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
;
357 int ins_free_records
;
361 struct ocfs2_merge_ctxt
{
362 enum ocfs2_contig_type c_contig_type
;
363 int c_has_empty_extent
;
364 int c_split_covers_rec
;
365 int c_used_tail_recs
;
369 * How many free extents have we got before we need more meta data?
371 int ocfs2_num_free_extents(struct ocfs2_super
*osb
,
373 struct ocfs2_dinode
*fe
)
376 struct ocfs2_extent_list
*el
;
377 struct ocfs2_extent_block
*eb
;
378 struct buffer_head
*eb_bh
= NULL
;
382 if (!OCFS2_IS_VALID_DINODE(fe
)) {
383 OCFS2_RO_ON_INVALID_DINODE(inode
->i_sb
, fe
);
388 if (fe
->i_last_eb_blk
) {
389 retval
= ocfs2_read_block(osb
, le64_to_cpu(fe
->i_last_eb_blk
),
390 &eb_bh
, OCFS2_BH_CACHED
, inode
);
395 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
398 el
= &fe
->id2
.i_list
;
400 BUG_ON(el
->l_tree_depth
!= 0);
402 retval
= le16_to_cpu(el
->l_count
) - le16_to_cpu(el
->l_next_free_rec
);
411 /* expects array to already be allocated
413 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
416 static int ocfs2_create_new_meta_bhs(struct ocfs2_super
*osb
,
420 struct ocfs2_alloc_context
*meta_ac
,
421 struct buffer_head
*bhs
[])
423 int count
, status
, i
;
424 u16 suballoc_bit_start
;
427 struct ocfs2_extent_block
*eb
;
432 while (count
< wanted
) {
433 status
= ocfs2_claim_metadata(osb
,
445 for(i
= count
; i
< (num_got
+ count
); i
++) {
446 bhs
[i
] = sb_getblk(osb
->sb
, first_blkno
);
447 if (bhs
[i
] == NULL
) {
452 ocfs2_set_new_buffer_uptodate(inode
, bhs
[i
]);
454 status
= ocfs2_journal_access(handle
, inode
, bhs
[i
],
455 OCFS2_JOURNAL_ACCESS_CREATE
);
461 memset(bhs
[i
]->b_data
, 0, osb
->sb
->s_blocksize
);
462 eb
= (struct ocfs2_extent_block
*) bhs
[i
]->b_data
;
463 /* Ok, setup the minimal stuff here. */
464 strcpy(eb
->h_signature
, OCFS2_EXTENT_BLOCK_SIGNATURE
);
465 eb
->h_blkno
= cpu_to_le64(first_blkno
);
466 eb
->h_fs_generation
= cpu_to_le32(osb
->fs_generation
);
467 eb
->h_suballoc_slot
= cpu_to_le16(osb
->slot_num
);
468 eb
->h_suballoc_bit
= cpu_to_le16(suballoc_bit_start
);
470 cpu_to_le16(ocfs2_extent_recs_per_eb(osb
->sb
));
472 suballoc_bit_start
++;
475 /* We'll also be dirtied by the caller, so
476 * this isn't absolutely necessary. */
477 status
= ocfs2_journal_dirty(handle
, bhs
[i
]);
490 for(i
= 0; i
< wanted
; i
++) {
501 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
503 * Returns the sum of the rightmost extent rec logical offset and
506 * ocfs2_add_branch() uses this to determine what logical cluster
507 * value should be populated into the leftmost new branch records.
509 * ocfs2_shift_tree_depth() uses this to determine the # clusters
510 * value for the new topmost tree record.
512 static inline u32
ocfs2_sum_rightmost_rec(struct ocfs2_extent_list
*el
)
516 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
518 return le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
519 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
523 * Add an entire tree branch to our inode. eb_bh is the extent block
524 * to start at, if we don't want to start the branch at the dinode
527 * last_eb_bh is required as we have to update it's next_leaf pointer
528 * for the new last extent block.
530 * the new branch will be 'empty' in the sense that every block will
531 * contain a single record with cluster count == 0.
533 static int ocfs2_add_branch(struct ocfs2_super
*osb
,
536 struct buffer_head
*fe_bh
,
537 struct buffer_head
*eb_bh
,
538 struct buffer_head
**last_eb_bh
,
539 struct ocfs2_alloc_context
*meta_ac
)
541 int status
, new_blocks
, i
;
542 u64 next_blkno
, new_last_eb_blk
;
543 struct buffer_head
*bh
;
544 struct buffer_head
**new_eb_bhs
= NULL
;
545 struct ocfs2_dinode
*fe
;
546 struct ocfs2_extent_block
*eb
;
547 struct ocfs2_extent_list
*eb_el
;
548 struct ocfs2_extent_list
*el
;
553 BUG_ON(!last_eb_bh
|| !*last_eb_bh
);
555 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
558 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
561 el
= &fe
->id2
.i_list
;
563 /* we never add a branch to a leaf. */
564 BUG_ON(!el
->l_tree_depth
);
566 new_blocks
= le16_to_cpu(el
->l_tree_depth
);
568 /* allocate the number of new eb blocks we need */
569 new_eb_bhs
= kcalloc(new_blocks
, sizeof(struct buffer_head
*),
577 status
= ocfs2_create_new_meta_bhs(osb
, handle
, inode
, new_blocks
,
578 meta_ac
, new_eb_bhs
);
584 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
585 new_cpos
= ocfs2_sum_rightmost_rec(&eb
->h_list
);
587 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
588 * linked with the rest of the tree.
589 * conversly, new_eb_bhs[0] is the new bottommost leaf.
591 * when we leave the loop, new_last_eb_blk will point to the
592 * newest leaf, and next_blkno will point to the topmost extent
594 next_blkno
= new_last_eb_blk
= 0;
595 for(i
= 0; i
< new_blocks
; i
++) {
597 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
598 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
599 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
605 status
= ocfs2_journal_access(handle
, inode
, bh
,
606 OCFS2_JOURNAL_ACCESS_CREATE
);
612 eb
->h_next_leaf_blk
= 0;
613 eb_el
->l_tree_depth
= cpu_to_le16(i
);
614 eb_el
->l_next_free_rec
= cpu_to_le16(1);
616 * This actually counts as an empty extent as
619 eb_el
->l_recs
[0].e_cpos
= cpu_to_le32(new_cpos
);
620 eb_el
->l_recs
[0].e_blkno
= cpu_to_le64(next_blkno
);
622 * eb_el isn't always an interior node, but even leaf
623 * nodes want a zero'd flags and reserved field so
624 * this gets the whole 32 bits regardless of use.
626 eb_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(0);
627 if (!eb_el
->l_tree_depth
)
628 new_last_eb_blk
= le64_to_cpu(eb
->h_blkno
);
630 status
= ocfs2_journal_dirty(handle
, bh
);
636 next_blkno
= le64_to_cpu(eb
->h_blkno
);
639 /* This is a bit hairy. We want to update up to three blocks
640 * here without leaving any of them in an inconsistent state
641 * in case of error. We don't have to worry about
642 * journal_dirty erroring as it won't unless we've aborted the
643 * handle (in which case we would never be here) so reserving
644 * the write with journal_access is all we need to do. */
645 status
= ocfs2_journal_access(handle
, inode
, *last_eb_bh
,
646 OCFS2_JOURNAL_ACCESS_WRITE
);
651 status
= ocfs2_journal_access(handle
, inode
, fe_bh
,
652 OCFS2_JOURNAL_ACCESS_WRITE
);
658 status
= ocfs2_journal_access(handle
, inode
, eb_bh
,
659 OCFS2_JOURNAL_ACCESS_WRITE
);
666 /* Link the new branch into the rest of the tree (el will
667 * either be on the fe, or the extent block passed in. */
668 i
= le16_to_cpu(el
->l_next_free_rec
);
669 el
->l_recs
[i
].e_blkno
= cpu_to_le64(next_blkno
);
670 el
->l_recs
[i
].e_cpos
= cpu_to_le32(new_cpos
);
671 el
->l_recs
[i
].e_int_clusters
= 0;
672 le16_add_cpu(&el
->l_next_free_rec
, 1);
674 /* fe needs a new last extent block pointer, as does the
675 * next_leaf on the previously last-extent-block. */
676 fe
->i_last_eb_blk
= cpu_to_le64(new_last_eb_blk
);
678 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
679 eb
->h_next_leaf_blk
= cpu_to_le64(new_last_eb_blk
);
681 status
= ocfs2_journal_dirty(handle
, *last_eb_bh
);
684 status
= ocfs2_journal_dirty(handle
, fe_bh
);
688 status
= ocfs2_journal_dirty(handle
, eb_bh
);
694 * Some callers want to track the rightmost leaf so pass it
698 get_bh(new_eb_bhs
[0]);
699 *last_eb_bh
= new_eb_bhs
[0];
704 for (i
= 0; i
< new_blocks
; i
++)
706 brelse(new_eb_bhs
[i
]);
715 * adds another level to the allocation tree.
716 * returns back the new extent block so you can add a branch to it
719 static int ocfs2_shift_tree_depth(struct ocfs2_super
*osb
,
722 struct buffer_head
*fe_bh
,
723 struct ocfs2_alloc_context
*meta_ac
,
724 struct buffer_head
**ret_new_eb_bh
)
728 struct buffer_head
*new_eb_bh
= NULL
;
729 struct ocfs2_dinode
*fe
;
730 struct ocfs2_extent_block
*eb
;
731 struct ocfs2_extent_list
*fe_el
;
732 struct ocfs2_extent_list
*eb_el
;
736 status
= ocfs2_create_new_meta_bhs(osb
, handle
, inode
, 1, meta_ac
,
743 eb
= (struct ocfs2_extent_block
*) new_eb_bh
->b_data
;
744 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
745 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
751 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
752 fe_el
= &fe
->id2
.i_list
;
754 status
= ocfs2_journal_access(handle
, inode
, new_eb_bh
,
755 OCFS2_JOURNAL_ACCESS_CREATE
);
761 /* copy the fe data into the new extent block */
762 eb_el
->l_tree_depth
= fe_el
->l_tree_depth
;
763 eb_el
->l_next_free_rec
= fe_el
->l_next_free_rec
;
764 for(i
= 0; i
< le16_to_cpu(fe_el
->l_next_free_rec
); i
++)
765 eb_el
->l_recs
[i
] = fe_el
->l_recs
[i
];
767 status
= ocfs2_journal_dirty(handle
, new_eb_bh
);
773 status
= ocfs2_journal_access(handle
, inode
, fe_bh
,
774 OCFS2_JOURNAL_ACCESS_WRITE
);
780 new_clusters
= ocfs2_sum_rightmost_rec(eb_el
);
783 le16_add_cpu(&fe_el
->l_tree_depth
, 1);
784 fe_el
->l_recs
[0].e_cpos
= 0;
785 fe_el
->l_recs
[0].e_blkno
= eb
->h_blkno
;
786 fe_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(new_clusters
);
787 for(i
= 1; i
< le16_to_cpu(fe_el
->l_next_free_rec
); i
++)
788 memset(&fe_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
789 fe_el
->l_next_free_rec
= cpu_to_le16(1);
791 /* If this is our 1st tree depth shift, then last_eb_blk
792 * becomes the allocated extent block */
793 if (fe_el
->l_tree_depth
== cpu_to_le16(1))
794 fe
->i_last_eb_blk
= eb
->h_blkno
;
796 status
= ocfs2_journal_dirty(handle
, fe_bh
);
802 *ret_new_eb_bh
= new_eb_bh
;
814 * Should only be called when there is no space left in any of the
815 * leaf nodes. What we want to do is find the lowest tree depth
816 * non-leaf extent block with room for new records. There are three
817 * valid results of this search:
819 * 1) a lowest extent block is found, then we pass it back in
820 * *lowest_eb_bh and return '0'
822 * 2) the search fails to find anything, but the dinode has room. We
823 * pass NULL back in *lowest_eb_bh, but still return '0'
825 * 3) the search fails to find anything AND the dinode is full, in
826 * which case we return > 0
828 * return status < 0 indicates an error.
830 static int ocfs2_find_branch_target(struct ocfs2_super
*osb
,
832 struct buffer_head
*fe_bh
,
833 struct buffer_head
**target_bh
)
837 struct ocfs2_dinode
*fe
;
838 struct ocfs2_extent_block
*eb
;
839 struct ocfs2_extent_list
*el
;
840 struct buffer_head
*bh
= NULL
;
841 struct buffer_head
*lowest_bh
= NULL
;
847 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
848 el
= &fe
->id2
.i_list
;
850 while(le16_to_cpu(el
->l_tree_depth
) > 1) {
851 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
852 ocfs2_error(inode
->i_sb
, "Dinode %llu has empty "
853 "extent list (next_free_rec == 0)",
854 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
858 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
859 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
861 ocfs2_error(inode
->i_sb
, "Dinode %llu has extent "
862 "list where extent # %d has no physical "
864 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, i
);
874 status
= ocfs2_read_block(osb
, blkno
, &bh
, OCFS2_BH_CACHED
,
881 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
882 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
883 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
889 if (le16_to_cpu(el
->l_next_free_rec
) <
890 le16_to_cpu(el
->l_count
)) {
898 /* If we didn't find one and the fe doesn't have any room,
901 && (fe
->id2
.i_list
.l_next_free_rec
== fe
->id2
.i_list
.l_count
))
904 *target_bh
= lowest_bh
;
914 * Grow a b-tree so that it has more records.
916 * We might shift the tree depth in which case existing paths should
917 * be considered invalid.
919 * Tree depth after the grow is returned via *final_depth.
921 * *last_eb_bh will be updated by ocfs2_add_branch().
923 static int ocfs2_grow_tree(struct inode
*inode
, handle_t
*handle
,
924 struct buffer_head
*di_bh
, int *final_depth
,
925 struct buffer_head
**last_eb_bh
,
926 struct ocfs2_alloc_context
*meta_ac
)
929 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
930 int depth
= le16_to_cpu(di
->id2
.i_list
.l_tree_depth
);
931 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
932 struct buffer_head
*bh
= NULL
;
934 BUG_ON(meta_ac
== NULL
);
936 shift
= ocfs2_find_branch_target(osb
, inode
, di_bh
, &bh
);
943 /* We traveled all the way to the bottom of the allocation tree
944 * and didn't find room for any more extents - we need to add
945 * another tree level */
948 mlog(0, "need to shift tree depth (current = %d)\n", depth
);
950 /* ocfs2_shift_tree_depth will return us a buffer with
951 * the new extent block (so we can pass that to
952 * ocfs2_add_branch). */
953 ret
= ocfs2_shift_tree_depth(osb
, handle
, inode
, di_bh
,
962 * Special case: we have room now if we shifted from
963 * tree_depth 0, so no more work needs to be done.
965 * We won't be calling add_branch, so pass
966 * back *last_eb_bh as the new leaf. At depth
967 * zero, it should always be null so there's
968 * no reason to brelse.
977 /* call ocfs2_add_branch to add the final part of the tree with
979 mlog(0, "add branch. bh = %p\n", bh
);
980 ret
= ocfs2_add_branch(osb
, handle
, inode
, di_bh
, bh
, last_eb_bh
,
989 *final_depth
= depth
;
995 * This is only valid for leaf nodes, which are the only ones that can
996 * have empty extents anyway.
998 static inline int ocfs2_is_empty_extent(struct ocfs2_extent_rec
*rec
)
1000 return !rec
->e_leaf_clusters
;
1004 * This function will discard the rightmost extent record.
1006 static void ocfs2_shift_records_right(struct ocfs2_extent_list
*el
)
1008 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1009 int count
= le16_to_cpu(el
->l_count
);
1010 unsigned int num_bytes
;
1013 /* This will cause us to go off the end of our extent list. */
1014 BUG_ON(next_free
>= count
);
1016 num_bytes
= sizeof(struct ocfs2_extent_rec
) * next_free
;
1018 memmove(&el
->l_recs
[1], &el
->l_recs
[0], num_bytes
);
1021 static void ocfs2_rotate_leaf(struct ocfs2_extent_list
*el
,
1022 struct ocfs2_extent_rec
*insert_rec
)
1024 int i
, insert_index
, next_free
, has_empty
, num_bytes
;
1025 u32 insert_cpos
= le32_to_cpu(insert_rec
->e_cpos
);
1026 struct ocfs2_extent_rec
*rec
;
1028 next_free
= le16_to_cpu(el
->l_next_free_rec
);
1029 has_empty
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
1033 /* The tree code before us didn't allow enough room in the leaf. */
1034 if (el
->l_next_free_rec
== el
->l_count
&& !has_empty
)
1038 * The easiest way to approach this is to just remove the
1039 * empty extent and temporarily decrement next_free.
1043 * If next_free was 1 (only an empty extent), this
1044 * loop won't execute, which is fine. We still want
1045 * the decrement above to happen.
1047 for(i
= 0; i
< (next_free
- 1); i
++)
1048 el
->l_recs
[i
] = el
->l_recs
[i
+1];
1054 * Figure out what the new record index should be.
1056 for(i
= 0; i
< next_free
; i
++) {
1057 rec
= &el
->l_recs
[i
];
1059 if (insert_cpos
< le32_to_cpu(rec
->e_cpos
))
1064 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1065 insert_cpos
, insert_index
, has_empty
, next_free
, le16_to_cpu(el
->l_count
));
1067 BUG_ON(insert_index
< 0);
1068 BUG_ON(insert_index
>= le16_to_cpu(el
->l_count
));
1069 BUG_ON(insert_index
> next_free
);
1072 * No need to memmove if we're just adding to the tail.
1074 if (insert_index
!= next_free
) {
1075 BUG_ON(next_free
>= le16_to_cpu(el
->l_count
));
1077 num_bytes
= next_free
- insert_index
;
1078 num_bytes
*= sizeof(struct ocfs2_extent_rec
);
1079 memmove(&el
->l_recs
[insert_index
+ 1],
1080 &el
->l_recs
[insert_index
],
1085 * Either we had an empty extent, and need to re-increment or
1086 * there was no empty extent on a non full rightmost leaf node,
1087 * in which case we still need to increment.
1090 el
->l_next_free_rec
= cpu_to_le16(next_free
);
1092 * Make sure none of the math above just messed up our tree.
1094 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) > le16_to_cpu(el
->l_count
));
1096 el
->l_recs
[insert_index
] = *insert_rec
;
1100 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list
*el
)
1102 int size
, num_recs
= le16_to_cpu(el
->l_next_free_rec
);
1104 BUG_ON(num_recs
== 0);
1106 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
1108 size
= num_recs
* sizeof(struct ocfs2_extent_rec
);
1109 memmove(&el
->l_recs
[0], &el
->l_recs
[1], size
);
1110 memset(&el
->l_recs
[num_recs
], 0,
1111 sizeof(struct ocfs2_extent_rec
));
1112 el
->l_next_free_rec
= cpu_to_le16(num_recs
);
1117 * Create an empty extent record .
1119 * l_next_free_rec may be updated.
1121 * If an empty extent already exists do nothing.
1123 static void ocfs2_create_empty_extent(struct ocfs2_extent_list
*el
)
1125 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1127 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
1132 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
1135 mlog_bug_on_msg(el
->l_count
== el
->l_next_free_rec
,
1136 "Asked to create an empty extent in a full list:\n"
1137 "count = %u, tree depth = %u",
1138 le16_to_cpu(el
->l_count
),
1139 le16_to_cpu(el
->l_tree_depth
));
1141 ocfs2_shift_records_right(el
);
1144 le16_add_cpu(&el
->l_next_free_rec
, 1);
1145 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1149 * For a rotation which involves two leaf nodes, the "root node" is
1150 * the lowest level tree node which contains a path to both leafs. This
1151 * resulting set of information can be used to form a complete "subtree"
1153 * This function is passed two full paths from the dinode down to a
1154 * pair of adjacent leaves. It's task is to figure out which path
1155 * index contains the subtree root - this can be the root index itself
1156 * in a worst-case rotation.
1158 * The array index of the subtree root is passed back.
1160 static int ocfs2_find_subtree_root(struct inode
*inode
,
1161 struct ocfs2_path
*left
,
1162 struct ocfs2_path
*right
)
1167 * Check that the caller passed in two paths from the same tree.
1169 BUG_ON(path_root_bh(left
) != path_root_bh(right
));
1175 * The caller didn't pass two adjacent paths.
1177 mlog_bug_on_msg(i
> left
->p_tree_depth
,
1178 "Inode %lu, left depth %u, right depth %u\n"
1179 "left leaf blk %llu, right leaf blk %llu\n",
1180 inode
->i_ino
, left
->p_tree_depth
,
1181 right
->p_tree_depth
,
1182 (unsigned long long)path_leaf_bh(left
)->b_blocknr
,
1183 (unsigned long long)path_leaf_bh(right
)->b_blocknr
);
1184 } while (left
->p_node
[i
].bh
->b_blocknr
==
1185 right
->p_node
[i
].bh
->b_blocknr
);
1190 typedef void (path_insert_t
)(void *, struct buffer_head
*);
1193 * Traverse a btree path in search of cpos, starting at root_el.
1195 * This code can be called with a cpos larger than the tree, in which
1196 * case it will return the rightmost path.
1198 static int __ocfs2_find_path(struct inode
*inode
,
1199 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1200 path_insert_t
*func
, void *data
)
1205 struct buffer_head
*bh
= NULL
;
1206 struct ocfs2_extent_block
*eb
;
1207 struct ocfs2_extent_list
*el
;
1208 struct ocfs2_extent_rec
*rec
;
1209 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1212 while (el
->l_tree_depth
) {
1213 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1214 ocfs2_error(inode
->i_sb
,
1215 "Inode %llu has empty extent list at "
1217 (unsigned long long)oi
->ip_blkno
,
1218 le16_to_cpu(el
->l_tree_depth
));
1224 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
) - 1; i
++) {
1225 rec
= &el
->l_recs
[i
];
1228 * In the case that cpos is off the allocation
1229 * tree, this should just wind up returning the
1232 range
= le32_to_cpu(rec
->e_cpos
) +
1233 ocfs2_rec_clusters(el
, rec
);
1234 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1238 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1240 ocfs2_error(inode
->i_sb
,
1241 "Inode %llu has bad blkno in extent list "
1242 "at depth %u (index %d)\n",
1243 (unsigned long long)oi
->ip_blkno
,
1244 le16_to_cpu(el
->l_tree_depth
), i
);
1251 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
), blkno
,
1252 &bh
, OCFS2_BH_CACHED
, inode
);
1258 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1260 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
1261 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
1266 if (le16_to_cpu(el
->l_next_free_rec
) >
1267 le16_to_cpu(el
->l_count
)) {
1268 ocfs2_error(inode
->i_sb
,
1269 "Inode %llu has bad count in extent list "
1270 "at block %llu (next free=%u, count=%u)\n",
1271 (unsigned long long)oi
->ip_blkno
,
1272 (unsigned long long)bh
->b_blocknr
,
1273 le16_to_cpu(el
->l_next_free_rec
),
1274 le16_to_cpu(el
->l_count
));
1285 * Catch any trailing bh that the loop didn't handle.
1293 * Given an initialized path (that is, it has a valid root extent
1294 * list), this function will traverse the btree in search of the path
1295 * which would contain cpos.
1297 * The path traveled is recorded in the path structure.
1299 * Note that this will not do any comparisons on leaf node extent
1300 * records, so it will work fine in the case that we just added a tree
1303 struct find_path_data
{
1305 struct ocfs2_path
*path
;
1307 static void find_path_ins(void *data
, struct buffer_head
*bh
)
1309 struct find_path_data
*fp
= data
;
1312 ocfs2_path_insert_eb(fp
->path
, fp
->index
, bh
);
1315 static int ocfs2_find_path(struct inode
*inode
, struct ocfs2_path
*path
,
1318 struct find_path_data data
;
1322 return __ocfs2_find_path(inode
, path_root_el(path
), cpos
,
1323 find_path_ins
, &data
);
1326 static void find_leaf_ins(void *data
, struct buffer_head
*bh
)
1328 struct ocfs2_extent_block
*eb
=(struct ocfs2_extent_block
*)bh
->b_data
;
1329 struct ocfs2_extent_list
*el
= &eb
->h_list
;
1330 struct buffer_head
**ret
= data
;
1332 /* We want to retain only the leaf block. */
1333 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
1339 * Find the leaf block in the tree which would contain cpos. No
1340 * checking of the actual leaf is done.
1342 * Some paths want to call this instead of allocating a path structure
1343 * and calling ocfs2_find_path().
1345 * This function doesn't handle non btree extent lists.
1347 int ocfs2_find_leaf(struct inode
*inode
, struct ocfs2_extent_list
*root_el
,
1348 u32 cpos
, struct buffer_head
**leaf_bh
)
1351 struct buffer_head
*bh
= NULL
;
1353 ret
= __ocfs2_find_path(inode
, root_el
, cpos
, find_leaf_ins
, &bh
);
1365 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1367 * Basically, we've moved stuff around at the bottom of the tree and
1368 * we need to fix up the extent records above the changes to reflect
1371 * left_rec: the record on the left.
1372 * left_child_el: is the child list pointed to by left_rec
1373 * right_rec: the record to the right of left_rec
1374 * right_child_el: is the child list pointed to by right_rec
1376 * By definition, this only works on interior nodes.
1378 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec
*left_rec
,
1379 struct ocfs2_extent_list
*left_child_el
,
1380 struct ocfs2_extent_rec
*right_rec
,
1381 struct ocfs2_extent_list
*right_child_el
)
1383 u32 left_clusters
, right_end
;
1386 * Interior nodes never have holes. Their cpos is the cpos of
1387 * the leftmost record in their child list. Their cluster
1388 * count covers the full theoretical range of their child list
1389 * - the range between their cpos and the cpos of the record
1390 * immediately to their right.
1392 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[0].e_cpos
);
1393 if (ocfs2_is_empty_extent(&right_child_el
->l_recs
[0])) {
1394 BUG_ON(le16_to_cpu(right_child_el
->l_next_free_rec
) <= 1);
1395 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[1].e_cpos
);
1397 left_clusters
-= le32_to_cpu(left_rec
->e_cpos
);
1398 left_rec
->e_int_clusters
= cpu_to_le32(left_clusters
);
1401 * Calculate the rightmost cluster count boundary before
1402 * moving cpos - we will need to adjust clusters after
1403 * updating e_cpos to keep the same highest cluster count.
1405 right_end
= le32_to_cpu(right_rec
->e_cpos
);
1406 right_end
+= le32_to_cpu(right_rec
->e_int_clusters
);
1408 right_rec
->e_cpos
= left_rec
->e_cpos
;
1409 le32_add_cpu(&right_rec
->e_cpos
, left_clusters
);
1411 right_end
-= le32_to_cpu(right_rec
->e_cpos
);
1412 right_rec
->e_int_clusters
= cpu_to_le32(right_end
);
1416 * Adjust the adjacent root node records involved in a
1417 * rotation. left_el_blkno is passed in as a key so that we can easily
1418 * find it's index in the root list.
1420 static void ocfs2_adjust_root_records(struct ocfs2_extent_list
*root_el
,
1421 struct ocfs2_extent_list
*left_el
,
1422 struct ocfs2_extent_list
*right_el
,
1427 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) <=
1428 le16_to_cpu(left_el
->l_tree_depth
));
1430 for(i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
) - 1; i
++) {
1431 if (le64_to_cpu(root_el
->l_recs
[i
].e_blkno
) == left_el_blkno
)
1436 * The path walking code should have never returned a root and
1437 * two paths which are not adjacent.
1439 BUG_ON(i
>= (le16_to_cpu(root_el
->l_next_free_rec
) - 1));
1441 ocfs2_adjust_adjacent_records(&root_el
->l_recs
[i
], left_el
,
1442 &root_el
->l_recs
[i
+ 1], right_el
);
1446 * We've changed a leaf block (in right_path) and need to reflect that
1447 * change back up the subtree.
1449 * This happens in multiple places:
1450 * - When we've moved an extent record from the left path leaf to the right
1451 * path leaf to make room for an empty extent in the left path leaf.
1452 * - When our insert into the right path leaf is at the leftmost edge
1453 * and requires an update of the path immediately to it's left. This
1454 * can occur at the end of some types of rotation and appending inserts.
1456 static void ocfs2_complete_edge_insert(struct inode
*inode
, handle_t
*handle
,
1457 struct ocfs2_path
*left_path
,
1458 struct ocfs2_path
*right_path
,
1462 struct ocfs2_extent_list
*el
, *left_el
, *right_el
;
1463 struct ocfs2_extent_rec
*left_rec
, *right_rec
;
1464 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
1467 * Update the counts and position values within all the
1468 * interior nodes to reflect the leaf rotation we just did.
1470 * The root node is handled below the loop.
1472 * We begin the loop with right_el and left_el pointing to the
1473 * leaf lists and work our way up.
1475 * NOTE: within this loop, left_el and right_el always refer
1476 * to the *child* lists.
1478 left_el
= path_leaf_el(left_path
);
1479 right_el
= path_leaf_el(right_path
);
1480 for(i
= left_path
->p_tree_depth
- 1; i
> subtree_index
; i
--) {
1481 mlog(0, "Adjust records at index %u\n", i
);
1484 * One nice property of knowing that all of these
1485 * nodes are below the root is that we only deal with
1486 * the leftmost right node record and the rightmost
1489 el
= left_path
->p_node
[i
].el
;
1490 idx
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
1491 left_rec
= &el
->l_recs
[idx
];
1493 el
= right_path
->p_node
[i
].el
;
1494 right_rec
= &el
->l_recs
[0];
1496 ocfs2_adjust_adjacent_records(left_rec
, left_el
, right_rec
,
1499 ret
= ocfs2_journal_dirty(handle
, left_path
->p_node
[i
].bh
);
1503 ret
= ocfs2_journal_dirty(handle
, right_path
->p_node
[i
].bh
);
1508 * Setup our list pointers now so that the current
1509 * parents become children in the next iteration.
1511 left_el
= left_path
->p_node
[i
].el
;
1512 right_el
= right_path
->p_node
[i
].el
;
1516 * At the root node, adjust the two adjacent records which
1517 * begin our path to the leaves.
1520 el
= left_path
->p_node
[subtree_index
].el
;
1521 left_el
= left_path
->p_node
[subtree_index
+ 1].el
;
1522 right_el
= right_path
->p_node
[subtree_index
+ 1].el
;
1524 ocfs2_adjust_root_records(el
, left_el
, right_el
,
1525 left_path
->p_node
[subtree_index
+ 1].bh
->b_blocknr
);
1527 root_bh
= left_path
->p_node
[subtree_index
].bh
;
1529 ret
= ocfs2_journal_dirty(handle
, root_bh
);
1534 static int ocfs2_rotate_subtree_right(struct inode
*inode
,
1536 struct ocfs2_path
*left_path
,
1537 struct ocfs2_path
*right_path
,
1541 struct buffer_head
*right_leaf_bh
;
1542 struct buffer_head
*left_leaf_bh
= NULL
;
1543 struct buffer_head
*root_bh
;
1544 struct ocfs2_extent_list
*right_el
, *left_el
;
1545 struct ocfs2_extent_rec move_rec
;
1547 left_leaf_bh
= path_leaf_bh(left_path
);
1548 left_el
= path_leaf_el(left_path
);
1550 if (left_el
->l_next_free_rec
!= left_el
->l_count
) {
1551 ocfs2_error(inode
->i_sb
,
1552 "Inode %llu has non-full interior leaf node %llu"
1554 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1555 (unsigned long long)left_leaf_bh
->b_blocknr
,
1556 le16_to_cpu(left_el
->l_next_free_rec
));
1561 * This extent block may already have an empty record, so we
1562 * return early if so.
1564 if (ocfs2_is_empty_extent(&left_el
->l_recs
[0]))
1567 root_bh
= left_path
->p_node
[subtree_index
].bh
;
1568 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
1570 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
1571 OCFS2_JOURNAL_ACCESS_WRITE
);
1577 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
1578 ret
= ocfs2_journal_access(handle
, inode
,
1579 right_path
->p_node
[i
].bh
,
1580 OCFS2_JOURNAL_ACCESS_WRITE
);
1586 ret
= ocfs2_journal_access(handle
, inode
,
1587 left_path
->p_node
[i
].bh
,
1588 OCFS2_JOURNAL_ACCESS_WRITE
);
1595 right_leaf_bh
= path_leaf_bh(right_path
);
1596 right_el
= path_leaf_el(right_path
);
1598 /* This is a code error, not a disk corruption. */
1599 mlog_bug_on_msg(!right_el
->l_next_free_rec
, "Inode %llu: Rotate fails "
1600 "because rightmost leaf block %llu is empty\n",
1601 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1602 (unsigned long long)right_leaf_bh
->b_blocknr
);
1604 ocfs2_create_empty_extent(right_el
);
1606 ret
= ocfs2_journal_dirty(handle
, right_leaf_bh
);
1612 /* Do the copy now. */
1613 i
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
1614 move_rec
= left_el
->l_recs
[i
];
1615 right_el
->l_recs
[0] = move_rec
;
1618 * Clear out the record we just copied and shift everything
1619 * over, leaving an empty extent in the left leaf.
1621 * We temporarily subtract from next_free_rec so that the
1622 * shift will lose the tail record (which is now defunct).
1624 le16_add_cpu(&left_el
->l_next_free_rec
, -1);
1625 ocfs2_shift_records_right(left_el
);
1626 memset(&left_el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1627 le16_add_cpu(&left_el
->l_next_free_rec
, 1);
1629 ret
= ocfs2_journal_dirty(handle
, left_leaf_bh
);
1635 ocfs2_complete_edge_insert(inode
, handle
, left_path
, right_path
,
1643 * Given a full path, determine what cpos value would return us a path
1644 * containing the leaf immediately to the left of the current one.
1646 * Will return zero if the path passed in is already the leftmost path.
1648 static int ocfs2_find_cpos_for_left_leaf(struct super_block
*sb
,
1649 struct ocfs2_path
*path
, u32
*cpos
)
1653 struct ocfs2_extent_list
*el
;
1655 BUG_ON(path
->p_tree_depth
== 0);
1659 blkno
= path_leaf_bh(path
)->b_blocknr
;
1661 /* Start at the tree node just above the leaf and work our way up. */
1662 i
= path
->p_tree_depth
- 1;
1664 el
= path
->p_node
[i
].el
;
1667 * Find the extent record just before the one in our
1670 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
1671 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
1675 * We've determined that the
1676 * path specified is already
1677 * the leftmost one - return a
1683 * The leftmost record points to our
1684 * leaf - we need to travel up the
1690 *cpos
= le32_to_cpu(el
->l_recs
[j
- 1].e_cpos
);
1691 *cpos
= *cpos
+ ocfs2_rec_clusters(el
,
1692 &el
->l_recs
[j
- 1]);
1699 * If we got here, we never found a valid node where
1700 * the tree indicated one should be.
1703 "Invalid extent tree at extent block %llu\n",
1704 (unsigned long long)blkno
);
1709 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
1718 * Extend the transaction by enough credits to complete the rotation,
1719 * and still leave at least the original number of credits allocated
1720 * to this transaction.
1722 static int ocfs2_extend_rotate_transaction(handle_t
*handle
, int subtree_depth
,
1724 struct ocfs2_path
*path
)
1726 int credits
= (path
->p_tree_depth
- subtree_depth
) * 2 + 1 + op_credits
;
1728 if (handle
->h_buffer_credits
< credits
)
1729 return ocfs2_extend_trans(handle
, credits
);
1735 * Trap the case where we're inserting into the theoretical range past
1736 * the _actual_ left leaf range. Otherwise, we'll rotate a record
1737 * whose cpos is less than ours into the right leaf.
1739 * It's only necessary to look at the rightmost record of the left
1740 * leaf because the logic that calls us should ensure that the
1741 * theoretical ranges in the path components above the leaves are
1744 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path
*left_path
,
1747 struct ocfs2_extent_list
*left_el
;
1748 struct ocfs2_extent_rec
*rec
;
1751 left_el
= path_leaf_el(left_path
);
1752 next_free
= le16_to_cpu(left_el
->l_next_free_rec
);
1753 rec
= &left_el
->l_recs
[next_free
- 1];
1755 if (insert_cpos
> le32_to_cpu(rec
->e_cpos
))
1760 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list
*el
, u32 cpos
)
1762 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1764 struct ocfs2_extent_rec
*rec
;
1769 rec
= &el
->l_recs
[0];
1770 if (ocfs2_is_empty_extent(rec
)) {
1774 rec
= &el
->l_recs
[1];
1777 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
1778 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1784 * Rotate all the records in a btree right one record, starting at insert_cpos.
1786 * The path to the rightmost leaf should be passed in.
1788 * The array is assumed to be large enough to hold an entire path (tree depth).
1790 * Upon succesful return from this function:
1792 * - The 'right_path' array will contain a path to the leaf block
1793 * whose range contains e_cpos.
1794 * - That leaf block will have a single empty extent in list index 0.
1795 * - In the case that the rotation requires a post-insert update,
1796 * *ret_left_path will contain a valid path which can be passed to
1797 * ocfs2_insert_path().
1799 static int ocfs2_rotate_tree_right(struct inode
*inode
,
1801 enum ocfs2_split_type split
,
1803 struct ocfs2_path
*right_path
,
1804 struct ocfs2_path
**ret_left_path
)
1806 int ret
, start
, orig_credits
= handle
->h_buffer_credits
;
1808 struct ocfs2_path
*left_path
= NULL
;
1810 *ret_left_path
= NULL
;
1812 left_path
= ocfs2_new_path(path_root_bh(right_path
),
1813 path_root_el(right_path
));
1820 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
, &cpos
);
1826 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos
, cpos
);
1829 * What we want to do here is:
1831 * 1) Start with the rightmost path.
1833 * 2) Determine a path to the leaf block directly to the left
1836 * 3) Determine the 'subtree root' - the lowest level tree node
1837 * which contains a path to both leaves.
1839 * 4) Rotate the subtree.
1841 * 5) Find the next subtree by considering the left path to be
1842 * the new right path.
1844 * The check at the top of this while loop also accepts
1845 * insert_cpos == cpos because cpos is only a _theoretical_
1846 * value to get us the left path - insert_cpos might very well
1847 * be filling that hole.
1849 * Stop at a cpos of '0' because we either started at the
1850 * leftmost branch (i.e., a tree with one branch and a
1851 * rotation inside of it), or we've gone as far as we can in
1852 * rotating subtrees.
1854 while (cpos
&& insert_cpos
<= cpos
) {
1855 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
1858 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
1864 mlog_bug_on_msg(path_leaf_bh(left_path
) ==
1865 path_leaf_bh(right_path
),
1866 "Inode %lu: error during insert of %u "
1867 "(left path cpos %u) results in two identical "
1868 "paths ending at %llu\n",
1869 inode
->i_ino
, insert_cpos
, cpos
,
1870 (unsigned long long)
1871 path_leaf_bh(left_path
)->b_blocknr
);
1873 if (split
== SPLIT_NONE
&&
1874 ocfs2_rotate_requires_path_adjustment(left_path
,
1878 * We've rotated the tree as much as we
1879 * should. The rest is up to
1880 * ocfs2_insert_path() to complete, after the
1881 * record insertion. We indicate this
1882 * situation by returning the left path.
1884 * The reason we don't adjust the records here
1885 * before the record insert is that an error
1886 * later might break the rule where a parent
1887 * record e_cpos will reflect the actual
1888 * e_cpos of the 1st nonempty record of the
1891 *ret_left_path
= left_path
;
1895 start
= ocfs2_find_subtree_root(inode
, left_path
, right_path
);
1897 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
1899 (unsigned long long) right_path
->p_node
[start
].bh
->b_blocknr
,
1900 right_path
->p_tree_depth
);
1902 ret
= ocfs2_extend_rotate_transaction(handle
, start
,
1903 orig_credits
, right_path
);
1909 ret
= ocfs2_rotate_subtree_right(inode
, handle
, left_path
,
1916 if (split
!= SPLIT_NONE
&&
1917 ocfs2_leftmost_rec_contains(path_leaf_el(right_path
),
1920 * A rotate moves the rightmost left leaf
1921 * record over to the leftmost right leaf
1922 * slot. If we're doing an extent split
1923 * instead of a real insert, then we have to
1924 * check that the extent to be split wasn't
1925 * just moved over. If it was, then we can
1926 * exit here, passing left_path back -
1927 * ocfs2_split_extent() is smart enough to
1928 * search both leaves.
1930 *ret_left_path
= left_path
;
1935 * There is no need to re-read the next right path
1936 * as we know that it'll be our current left
1937 * path. Optimize by copying values instead.
1939 ocfs2_mv_path(right_path
, left_path
);
1941 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
,
1950 ocfs2_free_path(left_path
);
1956 static void ocfs2_update_edge_lengths(struct inode
*inode
, handle_t
*handle
,
1957 struct ocfs2_path
*path
)
1960 struct ocfs2_extent_rec
*rec
;
1961 struct ocfs2_extent_list
*el
;
1962 struct ocfs2_extent_block
*eb
;
1965 /* Path should always be rightmost. */
1966 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
1967 BUG_ON(eb
->h_next_leaf_blk
!= 0ULL);
1970 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
1971 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1972 rec
= &el
->l_recs
[idx
];
1973 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
1975 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
1976 el
= path
->p_node
[i
].el
;
1977 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1978 rec
= &el
->l_recs
[idx
];
1980 rec
->e_int_clusters
= cpu_to_le32(range
);
1981 le32_add_cpu(&rec
->e_int_clusters
, -le32_to_cpu(rec
->e_cpos
));
1983 ocfs2_journal_dirty(handle
, path
->p_node
[i
].bh
);
1987 static void ocfs2_unlink_path(struct inode
*inode
, handle_t
*handle
,
1988 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
1989 struct ocfs2_path
*path
, int unlink_start
)
1992 struct ocfs2_extent_block
*eb
;
1993 struct ocfs2_extent_list
*el
;
1994 struct buffer_head
*bh
;
1996 for(i
= unlink_start
; i
< path_num_items(path
); i
++) {
1997 bh
= path
->p_node
[i
].bh
;
1999 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
2001 * Not all nodes might have had their final count
2002 * decremented by the caller - handle this here.
2005 if (le16_to_cpu(el
->l_next_free_rec
) > 1) {
2007 "Inode %llu, attempted to remove extent block "
2008 "%llu with %u records\n",
2009 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2010 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
2011 le16_to_cpu(el
->l_next_free_rec
));
2013 ocfs2_journal_dirty(handle
, bh
);
2014 ocfs2_remove_from_cache(inode
, bh
);
2018 el
->l_next_free_rec
= 0;
2019 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2021 ocfs2_journal_dirty(handle
, bh
);
2023 ret
= ocfs2_cache_extent_block_free(dealloc
, eb
);
2027 ocfs2_remove_from_cache(inode
, bh
);
2031 static void ocfs2_unlink_subtree(struct inode
*inode
, handle_t
*handle
,
2032 struct ocfs2_path
*left_path
,
2033 struct ocfs2_path
*right_path
,
2035 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2038 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2039 struct ocfs2_extent_list
*root_el
= left_path
->p_node
[subtree_index
].el
;
2040 struct ocfs2_extent_list
*el
;
2041 struct ocfs2_extent_block
*eb
;
2043 el
= path_leaf_el(left_path
);
2045 eb
= (struct ocfs2_extent_block
*)right_path
->p_node
[subtree_index
+ 1].bh
->b_data
;
2047 for(i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
2048 if (root_el
->l_recs
[i
].e_blkno
== eb
->h_blkno
)
2051 BUG_ON(i
>= le16_to_cpu(root_el
->l_next_free_rec
));
2053 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
2054 le16_add_cpu(&root_el
->l_next_free_rec
, -1);
2056 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2057 eb
->h_next_leaf_blk
= 0;
2059 ocfs2_journal_dirty(handle
, root_bh
);
2060 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2062 ocfs2_unlink_path(inode
, handle
, dealloc
, right_path
,
2066 static int ocfs2_rotate_subtree_left(struct inode
*inode
, handle_t
*handle
,
2067 struct ocfs2_path
*left_path
,
2068 struct ocfs2_path
*right_path
,
2070 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2073 int ret
, i
, del_right_subtree
= 0, right_has_empty
= 0;
2074 struct buffer_head
*root_bh
, *di_bh
= path_root_bh(right_path
);
2075 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
2076 struct ocfs2_extent_list
*right_leaf_el
, *left_leaf_el
;
2077 struct ocfs2_extent_block
*eb
;
2081 right_leaf_el
= path_leaf_el(right_path
);
2082 left_leaf_el
= path_leaf_el(left_path
);
2083 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2084 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2086 if (!ocfs2_is_empty_extent(&left_leaf_el
->l_recs
[0]))
2089 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(right_path
)->b_data
;
2090 if (ocfs2_is_empty_extent(&right_leaf_el
->l_recs
[0])) {
2092 * It's legal for us to proceed if the right leaf is
2093 * the rightmost one and it has an empty extent. There
2094 * are two cases to handle - whether the leaf will be
2095 * empty after removal or not. If the leaf isn't empty
2096 * then just remove the empty extent up front. The
2097 * next block will handle empty leaves by flagging
2100 * Non rightmost leaves will throw -EAGAIN and the
2101 * caller can manually move the subtree and retry.
2104 if (eb
->h_next_leaf_blk
!= 0ULL)
2107 if (le16_to_cpu(right_leaf_el
->l_next_free_rec
) > 1) {
2108 ret
= ocfs2_journal_access(handle
, inode
,
2109 path_leaf_bh(right_path
),
2110 OCFS2_JOURNAL_ACCESS_WRITE
);
2116 ocfs2_remove_empty_extent(right_leaf_el
);
2118 right_has_empty
= 1;
2121 if (eb
->h_next_leaf_blk
== 0ULL &&
2122 le16_to_cpu(right_leaf_el
->l_next_free_rec
) == 1) {
2124 * We have to update i_last_eb_blk during the meta
2127 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
2128 OCFS2_JOURNAL_ACCESS_WRITE
);
2134 del_right_subtree
= 1;
2138 * Getting here with an empty extent in the right path implies
2139 * that it's the rightmost path and will be deleted.
2141 BUG_ON(right_has_empty
&& !del_right_subtree
);
2143 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
2144 OCFS2_JOURNAL_ACCESS_WRITE
);
2150 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2151 ret
= ocfs2_journal_access(handle
, inode
,
2152 right_path
->p_node
[i
].bh
,
2153 OCFS2_JOURNAL_ACCESS_WRITE
);
2159 ret
= ocfs2_journal_access(handle
, inode
,
2160 left_path
->p_node
[i
].bh
,
2161 OCFS2_JOURNAL_ACCESS_WRITE
);
2168 if (!right_has_empty
) {
2170 * Only do this if we're moving a real
2171 * record. Otherwise, the action is delayed until
2172 * after removal of the right path in which case we
2173 * can do a simple shift to remove the empty extent.
2175 ocfs2_rotate_leaf(left_leaf_el
, &right_leaf_el
->l_recs
[0]);
2176 memset(&right_leaf_el
->l_recs
[0], 0,
2177 sizeof(struct ocfs2_extent_rec
));
2179 if (eb
->h_next_leaf_blk
== 0ULL) {
2181 * Move recs over to get rid of empty extent, decrease
2182 * next_free. This is allowed to remove the last
2183 * extent in our leaf (setting l_next_free_rec to
2184 * zero) - the delete code below won't care.
2186 ocfs2_remove_empty_extent(right_leaf_el
);
2189 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2192 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2196 if (del_right_subtree
) {
2197 ocfs2_unlink_subtree(inode
, handle
, left_path
, right_path
,
2198 subtree_index
, dealloc
);
2199 ocfs2_update_edge_lengths(inode
, handle
, left_path
);
2201 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2202 di
->i_last_eb_blk
= eb
->h_blkno
;
2205 * Removal of the extent in the left leaf was skipped
2206 * above so we could delete the right path
2209 if (right_has_empty
)
2210 ocfs2_remove_empty_extent(left_leaf_el
);
2212 ret
= ocfs2_journal_dirty(handle
, di_bh
);
2218 ocfs2_complete_edge_insert(inode
, handle
, left_path
, right_path
,
2226 * Given a full path, determine what cpos value would return us a path
2227 * containing the leaf immediately to the right of the current one.
2229 * Will return zero if the path passed in is already the rightmost path.
2231 * This looks similar, but is subtly different to
2232 * ocfs2_find_cpos_for_left_leaf().
2234 static int ocfs2_find_cpos_for_right_leaf(struct super_block
*sb
,
2235 struct ocfs2_path
*path
, u32
*cpos
)
2239 struct ocfs2_extent_list
*el
;
2243 if (path
->p_tree_depth
== 0)
2246 blkno
= path_leaf_bh(path
)->b_blocknr
;
2248 /* Start at the tree node just above the leaf and work our way up. */
2249 i
= path
->p_tree_depth
- 1;
2253 el
= path
->p_node
[i
].el
;
2256 * Find the extent record just after the one in our
2259 next_free
= le16_to_cpu(el
->l_next_free_rec
);
2260 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2261 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2262 if (j
== (next_free
- 1)) {
2265 * We've determined that the
2266 * path specified is already
2267 * the rightmost one - return a
2273 * The rightmost record points to our
2274 * leaf - we need to travel up the
2280 *cpos
= le32_to_cpu(el
->l_recs
[j
+ 1].e_cpos
);
2286 * If we got here, we never found a valid node where
2287 * the tree indicated one should be.
2290 "Invalid extent tree at extent block %llu\n",
2291 (unsigned long long)blkno
);
2296 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2304 static int ocfs2_rotate_rightmost_leaf_left(struct inode
*inode
,
2306 struct buffer_head
*bh
,
2307 struct ocfs2_extent_list
*el
)
2311 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2314 ret
= ocfs2_journal_access(handle
, inode
, bh
,
2315 OCFS2_JOURNAL_ACCESS_WRITE
);
2321 ocfs2_remove_empty_extent(el
);
2323 ret
= ocfs2_journal_dirty(handle
, bh
);
2331 static int __ocfs2_rotate_tree_left(struct inode
*inode
,
2332 handle_t
*handle
, int orig_credits
,
2333 struct ocfs2_path
*path
,
2334 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2335 struct ocfs2_path
**empty_extent_path
)
2337 int ret
, subtree_root
, deleted
;
2339 struct ocfs2_path
*left_path
= NULL
;
2340 struct ocfs2_path
*right_path
= NULL
;
2342 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path
)->l_recs
[0])));
2344 *empty_extent_path
= NULL
;
2346 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, path
,
2353 left_path
= ocfs2_new_path(path_root_bh(path
),
2354 path_root_el(path
));
2361 ocfs2_cp_path(left_path
, path
);
2363 right_path
= ocfs2_new_path(path_root_bh(path
),
2364 path_root_el(path
));
2371 while (right_cpos
) {
2372 ret
= ocfs2_find_path(inode
, right_path
, right_cpos
);
2378 subtree_root
= ocfs2_find_subtree_root(inode
, left_path
,
2381 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2383 (unsigned long long)
2384 right_path
->p_node
[subtree_root
].bh
->b_blocknr
,
2385 right_path
->p_tree_depth
);
2387 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_root
,
2388 orig_credits
, left_path
);
2394 ret
= ocfs2_rotate_subtree_left(inode
, handle
, left_path
,
2395 right_path
, subtree_root
,
2397 if (ret
== -EAGAIN
) {
2399 * The rotation has to temporarily stop due to
2400 * the right subtree having an empty
2401 * extent. Pass it back to the caller for a
2404 *empty_extent_path
= right_path
;
2414 * The subtree rotate might have removed records on
2415 * the rightmost edge. If so, then rotation is
2421 ocfs2_mv_path(left_path
, right_path
);
2423 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, left_path
,
2432 ocfs2_free_path(right_path
);
2433 ocfs2_free_path(left_path
);
2438 static int ocfs2_remove_rightmost_path(struct inode
*inode
, handle_t
*handle
,
2439 struct ocfs2_path
*path
,
2440 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2442 int ret
, subtree_index
;
2444 struct ocfs2_path
*left_path
= NULL
;
2445 struct ocfs2_dinode
*di
;
2446 struct ocfs2_extent_block
*eb
;
2447 struct ocfs2_extent_list
*el
;
2450 * XXX: This code assumes that the root is an inode, which is
2451 * true for now but may change as tree code gets generic.
2453 di
= (struct ocfs2_dinode
*)path_root_bh(path
)->b_data
;
2454 if (!OCFS2_IS_VALID_DINODE(di
)) {
2456 ocfs2_error(inode
->i_sb
,
2457 "Inode %llu has invalid path root",
2458 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
2463 * There's two ways we handle this depending on
2464 * whether path is the only existing one.
2466 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
2467 handle
->h_buffer_credits
,
2474 ret
= ocfs2_journal_access_path(inode
, handle
, path
);
2480 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
2488 * We have a path to the left of this one - it needs
2491 left_path
= ocfs2_new_path(path_root_bh(path
),
2492 path_root_el(path
));
2499 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
2505 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
2511 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
, path
);
2513 ocfs2_unlink_subtree(inode
, handle
, left_path
, path
,
2514 subtree_index
, dealloc
);
2515 ocfs2_update_edge_lengths(inode
, handle
, left_path
);
2517 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2518 di
->i_last_eb_blk
= eb
->h_blkno
;
2521 * 'path' is also the leftmost path which
2522 * means it must be the only one. This gets
2523 * handled differently because we want to
2524 * revert the inode back to having extents
2527 ocfs2_unlink_path(inode
, handle
, dealloc
, path
, 1);
2529 el
= &di
->id2
.i_list
;
2530 el
->l_tree_depth
= 0;
2531 el
->l_next_free_rec
= 0;
2532 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2534 di
->i_last_eb_blk
= 0;
2537 ocfs2_journal_dirty(handle
, path_root_bh(path
));
2540 ocfs2_free_path(left_path
);
2545 * Left rotation of btree records.
2547 * In many ways, this is (unsurprisingly) the opposite of right
2548 * rotation. We start at some non-rightmost path containing an empty
2549 * extent in the leaf block. The code works its way to the rightmost
2550 * path by rotating records to the left in every subtree.
2552 * This is used by any code which reduces the number of extent records
2553 * in a leaf. After removal, an empty record should be placed in the
2554 * leftmost list position.
2556 * This won't handle a length update of the rightmost path records if
2557 * the rightmost tree leaf record is removed so the caller is
2558 * responsible for detecting and correcting that.
2560 static int ocfs2_rotate_tree_left(struct inode
*inode
, handle_t
*handle
,
2561 struct ocfs2_path
*path
,
2562 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2564 int ret
, orig_credits
= handle
->h_buffer_credits
;
2565 struct ocfs2_path
*tmp_path
= NULL
, *restart_path
= NULL
;
2566 struct ocfs2_extent_block
*eb
;
2567 struct ocfs2_extent_list
*el
;
2569 el
= path_leaf_el(path
);
2570 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2573 if (path
->p_tree_depth
== 0) {
2574 rightmost_no_delete
:
2576 * In-inode extents. This is trivially handled, so do
2579 ret
= ocfs2_rotate_rightmost_leaf_left(inode
, handle
,
2581 path_leaf_el(path
));
2588 * Handle rightmost branch now. There's several cases:
2589 * 1) simple rotation leaving records in there. That's trivial.
2590 * 2) rotation requiring a branch delete - there's no more
2591 * records left. Two cases of this:
2592 * a) There are branches to the left.
2593 * b) This is also the leftmost (the only) branch.
2595 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
2596 * 2a) we need the left branch so that we can update it with the unlink
2597 * 2b) we need to bring the inode back to inline extents.
2600 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2602 if (eb
->h_next_leaf_blk
== 0) {
2604 * This gets a bit tricky if we're going to delete the
2605 * rightmost path. Get the other cases out of the way
2608 if (le16_to_cpu(el
->l_next_free_rec
) > 1)
2609 goto rightmost_no_delete
;
2611 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
2613 ocfs2_error(inode
->i_sb
,
2614 "Inode %llu has empty extent block at %llu",
2615 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2616 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
2621 * XXX: The caller can not trust "path" any more after
2622 * this as it will have been deleted. What do we do?
2624 * In theory the rotate-for-merge code will never get
2625 * here because it'll always ask for a rotate in a
2629 ret
= ocfs2_remove_rightmost_path(inode
, handle
, path
,
2637 * Now we can loop, remembering the path we get from -EAGAIN
2638 * and restarting from there.
2641 ret
= __ocfs2_rotate_tree_left(inode
, handle
, orig_credits
, path
,
2642 dealloc
, &restart_path
);
2643 if (ret
&& ret
!= -EAGAIN
) {
2648 while (ret
== -EAGAIN
) {
2649 tmp_path
= restart_path
;
2650 restart_path
= NULL
;
2652 ret
= __ocfs2_rotate_tree_left(inode
, handle
, orig_credits
,
2655 if (ret
&& ret
!= -EAGAIN
) {
2660 ocfs2_free_path(tmp_path
);
2668 ocfs2_free_path(tmp_path
);
2669 ocfs2_free_path(restart_path
);
2673 static void ocfs2_cleanup_merge(struct ocfs2_extent_list
*el
,
2676 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[index
];
2679 if (rec
->e_leaf_clusters
== 0) {
2681 * We consumed all of the merged-from record. An empty
2682 * extent cannot exist anywhere but the 1st array
2683 * position, so move things over if the merged-from
2684 * record doesn't occupy that position.
2686 * This creates a new empty extent so the caller
2687 * should be smart enough to have removed any existing
2691 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
2692 size
= index
* sizeof(struct ocfs2_extent_rec
);
2693 memmove(&el
->l_recs
[1], &el
->l_recs
[0], size
);
2697 * Always memset - the caller doesn't check whether it
2698 * created an empty extent, so there could be junk in
2701 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2706 * Remove split_rec clusters from the record at index and merge them
2707 * onto the beginning of the record at index + 1.
2709 static int ocfs2_merge_rec_right(struct inode
*inode
, struct buffer_head
*bh
,
2711 struct ocfs2_extent_rec
*split_rec
,
2712 struct ocfs2_extent_list
*el
, int index
)
2715 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
2716 struct ocfs2_extent_rec
*left_rec
;
2717 struct ocfs2_extent_rec
*right_rec
;
2719 BUG_ON(index
>= le16_to_cpu(el
->l_next_free_rec
));
2721 left_rec
= &el
->l_recs
[index
];
2722 right_rec
= &el
->l_recs
[index
+ 1];
2724 ret
= ocfs2_journal_access(handle
, inode
, bh
,
2725 OCFS2_JOURNAL_ACCESS_WRITE
);
2731 le16_add_cpu(&left_rec
->e_leaf_clusters
, -split_clusters
);
2733 le32_add_cpu(&right_rec
->e_cpos
, -split_clusters
);
2734 le64_add_cpu(&right_rec
->e_blkno
,
2735 -ocfs2_clusters_to_blocks(inode
->i_sb
, split_clusters
));
2736 le16_add_cpu(&right_rec
->e_leaf_clusters
, split_clusters
);
2738 ocfs2_cleanup_merge(el
, index
);
2740 ret
= ocfs2_journal_dirty(handle
, bh
);
2749 * Remove split_rec clusters from the record at index and merge them
2750 * onto the tail of the record at index - 1.
2752 static int ocfs2_merge_rec_left(struct inode
*inode
, struct buffer_head
*bh
,
2754 struct ocfs2_extent_rec
*split_rec
,
2755 struct ocfs2_extent_list
*el
, int index
)
2757 int ret
, has_empty_extent
= 0;
2758 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
2759 struct ocfs2_extent_rec
*left_rec
;
2760 struct ocfs2_extent_rec
*right_rec
;
2764 left_rec
= &el
->l_recs
[index
- 1];
2765 right_rec
= &el
->l_recs
[index
];
2766 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
2767 has_empty_extent
= 1;
2769 ret
= ocfs2_journal_access(handle
, inode
, bh
,
2770 OCFS2_JOURNAL_ACCESS_WRITE
);
2776 if (has_empty_extent
&& index
== 1) {
2778 * The easy case - we can just plop the record right in.
2780 *left_rec
= *split_rec
;
2782 has_empty_extent
= 0;
2784 le16_add_cpu(&left_rec
->e_leaf_clusters
, split_clusters
);
2787 le32_add_cpu(&right_rec
->e_cpos
, split_clusters
);
2788 le64_add_cpu(&right_rec
->e_blkno
,
2789 ocfs2_clusters_to_blocks(inode
->i_sb
, split_clusters
));
2790 le16_add_cpu(&right_rec
->e_leaf_clusters
, -split_clusters
);
2792 ocfs2_cleanup_merge(el
, index
);
2794 ret
= ocfs2_journal_dirty(handle
, bh
);
2802 static int ocfs2_try_to_merge_extent(struct inode
*inode
,
2804 struct ocfs2_path
*left_path
,
2806 struct ocfs2_extent_rec
*split_rec
,
2807 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2808 struct ocfs2_merge_ctxt
*ctxt
)
2811 int ret
= 0, delete_tail_recs
= 0;
2812 struct ocfs2_extent_list
*el
= path_leaf_el(left_path
);
2813 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
2815 BUG_ON(ctxt
->c_contig_type
== CONTIG_NONE
);
2817 if (ctxt
->c_split_covers_rec
) {
2820 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
||
2821 ctxt
->c_has_empty_extent
)
2824 if (ctxt
->c_has_empty_extent
) {
2826 * The merge code will need to create an empty
2827 * extent to take the place of the newly
2828 * emptied slot. Remove any pre-existing empty
2829 * extents - having more than one in a leaf is
2832 ret
= ocfs2_rotate_tree_left(inode
, handle
, left_path
,
2839 rec
= &el
->l_recs
[split_index
];
2843 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
) {
2845 * Left-right contig implies this.
2847 BUG_ON(!ctxt
->c_split_covers_rec
);
2848 BUG_ON(split_index
== 0);
2851 * Since the leftright insert always covers the entire
2852 * extent, this call will delete the insert record
2853 * entirely, resulting in an empty extent record added to
2856 * Since the adding of an empty extent shifts
2857 * everything back to the right, there's no need to
2858 * update split_index here.
2860 ret
= ocfs2_merge_rec_left(inode
, path_leaf_bh(left_path
),
2861 handle
, split_rec
, el
, split_index
);
2868 * We can only get this from logic error above.
2870 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
2873 * The left merge left us with an empty extent, remove
2876 ret
= ocfs2_rotate_tree_left(inode
, handle
, left_path
, dealloc
);
2882 rec
= &el
->l_recs
[split_index
];
2885 * Note that we don't pass split_rec here on purpose -
2886 * we've merged it into the left side.
2888 ret
= ocfs2_merge_rec_right(inode
, path_leaf_bh(left_path
),
2889 handle
, rec
, el
, split_index
);
2895 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
2897 ret
= ocfs2_rotate_tree_left(inode
, handle
, left_path
,
2900 * Error from this last rotate is not critical, so
2901 * print but don't bubble it up.
2908 * Merge a record to the left or right.
2910 * 'contig_type' is relative to the existing record,
2911 * so for example, if we're "right contig", it's to
2912 * the record on the left (hence the left merge).
2914 if (ctxt
->c_contig_type
== CONTIG_RIGHT
) {
2915 ret
= ocfs2_merge_rec_left(inode
,
2916 path_leaf_bh(left_path
),
2917 handle
, split_rec
, el
,
2924 ret
= ocfs2_merge_rec_right(inode
,
2925 path_leaf_bh(left_path
),
2926 handle
, split_rec
, el
,
2934 if (ctxt
->c_split_covers_rec
) {
2936 * The merge may have left an empty extent in
2937 * our leaf. Try to rotate it away.
2939 ret
= ocfs2_rotate_tree_left(inode
, handle
, left_path
,
2951 static void ocfs2_subtract_from_rec(struct super_block
*sb
,
2952 enum ocfs2_split_type split
,
2953 struct ocfs2_extent_rec
*rec
,
2954 struct ocfs2_extent_rec
*split_rec
)
2958 len_blocks
= ocfs2_clusters_to_blocks(sb
,
2959 le16_to_cpu(split_rec
->e_leaf_clusters
));
2961 if (split
== SPLIT_LEFT
) {
2963 * Region is on the left edge of the existing
2966 le32_add_cpu(&rec
->e_cpos
,
2967 le16_to_cpu(split_rec
->e_leaf_clusters
));
2968 le64_add_cpu(&rec
->e_blkno
, len_blocks
);
2969 le16_add_cpu(&rec
->e_leaf_clusters
,
2970 -le16_to_cpu(split_rec
->e_leaf_clusters
));
2973 * Region is on the right edge of the existing
2976 le16_add_cpu(&rec
->e_leaf_clusters
,
2977 -le16_to_cpu(split_rec
->e_leaf_clusters
));
2982 * Do the final bits of extent record insertion at the target leaf
2983 * list. If this leaf is part of an allocation tree, it is assumed
2984 * that the tree above has been prepared.
2986 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec
*insert_rec
,
2987 struct ocfs2_extent_list
*el
,
2988 struct ocfs2_insert_type
*insert
,
2989 struct inode
*inode
)
2991 int i
= insert
->ins_contig_index
;
2993 struct ocfs2_extent_rec
*rec
;
2995 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
2997 if (insert
->ins_split
!= SPLIT_NONE
) {
2998 i
= ocfs2_search_extent_list(el
, le32_to_cpu(insert_rec
->e_cpos
));
3000 rec
= &el
->l_recs
[i
];
3001 ocfs2_subtract_from_rec(inode
->i_sb
, insert
->ins_split
, rec
,
3007 * Contiguous insert - either left or right.
3009 if (insert
->ins_contig
!= CONTIG_NONE
) {
3010 rec
= &el
->l_recs
[i
];
3011 if (insert
->ins_contig
== CONTIG_LEFT
) {
3012 rec
->e_blkno
= insert_rec
->e_blkno
;
3013 rec
->e_cpos
= insert_rec
->e_cpos
;
3015 le16_add_cpu(&rec
->e_leaf_clusters
,
3016 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3021 * Handle insert into an empty leaf.
3023 if (le16_to_cpu(el
->l_next_free_rec
) == 0 ||
3024 ((le16_to_cpu(el
->l_next_free_rec
) == 1) &&
3025 ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3026 el
->l_recs
[0] = *insert_rec
;
3027 el
->l_next_free_rec
= cpu_to_le16(1);
3034 if (insert
->ins_appending
== APPEND_TAIL
) {
3035 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3036 rec
= &el
->l_recs
[i
];
3037 range
= le32_to_cpu(rec
->e_cpos
)
3038 + le16_to_cpu(rec
->e_leaf_clusters
);
3039 BUG_ON(le32_to_cpu(insert_rec
->e_cpos
) < range
);
3041 mlog_bug_on_msg(le16_to_cpu(el
->l_next_free_rec
) >=
3042 le16_to_cpu(el
->l_count
),
3043 "inode %lu, depth %u, count %u, next free %u, "
3044 "rec.cpos %u, rec.clusters %u, "
3045 "insert.cpos %u, insert.clusters %u\n",
3047 le16_to_cpu(el
->l_tree_depth
),
3048 le16_to_cpu(el
->l_count
),
3049 le16_to_cpu(el
->l_next_free_rec
),
3050 le32_to_cpu(el
->l_recs
[i
].e_cpos
),
3051 le16_to_cpu(el
->l_recs
[i
].e_leaf_clusters
),
3052 le32_to_cpu(insert_rec
->e_cpos
),
3053 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3055 el
->l_recs
[i
] = *insert_rec
;
3056 le16_add_cpu(&el
->l_next_free_rec
, 1);
3062 * Ok, we have to rotate.
3064 * At this point, it is safe to assume that inserting into an
3065 * empty leaf and appending to a leaf have both been handled
3068 * This leaf needs to have space, either by the empty 1st
3069 * extent record, or by virtue of an l_next_rec < l_count.
3071 ocfs2_rotate_leaf(el
, insert_rec
);
3074 static inline void ocfs2_update_dinode_clusters(struct inode
*inode
,
3075 struct ocfs2_dinode
*di
,
3078 le32_add_cpu(&di
->i_clusters
, clusters
);
3079 spin_lock(&OCFS2_I(inode
)->ip_lock
);
3080 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(di
->i_clusters
);
3081 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
3084 static void ocfs2_adjust_rightmost_records(struct inode
*inode
,
3086 struct ocfs2_path
*path
,
3087 struct ocfs2_extent_rec
*insert_rec
)
3089 int ret
, i
, next_free
;
3090 struct buffer_head
*bh
;
3091 struct ocfs2_extent_list
*el
;
3092 struct ocfs2_extent_rec
*rec
;
3095 * Update everything except the leaf block.
3097 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
3098 bh
= path
->p_node
[i
].bh
;
3099 el
= path
->p_node
[i
].el
;
3101 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3102 if (next_free
== 0) {
3103 ocfs2_error(inode
->i_sb
,
3104 "Dinode %llu has a bad extent list",
3105 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
3110 rec
= &el
->l_recs
[next_free
- 1];
3112 rec
->e_int_clusters
= insert_rec
->e_cpos
;
3113 le32_add_cpu(&rec
->e_int_clusters
,
3114 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3115 le32_add_cpu(&rec
->e_int_clusters
,
3116 -le32_to_cpu(rec
->e_cpos
));
3118 ret
= ocfs2_journal_dirty(handle
, bh
);
3125 static int ocfs2_append_rec_to_path(struct inode
*inode
, handle_t
*handle
,
3126 struct ocfs2_extent_rec
*insert_rec
,
3127 struct ocfs2_path
*right_path
,
3128 struct ocfs2_path
**ret_left_path
)
3131 struct ocfs2_extent_list
*el
;
3132 struct ocfs2_path
*left_path
= NULL
;
3134 *ret_left_path
= NULL
;
3137 * This shouldn't happen for non-trees. The extent rec cluster
3138 * count manipulation below only works for interior nodes.
3140 BUG_ON(right_path
->p_tree_depth
== 0);
3143 * If our appending insert is at the leftmost edge of a leaf,
3144 * then we might need to update the rightmost records of the
3147 el
= path_leaf_el(right_path
);
3148 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3149 if (next_free
== 0 ||
3150 (next_free
== 1 && ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3153 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
,
3160 mlog(0, "Append may need a left path update. cpos: %u, "
3161 "left_cpos: %u\n", le32_to_cpu(insert_rec
->e_cpos
),
3165 * No need to worry if the append is already in the
3169 left_path
= ocfs2_new_path(path_root_bh(right_path
),
3170 path_root_el(right_path
));
3177 ret
= ocfs2_find_path(inode
, left_path
, left_cpos
);
3184 * ocfs2_insert_path() will pass the left_path to the
3190 ret
= ocfs2_journal_access_path(inode
, handle
, right_path
);
3196 ocfs2_adjust_rightmost_records(inode
, handle
, right_path
, insert_rec
);
3198 *ret_left_path
= left_path
;
3202 ocfs2_free_path(left_path
);
3207 static void ocfs2_split_record(struct inode
*inode
,
3208 struct ocfs2_path
*left_path
,
3209 struct ocfs2_path
*right_path
,
3210 struct ocfs2_extent_rec
*split_rec
,
3211 enum ocfs2_split_type split
)
3214 u32 cpos
= le32_to_cpu(split_rec
->e_cpos
);
3215 struct ocfs2_extent_list
*left_el
= NULL
, *right_el
, *insert_el
, *el
;
3216 struct ocfs2_extent_rec
*rec
, *tmprec
;
3218 right_el
= path_leaf_el(right_path
);;
3220 left_el
= path_leaf_el(left_path
);
3223 insert_el
= right_el
;
3224 index
= ocfs2_search_extent_list(el
, cpos
);
3226 if (index
== 0 && left_path
) {
3227 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3230 * This typically means that the record
3231 * started in the left path but moved to the
3232 * right as a result of rotation. We either
3233 * move the existing record to the left, or we
3234 * do the later insert there.
3236 * In this case, the left path should always
3237 * exist as the rotate code will have passed
3238 * it back for a post-insert update.
3241 if (split
== SPLIT_LEFT
) {
3243 * It's a left split. Since we know
3244 * that the rotate code gave us an
3245 * empty extent in the left path, we
3246 * can just do the insert there.
3248 insert_el
= left_el
;
3251 * Right split - we have to move the
3252 * existing record over to the left
3253 * leaf. The insert will be into the
3254 * newly created empty extent in the
3257 tmprec
= &right_el
->l_recs
[index
];
3258 ocfs2_rotate_leaf(left_el
, tmprec
);
3261 memset(tmprec
, 0, sizeof(*tmprec
));
3262 index
= ocfs2_search_extent_list(left_el
, cpos
);
3263 BUG_ON(index
== -1);
3268 BUG_ON(!ocfs2_is_empty_extent(&left_el
->l_recs
[0]));
3270 * Left path is easy - we can just allow the insert to
3274 insert_el
= left_el
;
3275 index
= ocfs2_search_extent_list(el
, cpos
);
3276 BUG_ON(index
== -1);
3279 rec
= &el
->l_recs
[index
];
3280 ocfs2_subtract_from_rec(inode
->i_sb
, split
, rec
, split_rec
);
3281 ocfs2_rotate_leaf(insert_el
, split_rec
);
3285 * This function only does inserts on an allocation b-tree. For dinode
3286 * lists, ocfs2_insert_at_leaf() is called directly.
3288 * right_path is the path we want to do the actual insert
3289 * in. left_path should only be passed in if we need to update that
3290 * portion of the tree after an edge insert.
3292 static int ocfs2_insert_path(struct inode
*inode
,
3294 struct ocfs2_path
*left_path
,
3295 struct ocfs2_path
*right_path
,
3296 struct ocfs2_extent_rec
*insert_rec
,
3297 struct ocfs2_insert_type
*insert
)
3299 int ret
, subtree_index
;
3300 struct buffer_head
*leaf_bh
= path_leaf_bh(right_path
);
3303 * Pass both paths to the journal. The majority of inserts
3304 * will be touching all components anyway.
3306 ret
= ocfs2_journal_access_path(inode
, handle
, right_path
);
3313 int credits
= handle
->h_buffer_credits
;
3316 * There's a chance that left_path got passed back to
3317 * us without being accounted for in the
3318 * journal. Extend our transaction here to be sure we
3319 * can change those blocks.
3321 credits
+= left_path
->p_tree_depth
;
3323 ret
= ocfs2_extend_trans(handle
, credits
);
3329 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
3336 if (insert
->ins_split
!= SPLIT_NONE
) {
3338 * We could call ocfs2_insert_at_leaf() for some types
3339 * of splits, but it's easier to just let one seperate
3340 * function sort it all out.
3342 ocfs2_split_record(inode
, left_path
, right_path
,
3343 insert_rec
, insert
->ins_split
);
3345 ocfs2_insert_at_leaf(insert_rec
, path_leaf_el(right_path
),
3348 ret
= ocfs2_journal_dirty(handle
, leaf_bh
);
3354 * The rotate code has indicated that we need to fix
3355 * up portions of the tree after the insert.
3357 * XXX: Should we extend the transaction here?
3359 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
,
3361 ocfs2_complete_edge_insert(inode
, handle
, left_path
,
3362 right_path
, subtree_index
);
3370 static int ocfs2_do_insert_extent(struct inode
*inode
,
3372 struct buffer_head
*di_bh
,
3373 struct ocfs2_extent_rec
*insert_rec
,
3374 struct ocfs2_insert_type
*type
)
3376 int ret
, rotate
= 0;
3378 struct ocfs2_path
*right_path
= NULL
;
3379 struct ocfs2_path
*left_path
= NULL
;
3380 struct ocfs2_dinode
*di
;
3381 struct ocfs2_extent_list
*el
;
3383 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
3384 el
= &di
->id2
.i_list
;
3386 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
3387 OCFS2_JOURNAL_ACCESS_WRITE
);
3393 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
3394 ocfs2_insert_at_leaf(insert_rec
, el
, type
, inode
);
3395 goto out_update_clusters
;
3398 right_path
= ocfs2_new_inode_path(di_bh
);
3406 * Determine the path to start with. Rotations need the
3407 * rightmost path, everything else can go directly to the
3410 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
3411 if (type
->ins_appending
== APPEND_NONE
&&
3412 type
->ins_contig
== CONTIG_NONE
) {
3417 ret
= ocfs2_find_path(inode
, right_path
, cpos
);
3424 * Rotations and appends need special treatment - they modify
3425 * parts of the tree's above them.
3427 * Both might pass back a path immediate to the left of the
3428 * one being inserted to. This will be cause
3429 * ocfs2_insert_path() to modify the rightmost records of
3430 * left_path to account for an edge insert.
3432 * XXX: When modifying this code, keep in mind that an insert
3433 * can wind up skipping both of these two special cases...
3436 ret
= ocfs2_rotate_tree_right(inode
, handle
, type
->ins_split
,
3437 le32_to_cpu(insert_rec
->e_cpos
),
3438 right_path
, &left_path
);
3443 } else if (type
->ins_appending
== APPEND_TAIL
3444 && type
->ins_contig
!= CONTIG_LEFT
) {
3445 ret
= ocfs2_append_rec_to_path(inode
, handle
, insert_rec
,
3446 right_path
, &left_path
);
3453 ret
= ocfs2_insert_path(inode
, handle
, left_path
, right_path
,
3460 out_update_clusters
:
3461 if (type
->ins_split
== SPLIT_NONE
)
3462 ocfs2_update_dinode_clusters(inode
, di
,
3463 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3465 ret
= ocfs2_journal_dirty(handle
, di_bh
);
3470 ocfs2_free_path(left_path
);
3471 ocfs2_free_path(right_path
);
3476 static enum ocfs2_contig_type
3477 ocfs2_figure_merge_contig_type(struct inode
*inode
,
3478 struct ocfs2_extent_list
*el
, int index
,
3479 struct ocfs2_extent_rec
*split_rec
)
3481 struct ocfs2_extent_rec
*rec
;
3482 enum ocfs2_contig_type ret
= CONTIG_NONE
;
3485 * We're careful to check for an empty extent record here -
3486 * the merge code will know what to do if it sees one.
3490 rec
= &el
->l_recs
[index
- 1];
3491 if (index
== 1 && ocfs2_is_empty_extent(rec
)) {
3492 if (split_rec
->e_cpos
== el
->l_recs
[index
].e_cpos
)
3495 ret
= ocfs2_extent_contig(inode
, rec
, split_rec
);
3499 if (index
< (le16_to_cpu(el
->l_next_free_rec
) - 1)) {
3500 enum ocfs2_contig_type contig_type
;
3502 rec
= &el
->l_recs
[index
+ 1];
3503 contig_type
= ocfs2_extent_contig(inode
, rec
, split_rec
);
3505 if (contig_type
== CONTIG_LEFT
&& ret
== CONTIG_RIGHT
)
3506 ret
= CONTIG_LEFTRIGHT
;
3507 else if (ret
== CONTIG_NONE
)
3514 static void ocfs2_figure_contig_type(struct inode
*inode
,
3515 struct ocfs2_insert_type
*insert
,
3516 struct ocfs2_extent_list
*el
,
3517 struct ocfs2_extent_rec
*insert_rec
)
3520 enum ocfs2_contig_type contig_type
= CONTIG_NONE
;
3522 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3524 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
3525 contig_type
= ocfs2_extent_contig(inode
, &el
->l_recs
[i
],
3527 if (contig_type
!= CONTIG_NONE
) {
3528 insert
->ins_contig_index
= i
;
3532 insert
->ins_contig
= contig_type
;
3536 * This should only be called against the righmost leaf extent list.
3538 * ocfs2_figure_appending_type() will figure out whether we'll have to
3539 * insert at the tail of the rightmost leaf.
3541 * This should also work against the dinode list for tree's with 0
3542 * depth. If we consider the dinode list to be the rightmost leaf node
3543 * then the logic here makes sense.
3545 static void ocfs2_figure_appending_type(struct ocfs2_insert_type
*insert
,
3546 struct ocfs2_extent_list
*el
,
3547 struct ocfs2_extent_rec
*insert_rec
)
3550 u32 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
3551 struct ocfs2_extent_rec
*rec
;
3553 insert
->ins_appending
= APPEND_NONE
;
3555 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3557 if (!el
->l_next_free_rec
)
3558 goto set_tail_append
;
3560 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
3561 /* Were all records empty? */
3562 if (le16_to_cpu(el
->l_next_free_rec
) == 1)
3563 goto set_tail_append
;
3566 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3567 rec
= &el
->l_recs
[i
];
3570 (le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)))
3571 goto set_tail_append
;
3576 insert
->ins_appending
= APPEND_TAIL
;
3580 * Helper function called at the begining of an insert.
3582 * This computes a few things that are commonly used in the process of
3583 * inserting into the btree:
3584 * - Whether the new extent is contiguous with an existing one.
3585 * - The current tree depth.
3586 * - Whether the insert is an appending one.
3587 * - The total # of free records in the tree.
3589 * All of the information is stored on the ocfs2_insert_type
3592 static int ocfs2_figure_insert_type(struct inode
*inode
,
3593 struct buffer_head
*di_bh
,
3594 struct buffer_head
**last_eb_bh
,
3595 struct ocfs2_extent_rec
*insert_rec
,
3596 struct ocfs2_insert_type
*insert
)
3599 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
3600 struct ocfs2_extent_block
*eb
;
3601 struct ocfs2_extent_list
*el
;
3602 struct ocfs2_path
*path
= NULL
;
3603 struct buffer_head
*bh
= NULL
;
3605 insert
->ins_split
= SPLIT_NONE
;
3607 el
= &di
->id2
.i_list
;
3608 insert
->ins_tree_depth
= le16_to_cpu(el
->l_tree_depth
);
3610 if (el
->l_tree_depth
) {
3612 * If we have tree depth, we read in the
3613 * rightmost extent block ahead of time as
3614 * ocfs2_figure_insert_type() and ocfs2_add_branch()
3615 * may want it later.
3617 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
3618 le64_to_cpu(di
->i_last_eb_blk
), &bh
,
3619 OCFS2_BH_CACHED
, inode
);
3624 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
3629 * Unless we have a contiguous insert, we'll need to know if
3630 * there is room left in our allocation tree for another
3633 * XXX: This test is simplistic, we can search for empty
3634 * extent records too.
3636 insert
->ins_free_records
= le16_to_cpu(el
->l_count
) -
3637 le16_to_cpu(el
->l_next_free_rec
);
3639 if (!insert
->ins_tree_depth
) {
3640 ocfs2_figure_contig_type(inode
, insert
, el
, insert_rec
);
3641 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
3645 path
= ocfs2_new_inode_path(di_bh
);
3653 * In the case that we're inserting past what the tree
3654 * currently accounts for, ocfs2_find_path() will return for
3655 * us the rightmost tree path. This is accounted for below in
3656 * the appending code.
3658 ret
= ocfs2_find_path(inode
, path
, le32_to_cpu(insert_rec
->e_cpos
));
3664 el
= path_leaf_el(path
);
3667 * Now that we have the path, there's two things we want to determine:
3668 * 1) Contiguousness (also set contig_index if this is so)
3670 * 2) Are we doing an append? We can trivially break this up
3671 * into two types of appends: simple record append, or a
3672 * rotate inside the tail leaf.
3674 ocfs2_figure_contig_type(inode
, insert
, el
, insert_rec
);
3677 * The insert code isn't quite ready to deal with all cases of
3678 * left contiguousness. Specifically, if it's an insert into
3679 * the 1st record in a leaf, it will require the adjustment of
3680 * cluster count on the last record of the path directly to it's
3681 * left. For now, just catch that case and fool the layers
3682 * above us. This works just fine for tree_depth == 0, which
3683 * is why we allow that above.
3685 if (insert
->ins_contig
== CONTIG_LEFT
&&
3686 insert
->ins_contig_index
== 0)
3687 insert
->ins_contig
= CONTIG_NONE
;
3690 * Ok, so we can simply compare against last_eb to figure out
3691 * whether the path doesn't exist. This will only happen in
3692 * the case that we're doing a tail append, so maybe we can
3693 * take advantage of that information somehow.
3695 if (le64_to_cpu(di
->i_last_eb_blk
) == path_leaf_bh(path
)->b_blocknr
) {
3697 * Ok, ocfs2_find_path() returned us the rightmost
3698 * tree path. This might be an appending insert. There are
3700 * 1) We're doing a true append at the tail:
3701 * -This might even be off the end of the leaf
3702 * 2) We're "appending" by rotating in the tail
3704 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
3708 ocfs2_free_path(path
);
3718 * Insert an extent into an inode btree.
3720 * The caller needs to update fe->i_clusters
3722 int ocfs2_insert_extent(struct ocfs2_super
*osb
,
3724 struct inode
*inode
,
3725 struct buffer_head
*fe_bh
,
3730 struct ocfs2_alloc_context
*meta_ac
)
3733 struct buffer_head
*last_eb_bh
= NULL
;
3734 struct buffer_head
*bh
= NULL
;
3735 struct ocfs2_insert_type insert
= {0, };
3736 struct ocfs2_extent_rec rec
;
3738 mlog(0, "add %u clusters at position %u to inode %llu\n",
3739 new_clusters
, cpos
, (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
3741 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb
) &&
3742 (OCFS2_I(inode
)->ip_clusters
!= cpos
),
3743 "Device %s, asking for sparse allocation: inode %llu, "
3744 "cpos %u, clusters %u\n",
3746 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
,
3747 OCFS2_I(inode
)->ip_clusters
);
3749 memset(&rec
, 0, sizeof(rec
));
3750 rec
.e_cpos
= cpu_to_le32(cpos
);
3751 rec
.e_blkno
= cpu_to_le64(start_blk
);
3752 rec
.e_leaf_clusters
= cpu_to_le16(new_clusters
);
3753 rec
.e_flags
= flags
;
3755 status
= ocfs2_figure_insert_type(inode
, fe_bh
, &last_eb_bh
, &rec
,
3762 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
3763 "Insert.contig_index: %d, Insert.free_records: %d, "
3764 "Insert.tree_depth: %d\n",
3765 insert
.ins_appending
, insert
.ins_contig
, insert
.ins_contig_index
,
3766 insert
.ins_free_records
, insert
.ins_tree_depth
);
3768 if (insert
.ins_contig
== CONTIG_NONE
&& insert
.ins_free_records
== 0) {
3769 status
= ocfs2_grow_tree(inode
, handle
, fe_bh
,
3770 &insert
.ins_tree_depth
, &last_eb_bh
,
3778 /* Finally, we can add clusters. This might rotate the tree for us. */
3779 status
= ocfs2_do_insert_extent(inode
, handle
, fe_bh
, &rec
, &insert
);
3783 ocfs2_extent_map_insert_rec(inode
, &rec
);
3796 static void ocfs2_make_right_split_rec(struct super_block
*sb
,
3797 struct ocfs2_extent_rec
*split_rec
,
3799 struct ocfs2_extent_rec
*rec
)
3801 u32 rec_cpos
= le32_to_cpu(rec
->e_cpos
);
3802 u32 rec_range
= rec_cpos
+ le16_to_cpu(rec
->e_leaf_clusters
);
3804 memset(split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
3806 split_rec
->e_cpos
= cpu_to_le32(cpos
);
3807 split_rec
->e_leaf_clusters
= cpu_to_le16(rec_range
- cpos
);
3809 split_rec
->e_blkno
= rec
->e_blkno
;
3810 le64_add_cpu(&split_rec
->e_blkno
,
3811 ocfs2_clusters_to_blocks(sb
, cpos
- rec_cpos
));
3813 split_rec
->e_flags
= rec
->e_flags
;
3816 static int ocfs2_split_and_insert(struct inode
*inode
,
3818 struct ocfs2_path
*path
,
3819 struct buffer_head
*di_bh
,
3820 struct buffer_head
**last_eb_bh
,
3822 struct ocfs2_extent_rec
*orig_split_rec
,
3823 struct ocfs2_alloc_context
*meta_ac
)
3826 unsigned int insert_range
, rec_range
, do_leftright
= 0;
3827 struct ocfs2_extent_rec tmprec
;
3828 struct ocfs2_extent_list
*rightmost_el
;
3829 struct ocfs2_extent_rec rec
;
3830 struct ocfs2_extent_rec split_rec
= *orig_split_rec
;
3831 struct ocfs2_insert_type insert
;
3832 struct ocfs2_extent_block
*eb
;
3833 struct ocfs2_dinode
*di
;
3837 * Store a copy of the record on the stack - it might move
3838 * around as the tree is manipulated below.
3840 rec
= path_leaf_el(path
)->l_recs
[split_index
];
3842 di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
3843 rightmost_el
= &di
->id2
.i_list
;
3845 depth
= le16_to_cpu(rightmost_el
->l_tree_depth
);
3847 BUG_ON(!(*last_eb_bh
));
3848 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
3849 rightmost_el
= &eb
->h_list
;
3852 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
3853 le16_to_cpu(rightmost_el
->l_count
)) {
3854 int old_depth
= depth
;
3856 ret
= ocfs2_grow_tree(inode
, handle
, di_bh
, &depth
, last_eb_bh
,
3863 if (old_depth
!= depth
) {
3864 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
3865 rightmost_el
= &eb
->h_list
;
3869 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
3870 insert
.ins_appending
= APPEND_NONE
;
3871 insert
.ins_contig
= CONTIG_NONE
;
3872 insert
.ins_free_records
= le16_to_cpu(rightmost_el
->l_count
)
3873 - le16_to_cpu(rightmost_el
->l_next_free_rec
);
3874 insert
.ins_tree_depth
= depth
;
3876 insert_range
= le32_to_cpu(split_rec
.e_cpos
) +
3877 le16_to_cpu(split_rec
.e_leaf_clusters
);
3878 rec_range
= le32_to_cpu(rec
.e_cpos
) +
3879 le16_to_cpu(rec
.e_leaf_clusters
);
3881 if (split_rec
.e_cpos
== rec
.e_cpos
) {
3882 insert
.ins_split
= SPLIT_LEFT
;
3883 } else if (insert_range
== rec_range
) {
3884 insert
.ins_split
= SPLIT_RIGHT
;
3887 * Left/right split. We fake this as a right split
3888 * first and then make a second pass as a left split.
3890 insert
.ins_split
= SPLIT_RIGHT
;
3892 ocfs2_make_right_split_rec(inode
->i_sb
, &tmprec
, insert_range
,
3897 BUG_ON(do_leftright
);
3901 ret
= ocfs2_do_insert_extent(inode
, handle
, di_bh
, &split_rec
,
3908 if (do_leftright
== 1) {
3910 struct ocfs2_extent_list
*el
;
3913 split_rec
= *orig_split_rec
;
3915 ocfs2_reinit_path(path
, 1);
3917 cpos
= le32_to_cpu(split_rec
.e_cpos
);
3918 ret
= ocfs2_find_path(inode
, path
, cpos
);
3924 el
= path_leaf_el(path
);
3925 split_index
= ocfs2_search_extent_list(el
, cpos
);
3934 * Mark part or all of the extent record at split_index in the leaf
3935 * pointed to by path as written. This removes the unwritten
3938 * Care is taken to handle contiguousness so as to not grow the tree.
3940 * meta_ac is not strictly necessary - we only truly need it if growth
3941 * of the tree is required. All other cases will degrade into a less
3942 * optimal tree layout.
3944 * last_eb_bh should be the rightmost leaf block for any inode with a
3945 * 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.
3947 * This code is optimized for readability - several passes might be
3948 * made over certain portions of the tree. All of those blocks will
3949 * have been brought into cache (and pinned via the journal), so the
3950 * extra overhead is not expressed in terms of disk reads.
3952 static int __ocfs2_mark_extent_written(struct inode
*inode
,
3953 struct buffer_head
*di_bh
,
3955 struct ocfs2_path
*path
,
3957 struct ocfs2_extent_rec
*split_rec
,
3958 struct ocfs2_alloc_context
*meta_ac
,
3959 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3962 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
3963 struct buffer_head
*eb_bh
, *last_eb_bh
= NULL
;
3964 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
3965 struct ocfs2_merge_ctxt ctxt
;
3966 struct ocfs2_extent_list
*rightmost_el
;
3968 if (!rec
->e_flags
& OCFS2_EXT_UNWRITTEN
) {
3974 if (le32_to_cpu(rec
->e_cpos
) > le32_to_cpu(split_rec
->e_cpos
) ||
3975 ((le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)) <
3976 (le32_to_cpu(split_rec
->e_cpos
) + le16_to_cpu(split_rec
->e_leaf_clusters
)))) {
3982 eb_bh
= path_leaf_bh(path
);
3983 ret
= ocfs2_journal_access(handle
, inode
, eb_bh
,
3984 OCFS2_JOURNAL_ACCESS_WRITE
);
3990 ctxt
.c_contig_type
= ocfs2_figure_merge_contig_type(inode
, el
,
3995 * The core merge / split code wants to know how much room is
3996 * left in this inodes allocation tree, so we pass the
3997 * rightmost extent list.
3999 if (path
->p_tree_depth
) {
4000 struct ocfs2_extent_block
*eb
;
4001 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
4003 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
4004 le64_to_cpu(di
->i_last_eb_blk
),
4005 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
4011 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
4012 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
4013 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
4018 rightmost_el
= &eb
->h_list
;
4020 rightmost_el
= path_root_el(path
);
4022 ctxt
.c_used_tail_recs
= le16_to_cpu(rightmost_el
->l_next_free_rec
);
4023 if (ctxt
.c_used_tail_recs
> 0 &&
4024 ocfs2_is_empty_extent(&rightmost_el
->l_recs
[0]))
4025 ctxt
.c_used_tail_recs
--;
4027 if (rec
->e_cpos
== split_rec
->e_cpos
&&
4028 rec
->e_leaf_clusters
== split_rec
->e_leaf_clusters
)
4029 ctxt
.c_split_covers_rec
= 1;
4031 ctxt
.c_split_covers_rec
= 0;
4033 ctxt
.c_has_empty_extent
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
4035 mlog(0, "index: %d, contig: %u, used_tail_recs: %u, "
4036 "has_empty: %u, split_covers: %u\n", split_index
,
4037 ctxt
.c_contig_type
, ctxt
.c_used_tail_recs
,
4038 ctxt
.c_has_empty_extent
, ctxt
.c_split_covers_rec
);
4040 if (ctxt
.c_contig_type
== CONTIG_NONE
) {
4041 if (ctxt
.c_split_covers_rec
)
4042 el
->l_recs
[split_index
] = *split_rec
;
4044 ret
= ocfs2_split_and_insert(inode
, handle
, path
, di_bh
,
4045 &last_eb_bh
, split_index
,
4046 split_rec
, meta_ac
);
4050 ret
= ocfs2_try_to_merge_extent(inode
, handle
, path
,
4051 split_index
, split_rec
,
4057 ocfs2_journal_dirty(handle
, eb_bh
);
4065 * Mark the already-existing extent at cpos as written for len clusters.
4067 * If the existing extent is larger than the request, initiate a
4068 * split. An attempt will be made at merging with adjacent extents.
4070 * The caller is responsible for passing down meta_ac if we'll need it.
4072 int ocfs2_mark_extent_written(struct inode
*inode
, struct buffer_head
*di_bh
,
4073 handle_t
*handle
, u32 cpos
, u32 len
, u32 phys
,
4074 struct ocfs2_alloc_context
*meta_ac
,
4075 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
4078 u64 start_blkno
= ocfs2_clusters_to_blocks(inode
->i_sb
, phys
);
4079 struct ocfs2_extent_rec split_rec
;
4080 struct ocfs2_path
*left_path
= NULL
;
4081 struct ocfs2_extent_list
*el
;
4083 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4084 inode
->i_ino
, cpos
, len
, phys
, (unsigned long long)start_blkno
);
4086 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode
->i_sb
))) {
4087 ocfs2_error(inode
->i_sb
, "Inode %llu has unwritten extents "
4088 "that are being written to, but the feature bit "
4089 "is not set in the super block.",
4090 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
4096 * XXX: This should be fixed up so that we just re-insert the
4097 * next extent records.
4099 ocfs2_extent_map_trunc(inode
, 0);
4101 left_path
= ocfs2_new_inode_path(di_bh
);
4108 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
4113 el
= path_leaf_el(left_path
);
4115 index
= ocfs2_search_extent_list(el
, cpos
);
4116 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
4117 ocfs2_error(inode
->i_sb
,
4118 "Inode %llu has an extent at cpos %u which can no "
4119 "longer be found.\n",
4120 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
);
4125 memset(&split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4126 split_rec
.e_cpos
= cpu_to_le32(cpos
);
4127 split_rec
.e_leaf_clusters
= cpu_to_le16(len
);
4128 split_rec
.e_blkno
= cpu_to_le64(start_blkno
);
4129 split_rec
.e_flags
= path_leaf_el(left_path
)->l_recs
[index
].e_flags
;
4130 split_rec
.e_flags
&= ~OCFS2_EXT_UNWRITTEN
;
4132 ret
= __ocfs2_mark_extent_written(inode
, di_bh
, handle
, left_path
,
4133 index
, &split_rec
, meta_ac
, dealloc
);
4138 ocfs2_free_path(left_path
);
4142 static inline int ocfs2_truncate_log_needs_flush(struct ocfs2_super
*osb
)
4144 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
4145 struct ocfs2_dinode
*di
;
4146 struct ocfs2_truncate_log
*tl
;
4148 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4149 tl
= &di
->id2
.i_dealloc
;
4151 mlog_bug_on_msg(le16_to_cpu(tl
->tl_used
) > le16_to_cpu(tl
->tl_count
),
4152 "slot %d, invalid truncate log parameters: used = "
4153 "%u, count = %u\n", osb
->slot_num
,
4154 le16_to_cpu(tl
->tl_used
), le16_to_cpu(tl
->tl_count
));
4155 return le16_to_cpu(tl
->tl_used
) == le16_to_cpu(tl
->tl_count
);
4158 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log
*tl
,
4159 unsigned int new_start
)
4161 unsigned int tail_index
;
4162 unsigned int current_tail
;
4164 /* No records, nothing to coalesce */
4165 if (!le16_to_cpu(tl
->tl_used
))
4168 tail_index
= le16_to_cpu(tl
->tl_used
) - 1;
4169 current_tail
= le32_to_cpu(tl
->tl_recs
[tail_index
].t_start
);
4170 current_tail
+= le32_to_cpu(tl
->tl_recs
[tail_index
].t_clusters
);
4172 return current_tail
== new_start
;
4175 static int ocfs2_truncate_log_append(struct ocfs2_super
*osb
,
4178 unsigned int num_clusters
)
4181 unsigned int start_cluster
, tl_count
;
4182 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4183 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
4184 struct ocfs2_dinode
*di
;
4185 struct ocfs2_truncate_log
*tl
;
4187 mlog_entry("start_blk = %llu, num_clusters = %u\n",
4188 (unsigned long long)start_blk
, num_clusters
);
4190 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
4192 start_cluster
= ocfs2_blocks_to_clusters(osb
->sb
, start_blk
);
4194 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4195 tl
= &di
->id2
.i_dealloc
;
4196 if (!OCFS2_IS_VALID_DINODE(di
)) {
4197 OCFS2_RO_ON_INVALID_DINODE(osb
->sb
, di
);
4202 tl_count
= le16_to_cpu(tl
->tl_count
);
4203 mlog_bug_on_msg(tl_count
> ocfs2_truncate_recs_per_inode(osb
->sb
) ||
4205 "Truncate record count on #%llu invalid "
4206 "wanted %u, actual %u\n",
4207 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
4208 ocfs2_truncate_recs_per_inode(osb
->sb
),
4209 le16_to_cpu(tl
->tl_count
));
4211 /* Caller should have known to flush before calling us. */
4212 index
= le16_to_cpu(tl
->tl_used
);
4213 if (index
>= tl_count
) {
4219 status
= ocfs2_journal_access(handle
, tl_inode
, tl_bh
,
4220 OCFS2_JOURNAL_ACCESS_WRITE
);
4226 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
4227 "%llu (index = %d)\n", num_clusters
, start_cluster
,
4228 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
, index
);
4230 if (ocfs2_truncate_log_can_coalesce(tl
, start_cluster
)) {
4232 * Move index back to the record we are coalescing with.
4233 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
4237 num_clusters
+= le32_to_cpu(tl
->tl_recs
[index
].t_clusters
);
4238 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
4239 index
, le32_to_cpu(tl
->tl_recs
[index
].t_start
),
4242 tl
->tl_recs
[index
].t_start
= cpu_to_le32(start_cluster
);
4243 tl
->tl_used
= cpu_to_le16(index
+ 1);
4245 tl
->tl_recs
[index
].t_clusters
= cpu_to_le32(num_clusters
);
4247 status
= ocfs2_journal_dirty(handle
, tl_bh
);
4258 static int ocfs2_replay_truncate_records(struct ocfs2_super
*osb
,
4260 struct inode
*data_alloc_inode
,
4261 struct buffer_head
*data_alloc_bh
)
4265 unsigned int num_clusters
;
4267 struct ocfs2_truncate_rec rec
;
4268 struct ocfs2_dinode
*di
;
4269 struct ocfs2_truncate_log
*tl
;
4270 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4271 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
4275 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4276 tl
= &di
->id2
.i_dealloc
;
4277 i
= le16_to_cpu(tl
->tl_used
) - 1;
4279 /* Caller has given us at least enough credits to
4280 * update the truncate log dinode */
4281 status
= ocfs2_journal_access(handle
, tl_inode
, tl_bh
,
4282 OCFS2_JOURNAL_ACCESS_WRITE
);
4288 tl
->tl_used
= cpu_to_le16(i
);
4290 status
= ocfs2_journal_dirty(handle
, tl_bh
);
4296 /* TODO: Perhaps we can calculate the bulk of the
4297 * credits up front rather than extending like
4299 status
= ocfs2_extend_trans(handle
,
4300 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC
);
4306 rec
= tl
->tl_recs
[i
];
4307 start_blk
= ocfs2_clusters_to_blocks(data_alloc_inode
->i_sb
,
4308 le32_to_cpu(rec
.t_start
));
4309 num_clusters
= le32_to_cpu(rec
.t_clusters
);
4311 /* if start_blk is not set, we ignore the record as
4314 mlog(0, "free record %d, start = %u, clusters = %u\n",
4315 i
, le32_to_cpu(rec
.t_start
), num_clusters
);
4317 status
= ocfs2_free_clusters(handle
, data_alloc_inode
,
4318 data_alloc_bh
, start_blk
,
4333 /* Expects you to already be holding tl_inode->i_mutex */
4334 static int __ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
4337 unsigned int num_to_flush
;
4339 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4340 struct inode
*data_alloc_inode
= NULL
;
4341 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
4342 struct buffer_head
*data_alloc_bh
= NULL
;
4343 struct ocfs2_dinode
*di
;
4344 struct ocfs2_truncate_log
*tl
;
4348 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
4350 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4351 tl
= &di
->id2
.i_dealloc
;
4352 if (!OCFS2_IS_VALID_DINODE(di
)) {
4353 OCFS2_RO_ON_INVALID_DINODE(osb
->sb
, di
);
4358 num_to_flush
= le16_to_cpu(tl
->tl_used
);
4359 mlog(0, "Flush %u records from truncate log #%llu\n",
4360 num_to_flush
, (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
);
4361 if (!num_to_flush
) {
4366 data_alloc_inode
= ocfs2_get_system_file_inode(osb
,
4367 GLOBAL_BITMAP_SYSTEM_INODE
,
4368 OCFS2_INVALID_SLOT
);
4369 if (!data_alloc_inode
) {
4371 mlog(ML_ERROR
, "Could not get bitmap inode!\n");
4375 mutex_lock(&data_alloc_inode
->i_mutex
);
4377 status
= ocfs2_meta_lock(data_alloc_inode
, &data_alloc_bh
, 1);
4383 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
4384 if (IS_ERR(handle
)) {
4385 status
= PTR_ERR(handle
);
4390 status
= ocfs2_replay_truncate_records(osb
, handle
, data_alloc_inode
,
4395 ocfs2_commit_trans(osb
, handle
);
4398 brelse(data_alloc_bh
);
4399 ocfs2_meta_unlock(data_alloc_inode
, 1);
4402 mutex_unlock(&data_alloc_inode
->i_mutex
);
4403 iput(data_alloc_inode
);
4410 int ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
4413 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4415 mutex_lock(&tl_inode
->i_mutex
);
4416 status
= __ocfs2_flush_truncate_log(osb
);
4417 mutex_unlock(&tl_inode
->i_mutex
);
4422 static void ocfs2_truncate_log_worker(struct work_struct
*work
)
4425 struct ocfs2_super
*osb
=
4426 container_of(work
, struct ocfs2_super
,
4427 osb_truncate_log_wq
.work
);
4431 status
= ocfs2_flush_truncate_log(osb
);
4438 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
4439 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super
*osb
,
4442 if (osb
->osb_tl_inode
) {
4443 /* We want to push off log flushes while truncates are
4446 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
4448 queue_delayed_work(ocfs2_wq
, &osb
->osb_truncate_log_wq
,
4449 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL
);
4453 static int ocfs2_get_truncate_log_info(struct ocfs2_super
*osb
,
4455 struct inode
**tl_inode
,
4456 struct buffer_head
**tl_bh
)
4459 struct inode
*inode
= NULL
;
4460 struct buffer_head
*bh
= NULL
;
4462 inode
= ocfs2_get_system_file_inode(osb
,
4463 TRUNCATE_LOG_SYSTEM_INODE
,
4467 mlog(ML_ERROR
, "Could not get load truncate log inode!\n");
4471 status
= ocfs2_read_block(osb
, OCFS2_I(inode
)->ip_blkno
, &bh
,
4472 OCFS2_BH_CACHED
, inode
);
4486 /* called during the 1st stage of node recovery. we stamp a clean
4487 * truncate log and pass back a copy for processing later. if the
4488 * truncate log does not require processing, a *tl_copy is set to
4490 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super
*osb
,
4492 struct ocfs2_dinode
**tl_copy
)
4495 struct inode
*tl_inode
= NULL
;
4496 struct buffer_head
*tl_bh
= NULL
;
4497 struct ocfs2_dinode
*di
;
4498 struct ocfs2_truncate_log
*tl
;
4502 mlog(0, "recover truncate log from slot %d\n", slot_num
);
4504 status
= ocfs2_get_truncate_log_info(osb
, slot_num
, &tl_inode
, &tl_bh
);
4510 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
4511 tl
= &di
->id2
.i_dealloc
;
4512 if (!OCFS2_IS_VALID_DINODE(di
)) {
4513 OCFS2_RO_ON_INVALID_DINODE(tl_inode
->i_sb
, di
);
4518 if (le16_to_cpu(tl
->tl_used
)) {
4519 mlog(0, "We'll have %u logs to recover\n",
4520 le16_to_cpu(tl
->tl_used
));
4522 *tl_copy
= kmalloc(tl_bh
->b_size
, GFP_KERNEL
);
4529 /* Assuming the write-out below goes well, this copy
4530 * will be passed back to recovery for processing. */
4531 memcpy(*tl_copy
, tl_bh
->b_data
, tl_bh
->b_size
);
4533 /* All we need to do to clear the truncate log is set
4537 status
= ocfs2_write_block(osb
, tl_bh
, tl_inode
);
4550 if (status
< 0 && (*tl_copy
)) {
4559 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super
*osb
,
4560 struct ocfs2_dinode
*tl_copy
)
4564 unsigned int clusters
, num_recs
, start_cluster
;
4567 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4568 struct ocfs2_truncate_log
*tl
;
4572 if (OCFS2_I(tl_inode
)->ip_blkno
== le64_to_cpu(tl_copy
->i_blkno
)) {
4573 mlog(ML_ERROR
, "Asked to recover my own truncate log!\n");
4577 tl
= &tl_copy
->id2
.i_dealloc
;
4578 num_recs
= le16_to_cpu(tl
->tl_used
);
4579 mlog(0, "cleanup %u records from %llu\n", num_recs
,
4580 (unsigned long long)le64_to_cpu(tl_copy
->i_blkno
));
4582 mutex_lock(&tl_inode
->i_mutex
);
4583 for(i
= 0; i
< num_recs
; i
++) {
4584 if (ocfs2_truncate_log_needs_flush(osb
)) {
4585 status
= __ocfs2_flush_truncate_log(osb
);
4592 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
4593 if (IS_ERR(handle
)) {
4594 status
= PTR_ERR(handle
);
4599 clusters
= le32_to_cpu(tl
->tl_recs
[i
].t_clusters
);
4600 start_cluster
= le32_to_cpu(tl
->tl_recs
[i
].t_start
);
4601 start_blk
= ocfs2_clusters_to_blocks(osb
->sb
, start_cluster
);
4603 status
= ocfs2_truncate_log_append(osb
, handle
,
4604 start_blk
, clusters
);
4605 ocfs2_commit_trans(osb
, handle
);
4613 mutex_unlock(&tl_inode
->i_mutex
);
4619 void ocfs2_truncate_log_shutdown(struct ocfs2_super
*osb
)
4622 struct inode
*tl_inode
= osb
->osb_tl_inode
;
4627 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
4628 flush_workqueue(ocfs2_wq
);
4630 status
= ocfs2_flush_truncate_log(osb
);
4634 brelse(osb
->osb_tl_bh
);
4635 iput(osb
->osb_tl_inode
);
4641 int ocfs2_truncate_log_init(struct ocfs2_super
*osb
)
4644 struct inode
*tl_inode
= NULL
;
4645 struct buffer_head
*tl_bh
= NULL
;
4649 status
= ocfs2_get_truncate_log_info(osb
,
4656 /* ocfs2_truncate_log_shutdown keys on the existence of
4657 * osb->osb_tl_inode so we don't set any of the osb variables
4658 * until we're sure all is well. */
4659 INIT_DELAYED_WORK(&osb
->osb_truncate_log_wq
,
4660 ocfs2_truncate_log_worker
);
4661 osb
->osb_tl_bh
= tl_bh
;
4662 osb
->osb_tl_inode
= tl_inode
;
4669 * Delayed de-allocation of suballocator blocks.
4671 * Some sets of block de-allocations might involve multiple suballocator inodes.
4673 * The locking for this can get extremely complicated, especially when
4674 * the suballocator inodes to delete from aren't known until deep
4675 * within an unrelated codepath.
4677 * ocfs2_extent_block structures are a good example of this - an inode
4678 * btree could have been grown by any number of nodes each allocating
4679 * out of their own suballoc inode.
4681 * These structures allow the delay of block de-allocation until a
4682 * later time, when locking of multiple cluster inodes won't cause
4687 * Describes a single block free from a suballocator
4689 struct ocfs2_cached_block_free
{
4690 struct ocfs2_cached_block_free
*free_next
;
4692 unsigned int free_bit
;
4695 struct ocfs2_per_slot_free_list
{
4696 struct ocfs2_per_slot_free_list
*f_next_suballocator
;
4699 struct ocfs2_cached_block_free
*f_first
;
4702 static int ocfs2_free_cached_items(struct ocfs2_super
*osb
,
4705 struct ocfs2_cached_block_free
*head
)
4710 struct inode
*inode
;
4711 struct buffer_head
*di_bh
= NULL
;
4712 struct ocfs2_cached_block_free
*tmp
;
4714 inode
= ocfs2_get_system_file_inode(osb
, sysfile_type
, slot
);
4721 mutex_lock(&inode
->i_mutex
);
4723 ret
= ocfs2_meta_lock(inode
, &di_bh
, 1);
4729 handle
= ocfs2_start_trans(osb
, OCFS2_SUBALLOC_FREE
);
4730 if (IS_ERR(handle
)) {
4731 ret
= PTR_ERR(handle
);
4737 bg_blkno
= ocfs2_which_suballoc_group(head
->free_blk
,
4739 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
4740 head
->free_bit
, (unsigned long long)head
->free_blk
);
4742 ret
= ocfs2_free_suballoc_bits(handle
, inode
, di_bh
,
4743 head
->free_bit
, bg_blkno
, 1);
4749 ret
= ocfs2_extend_trans(handle
, OCFS2_SUBALLOC_FREE
);
4756 head
= head
->free_next
;
4761 ocfs2_commit_trans(osb
, handle
);
4764 ocfs2_meta_unlock(inode
, 1);
4767 mutex_unlock(&inode
->i_mutex
);
4771 /* Premature exit may have left some dangling items. */
4773 head
= head
->free_next
;
4780 int ocfs2_run_deallocs(struct ocfs2_super
*osb
,
4781 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
4784 struct ocfs2_per_slot_free_list
*fl
;
4789 while (ctxt
->c_first_suballocator
) {
4790 fl
= ctxt
->c_first_suballocator
;
4793 mlog(0, "Free items: (type %u, slot %d)\n",
4794 fl
->f_inode_type
, fl
->f_slot
);
4795 ret2
= ocfs2_free_cached_items(osb
, fl
->f_inode_type
,
4796 fl
->f_slot
, fl
->f_first
);
4803 ctxt
->c_first_suballocator
= fl
->f_next_suballocator
;
4810 static struct ocfs2_per_slot_free_list
*
4811 ocfs2_find_per_slot_free_list(int type
,
4813 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
4815 struct ocfs2_per_slot_free_list
*fl
= ctxt
->c_first_suballocator
;
4818 if (fl
->f_inode_type
== type
&& fl
->f_slot
== slot
)
4821 fl
= fl
->f_next_suballocator
;
4824 fl
= kmalloc(sizeof(*fl
), GFP_NOFS
);
4826 fl
->f_inode_type
= type
;
4829 fl
->f_next_suballocator
= ctxt
->c_first_suballocator
;
4831 ctxt
->c_first_suballocator
= fl
;
4836 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
4837 int type
, int slot
, u64 blkno
,
4841 struct ocfs2_per_slot_free_list
*fl
;
4842 struct ocfs2_cached_block_free
*item
;
4844 fl
= ocfs2_find_per_slot_free_list(type
, slot
, ctxt
);
4851 item
= kmalloc(sizeof(*item
), GFP_NOFS
);
4858 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
4859 type
, slot
, bit
, (unsigned long long)blkno
);
4861 item
->free_blk
= blkno
;
4862 item
->free_bit
= bit
;
4863 item
->free_next
= fl
->f_first
;
4872 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
4873 struct ocfs2_extent_block
*eb
)
4875 return ocfs2_cache_block_dealloc(ctxt
, EXTENT_ALLOC_SYSTEM_INODE
,
4876 le16_to_cpu(eb
->h_suballoc_slot
),
4877 le64_to_cpu(eb
->h_blkno
),
4878 le16_to_cpu(eb
->h_suballoc_bit
));
4881 /* This function will figure out whether the currently last extent
4882 * block will be deleted, and if it will, what the new last extent
4883 * block will be so we can update his h_next_leaf_blk field, as well
4884 * as the dinodes i_last_eb_blk */
4885 static int ocfs2_find_new_last_ext_blk(struct inode
*inode
,
4886 unsigned int clusters_to_del
,
4887 struct ocfs2_path
*path
,
4888 struct buffer_head
**new_last_eb
)
4890 int next_free
, ret
= 0;
4892 struct ocfs2_extent_rec
*rec
;
4893 struct ocfs2_extent_block
*eb
;
4894 struct ocfs2_extent_list
*el
;
4895 struct buffer_head
*bh
= NULL
;
4897 *new_last_eb
= NULL
;
4899 /* we have no tree, so of course, no last_eb. */
4900 if (!path
->p_tree_depth
)
4903 /* trunc to zero special case - this makes tree_depth = 0
4904 * regardless of what it is. */
4905 if (OCFS2_I(inode
)->ip_clusters
== clusters_to_del
)
4908 el
= path_leaf_el(path
);
4909 BUG_ON(!el
->l_next_free_rec
);
4912 * Make sure that this extent list will actually be empty
4913 * after we clear away the data. We can shortcut out if
4914 * there's more than one non-empty extent in the
4915 * list. Otherwise, a check of the remaining extent is
4918 next_free
= le16_to_cpu(el
->l_next_free_rec
);
4920 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
4924 /* We may have a valid extent in index 1, check it. */
4926 rec
= &el
->l_recs
[1];
4929 * Fall through - no more nonempty extents, so we want
4930 * to delete this leaf.
4936 rec
= &el
->l_recs
[0];
4941 * Check it we'll only be trimming off the end of this
4944 if (le16_to_cpu(rec
->e_leaf_clusters
) > clusters_to_del
)
4948 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
4954 ret
= ocfs2_find_leaf(inode
, path_root_el(path
), cpos
, &bh
);
4960 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
4962 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
4963 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
4969 get_bh(*new_last_eb
);
4970 mlog(0, "returning block %llu, (cpos: %u)\n",
4971 (unsigned long long)le64_to_cpu(eb
->h_blkno
), cpos
);
4979 * Trim some clusters off the rightmost edge of a tree. Only called
4982 * The caller needs to:
4983 * - start journaling of each path component.
4984 * - compute and fully set up any new last ext block
4986 static int ocfs2_trim_tree(struct inode
*inode
, struct ocfs2_path
*path
,
4987 handle_t
*handle
, struct ocfs2_truncate_context
*tc
,
4988 u32 clusters_to_del
, u64
*delete_start
)
4990 int ret
, i
, index
= path
->p_tree_depth
;
4993 struct buffer_head
*bh
;
4994 struct ocfs2_extent_list
*el
;
4995 struct ocfs2_extent_rec
*rec
;
4999 while (index
>= 0) {
5000 bh
= path
->p_node
[index
].bh
;
5001 el
= path
->p_node
[index
].el
;
5003 mlog(0, "traveling tree (index = %d, block = %llu)\n",
5004 index
, (unsigned long long)bh
->b_blocknr
);
5006 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
5009 (path
->p_tree_depth
- le16_to_cpu(el
->l_tree_depth
))) {
5010 ocfs2_error(inode
->i_sb
,
5011 "Inode %lu has invalid ext. block %llu",
5013 (unsigned long long)bh
->b_blocknr
);
5019 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
5020 rec
= &el
->l_recs
[i
];
5022 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
5023 "next = %u\n", i
, le32_to_cpu(rec
->e_cpos
),
5024 ocfs2_rec_clusters(el
, rec
),
5025 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
5026 le16_to_cpu(el
->l_next_free_rec
));
5028 BUG_ON(ocfs2_rec_clusters(el
, rec
) < clusters_to_del
);
5030 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
5032 * If the leaf block contains a single empty
5033 * extent and no records, we can just remove
5036 if (i
== 0 && ocfs2_is_empty_extent(rec
)) {
5038 sizeof(struct ocfs2_extent_rec
));
5039 el
->l_next_free_rec
= cpu_to_le16(0);
5045 * Remove any empty extents by shifting things
5046 * left. That should make life much easier on
5047 * the code below. This condition is rare
5048 * enough that we shouldn't see a performance
5051 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
5052 le16_add_cpu(&el
->l_next_free_rec
, -1);
5055 i
< le16_to_cpu(el
->l_next_free_rec
); i
++)
5056 el
->l_recs
[i
] = el
->l_recs
[i
+ 1];
5058 memset(&el
->l_recs
[i
], 0,
5059 sizeof(struct ocfs2_extent_rec
));
5062 * We've modified our extent list. The
5063 * simplest way to handle this change
5064 * is to being the search from the
5067 goto find_tail_record
;
5070 le16_add_cpu(&rec
->e_leaf_clusters
, -clusters_to_del
);
5073 * We'll use "new_edge" on our way back up the
5074 * tree to know what our rightmost cpos is.
5076 new_edge
= le16_to_cpu(rec
->e_leaf_clusters
);
5077 new_edge
+= le32_to_cpu(rec
->e_cpos
);
5080 * The caller will use this to delete data blocks.
5082 *delete_start
= le64_to_cpu(rec
->e_blkno
)
5083 + ocfs2_clusters_to_blocks(inode
->i_sb
,
5084 le16_to_cpu(rec
->e_leaf_clusters
));
5087 * If it's now empty, remove this record.
5089 if (le16_to_cpu(rec
->e_leaf_clusters
) == 0) {
5091 sizeof(struct ocfs2_extent_rec
));
5092 le16_add_cpu(&el
->l_next_free_rec
, -1);
5095 if (le64_to_cpu(rec
->e_blkno
) == deleted_eb
) {
5097 sizeof(struct ocfs2_extent_rec
));
5098 le16_add_cpu(&el
->l_next_free_rec
, -1);
5103 /* Can this actually happen? */
5104 if (le16_to_cpu(el
->l_next_free_rec
) == 0)
5108 * We never actually deleted any clusters
5109 * because our leaf was empty. There's no
5110 * reason to adjust the rightmost edge then.
5115 rec
->e_int_clusters
= cpu_to_le32(new_edge
);
5116 le32_add_cpu(&rec
->e_int_clusters
,
5117 -le32_to_cpu(rec
->e_cpos
));
5120 * A deleted child record should have been
5123 BUG_ON(le32_to_cpu(rec
->e_int_clusters
) == 0);
5127 ret
= ocfs2_journal_dirty(handle
, bh
);
5133 mlog(0, "extent list container %llu, after: record %d: "
5134 "(%u, %u, %llu), next = %u.\n",
5135 (unsigned long long)bh
->b_blocknr
, i
,
5136 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
),
5137 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
5138 le16_to_cpu(el
->l_next_free_rec
));
5141 * We must be careful to only attempt delete of an
5142 * extent block (and not the root inode block).
5144 if (index
> 0 && le16_to_cpu(el
->l_next_free_rec
) == 0) {
5145 struct ocfs2_extent_block
*eb
=
5146 (struct ocfs2_extent_block
*)bh
->b_data
;
5149 * Save this for use when processing the
5152 deleted_eb
= le64_to_cpu(eb
->h_blkno
);
5154 mlog(0, "deleting this extent block.\n");
5156 ocfs2_remove_from_cache(inode
, bh
);
5158 BUG_ON(ocfs2_rec_clusters(el
, &el
->l_recs
[0]));
5159 BUG_ON(le32_to_cpu(el
->l_recs
[0].e_cpos
));
5160 BUG_ON(le64_to_cpu(el
->l_recs
[0].e_blkno
));
5162 ret
= ocfs2_cache_extent_block_free(&tc
->tc_dealloc
, eb
);
5163 /* An error here is not fatal. */
5178 static int ocfs2_do_truncate(struct ocfs2_super
*osb
,
5179 unsigned int clusters_to_del
,
5180 struct inode
*inode
,
5181 struct buffer_head
*fe_bh
,
5183 struct ocfs2_truncate_context
*tc
,
5184 struct ocfs2_path
*path
)
5187 struct ocfs2_dinode
*fe
;
5188 struct ocfs2_extent_block
*last_eb
= NULL
;
5189 struct ocfs2_extent_list
*el
;
5190 struct buffer_head
*last_eb_bh
= NULL
;
5193 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
5195 status
= ocfs2_find_new_last_ext_blk(inode
, clusters_to_del
,
5203 * Each component will be touched, so we might as well journal
5204 * here to avoid having to handle errors later.
5206 status
= ocfs2_journal_access_path(inode
, handle
, path
);
5213 status
= ocfs2_journal_access(handle
, inode
, last_eb_bh
,
5214 OCFS2_JOURNAL_ACCESS_WRITE
);
5220 last_eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5223 el
= &(fe
->id2
.i_list
);
5226 * Lower levels depend on this never happening, but it's best
5227 * to check it up here before changing the tree.
5229 if (el
->l_tree_depth
&& el
->l_recs
[0].e_int_clusters
== 0) {
5230 ocfs2_error(inode
->i_sb
,
5231 "Inode %lu has an empty extent record, depth %u\n",
5232 inode
->i_ino
, le16_to_cpu(el
->l_tree_depth
));
5237 spin_lock(&OCFS2_I(inode
)->ip_lock
);
5238 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
) -
5240 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
5241 le32_add_cpu(&fe
->i_clusters
, -clusters_to_del
);
5243 status
= ocfs2_trim_tree(inode
, path
, handle
, tc
,
5244 clusters_to_del
, &delete_blk
);
5250 if (le32_to_cpu(fe
->i_clusters
) == 0) {
5251 /* trunc to zero is a special case. */
5252 el
->l_tree_depth
= 0;
5253 fe
->i_last_eb_blk
= 0;
5255 fe
->i_last_eb_blk
= last_eb
->h_blkno
;
5257 status
= ocfs2_journal_dirty(handle
, fe_bh
);
5264 /* If there will be a new last extent block, then by
5265 * definition, there cannot be any leaves to the right of
5267 last_eb
->h_next_leaf_blk
= 0;
5268 status
= ocfs2_journal_dirty(handle
, last_eb_bh
);
5276 status
= ocfs2_truncate_log_append(osb
, handle
, delete_blk
,
5290 static int ocfs2_writeback_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
5292 set_buffer_uptodate(bh
);
5293 mark_buffer_dirty(bh
);
5297 static int ocfs2_ordered_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
5299 set_buffer_uptodate(bh
);
5300 mark_buffer_dirty(bh
);
5301 return ocfs2_journal_dirty_data(handle
, bh
);
5304 static void ocfs2_zero_cluster_pages(struct inode
*inode
, loff_t isize
,
5305 struct page
**pages
, int numpages
,
5306 u64 phys
, handle_t
*handle
)
5308 int i
, ret
, partial
= 0;
5311 unsigned int from
, to
= PAGE_CACHE_SIZE
;
5312 struct super_block
*sb
= inode
->i_sb
;
5314 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb
)));
5319 from
= isize
& (PAGE_CACHE_SIZE
- 1); /* 1st page offset */
5320 if (PAGE_CACHE_SHIFT
> OCFS2_SB(sb
)->s_clustersize_bits
) {
5322 * Since 'from' has been capped to a value below page
5323 * size, this calculation won't be able to overflow
5326 to
= ocfs2_align_bytes_to_clusters(sb
, from
);
5329 * The truncate tail in this case should never contain
5330 * more than one page at maximum. The loop below also
5333 BUG_ON(numpages
!= 1);
5336 for(i
= 0; i
< numpages
; i
++) {
5339 BUG_ON(from
> PAGE_CACHE_SIZE
);
5340 BUG_ON(to
> PAGE_CACHE_SIZE
);
5342 ret
= ocfs2_map_page_blocks(page
, &phys
, inode
, from
, to
, 0);
5346 kaddr
= kmap_atomic(page
, KM_USER0
);
5347 memset(kaddr
+ from
, 0, to
- from
);
5348 kunmap_atomic(kaddr
, KM_USER0
);
5351 * Need to set the buffers we zero'd into uptodate
5352 * here if they aren't - ocfs2_map_page_blocks()
5353 * might've skipped some
5355 if (ocfs2_should_order_data(inode
)) {
5356 ret
= walk_page_buffers(handle
,
5359 ocfs2_ordered_zero_func
);
5363 ret
= walk_page_buffers(handle
, page_buffers(page
),
5365 ocfs2_writeback_zero_func
);
5371 SetPageUptodate(page
);
5373 flush_dcache_page(page
);
5376 * Every page after the 1st one should be completely zero'd.
5382 for (i
= 0; i
< numpages
; i
++) {
5385 mark_page_accessed(page
);
5386 page_cache_release(page
);
5391 static int ocfs2_grab_eof_pages(struct inode
*inode
, loff_t isize
, struct page
**pages
,
5392 int *num
, u64
*phys
)
5394 int i
, numpages
= 0, ret
= 0;
5395 unsigned int csize
= OCFS2_SB(inode
->i_sb
)->s_clustersize
;
5396 unsigned int ext_flags
;
5397 struct super_block
*sb
= inode
->i_sb
;
5398 struct address_space
*mapping
= inode
->i_mapping
;
5399 unsigned long index
;
5400 u64 next_cluster_bytes
;
5402 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb
)));
5404 /* Cluster boundary, so we don't need to grab any pages. */
5405 if ((isize
& (csize
- 1)) == 0)
5408 ret
= ocfs2_extent_map_get_blocks(inode
, isize
>> sb
->s_blocksize_bits
,
5409 phys
, NULL
, &ext_flags
);
5415 /* Tail is a hole. */
5419 /* Tail is marked as unwritten, we can count on write to zero
5421 if (ext_flags
& OCFS2_EXT_UNWRITTEN
)
5424 next_cluster_bytes
= ocfs2_align_bytes_to_clusters(inode
->i_sb
, isize
);
5425 index
= isize
>> PAGE_CACHE_SHIFT
;
5427 pages
[numpages
] = grab_cache_page(mapping
, index
);
5428 if (!pages
[numpages
]) {
5436 } while (index
< (next_cluster_bytes
>> PAGE_CACHE_SHIFT
));
5441 for (i
= 0; i
< numpages
; i
++) {
5443 unlock_page(pages
[i
]);
5444 page_cache_release(pages
[i
]);
5457 * Zero the area past i_size but still within an allocated
5458 * cluster. This avoids exposing nonzero data on subsequent file
5461 * We need to call this before i_size is updated on the inode because
5462 * otherwise block_write_full_page() will skip writeout of pages past
5463 * i_size. The new_i_size parameter is passed for this reason.
5465 int ocfs2_zero_tail_for_truncate(struct inode
*inode
, handle_t
*handle
,
5470 struct page
**pages
= NULL
;
5474 * File systems which don't support sparse files zero on every
5477 if (!ocfs2_sparse_alloc(OCFS2_SB(inode
->i_sb
)))
5480 pages
= kcalloc(ocfs2_pages_per_cluster(inode
->i_sb
),
5481 sizeof(struct page
*), GFP_NOFS
);
5482 if (pages
== NULL
) {
5488 ret
= ocfs2_grab_eof_pages(inode
, new_i_size
, pages
, &numpages
, &phys
);
5497 ocfs2_zero_cluster_pages(inode
, new_i_size
, pages
, numpages
, phys
,
5501 * Initiate writeout of the pages we zero'd here. We don't
5502 * wait on them - the truncate_inode_pages() call later will
5505 endbyte
= ocfs2_align_bytes_to_clusters(inode
->i_sb
, new_i_size
);
5506 ret
= do_sync_mapping_range(inode
->i_mapping
, new_i_size
,
5507 endbyte
- 1, SYNC_FILE_RANGE_WRITE
);
5519 * It is expected, that by the time you call this function,
5520 * inode->i_size and fe->i_size have been adjusted.
5522 * WARNING: This will kfree the truncate context
5524 int ocfs2_commit_truncate(struct ocfs2_super
*osb
,
5525 struct inode
*inode
,
5526 struct buffer_head
*fe_bh
,
5527 struct ocfs2_truncate_context
*tc
)
5529 int status
, i
, credits
, tl_sem
= 0;
5530 u32 clusters_to_del
, new_highest_cpos
, range
;
5531 struct ocfs2_extent_list
*el
;
5532 handle_t
*handle
= NULL
;
5533 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5534 struct ocfs2_path
*path
= NULL
;
5538 new_highest_cpos
= ocfs2_clusters_for_bytes(osb
->sb
,
5539 i_size_read(inode
));
5541 path
= ocfs2_new_inode_path(fe_bh
);
5548 ocfs2_extent_map_trunc(inode
, new_highest_cpos
);
5552 * Check that we still have allocation to delete.
5554 if (OCFS2_I(inode
)->ip_clusters
== 0) {
5560 * Truncate always works against the rightmost tree branch.
5562 status
= ocfs2_find_path(inode
, path
, UINT_MAX
);
5568 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
5569 OCFS2_I(inode
)->ip_clusters
, path
->p_tree_depth
);
5572 * By now, el will point to the extent list on the bottom most
5573 * portion of this tree. Only the tail record is considered in
5576 * We handle the following cases, in order:
5577 * - empty extent: delete the remaining branch
5578 * - remove the entire record
5579 * - remove a partial record
5580 * - no record needs to be removed (truncate has completed)
5582 el
= path_leaf_el(path
);
5583 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
5584 ocfs2_error(inode
->i_sb
,
5585 "Inode %llu has empty extent block at %llu\n",
5586 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
5587 (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
5592 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
5593 range
= le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
5594 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
5595 if (i
== 0 && ocfs2_is_empty_extent(&el
->l_recs
[i
])) {
5596 clusters_to_del
= 0;
5597 } else if (le32_to_cpu(el
->l_recs
[i
].e_cpos
) >= new_highest_cpos
) {
5598 clusters_to_del
= ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
5599 } else if (range
> new_highest_cpos
) {
5600 clusters_to_del
= (ocfs2_rec_clusters(el
, &el
->l_recs
[i
]) +
5601 le32_to_cpu(el
->l_recs
[i
].e_cpos
)) -
5608 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
5609 clusters_to_del
, (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
5611 BUG_ON(clusters_to_del
== 0);
5613 mutex_lock(&tl_inode
->i_mutex
);
5615 /* ocfs2_truncate_log_needs_flush guarantees us at least one
5616 * record is free for use. If there isn't any, we flush to get
5617 * an empty truncate log. */
5618 if (ocfs2_truncate_log_needs_flush(osb
)) {
5619 status
= __ocfs2_flush_truncate_log(osb
);
5626 credits
= ocfs2_calc_tree_trunc_credits(osb
->sb
, clusters_to_del
,
5627 (struct ocfs2_dinode
*)fe_bh
->b_data
,
5629 handle
= ocfs2_start_trans(osb
, credits
);
5630 if (IS_ERR(handle
)) {
5631 status
= PTR_ERR(handle
);
5637 status
= ocfs2_do_truncate(osb
, clusters_to_del
, inode
, fe_bh
, handle
,
5644 mutex_unlock(&tl_inode
->i_mutex
);
5647 ocfs2_commit_trans(osb
, handle
);
5650 ocfs2_reinit_path(path
, 1);
5653 * The check above will catch the case where we've truncated
5654 * away all allocation.
5660 ocfs2_schedule_truncate_log_flush(osb
, 1);
5663 mutex_unlock(&tl_inode
->i_mutex
);
5666 ocfs2_commit_trans(osb
, handle
);
5668 ocfs2_run_deallocs(osb
, &tc
->tc_dealloc
);
5670 ocfs2_free_path(path
);
5672 /* This will drop the ext_alloc cluster lock for us */
5673 ocfs2_free_truncate_context(tc
);
5680 * Expects the inode to already be locked.
5682 int ocfs2_prepare_truncate(struct ocfs2_super
*osb
,
5683 struct inode
*inode
,
5684 struct buffer_head
*fe_bh
,
5685 struct ocfs2_truncate_context
**tc
)
5688 unsigned int new_i_clusters
;
5689 struct ocfs2_dinode
*fe
;
5690 struct ocfs2_extent_block
*eb
;
5691 struct buffer_head
*last_eb_bh
= NULL
;
5697 new_i_clusters
= ocfs2_clusters_for_bytes(osb
->sb
,
5698 i_size_read(inode
));
5699 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
5701 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
5702 "%llu\n", le32_to_cpu(fe
->i_clusters
), new_i_clusters
,
5703 (unsigned long long)le64_to_cpu(fe
->i_size
));
5705 *tc
= kzalloc(sizeof(struct ocfs2_truncate_context
), GFP_KERNEL
);
5711 ocfs2_init_dealloc_ctxt(&(*tc
)->tc_dealloc
);
5713 if (fe
->id2
.i_list
.l_tree_depth
) {
5714 status
= ocfs2_read_block(osb
, le64_to_cpu(fe
->i_last_eb_blk
),
5715 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
5720 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5721 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
5722 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
5730 (*tc
)->tc_last_eb_bh
= last_eb_bh
;
5736 ocfs2_free_truncate_context(*tc
);
5743 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
)
5746 * The caller is responsible for completing deallocation
5747 * before freeing the context.
5749 if (tc
->tc_dealloc
.c_first_suballocator
!= NULL
)
5751 "Truncate completion has non-empty dealloc context\n");
5753 if (tc
->tc_last_eb_bh
)
5754 brelse(tc
->tc_last_eb_bh
);