context
*meta_ac
)
540 int status
, new_blocks
, i
;
541 u64 next_blkno
, new_last_eb_blk
;
542 struct buffer_head
*bh
;
543 struct buffer_head
**new_eb_bhs
= NULL
;
544 struct ocfs2_dinode
*fe
;
545 struct ocfs2_extent_block
*eb
;
546 struct ocfs2_extent_list
*eb_el
;
547 struct ocfs2_extent_list
*el
;
552 BUG_ON(!last_eb_bh
|| !*last_eb_bh
);
554 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
557 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
560 el
= &fe
->id2
.i_list
;
562 /* we never add a branch to a leaf. */
563 BUG_ON(!el
->l_tree_depth
);
565 new_blocks
= le16_to_cpu(el
->l_tree_depth
);
567 /* allocate the number of new eb blocks we need */
568 new_eb_bhs
= kcalloc(new_blocks
, sizeof(struct buffer_head
*),
576 status
= ocfs2_create_new_meta_bhs(osb
, handle
, inode
, new_blocks
,
577 meta_ac
, new_eb_bhs
);
583 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
584 new_cpos
= ocfs2_sum_rightmost_rec(&eb
->h_list
);
586 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
587 * linked with the rest of the tree.
588 * conversly, new_eb_bhs[0] is the new bottommost leaf.
590 * when we leave the loop, new_last_eb_blk will point to the
591 * newest leaf, and next_blkno will point to the topmost extent
593 next_blkno
= new_last_eb_blk
= 0;
594 for(i
= 0; i
< new_blocks
; i
++) {
596 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
597 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
598 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
604 status
= ocfs2_journal_access(handle
, inode
, bh
,
605 OCFS2_JOURNAL_ACCESS_CREATE
);
611 eb
->h_next_leaf_blk
= 0;
612 eb_el
->l_tree_depth
= cpu_to_le16(i
);
613 eb_el
->l_next_free_rec
= cpu_to_le16(1);
615 * This actually counts as an empty extent as
618 eb_el
->l_recs
[0].e_cpos
= cpu_to_le32(new_cpos
);
619 eb_el
->l_recs
[0].e_blkno
= cpu_to_le64(next_blkno
);
621 * eb_el isn't always an interior node, but even leaf
622 * nodes want a zero'd flags and reserved field so
623 * this gets the whole 32 bits regardless of use.
625 eb_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(0);
626 if (!eb_el
->l_tree_depth
)
627 new_last_eb_blk
= le64_to_cpu(eb
->h_blkno
);
629 status
= ocfs2_journal_dirty(handle
, bh
);
635 next_blkno
= le64_to_cpu(eb
->h_blkno
);
638 /* This is a bit hairy. We want to update up to three blocks
639 * here without leaving any of them in an inconsistent state
640 * in case of error. We don't have to worry about
641 * journal_dirty erroring as it won't unless we've aborted the
642 * handle (in which case we would never be here) so reserving
643 * the write with journal_access is all we need to do. */
644 status
= ocfs2_journal_access(handle
, inode
, *last_eb_bh
,
645 OCFS2_JOURNAL_ACCESS_WRITE
);
650 status
= ocfs2_journal_access(handle
, inode
, fe_bh
,
651 OCFS2_JOURNAL_ACCESS_WRITE
);
657 status
= ocfs2_journal_access(handle
, inode
, eb_bh
,
658 OCFS2_JOURNAL_ACCESS_WRITE
);
665 /* Link the new branch into the rest of the tree (el will
666 * either be on the fe, or the extent block passed in. */
667 i
= le16_to_cpu(el
->l_next_free_rec
);
668 el
->l_recs
[i
].e_blkno
= cpu_to_le64(next_blkno
);
669 el
->l_recs
[i
].e_cpos
= cpu_to_le32(new_cpos
);
670 el
->l_recs
[i
].e_int_clusters
= 0;
671 le16_add_cpu(&el
->l_next_free_rec
, 1);
673 /* fe needs a new last extent block pointer, as does the
674 * next_leaf on the previously last-extent-block. */
675 fe
->i_last_eb_blk
= cpu_to_le64(new_last_eb_blk
);
677 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
678 eb
->h_next_leaf_blk
= cpu_to_le64(new_last_eb_blk
);
680 status
= ocfs2_journal_dirty(handle
, *last_eb_bh
);
683 status
= ocfs2_journal_dirty(handle
, fe_bh
);
687 status
= ocfs2_journal_dirty(handle
, eb_bh
);
693 * Some callers want to track the rightmost leaf so pass it
697 get_bh(new_eb_bhs
[0]);
698 *last_eb_bh
= new_eb_bhs
[0];
703 for (i
= 0; i
< new_blocks
; i
++)
705 brelse(new_eb_bhs
[i
]);
714 * adds another level to the allocation tree.
715 * returns back the new extent block so you can add a branch to it
718 static int ocfs2_shift_tree_depth(struct ocfs2_super
*osb
,
721 struct buffer_head
*fe_bh
,
722 struct ocfs2_alloc_context
*meta_ac
,
723 struct buffer_head
**ret_new_eb_bh
)
727 struct buffer_head
*new_eb_bh
= NULL
;
728 struct ocfs2_dinode
*fe
;
729 struct ocfs2_extent_block
*eb
;
730 struct ocfs2_extent_list
*fe_el
;
731 struct ocfs2_extent_list
*eb_el
;
735 status
= ocfs2_create_new_meta_bhs(osb
, handle
, inode
, 1, meta_ac
,
742 eb
= (struct ocfs2_extent_block
*) new_eb_bh
->b_data
;
743 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
744 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
750 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
751 fe_el
= &fe
->id2
.i_list
;
753 status
= ocfs2_journal_access(handle
, inode
, new_eb_bh
,
754 OCFS2_JOURNAL_ACCESS_CREATE
);
760 /* copy the fe data into the new extent block */
761 eb_el
->l_tree_depth
= fe_el
->l_tree_depth
;
762 eb_el
->l_next_free_rec
= fe_el
->l_next_free_rec
;
763 for(i
= 0; i
< le16_to_cpu(fe_el
->l_next_free_rec
); i
++)
764 eb_el
->l_recs
[i
] = fe_el
->l_recs
[i
];
766 status
= ocfs2_journal_dirty(handle
, new_eb_bh
);
772 status
= ocfs2_journal_access(handle
, inode
, fe_bh
,
773 OCFS2_JOURNAL_ACCESS_WRITE
);
779 new_clusters
= ocfs2_sum_rightmost_rec(eb_el
);
782 le16_add_cpu(&fe_el
->l_tree_depth
, 1);
783 fe_el
->l_recs
[0].e_cpos
= 0;
784 fe_el
->l_recs
[0].e_blkno
= eb
->h_blkno
;
785 fe_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(new_clusters
);
786 for(i
= 1; i
< le16_to_cpu(fe_el
->l_next_free_rec
); i
++)
787 memset(&fe_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
788 fe_el
->l_next_free_rec
= cpu_to_le16(1);
790 /* If this is our 1st tree depth shift, then last_eb_blk
791 * becomes the allocated extent block */
792 if (fe_el
->l_tree_depth
== cpu_to_le16(1))
793 fe
->i_last_eb_blk
= eb
->h_blkno
;
795 status
= ocfs2_journal_dirty(handle
, fe_bh
);
801 *ret_new_eb_bh
= new_eb_bh
;
813 * Should only be called when there is no space left in any of the
814 * leaf nodes. What we want to do is find the lowest tree depth
815 * non-leaf extent block with room for new records. There are three
816 * valid results of this search:
818 * 1) a lowest extent block is found, then we pass it back in
819 * *lowest_eb_bh and return '0'
821 * 2) the search fails to find anything, but the dinode has room. We
822 * pass NULL back in *lowest_eb_bh, but still return '0'
824 * 3) the search fails to find anything AND the dinode is full, in
825 * which case we return > 0
827 * return status < 0 indicates an error.
829 static int ocfs2_find_branch_target(struct ocfs2_super
*osb
,
831 struct buffer_head
*fe_bh
,
832 struct buffer_head
**target_bh
)
836 struct ocfs2_dinode
*fe
;
837 struct ocfs2_extent_block
*eb
;
838 struct ocfs2_extent_list
*el
;
839 struct buffer_head
*bh
= NULL
;
840 struct buffer_head
*lowest_bh
= NULL
;
846 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
847 el
= &fe
->id2
.i_list
;
849 while(le16_to_cpu(el
->l_tree_depth
) > 1) {
850 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
851 ocfs2_error(inode
->i_sb
, "Dinode %llu has empty "
852 "extent list (next_free_rec == 0)",
853 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
857 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
858 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
860 ocfs2_error(inode
->i_sb
, "Dinode %llu has extent "
861 "list where extent # %d has no physical "
863 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, i
);
873 status
= ocfs2_read_block(osb
, blkno
, &bh
, OCFS2_BH_CACHED
,
880 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
881 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
882 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
888 if (le16_to_cpu(el
->l_next_free_rec
) <
889 le16_to_cpu(el
->l_count
)) {
897 /* If we didn't find one and the fe doesn't have any room,
900 && (fe
->id2
.i_list
.l_next_free_rec
== fe
->id2
.i_list
.l_count
))
903 *target_bh
= lowest_bh
;
913 * Grow a b-tree so that it has more records.
915 * We might shift the tree depth in which case existing paths should
916 * be considered invalid.
918 * Tree depth after the grow is returned via *final_depth.
920 * *last_eb_bh will be updated by ocfs2_add_branch().
922 static int ocfs2_grow_tree(struct inode
*inode
, handle_t
*handle
,
923 struct buffer_head
*di_bh
, int *final_depth
,
924 struct buffer_head
**last_eb_bh
,
925 struct ocfs2_alloc_context
*meta_ac
)
928 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
929 int depth
= le16_to_cpu(di
->id2
.i_list
.l_tree_depth
);
930 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
931 struct buffer_head
*bh
= NULL
;
933 BUG_ON(meta_ac
== NULL
);
935 shift
= ocfs2_find_branch_target(osb
, inode
, di_bh
, &bh
);
942 /* We traveled all the way to the bottom of the allocation tree
943 * and didn't find room for any more extents - we need to add
944 * another tree level */
947 mlog(0, "need to shift tree depth (current = %d)\n", depth
);
949 /* ocfs2_shift_tree_depth will return us a buffer with
950 * the new extent block (so we can pass that to
951 * ocfs2_add_branch). */
952 ret
= ocfs2_shift_tree_depth(osb
, handle
, inode
, di_bh
,
961 * Special case: we have room now if we shifted from
962 * tree_depth 0, so no more work needs to be done.
964 * We won't be calling add_branch, so pass
965 * back *last_eb_bh as the new leaf. At depth
966 * zero, it should always be null so there's
967 * no reason to brelse.
976 /* call ocfs2_add_branch to add the final part of the tree with
978 mlog(0, "add branch. bh = %p\n", bh
);
979 ret
= ocfs2_add_branch(osb
, handle
, inode
, di_bh
, bh
, last_eb_bh
,
988 *final_depth
= depth
;
994 * This function will discard the rightmost extent record.
996 static void ocfs2_shift_records_right(struct ocfs2_extent_list
*el
)
998 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
999 int count
= le16_to_cpu(el
->l_count
);
1000 unsigned int num_bytes
;
1003 /* This will cause us to go off the end of our extent list. */
1004 BUG_ON(next_free
>= count
);
1006 num_bytes
= sizeof(struct ocfs2_extent_rec
) * next_free
;
1008 memmove(&el
->l_recs
[1], &el
->l_recs
[0], num_bytes
);
1011 static void ocfs2_rotate_leaf(struct ocfs2_extent_list
*el
,
1012 struct ocfs2_extent_rec
*insert_rec
)
1014 int i
, insert_index
, next_free
, has_empty
, num_bytes
;
1015 u32 insert_cpos
= le32_to_cpu(insert_rec
->e_cpos
);
1016 struct ocfs2_extent_rec
*rec
;
1018 next_free
= le16_to_cpu(el
->l_next_free_rec
);
1019 has_empty
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
1023 /* The tree code before us didn't allow enough room in the leaf. */
1024 BUG_ON(el
->l_next_free_rec
== el
->l_count
&& !has_empty
);
1027 * The easiest way to approach this is to just remove the
1028 * empty extent and temporarily decrement next_free.
1032 * If next_free was 1 (only an empty extent), this
1033 * loop won't execute, which is fine. We still want
1034 * the decrement above to happen.
1036 for(i
= 0; i
< (next_free
- 1); i
++)
1037 el
->l_recs
[i
] = el
->l_recs
[i
+1];
1043 * Figure out what the new record index should be.
1045 for(i
= 0; i
< next_free
; i
++) {
1046 rec
= &el
->l_recs
[i
];
1048 if (insert_cpos
< le32_to_cpu(rec
->e_cpos
))
1053 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1054 insert_cpos
, insert_index
, has_empty
, next_free
, le16_to_cpu(el
->l_count
));
1056 BUG_ON(insert_index
< 0);
1057 BUG_ON(insert_index
>= le16_to_cpu(el
->l_count
));
1058 BUG_ON(insert_index
> next_free
);
1061 * No need to memmove if we're just adding to the tail.
1063 if (insert_index
!= next_free
) {
1064 BUG_ON(next_free
>= le16_to_cpu(el
->l_count
));
1066 num_bytes
= next_free
- insert_index
;
1067 num_bytes
*= sizeof(struct ocfs2_extent_rec
);
1068 memmove(&el
->l_recs
[insert_index
+ 1],
1069 &el
->l_recs
[insert_index
],
1074 * Either we had an empty extent, and need to re-increment or
1075 * there was no empty extent on a non full rightmost leaf node,
1076 * in which case we still need to increment.
1079 el
->l_next_free_rec
= cpu_to_le16(next_free
);
1081 * Make sure none of the math above just messed up our tree.
1083 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) > le16_to_cpu(el
->l_count
));
1085 el
->l_recs
[insert_index
] = *insert_rec
;
1089 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list
*el
)
1091 int size
, num_recs
= le16_to_cpu(el
->l_next_free_rec
);
1093 BUG_ON(num_recs
== 0);
1095 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
1097 size
= num_recs
* sizeof(struct ocfs2_extent_rec
);
1098 memmove(&el
->l_recs
[0], &el
->l_recs
[1], size
);
1099 memset(&el
->l_recs
[num_recs
], 0,
1100 sizeof(struct ocfs2_extent_rec
));
1101 el
->l_next_free_rec
= cpu_to_le16(num_recs
);
1106 * Create an empty extent record .
1108 * l_next_free_rec may be updated.
1110 * If an empty extent already exists do nothing.
1112 static void ocfs2_create_empty_extent(struct ocfs2_extent_list
*el
)
1114 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1116 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
1121 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
1124 mlog_bug_on_msg(el
->l_count
== el
->l_next_free_rec
,
1125 "Asked to create an empty extent in a full list:\n"
1126 "count = %u, tree depth = %u",
1127 le16_to_cpu(el
->l_count
),
1128 le16_to_cpu(el
->l_tree_depth
));
1130 ocfs2_shift_records_right(el
);
1133 le16_add_cpu(&el
->l_next_free_rec
, 1);
1134 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1138 * For a rotation which involves two leaf nodes, the "root node" is
1139 * the lowest level tree node which contains a path to both leafs. This
1140 * resulting set of information can be used to form a complete "subtree"
1142 * This function is passed two full paths from the dinode down to a
1143 * pair of adjacent leaves. It's task is to figure out which path
1144 * index contains the subtree root - this can be the root index itself
1145 * in a worst-case rotation.
1147 * The array index of the subtree root is passed back.
1149 static int ocfs2_find_subtree_root(struct inode
*inode
,
1150 struct ocfs2_path
*left
,
1151 struct ocfs2_path
*right
)
1156 * Check that the caller passed in two paths from the same tree.
1158 BUG_ON(path_root_bh(left
) != path_root_bh(right
));
1164 * The caller didn't pass two adjacent paths.
1166 mlog_bug_on_msg(i
> left
->p_tree_depth
,
1167 "Inode %lu, left depth %u, right depth %u\n"
1168 "left leaf blk %llu, right leaf blk %llu\n",
1169 inode
->i_ino
, left
->p_tree_depth
,
1170 right
->p_tree_depth
,
1171 (unsigned long long)path_leaf_bh(left
)->b_blocknr
,
1172 (unsigned long long)path_leaf_bh(right
)->b_blocknr
);
1173 } while (left
->p_node
[i
].bh
->b_blocknr
==
1174 right
->p_node
[i
].bh
->b_blocknr
);
1179 typedef void (path_insert_t
)(void *, struct buffer_head
*);
1182 * Traverse a btree path in search of cpos, starting at root_el.
1184 * This code can be called with a cpos larger than the tree, in which
1185 * case it will return the rightmost path.
1187 static int __ocfs2_find_path(struct inode
*inode
,
1188 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1189 path_insert_t
*func
, void *data
)
1194 struct buffer_head
*bh
= NULL
;
1195 struct ocfs2_extent_block
*eb
;
1196 struct ocfs2_extent_list
*el
;
1197 struct ocfs2_extent_rec
*rec
;
1198 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1201 while (el
->l_tree_depth
) {
1202 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1203 ocfs2_error(inode
->i_sb
,
1204 "Inode %llu has empty extent list at "
1206 (unsigned long long)oi
->ip_blkno
,
1207 le16_to_cpu(el
->l_tree_depth
));
1213 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
) - 1; i
++) {
1214 rec
= &el
->l_recs
[i
];
1217 * In the case that cpos is off the allocation
1218 * tree, this should just wind up returning the
1221 range
= le32_to_cpu(rec
->e_cpos
) +
1222 ocfs2_rec_clusters(el
, rec
);
1223 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1227 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1229 ocfs2_error(inode
->i_sb
,
1230 "Inode %llu has bad blkno in extent list "
1231 "at depth %u (index %d)\n",
1232 (unsigned long long)oi
->ip_blkno
,
1233 le16_to_cpu(el
->l_tree_depth
), i
);
1240 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
), blkno
,
1241 &bh
, OCFS2_BH_CACHED
, inode
);
1247 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1249 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
1250 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
1255 if (le16_to_cpu(el
->l_next_free_rec
) >
1256 le16_to_cpu(el
->l_count
)) {
1257 ocfs2_error(inode
->i_sb
,
1258 "Inode %llu has bad count in extent list "
1259 "at block %llu (next free=%u, count=%u)\n",
1260 (unsigned long long)oi
->ip_blkno
,
1261 (unsigned long long)bh
->b_blocknr
,
1262 le16_to_cpu(el
->l_next_free_rec
),
1263 le16_to_cpu(el
->l_count
));
1274 * Catch any trailing bh that the loop didn't handle.
1282 * Given an initialized path (that is, it has a valid root extent
1283 * list), this function will traverse the btree in search of the path
1284 * which would contain cpos.
1286 * The path traveled is recorded in the path structure.
1288 * Note that this will not do any comparisons on leaf node extent
1289 * records, so it will work fine in the case that we just added a tree
1292 struct find_path_data
{
1294 struct ocfs2_path
*path
;
1296 static void find_path_ins(void *data
, struct buffer_head
*bh
)
1298 struct find_path_data
*fp
= data
;
1301 ocfs2_path_insert_eb(fp
->path
, fp
->index
, bh
);
1304 static int ocfs2_find_path(struct inode
*inode
, struct ocfs2_path
*path
,
1307 struct find_path_data data
;
1311 return __ocfs2_find_path(inode
, path_root_el(path
), cpos
,
1312 find_path_ins
, &data
);
1315 static void find_leaf_ins(void *data
, struct buffer_head
*bh
)
1317 struct ocfs2_extent_block
*eb
=(struct ocfs2_extent_block
*)bh
->b_data
;
1318 struct ocfs2_extent_list
*el
= &eb
->h_list
;
1319 struct buffer_head
**ret
= data
;
1321 /* We want to retain only the leaf block. */
1322 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
1328 * Find the leaf block in the tree which would contain cpos. No
1329 * checking of the actual leaf is done.
1331 * Some paths want to call this instead of allocating a path structure
1332 * and calling ocfs2_find_path().
1334 * This function doesn't handle non btree extent lists.
1336 int ocfs2_find_leaf(struct inode
*inode
, struct ocfs2_extent_list
*root_el
,
1337 u32 cpos
, struct buffer_head
**leaf_bh
)
1340 struct buffer_head
*bh
= NULL
;
1342 ret
= __ocfs2_find_path(inode
, root_el
, cpos
, find_leaf_ins
, &bh
);
1354 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1356 * Basically, we've moved stuff around at the bottom of the tree and
1357 * we need to fix up the extent records above the changes to reflect
1360 * left_rec: the record on the left.
1361 * left_child_el: is the child list pointed to by left_rec
1362 * right_rec: the record to the right of left_rec
1363 * right_child_el: is the child list pointed to by right_rec
1365 * By definition, this only works on interior nodes.
1367 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec
*left_rec
,
1368 struct ocfs2_extent_list
*left_child_el
,
1369 struct ocfs2_extent_rec
*right_rec
,
1370 struct ocfs2_extent_list
*right_child_el
)
1372 u32 left_clusters
, right_end
;
1375 * Interior nodes never have holes. Their cpos is the cpos of
1376 * the leftmost record in their child list. Their cluster
1377 * count covers the full theoretical range of their child list
1378 * - the range between their cpos and the cpos of the record
1379 * immediately to their right.
1381 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[0].e_cpos
);
1382 if (ocfs2_is_empty_extent(&right_child_el
->l_recs
[0])) {
1383 BUG_ON(le16_to_cpu(right_child_el
->l_next_free_rec
) <= 1);
1384 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[1].e_cpos
);
1386 left_clusters
-= le32_to_cpu(left_rec
->e_cpos
);
1387 left_rec
->e_int_clusters
= cpu_to_le32(left_clusters
);
1390 * Calculate the rightmost cluster count boundary before
1391 * moving cpos - we will need to adjust clusters after
1392 * updating e_cpos to keep the same highest cluster count.
1394 right_end
= le32_to_cpu(right_rec
->e_cpos
);
1395 right_end
+= le32_to_cpu(right_rec
->e_int_clusters
);
1397 right_rec
->e_cpos
= left_rec
->e_cpos
;
1398 le32_add_cpu(&right_rec
->e_cpos
, left_clusters
);
1400 right_end
-= le32_to_cpu(right_rec
->e_cpos
);
1401 right_rec
->e_int_clusters
= cpu_to_le32(right_end
);
1405 * Adjust the adjacent root node records involved in a
1406 * rotation. left_el_blkno is passed in as a key so that we can easily
1407 * find it's index in the root list.
1409 static void ocfs2_adjust_root_records(struct ocfs2_extent_list
*root_el
,
1410 struct ocfs2_extent_list
*left_el
,
1411 struct ocfs2_extent_list
*right_el
,
1416 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) <=
1417 le16_to_cpu(left_el
->l_tree_depth
));
1419 for(i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
) - 1; i
++) {
1420 if (le64_to_cpu(root_el
->l_recs
[i
].e_blkno
) == left_el_blkno
)
1425 * The path walking code should have never returned a root and
1426 * two paths which are not adjacent.
1428 BUG_ON(i
>= (le16_to_cpu(root_el
->l_next_free_rec
) - 1));
1430 ocfs2_adjust_adjacent_records(&root_el
->l_recs
[i
], left_el
,
1431 &root_el
->l_recs
[i
+ 1], right_el
);
1435 * We've changed a leaf block (in right_path) and need to reflect that
1436 * change back up the subtree.
1438 * This happens in multiple places:
1439 * - When we've moved an extent record from the left path leaf to the right
1440 * path leaf to make room for an empty extent in the left path leaf.
1441 * - When our insert into the right path leaf is at the leftmost edge
1442 * and requires an update of the path immediately to it's left. This
1443 * can occur at the end of some types of rotation and appending inserts.
1444 * - When we've adjusted the last extent record in the left path leaf and the
1445 * 1st extent record in the right path leaf during cross extent block merge.
1447 static void ocfs2_complete_edge_insert(struct inode
*inode
, handle_t
*handle
,
1448 struct ocfs2_path
*left_path
,
1449 struct ocfs2_path
*right_path
,
1453 struct ocfs2_extent_list
*el
, *left_el
, *right_el
;
1454 struct ocfs2_extent_rec
*left_rec
, *right_rec
;
1455 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
1458 * Update the counts and position values within all the
1459 * interior nodes to reflect the leaf rotation we just did.
1461 * The root node is handled below the loop.
1463 * We begin the loop with right_el and left_el pointing to the
1464 * leaf lists and work our way up.
1466 * NOTE: within this loop, left_el and right_el always refer
1467 * to the *child* lists.
1469 left_el
= path_leaf_el(left_path
);
1470 right_el
= path_leaf_el(right_path
);
1471 for(i
= left_path
->p_tree_depth
- 1; i
> subtree_index
; i
--) {
1472 mlog(0, "Adjust records at index %u\n", i
);
1475 * One nice property of knowing that all of these
1476 * nodes are below the root is that we only deal with
1477 * the leftmost right node record and the rightmost
1480 el
= left_path
->p_node
[i
].el
;
1481 idx
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
1482 left_rec
= &el
->l_recs
[idx
];
1484 el
= right_path
->p_node
[i
].el
;
1485 right_rec
= &el
->l_recs
[0];
1487 ocfs2_adjust_adjacent_records(left_rec
, left_el
, right_rec
,
1490 ret
= ocfs2_journal_dirty(handle
, left_path
->p_node
[i
].bh
);
1494 ret
= ocfs2_journal_dirty(handle
, right_path
->p_node
[i
].bh
);
1499 * Setup our list pointers now so that the current
1500 * parents become children in the next iteration.
1502 left_el
= left_path
->p_node
[i
].el
;
1503 right_el
= right_path
->p_node
[i
].el
;
1507 * At the root node, adjust the two adjacent records which
1508 * begin our path to the leaves.
1511 el
= left_path
->p_node
[subtree_index
].el
;
1512 left_el
= left_path
->p_node
[subtree_index
+ 1].el
;
1513 right_el
= right_path
->p_node
[subtree_index
+ 1].el
;
1515 ocfs2_adjust_root_records(el
, left_el
, right_el
,
1516 left_path
->p_node
[subtree_index
+ 1].bh
->b_blocknr
);
1518 root_bh
= left_path
->p_node
[subtree_index
].bh
;
1520 ret
= ocfs2_journal_dirty(handle
, root_bh
);
1525 static int ocfs2_rotate_subtree_right(struct inode
*inode
,
1527 struct ocfs2_path
*left_path
,
1528 struct ocfs2_path
*right_path
,
1532 struct buffer_head
*right_leaf_bh
;
1533 struct buffer_head
*left_leaf_bh
= NULL
;
1534 struct buffer_head
*root_bh
;
1535 struct ocfs2_extent_list
*right_el
, *left_el
;
1536 struct ocfs2_extent_rec move_rec
;
1538 left_leaf_bh
= path_leaf_bh(left_path
);
1539 left_el
= path_leaf_el(left_path
);
1541 if (left_el
->l_next_free_rec
!= left_el
->l_count
) {
1542 ocfs2_error(inode
->i_sb
,
1543 "Inode %llu has non-full interior leaf node %llu"
1545 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1546 (unsigned long long)left_leaf_bh
->b_blocknr
,
1547 le16_to_cpu(left_el
->l_next_free_rec
));
1552 * This extent block may already have an empty record, so we
1553 * return early if so.
1555 if (ocfs2_is_empty_extent(&left_el
->l_recs
[0]))
1558 root_bh
= left_path
->p_node
[subtree_index
].bh
;
1559 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
1561 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
1562 OCFS2_JOURNAL_ACCESS_WRITE
);
1568 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
1569 ret
= ocfs2_journal_access(handle
, inode
,
1570 right_path
->p_node
[i
].bh
,
1571 OCFS2_JOURNAL_ACCESS_WRITE
);
1577 ret
= ocfs2_journal_access(handle
, inode
,
1578 left_path
->p_node
[i
].bh
,
1579 OCFS2_JOURNAL_ACCESS_WRITE
);
1586 right_leaf_bh
= path_leaf_bh(right_path
);
1587 right_el
= path_leaf_el(right_path
);
1589 /* This is a code error, not a disk corruption. */
1590 mlog_bug_on_msg(!right_el
->l_next_free_rec
, "Inode %llu: Rotate fails "
1591 "because rightmost leaf block %llu is empty\n",
1592 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1593 (unsigned long long)right_leaf_bh
->b_blocknr
);
1595 ocfs2_create_empty_extent(right_el
);
1597 ret
= ocfs2_journal_dirty(handle
, right_leaf_bh
);
1603 /* Do the copy now. */
1604 i
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
1605 move_rec
= left_el
->l_recs
[i
];
1606 right_el
->l_recs
[0] = move_rec
;
1609 * Clear out the record we just copied and shift everything
1610 * over, leaving an empty extent in the left leaf.
1612 * We temporarily subtract from next_free_rec so that the
1613 * shift will lose the tail record (which is now defunct).
1615 le16_add_cpu(&left_el
->l_next_free_rec
, -1);
1616 ocfs2_shift_records_right(left_el
);
1617 memset(&left_el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1618 le16_add_cpu(&left_el
->l_next_free_rec
, 1);
1620 ret
= ocfs2_journal_dirty(handle
, left_leaf_bh
);
1626 ocfs2_complete_edge_insert(inode
, handle
, left_path
, right_path
,
1634 * Given a full path, determine what cpos value would return us a path
1635 * containing the leaf immediately to the left of the current one.
1637 * Will return zero if the path passed in is already the leftmost path.
1639 static int ocfs2_find_cpos_for_left_leaf(struct super_block
*sb
,
1640 struct ocfs2_path
*path
, u32
*cpos
)
1644 struct ocfs2_extent_list
*el
;
1646 BUG_ON(path
->p_tree_depth
== 0);
1650 blkno
= path_leaf_bh(path
)->b_blocknr
;
1652 /* Start at the tree node just above the leaf and work our way up. */
1653 i
= path
->p_tree_depth
- 1;
1655 el
= path
->p_node
[i
].el
;
1658 * Find the extent record just before the one in our
1661 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
1662 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
1666 * We've determined that the
1667 * path specified is already
1668 * the leftmost one - return a
1674 * The leftmost record points to our
1675 * leaf - we need to travel up the
1681 *cpos
= le32_to_cpu(el
->l_recs
[j
- 1].e_cpos
);
1682 *cpos
= *cpos
+ ocfs2_rec_clusters(el
,
1683 &el
->l_recs
[j
- 1]);
1690 * If we got here, we never found a valid node where
1691 * the tree indicated one should be.
1694 "Invalid extent tree at extent block %llu\n",
1695 (unsigned long long)blkno
);
1700 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
1709 * Extend the transaction by enough credits to complete the rotation,
1710 * and still leave at least the original number of credits allocated
1711 * to this transaction.
1713 static int ocfs2_extend_rotate_transaction(handle_t
*handle
, int subtree_depth
,
1715 struct ocfs2_path
*path
)
1717 int credits
= (path
->p_tree_depth
- subtree_depth
) * 2 + 1 + op_credits
;
1719 if (handle
->h_buffer_credits
< credits
)
1720 return ocfs2_extend_trans(handle
, credits
);
1726 * Trap the case where we're inserting into the theoretical range past
1727 * the _actual_ left leaf range. Otherwise, we'll rotate a record
1728 * whose cpos is less than ours into the right leaf.
1730 * It's only necessary to look at the rightmost record of the left
1731 * leaf because the logic that calls us should ensure that the
1732 * theoretical ranges in the path components above the leaves are
1735 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path
*left_path
,
1738 struct ocfs2_extent_list
*left_el
;
1739 struct ocfs2_extent_rec
*rec
;
1742 left_el
= path_leaf_el(left_path
);
1743 next_free
= le16_to_cpu(left_el
->l_next_free_rec
);
1744 rec
= &left_el
->l_recs
[next_free
- 1];
1746 if (insert_cpos
> le32_to_cpu(rec
->e_cpos
))
1751 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list
*el
, u32 cpos
)
1753 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1755 struct ocfs2_extent_rec
*rec
;
1760 rec
= &el
->l_recs
[0];
1761 if (ocfs2_is_empty_extent(rec
)) {
1765 rec
= &el
->l_recs
[1];
1768 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
1769 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1775 * Rotate all the records in a btree right one record, starting at insert_cpos.
1777 * The path to the rightmost leaf should be passed in.
1779 * The array is assumed to be large enough to hold an entire path (tree depth).
1781 * Upon succesful return from this function:
1783 * - The 'right_path' array will contain a path to the leaf block
1784 * whose range contains e_cpos.
1785 * - That leaf block will have a single empty extent in list index 0.
1786 * - In the case that the rotation requires a post-insert update,
1787 * *ret_left_path will contain a valid path which can be passed to
1788 * ocfs2_insert_path().
1790 static int ocfs2_rotate_tree_right(struct inode
*inode
,
1792 enum ocfs2_split_type split
,
1794 struct ocfs2_path
*right_path
,
1795 struct ocfs2_path
**ret_left_path
)
1797 int ret
, start
, orig_credits
= handle
->h_buffer_credits
;
1799 struct ocfs2_path
*left_path
= NULL
;
1801 *ret_left_path
= NULL
;
1803 left_path
= ocfs2_new_path(path_root_bh(right_path
),
1804 path_root_el(right_path
));
1811 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
, &cpos
);
1817 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos
, cpos
);
1820 * What we want to do here is:
1822 * 1) Start with the rightmost path.
1824 * 2) Determine a path to the leaf block directly to the left
1827 * 3) Determine the 'subtree root' - the lowest level tree node
1828 * which contains a path to both leaves.
1830 * 4) Rotate the subtree.
1832 * 5) Find the next subtree by considering the left path to be
1833 * the new right path.
1835 * The check at the top of this while loop also accepts
1836 * insert_cpos == cpos because cpos is only a _theoretical_
1837 * value to get us the left path - insert_cpos might very well
1838 * be filling that hole.
1840 * Stop at a cpos of '0' because we either started at the
1841 * leftmost branch (i.e., a tree with one branch and a
1842 * rotation inside of it), or we've gone as far as we can in
1843 * rotating subtrees.
1845 while (cpos
&& insert_cpos
<= cpos
) {
1846 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
1849 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
1855 mlog_bug_on_msg(path_leaf_bh(left_path
) ==
1856 path_leaf_bh(right_path
),
1857 "Inode %lu: error during insert of %u "
1858 "(left path cpos %u) results in two identical "
1859 "paths ending at %llu\n",
1860 inode
->i_ino
, insert_cpos
, cpos
,
1861 (unsigned long long)
1862 path_leaf_bh(left_path
)->b_blocknr
);
1864 if (split
== SPLIT_NONE
&&
1865 ocfs2_rotate_requires_path_adjustment(left_path
,
1869 * We've rotated the tree as much as we
1870 * should. The rest is up to
1871 * ocfs2_insert_path() to complete, after the
1872 * record insertion. We indicate this
1873 * situation by returning the left path.
1875 * The reason we don't adjust the records here
1876 * before the record insert is that an error
1877 * later might break the rule where a parent
1878 * record e_cpos will reflect the actual
1879 * e_cpos of the 1st nonempty record of the
1882 *ret_left_path
= left_path
;
1886 start
= ocfs2_find_subtree_root(inode
, left_path
, right_path
);
1888 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
1890 (unsigned long long) right_path
->p_node
[start
].bh
->b_blocknr
,
1891 right_path
->p_tree_depth
);
1893 ret
= ocfs2_extend_rotate_transaction(handle
, start
,
1894 orig_credits
, right_path
);
1900 ret
= ocfs2_rotate_subtree_right(inode
, handle
, left_path
,
1907 if (split
!= SPLIT_NONE
&&
1908 ocfs2_leftmost_rec_contains(path_leaf_el(right_path
),
1911 * A rotate moves the rightmost left leaf
1912 * record over to the leftmost right leaf
1913 * slot. If we're doing an extent split
1914 * instead of a real insert, then we have to
1915 * check that the extent to be split wasn't
1916 * just moved over. If it was, then we can
1917 * exit here, passing left_path back -
1918 * ocfs2_split_extent() is smart enough to
1919 * search both leaves.
1921 *ret_left_path
= left_path
;
1926 * There is no need to re-read the next right path
1927 * as we know that it'll be our current left
1928 * path. Optimize by copying values instead.
1930 ocfs2_mv_path(right_path
, left_path
);
1932 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
,
1941 ocfs2_free_path(left_path
);
1947 static void ocfs2_update_edge_lengths(struct inode
*inode
, handle_t
*handle
,
1948 struct ocfs2_path
*path
)
1951 struct ocfs2_extent_rec
*rec
;
1952 struct ocfs2_extent_list
*el
;
1953 struct ocfs2_extent_block
*eb
;
1956 /* Path should always be rightmost. */
1957 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
1958 BUG_ON(eb
->h_next_leaf_blk
!= 0ULL);
1961 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
1962 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1963 rec
= &el
->l_recs
[idx
];
1964 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
1966 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
1967 el
= path
->p_node
[i
].el
;
1968 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1969 rec
= &el
->l_recs
[idx
];
1971 rec
->e_int_clusters
= cpu_to_le32(range
);
1972 le32_add_cpu(&rec
->e_int_clusters
, -le32_to_cpu(rec
->e_cpos
));
1974 ocfs2_journal_dirty(handle
, path
->p_node
[i
].bh
);
1978 static void ocfs2_unlink_path(struct inode
*inode
, handle_t
*handle
,
1979 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
1980 struct ocfs2_path
*path
, int unlink_start
)
1983 struct ocfs2_extent_block
*eb
;
1984 struct ocfs2_extent_list
*el
;
1985 struct buffer_head
*bh
;
1987 for(i
= unlink_start
; i
< path_num_items(path
); i
++) {
1988 bh
= path
->p_node
[i
].bh
;
1990 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
1992 * Not all nodes might have had their final count
1993 * decremented by the caller - handle this here.
1996 if (le16_to_cpu(el
->l_next_free_rec
) > 1) {
1998 "Inode %llu, attempted to remove extent block "
1999 "%llu with %u records\n",
2000 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2001 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
2002 le16_to_cpu(el
->l_next_free_rec
));
2004 ocfs2_journal_dirty(handle
, bh
);
2005 ocfs2_remove_from_cache(inode
, bh
);
2009 el
->l_next_free_rec
= 0;
2010 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2012 ocfs2_journal_dirty(handle
, bh
);
2014 ret
= ocfs2_cache_extent_block_free(dealloc
, eb
);
2018 ocfs2_remove_from_cache(inode
, bh
);
2022 static void ocfs2_unlink_subtree(struct inode
*inode
, handle_t
*handle
,
2023 struct ocfs2_path
*left_path
,
2024 struct ocfs2_path
*right_path
,
2026 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2029 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2030 struct ocfs2_extent_list
*root_el
= left_path
->p_node
[subtree_index
].el
;
2031 struct ocfs2_extent_list
*el
;
2032 struct ocfs2_extent_block
*eb
;
2034 el
= path_leaf_el(left_path
);
2036 eb
= (struct ocfs2_extent_block
*)right_path
->p_node
[subtree_index
+ 1].bh
->b_data
;
2038 for(i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
2039 if (root_el
->l_recs
[i
].e_blkno
== eb
->h_blkno
)
2042 BUG_ON(i
>= le16_to_cpu(root_el
->l_next_free_rec
));
2044 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
2045 le16_add_cpu(&root_el
->l_next_free_rec
, -1);
2047 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2048 eb
->h_next_leaf_blk
= 0;
2050 ocfs2_journal_dirty(handle
, root_bh
);
2051 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2053 ocfs2_unlink_path(inode
, handle
, dealloc
, right_path
,
2057 static int ocfs2_rotate_subtree_left(struct inode
*inode
, handle_t
*handle
,
2058 struct ocfs2_path
*left_path
,
2059 struct ocfs2_path
*right_path
,
2061 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2064 int ret
, i
, del_right_subtree
= 0, right_has_empty
= 0;
2065 struct buffer_head
*root_bh
, *di_bh
= path_root_bh(right_path
);
2066 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
2067 struct ocfs2_extent_list
*right_leaf_el
, *left_leaf_el
;
2068 struct ocfs2_extent_block
*eb
;
2072 right_leaf_el
= path_leaf_el(right_path
);
2073 left_leaf_el
= path_leaf_el(left_path
);
2074 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2075 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2077 if (!ocfs2_is_empty_extent(&left_leaf_el
->l_recs
[0]))
2080 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(right_path
)->b_data
;
2081 if (ocfs2_is_empty_extent(&right_leaf_el
->l_recs
[0])) {
2083 * It's legal for us to proceed if the right leaf is
2084 * the rightmost one and it has an empty extent. There
2085 * are two cases to handle - whether the leaf will be
2086 * empty after removal or not. If the leaf isn't empty
2087 * then just remove the empty extent up front. The
2088 * next block will handle empty leaves by flagging
2091 * Non rightmost leaves will throw -EAGAIN and the
2092 * caller can manually move the subtree and retry.
2095 if (eb
->h_next_leaf_blk
!= 0ULL)
2098 if (le16_to_cpu(right_leaf_el
->l_next_free_rec
) > 1) {
2099 ret
= ocfs2_journal_access(handle
, inode
,
2100 path_leaf_bh(right_path
),
2101 OCFS2_JOURNAL_ACCESS_WRITE
);
2107 ocfs2_remove_empty_extent(right_leaf_el
);
2109 right_has_empty
= 1;
2112 if (eb
->h_next_leaf_blk
== 0ULL &&
2113 le16_to_cpu(right_leaf_el
->l_next_free_rec
) == 1) {
2115 * We have to update i_last_eb_blk during the meta
2118 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
2119 OCFS2_JOURNAL_ACCESS_WRITE
);
2125 del_right_subtree
= 1;
2129 * Getting here with an empty extent in the right path implies
2130 * that it's the rightmost path and will be deleted.
2132 BUG_ON(right_has_empty
&& !del_right_subtree
);
2134 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
2135 OCFS2_JOURNAL_ACCESS_WRITE
);
2141 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2142 ret
= ocfs2_journal_access(handle
, inode
,
2143 right_path
->p_node
[i
].bh
,
2144 OCFS2_JOURNAL_ACCESS_WRITE
);
2150 ret
= ocfs2_journal_access(handle
, inode
,
2151 left_path
->p_node
[i
].bh
,
2152 OCFS2_JOURNAL_ACCESS_WRITE
);
2159 if (!right_has_empty
) {
2161 * Only do this if we're moving a real
2162 * record. Otherwise, the action is delayed until
2163 * after removal of the right path in which case we
2164 * can do a simple shift to remove the empty extent.
2166 ocfs2_rotate_leaf(left_leaf_el
, &right_leaf_el
->l_recs
[0]);
2167 memset(&right_leaf_el
->l_recs
[0], 0,
2168 sizeof(struct ocfs2_extent_rec
));
2170 if (eb
->h_next_leaf_blk
== 0ULL) {
2172 * Move recs over to get rid of empty extent, decrease
2173 * next_free. This is allowed to remove the last
2174 * extent in our leaf (setting l_next_free_rec to
2175 * zero) - the delete code below won't care.
2177 ocfs2_remove_empty_extent(right_leaf_el
);
2180 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2183 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2187 if (del_right_subtree
) {
2188 ocfs2_unlink_subtree(inode
, handle
, left_path
, right_path
,
2189 subtree_index
, dealloc
);
2190 ocfs2_update_edge_lengths(inode
, handle
, left_path
);
2192 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2193 di
->i_last_eb_blk
= eb
->h_blkno
;
2196 * Removal of the extent in the left leaf was skipped
2197 * above so we could delete the right path
2200 if (right_has_empty
)
2201 ocfs2_remove_empty_extent(left_leaf_el
);
2203 ret
= ocfs2_journal_dirty(handle
, di_bh
);
2209 ocfs2_complete_edge_insert(inode
, handle
, left_path
, right_path
,
2217 * Given a full path, determine what cpos value would return us a path
2218 * containing the leaf immediately to the right of the current one.
2220 * Will return zero if the path passed in is already the rightmost path.
2222 * This looks similar, but is subtly different to
2223 * ocfs2_find_cpos_for_left_leaf().
2225 static int ocfs2_find_cpos_for_right_leaf(struct super_block
*sb
,
2226 struct ocfs2_path
*path
, u32
*cpos
)
2230 struct ocfs2_extent_list
*el
;
2234 if (path
->p_tree_depth
== 0)
2237 blkno
= path_leaf_bh(path
)->b_blocknr
;
2239 /* Start at the tree node just above the leaf and work our way up. */
2240 i
= path
->p_tree_depth
- 1;
2244 el
= path
->p_node
[i
].el
;
2247 * Find the extent record just after the one in our
2250 next_free
= le16_to_cpu(el
->l_next_free_rec
);
2251 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2252 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2253 if (j
== (next_free
- 1)) {
2256 * We've determined that the
2257 * path specified is already
2258 * the rightmost one - return a
2264 * The rightmost record points to our
2265 * leaf - we need to travel up the
2271 *cpos
= le32_to_cpu(el
->l_recs
[j
+ 1].e_cpos
);
2277 * If we got here, we never found a valid node where
2278 * the tree indicated one should be.
2281 "Invalid extent tree at extent block %llu\n",
2282 (unsigned long long)blkno
);
2287 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2295 static int ocfs2_rotate_rightmost_leaf_left(struct inode
*inode
,
2297 struct buffer_head
*bh
,
2298 struct ocfs2_extent_list
*el
)
2302 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2305 ret
= ocfs2_journal_access(handle
, inode
, bh
,
2306 OCFS2_JOURNAL_ACCESS_WRITE
);
2312 ocfs2_remove_empty_extent(el
);
2314 ret
= ocfs2_journal_dirty(handle
, bh
);
2322 static int __ocfs2_rotate_tree_left(struct inode
*inode
,
2323 handle_t
*handle
, int orig_credits
,
2324 struct ocfs2_path
*path
,
2325 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2326 struct ocfs2_path
**empty_extent_path
)
2328 int ret
, subtree_root
, deleted
;
2330 struct ocfs2_path
*left_path
= NULL
;
2331 struct ocfs2_path
*right_path
= NULL
;
2333 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path
)->l_recs
[0])));
2335 *empty_extent_path
= NULL
;
2337 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, path
,
2344 left_path
= ocfs2_new_path(path_root_bh(path
),
2345 path_root_el(path
));
2352 ocfs2_cp_path(left_path
, path
);
2354 right_path
= ocfs2_new_path(path_root_bh(path
),
2355 path_root_el(path
));
2362 while (right_cpos
) {
2363 ret
= ocfs2_find_path(inode
, right_path
, right_cpos
);
2369 subtree_root
= ocfs2_find_subtree_root(inode
, left_path
,
2372 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2374 (unsigned long long)
2375 right_path
->p_node
[subtree_root
].bh
->b_blocknr
,
2376 right_path
->p_tree_depth
);
2378 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_root
,
2379 orig_credits
, left_path
);
2386 * Caller might still want to make changes to the
2387 * tree root, so re-add it to the journal here.
2389 ret
= ocfs2_journal_access(handle
, inode
,
2390 path_root_bh(left_path
),
2391 OCFS2_JOURNAL_ACCESS_WRITE
);
2397 ret
= ocfs2_rotate_subtree_left(inode
, handle
, left_path
,
2398 right_path
, subtree_root
,
2400 if (ret
== -EAGAIN
) {
2402 * The rotation has to temporarily stop due to
2403 * the right subtree having an empty
2404 * extent. Pass it back to the caller for a
2407 *empty_extent_path
= right_path
;
2417 * The subtree rotate might have removed records on
2418 * the rightmost edge. If so, then rotation is
2424 ocfs2_mv_path(left_path
, right_path
);
2426 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, left_path
,
2435 ocfs2_free_path(right_path
);
2436 ocfs2_free_path(left_path
);
2441 static int ocfs2_remove_rightmost_path(struct inode
*inode
, handle_t
*handle
,
2442 struct ocfs2_path
*path
,
2443 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2445 int ret
, subtree_index
;
2447 struct ocfs2_path
*left_path
= NULL
;
2448 struct ocfs2_dinode
*di
;
2449 struct ocfs2_extent_block
*eb
;
2450 struct ocfs2_extent_list
*el
;
2453 * XXX: This code assumes that the root is an inode, which is
2454 * true for now but may change as tree code gets generic.
2456 di
= (struct ocfs2_dinode
*)path_root_bh(path
)->b_data
;
2457 if (!OCFS2_IS_VALID_DINODE(di
)) {
2459 ocfs2_error(inode
->i_sb
,
2460 "Inode %llu has invalid path root",
2461 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
2466 * There's two ways we handle this depending on
2467 * whether path is the only existing one.
2469 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
2470 handle
->h_buffer_credits
,
2477 ret
= ocfs2_journal_access_path(inode
, handle
, path
);
2483 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
2491 * We have a path to the left of this one - it needs
2494 left_path
= ocfs2_new_path(path_root_bh(path
),
2495 path_root_el(path
));
2502 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
2508 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
2514 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
, path
);
2516 ocfs2_unlink_subtree(inode
, handle
, left_path
, path
,
2517 subtree_index
, dealloc
);
2518 ocfs2_update_edge_lengths(inode
, handle
, left_path
);
2520 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2521 di
->i_last_eb_blk
= eb
->h_blkno
;
2524 * 'path' is also the leftmost path which
2525 * means it must be the only one. This gets
2526 * handled differently because we want to
2527 * revert the inode back to having extents
2530 ocfs2_unlink_path(inode
, handle
, dealloc
, path
, 1);
2532 el
= &di
->id2
.i_list
;
2533 el
->l_tree_depth
= 0;
2534 el
->l_next_free_rec
= 0;
2535 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2537 di
->i_last_eb_blk
= 0;
2540 ocfs2_journal_dirty(handle
, path_root_bh(path
));
2543 ocfs2_free_path(left_path
);
2548 * Left rotation of btree records.
2550 * In many ways, this is (unsurprisingly) the opposite of right
2551 * rotation. We start at some non-rightmost path containing an empty
2552 * extent in the leaf block. The code works its way to the rightmost
2553 * path by rotating records to the left in every subtree.
2555 * This is used by any code which reduces the number of extent records
2556 * in a leaf. After removal, an empty record should be placed in the
2557 * leftmost list position.
2559 * This won't handle a length update of the rightmost path records if
2560 * the rightmost tree leaf record is removed so the caller is
2561 * responsible for detecting and correcting that.
2563 static int ocfs2_rotate_tree_left(struct inode
*inode
, handle_t
*handle
,
2564 struct ocfs2_path
*path
,
2565 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2567 int ret
, orig_credits
= handle
->h_buffer_credits
;
2568 struct ocfs2_path
*tmp_path
= NULL
, *restart_path
= NULL
;
2569 struct ocfs2_extent_block
*eb
;
2570 struct ocfs2_extent_list
*el
;
2572 el
= path_leaf_el(path
);
2573 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2576 if (path
->p_tree_depth
== 0) {
2577 rightmost_no_delete
:
2579 * In-inode extents. This is trivially handled, so do
2582 ret
= ocfs2_rotate_rightmost_leaf_left(inode
, handle
,
2584 path_leaf_el(path
));
2591 * Handle rightmost branch now. There's several cases:
2592 * 1) simple rotation leaving records in there. That's trivial.
2593 * 2) rotation requiring a branch delete - there's no more
2594 * records left. Two cases of this:
2595 * a) There are branches to the left.
2596 * b) This is also the leftmost (the only) branch.
2598 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
2599 * 2a) we need the left branch so that we can update it with the unlink
2600 * 2b) we need to bring the inode back to inline extents.
2603 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2605 if (eb
->h_next_leaf_blk
== 0) {
2607 * This gets a bit tricky if we're going to delete the
2608 * rightmost path. Get the other cases out of the way
2611 if (le16_to_cpu(el
->l_next_free_rec
) > 1)
2612 goto rightmost_no_delete
;
2614 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
2616 ocfs2_error(inode
->i_sb
,
2617 "Inode %llu has empty extent block at %llu",
2618 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2619 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
2624 * XXX: The caller can not trust "path" any more after
2625 * this as it will have been deleted. What do we do?
2627 * In theory the rotate-for-merge code will never get
2628 * here because it'll always ask for a rotate in a
2632 ret
= ocfs2_remove_rightmost_path(inode
, handle
, path
,
2640 * Now we can loop, remembering the path we get from -EAGAIN
2641 * and restarting from there.
2644 ret
= __ocfs2_rotate_tree_left(inode
, handle
, orig_credits
, path
,
2645 dealloc
, &restart_path
);
2646 if (ret
&& ret
!= -EAGAIN
) {
2651 while (ret
== -EAGAIN
) {
2652 tmp_path
= restart_path
;
2653 restart_path
= NULL
;
2655 ret
= __ocfs2_rotate_tree_left(inode
, handle
, orig_credits
,
2658 if (ret
&& ret
!= -EAGAIN
) {
2663 ocfs2_free_path(tmp_path
);
2671 ocfs2_free_path(tmp_path
);
2672 ocfs2_free_path(restart_path
);
2676 static void ocfs2_cleanup_merge(struct ocfs2_extent_list
*el
,
2679 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[index
];
2682 if (rec
->e_leaf_clusters
== 0) {
2684 * We consumed all of the merged-from record. An empty
2685 * extent cannot exist anywhere but the 1st array
2686 * position, so move things over if the merged-from
2687 * record doesn't occupy that position.
2689 * This creates a new empty extent so the caller
2690 * should be smart enough to have removed any existing
2694 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
2695 size
= index
* sizeof(struct ocfs2_extent_rec
);
2696 memmove(&el
->l_recs
[1], &el
->l_recs
[0], size
);
2700 * Always memset - the caller doesn't check whether it
2701 * created an empty extent, so there could be junk in
2704 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2708 static int ocfs2_get_right_path(struct inode
*inode
,
2709 struct ocfs2_path
*left_path
,
2710 struct ocfs2_path
**ret_right_path
)
2714 struct ocfs2_path
*right_path
= NULL
;
2715 struct ocfs2_extent_list
*left_el
;
2717 *ret_right_path
= NULL
;
2719 /* This function shouldn't be called for non-trees. */
2720 BUG_ON(left_path
->p_tree_depth
== 0);
2722 left_el
= path_leaf_el(left_path
);
2723 BUG_ON(left_el
->l_next_free_rec
!= left_el
->l_count
);
2725 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, left_path
,
2732 /* This function shouldn't be called for the rightmost leaf. */
2733 BUG_ON(right_cpos
== 0);
2735 right_path
= ocfs2_new_path(path_root_bh(left_path
),
2736 path_root_el(left_path
));
2743 ret
= ocfs2_find_path(inode
, right_path
, right_cpos
);
2749 *ret_right_path
= right_path
;
2752 ocfs2_free_path(right_path
);
2757 * Remove split_rec clusters from the record at index and merge them
2758 * onto the beginning of the record "next" to it.
2759 * For index < l_count - 1, the next means the extent rec at index + 1.
2760 * For index == l_count - 1, the "next" means the 1st extent rec of the
2761 * next extent block.
2763 static int ocfs2_merge_rec_right(struct inode
*inode
,
2764 struct ocfs2_path
*left_path
,
2766 struct ocfs2_extent_rec
*split_rec
,
2769 int ret
, next_free
, i
;
2770 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
2771 struct ocfs2_extent_rec
*left_rec
;
2772 struct ocfs2_extent_rec
*right_rec
;
2773 struct ocfs2_extent_list
*right_el
;
2774 struct ocfs2_path
*right_path
= NULL
;
2775 int subtree_index
= 0;
2776 struct ocfs2_extent_list
*el
= path_leaf_el(left_path
);
2777 struct buffer_head
*bh
= path_leaf_bh(left_path
);
2778 struct buffer_head
*root_bh
= NULL
;
2780 BUG_ON(index
>= le16_to_cpu(el
->l_next_free_rec
));
2781 left_rec
= &el
->l_recs
[index
];
2783 if (index
== le16_to_cpu(el
->l_next_free_rec
) - 1 &&
2784 le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
)) {
2785 /* we meet with a cross extent block merge. */
2786 ret
= ocfs2_get_right_path(inode
, left_path
, &right_path
);
2792 right_el
= path_leaf_el(right_path
);
2793 next_free
= le16_to_cpu(right_el
->l_next_free_rec
);
2794 BUG_ON(next_free
<= 0);
2795 right_rec
= &right_el
->l_recs
[0];
2796 if (ocfs2_is_empty_extent(right_rec
)) {
2797 BUG_ON(next_free
<= 1);
2798 right_rec
= &right_el
->l_recs
[1];
2801 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
2802 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
2803 le32_to_cpu(right_rec
->e_cpos
));
2805 subtree_index
= ocfs2_find_subtree_root(inode
,
2806 left_path
, right_path
);
2808 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
2809 handle
->h_buffer_credits
,
2816 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2817 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2819 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
2820 OCFS2_JOURNAL_ACCESS_WRITE
);
2826 for (i
= subtree_index
+ 1;
2827 i
< path_num_items(right_path
); i
++) {
2828 ret
= ocfs2_journal_access(handle
, inode
,
2829 right_path
->p_node
[i
].bh
,
2830 OCFS2_JOURNAL_ACCESS_WRITE
);
2836 ret
= ocfs2_journal_access(handle
, inode
,
2837 left_path
->p_node
[i
].bh
,
2838 OCFS2_JOURNAL_ACCESS_WRITE
);
2846 BUG_ON(index
== le16_to_cpu(el
->l_next_free_rec
) - 1);
2847 right_rec
= &el
->l_recs
[index
+ 1];
2850 ret
= ocfs2_journal_access(handle
, inode
, bh
,
2851 OCFS2_JOURNAL_ACCESS_WRITE
);
2857 le16_add_cpu(&left_rec
->e_leaf_clusters
, -split_clusters
);
2859 le32_add_cpu(&right_rec
->e_cpos
, -split_clusters
);
2860 le64_add_cpu(&right_rec
->e_blkno
,
2861 -ocfs2_clusters_to_blocks(inode
->i_sb
, split_clusters
));
2862 le16_add_cpu(&right_rec
->e_leaf_clusters
, split_clusters
);
2864 ocfs2_cleanup_merge(el
, index
);
2866 ret
= ocfs2_journal_dirty(handle
, bh
);
2871 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2875 ocfs2_complete_edge_insert(inode
, handle
, left_path
,
2876 right_path
, subtree_index
);
2880 ocfs2_free_path(right_path
);
2884 static int ocfs2_get_left_path(struct inode
*inode
,
2885 struct ocfs2_path
*right_path
,
2886 struct ocfs2_path
**ret_left_path
)
2890 struct ocfs2_path
*left_path
= NULL
;
2892 *ret_left_path
= NULL
;
2894 /* This function shouldn't be called for non-trees. */
2895 BUG_ON(right_path
->p_tree_depth
== 0);
2897 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
,
2898 right_path
, &left_cpos
);
2904 /* This function shouldn't be called for the leftmost leaf. */
2905 BUG_ON(left_cpos
== 0);
2907 left_path
= ocfs2_new_path(path_root_bh(right_path
),
2908 path_root_el(right_path
));
2915 ret
= ocfs2_find_path(inode
, left_path
, left_cpos
);
2921 *ret_left_path
= left_path
;
2924 ocfs2_free_path(left_path
);
2929 * Remove split_rec clusters from the record at index and merge them
2930 * onto the tail of the record "before" it.
2931 * For index > 0, the "before" means the extent rec at index - 1.
2933 * For index == 0, the "before" means the last record of the previous
2934 * extent block. And there is also a situation that we may need to
2935 * remove the rightmost leaf extent block in the right_path and change
2936 * the right path to indicate the new rightmost path.
2938 static int ocfs2_merge_rec_left(struct inode
*inode
,
2939 struct ocfs2_path
*right_path
,
2941 struct ocfs2_extent_rec
*split_rec
,
2942 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2945 int ret
, i
, subtree_index
= 0, has_empty_extent
= 0;
2946 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
2947 struct ocfs2_extent_rec
*left_rec
;
2948 struct ocfs2_extent_rec
*right_rec
;
2949 struct ocfs2_extent_list
*el
= path_leaf_el(right_path
);
2950 struct buffer_head
*bh
= path_leaf_bh(right_path
);
2951 struct buffer_head
*root_bh
= NULL
;
2952 struct ocfs2_path
*left_path
= NULL
;
2953 struct ocfs2_extent_list
*left_el
;
2957 right_rec
= &el
->l_recs
[index
];
2959 /* we meet with a cross extent block merge. */
2960 ret
= ocfs2_get_left_path(inode
, right_path
, &left_path
);
2966 left_el
= path_leaf_el(left_path
);
2967 BUG_ON(le16_to_cpu(left_el
->l_next_free_rec
) !=
2968 le16_to_cpu(left_el
->l_count
));
2970 left_rec
= &left_el
->l_recs
[
2971 le16_to_cpu(left_el
->l_next_free_rec
) - 1];
2972 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
2973 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
2974 le32_to_cpu(split_rec
->e_cpos
));
2976 subtree_index
= ocfs2_find_subtree_root(inode
,
2977 left_path
, right_path
);
2979 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
2980 handle
->h_buffer_credits
,
2987 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2988 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2990 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
2991 OCFS2_JOURNAL_ACCESS_WRITE
);
2997 for (i
= subtree_index
+ 1;
2998 i
< path_num_items(right_path
); i
++) {
2999 ret
= ocfs2_journal_access(handle
, inode
,
3000 right_path
->p_node
[i
].bh
,
3001 OCFS2_JOURNAL_ACCESS_WRITE
);
3007 ret
= ocfs2_journal_access(handle
, inode
,
3008 left_path
->p_node
[i
].bh
,
3009 OCFS2_JOURNAL_ACCESS_WRITE
);
3016 left_rec
= &el
->l_recs
[index
- 1];
3017 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
3018 has_empty_extent
= 1;
3021 ret
= ocfs2_journal_access(handle
, inode
, bh
,
3022 OCFS2_JOURNAL_ACCESS_WRITE
);
3028 if (has_empty_extent
&& index
== 1) {
3030 * The easy case - we can just plop the record right in.
3032 *left_rec
= *split_rec
;
3034 has_empty_extent
= 0;
3036 le16_add_cpu(&left_rec
->e_leaf_clusters
, split_clusters
);
3038 le32_add_cpu(&right_rec
->e_cpos
, split_clusters
);
3039 le64_add_cpu(&right_rec
->e_blkno
,
3040 ocfs2_clusters_to_blocks(inode
->i_sb
, split_clusters
));
3041 le16_add_cpu(&right_rec
->e_leaf_clusters
, -split_clusters
);
3043 ocfs2_cleanup_merge(el
, index
);
3045 ret
= ocfs2_journal_dirty(handle
, bh
);
3050 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
3055 * In the situation that the right_rec is empty and the extent
3056 * block is empty also, ocfs2_complete_edge_insert can't handle
3057 * it and we need to delete the right extent block.
3059 if (le16_to_cpu(right_rec
->e_leaf_clusters
) == 0 &&
3060 le16_to_cpu(el
->l_next_free_rec
) == 1) {
3062 ret
= ocfs2_remove_rightmost_path(inode
, handle
,
3063 right_path
, dealloc
);
3069 /* Now the rightmost extent block has been deleted.
3070 * So we use the new rightmost path.
3072 ocfs2_mv_path(right_path
, left_path
);
3075 ocfs2_complete_edge_insert(inode
, handle
, left_path
,
3076 right_path
, subtree_index
);
3080 ocfs2_free_path(left_path
);
3084 static int ocfs2_try_to_merge_extent(struct inode
*inode
,
3086 struct ocfs2_path
*path
,
3088 struct ocfs2_extent_rec
*split_rec
,
3089 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3090 struct ocfs2_merge_ctxt
*ctxt
)
3094 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
3095 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
3097 BUG_ON(ctxt
->c_contig_type
== CONTIG_NONE
);
3099 if (ctxt
->c_split_covers_rec
&& ctxt
->c_has_empty_extent
) {
3101 * The merge code will need to create an empty
3102 * extent to take the place of the newly
3103 * emptied slot. Remove any pre-existing empty
3104 * extents - having more than one in a leaf is
3107 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
,
3114 rec
= &el
->l_recs
[split_index
];
3117 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
) {
3119 * Left-right contig implies this.
3121 BUG_ON(!ctxt
->c_split_covers_rec
);
3124 * Since the leftright insert always covers the entire
3125 * extent, this call will delete the insert record
3126 * entirely, resulting in an empty extent record added to
3129 * Since the adding of an empty extent shifts
3130 * everything back to the right, there's no need to
3131 * update split_index here.
3133 * When the split_index is zero, we need to merge it to the
3134 * prevoius extent block. It is more efficient and easier
3135 * if we do merge_right first and merge_left later.
3137 ret
= ocfs2_merge_rec_right(inode
, path
,
3146 * We can only get this from logic error above.
3148 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
3150 /* The merge left us with an empty extent, remove it. */
3151 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
, dealloc
);
3157 rec
= &el
->l_recs
[split_index
];
3160 * Note that we don't pass split_rec here on purpose -
3161 * we've merged it into the rec already.
3163 ret
= ocfs2_merge_rec_left(inode
, path
,
3173 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
,
3176 * Error from this last rotate is not critical, so
3177 * print but don't bubble it up.
3184 * Merge a record to the left or right.
3186 * 'contig_type' is relative to the existing record,
3187 * so for example, if we're "right contig", it's to
3188 * the record on the left (hence the left merge).
3190 if (ctxt
->c_contig_type
== CONTIG_RIGHT
) {
3191 ret
= ocfs2_merge_rec_left(inode
,
3201 ret
= ocfs2_merge_rec_right(inode
,
3211 if (ctxt
->c_split_covers_rec
) {
3213 * The merge may have left an empty extent in
3214 * our leaf. Try to rotate it away.
3216 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
,
3228 static void ocfs2_subtract_from_rec(struct super_block
*sb
,
3229 enum ocfs2_split_type split
,
3230 struct ocfs2_extent_rec
*rec
,
3231 struct ocfs2_extent_rec
*split_rec
)
3235 len_blocks
= ocfs2_clusters_to_blocks(sb
,
3236 le16_to_cpu(split_rec
->e_leaf_clusters
));
3238 if (split
== SPLIT_LEFT
) {
3240 * Region is on the left edge of the existing
3243 le32_add_cpu(&rec
->e_cpos
,
3244 le16_to_cpu(split_rec
->e_leaf_clusters
));
3245 le64_add_cpu(&rec
->e_blkno
, len_blocks
);
3246 le16_add_cpu(&rec
->e_leaf_clusters
,
3247 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3250 * Region is on the right edge of the existing
3253 le16_add_cpu(&rec
->e_leaf_clusters
,
3254 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3259 * Do the final bits of extent record insertion at the target leaf
3260 * list. If this leaf is part of an allocation tree, it is assumed
3261 * that the tree above has been prepared.
3263 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec
*insert_rec
,
3264 struct ocfs2_extent_list
*el
,
3265 struct ocfs2_insert_type
*insert
,
3266 struct inode
*inode
)
3268 int i
= insert
->ins_contig_index
;
3270 struct ocfs2_extent_rec
*rec
;
3272 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3274 if (insert
->ins_split
!= SPLIT_NONE
) {
3275 i
= ocfs2_search_extent_list(el
, le32_to_cpu(insert_rec
->e_cpos
));
3277 rec
= &el
->l_recs
[i
];
3278 ocfs2_subtract_from_rec(inode
->i_sb
, insert
->ins_split
, rec
,
3284 * Contiguous insert - either left or right.
3286 if (insert
->ins_contig
!= CONTIG_NONE
) {
3287 rec
= &el
->l_recs
[i
];
3288 if (insert
->ins_contig
== CONTIG_LEFT
) {
3289 rec
->e_blkno
= insert_rec
->e_blkno
;
3290 rec
->e_cpos
= insert_rec
->e_cpos
;
3292 le16_add_cpu(&rec
->e_leaf_clusters
,
3293 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3298 * Handle insert into an empty leaf.
3300 if (le16_to_cpu(el
->l_next_free_rec
) == 0 ||
3301 ((le16_to_cpu(el
->l_next_free_rec
) == 1) &&
3302 ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3303 el
->l_recs
[0] = *insert_rec
;
3304 el
->l_next_free_rec
= cpu_to_le16(1);
3311 if (insert
->ins_appending
== APPEND_TAIL
) {
3312 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3313 rec
= &el
->l_recs
[i
];
3314 range
= le32_to_cpu(rec
->e_cpos
)
3315 + le16_to_cpu(rec
->e_leaf_clusters
);
3316 BUG_ON(le32_to_cpu(insert_rec
->e_cpos
) < range
);
3318 mlog_bug_on_msg(le16_to_cpu(el
->l_next_free_rec
) >=
3319 le16_to_cpu(el
->l_count
),
3320 "inode %lu, depth %u, count %u, next free %u, "
3321 "rec.cpos %u, rec.clusters %u, "
3322 "insert.cpos %u, insert.clusters %u\n",
3324 le16_to_cpu(el
->l_tree_depth
),
3325 le16_to_cpu(el
->l_count
),
3326 le16_to_cpu(el
->l_next_free_rec
),
3327 le32_to_cpu(el
->l_recs
[i
].e_cpos
),
3328 le16_to_cpu(el
->l_recs
[i
].e_leaf_clusters
),
3329 le32_to_cpu(insert_rec
->e_cpos
),
3330 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3332 el
->l_recs
[i
] = *insert_rec
;
3333 le16_add_cpu(&el
->l_next_free_rec
, 1);
3339 * Ok, we have to rotate.
3341 * At this point, it is safe to assume that inserting into an
3342 * empty leaf and appending to a leaf have both been handled
3345 * This leaf needs to have space, either by the empty 1st
3346 * extent record, or by virtue of an l_next_rec < l_count.
3348 ocfs2_rotate_leaf(el
, insert_rec
);
3351 static inline void ocfs2_update_dinode_clusters(struct inode
*inode
,
3352 struct ocfs2_dinode
*di
,
3355 le32_add_cpu(&di
->i_clusters
, clusters
);
3356 spin_lock(&OCFS2_I(inode
)->ip_lock
);
3357 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(di
->i_clusters
);
3358 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
3361 static void ocfs2_adjust_rightmost_records(struct inode
*inode
,
3363 struct ocfs2_path
*path
,
3364 struct ocfs2_extent_rec
*insert_rec
)
3366 int ret
, i
, next_free
;
3367 struct buffer_head
*bh
;
3368 struct ocfs2_extent_list
*el
;
3369 struct ocfs2_extent_rec
*rec
;
3372 * Update everything except the leaf block.
3374 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
3375 bh
= path
->p_node
[i
].bh
;
3376 el
= path
->p_node
[i
].el
;
3378 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3379 if (next_free
== 0) {
3380 ocfs2_error(inode
->i_sb
,
3381 "Dinode %llu has a bad extent list",
3382 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
3387 rec
= &el
->l_recs
[next_free
- 1];
3389 rec
->e_int_clusters
= insert_rec
->e_cpos
;
3390 le32_add_cpu(&rec
->e_int_clusters
,
3391 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3392 le32_add_cpu(&rec
->e_int_clusters
,
3393 -le32_to_cpu(rec
->e_cpos
));
3395 ret
= ocfs2_journal_dirty(handle
, bh
);
3402 static int ocfs2_append_rec_to_path(struct inode
*inode
, handle_t
*handle
,
3403 struct ocfs2_extent_rec
*insert_rec
,
3404 struct ocfs2_path
*right_path
,
3405 struct ocfs2_path
**ret_left_path
)
3408 struct ocfs2_extent_list
*el
;
3409 struct ocfs2_path
*left_path
= NULL
;
3411 *ret_left_path
= NULL
;
3414 * This shouldn't happen for non-trees. The extent rec cluster
3415 * count manipulation below only works for interior nodes.
3417 BUG_ON(right_path
->p_tree_depth
== 0);
3420 * If our appending insert is at the leftmost edge of a leaf,
3421 * then we might need to update the rightmost records of the
3424 el
= path_leaf_el(right_path
);
3425 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3426 if (next_free
== 0 ||
3427 (next_free
== 1 && ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3430 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
,
3437 mlog(0, "Append may need a left path update. cpos: %u, "
3438 "left_cpos: %u\n", le32_to_cpu(insert_rec
->e_cpos
),
3442 * No need to worry if the append is already in the
3446 left_path
= ocfs2_new_path(path_root_bh(right_path
),
3447 path_root_el(right_path
));
3454 ret
= ocfs2_find_path(inode
, left_path
, left_cpos
);
3461 * ocfs2_insert_path() will pass the left_path to the
3467 ret
= ocfs2_journal_access_path(inode
, handle
, right_path
);
3473 ocfs2_adjust_rightmost_records(inode
, handle
, right_path
, insert_rec
);
3475 *ret_left_path
= left_path
;
3479 ocfs2_free_path(left_path
);
3484 static void ocfs2_split_record(struct inode
*inode
,
3485 struct ocfs2_path
*left_path
,
3486 struct ocfs2_path
*right_path
,
3487 struct ocfs2_extent_rec
*split_rec
,
3488 enum ocfs2_split_type split
)
3491 u32 cpos
= le32_to_cpu(split_rec
->e_cpos
);
3492 struct ocfs2_extent_list
*left_el
= NULL
, *right_el
, *insert_el
, *el
;
3493 struct ocfs2_extent_rec
*rec
, *tmprec
;
3495 right_el
= path_leaf_el(right_path
);;
3497 left_el
= path_leaf_el(left_path
);
3500 insert_el
= right_el
;
3501 index
= ocfs2_search_extent_list(el
, cpos
);
3503 if (index
== 0 && left_path
) {
3504 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3507 * This typically means that the record
3508 * started in the left path but moved to the
3509 * right as a result of rotation. We either
3510 * move the existing record to the left, or we
3511 * do the later insert there.
3513 * In this case, the left path should always
3514 * exist as the rotate code will have passed
3515 * it back for a post-insert update.
3518 if (split
== SPLIT_LEFT
) {
3520 * It's a left split. Since we know
3521 * that the rotate code gave us an
3522 * empty extent in the left path, we
3523 * can just do the insert there.
3525 insert_el
= left_el
;
3528 * Right split - we have to move the
3529 * existing record over to the left
3530 * leaf. The insert will be into the
3531 * newly created empty extent in the
3534 tmprec
= &right_el
->l_recs
[index
];
3535 ocfs2_rotate_leaf(left_el
, tmprec
);
3538 memset(tmprec
, 0, sizeof(*tmprec
));
3539 index
= ocfs2_search_extent_list(left_el
, cpos
);
3540 BUG_ON(index
== -1);
3545 BUG_ON(!ocfs2_is_empty_extent(&left_el
->l_recs
[0]));
3547 * Left path is easy - we can just allow the insert to
3551 insert_el
= left_el
;
3552 index
= ocfs2_search_extent_list(el
, cpos
);
3553 BUG_ON(index
== -1);
3556 rec
= &el
->l_recs
[index
];
3557 ocfs2_subtract_from_rec(inode
->i_sb
, split
, rec
, split_rec
);
3558 ocfs2_rotate_leaf(insert_el
, split_rec
);
3562 * This function only does inserts on an allocation b-tree. For dinode
3563 * lists, ocfs2_insert_at_leaf() is called directly.
3565 * right_path is the path we want to do the actual insert
3566 * in. left_path should only be passed in if we need to update that
3567 * portion of the tree after an edge insert.
3569 static int ocfs2_insert_path(struct inode
*inode
,
3571 struct ocfs2_path
*left_path
,
3572 struct ocfs2_path
*right_path
,
3573 struct ocfs2_extent_rec
*insert_rec
,
3574 struct ocfs2_insert_type
*insert
)
3576 int ret
, subtree_index
;
3577 struct buffer_head
*leaf_bh
= path_leaf_bh(right_path
);
3580 int credits
= handle
->h_buffer_credits
;
3583 * There's a chance that left_path got passed back to
3584 * us without being accounted for in the
3585 * journal. Extend our transaction here to be sure we
3586 * can change those blocks.
3588 credits
+= left_path
->p_tree_depth
;
3590 ret
= ocfs2_extend_trans(handle
, credits
);
3596 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
3604 * Pass both paths to the journal. The majority of inserts
3605 * will be touching all components anyway.
3607 ret
= ocfs2_journal_access_path(inode
, handle
, right_path
);
3613 if (insert
->ins_split
!= SPLIT_NONE
) {
3615 * We could call ocfs2_insert_at_leaf() for some types
3616 * of splits, but it's easier to just let one separate
3617 * function sort it all out.
3619 ocfs2_split_record(inode
, left_path
, right_path
,
3620 insert_rec
, insert
->ins_split
);
3623 * Split might have modified either leaf and we don't
3624 * have a guarantee that the later edge insert will
3625 * dirty this for us.
3628 ret
= ocfs2_journal_dirty(handle
,
3629 path_leaf_bh(left_path
));
3633 ocfs2_insert_at_leaf(insert_rec
, path_leaf_el(right_path
),
3636 ret
= ocfs2_journal_dirty(handle
, leaf_bh
);
3642 * The rotate code has indicated that we need to fix
3643 * up portions of the tree after the insert.
3645 * XXX: Should we extend the transaction here?
3647 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
,
3649 ocfs2_complete_edge_insert(inode
, handle
, left_path
,
3650 right_path
, subtree_index
);
3658 static int ocfs2_do_insert_extent(struct inode
*inode
,
3660 struct buffer_head
*di_bh
,
3661 struct ocfs2_extent_rec
*insert_rec
,
3662 struct ocfs2_insert_type
*type
)
3664 int ret
, rotate
= 0;
3666 struct ocfs2_path
*right_path
= NULL
;
3667 struct ocfs2_path
*left_path
= NULL
;
3668 struct ocfs2_dinode
*di
;
3669 struct ocfs2_extent_list
*el
;
3671 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
3672 el
= &di
->id2
.i_list
;
3674 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
3675 OCFS2_JOURNAL_ACCESS_WRITE
);
3681 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
3682 ocfs2_insert_at_leaf(insert_rec
, el
, type
, inode
);
3683 goto out_update_clusters
;
3686 right_path
= ocfs2_new_inode_path(di_bh
);
3694 * Determine the path to start with. Rotations need the
3695 * rightmost path, everything else can go directly to the
3698 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
3699 if (type
->ins_appending
== APPEND_NONE
&&
3700 type
->ins_contig
== CONTIG_NONE
) {
3705 ret
= ocfs2_find_path(inode
, right_path
, cpos
);
3712 * Rotations and appends need special treatment - they modify
3713 * parts of the tree's above them.
3715 * Both might pass back a path immediate to the left of the
3716 * one being inserted to. This will be cause
3717 * ocfs2_insert_path() to modify the rightmost records of
3718 * left_path to account for an edge insert.
3720 * XXX: When modifying this code, keep in mind that an insert
3721 * can wind up skipping both of these two special cases...
3724 ret
= ocfs2_rotate_tree_right(inode
, handle
, type
->ins_split
,
3725 le32_to_cpu(insert_rec
->e_cpos
),
3726 right_path
, &left_path
);
3733 * ocfs2_rotate_tree_right() might have extended the
3734 * transaction without re-journaling our tree root.
3736 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
3737 OCFS2_JOURNAL_ACCESS_WRITE
);
3742 } else if (type
->ins_appending
== APPEND_TAIL
3743 && type
->ins_contig
!= CONTIG_LEFT
) {
3744 ret
= ocfs2_append_rec_to_path(inode
, handle
, insert_rec
,
3745 right_path
, &left_path
);
3752 ret
= ocfs2_insert_path(inode
, handle
, left_path
, right_path
,
3759 out_update_clusters
:
3760 if (type
->ins_split
== SPLIT_NONE
)
3761 ocfs2_update_dinode_clusters(inode
, di
,
3762 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3764 ret
= ocfs2_journal_dirty(handle
, di_bh
);
3769 ocfs2_free_path(left_path
);
3770 ocfs2_free_path(right_path
);
3775 static enum ocfs2_contig_type
3776 ocfs2_figure_merge_contig_type(struct inode
*inode
, struct ocfs2_path
*path
,
3777 struct ocfs2_extent_list
*el
, int index
,
3778 struct ocfs2_extent_rec
*split_rec
)
3781 enum ocfs2_contig_type ret
= CONTIG_NONE
;
3782 u32 left_cpos
, right_cpos
;
3783 struct ocfs2_extent_rec
*rec
= NULL
;
3784 struct ocfs2_extent_list
*new_el
;
3785 struct ocfs2_path
*left_path
= NULL
, *right_path
= NULL
;
3786 struct buffer_head
*bh
;
3787 struct ocfs2_extent_block
*eb
;
3790 rec
= &el
->l_recs
[index
- 1];
3791 } else if (path
->p_tree_depth
> 0) {
3792 status
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
,
3797 if (left_cpos
!= 0) {
3798 left_path
= ocfs2_new_path(path_root_bh(path
),
3799 path_root_el(path
));
3803 status
= ocfs2_find_path(inode
, left_path
, left_cpos
);
3807 new_el
= path_leaf_el(left_path
);
3809 if (le16_to_cpu(new_el
->l_next_free_rec
) !=
3810 le16_to_cpu(new_el
->l_count
)) {
3811 bh
= path_leaf_bh(left_path
);
3812 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
3813 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
,
3817 rec
= &new_el
->l_recs
[
3818 le16_to_cpu(new_el
->l_next_free_rec
) - 1];
3823 * We're careful to check for an empty extent record here -
3824 * the merge code will know what to do if it sees one.
3827 if (index
== 1 && ocfs2_is_empty_extent(rec
)) {
3828 if (split_rec
->e_cpos
== el
->l_recs
[index
].e_cpos
)
3831 ret
= ocfs2_extent_contig(inode
, rec
, split_rec
);
3836 if (index
< (le16_to_cpu(el
->l_next_free_rec
) - 1))
3837 rec
= &el
->l_recs
[index
+ 1];
3838 else if (le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
) &&
3839 path
->p_tree_depth
> 0) {
3840 status
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
,
3845 if (right_cpos
== 0)
3848 right_path
= ocfs2_new_path(path_root_bh(path
),
3849 path_root_el(path
));
3853 status
= ocfs2_find_path(inode
, right_path
, right_cpos
);
3857 new_el
= path_leaf_el(right_path
);
3858 rec
= &new_el
->l_recs
[0];
3859 if (ocfs2_is_empty_extent(rec
)) {
3860 if (le16_to_cpu(new_el
->l_next_free_rec
) <= 1) {
3861 bh
= path_leaf_bh(right_path
);
3862 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
3863 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
,
3867 rec
= &new_el
->l_recs
[1];
3872 enum ocfs2_contig_type contig_type
;
3874 contig_type
= ocfs2_extent_contig(inode
, rec
, split_rec
);
3876 if (contig_type
== CONTIG_LEFT
&& ret
== CONTIG_RIGHT
)
3877 ret
= CONTIG_LEFTRIGHT
;
3878 else if (ret
== CONTIG_NONE
)
3884 ocfs2_free_path(left_path
);
3886 ocfs2_free_path(right_path
);
3891 static void ocfs2_figure_contig_type(struct inode
*inode
,
3892 struct ocfs2_insert_type
*insert
,
3893 struct ocfs2_extent_list
*el
,
3894 struct ocfs2_extent_rec
*insert_rec
)
3897 enum ocfs2_contig_type contig_type
= CONTIG_NONE
;
3899 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3901 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
3902 contig_type
= ocfs2_extent_contig(inode
, &el
->l_recs
[i
],
3904 if (contig_type
!= CONTIG_NONE
) {
3905 insert
->ins_contig_index
= i
;
3909 insert
->ins_contig
= contig_type
;
3913 * This should only be called against the righmost leaf extent list.
3915 * ocfs2_figure_appending_type() will figure out whether we'll have to
3916 * insert at the tail of the rightmost leaf.
3918 * This should also work against the dinode list for tree's with 0
3919 * depth. If we consider the dinode list to be the rightmost leaf node
3920 * then the logic here makes sense.
3922 static void ocfs2_figure_appending_type(struct ocfs2_insert_type
*insert
,
3923 struct ocfs2_extent_list
*el
,
3924 struct ocfs2_extent_rec
*insert_rec
)
3927 u32 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
3928 struct ocfs2_extent_rec
*rec
;
3930 insert
->ins_appending
= APPEND_NONE
;
3932 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3934 if (!el
->l_next_free_rec
)
3935 goto set_tail_append
;
3937 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
3938 /* Were all records empty? */
3939 if (le16_to_cpu(el
->l_next_free_rec
) == 1)
3940 goto set_tail_append
;
3943 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3944 rec
= &el
->l_recs
[i
];
3947 (le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)))
3948 goto set_tail_append
;
3953 insert
->ins_appending
= APPEND_TAIL
;
3957 * Helper function called at the begining of an insert.
3959 * This computes a few things that are commonly used in the process of
3960 * inserting into the btree:
3961 * - Whether the new extent is contiguous with an existing one.
3962 * - The current tree depth.
3963 * - Whether the insert is an appending one.
3964 * - The total # of free records in the tree.
3966 * All of the information is stored on the ocfs2_insert_type
3969 static int ocfs2_figure_insert_type(struct inode
*inode
,
3970 struct buffer_head
*di_bh
,
3971 struct buffer_head
**last_eb_bh
,
3972 struct ocfs2_extent_rec
*insert_rec
,
3974 struct ocfs2_insert_type
*insert
)
3977 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
3978 struct ocfs2_extent_block
*eb
;
3979 struct ocfs2_extent_list
*el
;
3980 struct ocfs2_path
*path
= NULL
;
3981 struct buffer_head
*bh
= NULL
;
3983 insert
->ins_split
= SPLIT_NONE
;
3985 el
= &di
->id2
.i_list
;
3986 insert
->ins_tree_depth
= le16_to_cpu(el
->l_tree_depth
);
3988 if (el
->l_tree_depth
) {
3990 * If we have tree depth, we read in the
3991 * rightmost extent block ahead of time as
3992 * ocfs2_figure_insert_type() and ocfs2_add_branch()
3993 * may want it later.
3995 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
3996 le64_to_cpu(di
->i_last_eb_blk
), &bh
,
3997 OCFS2_BH_CACHED
, inode
);
4002 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
4007 * Unless we have a contiguous insert, we'll need to know if
4008 * there is room left in our allocation tree for another
4011 * XXX: This test is simplistic, we can search for empty
4012 * extent records too.
4014 *free_records
= le16_to_cpu(el
->l_count
) -
4015 le16_to_cpu(el
->l_next_free_rec
);
4017 if (!insert
->ins_tree_depth
) {
4018 ocfs2_figure_contig_type(inode
, insert
, el
, insert_rec
);
4019 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4023 path
= ocfs2_new_inode_path(di_bh
);
4031 * In the case that we're inserting past what the tree
4032 * currently accounts for, ocfs2_find_path() will return for
4033 * us the rightmost tree path. This is accounted for below in
4034 * the appending code.
4036 ret
= ocfs2_find_path(inode
, path
, le32_to_cpu(insert_rec
->e_cpos
));
4042 el
= path_leaf_el(path
);
4045 * Now that we have the path, there's two things we want to determine:
4046 * 1) Contiguousness (also set contig_index if this is so)
4048 * 2) Are we doing an append? We can trivially break this up
4049 * into two types of appends: simple record append, or a
4050 * rotate inside the tail leaf.
4052 ocfs2_figure_contig_type(inode
, insert
, el
, insert_rec
);
4055 * The insert code isn't quite ready to deal with all cases of
4056 * left contiguousness. Specifically, if it's an insert into
4057 * the 1st record in a leaf, it will require the adjustment of
4058 * cluster count on the last record of the path directly to it's
4059 * left. For now, just catch that case and fool the layers
4060 * above us. This works just fine for tree_depth == 0, which
4061 * is why we allow that above.
4063 if (insert
->ins_contig
== CONTIG_LEFT
&&
4064 insert
->ins_contig_index
== 0)
4065 insert
->ins_contig
= CONTIG_NONE
;
4068 * Ok, so we can simply compare against last_eb to figure out
4069 * whether the path doesn't exist. This will only happen in
4070 * the case that we're doing a tail append, so maybe we can
4071 * take advantage of that information somehow.
4073 if (le64_to_cpu(di
->i_last_eb_blk
) == path_leaf_bh(path
)->b_blocknr
) {
4075 * Ok, ocfs2_find_path() returned us the rightmost
4076 * tree path. This might be an appending insert. There are
4078 * 1) We're doing a true append at the tail:
4079 * -This might even be off the end of the leaf
4080 * 2) We're "appending" by rotating in the tail
4082 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4086 ocfs2_free_path(path
);
4096 * Insert an extent into an inode btree.
4098 * The caller needs to update fe->i_clusters
4100 int ocfs2_insert_extent(struct ocfs2_super
*osb
,
4102 struct inode
*inode
,
4103 struct buffer_head
*fe_bh
,
4108 struct ocfs2_alloc_context
*meta_ac
)
4111 int uninitialized_var(free_records
);
4112 struct buffer_head
*last_eb_bh
= NULL
;
4113 struct ocfs2_insert_type insert
= {0, };
4114 struct ocfs2_extent_rec rec
;
4116 BUG_ON(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
);
4118 mlog(0, "add %u clusters at position %u to inode %llu\n",
4119 new_clusters
, cpos
, (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
4121 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb
) &&
4122 (OCFS2_I(inode
)->ip_clusters
!= cpos
),
4123 "Device %s, asking for sparse allocation: inode %llu, "
4124 "cpos %u, clusters %u\n",
4126 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
,
4127 OCFS2_I(inode
)->ip_clusters
);
4129 memset(&rec
, 0, sizeof(rec
));
4130 rec
.e_cpos
= cpu_to_le32(cpos
);
4131 rec
.e_blkno
= cpu_to_le64(start_blk
);
4132 rec
.e_leaf_clusters
= cpu_to_le16(new_clusters
);
4133 rec
.e_flags
= flags
;
4135 status
= ocfs2_figure_insert_type(inode
, fe_bh
, &last_eb_bh
, &rec
,
4136 &free_records
, &insert
);
4142 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4143 "Insert.contig_index: %d, Insert.free_records: %d, "
4144 "Insert.tree_depth: %d\n",
4145 insert
.ins_appending
, insert
.ins_contig
, insert
.ins_contig_index
,
4146 free_records
, insert
.ins_tree_depth
);
4148 if (insert
.ins_contig
== CONTIG_NONE
&& free_records
== 0) {
4149 status
= ocfs2_grow_tree(inode
, handle
, fe_bh
,
4150 &insert
.ins_tree_depth
, &last_eb_bh
,
4158 /* Finally, we can add clusters. This might rotate the tree for us. */
4159 status
= ocfs2_do_insert_extent(inode
, handle
, fe_bh
, &rec
, &insert
);
4163 ocfs2_extent_map_insert_rec(inode
, &rec
);
4173 static void ocfs2_make_right_split_rec(struct super_block
*sb
,
4174 struct ocfs2_extent_rec
*split_rec
,
4176 struct ocfs2_extent_rec
*rec
)
4178 u32 rec_cpos
= le32_to_cpu(rec
->e_cpos
);
4179 u32 rec_range
= rec_cpos
+ le16_to_cpu(rec
->e_leaf_clusters
);
4181 memset(split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4183 split_rec
->e_cpos
= cpu_to_le32(cpos
);
4184 split_rec
->e_leaf_clusters
= cpu_to_le16(rec_range
- cpos
);
4186 split_rec
->e_blkno
= rec
->e_blkno
;
4187 le64_add_cpu(&split_rec
->e_blkno
,
4188 ocfs2_clusters_to_blocks(sb
, cpos
- rec_cpos
));
4190 split_rec
->e_flags
= rec
->e_flags
;
4193 static int ocfs2_split_and_insert(struct inode
*inode
,
4195 struct ocfs2_path
*path
,
4196 struct buffer_head
*di_bh
,
4197 struct buffer_head
**last_eb_bh
,
4199 struct ocfs2_extent_rec
*orig_split_rec
,
4200 struct ocfs2_alloc_context
*meta_ac
)
4203 unsigned int insert_range
, rec_range
, do_leftright
= 0;
4204 struct ocfs2_extent_rec tmprec
;
4205 struct ocfs2_extent_list
*rightmost_el
;
4206 struct ocfs2_extent_rec rec
;
4207 struct ocfs2_extent_rec split_rec
= *orig_split_rec
;
4208 struct ocfs2_insert_type insert
;
4209 struct ocfs2_extent_block
*eb
;
4210 struct ocfs2_dinode
*di
;
4214 * Store a copy of the record on the stack - it might move
4215 * around as the tree is manipulated below.
4217 rec
= path_leaf_el(path
)->l_recs
[split_index
];
4219 di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
4220 rightmost_el
= &di
->id2
.i_list
;
4222 depth
= le16_to_cpu(rightmost_el
->l_tree_depth
);
4224 BUG_ON(!(*last_eb_bh
));
4225 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
4226 rightmost_el
= &eb
->h_list
;
4229 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4230 le16_to_cpu(rightmost_el
->l_count
)) {
4231 ret
= ocfs2_grow_tree(inode
, handle
, di_bh
, &depth
, last_eb_bh
,
4239 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4240 insert
.ins_appending
= APPEND_NONE
;
4241 insert
.ins_contig
= CONTIG_NONE
;
4242 insert
.ins_tree_depth
= depth
;
4244 insert_range
= le32_to_cpu(split_rec
.e_cpos
) +
4245 le16_to_cpu(split_rec
.e_leaf_clusters
);
4246 rec_range
= le32_to_cpu(rec
.e_cpos
) +
4247 le16_to_cpu(rec
.e_leaf_clusters
);
4249 if (split_rec
.e_cpos
== rec
.e_cpos
) {
4250 insert
.ins_split
= SPLIT_LEFT
;
4251 } else if (insert_range
== rec_range
) {
4252 insert
.ins_split
= SPLIT_RIGHT
;
4255 * Left/right split. We fake this as a right split
4256 * first and then make a second pass as a left split.
4258 insert
.ins_split
= SPLIT_RIGHT
;
4260 ocfs2_make_right_split_rec(inode
->i_sb
, &tmprec
, insert_range
,
4265 BUG_ON(do_leftright
);
4269 ret
= ocfs2_do_insert_extent(inode
, handle
, di_bh
, &split_rec
,
4276 if (do_leftright
== 1) {
4278 struct ocfs2_extent_list
*el
;
4281 split_rec
= *orig_split_rec
;
4283 ocfs2_reinit_path(path
, 1);
4285 cpos
= le32_to_cpu(split_rec
.e_cpos
);
4286 ret
= ocfs2_find_path(inode
, path
, cpos
);
4292 el
= path_leaf_el(path
);
4293 split_index
= ocfs2_search_extent_list(el
, cpos
);
4302 * Mark part or all of the extent record at split_index in the leaf
4303 * pointed to by path as written. This removes the unwritten
4306 * Care is taken to handle contiguousness so as to not grow the tree.
4308 * meta_ac is not strictly necessary - we only truly need it if growth
4309 * of the tree is required. All other cases will degrade into a less
4310 * optimal tree layout.
4312 * last_eb_bh should be the rightmost leaf block for any inode with a
4313 * btree. Since a split may grow the tree or a merge might shrink it, the caller cannot trust the contents of that buffer after this call.
4315 * This code is optimized for readability - several passes might be
4316 * made over certain portions of the tree. All of those blocks will
4317 * have been brought into cache (and pinned via the journal), so the
4318 * extra overhead is not expressed in terms of disk reads.
4320 static int __ocfs2_mark_extent_written(struct inode
*inode
,
4321 struct buffer_head
*di_bh
,
4323 struct ocfs2_path
*path
,
4325 struct ocfs2_extent_rec
*split_rec
,
4326 struct ocfs2_alloc_context
*meta_ac
,
4327 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
4330 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
4331 struct buffer_head
*last_eb_bh
= NULL
;
4332 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
4333 struct ocfs2_merge_ctxt ctxt
;
4334 struct ocfs2_extent_list
*rightmost_el
;
4336 if (!(rec
->e_flags
& OCFS2_EXT_UNWRITTEN
)) {
4342 if (le32_to_cpu(rec
->e_cpos
) > le32_to_cpu(split_rec
->e_cpos
) ||
4343 ((le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)) <
4344 (le32_to_cpu(split_rec
->e_cpos
) + le16_to_cpu(split_rec
->e_leaf_clusters
)))) {
4350 ctxt
.c_contig_type
= ocfs2_figure_merge_contig_type(inode
, path
, el
,
4355 * The core merge / split code wants to know how much room is
4356 * left in this inodes allocation tree, so we pass the
4357 * rightmost extent list.
4359 if (path
->p_tree_depth
) {
4360 struct ocfs2_extent_block
*eb
;
4361 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
4363 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
4364 le64_to_cpu(di
->i_last_eb_blk
),
4365 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
4371 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
4372 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
4373 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
4378 rightmost_el
= &eb
->h_list
;
4380 rightmost_el
= path_root_el(path
);
4382 if (rec
->e_cpos
== split_rec
->e_cpos
&&
4383 rec
->e_leaf_clusters
== split_rec
->e_leaf_clusters
)
4384 ctxt
.c_split_covers_rec
= 1;
4386 ctxt
.c_split_covers_rec
= 0;
4388 ctxt
.c_has_empty_extent
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
4390 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
4391 split_index
, ctxt
.c_contig_type
, ctxt
.c_has_empty_extent
,
4392 ctxt
.c_split_covers_rec
);
4394 if (ctxt
.c_contig_type
== CONTIG_NONE
) {
4395 if (ctxt
.c_split_covers_rec
)
4396 el
->l_recs
[split_index
] = *split_rec
;
4398 ret
= ocfs2_split_and_insert(inode
, handle
, path
, di_bh
,
4399 &last_eb_bh
, split_index
,
4400 split_rec
, meta_ac
);
4404 ret
= ocfs2_try_to_merge_extent(inode
, handle
, path
,
4405 split_index
, split_rec
,
4417 * Mark the already-existing extent at cpos as written for len clusters.
4419 * If the existing extent is larger than the request, initiate a
4420 * split. An attempt will be made at merging with adjacent extents.
4422 * The caller is responsible for passing down meta_ac if we'll need it.
4424 int ocfs2_mark_extent_written(struct inode
*inode
, struct buffer_head
*di_bh
,
4425 handle_t
*handle
, u32 cpos
, u32 len
, u32 phys
,
4426 struct ocfs2_alloc_context
*meta_ac
,
4427 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
4430 u64 start_blkno
= ocfs2_clusters_to_blocks(inode
->i_sb
, phys
);
4431 struct ocfs2_extent_rec split_rec
;
4432 struct ocfs2_path
*left_path
= NULL
;
4433 struct ocfs2_extent_list
*el
;
4435 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4436 inode
->i_ino
, cpos
, len
, phys
, (unsigned long long)start_blkno
);
4438 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode
->i_sb
))) {
4439 ocfs2_error(inode
->i_sb
, "Inode %llu has unwritten extents "
4440 "that are being written to, but the feature bit "
4441 "is not set in the super block.",
4442 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
4448 * XXX: This should be fixed up so that we just re-insert the
4449 * next extent records.
4451 ocfs2_extent_map_trunc(inode
, 0);
4453 left_path
= ocfs2_new_inode_path(di_bh
);
4460 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
4465 el
= path_leaf_el(left_path
);
4467 index
= ocfs2_search_extent_list(el
, cpos
);
4468 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
4469 ocfs2_error(inode
->i_sb
,
4470 "Inode %llu has an extent at cpos %u which can no "
4471 "longer be found.\n",
4472 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
);
4477 memset(&split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4478 split_rec
.e_cpos
= cpu_to_le32(cpos
);
4479 split_rec
.e_leaf_clusters
= cpu_to_le16(len
);
4480 split_rec
.e_blkno
= cpu_to_le64(start_blkno
);
4481 split_rec
.e_flags
= path_leaf_el(left_path
)->l_recs
[index
].e_flags
;
4482 split_rec
.e_flags
&= ~OCFS2_EXT_UNWRITTEN
;
4484 ret
= __ocfs2_mark_extent_written(inode
, di_bh
, handle
, left_path
,
4485 index
, &split_rec
, meta_ac
, dealloc
);
4490 ocfs2_free_path(left_path
);
4494 static int ocfs2_split_tree(struct inode
*inode
, struct buffer_head
*di_bh
,
4495 handle_t
*handle
, struct ocfs2_path
*path
,
4496 int index
, u32 new_range
,
4497 struct ocfs2_alloc_context
*meta_ac
)
4499 int ret
, depth
, credits
= handle
->h_buffer_credits
;
4500 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
4501 struct buffer_head
*last_eb_bh
= NULL
;
4502 struct ocfs2_extent_block
*eb
;
4503 struct ocfs2_extent_list
*rightmost_el
, *el
;
4504 struct ocfs2_extent_rec split_rec
;
4505 struct ocfs2_extent_rec
*rec
;
4506 struct ocfs2_insert_type insert
;
4509 * Setup the record to split before we grow the tree.
4511 el
= path_leaf_el(path
);
4512 rec
= &el
->l_recs
[index
];
4513 ocfs2_make_right_split_rec(inode
->i_sb
, &split_rec
, new_range
, rec
);
4515 depth
= path
->p_tree_depth
;
4517 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
4518 le64_to_cpu(di
->i_last_eb_blk
),
4519 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
4525 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
4526 rightmost_el
= &eb
->h_list
;
4528 rightmost_el
= path_leaf_el(path
);
4530 credits
+= path
->p_tree_depth
+
4531 ocfs2_extend_meta_needed(&di
->id2
.i_list
);
4532 ret
= ocfs2_extend_trans(handle
, credits
);
4538 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4539 le16_to_cpu(rightmost_el
->l_count
)) {
4540 ret
= ocfs2_grow_tree(inode
, handle
, di_bh
, &depth
, &last_eb_bh
,
4548 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4549 insert
.ins_appending
= APPEND_NONE
;
4550 insert
.ins_contig
= CONTIG_NONE
;
4551 insert
.ins_split
= SPLIT_RIGHT
;
4552 insert
.ins_tree_depth
= depth
;
4554 ret
= ocfs2_do_insert_extent(inode
, handle
, di_bh
, &split_rec
, &insert
);
4563 static int ocfs2_truncate_rec(struct inode
*inode
, handle_t
*handle
,
4564 struct ocfs2_path
*path
, int index
,
4565 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
4569 u32 left_cpos
, rec_range
, trunc_range
;
4570 int wants_rotate
= 0, is_rightmost_tree_rec
= 0;
4571 struct super_block
*sb
= inode
->i_sb
;
4572 struct ocfs2_path
*left_path
= NULL
;
4573 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
4574 struct ocfs2_extent_rec
*rec
;
4575 struct ocfs2_extent_block
*eb
;
4577 if (ocfs2_is_empty_extent(&el
->l_recs
[0]) && index
> 0) {
4578 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
, dealloc
);
4587 if (index
== (le16_to_cpu(el
->l_next_free_rec
) - 1) &&
4588 path
->p_tree_depth
) {
4590 * Check whether this is the rightmost tree record. If
4591 * we remove all of this record or part of its right
4592 * edge then an update of the record lengths above it
4595 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
4596 if (eb
->h_next_leaf_blk
== 0)
4597 is_rightmost_tree_rec
= 1;
4600 rec
= &el
->l_recs
[index
];
4601 if (index
== 0 && path
->p_tree_depth
&&
4602 le32_to_cpu(rec
->e_cpos
) == cpos
) {
4604 * Changing the leftmost offset (via partial or whole
4605 * record truncate) of an interior (or rightmost) path
4606 * means we have to update the subtree that is formed
4607 * by this leaf and the one to it's left.
4609 * There are two cases we can skip:
4610 * 1) Path is the leftmost one in our inode tree.
4611 * 2) The leaf is rightmost and will be empty after
4612 * we remove the extent record - the rotate code
4613 * knows how to update the newly formed edge.
4616 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
,
4623 if (left_cpos
&& le16_to_cpu(el
->l_next_free_rec
) > 1) {
4624 left_path
= ocfs2_new_path(path_root_bh(path
),
4625 path_root_el(path
));
4632 ret
= ocfs2_find_path(inode
, left_path
, left_cpos
);
4640 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
4641 handle
->h_buffer_credits
,
4648 ret
= ocfs2_journal_access_path(inode
, handle
, path
);
4654 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
4660 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
4661 trunc_range
= cpos
+ len
;
4663 if (le32_to_cpu(rec
->e_cpos
) == cpos
&& rec_range
== trunc_range
) {
4666 memset(rec
, 0, sizeof(*rec
));
4667 ocfs2_cleanup_merge(el
, index
);
4670 next_free
= le16_to_cpu(el
->l_next_free_rec
);
4671 if (is_rightmost_tree_rec
&& next_free
> 1) {
4673 * We skip the edge update if this path will
4674 * be deleted by the rotate code.
4676 rec
= &el
->l_recs
[next_free
- 1];
4677 ocfs2_adjust_rightmost_records(inode
, handle
, path
,
4680 } else if (le32_to_cpu(rec
->e_cpos
) == cpos
) {
4681 /* Remove leftmost portion of the record. */
4682 le32_add_cpu(&rec
->e_cpos
, len
);
4683 le64_add_cpu(&rec
->e_blkno
, ocfs2_clusters_to_blocks(sb
, len
));
4684 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
4685 } else if (rec_range
== trunc_range
) {
4686 /* Remove rightmost portion of the record */
4687 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
4688 if (is_rightmost_tree_rec
)
4689 ocfs2_adjust_rightmost_records(inode
, handle
, path
, rec
);
4691 /* Caller should have trapped this. */
4692 mlog(ML_ERROR
, "Inode %llu: Invalid record truncate: (%u, %u) "
4693 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
4694 le32_to_cpu(rec
->e_cpos
),
4695 le16_to_cpu(rec
->e_leaf_clusters
), cpos
, len
);
4702 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
, path
);
4703 ocfs2_complete_edge_insert(inode
, handle
, left_path
, path
,
4707 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
4709 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
, dealloc
);
4716 ocfs2_free_path(left_path
);
4720 int ocfs2_remove_extent(struct inode
*inode
, struct buffer_head
*di_bh
,
4721 u32 cpos
, u32 len
, handle_t
*handle
,
4722 struct ocfs2_alloc_context
*meta_ac
,
4723 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
4726 u32 rec_range
, trunc_range
;
4727 struct ocfs2_extent_rec
*rec
;
4728 struct ocfs2_extent_list
*el
;
4729 struct ocfs2_path
*path
;
4731 ocfs2_extent_map_trunc(inode
, 0);
4733 path
= ocfs2_new_inode_path(di_bh
);
4740 ret
= ocfs2_find_path(inode
, path
, cpos
);
4746 el
= path_leaf_el(path
);
4747 index
= ocfs2_search_extent_list(el
, cpos
);
4748 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
4749 ocfs2_error(inode
->i_sb
,
4750 "Inode %llu has an extent at cpos %u which can no "
4751 "longer be found.\n",
4752 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
);
4758 * We have 3 cases of extent removal:
4759 * 1) Range covers the entire extent rec
4760 * 2) Range begins or ends on one edge of the extent rec
4761 * 3) Range is in the middle of the extent rec (no shared edges)
4763 * For case 1 we remove the extent rec and left rotate to
4766 * For case 2 we just shrink the existing extent rec, with a
4767 * tree update if the shrinking edge is also the edge of an
4770 * For case 3 we do a right split to turn the extent rec into
4771 * something case 2 can handle.
4773 rec
= &el
->l_recs
[index
];
4774 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
4775 trunc_range
= cpos
+ len
;
4777 BUG_ON(cpos
< le32_to_cpu(rec
->e_cpos
) || trunc_range
> rec_range
);
4779 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
4780 "(cpos %u, len %u)\n",
4781 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
, len
, index
,
4782 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
));
4784 if (le32_to_cpu(rec
->e_cpos
) == cpos
|| rec_range
== trunc_range
) {
4785 ret
= ocfs2_truncate_rec(inode
, handle
, path
, index
, dealloc
,
4792 ret
= ocfs2_split_tree(inode
, di_bh
, handle
, path
, index
,
4793 trunc_range
, meta_ac
);
4800 * The split could have manipulated the tree enough to
4801 * move the record location, so we have to look for it again.
4803 ocfs2_reinit_path(path
, 1);
4805 ret
= ocfs2_find_path(inode
, path
, cpos
);
4811 el
= path_leaf_el(path
);
4812 index
= ocfs2_search_extent_list(el
, cpos
);
4813 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
4814 ocfs2_error(inode
->i_sb
,
4815 "Inode %llu: split at cpos %u lost record.",
4816 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
4823 * Double check our values here. If anything is fishy,
4824 * it's easier to catch it at the top level.
4826 rec
= &el
->l_recs
[index
];
4827 rec_range
= le32_to_cpu(rec
->e_cpos
) +
4828 ocfs2_rec_clusters(el
, rec
);
4829 if (rec_range
!= trunc_range
) {
4830 ocfs2_error(inode
->i_sb
,
4831 "Inode %llu: error after split at cpos %u"
4832 "trunc len %u, existing record is (%u,%u)",
4833 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
4834 cpos
, len
, le32_to_cpu(rec
->e_cpos
),
4835 ocfs2_rec_clusters(el
, rec
));
4840 ret
= ocfs2_truncate_rec(inode
, handle
, path
, index
, dealloc
,
4849 ocfs2_free_path(path
);
4853 int ocfs2_truncate_log_needs_flush(struct ocfs2_super
*osb
)
4855 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
4856 struct ocfs2_dinode
*di
;
4857 struct ocfs2_truncate_log
*tl
;
4859 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4860 tl
= &di
->id2
.i_dealloc
;
4862 mlog_bug_on_msg(le16_to_cpu(tl
->tl_used
) > le16_to_cpu(tl
->tl_count
),
4863 "slot %d, invalid truncate log parameters: used = "
4864 "%u, count = %u\n", osb
->slot_num
,
4865 le16_to_cpu(tl
->tl_used
), le16_to_cpu(tl
->tl_count
));
4866 return le16_to_cpu(tl
->tl_used
) == le16_to_cpu(tl
->tl_count
);
4869 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log
*tl
,
4870 unsigned int new_start
)
4872 unsigned int tail_index
;
4873 unsigned int current_tail
;
4875 /* No records, nothing to coalesce */
4876 if (!le16_to_cpu(tl
->tl_used
))
4879 tail_index
= le16_to_cpu(tl
->tl_used
) - 1;
4880 current_tail
= le32_to_cpu(tl
->tl_recs
[tail_index
].t_start
);
4881 current_tail
+= le32_to_cpu(tl
->tl_recs
[tail_index
].t_clusters
);
4883 return current_tail
== new_start
;
4886 int ocfs2_truncate_log_append(struct ocfs2_super
*osb
,
4889 unsigned int num_clusters
)
4892 unsigned int start_cluster
, tl_count
;
4893 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4894 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
4895 struct ocfs2_dinode
*di
;
4896 struct ocfs2_truncate_log
*tl
;
4898 mlog_entry("start_blk = %llu, num_clusters = %u\n",
4899 (unsigned long long)start_blk
, num_clusters
);
4901 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
4903 start_cluster
= ocfs2_blocks_to_clusters(osb
->sb
, start_blk
);
4905 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4906 tl
= &di
->id2
.i_dealloc
;
4907 if (!OCFS2_IS_VALID_DINODE(di
)) {
4908 OCFS2_RO_ON_INVALID_DINODE(osb
->sb
, di
);
4913 tl_count
= le16_to_cpu(tl
->tl_count
);
4914 mlog_bug_on_msg(tl_count
> ocfs2_truncate_recs_per_inode(osb
->sb
) ||
4916 "Truncate record count on #%llu invalid "
4917 "wanted %u, actual %u\n",
4918 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
4919 ocfs2_truncate_recs_per_inode(osb
->sb
),
4920 le16_to_cpu(tl
->tl_count
));
4922 /* Caller should have known to flush before calling us. */
4923 index
= le16_to_cpu(tl
->tl_used
);
4924 if (index
>= tl_count
) {
4930 status
= ocfs2_journal_access(handle
, tl_inode
, tl_bh
,
4931 OCFS2_JOURNAL_ACCESS_WRITE
);
4937 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
4938 "%llu (index = %d)\n", num_clusters
, start_cluster
,
4939 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
, index
);
4941 if (ocfs2_truncate_log_can_coalesce(tl
, start_cluster
)) {
4943 * Move index back to the record we are coalescing with.
4944 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
4948 num_clusters
+= le32_to_cpu(tl
->tl_recs
[index
].t_clusters
);
4949 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
4950 index
, le32_to_cpu(tl
->tl_recs
[index
].t_start
),
4953 tl
->tl_recs
[index
].t_start
= cpu_to_le32(start_cluster
);
4954 tl
->tl_used
= cpu_to_le16(index
+ 1);
4956 tl
->tl_recs
[index
].t_clusters
= cpu_to_le32(num_clusters
);
4958 status
= ocfs2_journal_dirty(handle
, tl_bh
);
4969 static int ocfs2_replay_truncate_records(struct ocfs2_super
*osb
,
4971 struct inode
*data_alloc_inode
,
4972 struct buffer_head
*data_alloc_bh
)
4976 unsigned int num_clusters
;
4978 struct ocfs2_truncate_rec rec
;
4979 struct ocfs2_dinode
*di
;
4980 struct ocfs2_truncate_log
*tl
;
4981 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4982 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
4986 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4987 tl
= &di
->id2
.i_dealloc
;
4988 i
= le16_to_cpu(tl
->tl_used
) - 1;
4990 /* Caller has given us at least enough credits to
4991 * update the truncate log dinode */
4992 status
= ocfs2_journal_access(handle
, tl_inode
, tl_bh
,
4993 OCFS2_JOURNAL_ACCESS_WRITE
);
4999 tl
->tl_used
= cpu_to_le16(i
);
5001 status
= ocfs2_journal_dirty(handle
, tl_bh
);
5007 /* TODO: Perhaps we can calculate the bulk of the
5008 * credits up front rather than extending like
5010 status
= ocfs2_extend_trans(handle
,
5011 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC
);
5017 rec
= tl
->tl_recs
[i
];
5018 start_blk
= ocfs2_clusters_to_blocks(data_alloc_inode
->i_sb
,
5019 le32_to_cpu(rec
.t_start
));
5020 num_clusters
= le32_to_cpu(rec
.t_clusters
);
5022 /* if start_blk is not set, we ignore the record as
5025 mlog(0, "free record %d, start = %u, clusters = %u\n",
5026 i
, le32_to_cpu(rec
.t_start
), num_clusters
);
5028 status
= ocfs2_free_clusters(handle
, data_alloc_inode
,
5029 data_alloc_bh
, start_blk
,
5044 /* Expects you to already be holding tl_inode->i_mutex */
5045 int __ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5048 unsigned int num_to_flush
;
5050 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5051 struct inode
*data_alloc_inode
= NULL
;
5052 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5053 struct buffer_head
*data_alloc_bh
= NULL
;
5054 struct ocfs2_dinode
*di
;
5055 struct ocfs2_truncate_log
*tl
;
5059 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
5061 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5062 tl
= &di
->id2
.i_dealloc
;
5063 if (!OCFS2_IS_VALID_DINODE(di
)) {
5064 OCFS2_RO_ON_INVALID_DINODE(osb
->sb
, di
);
5069 num_to_flush
= le16_to_cpu(tl
->tl_used
);
5070 mlog(0, "Flush %u records from truncate log #%llu\n",
5071 num_to_flush
, (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
);
5072 if (!num_to_flush
) {
5077 data_alloc_inode
= ocfs2_get_system_file_inode(osb
,
5078 GLOBAL_BITMAP_SYSTEM_INODE
,
5079 OCFS2_INVALID_SLOT
);
5080 if (!data_alloc_inode
) {
5082 mlog(ML_ERROR
, "Could not get bitmap inode!\n");
5086 mutex_lock(&data_alloc_inode
->i_mutex
);
5088 status
= ocfs2_inode_lock(data_alloc_inode
, &data_alloc_bh
, 1);
5094 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
5095 if (IS_ERR(handle
)) {
5096 status
= PTR_ERR(handle
);
5101 status
= ocfs2_replay_truncate_records(osb
, handle
, data_alloc_inode
,
5106 ocfs2_commit_trans(osb
, handle
);
5109 brelse(data_alloc_bh
);
5110 ocfs2_inode_unlock(data_alloc_inode
, 1);
5113 mutex_unlock(&data_alloc_inode
->i_mutex
);
5114 iput(data_alloc_inode
);
5121 int ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5124 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5126 mutex_lock(&tl_inode
->i_mutex
);
5127 status
= __ocfs2_flush_truncate_log(osb
);
5128 mutex_unlock(&tl_inode
->i_mutex
);
5133 static void ocfs2_truncate_log_worker(struct work_struct
*work
)
5136 struct ocfs2_super
*osb
=
5137 container_of(work
, struct ocfs2_super
,
5138 osb_truncate_log_wq
.work
);
5142 status
= ocfs2_flush_truncate_log(osb
);
5146 ocfs2_init_inode_steal_slot(osb
);
5151 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5152 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super
*osb
,
5155 if (osb
->osb_tl_inode
) {
5156 /* We want to push off log flushes while truncates are
5159 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
5161 queue_delayed_work(ocfs2_wq
, &osb
->osb_truncate_log_wq
,
5162 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL
);
5166 static int ocfs2_get_truncate_log_info(struct ocfs2_super
*osb
,
5168 struct inode
**tl_inode
,
5169 struct buffer_head
**tl_bh
)
5172 struct inode
*inode
= NULL
;
5173 struct buffer_head
*bh
= NULL
;
5175 inode
= ocfs2_get_system_file_inode(osb
,
5176 TRUNCATE_LOG_SYSTEM_INODE
,
5180 mlog(ML_ERROR
, "Could not get load truncate log inode!\n");
5184 status
= ocfs2_read_block(osb
, OCFS2_I(inode
)->ip_blkno
, &bh
,
5185 OCFS2_BH_CACHED
, inode
);
5199 /* called during the 1st stage of node recovery. we stamp a clean
5200 * truncate log and pass back a copy for processing later. if the
5201 * truncate log does not require processing, a *tl_copy is set to
5203 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super
*osb
,
5205 struct ocfs2_dinode
**tl_copy
)
5208 struct inode
*tl_inode
= NULL
;
5209 struct buffer_head
*tl_bh
= NULL
;
5210 struct ocfs2_dinode
*di
;
5211 struct ocfs2_truncate_log
*tl
;
5215 mlog(0, "recover truncate log from slot %d\n", slot_num
);
5217 status
= ocfs2_get_truncate_log_info(osb
, slot_num
, &tl_inode
, &tl_bh
);
5223 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5224 tl
= &di
->id2
.i_dealloc
;
5225 if (!OCFS2_IS_VALID_DINODE(di
)) {
5226 OCFS2_RO_ON_INVALID_DINODE(tl_inode
->i_sb
, di
);
5231 if (le16_to_cpu(tl
->tl_used
)) {
5232 mlog(0, "We'll have %u logs to recover\n",
5233 le16_to_cpu(tl
->tl_used
));
5235 *tl_copy
= kmalloc(tl_bh
->b_size
, GFP_KERNEL
);
5242 /* Assuming the write-out below goes well, this copy
5243 * will be passed back to recovery for processing. */
5244 memcpy(*tl_copy
, tl_bh
->b_data
, tl_bh
->b_size
);
5246 /* All we need to do to clear the truncate log is set
5250 status
= ocfs2_write_block(osb
, tl_bh
, tl_inode
);
5263 if (status
< 0 && (*tl_copy
)) {
5272 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super
*osb
,
5273 struct ocfs2_dinode
*tl_copy
)
5277 unsigned int clusters
, num_recs
, start_cluster
;
5280 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5281 struct ocfs2_truncate_log
*tl
;
5285 if (OCFS2_I(tl_inode
)->ip_blkno
== le64_to_cpu(tl_copy
->i_blkno
)) {
5286 mlog(ML_ERROR
, "Asked to recover my own truncate log!\n");
5290 tl
= &tl_copy
->id2
.i_dealloc
;
5291 num_recs
= le16_to_cpu(tl
->tl_used
);
5292 mlog(0, "cleanup %u records from %llu\n", num_recs
,
5293 (unsigned long long)le64_to_cpu(tl_copy
->i_blkno
));
5295 mutex_lock(&tl_inode
->i_mutex
);
5296 for(i
= 0; i
< num_recs
; i
++) {
5297 if (ocfs2_truncate_log_needs_flush(osb
)) {
5298 status
= __ocfs2_flush_truncate_log(osb
);
5305 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
5306 if (IS_ERR(handle
)) {
5307 status
= PTR_ERR(handle
);
5312 clusters
= le32_to_cpu(tl
->tl_recs
[i
].t_clusters
);
5313 start_cluster
= le32_to_cpu(tl
->tl_recs
[i
].t_start
);
5314 start_blk
= ocfs2_clusters_to_blocks(osb
->sb
, start_cluster
);
5316 status
= ocfs2_truncate_log_append(osb
, handle
,
5317 start_blk
, clusters
);
5318 ocfs2_commit_trans(osb
, handle
);
5326 mutex_unlock(&tl_inode
->i_mutex
);
5332 void ocfs2_truncate_log_shutdown(struct ocfs2_super
*osb
)
5335 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5340 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
5341 flush_workqueue(ocfs2_wq
);
5343 status
= ocfs2_flush_truncate_log(osb
);
5347 brelse(osb
->osb_tl_bh
);
5348 iput(osb
->osb_tl_inode
);
5354 int ocfs2_truncate_log_init(struct ocfs2_super
*osb
)
5357 struct inode
*tl_inode
= NULL
;
5358 struct buffer_head
*tl_bh
= NULL
;
5362 status
= ocfs2_get_truncate_log_info(osb
,
5369 /* ocfs2_truncate_log_shutdown keys on the existence of
5370 * osb->osb_tl_inode so we don't set any of the osb variables
5371 * until we're sure all is well. */
5372 INIT_DELAYED_WORK(&osb
->osb_truncate_log_wq
,
5373 ocfs2_truncate_log_worker
);
5374 osb
->osb_tl_bh
= tl_bh
;
5375 osb
->osb_tl_inode
= tl_inode
;
5382 * Delayed de-allocation of suballocator blocks.
5384 * Some sets of block de-allocations might involve multiple suballocator inodes.
5386 * The locking for this can get extremely complicated, especially when
5387 * the suballocator inodes to delete from aren't known until deep
5388 * within an unrelated codepath.
5390 * ocfs2_extent_block structures are a good example of this - an inode
5391 * btree could have been grown by any number of nodes each allocating
5392 * out of their own suballoc inode.
5394 * These structures allow the delay of block de-allocation until a
5395 * later time, when locking of multiple cluster inodes won't cause
5400 * Describes a single block free from a suballocator
5402 struct ocfs2_cached_block_free
{
5403 struct ocfs2_cached_block_free
*free_next
;
5405 unsigned int free_bit
;
5408 struct ocfs2_per_slot_free_list
{
5409 struct ocfs2_per_slot_free_list
*f_next_suballocator
;
5412 struct ocfs2_cached_block_free
*f_first
;
5415 static int ocfs2_free_cached_items(struct ocfs2_super
*osb
,
5418 struct ocfs2_cached_block_free
*head
)
5423 struct inode
*inode
;
5424 struct buffer_head
*di_bh
= NULL
;
5425 struct ocfs2_cached_block_free
*tmp
;
5427 inode
= ocfs2_get_system_file_inode(osb
, sysfile_type
, slot
);
5434 mutex_lock(&inode
->i_mutex
);
5436 ret
= ocfs2_inode_lock(inode
, &di_bh
, 1);
5442 handle
= ocfs2_start_trans(osb
, OCFS2_SUBALLOC_FREE
);
5443 if (IS_ERR(handle
)) {
5444 ret
= PTR_ERR(handle
);
5450 bg_blkno
= ocfs2_which_suballoc_group(head
->free_blk
,
5452 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
5453 head
->free_bit
, (unsigned long long)head
->free_blk
);
5455 ret
= ocfs2_free_suballoc_bits(handle
, inode
, di_bh
,
5456 head
->free_bit
, bg_blkno
, 1);
5462 ret
= ocfs2_extend_trans(handle
, OCFS2_SUBALLOC_FREE
);
5469 head
= head
->free_next
;
5474 ocfs2_commit_trans(osb
, handle
);
5477 ocfs2_inode_unlock(inode
, 1);
5480 mutex_unlock(&inode
->i_mutex
);
5484 /* Premature exit may have left some dangling items. */
5486 head
= head
->free_next
;
5493 int ocfs2_run_deallocs(struct ocfs2_super
*osb
,
5494 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
5497 struct ocfs2_per_slot_free_list
*fl
;
5502 while (ctxt
->c_first_suballocator
) {
5503 fl
= ctxt
->c_first_suballocator
;
5506 mlog(0, "Free items: (type %u, slot %d)\n",
5507 fl
->f_inode_type
, fl
->f_slot
);
5508 ret2
= ocfs2_free_cached_items(osb
, fl
->f_inode_type
,
5509 fl
->f_slot
, fl
->f_first
);
5516 ctxt
->c_first_suballocator
= fl
->f_next_suballocator
;
5523 static struct ocfs2_per_slot_free_list
*
5524 ocfs2_find_per_slot_free_list(int type
,
5526 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
5528 struct ocfs2_per_slot_free_list
*fl
= ctxt
->c_first_suballocator
;
5531 if (fl
->f_inode_type
== type
&& fl
->f_slot
== slot
)
5534 fl
= fl
->f_next_suballocator
;
5537 fl
= kmalloc(sizeof(*fl
), GFP_NOFS
);
5539 fl
->f_inode_type
= type
;
5542 fl
->f_next_suballocator
= ctxt
->c_first_suballocator
;
5544 ctxt
->c_first_suballocator
= fl
;
5549 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
5550 int type
, int slot
, u64 blkno
,
5554 struct ocfs2_per_slot_free_list
*fl
;
5555 struct ocfs2_cached_block_free
*item
;
5557 fl
= ocfs2_find_per_slot_free_list(type
, slot
, ctxt
);
5564 item
= kmalloc(sizeof(*item
), GFP_NOFS
);
5571 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
5572 type
, slot
, bit
, (unsigned long long)blkno
);
5574 item
->free_blk
= blkno
;
5575 item
->free_bit
= bit
;
5576 item
->free_next
= fl
->f_first
;
5585 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
5586 struct ocfs2_extent_block
*eb
)
5588 return ocfs2_cache_block_dealloc(ctxt
, EXTENT_ALLOC_SYSTEM_INODE
,
5589 le16_to_cpu(eb
->h_suballoc_slot
),
5590 le64_to_cpu(eb
->h_blkno
),
5591 le16_to_cpu(eb
->h_suballoc_bit
));
5594 /* This function will figure out whether the currently last extent
5595 * block will be deleted, and if it will, what the new last extent
5596 * block will be so we can update his h_next_leaf_blk field, as well
5597 * as the dinodes i_last_eb_blk */
5598 static int ocfs2_find_new_last_ext_blk(struct inode
*inode
,
5599 unsigned int clusters_to_del
,
5600 struct ocfs2_path
*path
,
5601 struct buffer_head
**new_last_eb
)
5603 int next_free
, ret
= 0;
5605 struct ocfs2_extent_rec
*rec
;
5606 struct ocfs2_extent_block
*eb
;
5607 struct ocfs2_extent_list
*el
;
5608 struct buffer_head
*bh
= NULL
;
5610 *new_last_eb
= NULL
;
5612 /* we have no tree, so of course, no last_eb. */
5613 if (!path
->p_tree_depth
)
5616 /* trunc to zero special case - this makes tree_depth = 0
5617 * regardless of what it is. */
5618 if (OCFS2_I(inode
)->ip_clusters
== clusters_to_del
)
5621 el
= path_leaf_el(path
);
5622 BUG_ON(!el
->l_next_free_rec
);
5625 * Make sure that this extent list will actually be empty
5626 * after we clear away the data. We can shortcut out if
5627 * there's more than one non-empty extent in the
5628 * list. Otherwise, a check of the remaining extent is
5631 next_free
= le16_to_cpu(el
->l_next_free_rec
);
5633 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
5637 /* We may have a valid extent in index 1, check it. */
5639 rec
= &el
->l_recs
[1];
5642 * Fall through - no more nonempty extents, so we want
5643 * to delete this leaf.
5649 rec
= &el
->l_recs
[0];
5654 * Check it we'll only be trimming off the end of this
5657 if (le16_to_cpu(rec
->e_leaf_clusters
) > clusters_to_del
)
5661 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
5667 ret
= ocfs2_find_leaf(inode
, path_root_el(path
), cpos
, &bh
);
5673 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
5675 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
5676 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
5682 get_bh(*new_last_eb
);
5683 mlog(0, "returning block %llu, (cpos: %u)\n",
5684 (unsigned long long)le64_to_cpu(eb
->h_blkno
), cpos
);
5692 * Trim some clusters off the rightmost edge of a tree. Only called
5695 * The caller needs to:
5696 * - start journaling of each path component.
5697 * - compute and fully set up any new last ext block
5699 static int ocfs2_trim_tree(struct inode
*inode
, struct ocfs2_path
*path
,
5700 handle_t
*handle
, struct ocfs2_truncate_context
*tc
,
5701 u32 clusters_to_del
, u64
*delete_start
)
5703 int ret
, i
, index
= path
->p_tree_depth
;
5706 struct buffer_head
*bh
;
5707 struct ocfs2_extent_list
*el
;
5708 struct ocfs2_extent_rec
*rec
;
5712 while (index
>= 0) {
5713 bh
= path
->p_node
[index
].bh
;
5714 el
= path
->p_node
[index
].el
;
5716 mlog(0, "traveling tree (index = %d, block = %llu)\n",
5717 index
, (unsigned long long)bh
->b_blocknr
);
5719 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
5722 (path
->p_tree_depth
- le16_to_cpu(el
->l_tree_depth
))) {
5723 ocfs2_error(inode
->i_sb
,
5724 "Inode %lu has invalid ext. block %llu",
5726 (unsigned long long)bh
->b_blocknr
);
5732 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
5733 rec
= &el
->l_recs
[i
];
5735 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
5736 "next = %u\n", i
, le32_to_cpu(rec
->e_cpos
),
5737 ocfs2_rec_clusters(el
, rec
),
5738 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
5739 le16_to_cpu(el
->l_next_free_rec
));
5741 BUG_ON(ocfs2_rec_clusters(el
, rec
) < clusters_to_del
);
5743 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
5745 * If the leaf block contains a single empty
5746 * extent and no records, we can just remove
5749 if (i
== 0 && ocfs2_is_empty_extent(rec
)) {
5751 sizeof(struct ocfs2_extent_rec
));
5752 el
->l_next_free_rec
= cpu_to_le16(0);
5758 * Remove any empty extents by shifting things
5759 * left. That should make life much easier on
5760 * the code below. This condition is rare
5761 * enough that we shouldn't see a performance
5764 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
5765 le16_add_cpu(&el
->l_next_free_rec
, -1);
5768 i
< le16_to_cpu(el
->l_next_free_rec
); i
++)
5769 el
->l_recs
[i
] = el
->l_recs
[i
+ 1];
5771 memset(&el
->l_recs
[i
], 0,
5772 sizeof(struct ocfs2_extent_rec
));
5775 * We've modified our extent list. The
5776 * simplest way to handle this change
5777 * is to being the search from the
5780 goto find_tail_record
;
5783 le16_add_cpu(&rec
->e_leaf_clusters
, -clusters_to_del
);
5786 * We'll use "new_edge" on our way back up the
5787 * tree to know what our rightmost cpos is.
5789 new_edge
= le16_to_cpu(rec
->e_leaf_clusters
);
5790 new_edge
+= le32_to_cpu(rec
->e_cpos
);
5793 * The caller will use this to delete data blocks.
5795 *delete_start
= le64_to_cpu(rec
->e_blkno
)
5796 + ocfs2_clusters_to_blocks(inode
->i_sb
,
5797 le16_to_cpu(rec
->e_leaf_clusters
));
5800 * If it's now empty, remove this record.
5802 if (le16_to_cpu(rec
->e_leaf_clusters
) == 0) {
5804 sizeof(struct ocfs2_extent_rec
));
5805 le16_add_cpu(&el
->l_next_free_rec
, -1);
5808 if (le64_to_cpu(rec
->e_blkno
) == deleted_eb
) {
5810 sizeof(struct ocfs2_extent_rec
));
5811 le16_add_cpu(&el
->l_next_free_rec
, -1);
5816 /* Can this actually happen? */
5817 if (le16_to_cpu(el
->l_next_free_rec
) == 0)
5821 * We never actually deleted any clusters
5822 * because our leaf was empty. There's no
5823 * reason to adjust the rightmost edge then.
5828 rec
->e_int_clusters
= cpu_to_le32(new_edge
);
5829 le32_add_cpu(&rec
->e_int_clusters
,
5830 -le32_to_cpu(rec
->e_cpos
));
5833 * A deleted child record should have been
5836 BUG_ON(le32_to_cpu(rec
->e_int_clusters
) == 0);
5840 ret
= ocfs2_journal_dirty(handle
, bh
);
5846 mlog(0, "extent list container %llu, after: record %d: "
5847 "(%u, %u, %llu), next = %u.\n",
5848 (unsigned long long)bh
->b_blocknr
, i
,
5849 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
),
5850 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
5851 le16_to_cpu(el
->l_next_free_rec
));
5854 * We must be careful to only attempt delete of an
5855 * extent block (and not the root inode block).
5857 if (index
> 0 && le16_to_cpu(el
->l_next_free_rec
) == 0) {
5858 struct ocfs2_extent_block
*eb
=
5859 (struct ocfs2_extent_block
*)bh
->b_data
;
5862 * Save this for use when processing the
5865 deleted_eb
= le64_to_cpu(eb
->h_blkno
);
5867 mlog(0, "deleting this extent block.\n");
5869 ocfs2_remove_from_cache(inode
, bh
);
5871 BUG_ON(ocfs2_rec_clusters(el
, &el
->l_recs
[0]));
5872 BUG_ON(le32_to_cpu(el
->l_recs
[0].e_cpos
));
5873 BUG_ON(le64_to_cpu(el
->l_recs
[0].e_blkno
));
5875 ret
= ocfs2_cache_extent_block_free(&tc
->tc_dealloc
, eb
);
5876 /* An error here is not fatal. */
5891 static int ocfs2_do_truncate(struct ocfs2_super
*osb
,
5892 unsigned int clusters_to_del
,
5893 struct inode
*inode
,
5894 struct buffer_head
*fe_bh
,
5896 struct ocfs2_truncate_context
*tc
,
5897 struct ocfs2_path
*path
)
5900 struct ocfs2_dinode
*fe
;
5901 struct ocfs2_extent_block
*last_eb
= NULL
;
5902 struct ocfs2_extent_list
*el
;
5903 struct buffer_head
*last_eb_bh
= NULL
;
5906 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
5908 status
= ocfs2_find_new_last_ext_blk(inode
, clusters_to_del
,
5916 * Each component will be touched, so we might as well journal
5917 * here to avoid having to handle errors later.
5919 status
= ocfs2_journal_access_path(inode
, handle
, path
);
5926 status
= ocfs2_journal_access(handle
, inode
, last_eb_bh
,
5927 OCFS2_JOURNAL_ACCESS_WRITE
);
5933 last_eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5936 el
= &(fe
->id2
.i_list
);
5939 * Lower levels depend on this never happening, but it's best
5940 * to check it up here before changing the tree.
5942 if (el
->l_tree_depth
&& el
->l_recs
[0].e_int_clusters
== 0) {
5943 ocfs2_error(inode
->i_sb
,
5944 "Inode %lu has an empty extent record, depth %u\n",
5945 inode
->i_ino
, le16_to_cpu(el
->l_tree_depth
));
5950 spin_lock(&OCFS2_I(inode
)->ip_lock
);
5951 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
) -
5953 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
5954 le32_add_cpu(&fe
->i_clusters
, -clusters_to_del
);
5955 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
5957 status
= ocfs2_trim_tree(inode
, path
, handle
, tc
,
5958 clusters_to_del
, &delete_blk
);
5964 if (le32_to_cpu(fe
->i_clusters
) == 0) {
5965 /* trunc to zero is a special case. */
5966 el
->l_tree_depth
= 0;
5967 fe
->i_last_eb_blk
= 0;
5969 fe
->i_last_eb_blk
= last_eb
->h_blkno
;
5971 status
= ocfs2_journal_dirty(handle
, fe_bh
);
5978 /* If there will be a new last extent block, then by
5979 * definition, there cannot be any leaves to the right of
5981 last_eb
->h_next_leaf_blk
= 0;
5982 status
= ocfs2_journal_dirty(handle
, last_eb_bh
);
5990 status
= ocfs2_truncate_log_append(osb
, handle
, delete_blk
,
6004 static int ocfs2_writeback_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
6006 set_buffer_uptodate(bh
);
6007 mark_buffer_dirty(bh
);
6011 static int ocfs2_ordered_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
6013 set_buffer_uptodate(bh
);
6014 mark_buffer_dirty(bh
);
6015 return ocfs2_journal_dirty_data(handle
, bh
);
6018 static void ocfs2_map_and_dirty_page(struct inode
*inode
, handle_t
*handle
,
6019 unsigned int from
, unsigned int to
,
6020 struct page
*page
, int zero
, u64
*phys
)
6022 int ret
, partial
= 0;
6024 ret
= ocfs2_map_page_blocks(page
, phys
, inode
, from
, to
, 0);
6029 zero_user_segment(page
, from
, to
);
6032 * Need to set the buffers we zero'd into uptodate
6033 * here if they aren't - ocfs2_map_page_blocks()
6034 * might've skipped some
6036 if (ocfs2_should_order_data(inode
)) {
6037 ret
= walk_page_buffers(handle
,
6040 ocfs2_ordered_zero_func
);
6044 ret
= walk_page_buffers(handle
, page_buffers(page
),
6046 ocfs2_writeback_zero_func
);
6052 SetPageUptodate(page
);
6054 flush_dcache_page(page
);
6057 static void ocfs2_zero_cluster_pages(struct inode
*inode
, loff_t start
,
6058 loff_t end
, struct page
**pages
,
6059 int numpages
, u64 phys
, handle_t
*handle
)
6063 unsigned int from
, to
= PAGE_CACHE_SIZE
;
6064 struct super_block
*sb
= inode
->i_sb
;
6066 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb
)));
6071 to
= PAGE_CACHE_SIZE
;
6072 for(i
= 0; i
< numpages
; i
++) {
6075 from
= start
& (PAGE_CACHE_SIZE
- 1);
6076 if ((end
>> PAGE_CACHE_SHIFT
) == page
->index
)
6077 to
= end
& (PAGE_CACHE_SIZE
- 1);
6079 BUG_ON(from
> PAGE_CACHE_SIZE
);
6080 BUG_ON(to
> PAGE_CACHE_SIZE
);
6082 ocfs2_map_and_dirty_page(inode
, handle
, from
, to
, page
, 1,
6085 start
= (page
->index
+ 1) << PAGE_CACHE_SHIFT
;
6089 ocfs2_unlock_and_free_pages(pages
, numpages
);
6092 static int ocfs2_grab_eof_pages(struct inode
*inode
, loff_t start
, loff_t end
,
6093 struct page
**pages
, int *num
)
6095 int numpages
, ret
= 0;
6096 struct super_block
*sb
= inode
->i_sb
;
6097 struct address_space
*mapping
= inode
->i_mapping
;
6098 unsigned long index
;
6099 loff_t last_page_bytes
;
6101 BUG_ON(start
> end
);
6103 BUG_ON(start
>> OCFS2_SB(sb
)->s_clustersize_bits
!=
6104 (end
- 1) >> OCFS2_SB(sb
)->s_clustersize_bits
);
6107 last_page_bytes
= PAGE_ALIGN(end
);
6108 index
= start
>> PAGE_CACHE_SHIFT
;
6110 pages
[numpages
] = grab_cache_page(mapping
, index
);
6111 if (!pages
[numpages
]) {
6119 } while (index
< (last_page_bytes
>> PAGE_CACHE_SHIFT
));
6124 ocfs2_unlock_and_free_pages(pages
, numpages
);
6134 * Zero the area past i_size but still within an allocated
6135 * cluster. This avoids exposing nonzero data on subsequent file
6138 * We need to call this before i_size is updated on the inode because
6139 * otherwise block_write_full_page() will skip writeout of pages past
6140 * i_size. The new_i_size parameter is passed for this reason.
6142 int ocfs2_zero_range_for_truncate(struct inode
*inode
, handle_t
*handle
,
6143 u64 range_start
, u64 range_end
)
6145 int ret
= 0, numpages
;
6146 struct page
**pages
= NULL
;
6148 unsigned int ext_flags
;
6149 struct super_block
*sb
= inode
->i_sb
;
6152 * File systems which don't support sparse files zero on every
6155 if (!ocfs2_sparse_alloc(OCFS2_SB(sb
)))
6158 pages
= kcalloc(ocfs2_pages_per_cluster(sb
),
6159 sizeof(struct page
*), GFP_NOFS
);
6160 if (pages
== NULL
) {
6166 if (range_start
== range_end
)
6169 ret
= ocfs2_extent_map_get_blocks(inode
,
6170 range_start
>> sb
->s_blocksize_bits
,
6171 &phys
, NULL
, &ext_flags
);
6178 * Tail is a hole, or is marked unwritten. In either case, we
6179 * can count on read and write to return/push zero's.
6181 if (phys
== 0 || ext_flags
& OCFS2_EXT_UNWRITTEN
)
6184 ret
= ocfs2_grab_eof_pages(inode
, range_start
, range_end
, pages
,
6191 ocfs2_zero_cluster_pages(inode
, range_start
, range_end
, pages
,
6192 numpages
, phys
, handle
);
6195 * Initiate writeout of the pages we zero'd here. We don't
6196 * wait on them - the truncate_inode_pages() call later will
6199 ret
= do_sync_mapping_range(inode
->i_mapping
, range_start
,
6200 range_end
- 1, SYNC_FILE_RANGE_WRITE
);
6211 static void ocfs2_zero_dinode_id2(struct inode
*inode
, struct ocfs2_dinode
*di
)
6213 unsigned int blocksize
= 1 << inode
->i_sb
->s_blocksize_bits
;
6215 memset(&di
->id2
, 0, blocksize
- offsetof(struct ocfs2_dinode
, id2
));
6218 void ocfs2_dinode_new_extent_list(struct inode
*inode
,
6219 struct ocfs2_dinode
*di
)
6221 ocfs2_zero_dinode_id2(inode
, di
);
6222 di
->id2
.i_list
.l_tree_depth
= 0;
6223 di
->id2
.i_list
.l_next_free_rec
= 0;
6224 di
->id2
.i_list
.l_count
= cpu_to_le16(ocfs2_extent_recs_per_inode(inode
->i_sb
));
6227 void ocfs2_set_inode_data_inline(struct inode
*inode
, struct ocfs2_dinode
*di
)
6229 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
6230 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
6232 spin_lock(&oi
->ip_lock
);
6233 oi
->ip_dyn_features
|= OCFS2_INLINE_DATA_FL
;
6234 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
6235 spin_unlock(&oi
->ip_lock
);
6238 * We clear the entire i_data structure here so that all
6239 * fields can be properly initialized.
6241 ocfs2_zero_dinode_id2(inode
, di
);
6243 idata
->id_count
= cpu_to_le16(ocfs2_max_inline_data(inode
->i_sb
));
6246 int ocfs2_convert_inline_data_to_extents(struct inode
*inode
,
6247 struct buffer_head
*di_bh
)
6249 int ret
, i
, has_data
, num_pages
= 0;
6251 u64
uninitialized_var(block
);
6252 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
6253 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
6254 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
6255 struct ocfs2_alloc_context
*data_ac
= NULL
;
6256 struct page
**pages
= NULL
;
6257 loff_t end
= osb
->s_clustersize
;
6259 has_data
= i_size_read(inode
) ? 1 : 0;
6262 pages
= kcalloc(ocfs2_pages_per_cluster(osb
->sb
),
6263 sizeof(struct page
*), GFP_NOFS
);
6264 if (pages
== NULL
) {
6270 ret
= ocfs2_reserve_clusters(osb
, 1, &data_ac
);
6277 handle
= ocfs2_start_trans(osb
, OCFS2_INLINE_TO_EXTENTS_CREDITS
);
6278 if (IS_ERR(handle
)) {
6279 ret
= PTR_ERR(handle
);
6284 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
6285 OCFS2_JOURNAL_ACCESS_WRITE
);
6293 unsigned int page_end
;
6296 ret
= ocfs2_claim_clusters(osb
, handle
, data_ac
, 1, &bit_off
,
6304 * Save two copies, one for insert, and one that can
6305 * be changed by ocfs2_map_and_dirty_page() below.
6307 block
= phys
= ocfs2_clusters_to_blocks(inode
->i_sb
, bit_off
);
6310 * Non sparse file systems zero on extend, so no need
6313 if (!ocfs2_sparse_alloc(osb
) &&
6314 PAGE_CACHE_SIZE
< osb
->s_clustersize
)
6315 end
= PAGE_CACHE_SIZE
;
6317 ret
= ocfs2_grab_eof_pages(inode
, 0, end
, pages
, &num_pages
);
6324 * This should populate the 1st page for us and mark
6327 ret
= ocfs2_read_inline_data(inode
, pages
[0], di_bh
);
6333 page_end
= PAGE_CACHE_SIZE
;
6334 if (PAGE_CACHE_SIZE
> osb
->s_clustersize
)
6335 page_end
= osb
->s_clustersize
;
6337 for (i
= 0; i
< num_pages
; i
++)
6338 ocfs2_map_and_dirty_page(inode
, handle
, 0, page_end
,
6339 pages
[i
], i
> 0, &phys
);
6342 spin_lock(&oi
->ip_lock
);
6343 oi
->ip_dyn_features
&= ~OCFS2_INLINE_DATA_FL
;
6344 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
6345 spin_unlock(&oi
->ip_lock
);
6347 ocfs2_dinode_new_extent_list(inode
, di
);
6349 ocfs2_journal_dirty(handle
, di_bh
);
6353 * An error at this point should be extremely rare. If
6354 * this proves to be false, we could always re-build
6355 * the in-inode data from our pages.
6357 ret
= ocfs2_insert_extent(osb
, handle
, inode
, di_bh
,
6358 0, block
, 1, 0, NULL
);
6364 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
6368 ocfs2_commit_trans(osb
, handle
);
6372 ocfs2_free_alloc_context(data_ac
);
6376 ocfs2_unlock_and_free_pages(pages
, num_pages
);
6384 * It is expected, that by the time you call this function,
6385 * inode->i_size and fe->i_size have been adjusted.
6387 * WARNING: This will kfree the truncate context
6389 int ocfs2_commit_truncate(struct ocfs2_super
*osb
,
6390 struct inode
*inode
,
6391 struct buffer_head
*fe_bh
,
6392 struct ocfs2_truncate_context
*tc
)
6394 int status
, i
, credits
, tl_sem
= 0;
6395 u32 clusters_to_del
, new_highest_cpos
, range
;
6396 struct ocfs2_extent_list
*el
;
6397 handle_t
*handle
= NULL
;
6398 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6399 struct ocfs2_path
*path
= NULL
;
6403 new_highest_cpos
= ocfs2_clusters_for_bytes(osb
->sb
,
6404 i_size_read(inode
));
6406 path
= ocfs2_new_inode_path(fe_bh
);
6413 ocfs2_extent_map_trunc(inode
, new_highest_cpos
);
6417 * Check that we still have allocation to delete.
6419 if (OCFS2_I(inode
)->ip_clusters
== 0) {
6425 * Truncate always works against the rightmost tree branch.
6427 status
= ocfs2_find_path(inode
, path
, UINT_MAX
);
6433 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
6434 OCFS2_I(inode
)->ip_clusters
, path
->p_tree_depth
);
6437 * By now, el will point to the extent list on the bottom most
6438 * portion of this tree. Only the tail record is considered in
6441 * We handle the following cases, in order:
6442 * - empty extent: delete the remaining branch
6443 * - remove the entire record
6444 * - remove a partial record
6445 * - no record needs to be removed (truncate has completed)
6447 el
= path_leaf_el(path
);
6448 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
6449 ocfs2_error(inode
->i_sb
,
6450 "Inode %llu has empty extent block at %llu\n",
6451 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
6452 (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
6457 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
6458 range
= le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
6459 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
6460 if (i
== 0 && ocfs2_is_empty_extent(&el
->l_recs
[i
])) {
6461 clusters_to_del
= 0;
6462 } else if (le32_to_cpu(el
->l_recs
[i
].e_cpos
) >= new_highest_cpos
) {
6463 clusters_to_del
= ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
6464 } else if (range
> new_highest_cpos
) {
6465 clusters_to_del
= (ocfs2_rec_clusters(el
, &el
->l_recs
[i
]) +
6466 le32_to_cpu(el
->l_recs
[i
].e_cpos
)) -
6473 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
6474 clusters_to_del
, (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
6476 mutex_lock(&tl_inode
->i_mutex
);
6478 /* ocfs2_truncate_log_needs_flush guarantees us at least one
6479 * record is free for use. If there isn't any, we flush to get
6480 * an empty truncate log. */
6481 if (ocfs2_truncate_log_needs_flush(osb
)) {
6482 status
= __ocfs2_flush_truncate_log(osb
);
6489 credits
= ocfs2_calc_tree_trunc_credits(osb
->sb
, clusters_to_del
,
6490 (struct ocfs2_dinode
*)fe_bh
->b_data
,
6492 handle
= ocfs2_start_trans(osb
, credits
);
6493 if (IS_ERR(handle
)) {
6494 status
= PTR_ERR(handle
);
6500 status
= ocfs2_do_truncate(osb
, clusters_to_del
, inode
, fe_bh
, handle
,
6507 mutex_unlock(&tl_inode
->i_mutex
);
6510 ocfs2_commit_trans(osb
, handle
);
6513 ocfs2_reinit_path(path
, 1);
6516 * The check above will catch the case where we've truncated
6517 * away all allocation.
6523 ocfs2_schedule_truncate_log_flush(osb
, 1);
6526 mutex_unlock(&tl_inode
->i_mutex
);
6529 ocfs2_commit_trans(osb
, handle
);
6531 ocfs2_run_deallocs(osb
, &tc
->tc_dealloc
);
6533 ocfs2_free_path(path
);
6535 /* This will drop the ext_alloc cluster lock for us */
6536 ocfs2_free_truncate_context(tc
);
6543 * Expects the inode to already be locked.
6545 int ocfs2_prepare_truncate(struct ocfs2_super
*osb
,
6546 struct inode
*inode
,
6547 struct buffer_head
*fe_bh
,
6548 struct ocfs2_truncate_context
**tc
)
6551 unsigned int new_i_clusters
;
6552 struct ocfs2_dinode
*fe
;
6553 struct ocfs2_extent_block
*eb
;
6554 struct buffer_head
*last_eb_bh
= NULL
;
6560 new_i_clusters
= ocfs2_clusters_for_bytes(osb
->sb
,
6561 i_size_read(inode
));
6562 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
6564 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
6565 "%llu\n", le32_to_cpu(fe
->i_clusters
), new_i_clusters
,
6566 (unsigned long long)le64_to_cpu(fe
->i_size
));
6568 *tc
= kzalloc(sizeof(struct ocfs2_truncate_context
), GFP_KERNEL
);
6574 ocfs2_init_dealloc_ctxt(&(*tc
)->tc_dealloc
);
6576 if (fe
->id2
.i_list
.l_tree_depth
) {
6577 status
= ocfs2_read_block(osb
, le64_to_cpu(fe
->i_last_eb_blk
),
6578 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
6583 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
6584 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
6585 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
6593 (*tc
)->tc_last_eb_bh
= last_eb_bh
;
6599 ocfs2_free_truncate_context(*tc
);
6607 * 'start' is inclusive, 'end' is not.
6609 int ocfs2_truncate_inline(struct inode
*inode
, struct buffer_head
*di_bh
,
6610 unsigned int start
, unsigned int end
, int trunc
)
6613 unsigned int numbytes
;
6615 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
6616 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
6617 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
6619 if (end
> i_size_read(inode
))
6620 end
= i_size_read(inode
);
6622 BUG_ON(start
>= end
);
6624 if (!(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
) ||
6625 !(le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_DATA_FL
) ||
6626 !ocfs2_supports_inline_data(osb
)) {
6627 ocfs2_error(inode
->i_sb
,
6628 "Inline data flags for inode %llu don't agree! "
6629 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
6630 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
6631 le16_to_cpu(di
->i_dyn_features
),
6632 OCFS2_I(inode
)->ip_dyn_features
,
6633 osb
->s_feature_incompat
);
6638 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
6639 if (IS_ERR(handle
)) {
6640 ret
= PTR_ERR(handle
);
6645 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
6646 OCFS2_JOURNAL_ACCESS_WRITE
);
6652 numbytes
= end
- start
;
6653 memset(idata
->id_data
+ start
, 0, numbytes
);
6656 * No need to worry about the data page here - it's been
6657 * truncated already and inline data doesn't need it for
6658 * pushing zero's to disk, so we'll let readpage pick it up
6662 i_size_write(inode
, start
);
6663 di
->i_size
= cpu_to_le64(start
);
6666 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
6667 inode
->i_ctime
= inode
->i_mtime
= CURRENT_TIME
;
6669 di
->i_ctime
= di
->i_mtime
= cpu_to_le64(inode
->i_ctime
.tv_sec
);
6670 di
->i_ctime_nsec
= di
->i_mtime_nsec
= cpu_to_le32(inode
->i_ctime
.tv_nsec
);
6672 ocfs2_journal_dirty(handle
, di_bh
);
6675 ocfs2_commit_trans(osb
, handle
);
6681 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
)
6684 * The caller is responsible for completing deallocation
6685 * before freeing the context.
6687 if (tc
->tc_dealloc
.c_first_suballocator
!= NULL
)
6689 "Truncate completion has non-empty dealloc context\n");
6691 if (tc
->tc_last_eb_bh
)
6692 brelse(tc
->tc_last_eb_bh
);