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"
54 * Operations for a specific extent tree type.
56 * To implement an on-disk btree (extent tree) type in ocfs2, add
57 * an ocfs2_extent_tree_operations structure and the matching
58 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
59 * for the allocation portion of the extent tree.
61 struct ocfs2_extent_tree_operations
{
63 * last_eb_blk is the block number of the right most leaf extent
64 * block. Most on-disk structures containing an extent tree store
65 * this value for fast access. The ->eo_set_last_eb_blk() and
66 * ->eo_get_last_eb_blk() operations access this value. They are
69 void (*eo_set_last_eb_blk
)(struct ocfs2_extent_tree
*et
,
71 u64 (*eo_get_last_eb_blk
)(struct ocfs2_extent_tree
*et
);
74 * The on-disk structure usually keeps track of how many total
75 * clusters are stored in this extent tree. This function updates
76 * that value. new_clusters is the delta, and must be
77 * added to the total. Required.
79 void (*eo_update_clusters
)(struct inode
*inode
,
80 struct ocfs2_extent_tree
*et
,
84 * If ->eo_insert_check() exists, it is called before rec is
85 * inserted into the extent tree. It is optional.
87 int (*eo_insert_check
)(struct inode
*inode
,
88 struct ocfs2_extent_tree
*et
,
89 struct ocfs2_extent_rec
*rec
);
90 int (*eo_sanity_check
)(struct inode
*inode
, struct ocfs2_extent_tree
*et
);
93 * --------------------------------------------------------------
94 * The remaining are internal to ocfs2_extent_tree and don't have
99 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
102 void (*eo_fill_root_el
)(struct ocfs2_extent_tree
*et
);
105 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
106 * it exists. If it does not, et->et_max_leaf_clusters is set
107 * to 0 (unlimited). Optional.
109 void (*eo_fill_max_leaf_clusters
)(struct inode
*inode
,
110 struct ocfs2_extent_tree
*et
);
115 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
118 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
);
119 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
121 static void ocfs2_dinode_update_clusters(struct inode
*inode
,
122 struct ocfs2_extent_tree
*et
,
124 static int ocfs2_dinode_insert_check(struct inode
*inode
,
125 struct ocfs2_extent_tree
*et
,
126 struct ocfs2_extent_rec
*rec
);
127 static int ocfs2_dinode_sanity_check(struct inode
*inode
,
128 struct ocfs2_extent_tree
*et
);
129 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
);
130 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops
= {
131 .eo_set_last_eb_blk
= ocfs2_dinode_set_last_eb_blk
,
132 .eo_get_last_eb_blk
= ocfs2_dinode_get_last_eb_blk
,
133 .eo_update_clusters
= ocfs2_dinode_update_clusters
,
134 .eo_insert_check
= ocfs2_dinode_insert_check
,
135 .eo_sanity_check
= ocfs2_dinode_sanity_check
,
136 .eo_fill_root_el
= ocfs2_dinode_fill_root_el
,
139 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
142 struct ocfs2_dinode
*di
= et
->et_object
;
144 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
145 di
->i_last_eb_blk
= cpu_to_le64(blkno
);
148 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
150 struct ocfs2_dinode
*di
= et
->et_object
;
152 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
153 return le64_to_cpu(di
->i_last_eb_blk
);
156 static void ocfs2_dinode_update_clusters(struct inode
*inode
,
157 struct ocfs2_extent_tree
*et
,
160 struct ocfs2_dinode
*di
= et
->et_object
;
162 le32_add_cpu(&di
->i_clusters
, clusters
);
163 spin_lock(&OCFS2_I(inode
)->ip_lock
);
164 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(di
->i_clusters
);
165 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
168 static int ocfs2_dinode_insert_check(struct inode
*inode
,
169 struct ocfs2_extent_tree
*et
,
170 struct ocfs2_extent_rec
*rec
)
172 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
174 BUG_ON(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
);
175 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb
) &&
176 (OCFS2_I(inode
)->ip_clusters
!= rec
->e_cpos
),
177 "Device %s, asking for sparse allocation: inode %llu, "
178 "cpos %u, clusters %u\n",
180 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
182 OCFS2_I(inode
)->ip_clusters
);
187 static int ocfs2_dinode_sanity_check(struct inode
*inode
,
188 struct ocfs2_extent_tree
*et
)
191 struct ocfs2_dinode
*di
;
193 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
196 if (!OCFS2_IS_VALID_DINODE(di
)) {
198 ocfs2_error(inode
->i_sb
,
199 "Inode %llu has invalid path root",
200 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
206 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
)
208 struct ocfs2_dinode
*di
= et
->et_object
;
210 et
->et_root_el
= &di
->id2
.i_list
;
214 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree
*et
)
216 struct ocfs2_xattr_value_root
*xv
= et
->et_object
;
218 et
->et_root_el
= &xv
->xr_list
;
221 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
224 struct ocfs2_xattr_value_root
*xv
=
225 (struct ocfs2_xattr_value_root
*)et
->et_object
;
227 xv
->xr_last_eb_blk
= cpu_to_le64(blkno
);
230 static u64
ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
232 struct ocfs2_xattr_value_root
*xv
=
233 (struct ocfs2_xattr_value_root
*) et
->et_object
;
235 return le64_to_cpu(xv
->xr_last_eb_blk
);
238 static void ocfs2_xattr_value_update_clusters(struct inode
*inode
,
239 struct ocfs2_extent_tree
*et
,
242 struct ocfs2_xattr_value_root
*xv
=
243 (struct ocfs2_xattr_value_root
*)et
->et_object
;
245 le32_add_cpu(&xv
->xr_clusters
, clusters
);
248 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops
= {
249 .eo_set_last_eb_blk
= ocfs2_xattr_value_set_last_eb_blk
,
250 .eo_get_last_eb_blk
= ocfs2_xattr_value_get_last_eb_blk
,
251 .eo_update_clusters
= ocfs2_xattr_value_update_clusters
,
252 .eo_fill_root_el
= ocfs2_xattr_value_fill_root_el
,
255 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree
*et
)
257 struct ocfs2_xattr_block
*xb
= et
->et_object
;
259 et
->et_root_el
= &xb
->xb_attrs
.xb_root
.xt_list
;
262 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct inode
*inode
,
263 struct ocfs2_extent_tree
*et
)
265 et
->et_max_leaf_clusters
=
266 ocfs2_clusters_for_bytes(inode
->i_sb
,
267 OCFS2_MAX_XATTR_TREE_LEAF_SIZE
);
270 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
273 struct ocfs2_xattr_block
*xb
= et
->et_object
;
274 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
276 xt
->xt_last_eb_blk
= cpu_to_le64(blkno
);
279 static u64
ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
281 struct ocfs2_xattr_block
*xb
= et
->et_object
;
282 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
284 return le64_to_cpu(xt
->xt_last_eb_blk
);
287 static void ocfs2_xattr_tree_update_clusters(struct inode
*inode
,
288 struct ocfs2_extent_tree
*et
,
291 struct ocfs2_xattr_block
*xb
= et
->et_object
;
293 le32_add_cpu(&xb
->xb_attrs
.xb_root
.xt_clusters
, clusters
);
296 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops
= {
297 .eo_set_last_eb_blk
= ocfs2_xattr_tree_set_last_eb_blk
,
298 .eo_get_last_eb_blk
= ocfs2_xattr_tree_get_last_eb_blk
,
299 .eo_update_clusters
= ocfs2_xattr_tree_update_clusters
,
300 .eo_fill_root_el
= ocfs2_xattr_tree_fill_root_el
,
301 .eo_fill_max_leaf_clusters
= ocfs2_xattr_tree_fill_max_leaf_clusters
,
304 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree
*et
,
306 struct buffer_head
*bh
,
308 struct ocfs2_extent_tree_operations
*ops
)
313 obj
= (void *)bh
->b_data
;
316 et
->et_ops
->eo_fill_root_el(et
);
317 if (!et
->et_ops
->eo_fill_max_leaf_clusters
)
318 et
->et_max_leaf_clusters
= 0;
320 et
->et_ops
->eo_fill_max_leaf_clusters(inode
, et
);
323 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree
*et
,
325 struct buffer_head
*bh
)
327 __ocfs2_init_extent_tree(et
, inode
, bh
, NULL
, &ocfs2_dinode_et_ops
);
330 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree
*et
,
332 struct buffer_head
*bh
)
334 __ocfs2_init_extent_tree(et
, inode
, bh
, NULL
,
335 &ocfs2_xattr_tree_et_ops
);
338 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree
*et
,
340 struct buffer_head
*bh
,
341 struct ocfs2_xattr_value_root
*xv
)
343 __ocfs2_init_extent_tree(et
, inode
, bh
, xv
,
344 &ocfs2_xattr_value_et_ops
);
347 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
350 et
->et_ops
->eo_set_last_eb_blk(et
, new_last_eb_blk
);
353 static inline u64
ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
355 return et
->et_ops
->eo_get_last_eb_blk(et
);
358 static inline void ocfs2_et_update_clusters(struct inode
*inode
,
359 struct ocfs2_extent_tree
*et
,
362 et
->et_ops
->eo_update_clusters(inode
, et
, clusters
);
365 static inline int ocfs2_et_insert_check(struct inode
*inode
,
366 struct ocfs2_extent_tree
*et
,
367 struct ocfs2_extent_rec
*rec
)
371 if (et
->et_ops
->eo_insert_check
)
372 ret
= et
->et_ops
->eo_insert_check(inode
, et
, rec
);
376 static inline int ocfs2_et_sanity_check(struct inode
*inode
,
377 struct ocfs2_extent_tree
*et
)
381 if (et
->et_ops
->eo_sanity_check
)
382 ret
= et
->et_ops
->eo_sanity_check(inode
, et
);
386 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
);
387 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
388 struct ocfs2_extent_block
*eb
);
391 * Structures which describe a path through a btree, and functions to
394 * The idea here is to be as generic as possible with the tree
397 struct ocfs2_path_item
{
398 struct buffer_head
*bh
;
399 struct ocfs2_extent_list
*el
;
402 #define OCFS2_MAX_PATH_DEPTH 5
406 struct ocfs2_path_item p_node
[OCFS2_MAX_PATH_DEPTH
];
409 #define path_root_bh(_path) ((_path)->p_node[0].bh)
410 #define path_root_el(_path) ((_path)->p_node[0].el)
411 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
412 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
413 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
416 * Reset the actual path elements so that we can re-use the structure
417 * to build another path. Generally, this involves freeing the buffer
420 static void ocfs2_reinit_path(struct ocfs2_path
*path
, int keep_root
)
422 int i
, start
= 0, depth
= 0;
423 struct ocfs2_path_item
*node
;
428 for(i
= start
; i
< path_num_items(path
); i
++) {
429 node
= &path
->p_node
[i
];
437 * Tree depth may change during truncate, or insert. If we're
438 * keeping the root extent list, then make sure that our path
439 * structure reflects the proper depth.
442 depth
= le16_to_cpu(path_root_el(path
)->l_tree_depth
);
444 path
->p_tree_depth
= depth
;
447 static void ocfs2_free_path(struct ocfs2_path
*path
)
450 ocfs2_reinit_path(path
, 0);
456 * All the elements of src into dest. After this call, src could be freed
457 * without affecting dest.
459 * Both paths should have the same root. Any non-root elements of dest
462 static void ocfs2_cp_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
466 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
467 BUG_ON(path_root_el(dest
) != path_root_el(src
));
469 ocfs2_reinit_path(dest
, 1);
471 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
472 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
473 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
475 if (dest
->p_node
[i
].bh
)
476 get_bh(dest
->p_node
[i
].bh
);
481 * Make the *dest path the same as src and re-initialize src path to
484 static void ocfs2_mv_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
488 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
490 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
491 brelse(dest
->p_node
[i
].bh
);
493 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
494 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
496 src
->p_node
[i
].bh
= NULL
;
497 src
->p_node
[i
].el
= NULL
;
502 * Insert an extent block at given index.
504 * This will not take an additional reference on eb_bh.
506 static inline void ocfs2_path_insert_eb(struct ocfs2_path
*path
, int index
,
507 struct buffer_head
*eb_bh
)
509 struct ocfs2_extent_block
*eb
= (struct ocfs2_extent_block
*)eb_bh
->b_data
;
512 * Right now, no root bh is an extent block, so this helps
513 * catch code errors with dinode trees. The assertion can be
514 * safely removed if we ever need to insert extent block
515 * structures at the root.
519 path
->p_node
[index
].bh
= eb_bh
;
520 path
->p_node
[index
].el
= &eb
->h_list
;
523 static struct ocfs2_path
*ocfs2_new_path(struct buffer_head
*root_bh
,
524 struct ocfs2_extent_list
*root_el
)
526 struct ocfs2_path
*path
;
528 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) >= OCFS2_MAX_PATH_DEPTH
);
530 path
= kzalloc(sizeof(*path
), GFP_NOFS
);
532 path
->p_tree_depth
= le16_to_cpu(root_el
->l_tree_depth
);
534 path_root_bh(path
) = root_bh
;
535 path_root_el(path
) = root_el
;
542 * Convenience function to journal all components in a path.
544 static int ocfs2_journal_access_path(struct inode
*inode
, handle_t
*handle
,
545 struct ocfs2_path
*path
)
552 for(i
= 0; i
< path_num_items(path
); i
++) {
553 ret
= ocfs2_journal_access(handle
, inode
, path
->p_node
[i
].bh
,
554 OCFS2_JOURNAL_ACCESS_WRITE
);
566 * Return the index of the extent record which contains cluster #v_cluster.
567 * -1 is returned if it was not found.
569 * Should work fine on interior and exterior nodes.
571 int ocfs2_search_extent_list(struct ocfs2_extent_list
*el
, u32 v_cluster
)
575 struct ocfs2_extent_rec
*rec
;
576 u32 rec_end
, rec_start
, clusters
;
578 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
579 rec
= &el
->l_recs
[i
];
581 rec_start
= le32_to_cpu(rec
->e_cpos
);
582 clusters
= ocfs2_rec_clusters(el
, rec
);
584 rec_end
= rec_start
+ clusters
;
586 if (v_cluster
>= rec_start
&& v_cluster
< rec_end
) {
595 enum ocfs2_contig_type
{
604 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
605 * ocfs2_extent_contig only work properly against leaf nodes!
607 static int ocfs2_block_extent_contig(struct super_block
*sb
,
608 struct ocfs2_extent_rec
*ext
,
611 u64 blk_end
= le64_to_cpu(ext
->e_blkno
);
613 blk_end
+= ocfs2_clusters_to_blocks(sb
,
614 le16_to_cpu(ext
->e_leaf_clusters
));
616 return blkno
== blk_end
;
619 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec
*left
,
620 struct ocfs2_extent_rec
*right
)
624 left_range
= le32_to_cpu(left
->e_cpos
) +
625 le16_to_cpu(left
->e_leaf_clusters
);
627 return (left_range
== le32_to_cpu(right
->e_cpos
));
630 static enum ocfs2_contig_type
631 ocfs2_extent_contig(struct inode
*inode
,
632 struct ocfs2_extent_rec
*ext
,
633 struct ocfs2_extent_rec
*insert_rec
)
635 u64 blkno
= le64_to_cpu(insert_rec
->e_blkno
);
638 * Refuse to coalesce extent records with different flag
639 * fields - we don't want to mix unwritten extents with user
642 if (ext
->e_flags
!= insert_rec
->e_flags
)
645 if (ocfs2_extents_adjacent(ext
, insert_rec
) &&
646 ocfs2_block_extent_contig(inode
->i_sb
, ext
, blkno
))
649 blkno
= le64_to_cpu(ext
->e_blkno
);
650 if (ocfs2_extents_adjacent(insert_rec
, ext
) &&
651 ocfs2_block_extent_contig(inode
->i_sb
, insert_rec
, blkno
))
658 * NOTE: We can have pretty much any combination of contiguousness and
661 * The usefulness of APPEND_TAIL is more in that it lets us know that
662 * we'll have to update the path to that leaf.
664 enum ocfs2_append_type
{
669 enum ocfs2_split_type
{
675 struct ocfs2_insert_type
{
676 enum ocfs2_split_type ins_split
;
677 enum ocfs2_append_type ins_appending
;
678 enum ocfs2_contig_type ins_contig
;
679 int ins_contig_index
;
683 struct ocfs2_merge_ctxt
{
684 enum ocfs2_contig_type c_contig_type
;
685 int c_has_empty_extent
;
686 int c_split_covers_rec
;
690 * How many free extents have we got before we need more meta data?
692 int ocfs2_num_free_extents(struct ocfs2_super
*osb
,
694 struct ocfs2_extent_tree
*et
)
697 struct ocfs2_extent_list
*el
= NULL
;
698 struct ocfs2_extent_block
*eb
;
699 struct buffer_head
*eb_bh
= NULL
;
705 last_eb_blk
= ocfs2_et_get_last_eb_blk(et
);
708 retval
= ocfs2_read_block(osb
, last_eb_blk
,
709 &eb_bh
, OCFS2_BH_CACHED
, inode
);
714 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
718 BUG_ON(el
->l_tree_depth
!= 0);
720 retval
= le16_to_cpu(el
->l_count
) - le16_to_cpu(el
->l_next_free_rec
);
729 /* expects array to already be allocated
731 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
734 static int ocfs2_create_new_meta_bhs(struct ocfs2_super
*osb
,
738 struct ocfs2_alloc_context
*meta_ac
,
739 struct buffer_head
*bhs
[])
741 int count
, status
, i
;
742 u16 suballoc_bit_start
;
745 struct ocfs2_extent_block
*eb
;
750 while (count
< wanted
) {
751 status
= ocfs2_claim_metadata(osb
,
763 for(i
= count
; i
< (num_got
+ count
); i
++) {
764 bhs
[i
] = sb_getblk(osb
->sb
, first_blkno
);
765 if (bhs
[i
] == NULL
) {
770 ocfs2_set_new_buffer_uptodate(inode
, bhs
[i
]);
772 status
= ocfs2_journal_access(handle
, inode
, bhs
[i
],
773 OCFS2_JOURNAL_ACCESS_CREATE
);
779 memset(bhs
[i
]->b_data
, 0, osb
->sb
->s_blocksize
);
780 eb
= (struct ocfs2_extent_block
*) bhs
[i
]->b_data
;
781 /* Ok, setup the minimal stuff here. */
782 strcpy(eb
->h_signature
, OCFS2_EXTENT_BLOCK_SIGNATURE
);
783 eb
->h_blkno
= cpu_to_le64(first_blkno
);
784 eb
->h_fs_generation
= cpu_to_le32(osb
->fs_generation
);
785 eb
->h_suballoc_slot
= cpu_to_le16(osb
->slot_num
);
786 eb
->h_suballoc_bit
= cpu_to_le16(suballoc_bit_start
);
788 cpu_to_le16(ocfs2_extent_recs_per_eb(osb
->sb
));
790 suballoc_bit_start
++;
793 /* We'll also be dirtied by the caller, so
794 * this isn't absolutely necessary. */
795 status
= ocfs2_journal_dirty(handle
, bhs
[i
]);
808 for(i
= 0; i
< wanted
; i
++) {
819 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
821 * Returns the sum of the rightmost extent rec logical offset and
824 * ocfs2_add_branch() uses this to determine what logical cluster
825 * value should be populated into the leftmost new branch records.
827 * ocfs2_shift_tree_depth() uses this to determine the # clusters
828 * value for the new topmost tree record.
830 static inline u32
ocfs2_sum_rightmost_rec(struct ocfs2_extent_list
*el
)
834 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
836 return le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
837 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
841 * Add an entire tree branch to our inode. eb_bh is the extent block
842 * to start at, if we don't want to start the branch at the dinode
845 * last_eb_bh is required as we have to update it's next_leaf pointer
846 * for the new last extent block.
848 * the new branch will be 'empty' in the sense that every block will
849 * contain a single record with cluster count == 0.
851 static int ocfs2_add_branch(struct ocfs2_super
*osb
,
854 struct ocfs2_extent_tree
*et
,
855 struct buffer_head
*eb_bh
,
856 struct buffer_head
**last_eb_bh
,
857 struct ocfs2_alloc_context
*meta_ac
)
859 int status
, new_blocks
, i
;
860 u64 next_blkno
, new_last_eb_blk
;
861 struct buffer_head
*bh
;
862 struct buffer_head
**new_eb_bhs
= NULL
;
863 struct ocfs2_extent_block
*eb
;
864 struct ocfs2_extent_list
*eb_el
;
865 struct ocfs2_extent_list
*el
;
870 BUG_ON(!last_eb_bh
|| !*last_eb_bh
);
873 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
878 /* we never add a branch to a leaf. */
879 BUG_ON(!el
->l_tree_depth
);
881 new_blocks
= le16_to_cpu(el
->l_tree_depth
);
883 /* allocate the number of new eb blocks we need */
884 new_eb_bhs
= kcalloc(new_blocks
, sizeof(struct buffer_head
*),
892 status
= ocfs2_create_new_meta_bhs(osb
, handle
, inode
, new_blocks
,
893 meta_ac
, new_eb_bhs
);
899 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
900 new_cpos
= ocfs2_sum_rightmost_rec(&eb
->h_list
);
902 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
903 * linked with the rest of the tree.
904 * conversly, new_eb_bhs[0] is the new bottommost leaf.
906 * when we leave the loop, new_last_eb_blk will point to the
907 * newest leaf, and next_blkno will point to the topmost extent
909 next_blkno
= new_last_eb_blk
= 0;
910 for(i
= 0; i
< new_blocks
; i
++) {
912 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
913 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
914 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
920 status
= ocfs2_journal_access(handle
, inode
, bh
,
921 OCFS2_JOURNAL_ACCESS_CREATE
);
927 eb
->h_next_leaf_blk
= 0;
928 eb_el
->l_tree_depth
= cpu_to_le16(i
);
929 eb_el
->l_next_free_rec
= cpu_to_le16(1);
931 * This actually counts as an empty extent as
934 eb_el
->l_recs
[0].e_cpos
= cpu_to_le32(new_cpos
);
935 eb_el
->l_recs
[0].e_blkno
= cpu_to_le64(next_blkno
);
937 * eb_el isn't always an interior node, but even leaf
938 * nodes want a zero'd flags and reserved field so
939 * this gets the whole 32 bits regardless of use.
941 eb_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(0);
942 if (!eb_el
->l_tree_depth
)
943 new_last_eb_blk
= le64_to_cpu(eb
->h_blkno
);
945 status
= ocfs2_journal_dirty(handle
, bh
);
951 next_blkno
= le64_to_cpu(eb
->h_blkno
);
954 /* This is a bit hairy. We want to update up to three blocks
955 * here without leaving any of them in an inconsistent state
956 * in case of error. We don't have to worry about
957 * journal_dirty erroring as it won't unless we've aborted the
958 * handle (in which case we would never be here) so reserving
959 * the write with journal_access is all we need to do. */
960 status
= ocfs2_journal_access(handle
, inode
, *last_eb_bh
,
961 OCFS2_JOURNAL_ACCESS_WRITE
);
966 status
= ocfs2_journal_access(handle
, inode
, et
->et_root_bh
,
967 OCFS2_JOURNAL_ACCESS_WRITE
);
973 status
= ocfs2_journal_access(handle
, inode
, eb_bh
,
974 OCFS2_JOURNAL_ACCESS_WRITE
);
981 /* Link the new branch into the rest of the tree (el will
982 * either be on the root_bh, or the extent block passed in. */
983 i
= le16_to_cpu(el
->l_next_free_rec
);
984 el
->l_recs
[i
].e_blkno
= cpu_to_le64(next_blkno
);
985 el
->l_recs
[i
].e_cpos
= cpu_to_le32(new_cpos
);
986 el
->l_recs
[i
].e_int_clusters
= 0;
987 le16_add_cpu(&el
->l_next_free_rec
, 1);
989 /* fe needs a new last extent block pointer, as does the
990 * next_leaf on the previously last-extent-block. */
991 ocfs2_et_set_last_eb_blk(et
, new_last_eb_blk
);
993 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
994 eb
->h_next_leaf_blk
= cpu_to_le64(new_last_eb_blk
);
996 status
= ocfs2_journal_dirty(handle
, *last_eb_bh
);
999 status
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1003 status
= ocfs2_journal_dirty(handle
, eb_bh
);
1009 * Some callers want to track the rightmost leaf so pass it
1012 brelse(*last_eb_bh
);
1013 get_bh(new_eb_bhs
[0]);
1014 *last_eb_bh
= new_eb_bhs
[0];
1019 for (i
= 0; i
< new_blocks
; i
++)
1021 brelse(new_eb_bhs
[i
]);
1030 * adds another level to the allocation tree.
1031 * returns back the new extent block so you can add a branch to it
1034 static int ocfs2_shift_tree_depth(struct ocfs2_super
*osb
,
1036 struct inode
*inode
,
1037 struct ocfs2_extent_tree
*et
,
1038 struct ocfs2_alloc_context
*meta_ac
,
1039 struct buffer_head
**ret_new_eb_bh
)
1043 struct buffer_head
*new_eb_bh
= NULL
;
1044 struct ocfs2_extent_block
*eb
;
1045 struct ocfs2_extent_list
*root_el
;
1046 struct ocfs2_extent_list
*eb_el
;
1050 status
= ocfs2_create_new_meta_bhs(osb
, handle
, inode
, 1, meta_ac
,
1057 eb
= (struct ocfs2_extent_block
*) new_eb_bh
->b_data
;
1058 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
1059 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
1064 eb_el
= &eb
->h_list
;
1065 root_el
= et
->et_root_el
;
1067 status
= ocfs2_journal_access(handle
, inode
, new_eb_bh
,
1068 OCFS2_JOURNAL_ACCESS_CREATE
);
1074 /* copy the root extent list data into the new extent block */
1075 eb_el
->l_tree_depth
= root_el
->l_tree_depth
;
1076 eb_el
->l_next_free_rec
= root_el
->l_next_free_rec
;
1077 for (i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1078 eb_el
->l_recs
[i
] = root_el
->l_recs
[i
];
1080 status
= ocfs2_journal_dirty(handle
, new_eb_bh
);
1086 status
= ocfs2_journal_access(handle
, inode
, et
->et_root_bh
,
1087 OCFS2_JOURNAL_ACCESS_WRITE
);
1093 new_clusters
= ocfs2_sum_rightmost_rec(eb_el
);
1095 /* update root_bh now */
1096 le16_add_cpu(&root_el
->l_tree_depth
, 1);
1097 root_el
->l_recs
[0].e_cpos
= 0;
1098 root_el
->l_recs
[0].e_blkno
= eb
->h_blkno
;
1099 root_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(new_clusters
);
1100 for (i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1101 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
1102 root_el
->l_next_free_rec
= cpu_to_le16(1);
1104 /* If this is our 1st tree depth shift, then last_eb_blk
1105 * becomes the allocated extent block */
1106 if (root_el
->l_tree_depth
== cpu_to_le16(1))
1107 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
1109 status
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1115 *ret_new_eb_bh
= new_eb_bh
;
1127 * Should only be called when there is no space left in any of the
1128 * leaf nodes. What we want to do is find the lowest tree depth
1129 * non-leaf extent block with room for new records. There are three
1130 * valid results of this search:
1132 * 1) a lowest extent block is found, then we pass it back in
1133 * *lowest_eb_bh and return '0'
1135 * 2) the search fails to find anything, but the root_el has room. We
1136 * pass NULL back in *lowest_eb_bh, but still return '0'
1138 * 3) the search fails to find anything AND the root_el is full, in
1139 * which case we return > 0
1141 * return status < 0 indicates an error.
1143 static int ocfs2_find_branch_target(struct ocfs2_super
*osb
,
1144 struct inode
*inode
,
1145 struct ocfs2_extent_tree
*et
,
1146 struct buffer_head
**target_bh
)
1150 struct ocfs2_extent_block
*eb
;
1151 struct ocfs2_extent_list
*el
;
1152 struct buffer_head
*bh
= NULL
;
1153 struct buffer_head
*lowest_bh
= NULL
;
1159 el
= et
->et_root_el
;
1161 while(le16_to_cpu(el
->l_tree_depth
) > 1) {
1162 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1163 ocfs2_error(inode
->i_sb
, "Dinode %llu has empty "
1164 "extent list (next_free_rec == 0)",
1165 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1169 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1170 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1172 ocfs2_error(inode
->i_sb
, "Dinode %llu has extent "
1173 "list where extent # %d has no physical "
1175 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, i
);
1185 status
= ocfs2_read_block(osb
, blkno
, &bh
, OCFS2_BH_CACHED
,
1192 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1193 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
1194 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
1200 if (le16_to_cpu(el
->l_next_free_rec
) <
1201 le16_to_cpu(el
->l_count
)) {
1209 /* If we didn't find one and the fe doesn't have any room,
1210 * then return '1' */
1211 el
= et
->et_root_el
;
1212 if (!lowest_bh
&& (el
->l_next_free_rec
== el
->l_count
))
1215 *target_bh
= lowest_bh
;
1225 * Grow a b-tree so that it has more records.
1227 * We might shift the tree depth in which case existing paths should
1228 * be considered invalid.
1230 * Tree depth after the grow is returned via *final_depth.
1232 * *last_eb_bh will be updated by ocfs2_add_branch().
1234 static int ocfs2_grow_tree(struct inode
*inode
, handle_t
*handle
,
1235 struct ocfs2_extent_tree
*et
, int *final_depth
,
1236 struct buffer_head
**last_eb_bh
,
1237 struct ocfs2_alloc_context
*meta_ac
)
1240 struct ocfs2_extent_list
*el
= et
->et_root_el
;
1241 int depth
= le16_to_cpu(el
->l_tree_depth
);
1242 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1243 struct buffer_head
*bh
= NULL
;
1245 BUG_ON(meta_ac
== NULL
);
1247 shift
= ocfs2_find_branch_target(osb
, inode
, et
, &bh
);
1254 /* We traveled all the way to the bottom of the allocation tree
1255 * and didn't find room for any more extents - we need to add
1256 * another tree level */
1259 mlog(0, "need to shift tree depth (current = %d)\n", depth
);
1261 /* ocfs2_shift_tree_depth will return us a buffer with
1262 * the new extent block (so we can pass that to
1263 * ocfs2_add_branch). */
1264 ret
= ocfs2_shift_tree_depth(osb
, handle
, inode
, et
,
1273 * Special case: we have room now if we shifted from
1274 * tree_depth 0, so no more work needs to be done.
1276 * We won't be calling add_branch, so pass
1277 * back *last_eb_bh as the new leaf. At depth
1278 * zero, it should always be null so there's
1279 * no reason to brelse.
1281 BUG_ON(*last_eb_bh
);
1288 /* call ocfs2_add_branch to add the final part of the tree with
1290 mlog(0, "add branch. bh = %p\n", bh
);
1291 ret
= ocfs2_add_branch(osb
, handle
, inode
, et
, bh
, last_eb_bh
,
1300 *final_depth
= depth
;
1306 * This function will discard the rightmost extent record.
1308 static void ocfs2_shift_records_right(struct ocfs2_extent_list
*el
)
1310 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1311 int count
= le16_to_cpu(el
->l_count
);
1312 unsigned int num_bytes
;
1315 /* This will cause us to go off the end of our extent list. */
1316 BUG_ON(next_free
>= count
);
1318 num_bytes
= sizeof(struct ocfs2_extent_rec
) * next_free
;
1320 memmove(&el
->l_recs
[1], &el
->l_recs
[0], num_bytes
);
1323 static void ocfs2_rotate_leaf(struct ocfs2_extent_list
*el
,
1324 struct ocfs2_extent_rec
*insert_rec
)
1326 int i
, insert_index
, next_free
, has_empty
, num_bytes
;
1327 u32 insert_cpos
= le32_to_cpu(insert_rec
->e_cpos
);
1328 struct ocfs2_extent_rec
*rec
;
1330 next_free
= le16_to_cpu(el
->l_next_free_rec
);
1331 has_empty
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
1335 /* The tree code before us didn't allow enough room in the leaf. */
1336 BUG_ON(el
->l_next_free_rec
== el
->l_count
&& !has_empty
);
1339 * The easiest way to approach this is to just remove the
1340 * empty extent and temporarily decrement next_free.
1344 * If next_free was 1 (only an empty extent), this
1345 * loop won't execute, which is fine. We still want
1346 * the decrement above to happen.
1348 for(i
= 0; i
< (next_free
- 1); i
++)
1349 el
->l_recs
[i
] = el
->l_recs
[i
+1];
1355 * Figure out what the new record index should be.
1357 for(i
= 0; i
< next_free
; i
++) {
1358 rec
= &el
->l_recs
[i
];
1360 if (insert_cpos
< le32_to_cpu(rec
->e_cpos
))
1365 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1366 insert_cpos
, insert_index
, has_empty
, next_free
, le16_to_cpu(el
->l_count
));
1368 BUG_ON(insert_index
< 0);
1369 BUG_ON(insert_index
>= le16_to_cpu(el
->l_count
));
1370 BUG_ON(insert_index
> next_free
);
1373 * No need to memmove if we're just adding to the tail.
1375 if (insert_index
!= next_free
) {
1376 BUG_ON(next_free
>= le16_to_cpu(el
->l_count
));
1378 num_bytes
= next_free
- insert_index
;
1379 num_bytes
*= sizeof(struct ocfs2_extent_rec
);
1380 memmove(&el
->l_recs
[insert_index
+ 1],
1381 &el
->l_recs
[insert_index
],
1386 * Either we had an empty extent, and need to re-increment or
1387 * there was no empty extent on a non full rightmost leaf node,
1388 * in which case we still need to increment.
1391 el
->l_next_free_rec
= cpu_to_le16(next_free
);
1393 * Make sure none of the math above just messed up our tree.
1395 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) > le16_to_cpu(el
->l_count
));
1397 el
->l_recs
[insert_index
] = *insert_rec
;
1401 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list
*el
)
1403 int size
, num_recs
= le16_to_cpu(el
->l_next_free_rec
);
1405 BUG_ON(num_recs
== 0);
1407 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
1409 size
= num_recs
* sizeof(struct ocfs2_extent_rec
);
1410 memmove(&el
->l_recs
[0], &el
->l_recs
[1], size
);
1411 memset(&el
->l_recs
[num_recs
], 0,
1412 sizeof(struct ocfs2_extent_rec
));
1413 el
->l_next_free_rec
= cpu_to_le16(num_recs
);
1418 * Create an empty extent record .
1420 * l_next_free_rec may be updated.
1422 * If an empty extent already exists do nothing.
1424 static void ocfs2_create_empty_extent(struct ocfs2_extent_list
*el
)
1426 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1428 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
1433 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
1436 mlog_bug_on_msg(el
->l_count
== el
->l_next_free_rec
,
1437 "Asked to create an empty extent in a full list:\n"
1438 "count = %u, tree depth = %u",
1439 le16_to_cpu(el
->l_count
),
1440 le16_to_cpu(el
->l_tree_depth
));
1442 ocfs2_shift_records_right(el
);
1445 le16_add_cpu(&el
->l_next_free_rec
, 1);
1446 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1450 * For a rotation which involves two leaf nodes, the "root node" is
1451 * the lowest level tree node which contains a path to both leafs. This
1452 * resulting set of information can be used to form a complete "subtree"
1454 * This function is passed two full paths from the dinode down to a
1455 * pair of adjacent leaves. It's task is to figure out which path
1456 * index contains the subtree root - this can be the root index itself
1457 * in a worst-case rotation.
1459 * The array index of the subtree root is passed back.
1461 static int ocfs2_find_subtree_root(struct inode
*inode
,
1462 struct ocfs2_path
*left
,
1463 struct ocfs2_path
*right
)
1468 * Check that the caller passed in two paths from the same tree.
1470 BUG_ON(path_root_bh(left
) != path_root_bh(right
));
1476 * The caller didn't pass two adjacent paths.
1478 mlog_bug_on_msg(i
> left
->p_tree_depth
,
1479 "Inode %lu, left depth %u, right depth %u\n"
1480 "left leaf blk %llu, right leaf blk %llu\n",
1481 inode
->i_ino
, left
->p_tree_depth
,
1482 right
->p_tree_depth
,
1483 (unsigned long long)path_leaf_bh(left
)->b_blocknr
,
1484 (unsigned long long)path_leaf_bh(right
)->b_blocknr
);
1485 } while (left
->p_node
[i
].bh
->b_blocknr
==
1486 right
->p_node
[i
].bh
->b_blocknr
);
1491 typedef void (path_insert_t
)(void *, struct buffer_head
*);
1494 * Traverse a btree path in search of cpos, starting at root_el.
1496 * This code can be called with a cpos larger than the tree, in which
1497 * case it will return the rightmost path.
1499 static int __ocfs2_find_path(struct inode
*inode
,
1500 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1501 path_insert_t
*func
, void *data
)
1506 struct buffer_head
*bh
= NULL
;
1507 struct ocfs2_extent_block
*eb
;
1508 struct ocfs2_extent_list
*el
;
1509 struct ocfs2_extent_rec
*rec
;
1510 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1513 while (el
->l_tree_depth
) {
1514 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1515 ocfs2_error(inode
->i_sb
,
1516 "Inode %llu has empty extent list at "
1518 (unsigned long long)oi
->ip_blkno
,
1519 le16_to_cpu(el
->l_tree_depth
));
1525 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
) - 1; i
++) {
1526 rec
= &el
->l_recs
[i
];
1529 * In the case that cpos is off the allocation
1530 * tree, this should just wind up returning the
1533 range
= le32_to_cpu(rec
->e_cpos
) +
1534 ocfs2_rec_clusters(el
, rec
);
1535 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1539 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1541 ocfs2_error(inode
->i_sb
,
1542 "Inode %llu has bad blkno in extent list "
1543 "at depth %u (index %d)\n",
1544 (unsigned long long)oi
->ip_blkno
,
1545 le16_to_cpu(el
->l_tree_depth
), i
);
1552 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
), blkno
,
1553 &bh
, OCFS2_BH_CACHED
, inode
);
1559 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1561 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
1562 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
1567 if (le16_to_cpu(el
->l_next_free_rec
) >
1568 le16_to_cpu(el
->l_count
)) {
1569 ocfs2_error(inode
->i_sb
,
1570 "Inode %llu has bad count in extent list "
1571 "at block %llu (next free=%u, count=%u)\n",
1572 (unsigned long long)oi
->ip_blkno
,
1573 (unsigned long long)bh
->b_blocknr
,
1574 le16_to_cpu(el
->l_next_free_rec
),
1575 le16_to_cpu(el
->l_count
));
1586 * Catch any trailing bh that the loop didn't handle.
1594 * Given an initialized path (that is, it has a valid root extent
1595 * list), this function will traverse the btree in search of the path
1596 * which would contain cpos.
1598 * The path traveled is recorded in the path structure.
1600 * Note that this will not do any comparisons on leaf node extent
1601 * records, so it will work fine in the case that we just added a tree
1604 struct find_path_data
{
1606 struct ocfs2_path
*path
;
1608 static void find_path_ins(void *data
, struct buffer_head
*bh
)
1610 struct find_path_data
*fp
= data
;
1613 ocfs2_path_insert_eb(fp
->path
, fp
->index
, bh
);
1616 static int ocfs2_find_path(struct inode
*inode
, struct ocfs2_path
*path
,
1619 struct find_path_data data
;
1623 return __ocfs2_find_path(inode
, path_root_el(path
), cpos
,
1624 find_path_ins
, &data
);
1627 static void find_leaf_ins(void *data
, struct buffer_head
*bh
)
1629 struct ocfs2_extent_block
*eb
=(struct ocfs2_extent_block
*)bh
->b_data
;
1630 struct ocfs2_extent_list
*el
= &eb
->h_list
;
1631 struct buffer_head
**ret
= data
;
1633 /* We want to retain only the leaf block. */
1634 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
1640 * Find the leaf block in the tree which would contain cpos. No
1641 * checking of the actual leaf is done.
1643 * Some paths want to call this instead of allocating a path structure
1644 * and calling ocfs2_find_path().
1646 * This function doesn't handle non btree extent lists.
1648 int ocfs2_find_leaf(struct inode
*inode
, struct ocfs2_extent_list
*root_el
,
1649 u32 cpos
, struct buffer_head
**leaf_bh
)
1652 struct buffer_head
*bh
= NULL
;
1654 ret
= __ocfs2_find_path(inode
, root_el
, cpos
, find_leaf_ins
, &bh
);
1666 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1668 * Basically, we've moved stuff around at the bottom of the tree and
1669 * we need to fix up the extent records above the changes to reflect
1672 * left_rec: the record on the left.
1673 * left_child_el: is the child list pointed to by left_rec
1674 * right_rec: the record to the right of left_rec
1675 * right_child_el: is the child list pointed to by right_rec
1677 * By definition, this only works on interior nodes.
1679 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec
*left_rec
,
1680 struct ocfs2_extent_list
*left_child_el
,
1681 struct ocfs2_extent_rec
*right_rec
,
1682 struct ocfs2_extent_list
*right_child_el
)
1684 u32 left_clusters
, right_end
;
1687 * Interior nodes never have holes. Their cpos is the cpos of
1688 * the leftmost record in their child list. Their cluster
1689 * count covers the full theoretical range of their child list
1690 * - the range between their cpos and the cpos of the record
1691 * immediately to their right.
1693 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[0].e_cpos
);
1694 if (ocfs2_is_empty_extent(&right_child_el
->l_recs
[0])) {
1695 BUG_ON(le16_to_cpu(right_child_el
->l_next_free_rec
) <= 1);
1696 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[1].e_cpos
);
1698 left_clusters
-= le32_to_cpu(left_rec
->e_cpos
);
1699 left_rec
->e_int_clusters
= cpu_to_le32(left_clusters
);
1702 * Calculate the rightmost cluster count boundary before
1703 * moving cpos - we will need to adjust clusters after
1704 * updating e_cpos to keep the same highest cluster count.
1706 right_end
= le32_to_cpu(right_rec
->e_cpos
);
1707 right_end
+= le32_to_cpu(right_rec
->e_int_clusters
);
1709 right_rec
->e_cpos
= left_rec
->e_cpos
;
1710 le32_add_cpu(&right_rec
->e_cpos
, left_clusters
);
1712 right_end
-= le32_to_cpu(right_rec
->e_cpos
);
1713 right_rec
->e_int_clusters
= cpu_to_le32(right_end
);
1717 * Adjust the adjacent root node records involved in a
1718 * rotation. left_el_blkno is passed in as a key so that we can easily
1719 * find it's index in the root list.
1721 static void ocfs2_adjust_root_records(struct ocfs2_extent_list
*root_el
,
1722 struct ocfs2_extent_list
*left_el
,
1723 struct ocfs2_extent_list
*right_el
,
1728 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) <=
1729 le16_to_cpu(left_el
->l_tree_depth
));
1731 for(i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
) - 1; i
++) {
1732 if (le64_to_cpu(root_el
->l_recs
[i
].e_blkno
) == left_el_blkno
)
1737 * The path walking code should have never returned a root and
1738 * two paths which are not adjacent.
1740 BUG_ON(i
>= (le16_to_cpu(root_el
->l_next_free_rec
) - 1));
1742 ocfs2_adjust_adjacent_records(&root_el
->l_recs
[i
], left_el
,
1743 &root_el
->l_recs
[i
+ 1], right_el
);
1747 * We've changed a leaf block (in right_path) and need to reflect that
1748 * change back up the subtree.
1750 * This happens in multiple places:
1751 * - When we've moved an extent record from the left path leaf to the right
1752 * path leaf to make room for an empty extent in the left path leaf.
1753 * - When our insert into the right path leaf is at the leftmost edge
1754 * and requires an update of the path immediately to it's left. This
1755 * can occur at the end of some types of rotation and appending inserts.
1756 * - When we've adjusted the last extent record in the left path leaf and the
1757 * 1st extent record in the right path leaf during cross extent block merge.
1759 static void ocfs2_complete_edge_insert(struct inode
*inode
, handle_t
*handle
,
1760 struct ocfs2_path
*left_path
,
1761 struct ocfs2_path
*right_path
,
1765 struct ocfs2_extent_list
*el
, *left_el
, *right_el
;
1766 struct ocfs2_extent_rec
*left_rec
, *right_rec
;
1767 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
1770 * Update the counts and position values within all the
1771 * interior nodes to reflect the leaf rotation we just did.
1773 * The root node is handled below the loop.
1775 * We begin the loop with right_el and left_el pointing to the
1776 * leaf lists and work our way up.
1778 * NOTE: within this loop, left_el and right_el always refer
1779 * to the *child* lists.
1781 left_el
= path_leaf_el(left_path
);
1782 right_el
= path_leaf_el(right_path
);
1783 for(i
= left_path
->p_tree_depth
- 1; i
> subtree_index
; i
--) {
1784 mlog(0, "Adjust records at index %u\n", i
);
1787 * One nice property of knowing that all of these
1788 * nodes are below the root is that we only deal with
1789 * the leftmost right node record and the rightmost
1792 el
= left_path
->p_node
[i
].el
;
1793 idx
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
1794 left_rec
= &el
->l_recs
[idx
];
1796 el
= right_path
->p_node
[i
].el
;
1797 right_rec
= &el
->l_recs
[0];
1799 ocfs2_adjust_adjacent_records(left_rec
, left_el
, right_rec
,
1802 ret
= ocfs2_journal_dirty(handle
, left_path
->p_node
[i
].bh
);
1806 ret
= ocfs2_journal_dirty(handle
, right_path
->p_node
[i
].bh
);
1811 * Setup our list pointers now so that the current
1812 * parents become children in the next iteration.
1814 left_el
= left_path
->p_node
[i
].el
;
1815 right_el
= right_path
->p_node
[i
].el
;
1819 * At the root node, adjust the two adjacent records which
1820 * begin our path to the leaves.
1823 el
= left_path
->p_node
[subtree_index
].el
;
1824 left_el
= left_path
->p_node
[subtree_index
+ 1].el
;
1825 right_el
= right_path
->p_node
[subtree_index
+ 1].el
;
1827 ocfs2_adjust_root_records(el
, left_el
, right_el
,
1828 left_path
->p_node
[subtree_index
+ 1].bh
->b_blocknr
);
1830 root_bh
= left_path
->p_node
[subtree_index
].bh
;
1832 ret
= ocfs2_journal_dirty(handle
, root_bh
);
1837 static int ocfs2_rotate_subtree_right(struct inode
*inode
,
1839 struct ocfs2_path
*left_path
,
1840 struct ocfs2_path
*right_path
,
1844 struct buffer_head
*right_leaf_bh
;
1845 struct buffer_head
*left_leaf_bh
= NULL
;
1846 struct buffer_head
*root_bh
;
1847 struct ocfs2_extent_list
*right_el
, *left_el
;
1848 struct ocfs2_extent_rec move_rec
;
1850 left_leaf_bh
= path_leaf_bh(left_path
);
1851 left_el
= path_leaf_el(left_path
);
1853 if (left_el
->l_next_free_rec
!= left_el
->l_count
) {
1854 ocfs2_error(inode
->i_sb
,
1855 "Inode %llu has non-full interior leaf node %llu"
1857 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1858 (unsigned long long)left_leaf_bh
->b_blocknr
,
1859 le16_to_cpu(left_el
->l_next_free_rec
));
1864 * This extent block may already have an empty record, so we
1865 * return early if so.
1867 if (ocfs2_is_empty_extent(&left_el
->l_recs
[0]))
1870 root_bh
= left_path
->p_node
[subtree_index
].bh
;
1871 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
1873 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
1874 OCFS2_JOURNAL_ACCESS_WRITE
);
1880 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
1881 ret
= ocfs2_journal_access(handle
, inode
,
1882 right_path
->p_node
[i
].bh
,
1883 OCFS2_JOURNAL_ACCESS_WRITE
);
1889 ret
= ocfs2_journal_access(handle
, inode
,
1890 left_path
->p_node
[i
].bh
,
1891 OCFS2_JOURNAL_ACCESS_WRITE
);
1898 right_leaf_bh
= path_leaf_bh(right_path
);
1899 right_el
= path_leaf_el(right_path
);
1901 /* This is a code error, not a disk corruption. */
1902 mlog_bug_on_msg(!right_el
->l_next_free_rec
, "Inode %llu: Rotate fails "
1903 "because rightmost leaf block %llu is empty\n",
1904 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1905 (unsigned long long)right_leaf_bh
->b_blocknr
);
1907 ocfs2_create_empty_extent(right_el
);
1909 ret
= ocfs2_journal_dirty(handle
, right_leaf_bh
);
1915 /* Do the copy now. */
1916 i
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
1917 move_rec
= left_el
->l_recs
[i
];
1918 right_el
->l_recs
[0] = move_rec
;
1921 * Clear out the record we just copied and shift everything
1922 * over, leaving an empty extent in the left leaf.
1924 * We temporarily subtract from next_free_rec so that the
1925 * shift will lose the tail record (which is now defunct).
1927 le16_add_cpu(&left_el
->l_next_free_rec
, -1);
1928 ocfs2_shift_records_right(left_el
);
1929 memset(&left_el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1930 le16_add_cpu(&left_el
->l_next_free_rec
, 1);
1932 ret
= ocfs2_journal_dirty(handle
, left_leaf_bh
);
1938 ocfs2_complete_edge_insert(inode
, handle
, left_path
, right_path
,
1946 * Given a full path, determine what cpos value would return us a path
1947 * containing the leaf immediately to the left of the current one.
1949 * Will return zero if the path passed in is already the leftmost path.
1951 static int ocfs2_find_cpos_for_left_leaf(struct super_block
*sb
,
1952 struct ocfs2_path
*path
, u32
*cpos
)
1956 struct ocfs2_extent_list
*el
;
1958 BUG_ON(path
->p_tree_depth
== 0);
1962 blkno
= path_leaf_bh(path
)->b_blocknr
;
1964 /* Start at the tree node just above the leaf and work our way up. */
1965 i
= path
->p_tree_depth
- 1;
1967 el
= path
->p_node
[i
].el
;
1970 * Find the extent record just before the one in our
1973 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
1974 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
1978 * We've determined that the
1979 * path specified is already
1980 * the leftmost one - return a
1986 * The leftmost record points to our
1987 * leaf - we need to travel up the
1993 *cpos
= le32_to_cpu(el
->l_recs
[j
- 1].e_cpos
);
1994 *cpos
= *cpos
+ ocfs2_rec_clusters(el
,
1995 &el
->l_recs
[j
- 1]);
2002 * If we got here, we never found a valid node where
2003 * the tree indicated one should be.
2006 "Invalid extent tree at extent block %llu\n",
2007 (unsigned long long)blkno
);
2012 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2021 * Extend the transaction by enough credits to complete the rotation,
2022 * and still leave at least the original number of credits allocated
2023 * to this transaction.
2025 static int ocfs2_extend_rotate_transaction(handle_t
*handle
, int subtree_depth
,
2027 struct ocfs2_path
*path
)
2029 int credits
= (path
->p_tree_depth
- subtree_depth
) * 2 + 1 + op_credits
;
2031 if (handle
->h_buffer_credits
< credits
)
2032 return ocfs2_extend_trans(handle
, credits
);
2038 * Trap the case where we're inserting into the theoretical range past
2039 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2040 * whose cpos is less than ours into the right leaf.
2042 * It's only necessary to look at the rightmost record of the left
2043 * leaf because the logic that calls us should ensure that the
2044 * theoretical ranges in the path components above the leaves are
2047 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path
*left_path
,
2050 struct ocfs2_extent_list
*left_el
;
2051 struct ocfs2_extent_rec
*rec
;
2054 left_el
= path_leaf_el(left_path
);
2055 next_free
= le16_to_cpu(left_el
->l_next_free_rec
);
2056 rec
= &left_el
->l_recs
[next_free
- 1];
2058 if (insert_cpos
> le32_to_cpu(rec
->e_cpos
))
2063 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list
*el
, u32 cpos
)
2065 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
2067 struct ocfs2_extent_rec
*rec
;
2072 rec
= &el
->l_recs
[0];
2073 if (ocfs2_is_empty_extent(rec
)) {
2077 rec
= &el
->l_recs
[1];
2080 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2081 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
2087 * Rotate all the records in a btree right one record, starting at insert_cpos.
2089 * The path to the rightmost leaf should be passed in.
2091 * The array is assumed to be large enough to hold an entire path (tree depth).
2093 * Upon succesful return from this function:
2095 * - The 'right_path' array will contain a path to the leaf block
2096 * whose range contains e_cpos.
2097 * - That leaf block will have a single empty extent in list index 0.
2098 * - In the case that the rotation requires a post-insert update,
2099 * *ret_left_path will contain a valid path which can be passed to
2100 * ocfs2_insert_path().
2102 static int ocfs2_rotate_tree_right(struct inode
*inode
,
2104 enum ocfs2_split_type split
,
2106 struct ocfs2_path
*right_path
,
2107 struct ocfs2_path
**ret_left_path
)
2109 int ret
, start
, orig_credits
= handle
->h_buffer_credits
;
2111 struct ocfs2_path
*left_path
= NULL
;
2113 *ret_left_path
= NULL
;
2115 left_path
= ocfs2_new_path(path_root_bh(right_path
),
2116 path_root_el(right_path
));
2123 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
, &cpos
);
2129 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos
, cpos
);
2132 * What we want to do here is:
2134 * 1) Start with the rightmost path.
2136 * 2) Determine a path to the leaf block directly to the left
2139 * 3) Determine the 'subtree root' - the lowest level tree node
2140 * which contains a path to both leaves.
2142 * 4) Rotate the subtree.
2144 * 5) Find the next subtree by considering the left path to be
2145 * the new right path.
2147 * The check at the top of this while loop also accepts
2148 * insert_cpos == cpos because cpos is only a _theoretical_
2149 * value to get us the left path - insert_cpos might very well
2150 * be filling that hole.
2152 * Stop at a cpos of '0' because we either started at the
2153 * leftmost branch (i.e., a tree with one branch and a
2154 * rotation inside of it), or we've gone as far as we can in
2155 * rotating subtrees.
2157 while (cpos
&& insert_cpos
<= cpos
) {
2158 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2161 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
2167 mlog_bug_on_msg(path_leaf_bh(left_path
) ==
2168 path_leaf_bh(right_path
),
2169 "Inode %lu: error during insert of %u "
2170 "(left path cpos %u) results in two identical "
2171 "paths ending at %llu\n",
2172 inode
->i_ino
, insert_cpos
, cpos
,
2173 (unsigned long long)
2174 path_leaf_bh(left_path
)->b_blocknr
);
2176 if (split
== SPLIT_NONE
&&
2177 ocfs2_rotate_requires_path_adjustment(left_path
,
2181 * We've rotated the tree as much as we
2182 * should. The rest is up to
2183 * ocfs2_insert_path() to complete, after the
2184 * record insertion. We indicate this
2185 * situation by returning the left path.
2187 * The reason we don't adjust the records here
2188 * before the record insert is that an error
2189 * later might break the rule where a parent
2190 * record e_cpos will reflect the actual
2191 * e_cpos of the 1st nonempty record of the
2194 *ret_left_path
= left_path
;
2198 start
= ocfs2_find_subtree_root(inode
, left_path
, right_path
);
2200 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2202 (unsigned long long) right_path
->p_node
[start
].bh
->b_blocknr
,
2203 right_path
->p_tree_depth
);
2205 ret
= ocfs2_extend_rotate_transaction(handle
, start
,
2206 orig_credits
, right_path
);
2212 ret
= ocfs2_rotate_subtree_right(inode
, handle
, left_path
,
2219 if (split
!= SPLIT_NONE
&&
2220 ocfs2_leftmost_rec_contains(path_leaf_el(right_path
),
2223 * A rotate moves the rightmost left leaf
2224 * record over to the leftmost right leaf
2225 * slot. If we're doing an extent split
2226 * instead of a real insert, then we have to
2227 * check that the extent to be split wasn't
2228 * just moved over. If it was, then we can
2229 * exit here, passing left_path back -
2230 * ocfs2_split_extent() is smart enough to
2231 * search both leaves.
2233 *ret_left_path
= left_path
;
2238 * There is no need to re-read the next right path
2239 * as we know that it'll be our current left
2240 * path. Optimize by copying values instead.
2242 ocfs2_mv_path(right_path
, left_path
);
2244 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
,
2253 ocfs2_free_path(left_path
);
2259 static void ocfs2_update_edge_lengths(struct inode
*inode
, handle_t
*handle
,
2260 struct ocfs2_path
*path
)
2263 struct ocfs2_extent_rec
*rec
;
2264 struct ocfs2_extent_list
*el
;
2265 struct ocfs2_extent_block
*eb
;
2268 /* Path should always be rightmost. */
2269 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2270 BUG_ON(eb
->h_next_leaf_blk
!= 0ULL);
2273 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
2274 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2275 rec
= &el
->l_recs
[idx
];
2276 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2278 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
2279 el
= path
->p_node
[i
].el
;
2280 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2281 rec
= &el
->l_recs
[idx
];
2283 rec
->e_int_clusters
= cpu_to_le32(range
);
2284 le32_add_cpu(&rec
->e_int_clusters
, -le32_to_cpu(rec
->e_cpos
));
2286 ocfs2_journal_dirty(handle
, path
->p_node
[i
].bh
);
2290 static void ocfs2_unlink_path(struct inode
*inode
, handle_t
*handle
,
2291 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2292 struct ocfs2_path
*path
, int unlink_start
)
2295 struct ocfs2_extent_block
*eb
;
2296 struct ocfs2_extent_list
*el
;
2297 struct buffer_head
*bh
;
2299 for(i
= unlink_start
; i
< path_num_items(path
); i
++) {
2300 bh
= path
->p_node
[i
].bh
;
2302 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
2304 * Not all nodes might have had their final count
2305 * decremented by the caller - handle this here.
2308 if (le16_to_cpu(el
->l_next_free_rec
) > 1) {
2310 "Inode %llu, attempted to remove extent block "
2311 "%llu with %u records\n",
2312 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2313 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
2314 le16_to_cpu(el
->l_next_free_rec
));
2316 ocfs2_journal_dirty(handle
, bh
);
2317 ocfs2_remove_from_cache(inode
, bh
);
2321 el
->l_next_free_rec
= 0;
2322 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2324 ocfs2_journal_dirty(handle
, bh
);
2326 ret
= ocfs2_cache_extent_block_free(dealloc
, eb
);
2330 ocfs2_remove_from_cache(inode
, bh
);
2334 static void ocfs2_unlink_subtree(struct inode
*inode
, handle_t
*handle
,
2335 struct ocfs2_path
*left_path
,
2336 struct ocfs2_path
*right_path
,
2338 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2341 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2342 struct ocfs2_extent_list
*root_el
= left_path
->p_node
[subtree_index
].el
;
2343 struct ocfs2_extent_list
*el
;
2344 struct ocfs2_extent_block
*eb
;
2346 el
= path_leaf_el(left_path
);
2348 eb
= (struct ocfs2_extent_block
*)right_path
->p_node
[subtree_index
+ 1].bh
->b_data
;
2350 for(i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
2351 if (root_el
->l_recs
[i
].e_blkno
== eb
->h_blkno
)
2354 BUG_ON(i
>= le16_to_cpu(root_el
->l_next_free_rec
));
2356 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
2357 le16_add_cpu(&root_el
->l_next_free_rec
, -1);
2359 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2360 eb
->h_next_leaf_blk
= 0;
2362 ocfs2_journal_dirty(handle
, root_bh
);
2363 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2365 ocfs2_unlink_path(inode
, handle
, dealloc
, right_path
,
2369 static int ocfs2_rotate_subtree_left(struct inode
*inode
, handle_t
*handle
,
2370 struct ocfs2_path
*left_path
,
2371 struct ocfs2_path
*right_path
,
2373 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2375 struct ocfs2_extent_tree
*et
)
2377 int ret
, i
, del_right_subtree
= 0, right_has_empty
= 0;
2378 struct buffer_head
*root_bh
, *et_root_bh
= path_root_bh(right_path
);
2379 struct ocfs2_extent_list
*right_leaf_el
, *left_leaf_el
;
2380 struct ocfs2_extent_block
*eb
;
2384 right_leaf_el
= path_leaf_el(right_path
);
2385 left_leaf_el
= path_leaf_el(left_path
);
2386 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2387 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2389 if (!ocfs2_is_empty_extent(&left_leaf_el
->l_recs
[0]))
2392 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(right_path
)->b_data
;
2393 if (ocfs2_is_empty_extent(&right_leaf_el
->l_recs
[0])) {
2395 * It's legal for us to proceed if the right leaf is
2396 * the rightmost one and it has an empty extent. There
2397 * are two cases to handle - whether the leaf will be
2398 * empty after removal or not. If the leaf isn't empty
2399 * then just remove the empty extent up front. The
2400 * next block will handle empty leaves by flagging
2403 * Non rightmost leaves will throw -EAGAIN and the
2404 * caller can manually move the subtree and retry.
2407 if (eb
->h_next_leaf_blk
!= 0ULL)
2410 if (le16_to_cpu(right_leaf_el
->l_next_free_rec
) > 1) {
2411 ret
= ocfs2_journal_access(handle
, inode
,
2412 path_leaf_bh(right_path
),
2413 OCFS2_JOURNAL_ACCESS_WRITE
);
2419 ocfs2_remove_empty_extent(right_leaf_el
);
2421 right_has_empty
= 1;
2424 if (eb
->h_next_leaf_blk
== 0ULL &&
2425 le16_to_cpu(right_leaf_el
->l_next_free_rec
) == 1) {
2427 * We have to update i_last_eb_blk during the meta
2430 ret
= ocfs2_journal_access(handle
, inode
, et_root_bh
,
2431 OCFS2_JOURNAL_ACCESS_WRITE
);
2437 del_right_subtree
= 1;
2441 * Getting here with an empty extent in the right path implies
2442 * that it's the rightmost path and will be deleted.
2444 BUG_ON(right_has_empty
&& !del_right_subtree
);
2446 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
2447 OCFS2_JOURNAL_ACCESS_WRITE
);
2453 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2454 ret
= ocfs2_journal_access(handle
, inode
,
2455 right_path
->p_node
[i
].bh
,
2456 OCFS2_JOURNAL_ACCESS_WRITE
);
2462 ret
= ocfs2_journal_access(handle
, inode
,
2463 left_path
->p_node
[i
].bh
,
2464 OCFS2_JOURNAL_ACCESS_WRITE
);
2471 if (!right_has_empty
) {
2473 * Only do this if we're moving a real
2474 * record. Otherwise, the action is delayed until
2475 * after removal of the right path in which case we
2476 * can do a simple shift to remove the empty extent.
2478 ocfs2_rotate_leaf(left_leaf_el
, &right_leaf_el
->l_recs
[0]);
2479 memset(&right_leaf_el
->l_recs
[0], 0,
2480 sizeof(struct ocfs2_extent_rec
));
2482 if (eb
->h_next_leaf_blk
== 0ULL) {
2484 * Move recs over to get rid of empty extent, decrease
2485 * next_free. This is allowed to remove the last
2486 * extent in our leaf (setting l_next_free_rec to
2487 * zero) - the delete code below won't care.
2489 ocfs2_remove_empty_extent(right_leaf_el
);
2492 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2495 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2499 if (del_right_subtree
) {
2500 ocfs2_unlink_subtree(inode
, handle
, left_path
, right_path
,
2501 subtree_index
, dealloc
);
2502 ocfs2_update_edge_lengths(inode
, handle
, left_path
);
2504 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2505 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
2508 * Removal of the extent in the left leaf was skipped
2509 * above so we could delete the right path
2512 if (right_has_empty
)
2513 ocfs2_remove_empty_extent(left_leaf_el
);
2515 ret
= ocfs2_journal_dirty(handle
, et_root_bh
);
2521 ocfs2_complete_edge_insert(inode
, handle
, left_path
, right_path
,
2529 * Given a full path, determine what cpos value would return us a path
2530 * containing the leaf immediately to the right of the current one.
2532 * Will return zero if the path passed in is already the rightmost path.
2534 * This looks similar, but is subtly different to
2535 * ocfs2_find_cpos_for_left_leaf().
2537 static int ocfs2_find_cpos_for_right_leaf(struct super_block
*sb
,
2538 struct ocfs2_path
*path
, u32
*cpos
)
2542 struct ocfs2_extent_list
*el
;
2546 if (path
->p_tree_depth
== 0)
2549 blkno
= path_leaf_bh(path
)->b_blocknr
;
2551 /* Start at the tree node just above the leaf and work our way up. */
2552 i
= path
->p_tree_depth
- 1;
2556 el
= path
->p_node
[i
].el
;
2559 * Find the extent record just after the one in our
2562 next_free
= le16_to_cpu(el
->l_next_free_rec
);
2563 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2564 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2565 if (j
== (next_free
- 1)) {
2568 * We've determined that the
2569 * path specified is already
2570 * the rightmost one - return a
2576 * The rightmost record points to our
2577 * leaf - we need to travel up the
2583 *cpos
= le32_to_cpu(el
->l_recs
[j
+ 1].e_cpos
);
2589 * If we got here, we never found a valid node where
2590 * the tree indicated one should be.
2593 "Invalid extent tree at extent block %llu\n",
2594 (unsigned long long)blkno
);
2599 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2607 static int ocfs2_rotate_rightmost_leaf_left(struct inode
*inode
,
2609 struct buffer_head
*bh
,
2610 struct ocfs2_extent_list
*el
)
2614 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2617 ret
= ocfs2_journal_access(handle
, inode
, bh
,
2618 OCFS2_JOURNAL_ACCESS_WRITE
);
2624 ocfs2_remove_empty_extent(el
);
2626 ret
= ocfs2_journal_dirty(handle
, bh
);
2634 static int __ocfs2_rotate_tree_left(struct inode
*inode
,
2635 handle_t
*handle
, int orig_credits
,
2636 struct ocfs2_path
*path
,
2637 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2638 struct ocfs2_path
**empty_extent_path
,
2639 struct ocfs2_extent_tree
*et
)
2641 int ret
, subtree_root
, deleted
;
2643 struct ocfs2_path
*left_path
= NULL
;
2644 struct ocfs2_path
*right_path
= NULL
;
2646 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path
)->l_recs
[0])));
2648 *empty_extent_path
= NULL
;
2650 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, path
,
2657 left_path
= ocfs2_new_path(path_root_bh(path
),
2658 path_root_el(path
));
2665 ocfs2_cp_path(left_path
, path
);
2667 right_path
= ocfs2_new_path(path_root_bh(path
),
2668 path_root_el(path
));
2675 while (right_cpos
) {
2676 ret
= ocfs2_find_path(inode
, right_path
, right_cpos
);
2682 subtree_root
= ocfs2_find_subtree_root(inode
, left_path
,
2685 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2687 (unsigned long long)
2688 right_path
->p_node
[subtree_root
].bh
->b_blocknr
,
2689 right_path
->p_tree_depth
);
2691 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_root
,
2692 orig_credits
, left_path
);
2699 * Caller might still want to make changes to the
2700 * tree root, so re-add it to the journal here.
2702 ret
= ocfs2_journal_access(handle
, inode
,
2703 path_root_bh(left_path
),
2704 OCFS2_JOURNAL_ACCESS_WRITE
);
2710 ret
= ocfs2_rotate_subtree_left(inode
, handle
, left_path
,
2711 right_path
, subtree_root
,
2712 dealloc
, &deleted
, et
);
2713 if (ret
== -EAGAIN
) {
2715 * The rotation has to temporarily stop due to
2716 * the right subtree having an empty
2717 * extent. Pass it back to the caller for a
2720 *empty_extent_path
= right_path
;
2730 * The subtree rotate might have removed records on
2731 * the rightmost edge. If so, then rotation is
2737 ocfs2_mv_path(left_path
, right_path
);
2739 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, left_path
,
2748 ocfs2_free_path(right_path
);
2749 ocfs2_free_path(left_path
);
2754 static int ocfs2_remove_rightmost_path(struct inode
*inode
, handle_t
*handle
,
2755 struct ocfs2_path
*path
,
2756 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2757 struct ocfs2_extent_tree
*et
)
2759 int ret
, subtree_index
;
2761 struct ocfs2_path
*left_path
= NULL
;
2762 struct ocfs2_extent_block
*eb
;
2763 struct ocfs2_extent_list
*el
;
2766 ret
= ocfs2_et_sanity_check(inode
, et
);
2770 * There's two ways we handle this depending on
2771 * whether path is the only existing one.
2773 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
2774 handle
->h_buffer_credits
,
2781 ret
= ocfs2_journal_access_path(inode
, handle
, path
);
2787 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
2795 * We have a path to the left of this one - it needs
2798 left_path
= ocfs2_new_path(path_root_bh(path
),
2799 path_root_el(path
));
2806 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
2812 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
2818 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
, path
);
2820 ocfs2_unlink_subtree(inode
, handle
, left_path
, path
,
2821 subtree_index
, dealloc
);
2822 ocfs2_update_edge_lengths(inode
, handle
, left_path
);
2824 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2825 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
2828 * 'path' is also the leftmost path which
2829 * means it must be the only one. This gets
2830 * handled differently because we want to
2831 * revert the inode back to having extents
2834 ocfs2_unlink_path(inode
, handle
, dealloc
, path
, 1);
2836 el
= et
->et_root_el
;
2837 el
->l_tree_depth
= 0;
2838 el
->l_next_free_rec
= 0;
2839 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2841 ocfs2_et_set_last_eb_blk(et
, 0);
2844 ocfs2_journal_dirty(handle
, path_root_bh(path
));
2847 ocfs2_free_path(left_path
);
2852 * Left rotation of btree records.
2854 * In many ways, this is (unsurprisingly) the opposite of right
2855 * rotation. We start at some non-rightmost path containing an empty
2856 * extent in the leaf block. The code works its way to the rightmost
2857 * path by rotating records to the left in every subtree.
2859 * This is used by any code which reduces the number of extent records
2860 * in a leaf. After removal, an empty record should be placed in the
2861 * leftmost list position.
2863 * This won't handle a length update of the rightmost path records if
2864 * the rightmost tree leaf record is removed so the caller is
2865 * responsible for detecting and correcting that.
2867 static int ocfs2_rotate_tree_left(struct inode
*inode
, handle_t
*handle
,
2868 struct ocfs2_path
*path
,
2869 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2870 struct ocfs2_extent_tree
*et
)
2872 int ret
, orig_credits
= handle
->h_buffer_credits
;
2873 struct ocfs2_path
*tmp_path
= NULL
, *restart_path
= NULL
;
2874 struct ocfs2_extent_block
*eb
;
2875 struct ocfs2_extent_list
*el
;
2877 el
= path_leaf_el(path
);
2878 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2881 if (path
->p_tree_depth
== 0) {
2882 rightmost_no_delete
:
2884 * Inline extents. This is trivially handled, so do
2887 ret
= ocfs2_rotate_rightmost_leaf_left(inode
, handle
,
2889 path_leaf_el(path
));
2896 * Handle rightmost branch now. There's several cases:
2897 * 1) simple rotation leaving records in there. That's trivial.
2898 * 2) rotation requiring a branch delete - there's no more
2899 * records left. Two cases of this:
2900 * a) There are branches to the left.
2901 * b) This is also the leftmost (the only) branch.
2903 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
2904 * 2a) we need the left branch so that we can update it with the unlink
2905 * 2b) we need to bring the inode back to inline extents.
2908 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2910 if (eb
->h_next_leaf_blk
== 0) {
2912 * This gets a bit tricky if we're going to delete the
2913 * rightmost path. Get the other cases out of the way
2916 if (le16_to_cpu(el
->l_next_free_rec
) > 1)
2917 goto rightmost_no_delete
;
2919 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
2921 ocfs2_error(inode
->i_sb
,
2922 "Inode %llu has empty extent block at %llu",
2923 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2924 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
2929 * XXX: The caller can not trust "path" any more after
2930 * this as it will have been deleted. What do we do?
2932 * In theory the rotate-for-merge code will never get
2933 * here because it'll always ask for a rotate in a
2937 ret
= ocfs2_remove_rightmost_path(inode
, handle
, path
,
2945 * Now we can loop, remembering the path we get from -EAGAIN
2946 * and restarting from there.
2949 ret
= __ocfs2_rotate_tree_left(inode
, handle
, orig_credits
, path
,
2950 dealloc
, &restart_path
, et
);
2951 if (ret
&& ret
!= -EAGAIN
) {
2956 while (ret
== -EAGAIN
) {
2957 tmp_path
= restart_path
;
2958 restart_path
= NULL
;
2960 ret
= __ocfs2_rotate_tree_left(inode
, handle
, orig_credits
,
2963 if (ret
&& ret
!= -EAGAIN
) {
2968 ocfs2_free_path(tmp_path
);
2976 ocfs2_free_path(tmp_path
);
2977 ocfs2_free_path(restart_path
);
2981 static void ocfs2_cleanup_merge(struct ocfs2_extent_list
*el
,
2984 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[index
];
2987 if (rec
->e_leaf_clusters
== 0) {
2989 * We consumed all of the merged-from record. An empty
2990 * extent cannot exist anywhere but the 1st array
2991 * position, so move things over if the merged-from
2992 * record doesn't occupy that position.
2994 * This creates a new empty extent so the caller
2995 * should be smart enough to have removed any existing
2999 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3000 size
= index
* sizeof(struct ocfs2_extent_rec
);
3001 memmove(&el
->l_recs
[1], &el
->l_recs
[0], size
);
3005 * Always memset - the caller doesn't check whether it
3006 * created an empty extent, so there could be junk in
3009 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3013 static int ocfs2_get_right_path(struct inode
*inode
,
3014 struct ocfs2_path
*left_path
,
3015 struct ocfs2_path
**ret_right_path
)
3019 struct ocfs2_path
*right_path
= NULL
;
3020 struct ocfs2_extent_list
*left_el
;
3022 *ret_right_path
= NULL
;
3024 /* This function shouldn't be called for non-trees. */
3025 BUG_ON(left_path
->p_tree_depth
== 0);
3027 left_el
= path_leaf_el(left_path
);
3028 BUG_ON(left_el
->l_next_free_rec
!= left_el
->l_count
);
3030 ret
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
, left_path
,
3037 /* This function shouldn't be called for the rightmost leaf. */
3038 BUG_ON(right_cpos
== 0);
3040 right_path
= ocfs2_new_path(path_root_bh(left_path
),
3041 path_root_el(left_path
));
3048 ret
= ocfs2_find_path(inode
, right_path
, right_cpos
);
3054 *ret_right_path
= right_path
;
3057 ocfs2_free_path(right_path
);
3062 * Remove split_rec clusters from the record at index and merge them
3063 * onto the beginning of the record "next" to it.
3064 * For index < l_count - 1, the next means the extent rec at index + 1.
3065 * For index == l_count - 1, the "next" means the 1st extent rec of the
3066 * next extent block.
3068 static int ocfs2_merge_rec_right(struct inode
*inode
,
3069 struct ocfs2_path
*left_path
,
3071 struct ocfs2_extent_rec
*split_rec
,
3074 int ret
, next_free
, i
;
3075 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3076 struct ocfs2_extent_rec
*left_rec
;
3077 struct ocfs2_extent_rec
*right_rec
;
3078 struct ocfs2_extent_list
*right_el
;
3079 struct ocfs2_path
*right_path
= NULL
;
3080 int subtree_index
= 0;
3081 struct ocfs2_extent_list
*el
= path_leaf_el(left_path
);
3082 struct buffer_head
*bh
= path_leaf_bh(left_path
);
3083 struct buffer_head
*root_bh
= NULL
;
3085 BUG_ON(index
>= le16_to_cpu(el
->l_next_free_rec
));
3086 left_rec
= &el
->l_recs
[index
];
3088 if (index
== le16_to_cpu(el
->l_next_free_rec
) - 1 &&
3089 le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
)) {
3090 /* we meet with a cross extent block merge. */
3091 ret
= ocfs2_get_right_path(inode
, left_path
, &right_path
);
3097 right_el
= path_leaf_el(right_path
);
3098 next_free
= le16_to_cpu(right_el
->l_next_free_rec
);
3099 BUG_ON(next_free
<= 0);
3100 right_rec
= &right_el
->l_recs
[0];
3101 if (ocfs2_is_empty_extent(right_rec
)) {
3102 BUG_ON(next_free
<= 1);
3103 right_rec
= &right_el
->l_recs
[1];
3106 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3107 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3108 le32_to_cpu(right_rec
->e_cpos
));
3110 subtree_index
= ocfs2_find_subtree_root(inode
,
3111 left_path
, right_path
);
3113 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3114 handle
->h_buffer_credits
,
3121 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3122 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3124 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
3125 OCFS2_JOURNAL_ACCESS_WRITE
);
3131 for (i
= subtree_index
+ 1;
3132 i
< path_num_items(right_path
); i
++) {
3133 ret
= ocfs2_journal_access(handle
, inode
,
3134 right_path
->p_node
[i
].bh
,
3135 OCFS2_JOURNAL_ACCESS_WRITE
);
3141 ret
= ocfs2_journal_access(handle
, inode
,
3142 left_path
->p_node
[i
].bh
,
3143 OCFS2_JOURNAL_ACCESS_WRITE
);
3151 BUG_ON(index
== le16_to_cpu(el
->l_next_free_rec
) - 1);
3152 right_rec
= &el
->l_recs
[index
+ 1];
3155 ret
= ocfs2_journal_access(handle
, inode
, bh
,
3156 OCFS2_JOURNAL_ACCESS_WRITE
);
3162 le16_add_cpu(&left_rec
->e_leaf_clusters
, -split_clusters
);
3164 le32_add_cpu(&right_rec
->e_cpos
, -split_clusters
);
3165 le64_add_cpu(&right_rec
->e_blkno
,
3166 -ocfs2_clusters_to_blocks(inode
->i_sb
, split_clusters
));
3167 le16_add_cpu(&right_rec
->e_leaf_clusters
, split_clusters
);
3169 ocfs2_cleanup_merge(el
, index
);
3171 ret
= ocfs2_journal_dirty(handle
, bh
);
3176 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
3180 ocfs2_complete_edge_insert(inode
, handle
, left_path
,
3181 right_path
, subtree_index
);
3185 ocfs2_free_path(right_path
);
3189 static int ocfs2_get_left_path(struct inode
*inode
,
3190 struct ocfs2_path
*right_path
,
3191 struct ocfs2_path
**ret_left_path
)
3195 struct ocfs2_path
*left_path
= NULL
;
3197 *ret_left_path
= NULL
;
3199 /* This function shouldn't be called for non-trees. */
3200 BUG_ON(right_path
->p_tree_depth
== 0);
3202 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
,
3203 right_path
, &left_cpos
);
3209 /* This function shouldn't be called for the leftmost leaf. */
3210 BUG_ON(left_cpos
== 0);
3212 left_path
= ocfs2_new_path(path_root_bh(right_path
),
3213 path_root_el(right_path
));
3220 ret
= ocfs2_find_path(inode
, left_path
, left_cpos
);
3226 *ret_left_path
= left_path
;
3229 ocfs2_free_path(left_path
);
3234 * Remove split_rec clusters from the record at index and merge them
3235 * onto the tail of the record "before" it.
3236 * For index > 0, the "before" means the extent rec at index - 1.
3238 * For index == 0, the "before" means the last record of the previous
3239 * extent block. And there is also a situation that we may need to
3240 * remove the rightmost leaf extent block in the right_path and change
3241 * the right path to indicate the new rightmost path.
3243 static int ocfs2_merge_rec_left(struct inode
*inode
,
3244 struct ocfs2_path
*right_path
,
3246 struct ocfs2_extent_rec
*split_rec
,
3247 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3248 struct ocfs2_extent_tree
*et
,
3251 int ret
, i
, subtree_index
= 0, has_empty_extent
= 0;
3252 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3253 struct ocfs2_extent_rec
*left_rec
;
3254 struct ocfs2_extent_rec
*right_rec
;
3255 struct ocfs2_extent_list
*el
= path_leaf_el(right_path
);
3256 struct buffer_head
*bh
= path_leaf_bh(right_path
);
3257 struct buffer_head
*root_bh
= NULL
;
3258 struct ocfs2_path
*left_path
= NULL
;
3259 struct ocfs2_extent_list
*left_el
;
3263 right_rec
= &el
->l_recs
[index
];
3265 /* we meet with a cross extent block merge. */
3266 ret
= ocfs2_get_left_path(inode
, right_path
, &left_path
);
3272 left_el
= path_leaf_el(left_path
);
3273 BUG_ON(le16_to_cpu(left_el
->l_next_free_rec
) !=
3274 le16_to_cpu(left_el
->l_count
));
3276 left_rec
= &left_el
->l_recs
[
3277 le16_to_cpu(left_el
->l_next_free_rec
) - 1];
3278 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3279 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3280 le32_to_cpu(split_rec
->e_cpos
));
3282 subtree_index
= ocfs2_find_subtree_root(inode
,
3283 left_path
, right_path
);
3285 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3286 handle
->h_buffer_credits
,
3293 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3294 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3296 ret
= ocfs2_journal_access(handle
, inode
, root_bh
,
3297 OCFS2_JOURNAL_ACCESS_WRITE
);
3303 for (i
= subtree_index
+ 1;
3304 i
< path_num_items(right_path
); i
++) {
3305 ret
= ocfs2_journal_access(handle
, inode
,
3306 right_path
->p_node
[i
].bh
,
3307 OCFS2_JOURNAL_ACCESS_WRITE
);
3313 ret
= ocfs2_journal_access(handle
, inode
,
3314 left_path
->p_node
[i
].bh
,
3315 OCFS2_JOURNAL_ACCESS_WRITE
);
3322 left_rec
= &el
->l_recs
[index
- 1];
3323 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
3324 has_empty_extent
= 1;
3327 ret
= ocfs2_journal_access(handle
, inode
, bh
,
3328 OCFS2_JOURNAL_ACCESS_WRITE
);
3334 if (has_empty_extent
&& index
== 1) {
3336 * The easy case - we can just plop the record right in.
3338 *left_rec
= *split_rec
;
3340 has_empty_extent
= 0;
3342 le16_add_cpu(&left_rec
->e_leaf_clusters
, split_clusters
);
3344 le32_add_cpu(&right_rec
->e_cpos
, split_clusters
);
3345 le64_add_cpu(&right_rec
->e_blkno
,
3346 ocfs2_clusters_to_blocks(inode
->i_sb
, split_clusters
));
3347 le16_add_cpu(&right_rec
->e_leaf_clusters
, -split_clusters
);
3349 ocfs2_cleanup_merge(el
, index
);
3351 ret
= ocfs2_journal_dirty(handle
, bh
);
3356 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
3361 * In the situation that the right_rec is empty and the extent
3362 * block is empty also, ocfs2_complete_edge_insert can't handle
3363 * it and we need to delete the right extent block.
3365 if (le16_to_cpu(right_rec
->e_leaf_clusters
) == 0 &&
3366 le16_to_cpu(el
->l_next_free_rec
) == 1) {
3368 ret
= ocfs2_remove_rightmost_path(inode
, handle
,
3376 /* Now the rightmost extent block has been deleted.
3377 * So we use the new rightmost path.
3379 ocfs2_mv_path(right_path
, left_path
);
3382 ocfs2_complete_edge_insert(inode
, handle
, left_path
,
3383 right_path
, subtree_index
);
3387 ocfs2_free_path(left_path
);
3391 static int ocfs2_try_to_merge_extent(struct inode
*inode
,
3393 struct ocfs2_path
*path
,
3395 struct ocfs2_extent_rec
*split_rec
,
3396 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3397 struct ocfs2_merge_ctxt
*ctxt
,
3398 struct ocfs2_extent_tree
*et
)
3402 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
3403 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
3405 BUG_ON(ctxt
->c_contig_type
== CONTIG_NONE
);
3407 if (ctxt
->c_split_covers_rec
&& ctxt
->c_has_empty_extent
) {
3409 * The merge code will need to create an empty
3410 * extent to take the place of the newly
3411 * emptied slot. Remove any pre-existing empty
3412 * extents - having more than one in a leaf is
3415 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
,
3422 rec
= &el
->l_recs
[split_index
];
3425 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
) {
3427 * Left-right contig implies this.
3429 BUG_ON(!ctxt
->c_split_covers_rec
);
3432 * Since the leftright insert always covers the entire
3433 * extent, this call will delete the insert record
3434 * entirely, resulting in an empty extent record added to
3437 * Since the adding of an empty extent shifts
3438 * everything back to the right, there's no need to
3439 * update split_index here.
3441 * When the split_index is zero, we need to merge it to the
3442 * prevoius extent block. It is more efficient and easier
3443 * if we do merge_right first and merge_left later.
3445 ret
= ocfs2_merge_rec_right(inode
, path
,
3454 * We can only get this from logic error above.
3456 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
3458 /* The merge left us with an empty extent, remove it. */
3459 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
,
3466 rec
= &el
->l_recs
[split_index
];
3469 * Note that we don't pass split_rec here on purpose -
3470 * we've merged it into the rec already.
3472 ret
= ocfs2_merge_rec_left(inode
, path
,
3482 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
,
3485 * Error from this last rotate is not critical, so
3486 * print but don't bubble it up.
3493 * Merge a record to the left or right.
3495 * 'contig_type' is relative to the existing record,
3496 * so for example, if we're "right contig", it's to
3497 * the record on the left (hence the left merge).
3499 if (ctxt
->c_contig_type
== CONTIG_RIGHT
) {
3500 ret
= ocfs2_merge_rec_left(inode
,
3510 ret
= ocfs2_merge_rec_right(inode
,
3520 if (ctxt
->c_split_covers_rec
) {
3522 * The merge may have left an empty extent in
3523 * our leaf. Try to rotate it away.
3525 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
,
3537 static void ocfs2_subtract_from_rec(struct super_block
*sb
,
3538 enum ocfs2_split_type split
,
3539 struct ocfs2_extent_rec
*rec
,
3540 struct ocfs2_extent_rec
*split_rec
)
3544 len_blocks
= ocfs2_clusters_to_blocks(sb
,
3545 le16_to_cpu(split_rec
->e_leaf_clusters
));
3547 if (split
== SPLIT_LEFT
) {
3549 * Region is on the left edge of the existing
3552 le32_add_cpu(&rec
->e_cpos
,
3553 le16_to_cpu(split_rec
->e_leaf_clusters
));
3554 le64_add_cpu(&rec
->e_blkno
, len_blocks
);
3555 le16_add_cpu(&rec
->e_leaf_clusters
,
3556 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3559 * Region is on the right edge of the existing
3562 le16_add_cpu(&rec
->e_leaf_clusters
,
3563 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3568 * Do the final bits of extent record insertion at the target leaf
3569 * list. If this leaf is part of an allocation tree, it is assumed
3570 * that the tree above has been prepared.
3572 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec
*insert_rec
,
3573 struct ocfs2_extent_list
*el
,
3574 struct ocfs2_insert_type
*insert
,
3575 struct inode
*inode
)
3577 int i
= insert
->ins_contig_index
;
3579 struct ocfs2_extent_rec
*rec
;
3581 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3583 if (insert
->ins_split
!= SPLIT_NONE
) {
3584 i
= ocfs2_search_extent_list(el
, le32_to_cpu(insert_rec
->e_cpos
));
3586 rec
= &el
->l_recs
[i
];
3587 ocfs2_subtract_from_rec(inode
->i_sb
, insert
->ins_split
, rec
,
3593 * Contiguous insert - either left or right.
3595 if (insert
->ins_contig
!= CONTIG_NONE
) {
3596 rec
= &el
->l_recs
[i
];
3597 if (insert
->ins_contig
== CONTIG_LEFT
) {
3598 rec
->e_blkno
= insert_rec
->e_blkno
;
3599 rec
->e_cpos
= insert_rec
->e_cpos
;
3601 le16_add_cpu(&rec
->e_leaf_clusters
,
3602 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3607 * Handle insert into an empty leaf.
3609 if (le16_to_cpu(el
->l_next_free_rec
) == 0 ||
3610 ((le16_to_cpu(el
->l_next_free_rec
) == 1) &&
3611 ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3612 el
->l_recs
[0] = *insert_rec
;
3613 el
->l_next_free_rec
= cpu_to_le16(1);
3620 if (insert
->ins_appending
== APPEND_TAIL
) {
3621 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3622 rec
= &el
->l_recs
[i
];
3623 range
= le32_to_cpu(rec
->e_cpos
)
3624 + le16_to_cpu(rec
->e_leaf_clusters
);
3625 BUG_ON(le32_to_cpu(insert_rec
->e_cpos
) < range
);
3627 mlog_bug_on_msg(le16_to_cpu(el
->l_next_free_rec
) >=
3628 le16_to_cpu(el
->l_count
),
3629 "inode %lu, depth %u, count %u, next free %u, "
3630 "rec.cpos %u, rec.clusters %u, "
3631 "insert.cpos %u, insert.clusters %u\n",
3633 le16_to_cpu(el
->l_tree_depth
),
3634 le16_to_cpu(el
->l_count
),
3635 le16_to_cpu(el
->l_next_free_rec
),
3636 le32_to_cpu(el
->l_recs
[i
].e_cpos
),
3637 le16_to_cpu(el
->l_recs
[i
].e_leaf_clusters
),
3638 le32_to_cpu(insert_rec
->e_cpos
),
3639 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3641 el
->l_recs
[i
] = *insert_rec
;
3642 le16_add_cpu(&el
->l_next_free_rec
, 1);
3648 * Ok, we have to rotate.
3650 * At this point, it is safe to assume that inserting into an
3651 * empty leaf and appending to a leaf have both been handled
3654 * This leaf needs to have space, either by the empty 1st
3655 * extent record, or by virtue of an l_next_rec < l_count.
3657 ocfs2_rotate_leaf(el
, insert_rec
);
3660 static void ocfs2_adjust_rightmost_records(struct inode
*inode
,
3662 struct ocfs2_path
*path
,
3663 struct ocfs2_extent_rec
*insert_rec
)
3665 int ret
, i
, next_free
;
3666 struct buffer_head
*bh
;
3667 struct ocfs2_extent_list
*el
;
3668 struct ocfs2_extent_rec
*rec
;
3671 * Update everything except the leaf block.
3673 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
3674 bh
= path
->p_node
[i
].bh
;
3675 el
= path
->p_node
[i
].el
;
3677 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3678 if (next_free
== 0) {
3679 ocfs2_error(inode
->i_sb
,
3680 "Dinode %llu has a bad extent list",
3681 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
3686 rec
= &el
->l_recs
[next_free
- 1];
3688 rec
->e_int_clusters
= insert_rec
->e_cpos
;
3689 le32_add_cpu(&rec
->e_int_clusters
,
3690 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3691 le32_add_cpu(&rec
->e_int_clusters
,
3692 -le32_to_cpu(rec
->e_cpos
));
3694 ret
= ocfs2_journal_dirty(handle
, bh
);
3701 static int ocfs2_append_rec_to_path(struct inode
*inode
, handle_t
*handle
,
3702 struct ocfs2_extent_rec
*insert_rec
,
3703 struct ocfs2_path
*right_path
,
3704 struct ocfs2_path
**ret_left_path
)
3707 struct ocfs2_extent_list
*el
;
3708 struct ocfs2_path
*left_path
= NULL
;
3710 *ret_left_path
= NULL
;
3713 * This shouldn't happen for non-trees. The extent rec cluster
3714 * count manipulation below only works for interior nodes.
3716 BUG_ON(right_path
->p_tree_depth
== 0);
3719 * If our appending insert is at the leftmost edge of a leaf,
3720 * then we might need to update the rightmost records of the
3723 el
= path_leaf_el(right_path
);
3724 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3725 if (next_free
== 0 ||
3726 (next_free
== 1 && ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3729 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, right_path
,
3736 mlog(0, "Append may need a left path update. cpos: %u, "
3737 "left_cpos: %u\n", le32_to_cpu(insert_rec
->e_cpos
),
3741 * No need to worry if the append is already in the
3745 left_path
= ocfs2_new_path(path_root_bh(right_path
),
3746 path_root_el(right_path
));
3753 ret
= ocfs2_find_path(inode
, left_path
, left_cpos
);
3760 * ocfs2_insert_path() will pass the left_path to the
3766 ret
= ocfs2_journal_access_path(inode
, handle
, right_path
);
3772 ocfs2_adjust_rightmost_records(inode
, handle
, right_path
, insert_rec
);
3774 *ret_left_path
= left_path
;
3778 ocfs2_free_path(left_path
);
3783 static void ocfs2_split_record(struct inode
*inode
,
3784 struct ocfs2_path
*left_path
,
3785 struct ocfs2_path
*right_path
,
3786 struct ocfs2_extent_rec
*split_rec
,
3787 enum ocfs2_split_type split
)
3790 u32 cpos
= le32_to_cpu(split_rec
->e_cpos
);
3791 struct ocfs2_extent_list
*left_el
= NULL
, *right_el
, *insert_el
, *el
;
3792 struct ocfs2_extent_rec
*rec
, *tmprec
;
3794 right_el
= path_leaf_el(right_path
);;
3796 left_el
= path_leaf_el(left_path
);
3799 insert_el
= right_el
;
3800 index
= ocfs2_search_extent_list(el
, cpos
);
3802 if (index
== 0 && left_path
) {
3803 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3806 * This typically means that the record
3807 * started in the left path but moved to the
3808 * right as a result of rotation. We either
3809 * move the existing record to the left, or we
3810 * do the later insert there.
3812 * In this case, the left path should always
3813 * exist as the rotate code will have passed
3814 * it back for a post-insert update.
3817 if (split
== SPLIT_LEFT
) {
3819 * It's a left split. Since we know
3820 * that the rotate code gave us an
3821 * empty extent in the left path, we
3822 * can just do the insert there.
3824 insert_el
= left_el
;
3827 * Right split - we have to move the
3828 * existing record over to the left
3829 * leaf. The insert will be into the
3830 * newly created empty extent in the
3833 tmprec
= &right_el
->l_recs
[index
];
3834 ocfs2_rotate_leaf(left_el
, tmprec
);
3837 memset(tmprec
, 0, sizeof(*tmprec
));
3838 index
= ocfs2_search_extent_list(left_el
, cpos
);
3839 BUG_ON(index
== -1);
3844 BUG_ON(!ocfs2_is_empty_extent(&left_el
->l_recs
[0]));
3846 * Left path is easy - we can just allow the insert to
3850 insert_el
= left_el
;
3851 index
= ocfs2_search_extent_list(el
, cpos
);
3852 BUG_ON(index
== -1);
3855 rec
= &el
->l_recs
[index
];
3856 ocfs2_subtract_from_rec(inode
->i_sb
, split
, rec
, split_rec
);
3857 ocfs2_rotate_leaf(insert_el
, split_rec
);
3861 * This function only does inserts on an allocation b-tree. For tree
3862 * depth = 0, ocfs2_insert_at_leaf() is called directly.
3864 * right_path is the path we want to do the actual insert
3865 * in. left_path should only be passed in if we need to update that
3866 * portion of the tree after an edge insert.
3868 static int ocfs2_insert_path(struct inode
*inode
,
3870 struct ocfs2_path
*left_path
,
3871 struct ocfs2_path
*right_path
,
3872 struct ocfs2_extent_rec
*insert_rec
,
3873 struct ocfs2_insert_type
*insert
)
3875 int ret
, subtree_index
;
3876 struct buffer_head
*leaf_bh
= path_leaf_bh(right_path
);
3879 int credits
= handle
->h_buffer_credits
;
3882 * There's a chance that left_path got passed back to
3883 * us without being accounted for in the
3884 * journal. Extend our transaction here to be sure we
3885 * can change those blocks.
3887 credits
+= left_path
->p_tree_depth
;
3889 ret
= ocfs2_extend_trans(handle
, credits
);
3895 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
3903 * Pass both paths to the journal. The majority of inserts
3904 * will be touching all components anyway.
3906 ret
= ocfs2_journal_access_path(inode
, handle
, right_path
);
3912 if (insert
->ins_split
!= SPLIT_NONE
) {
3914 * We could call ocfs2_insert_at_leaf() for some types
3915 * of splits, but it's easier to just let one separate
3916 * function sort it all out.
3918 ocfs2_split_record(inode
, left_path
, right_path
,
3919 insert_rec
, insert
->ins_split
);
3922 * Split might have modified either leaf and we don't
3923 * have a guarantee that the later edge insert will
3924 * dirty this for us.
3927 ret
= ocfs2_journal_dirty(handle
,
3928 path_leaf_bh(left_path
));
3932 ocfs2_insert_at_leaf(insert_rec
, path_leaf_el(right_path
),
3935 ret
= ocfs2_journal_dirty(handle
, leaf_bh
);
3941 * The rotate code has indicated that we need to fix
3942 * up portions of the tree after the insert.
3944 * XXX: Should we extend the transaction here?
3946 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
,
3948 ocfs2_complete_edge_insert(inode
, handle
, left_path
,
3949 right_path
, subtree_index
);
3957 static int ocfs2_do_insert_extent(struct inode
*inode
,
3959 struct ocfs2_extent_tree
*et
,
3960 struct ocfs2_extent_rec
*insert_rec
,
3961 struct ocfs2_insert_type
*type
)
3963 int ret
, rotate
= 0;
3965 struct ocfs2_path
*right_path
= NULL
;
3966 struct ocfs2_path
*left_path
= NULL
;
3967 struct ocfs2_extent_list
*el
;
3969 el
= et
->et_root_el
;
3971 ret
= ocfs2_journal_access(handle
, inode
, et
->et_root_bh
,
3972 OCFS2_JOURNAL_ACCESS_WRITE
);
3978 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
3979 ocfs2_insert_at_leaf(insert_rec
, el
, type
, inode
);
3980 goto out_update_clusters
;
3983 right_path
= ocfs2_new_path(et
->et_root_bh
, et
->et_root_el
);
3991 * Determine the path to start with. Rotations need the
3992 * rightmost path, everything else can go directly to the
3995 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
3996 if (type
->ins_appending
== APPEND_NONE
&&
3997 type
->ins_contig
== CONTIG_NONE
) {
4002 ret
= ocfs2_find_path(inode
, right_path
, cpos
);
4009 * Rotations and appends need special treatment - they modify
4010 * parts of the tree's above them.
4012 * Both might pass back a path immediate to the left of the
4013 * one being inserted to. This will be cause
4014 * ocfs2_insert_path() to modify the rightmost records of
4015 * left_path to account for an edge insert.
4017 * XXX: When modifying this code, keep in mind that an insert
4018 * can wind up skipping both of these two special cases...
4021 ret
= ocfs2_rotate_tree_right(inode
, handle
, type
->ins_split
,
4022 le32_to_cpu(insert_rec
->e_cpos
),
4023 right_path
, &left_path
);
4030 * ocfs2_rotate_tree_right() might have extended the
4031 * transaction without re-journaling our tree root.
4033 ret
= ocfs2_journal_access(handle
, inode
, et
->et_root_bh
,
4034 OCFS2_JOURNAL_ACCESS_WRITE
);
4039 } else if (type
->ins_appending
== APPEND_TAIL
4040 && type
->ins_contig
!= CONTIG_LEFT
) {
4041 ret
= ocfs2_append_rec_to_path(inode
, handle
, insert_rec
,
4042 right_path
, &left_path
);
4049 ret
= ocfs2_insert_path(inode
, handle
, left_path
, right_path
,
4056 out_update_clusters
:
4057 if (type
->ins_split
== SPLIT_NONE
)
4058 ocfs2_et_update_clusters(inode
, et
,
4059 le16_to_cpu(insert_rec
->e_leaf_clusters
));
4061 ret
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4066 ocfs2_free_path(left_path
);
4067 ocfs2_free_path(right_path
);
4072 static enum ocfs2_contig_type
4073 ocfs2_figure_merge_contig_type(struct inode
*inode
, struct ocfs2_path
*path
,
4074 struct ocfs2_extent_list
*el
, int index
,
4075 struct ocfs2_extent_rec
*split_rec
)
4078 enum ocfs2_contig_type ret
= CONTIG_NONE
;
4079 u32 left_cpos
, right_cpos
;
4080 struct ocfs2_extent_rec
*rec
= NULL
;
4081 struct ocfs2_extent_list
*new_el
;
4082 struct ocfs2_path
*left_path
= NULL
, *right_path
= NULL
;
4083 struct buffer_head
*bh
;
4084 struct ocfs2_extent_block
*eb
;
4087 rec
= &el
->l_recs
[index
- 1];
4088 } else if (path
->p_tree_depth
> 0) {
4089 status
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
,
4094 if (left_cpos
!= 0) {
4095 left_path
= ocfs2_new_path(path_root_bh(path
),
4096 path_root_el(path
));
4100 status
= ocfs2_find_path(inode
, left_path
, left_cpos
);
4104 new_el
= path_leaf_el(left_path
);
4106 if (le16_to_cpu(new_el
->l_next_free_rec
) !=
4107 le16_to_cpu(new_el
->l_count
)) {
4108 bh
= path_leaf_bh(left_path
);
4109 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4110 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
,
4114 rec
= &new_el
->l_recs
[
4115 le16_to_cpu(new_el
->l_next_free_rec
) - 1];
4120 * We're careful to check for an empty extent record here -
4121 * the merge code will know what to do if it sees one.
4124 if (index
== 1 && ocfs2_is_empty_extent(rec
)) {
4125 if (split_rec
->e_cpos
== el
->l_recs
[index
].e_cpos
)
4128 ret
= ocfs2_extent_contig(inode
, rec
, split_rec
);
4133 if (index
< (le16_to_cpu(el
->l_next_free_rec
) - 1))
4134 rec
= &el
->l_recs
[index
+ 1];
4135 else if (le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
) &&
4136 path
->p_tree_depth
> 0) {
4137 status
= ocfs2_find_cpos_for_right_leaf(inode
->i_sb
,
4142 if (right_cpos
== 0)
4145 right_path
= ocfs2_new_path(path_root_bh(path
),
4146 path_root_el(path
));
4150 status
= ocfs2_find_path(inode
, right_path
, right_cpos
);
4154 new_el
= path_leaf_el(right_path
);
4155 rec
= &new_el
->l_recs
[0];
4156 if (ocfs2_is_empty_extent(rec
)) {
4157 if (le16_to_cpu(new_el
->l_next_free_rec
) <= 1) {
4158 bh
= path_leaf_bh(right_path
);
4159 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4160 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
,
4164 rec
= &new_el
->l_recs
[1];
4169 enum ocfs2_contig_type contig_type
;
4171 contig_type
= ocfs2_extent_contig(inode
, rec
, split_rec
);
4173 if (contig_type
== CONTIG_LEFT
&& ret
== CONTIG_RIGHT
)
4174 ret
= CONTIG_LEFTRIGHT
;
4175 else if (ret
== CONTIG_NONE
)
4181 ocfs2_free_path(left_path
);
4183 ocfs2_free_path(right_path
);
4188 static void ocfs2_figure_contig_type(struct inode
*inode
,
4189 struct ocfs2_insert_type
*insert
,
4190 struct ocfs2_extent_list
*el
,
4191 struct ocfs2_extent_rec
*insert_rec
,
4192 struct ocfs2_extent_tree
*et
)
4195 enum ocfs2_contig_type contig_type
= CONTIG_NONE
;
4197 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4199 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
4200 contig_type
= ocfs2_extent_contig(inode
, &el
->l_recs
[i
],
4202 if (contig_type
!= CONTIG_NONE
) {
4203 insert
->ins_contig_index
= i
;
4207 insert
->ins_contig
= contig_type
;
4209 if (insert
->ins_contig
!= CONTIG_NONE
) {
4210 struct ocfs2_extent_rec
*rec
=
4211 &el
->l_recs
[insert
->ins_contig_index
];
4212 unsigned int len
= le16_to_cpu(rec
->e_leaf_clusters
) +
4213 le16_to_cpu(insert_rec
->e_leaf_clusters
);
4216 * Caller might want us to limit the size of extents, don't
4217 * calculate contiguousness if we might exceed that limit.
4219 if (et
->et_max_leaf_clusters
&&
4220 (len
> et
->et_max_leaf_clusters
))
4221 insert
->ins_contig
= CONTIG_NONE
;
4226 * This should only be called against the righmost leaf extent list.
4228 * ocfs2_figure_appending_type() will figure out whether we'll have to
4229 * insert at the tail of the rightmost leaf.
4231 * This should also work against the root extent list for tree's with 0
4232 * depth. If we consider the root extent list to be the rightmost leaf node
4233 * then the logic here makes sense.
4235 static void ocfs2_figure_appending_type(struct ocfs2_insert_type
*insert
,
4236 struct ocfs2_extent_list
*el
,
4237 struct ocfs2_extent_rec
*insert_rec
)
4240 u32 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4241 struct ocfs2_extent_rec
*rec
;
4243 insert
->ins_appending
= APPEND_NONE
;
4245 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4247 if (!el
->l_next_free_rec
)
4248 goto set_tail_append
;
4250 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
4251 /* Were all records empty? */
4252 if (le16_to_cpu(el
->l_next_free_rec
) == 1)
4253 goto set_tail_append
;
4256 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
4257 rec
= &el
->l_recs
[i
];
4260 (le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)))
4261 goto set_tail_append
;
4266 insert
->ins_appending
= APPEND_TAIL
;
4270 * Helper function called at the begining of an insert.
4272 * This computes a few things that are commonly used in the process of
4273 * inserting into the btree:
4274 * - Whether the new extent is contiguous with an existing one.
4275 * - The current tree depth.
4276 * - Whether the insert is an appending one.
4277 * - The total # of free records in the tree.
4279 * All of the information is stored on the ocfs2_insert_type
4282 static int ocfs2_figure_insert_type(struct inode
*inode
,
4283 struct ocfs2_extent_tree
*et
,
4284 struct buffer_head
**last_eb_bh
,
4285 struct ocfs2_extent_rec
*insert_rec
,
4287 struct ocfs2_insert_type
*insert
)
4290 struct ocfs2_extent_block
*eb
;
4291 struct ocfs2_extent_list
*el
;
4292 struct ocfs2_path
*path
= NULL
;
4293 struct buffer_head
*bh
= NULL
;
4295 insert
->ins_split
= SPLIT_NONE
;
4297 el
= et
->et_root_el
;
4298 insert
->ins_tree_depth
= le16_to_cpu(el
->l_tree_depth
);
4300 if (el
->l_tree_depth
) {
4302 * If we have tree depth, we read in the
4303 * rightmost extent block ahead of time as
4304 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4305 * may want it later.
4307 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
4308 ocfs2_et_get_last_eb_blk(et
), &bh
,
4309 OCFS2_BH_CACHED
, inode
);
4314 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
4319 * Unless we have a contiguous insert, we'll need to know if
4320 * there is room left in our allocation tree for another
4323 * XXX: This test is simplistic, we can search for empty
4324 * extent records too.
4326 *free_records
= le16_to_cpu(el
->l_count
) -
4327 le16_to_cpu(el
->l_next_free_rec
);
4329 if (!insert
->ins_tree_depth
) {
4330 ocfs2_figure_contig_type(inode
, insert
, el
, insert_rec
, et
);
4331 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4335 path
= ocfs2_new_path(et
->et_root_bh
, et
->et_root_el
);
4343 * In the case that we're inserting past what the tree
4344 * currently accounts for, ocfs2_find_path() will return for
4345 * us the rightmost tree path. This is accounted for below in
4346 * the appending code.
4348 ret
= ocfs2_find_path(inode
, path
, le32_to_cpu(insert_rec
->e_cpos
));
4354 el
= path_leaf_el(path
);
4357 * Now that we have the path, there's two things we want to determine:
4358 * 1) Contiguousness (also set contig_index if this is so)
4360 * 2) Are we doing an append? We can trivially break this up
4361 * into two types of appends: simple record append, or a
4362 * rotate inside the tail leaf.
4364 ocfs2_figure_contig_type(inode
, insert
, el
, insert_rec
, et
);
4367 * The insert code isn't quite ready to deal with all cases of
4368 * left contiguousness. Specifically, if it's an insert into
4369 * the 1st record in a leaf, it will require the adjustment of
4370 * cluster count on the last record of the path directly to it's
4371 * left. For now, just catch that case and fool the layers
4372 * above us. This works just fine for tree_depth == 0, which
4373 * is why we allow that above.
4375 if (insert
->ins_contig
== CONTIG_LEFT
&&
4376 insert
->ins_contig_index
== 0)
4377 insert
->ins_contig
= CONTIG_NONE
;
4380 * Ok, so we can simply compare against last_eb to figure out
4381 * whether the path doesn't exist. This will only happen in
4382 * the case that we're doing a tail append, so maybe we can
4383 * take advantage of that information somehow.
4385 if (ocfs2_et_get_last_eb_blk(et
) ==
4386 path_leaf_bh(path
)->b_blocknr
) {
4388 * Ok, ocfs2_find_path() returned us the rightmost
4389 * tree path. This might be an appending insert. There are
4391 * 1) We're doing a true append at the tail:
4392 * -This might even be off the end of the leaf
4393 * 2) We're "appending" by rotating in the tail
4395 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4399 ocfs2_free_path(path
);
4409 * Insert an extent into an inode btree.
4411 * The caller needs to update fe->i_clusters
4413 int ocfs2_insert_extent(struct ocfs2_super
*osb
,
4415 struct inode
*inode
,
4416 struct ocfs2_extent_tree
*et
,
4421 struct ocfs2_alloc_context
*meta_ac
)
4424 int uninitialized_var(free_records
);
4425 struct buffer_head
*last_eb_bh
= NULL
;
4426 struct ocfs2_insert_type insert
= {0, };
4427 struct ocfs2_extent_rec rec
;
4429 mlog(0, "add %u clusters at position %u to inode %llu\n",
4430 new_clusters
, cpos
, (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
4432 memset(&rec
, 0, sizeof(rec
));
4433 rec
.e_cpos
= cpu_to_le32(cpos
);
4434 rec
.e_blkno
= cpu_to_le64(start_blk
);
4435 rec
.e_leaf_clusters
= cpu_to_le16(new_clusters
);
4436 rec
.e_flags
= flags
;
4437 status
= ocfs2_et_insert_check(inode
, et
, &rec
);
4443 status
= ocfs2_figure_insert_type(inode
, et
, &last_eb_bh
, &rec
,
4444 &free_records
, &insert
);
4450 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4451 "Insert.contig_index: %d, Insert.free_records: %d, "
4452 "Insert.tree_depth: %d\n",
4453 insert
.ins_appending
, insert
.ins_contig
, insert
.ins_contig_index
,
4454 free_records
, insert
.ins_tree_depth
);
4456 if (insert
.ins_contig
== CONTIG_NONE
&& free_records
== 0) {
4457 status
= ocfs2_grow_tree(inode
, handle
, et
,
4458 &insert
.ins_tree_depth
, &last_eb_bh
,
4466 /* Finally, we can add clusters. This might rotate the tree for us. */
4467 status
= ocfs2_do_insert_extent(inode
, handle
, et
, &rec
, &insert
);
4470 else if (et
->et_ops
== &ocfs2_dinode_et_ops
)
4471 ocfs2_extent_map_insert_rec(inode
, &rec
);
4482 * Allcate and add clusters into the extent b-tree.
4483 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4484 * The extent b-tree's root is specified by et, and
4485 * it is not limited to the file storage. Any extent tree can use this
4486 * function if it implements the proper ocfs2_extent_tree.
4488 int ocfs2_add_clusters_in_btree(struct ocfs2_super
*osb
,
4489 struct inode
*inode
,
4490 u32
*logical_offset
,
4491 u32 clusters_to_add
,
4493 struct ocfs2_extent_tree
*et
,
4495 struct ocfs2_alloc_context
*data_ac
,
4496 struct ocfs2_alloc_context
*meta_ac
,
4497 enum ocfs2_alloc_restarted
*reason_ret
)
4501 enum ocfs2_alloc_restarted reason
= RESTART_NONE
;
4502 u32 bit_off
, num_bits
;
4506 BUG_ON(!clusters_to_add
);
4509 flags
= OCFS2_EXT_UNWRITTEN
;
4511 free_extents
= ocfs2_num_free_extents(osb
, inode
, et
);
4512 if (free_extents
< 0) {
4513 status
= free_extents
;
4518 /* there are two cases which could cause us to EAGAIN in the
4519 * we-need-more-metadata case:
4520 * 1) we haven't reserved *any*
4521 * 2) we are so fragmented, we've needed to add metadata too
4523 if (!free_extents
&& !meta_ac
) {
4524 mlog(0, "we haven't reserved any metadata!\n");
4526 reason
= RESTART_META
;
4528 } else if ((!free_extents
)
4529 && (ocfs2_alloc_context_bits_left(meta_ac
)
4530 < ocfs2_extend_meta_needed(et
->et_root_el
))) {
4531 mlog(0, "filesystem is really fragmented...\n");
4533 reason
= RESTART_META
;
4537 status
= __ocfs2_claim_clusters(osb
, handle
, data_ac
, 1,
4538 clusters_to_add
, &bit_off
, &num_bits
);
4540 if (status
!= -ENOSPC
)
4545 BUG_ON(num_bits
> clusters_to_add
);
4547 /* reserve our write early -- insert_extent may update the inode */
4548 status
= ocfs2_journal_access(handle
, inode
, et
->et_root_bh
,
4549 OCFS2_JOURNAL_ACCESS_WRITE
);
4555 block
= ocfs2_clusters_to_blocks(osb
->sb
, bit_off
);
4556 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
4557 num_bits
, bit_off
, (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
4558 status
= ocfs2_insert_extent(osb
, handle
, inode
, et
,
4559 *logical_offset
, block
,
4560 num_bits
, flags
, meta_ac
);
4566 status
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4572 clusters_to_add
-= num_bits
;
4573 *logical_offset
+= num_bits
;
4575 if (clusters_to_add
) {
4576 mlog(0, "need to alloc once more, wanted = %u\n",
4579 reason
= RESTART_TRANS
;
4585 *reason_ret
= reason
;
4589 static void ocfs2_make_right_split_rec(struct super_block
*sb
,
4590 struct ocfs2_extent_rec
*split_rec
,
4592 struct ocfs2_extent_rec
*rec
)
4594 u32 rec_cpos
= le32_to_cpu(rec
->e_cpos
);
4595 u32 rec_range
= rec_cpos
+ le16_to_cpu(rec
->e_leaf_clusters
);
4597 memset(split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4599 split_rec
->e_cpos
= cpu_to_le32(cpos
);
4600 split_rec
->e_leaf_clusters
= cpu_to_le16(rec_range
- cpos
);
4602 split_rec
->e_blkno
= rec
->e_blkno
;
4603 le64_add_cpu(&split_rec
->e_blkno
,
4604 ocfs2_clusters_to_blocks(sb
, cpos
- rec_cpos
));
4606 split_rec
->e_flags
= rec
->e_flags
;
4609 static int ocfs2_split_and_insert(struct inode
*inode
,
4611 struct ocfs2_path
*path
,
4612 struct ocfs2_extent_tree
*et
,
4613 struct buffer_head
**last_eb_bh
,
4615 struct ocfs2_extent_rec
*orig_split_rec
,
4616 struct ocfs2_alloc_context
*meta_ac
)
4619 unsigned int insert_range
, rec_range
, do_leftright
= 0;
4620 struct ocfs2_extent_rec tmprec
;
4621 struct ocfs2_extent_list
*rightmost_el
;
4622 struct ocfs2_extent_rec rec
;
4623 struct ocfs2_extent_rec split_rec
= *orig_split_rec
;
4624 struct ocfs2_insert_type insert
;
4625 struct ocfs2_extent_block
*eb
;
4629 * Store a copy of the record on the stack - it might move
4630 * around as the tree is manipulated below.
4632 rec
= path_leaf_el(path
)->l_recs
[split_index
];
4634 rightmost_el
= et
->et_root_el
;
4636 depth
= le16_to_cpu(rightmost_el
->l_tree_depth
);
4638 BUG_ON(!(*last_eb_bh
));
4639 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
4640 rightmost_el
= &eb
->h_list
;
4643 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4644 le16_to_cpu(rightmost_el
->l_count
)) {
4645 ret
= ocfs2_grow_tree(inode
, handle
, et
,
4646 &depth
, last_eb_bh
, meta_ac
);
4653 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4654 insert
.ins_appending
= APPEND_NONE
;
4655 insert
.ins_contig
= CONTIG_NONE
;
4656 insert
.ins_tree_depth
= depth
;
4658 insert_range
= le32_to_cpu(split_rec
.e_cpos
) +
4659 le16_to_cpu(split_rec
.e_leaf_clusters
);
4660 rec_range
= le32_to_cpu(rec
.e_cpos
) +
4661 le16_to_cpu(rec
.e_leaf_clusters
);
4663 if (split_rec
.e_cpos
== rec
.e_cpos
) {
4664 insert
.ins_split
= SPLIT_LEFT
;
4665 } else if (insert_range
== rec_range
) {
4666 insert
.ins_split
= SPLIT_RIGHT
;
4669 * Left/right split. We fake this as a right split
4670 * first and then make a second pass as a left split.
4672 insert
.ins_split
= SPLIT_RIGHT
;
4674 ocfs2_make_right_split_rec(inode
->i_sb
, &tmprec
, insert_range
,
4679 BUG_ON(do_leftright
);
4683 ret
= ocfs2_do_insert_extent(inode
, handle
, et
, &split_rec
, &insert
);
4689 if (do_leftright
== 1) {
4691 struct ocfs2_extent_list
*el
;
4694 split_rec
= *orig_split_rec
;
4696 ocfs2_reinit_path(path
, 1);
4698 cpos
= le32_to_cpu(split_rec
.e_cpos
);
4699 ret
= ocfs2_find_path(inode
, path
, cpos
);
4705 el
= path_leaf_el(path
);
4706 split_index
= ocfs2_search_extent_list(el
, cpos
);
4715 * Mark part or all of the extent record at split_index in the leaf
4716 * pointed to by path as written. This removes the unwritten
4719 * Care is taken to handle contiguousness so as to not grow the tree.
4721 * meta_ac is not strictly necessary - we only truly need it if growth
4722 * of the tree is required. All other cases will degrade into a less
4723 * optimal tree layout.
4725 * last_eb_bh should be the rightmost leaf block for any extent
4726 * btree. Since a split may grow the tree or a merge might shrink it,
4727 * the caller cannot trust the contents of that buffer after this call.
4729 * This code is optimized for readability - several passes might be
4730 * made over certain portions of the tree. All of those blocks will
4731 * have been brought into cache (and pinned via the journal), so the
4732 * extra overhead is not expressed in terms of disk reads.
4734 static int __ocfs2_mark_extent_written(struct inode
*inode
,
4735 struct ocfs2_extent_tree
*et
,
4737 struct ocfs2_path
*path
,
4739 struct ocfs2_extent_rec
*split_rec
,
4740 struct ocfs2_alloc_context
*meta_ac
,
4741 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
4744 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
4745 struct buffer_head
*last_eb_bh
= NULL
;
4746 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
4747 struct ocfs2_merge_ctxt ctxt
;
4748 struct ocfs2_extent_list
*rightmost_el
;
4750 if (!(rec
->e_flags
& OCFS2_EXT_UNWRITTEN
)) {
4756 if (le32_to_cpu(rec
->e_cpos
) > le32_to_cpu(split_rec
->e_cpos
) ||
4757 ((le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)) <
4758 (le32_to_cpu(split_rec
->e_cpos
) + le16_to_cpu(split_rec
->e_leaf_clusters
)))) {
4764 ctxt
.c_contig_type
= ocfs2_figure_merge_contig_type(inode
, path
, el
,
4769 * The core merge / split code wants to know how much room is
4770 * left in this inodes allocation tree, so we pass the
4771 * rightmost extent list.
4773 if (path
->p_tree_depth
) {
4774 struct ocfs2_extent_block
*eb
;
4776 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
4777 ocfs2_et_get_last_eb_blk(et
),
4778 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
4784 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
4785 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
4786 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
4791 rightmost_el
= &eb
->h_list
;
4793 rightmost_el
= path_root_el(path
);
4795 if (rec
->e_cpos
== split_rec
->e_cpos
&&
4796 rec
->e_leaf_clusters
== split_rec
->e_leaf_clusters
)
4797 ctxt
.c_split_covers_rec
= 1;
4799 ctxt
.c_split_covers_rec
= 0;
4801 ctxt
.c_has_empty_extent
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
4803 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
4804 split_index
, ctxt
.c_contig_type
, ctxt
.c_has_empty_extent
,
4805 ctxt
.c_split_covers_rec
);
4807 if (ctxt
.c_contig_type
== CONTIG_NONE
) {
4808 if (ctxt
.c_split_covers_rec
)
4809 el
->l_recs
[split_index
] = *split_rec
;
4811 ret
= ocfs2_split_and_insert(inode
, handle
, path
, et
,
4812 &last_eb_bh
, split_index
,
4813 split_rec
, meta_ac
);
4817 ret
= ocfs2_try_to_merge_extent(inode
, handle
, path
,
4818 split_index
, split_rec
,
4819 dealloc
, &ctxt
, et
);
4830 * Mark the already-existing extent at cpos as written for len clusters.
4832 * If the existing extent is larger than the request, initiate a
4833 * split. An attempt will be made at merging with adjacent extents.
4835 * The caller is responsible for passing down meta_ac if we'll need it.
4837 int ocfs2_mark_extent_written(struct inode
*inode
,
4838 struct ocfs2_extent_tree
*et
,
4839 handle_t
*handle
, u32 cpos
, u32 len
, u32 phys
,
4840 struct ocfs2_alloc_context
*meta_ac
,
4841 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
4844 u64 start_blkno
= ocfs2_clusters_to_blocks(inode
->i_sb
, phys
);
4845 struct ocfs2_extent_rec split_rec
;
4846 struct ocfs2_path
*left_path
= NULL
;
4847 struct ocfs2_extent_list
*el
;
4849 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4850 inode
->i_ino
, cpos
, len
, phys
, (unsigned long long)start_blkno
);
4852 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode
->i_sb
))) {
4853 ocfs2_error(inode
->i_sb
, "Inode %llu has unwritten extents "
4854 "that are being written to, but the feature bit "
4855 "is not set in the super block.",
4856 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
4862 * XXX: This should be fixed up so that we just re-insert the
4863 * next extent records.
4865 * XXX: This is a hack on the extent tree, maybe it should be
4868 if (et
->et_ops
== &ocfs2_dinode_et_ops
)
4869 ocfs2_extent_map_trunc(inode
, 0);
4871 left_path
= ocfs2_new_path(et
->et_root_bh
, et
->et_root_el
);
4878 ret
= ocfs2_find_path(inode
, left_path
, cpos
);
4883 el
= path_leaf_el(left_path
);
4885 index
= ocfs2_search_extent_list(el
, cpos
);
4886 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
4887 ocfs2_error(inode
->i_sb
,
4888 "Inode %llu has an extent at cpos %u which can no "
4889 "longer be found.\n",
4890 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
);
4895 memset(&split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4896 split_rec
.e_cpos
= cpu_to_le32(cpos
);
4897 split_rec
.e_leaf_clusters
= cpu_to_le16(len
);
4898 split_rec
.e_blkno
= cpu_to_le64(start_blkno
);
4899 split_rec
.e_flags
= path_leaf_el(left_path
)->l_recs
[index
].e_flags
;
4900 split_rec
.e_flags
&= ~OCFS2_EXT_UNWRITTEN
;
4902 ret
= __ocfs2_mark_extent_written(inode
, et
, handle
, left_path
,
4903 index
, &split_rec
, meta_ac
,
4909 ocfs2_free_path(left_path
);
4913 static int ocfs2_split_tree(struct inode
*inode
, struct ocfs2_extent_tree
*et
,
4914 handle_t
*handle
, struct ocfs2_path
*path
,
4915 int index
, u32 new_range
,
4916 struct ocfs2_alloc_context
*meta_ac
)
4918 int ret
, depth
, credits
= handle
->h_buffer_credits
;
4919 struct buffer_head
*last_eb_bh
= NULL
;
4920 struct ocfs2_extent_block
*eb
;
4921 struct ocfs2_extent_list
*rightmost_el
, *el
;
4922 struct ocfs2_extent_rec split_rec
;
4923 struct ocfs2_extent_rec
*rec
;
4924 struct ocfs2_insert_type insert
;
4927 * Setup the record to split before we grow the tree.
4929 el
= path_leaf_el(path
);
4930 rec
= &el
->l_recs
[index
];
4931 ocfs2_make_right_split_rec(inode
->i_sb
, &split_rec
, new_range
, rec
);
4933 depth
= path
->p_tree_depth
;
4935 ret
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
4936 ocfs2_et_get_last_eb_blk(et
),
4937 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
4943 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
4944 rightmost_el
= &eb
->h_list
;
4946 rightmost_el
= path_leaf_el(path
);
4948 credits
+= path
->p_tree_depth
+
4949 ocfs2_extend_meta_needed(et
->et_root_el
);
4950 ret
= ocfs2_extend_trans(handle
, credits
);
4956 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4957 le16_to_cpu(rightmost_el
->l_count
)) {
4958 ret
= ocfs2_grow_tree(inode
, handle
, et
, &depth
, &last_eb_bh
,
4966 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4967 insert
.ins_appending
= APPEND_NONE
;
4968 insert
.ins_contig
= CONTIG_NONE
;
4969 insert
.ins_split
= SPLIT_RIGHT
;
4970 insert
.ins_tree_depth
= depth
;
4972 ret
= ocfs2_do_insert_extent(inode
, handle
, et
, &split_rec
, &insert
);
4981 static int ocfs2_truncate_rec(struct inode
*inode
, handle_t
*handle
,
4982 struct ocfs2_path
*path
, int index
,
4983 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
4985 struct ocfs2_extent_tree
*et
)
4988 u32 left_cpos
, rec_range
, trunc_range
;
4989 int wants_rotate
= 0, is_rightmost_tree_rec
= 0;
4990 struct super_block
*sb
= inode
->i_sb
;
4991 struct ocfs2_path
*left_path
= NULL
;
4992 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
4993 struct ocfs2_extent_rec
*rec
;
4994 struct ocfs2_extent_block
*eb
;
4996 if (ocfs2_is_empty_extent(&el
->l_recs
[0]) && index
> 0) {
4997 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
, dealloc
, et
);
5006 if (index
== (le16_to_cpu(el
->l_next_free_rec
) - 1) &&
5007 path
->p_tree_depth
) {
5009 * Check whether this is the rightmost tree record. If
5010 * we remove all of this record or part of its right
5011 * edge then an update of the record lengths above it
5014 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
5015 if (eb
->h_next_leaf_blk
== 0)
5016 is_rightmost_tree_rec
= 1;
5019 rec
= &el
->l_recs
[index
];
5020 if (index
== 0 && path
->p_tree_depth
&&
5021 le32_to_cpu(rec
->e_cpos
) == cpos
) {
5023 * Changing the leftmost offset (via partial or whole
5024 * record truncate) of an interior (or rightmost) path
5025 * means we have to update the subtree that is formed
5026 * by this leaf and the one to it's left.
5028 * There are two cases we can skip:
5029 * 1) Path is the leftmost one in our inode tree.
5030 * 2) The leaf is rightmost and will be empty after
5031 * we remove the extent record - the rotate code
5032 * knows how to update the newly formed edge.
5035 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
,
5042 if (left_cpos
&& le16_to_cpu(el
->l_next_free_rec
) > 1) {
5043 left_path
= ocfs2_new_path(path_root_bh(path
),
5044 path_root_el(path
));
5051 ret
= ocfs2_find_path(inode
, left_path
, left_cpos
);
5059 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
5060 handle
->h_buffer_credits
,
5067 ret
= ocfs2_journal_access_path(inode
, handle
, path
);
5073 ret
= ocfs2_journal_access_path(inode
, handle
, left_path
);
5079 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5080 trunc_range
= cpos
+ len
;
5082 if (le32_to_cpu(rec
->e_cpos
) == cpos
&& rec_range
== trunc_range
) {
5085 memset(rec
, 0, sizeof(*rec
));
5086 ocfs2_cleanup_merge(el
, index
);
5089 next_free
= le16_to_cpu(el
->l_next_free_rec
);
5090 if (is_rightmost_tree_rec
&& next_free
> 1) {
5092 * We skip the edge update if this path will
5093 * be deleted by the rotate code.
5095 rec
= &el
->l_recs
[next_free
- 1];
5096 ocfs2_adjust_rightmost_records(inode
, handle
, path
,
5099 } else if (le32_to_cpu(rec
->e_cpos
) == cpos
) {
5100 /* Remove leftmost portion of the record. */
5101 le32_add_cpu(&rec
->e_cpos
, len
);
5102 le64_add_cpu(&rec
->e_blkno
, ocfs2_clusters_to_blocks(sb
, len
));
5103 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5104 } else if (rec_range
== trunc_range
) {
5105 /* Remove rightmost portion of the record */
5106 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5107 if (is_rightmost_tree_rec
)
5108 ocfs2_adjust_rightmost_records(inode
, handle
, path
, rec
);
5110 /* Caller should have trapped this. */
5111 mlog(ML_ERROR
, "Inode %llu: Invalid record truncate: (%u, %u) "
5112 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
5113 le32_to_cpu(rec
->e_cpos
),
5114 le16_to_cpu(rec
->e_leaf_clusters
), cpos
, len
);
5121 subtree_index
= ocfs2_find_subtree_root(inode
, left_path
, path
);
5122 ocfs2_complete_edge_insert(inode
, handle
, left_path
, path
,
5126 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5128 ret
= ocfs2_rotate_tree_left(inode
, handle
, path
, dealloc
, et
);
5135 ocfs2_free_path(left_path
);
5139 int ocfs2_remove_extent(struct inode
*inode
,
5140 struct ocfs2_extent_tree
*et
,
5141 u32 cpos
, u32 len
, handle_t
*handle
,
5142 struct ocfs2_alloc_context
*meta_ac
,
5143 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5146 u32 rec_range
, trunc_range
;
5147 struct ocfs2_extent_rec
*rec
;
5148 struct ocfs2_extent_list
*el
;
5149 struct ocfs2_path
*path
= NULL
;
5151 ocfs2_extent_map_trunc(inode
, 0);
5153 path
= ocfs2_new_path(et
->et_root_bh
, et
->et_root_el
);
5160 ret
= ocfs2_find_path(inode
, path
, cpos
);
5166 el
= path_leaf_el(path
);
5167 index
= ocfs2_search_extent_list(el
, cpos
);
5168 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5169 ocfs2_error(inode
->i_sb
,
5170 "Inode %llu has an extent at cpos %u which can no "
5171 "longer be found.\n",
5172 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
);
5178 * We have 3 cases of extent removal:
5179 * 1) Range covers the entire extent rec
5180 * 2) Range begins or ends on one edge of the extent rec
5181 * 3) Range is in the middle of the extent rec (no shared edges)
5183 * For case 1 we remove the extent rec and left rotate to
5186 * For case 2 we just shrink the existing extent rec, with a
5187 * tree update if the shrinking edge is also the edge of an
5190 * For case 3 we do a right split to turn the extent rec into
5191 * something case 2 can handle.
5193 rec
= &el
->l_recs
[index
];
5194 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5195 trunc_range
= cpos
+ len
;
5197 BUG_ON(cpos
< le32_to_cpu(rec
->e_cpos
) || trunc_range
> rec_range
);
5199 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
5200 "(cpos %u, len %u)\n",
5201 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, cpos
, len
, index
,
5202 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
));
5204 if (le32_to_cpu(rec
->e_cpos
) == cpos
|| rec_range
== trunc_range
) {
5205 ret
= ocfs2_truncate_rec(inode
, handle
, path
, index
, dealloc
,
5212 ret
= ocfs2_split_tree(inode
, et
, handle
, path
, index
,
5213 trunc_range
, meta_ac
);
5220 * The split could have manipulated the tree enough to
5221 * move the record location, so we have to look for it again.
5223 ocfs2_reinit_path(path
, 1);
5225 ret
= ocfs2_find_path(inode
, path
, cpos
);
5231 el
= path_leaf_el(path
);
5232 index
= ocfs2_search_extent_list(el
, cpos
);
5233 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5234 ocfs2_error(inode
->i_sb
,
5235 "Inode %llu: split at cpos %u lost record.",
5236 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
5243 * Double check our values here. If anything is fishy,
5244 * it's easier to catch it at the top level.
5246 rec
= &el
->l_recs
[index
];
5247 rec_range
= le32_to_cpu(rec
->e_cpos
) +
5248 ocfs2_rec_clusters(el
, rec
);
5249 if (rec_range
!= trunc_range
) {
5250 ocfs2_error(inode
->i_sb
,
5251 "Inode %llu: error after split at cpos %u"
5252 "trunc len %u, existing record is (%u,%u)",
5253 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
5254 cpos
, len
, le32_to_cpu(rec
->e_cpos
),
5255 ocfs2_rec_clusters(el
, rec
));
5260 ret
= ocfs2_truncate_rec(inode
, handle
, path
, index
, dealloc
,
5269 ocfs2_free_path(path
);
5273 int ocfs2_truncate_log_needs_flush(struct ocfs2_super
*osb
)
5275 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5276 struct ocfs2_dinode
*di
;
5277 struct ocfs2_truncate_log
*tl
;
5279 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5280 tl
= &di
->id2
.i_dealloc
;
5282 mlog_bug_on_msg(le16_to_cpu(tl
->tl_used
) > le16_to_cpu(tl
->tl_count
),
5283 "slot %d, invalid truncate log parameters: used = "
5284 "%u, count = %u\n", osb
->slot_num
,
5285 le16_to_cpu(tl
->tl_used
), le16_to_cpu(tl
->tl_count
));
5286 return le16_to_cpu(tl
->tl_used
) == le16_to_cpu(tl
->tl_count
);
5289 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log
*tl
,
5290 unsigned int new_start
)
5292 unsigned int tail_index
;
5293 unsigned int current_tail
;
5295 /* No records, nothing to coalesce */
5296 if (!le16_to_cpu(tl
->tl_used
))
5299 tail_index
= le16_to_cpu(tl
->tl_used
) - 1;
5300 current_tail
= le32_to_cpu(tl
->tl_recs
[tail_index
].t_start
);
5301 current_tail
+= le32_to_cpu(tl
->tl_recs
[tail_index
].t_clusters
);
5303 return current_tail
== new_start
;
5306 int ocfs2_truncate_log_append(struct ocfs2_super
*osb
,
5309 unsigned int num_clusters
)
5312 unsigned int start_cluster
, tl_count
;
5313 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5314 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5315 struct ocfs2_dinode
*di
;
5316 struct ocfs2_truncate_log
*tl
;
5318 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5319 (unsigned long long)start_blk
, num_clusters
);
5321 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
5323 start_cluster
= ocfs2_blocks_to_clusters(osb
->sb
, start_blk
);
5325 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5326 tl
= &di
->id2
.i_dealloc
;
5327 if (!OCFS2_IS_VALID_DINODE(di
)) {
5328 OCFS2_RO_ON_INVALID_DINODE(osb
->sb
, di
);
5333 tl_count
= le16_to_cpu(tl
->tl_count
);
5334 mlog_bug_on_msg(tl_count
> ocfs2_truncate_recs_per_inode(osb
->sb
) ||
5336 "Truncate record count on #%llu invalid "
5337 "wanted %u, actual %u\n",
5338 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5339 ocfs2_truncate_recs_per_inode(osb
->sb
),
5340 le16_to_cpu(tl
->tl_count
));
5342 /* Caller should have known to flush before calling us. */
5343 index
= le16_to_cpu(tl
->tl_used
);
5344 if (index
>= tl_count
) {
5350 status
= ocfs2_journal_access(handle
, tl_inode
, tl_bh
,
5351 OCFS2_JOURNAL_ACCESS_WRITE
);
5357 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5358 "%llu (index = %d)\n", num_clusters
, start_cluster
,
5359 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
, index
);
5361 if (ocfs2_truncate_log_can_coalesce(tl
, start_cluster
)) {
5363 * Move index back to the record we are coalescing with.
5364 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5368 num_clusters
+= le32_to_cpu(tl
->tl_recs
[index
].t_clusters
);
5369 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5370 index
, le32_to_cpu(tl
->tl_recs
[index
].t_start
),
5373 tl
->tl_recs
[index
].t_start
= cpu_to_le32(start_cluster
);
5374 tl
->tl_used
= cpu_to_le16(index
+ 1);
5376 tl
->tl_recs
[index
].t_clusters
= cpu_to_le32(num_clusters
);
5378 status
= ocfs2_journal_dirty(handle
, tl_bh
);
5389 static int ocfs2_replay_truncate_records(struct ocfs2_super
*osb
,
5391 struct inode
*data_alloc_inode
,
5392 struct buffer_head
*data_alloc_bh
)
5396 unsigned int num_clusters
;
5398 struct ocfs2_truncate_rec rec
;
5399 struct ocfs2_dinode
*di
;
5400 struct ocfs2_truncate_log
*tl
;
5401 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5402 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5406 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5407 tl
= &di
->id2
.i_dealloc
;
5408 i
= le16_to_cpu(tl
->tl_used
) - 1;
5410 /* Caller has given us at least enough credits to
5411 * update the truncate log dinode */
5412 status
= ocfs2_journal_access(handle
, tl_inode
, tl_bh
,
5413 OCFS2_JOURNAL_ACCESS_WRITE
);
5419 tl
->tl_used
= cpu_to_le16(i
);
5421 status
= ocfs2_journal_dirty(handle
, tl_bh
);
5427 /* TODO: Perhaps we can calculate the bulk of the
5428 * credits up front rather than extending like
5430 status
= ocfs2_extend_trans(handle
,
5431 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC
);
5437 rec
= tl
->tl_recs
[i
];
5438 start_blk
= ocfs2_clusters_to_blocks(data_alloc_inode
->i_sb
,
5439 le32_to_cpu(rec
.t_start
));
5440 num_clusters
= le32_to_cpu(rec
.t_clusters
);
5442 /* if start_blk is not set, we ignore the record as
5445 mlog(0, "free record %d, start = %u, clusters = %u\n",
5446 i
, le32_to_cpu(rec
.t_start
), num_clusters
);
5448 status
= ocfs2_free_clusters(handle
, data_alloc_inode
,
5449 data_alloc_bh
, start_blk
,
5464 /* Expects you to already be holding tl_inode->i_mutex */
5465 int __ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5468 unsigned int num_to_flush
;
5470 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5471 struct inode
*data_alloc_inode
= NULL
;
5472 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5473 struct buffer_head
*data_alloc_bh
= NULL
;
5474 struct ocfs2_dinode
*di
;
5475 struct ocfs2_truncate_log
*tl
;
5479 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
5481 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5482 tl
= &di
->id2
.i_dealloc
;
5483 if (!OCFS2_IS_VALID_DINODE(di
)) {
5484 OCFS2_RO_ON_INVALID_DINODE(osb
->sb
, di
);
5489 num_to_flush
= le16_to_cpu(tl
->tl_used
);
5490 mlog(0, "Flush %u records from truncate log #%llu\n",
5491 num_to_flush
, (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
);
5492 if (!num_to_flush
) {
5497 data_alloc_inode
= ocfs2_get_system_file_inode(osb
,
5498 GLOBAL_BITMAP_SYSTEM_INODE
,
5499 OCFS2_INVALID_SLOT
);
5500 if (!data_alloc_inode
) {
5502 mlog(ML_ERROR
, "Could not get bitmap inode!\n");
5506 mutex_lock(&data_alloc_inode
->i_mutex
);
5508 status
= ocfs2_inode_lock(data_alloc_inode
, &data_alloc_bh
, 1);
5514 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
5515 if (IS_ERR(handle
)) {
5516 status
= PTR_ERR(handle
);
5521 status
= ocfs2_replay_truncate_records(osb
, handle
, data_alloc_inode
,
5526 ocfs2_commit_trans(osb
, handle
);
5529 brelse(data_alloc_bh
);
5530 ocfs2_inode_unlock(data_alloc_inode
, 1);
5533 mutex_unlock(&data_alloc_inode
->i_mutex
);
5534 iput(data_alloc_inode
);
5541 int ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5544 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5546 mutex_lock(&tl_inode
->i_mutex
);
5547 status
= __ocfs2_flush_truncate_log(osb
);
5548 mutex_unlock(&tl_inode
->i_mutex
);
5553 static void ocfs2_truncate_log_worker(struct work_struct
*work
)
5556 struct ocfs2_super
*osb
=
5557 container_of(work
, struct ocfs2_super
,
5558 osb_truncate_log_wq
.work
);
5562 status
= ocfs2_flush_truncate_log(osb
);
5566 ocfs2_init_inode_steal_slot(osb
);
5571 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5572 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super
*osb
,
5575 if (osb
->osb_tl_inode
) {
5576 /* We want to push off log flushes while truncates are
5579 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
5581 queue_delayed_work(ocfs2_wq
, &osb
->osb_truncate_log_wq
,
5582 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL
);
5586 static int ocfs2_get_truncate_log_info(struct ocfs2_super
*osb
,
5588 struct inode
**tl_inode
,
5589 struct buffer_head
**tl_bh
)
5592 struct inode
*inode
= NULL
;
5593 struct buffer_head
*bh
= NULL
;
5595 inode
= ocfs2_get_system_file_inode(osb
,
5596 TRUNCATE_LOG_SYSTEM_INODE
,
5600 mlog(ML_ERROR
, "Could not get load truncate log inode!\n");
5604 status
= ocfs2_read_block(osb
, OCFS2_I(inode
)->ip_blkno
, &bh
,
5605 OCFS2_BH_CACHED
, inode
);
5619 /* called during the 1st stage of node recovery. we stamp a clean
5620 * truncate log and pass back a copy for processing later. if the
5621 * truncate log does not require processing, a *tl_copy is set to
5623 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super
*osb
,
5625 struct ocfs2_dinode
**tl_copy
)
5628 struct inode
*tl_inode
= NULL
;
5629 struct buffer_head
*tl_bh
= NULL
;
5630 struct ocfs2_dinode
*di
;
5631 struct ocfs2_truncate_log
*tl
;
5635 mlog(0, "recover truncate log from slot %d\n", slot_num
);
5637 status
= ocfs2_get_truncate_log_info(osb
, slot_num
, &tl_inode
, &tl_bh
);
5643 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5644 tl
= &di
->id2
.i_dealloc
;
5645 if (!OCFS2_IS_VALID_DINODE(di
)) {
5646 OCFS2_RO_ON_INVALID_DINODE(tl_inode
->i_sb
, di
);
5651 if (le16_to_cpu(tl
->tl_used
)) {
5652 mlog(0, "We'll have %u logs to recover\n",
5653 le16_to_cpu(tl
->tl_used
));
5655 *tl_copy
= kmalloc(tl_bh
->b_size
, GFP_KERNEL
);
5662 /* Assuming the write-out below goes well, this copy
5663 * will be passed back to recovery for processing. */
5664 memcpy(*tl_copy
, tl_bh
->b_data
, tl_bh
->b_size
);
5666 /* All we need to do to clear the truncate log is set
5670 status
= ocfs2_write_block(osb
, tl_bh
, tl_inode
);
5683 if (status
< 0 && (*tl_copy
)) {
5692 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super
*osb
,
5693 struct ocfs2_dinode
*tl_copy
)
5697 unsigned int clusters
, num_recs
, start_cluster
;
5700 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5701 struct ocfs2_truncate_log
*tl
;
5705 if (OCFS2_I(tl_inode
)->ip_blkno
== le64_to_cpu(tl_copy
->i_blkno
)) {
5706 mlog(ML_ERROR
, "Asked to recover my own truncate log!\n");
5710 tl
= &tl_copy
->id2
.i_dealloc
;
5711 num_recs
= le16_to_cpu(tl
->tl_used
);
5712 mlog(0, "cleanup %u records from %llu\n", num_recs
,
5713 (unsigned long long)le64_to_cpu(tl_copy
->i_blkno
));
5715 mutex_lock(&tl_inode
->i_mutex
);
5716 for(i
= 0; i
< num_recs
; i
++) {
5717 if (ocfs2_truncate_log_needs_flush(osb
)) {
5718 status
= __ocfs2_flush_truncate_log(osb
);
5725 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
5726 if (IS_ERR(handle
)) {
5727 status
= PTR_ERR(handle
);
5732 clusters
= le32_to_cpu(tl
->tl_recs
[i
].t_clusters
);
5733 start_cluster
= le32_to_cpu(tl
->tl_recs
[i
].t_start
);
5734 start_blk
= ocfs2_clusters_to_blocks(osb
->sb
, start_cluster
);
5736 status
= ocfs2_truncate_log_append(osb
, handle
,
5737 start_blk
, clusters
);
5738 ocfs2_commit_trans(osb
, handle
);
5746 mutex_unlock(&tl_inode
->i_mutex
);
5752 void ocfs2_truncate_log_shutdown(struct ocfs2_super
*osb
)
5755 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5760 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
5761 flush_workqueue(ocfs2_wq
);
5763 status
= ocfs2_flush_truncate_log(osb
);
5767 brelse(osb
->osb_tl_bh
);
5768 iput(osb
->osb_tl_inode
);
5774 int ocfs2_truncate_log_init(struct ocfs2_super
*osb
)
5777 struct inode
*tl_inode
= NULL
;
5778 struct buffer_head
*tl_bh
= NULL
;
5782 status
= ocfs2_get_truncate_log_info(osb
,
5789 /* ocfs2_truncate_log_shutdown keys on the existence of
5790 * osb->osb_tl_inode so we don't set any of the osb variables
5791 * until we're sure all is well. */
5792 INIT_DELAYED_WORK(&osb
->osb_truncate_log_wq
,
5793 ocfs2_truncate_log_worker
);
5794 osb
->osb_tl_bh
= tl_bh
;
5795 osb
->osb_tl_inode
= tl_inode
;
5802 * Delayed de-allocation of suballocator blocks.
5804 * Some sets of block de-allocations might involve multiple suballocator inodes.
5806 * The locking for this can get extremely complicated, especially when
5807 * the suballocator inodes to delete from aren't known until deep
5808 * within an unrelated codepath.
5810 * ocfs2_extent_block structures are a good example of this - an inode
5811 * btree could have been grown by any number of nodes each allocating
5812 * out of their own suballoc inode.
5814 * These structures allow the delay of block de-allocation until a
5815 * later time, when locking of multiple cluster inodes won't cause
5820 * Describes a single block free from a suballocator
5822 struct ocfs2_cached_block_free
{
5823 struct ocfs2_cached_block_free
*free_next
;
5825 unsigned int free_bit
;
5828 struct ocfs2_per_slot_free_list
{
5829 struct ocfs2_per_slot_free_list
*f_next_suballocator
;
5832 struct ocfs2_cached_block_free
*f_first
;
5835 static int ocfs2_free_cached_items(struct ocfs2_super
*osb
,
5838 struct ocfs2_cached_block_free
*head
)
5843 struct inode
*inode
;
5844 struct buffer_head
*di_bh
= NULL
;
5845 struct ocfs2_cached_block_free
*tmp
;
5847 inode
= ocfs2_get_system_file_inode(osb
, sysfile_type
, slot
);
5854 mutex_lock(&inode
->i_mutex
);
5856 ret
= ocfs2_inode_lock(inode
, &di_bh
, 1);
5862 handle
= ocfs2_start_trans(osb
, OCFS2_SUBALLOC_FREE
);
5863 if (IS_ERR(handle
)) {
5864 ret
= PTR_ERR(handle
);
5870 bg_blkno
= ocfs2_which_suballoc_group(head
->free_blk
,
5872 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
5873 head
->free_bit
, (unsigned long long)head
->free_blk
);
5875 ret
= ocfs2_free_suballoc_bits(handle
, inode
, di_bh
,
5876 head
->free_bit
, bg_blkno
, 1);
5882 ret
= ocfs2_extend_trans(handle
, OCFS2_SUBALLOC_FREE
);
5889 head
= head
->free_next
;
5894 ocfs2_commit_trans(osb
, handle
);
5897 ocfs2_inode_unlock(inode
, 1);
5900 mutex_unlock(&inode
->i_mutex
);
5904 /* Premature exit may have left some dangling items. */
5906 head
= head
->free_next
;
5913 int ocfs2_run_deallocs(struct ocfs2_super
*osb
,
5914 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
5917 struct ocfs2_per_slot_free_list
*fl
;
5922 while (ctxt
->c_first_suballocator
) {
5923 fl
= ctxt
->c_first_suballocator
;
5926 mlog(0, "Free items: (type %u, slot %d)\n",
5927 fl
->f_inode_type
, fl
->f_slot
);
5928 ret2
= ocfs2_free_cached_items(osb
, fl
->f_inode_type
,
5929 fl
->f_slot
, fl
->f_first
);
5936 ctxt
->c_first_suballocator
= fl
->f_next_suballocator
;
5943 static struct ocfs2_per_slot_free_list
*
5944 ocfs2_find_per_slot_free_list(int type
,
5946 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
5948 struct ocfs2_per_slot_free_list
*fl
= ctxt
->c_first_suballocator
;
5951 if (fl
->f_inode_type
== type
&& fl
->f_slot
== slot
)
5954 fl
= fl
->f_next_suballocator
;
5957 fl
= kmalloc(sizeof(*fl
), GFP_NOFS
);
5959 fl
->f_inode_type
= type
;
5962 fl
->f_next_suballocator
= ctxt
->c_first_suballocator
;
5964 ctxt
->c_first_suballocator
= fl
;
5969 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
5970 int type
, int slot
, u64 blkno
,
5974 struct ocfs2_per_slot_free_list
*fl
;
5975 struct ocfs2_cached_block_free
*item
;
5977 fl
= ocfs2_find_per_slot_free_list(type
, slot
, ctxt
);
5984 item
= kmalloc(sizeof(*item
), GFP_NOFS
);
5991 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
5992 type
, slot
, bit
, (unsigned long long)blkno
);
5994 item
->free_blk
= blkno
;
5995 item
->free_bit
= bit
;
5996 item
->free_next
= fl
->f_first
;
6005 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6006 struct ocfs2_extent_block
*eb
)
6008 return ocfs2_cache_block_dealloc(ctxt
, EXTENT_ALLOC_SYSTEM_INODE
,
6009 le16_to_cpu(eb
->h_suballoc_slot
),
6010 le64_to_cpu(eb
->h_blkno
),
6011 le16_to_cpu(eb
->h_suballoc_bit
));
6014 /* This function will figure out whether the currently last extent
6015 * block will be deleted, and if it will, what the new last extent
6016 * block will be so we can update his h_next_leaf_blk field, as well
6017 * as the dinodes i_last_eb_blk */
6018 static int ocfs2_find_new_last_ext_blk(struct inode
*inode
,
6019 unsigned int clusters_to_del
,
6020 struct ocfs2_path
*path
,
6021 struct buffer_head
**new_last_eb
)
6023 int next_free
, ret
= 0;
6025 struct ocfs2_extent_rec
*rec
;
6026 struct ocfs2_extent_block
*eb
;
6027 struct ocfs2_extent_list
*el
;
6028 struct buffer_head
*bh
= NULL
;
6030 *new_last_eb
= NULL
;
6032 /* we have no tree, so of course, no last_eb. */
6033 if (!path
->p_tree_depth
)
6036 /* trunc to zero special case - this makes tree_depth = 0
6037 * regardless of what it is. */
6038 if (OCFS2_I(inode
)->ip_clusters
== clusters_to_del
)
6041 el
= path_leaf_el(path
);
6042 BUG_ON(!el
->l_next_free_rec
);
6045 * Make sure that this extent list will actually be empty
6046 * after we clear away the data. We can shortcut out if
6047 * there's more than one non-empty extent in the
6048 * list. Otherwise, a check of the remaining extent is
6051 next_free
= le16_to_cpu(el
->l_next_free_rec
);
6053 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
6057 /* We may have a valid extent in index 1, check it. */
6059 rec
= &el
->l_recs
[1];
6062 * Fall through - no more nonempty extents, so we want
6063 * to delete this leaf.
6069 rec
= &el
->l_recs
[0];
6074 * Check it we'll only be trimming off the end of this
6077 if (le16_to_cpu(rec
->e_leaf_clusters
) > clusters_to_del
)
6081 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
6087 ret
= ocfs2_find_leaf(inode
, path_root_el(path
), cpos
, &bh
);
6093 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
6095 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
6096 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
6102 get_bh(*new_last_eb
);
6103 mlog(0, "returning block %llu, (cpos: %u)\n",
6104 (unsigned long long)le64_to_cpu(eb
->h_blkno
), cpos
);
6112 * Trim some clusters off the rightmost edge of a tree. Only called
6115 * The caller needs to:
6116 * - start journaling of each path component.
6117 * - compute and fully set up any new last ext block
6119 static int ocfs2_trim_tree(struct inode
*inode
, struct ocfs2_path
*path
,
6120 handle_t
*handle
, struct ocfs2_truncate_context
*tc
,
6121 u32 clusters_to_del
, u64
*delete_start
)
6123 int ret
, i
, index
= path
->p_tree_depth
;
6126 struct buffer_head
*bh
;
6127 struct ocfs2_extent_list
*el
;
6128 struct ocfs2_extent_rec
*rec
;
6132 while (index
>= 0) {
6133 bh
= path
->p_node
[index
].bh
;
6134 el
= path
->p_node
[index
].el
;
6136 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6137 index
, (unsigned long long)bh
->b_blocknr
);
6139 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
6142 (path
->p_tree_depth
- le16_to_cpu(el
->l_tree_depth
))) {
6143 ocfs2_error(inode
->i_sb
,
6144 "Inode %lu has invalid ext. block %llu",
6146 (unsigned long long)bh
->b_blocknr
);
6152 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
6153 rec
= &el
->l_recs
[i
];
6155 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6156 "next = %u\n", i
, le32_to_cpu(rec
->e_cpos
),
6157 ocfs2_rec_clusters(el
, rec
),
6158 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
6159 le16_to_cpu(el
->l_next_free_rec
));
6161 BUG_ON(ocfs2_rec_clusters(el
, rec
) < clusters_to_del
);
6163 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
6165 * If the leaf block contains a single empty
6166 * extent and no records, we can just remove
6169 if (i
== 0 && ocfs2_is_empty_extent(rec
)) {
6171 sizeof(struct ocfs2_extent_rec
));
6172 el
->l_next_free_rec
= cpu_to_le16(0);
6178 * Remove any empty extents by shifting things
6179 * left. That should make life much easier on
6180 * the code below. This condition is rare
6181 * enough that we shouldn't see a performance
6184 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
6185 le16_add_cpu(&el
->l_next_free_rec
, -1);
6188 i
< le16_to_cpu(el
->l_next_free_rec
); i
++)
6189 el
->l_recs
[i
] = el
->l_recs
[i
+ 1];
6191 memset(&el
->l_recs
[i
], 0,
6192 sizeof(struct ocfs2_extent_rec
));
6195 * We've modified our extent list. The
6196 * simplest way to handle this change
6197 * is to being the search from the
6200 goto find_tail_record
;
6203 le16_add_cpu(&rec
->e_leaf_clusters
, -clusters_to_del
);
6206 * We'll use "new_edge" on our way back up the
6207 * tree to know what our rightmost cpos is.
6209 new_edge
= le16_to_cpu(rec
->e_leaf_clusters
);
6210 new_edge
+= le32_to_cpu(rec
->e_cpos
);
6213 * The caller will use this to delete data blocks.
6215 *delete_start
= le64_to_cpu(rec
->e_blkno
)
6216 + ocfs2_clusters_to_blocks(inode
->i_sb
,
6217 le16_to_cpu(rec
->e_leaf_clusters
));
6220 * If it's now empty, remove this record.
6222 if (le16_to_cpu(rec
->e_leaf_clusters
) == 0) {
6224 sizeof(struct ocfs2_extent_rec
));
6225 le16_add_cpu(&el
->l_next_free_rec
, -1);
6228 if (le64_to_cpu(rec
->e_blkno
) == deleted_eb
) {
6230 sizeof(struct ocfs2_extent_rec
));
6231 le16_add_cpu(&el
->l_next_free_rec
, -1);
6236 /* Can this actually happen? */
6237 if (le16_to_cpu(el
->l_next_free_rec
) == 0)
6241 * We never actually deleted any clusters
6242 * because our leaf was empty. There's no
6243 * reason to adjust the rightmost edge then.
6248 rec
->e_int_clusters
= cpu_to_le32(new_edge
);
6249 le32_add_cpu(&rec
->e_int_clusters
,
6250 -le32_to_cpu(rec
->e_cpos
));
6253 * A deleted child record should have been
6256 BUG_ON(le32_to_cpu(rec
->e_int_clusters
) == 0);
6260 ret
= ocfs2_journal_dirty(handle
, bh
);
6266 mlog(0, "extent list container %llu, after: record %d: "
6267 "(%u, %u, %llu), next = %u.\n",
6268 (unsigned long long)bh
->b_blocknr
, i
,
6269 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
),
6270 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
6271 le16_to_cpu(el
->l_next_free_rec
));
6274 * We must be careful to only attempt delete of an
6275 * extent block (and not the root inode block).
6277 if (index
> 0 && le16_to_cpu(el
->l_next_free_rec
) == 0) {
6278 struct ocfs2_extent_block
*eb
=
6279 (struct ocfs2_extent_block
*)bh
->b_data
;
6282 * Save this for use when processing the
6285 deleted_eb
= le64_to_cpu(eb
->h_blkno
);
6287 mlog(0, "deleting this extent block.\n");
6289 ocfs2_remove_from_cache(inode
, bh
);
6291 BUG_ON(ocfs2_rec_clusters(el
, &el
->l_recs
[0]));
6292 BUG_ON(le32_to_cpu(el
->l_recs
[0].e_cpos
));
6293 BUG_ON(le64_to_cpu(el
->l_recs
[0].e_blkno
));
6295 ret
= ocfs2_cache_extent_block_free(&tc
->tc_dealloc
, eb
);
6296 /* An error here is not fatal. */
6311 static int ocfs2_do_truncate(struct ocfs2_super
*osb
,
6312 unsigned int clusters_to_del
,
6313 struct inode
*inode
,
6314 struct buffer_head
*fe_bh
,
6316 struct ocfs2_truncate_context
*tc
,
6317 struct ocfs2_path
*path
)
6320 struct ocfs2_dinode
*fe
;
6321 struct ocfs2_extent_block
*last_eb
= NULL
;
6322 struct ocfs2_extent_list
*el
;
6323 struct buffer_head
*last_eb_bh
= NULL
;
6326 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
6328 status
= ocfs2_find_new_last_ext_blk(inode
, clusters_to_del
,
6336 * Each component will be touched, so we might as well journal
6337 * here to avoid having to handle errors later.
6339 status
= ocfs2_journal_access_path(inode
, handle
, path
);
6346 status
= ocfs2_journal_access(handle
, inode
, last_eb_bh
,
6347 OCFS2_JOURNAL_ACCESS_WRITE
);
6353 last_eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
6356 el
= &(fe
->id2
.i_list
);
6359 * Lower levels depend on this never happening, but it's best
6360 * to check it up here before changing the tree.
6362 if (el
->l_tree_depth
&& el
->l_recs
[0].e_int_clusters
== 0) {
6363 ocfs2_error(inode
->i_sb
,
6364 "Inode %lu has an empty extent record, depth %u\n",
6365 inode
->i_ino
, le16_to_cpu(el
->l_tree_depth
));
6370 spin_lock(&OCFS2_I(inode
)->ip_lock
);
6371 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
) -
6373 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
6374 le32_add_cpu(&fe
->i_clusters
, -clusters_to_del
);
6375 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
6377 status
= ocfs2_trim_tree(inode
, path
, handle
, tc
,
6378 clusters_to_del
, &delete_blk
);
6384 if (le32_to_cpu(fe
->i_clusters
) == 0) {
6385 /* trunc to zero is a special case. */
6386 el
->l_tree_depth
= 0;
6387 fe
->i_last_eb_blk
= 0;
6389 fe
->i_last_eb_blk
= last_eb
->h_blkno
;
6391 status
= ocfs2_journal_dirty(handle
, fe_bh
);
6398 /* If there will be a new last extent block, then by
6399 * definition, there cannot be any leaves to the right of
6401 last_eb
->h_next_leaf_blk
= 0;
6402 status
= ocfs2_journal_dirty(handle
, last_eb_bh
);
6410 status
= ocfs2_truncate_log_append(osb
, handle
, delete_blk
,
6424 static int ocfs2_writeback_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
6426 set_buffer_uptodate(bh
);
6427 mark_buffer_dirty(bh
);
6431 static int ocfs2_ordered_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
6433 set_buffer_uptodate(bh
);
6434 mark_buffer_dirty(bh
);
6435 return ocfs2_journal_dirty_data(handle
, bh
);
6438 static void ocfs2_map_and_dirty_page(struct inode
*inode
, handle_t
*handle
,
6439 unsigned int from
, unsigned int to
,
6440 struct page
*page
, int zero
, u64
*phys
)
6442 int ret
, partial
= 0;
6444 ret
= ocfs2_map_page_blocks(page
, phys
, inode
, from
, to
, 0);
6449 zero_user_segment(page
, from
, to
);
6452 * Need to set the buffers we zero'd into uptodate
6453 * here if they aren't - ocfs2_map_page_blocks()
6454 * might've skipped some
6456 if (ocfs2_should_order_data(inode
)) {
6457 ret
= walk_page_buffers(handle
,
6460 ocfs2_ordered_zero_func
);
6464 ret
= walk_page_buffers(handle
, page_buffers(page
),
6466 ocfs2_writeback_zero_func
);
6472 SetPageUptodate(page
);
6474 flush_dcache_page(page
);
6477 static void ocfs2_zero_cluster_pages(struct inode
*inode
, loff_t start
,
6478 loff_t end
, struct page
**pages
,
6479 int numpages
, u64 phys
, handle_t
*handle
)
6483 unsigned int from
, to
= PAGE_CACHE_SIZE
;
6484 struct super_block
*sb
= inode
->i_sb
;
6486 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb
)));
6491 to
= PAGE_CACHE_SIZE
;
6492 for(i
= 0; i
< numpages
; i
++) {
6495 from
= start
& (PAGE_CACHE_SIZE
- 1);
6496 if ((end
>> PAGE_CACHE_SHIFT
) == page
->index
)
6497 to
= end
& (PAGE_CACHE_SIZE
- 1);
6499 BUG_ON(from
> PAGE_CACHE_SIZE
);
6500 BUG_ON(to
> PAGE_CACHE_SIZE
);
6502 ocfs2_map_and_dirty_page(inode
, handle
, from
, to
, page
, 1,
6505 start
= (page
->index
+ 1) << PAGE_CACHE_SHIFT
;
6509 ocfs2_unlock_and_free_pages(pages
, numpages
);
6512 static int ocfs2_grab_eof_pages(struct inode
*inode
, loff_t start
, loff_t end
,
6513 struct page
**pages
, int *num
)
6515 int numpages
, ret
= 0;
6516 struct super_block
*sb
= inode
->i_sb
;
6517 struct address_space
*mapping
= inode
->i_mapping
;
6518 unsigned long index
;
6519 loff_t last_page_bytes
;
6521 BUG_ON(start
> end
);
6523 BUG_ON(start
>> OCFS2_SB(sb
)->s_clustersize_bits
!=
6524 (end
- 1) >> OCFS2_SB(sb
)->s_clustersize_bits
);
6527 last_page_bytes
= PAGE_ALIGN(end
);
6528 index
= start
>> PAGE_CACHE_SHIFT
;
6530 pages
[numpages
] = grab_cache_page(mapping
, index
);
6531 if (!pages
[numpages
]) {
6539 } while (index
< (last_page_bytes
>> PAGE_CACHE_SHIFT
));
6544 ocfs2_unlock_and_free_pages(pages
, numpages
);
6554 * Zero the area past i_size but still within an allocated
6555 * cluster. This avoids exposing nonzero data on subsequent file
6558 * We need to call this before i_size is updated on the inode because
6559 * otherwise block_write_full_page() will skip writeout of pages past
6560 * i_size. The new_i_size parameter is passed for this reason.
6562 int ocfs2_zero_range_for_truncate(struct inode
*inode
, handle_t
*handle
,
6563 u64 range_start
, u64 range_end
)
6565 int ret
= 0, numpages
;
6566 struct page
**pages
= NULL
;
6568 unsigned int ext_flags
;
6569 struct super_block
*sb
= inode
->i_sb
;
6572 * File systems which don't support sparse files zero on every
6575 if (!ocfs2_sparse_alloc(OCFS2_SB(sb
)))
6578 pages
= kcalloc(ocfs2_pages_per_cluster(sb
),
6579 sizeof(struct page
*), GFP_NOFS
);
6580 if (pages
== NULL
) {
6586 if (range_start
== range_end
)
6589 ret
= ocfs2_extent_map_get_blocks(inode
,
6590 range_start
>> sb
->s_blocksize_bits
,
6591 &phys
, NULL
, &ext_flags
);
6598 * Tail is a hole, or is marked unwritten. In either case, we
6599 * can count on read and write to return/push zero's.
6601 if (phys
== 0 || ext_flags
& OCFS2_EXT_UNWRITTEN
)
6604 ret
= ocfs2_grab_eof_pages(inode
, range_start
, range_end
, pages
,
6611 ocfs2_zero_cluster_pages(inode
, range_start
, range_end
, pages
,
6612 numpages
, phys
, handle
);
6615 * Initiate writeout of the pages we zero'd here. We don't
6616 * wait on them - the truncate_inode_pages() call later will
6619 ret
= do_sync_mapping_range(inode
->i_mapping
, range_start
,
6620 range_end
- 1, SYNC_FILE_RANGE_WRITE
);
6631 static void ocfs2_zero_dinode_id2_with_xattr(struct inode
*inode
,
6632 struct ocfs2_dinode
*di
)
6634 unsigned int blocksize
= 1 << inode
->i_sb
->s_blocksize_bits
;
6635 unsigned int xattrsize
= le16_to_cpu(di
->i_xattr_inline_size
);
6637 if (le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_XATTR_FL
)
6638 memset(&di
->id2
, 0, blocksize
-
6639 offsetof(struct ocfs2_dinode
, id2
) -
6642 memset(&di
->id2
, 0, blocksize
-
6643 offsetof(struct ocfs2_dinode
, id2
));
6646 void ocfs2_dinode_new_extent_list(struct inode
*inode
,
6647 struct ocfs2_dinode
*di
)
6649 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
6650 di
->id2
.i_list
.l_tree_depth
= 0;
6651 di
->id2
.i_list
.l_next_free_rec
= 0;
6652 di
->id2
.i_list
.l_count
= cpu_to_le16(
6653 ocfs2_extent_recs_per_inode_with_xattr(inode
->i_sb
, di
));
6656 void ocfs2_set_inode_data_inline(struct inode
*inode
, struct ocfs2_dinode
*di
)
6658 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
6659 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
6661 spin_lock(&oi
->ip_lock
);
6662 oi
->ip_dyn_features
|= OCFS2_INLINE_DATA_FL
;
6663 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
6664 spin_unlock(&oi
->ip_lock
);
6667 * We clear the entire i_data structure here so that all
6668 * fields can be properly initialized.
6670 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
6672 idata
->id_count
= cpu_to_le16(
6673 ocfs2_max_inline_data_with_xattr(inode
->i_sb
, di
));
6676 int ocfs2_convert_inline_data_to_extents(struct inode
*inode
,
6677 struct buffer_head
*di_bh
)
6679 int ret
, i
, has_data
, num_pages
= 0;
6681 u64
uninitialized_var(block
);
6682 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
6683 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
6684 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
6685 struct ocfs2_alloc_context
*data_ac
= NULL
;
6686 struct page
**pages
= NULL
;
6687 loff_t end
= osb
->s_clustersize
;
6688 struct ocfs2_extent_tree et
;
6690 has_data
= i_size_read(inode
) ? 1 : 0;
6693 pages
= kcalloc(ocfs2_pages_per_cluster(osb
->sb
),
6694 sizeof(struct page
*), GFP_NOFS
);
6695 if (pages
== NULL
) {
6701 ret
= ocfs2_reserve_clusters(osb
, 1, &data_ac
);
6708 handle
= ocfs2_start_trans(osb
, OCFS2_INLINE_TO_EXTENTS_CREDITS
);
6709 if (IS_ERR(handle
)) {
6710 ret
= PTR_ERR(handle
);
6715 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
6716 OCFS2_JOURNAL_ACCESS_WRITE
);
6724 unsigned int page_end
;
6727 ret
= ocfs2_claim_clusters(osb
, handle
, data_ac
, 1, &bit_off
,
6735 * Save two copies, one for insert, and one that can
6736 * be changed by ocfs2_map_and_dirty_page() below.
6738 block
= phys
= ocfs2_clusters_to_blocks(inode
->i_sb
, bit_off
);
6741 * Non sparse file systems zero on extend, so no need
6744 if (!ocfs2_sparse_alloc(osb
) &&
6745 PAGE_CACHE_SIZE
< osb
->s_clustersize
)
6746 end
= PAGE_CACHE_SIZE
;
6748 ret
= ocfs2_grab_eof_pages(inode
, 0, end
, pages
, &num_pages
);
6755 * This should populate the 1st page for us and mark
6758 ret
= ocfs2_read_inline_data(inode
, pages
[0], di_bh
);
6764 page_end
= PAGE_CACHE_SIZE
;
6765 if (PAGE_CACHE_SIZE
> osb
->s_clustersize
)
6766 page_end
= osb
->s_clustersize
;
6768 for (i
= 0; i
< num_pages
; i
++)
6769 ocfs2_map_and_dirty_page(inode
, handle
, 0, page_end
,
6770 pages
[i
], i
> 0, &phys
);
6773 spin_lock(&oi
->ip_lock
);
6774 oi
->ip_dyn_features
&= ~OCFS2_INLINE_DATA_FL
;
6775 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
6776 spin_unlock(&oi
->ip_lock
);
6778 ocfs2_dinode_new_extent_list(inode
, di
);
6780 ocfs2_journal_dirty(handle
, di_bh
);
6784 * An error at this point should be extremely rare. If
6785 * this proves to be false, we could always re-build
6786 * the in-inode data from our pages.
6788 ocfs2_init_dinode_extent_tree(&et
, inode
, di_bh
);
6789 ret
= ocfs2_insert_extent(osb
, handle
, inode
, &et
,
6790 0, block
, 1, 0, NULL
);
6796 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
6800 ocfs2_commit_trans(osb
, handle
);
6804 ocfs2_free_alloc_context(data_ac
);
6808 ocfs2_unlock_and_free_pages(pages
, num_pages
);
6816 * It is expected, that by the time you call this function,
6817 * inode->i_size and fe->i_size have been adjusted.
6819 * WARNING: This will kfree the truncate context
6821 int ocfs2_commit_truncate(struct ocfs2_super
*osb
,
6822 struct inode
*inode
,
6823 struct buffer_head
*fe_bh
,
6824 struct ocfs2_truncate_context
*tc
)
6826 int status
, i
, credits
, tl_sem
= 0;
6827 u32 clusters_to_del
, new_highest_cpos
, range
;
6828 struct ocfs2_extent_list
*el
;
6829 handle_t
*handle
= NULL
;
6830 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6831 struct ocfs2_path
*path
= NULL
;
6832 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)fe_bh
->b_data
;
6836 new_highest_cpos
= ocfs2_clusters_for_bytes(osb
->sb
,
6837 i_size_read(inode
));
6839 path
= ocfs2_new_path(fe_bh
, &di
->id2
.i_list
);
6846 ocfs2_extent_map_trunc(inode
, new_highest_cpos
);
6850 * Check that we still have allocation to delete.
6852 if (OCFS2_I(inode
)->ip_clusters
== 0) {
6858 * Truncate always works against the rightmost tree branch.
6860 status
= ocfs2_find_path(inode
, path
, UINT_MAX
);
6866 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
6867 OCFS2_I(inode
)->ip_clusters
, path
->p_tree_depth
);
6870 * By now, el will point to the extent list on the bottom most
6871 * portion of this tree. Only the tail record is considered in
6874 * We handle the following cases, in order:
6875 * - empty extent: delete the remaining branch
6876 * - remove the entire record
6877 * - remove a partial record
6878 * - no record needs to be removed (truncate has completed)
6880 el
= path_leaf_el(path
);
6881 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
6882 ocfs2_error(inode
->i_sb
,
6883 "Inode %llu has empty extent block at %llu\n",
6884 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
6885 (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
6890 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
6891 range
= le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
6892 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
6893 if (i
== 0 && ocfs2_is_empty_extent(&el
->l_recs
[i
])) {
6894 clusters_to_del
= 0;
6895 } else if (le32_to_cpu(el
->l_recs
[i
].e_cpos
) >= new_highest_cpos
) {
6896 clusters_to_del
= ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
6897 } else if (range
> new_highest_cpos
) {
6898 clusters_to_del
= (ocfs2_rec_clusters(el
, &el
->l_recs
[i
]) +
6899 le32_to_cpu(el
->l_recs
[i
].e_cpos
)) -
6906 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
6907 clusters_to_del
, (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
6909 mutex_lock(&tl_inode
->i_mutex
);
6911 /* ocfs2_truncate_log_needs_flush guarantees us at least one
6912 * record is free for use. If there isn't any, we flush to get
6913 * an empty truncate log. */
6914 if (ocfs2_truncate_log_needs_flush(osb
)) {
6915 status
= __ocfs2_flush_truncate_log(osb
);
6922 credits
= ocfs2_calc_tree_trunc_credits(osb
->sb
, clusters_to_del
,
6923 (struct ocfs2_dinode
*)fe_bh
->b_data
,
6925 handle
= ocfs2_start_trans(osb
, credits
);
6926 if (IS_ERR(handle
)) {
6927 status
= PTR_ERR(handle
);
6933 status
= ocfs2_do_truncate(osb
, clusters_to_del
, inode
, fe_bh
, handle
,
6940 mutex_unlock(&tl_inode
->i_mutex
);
6943 ocfs2_commit_trans(osb
, handle
);
6946 ocfs2_reinit_path(path
, 1);
6949 * The check above will catch the case where we've truncated
6950 * away all allocation.
6956 ocfs2_schedule_truncate_log_flush(osb
, 1);
6959 mutex_unlock(&tl_inode
->i_mutex
);
6962 ocfs2_commit_trans(osb
, handle
);
6964 ocfs2_run_deallocs(osb
, &tc
->tc_dealloc
);
6966 ocfs2_free_path(path
);
6968 /* This will drop the ext_alloc cluster lock for us */
6969 ocfs2_free_truncate_context(tc
);
6976 * Expects the inode to already be locked.
6978 int ocfs2_prepare_truncate(struct ocfs2_super
*osb
,
6979 struct inode
*inode
,
6980 struct buffer_head
*fe_bh
,
6981 struct ocfs2_truncate_context
**tc
)
6984 unsigned int new_i_clusters
;
6985 struct ocfs2_dinode
*fe
;
6986 struct ocfs2_extent_block
*eb
;
6987 struct buffer_head
*last_eb_bh
= NULL
;
6993 new_i_clusters
= ocfs2_clusters_for_bytes(osb
->sb
,
6994 i_size_read(inode
));
6995 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
6997 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
6998 "%llu\n", le32_to_cpu(fe
->i_clusters
), new_i_clusters
,
6999 (unsigned long long)le64_to_cpu(fe
->i_size
));
7001 *tc
= kzalloc(sizeof(struct ocfs2_truncate_context
), GFP_KERNEL
);
7007 ocfs2_init_dealloc_ctxt(&(*tc
)->tc_dealloc
);
7009 if (fe
->id2
.i_list
.l_tree_depth
) {
7010 status
= ocfs2_read_block(osb
, le64_to_cpu(fe
->i_last_eb_blk
),
7011 &last_eb_bh
, OCFS2_BH_CACHED
, inode
);
7016 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
7017 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
7018 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode
->i_sb
, eb
);
7026 (*tc
)->tc_last_eb_bh
= last_eb_bh
;
7032 ocfs2_free_truncate_context(*tc
);
7040 * 'start' is inclusive, 'end' is not.
7042 int ocfs2_truncate_inline(struct inode
*inode
, struct buffer_head
*di_bh
,
7043 unsigned int start
, unsigned int end
, int trunc
)
7046 unsigned int numbytes
;
7048 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7049 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7050 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7052 if (end
> i_size_read(inode
))
7053 end
= i_size_read(inode
);
7055 BUG_ON(start
>= end
);
7057 if (!(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
) ||
7058 !(le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_DATA_FL
) ||
7059 !ocfs2_supports_inline_data(osb
)) {
7060 ocfs2_error(inode
->i_sb
,
7061 "Inline data flags for inode %llu don't agree! "
7062 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7063 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7064 le16_to_cpu(di
->i_dyn_features
),
7065 OCFS2_I(inode
)->ip_dyn_features
,
7066 osb
->s_feature_incompat
);
7071 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
7072 if (IS_ERR(handle
)) {
7073 ret
= PTR_ERR(handle
);
7078 ret
= ocfs2_journal_access(handle
, inode
, di_bh
,
7079 OCFS2_JOURNAL_ACCESS_WRITE
);
7085 numbytes
= end
- start
;
7086 memset(idata
->id_data
+ start
, 0, numbytes
);
7089 * No need to worry about the data page here - it's been
7090 * truncated already and inline data doesn't need it for
7091 * pushing zero's to disk, so we'll let readpage pick it up
7095 i_size_write(inode
, start
);
7096 di
->i_size
= cpu_to_le64(start
);
7099 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7100 inode
->i_ctime
= inode
->i_mtime
= CURRENT_TIME
;
7102 di
->i_ctime
= di
->i_mtime
= cpu_to_le64(inode
->i_ctime
.tv_sec
);
7103 di
->i_ctime_nsec
= di
->i_mtime_nsec
= cpu_to_le32(inode
->i_ctime
.tv_nsec
);
7105 ocfs2_journal_dirty(handle
, di_bh
);
7108 ocfs2_commit_trans(osb
, handle
);
7114 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
)
7117 * The caller is responsible for completing deallocation
7118 * before freeing the context.
7120 if (tc
->tc_dealloc
.c_first_suballocator
!= NULL
)
7122 "Truncate completion has non-empty dealloc context\n");
7124 if (tc
->tc_last_eb_bh
)
7125 brelse(tc
->tc_last_eb_bh
);