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_get_<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_get_extent_tree(struct ocfs2_extent_tree *et,
306 struct buffer_head *bh,
308 struct ocfs2_extent_tree_operations *ops)
314 obj = (void *)bh->b_data;
317 et->et_ops->eo_fill_root_el(et);
318 if (!et->et_ops->eo_fill_max_leaf_clusters)
319 et->et_max_leaf_clusters = 0;
321 et->et_ops->eo_fill_max_leaf_clusters(inode, et);
324 void ocfs2_get_dinode_extent_tree(struct ocfs2_extent_tree *et,
326 struct buffer_head *bh)
328 __ocfs2_get_extent_tree(et, inode, bh, NULL, &ocfs2_dinode_et_ops);
331 void ocfs2_get_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
333 struct buffer_head *bh)
335 __ocfs2_get_extent_tree(et, inode, bh, NULL,
336 &ocfs2_xattr_tree_et_ops);
339 void ocfs2_get_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
341 struct buffer_head *bh,
342 struct ocfs2_xattr_value_root *xv)
344 __ocfs2_get_extent_tree(et, inode, bh, xv,
345 &ocfs2_xattr_value_et_ops);
348 void ocfs2_put_extent_tree(struct ocfs2_extent_tree *et)
350 brelse(et->et_root_bh);
353 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
356 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
359 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
361 return et->et_ops->eo_get_last_eb_blk(et);
364 static inline void ocfs2_et_update_clusters(struct inode *inode,
365 struct ocfs2_extent_tree *et,
368 et->et_ops->eo_update_clusters(inode, et, clusters);
371 static inline int ocfs2_et_insert_check(struct inode *inode,
372 struct ocfs2_extent_tree *et,
373 struct ocfs2_extent_rec *rec)
377 if (et->et_ops->eo_insert_check)
378 ret = et->et_ops->eo_insert_check(inode, et, rec);
382 static inline int ocfs2_et_sanity_check(struct inode *inode,
383 struct ocfs2_extent_tree *et)
387 if (et->et_ops->eo_sanity_check)
388 ret = et->et_ops->eo_sanity_check(inode, et);
392 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
393 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
394 struct ocfs2_extent_block *eb);
397 * Structures which describe a path through a btree, and functions to
400 * The idea here is to be as generic as possible with the tree
403 struct ocfs2_path_item {
404 struct buffer_head *bh;
405 struct ocfs2_extent_list *el;
408 #define OCFS2_MAX_PATH_DEPTH 5
412 struct ocfs2_path_item p_node[OCFS2_MAX_PATH_DEPTH];
415 #define path_root_bh(_path) ((_path)->p_node[0].bh)
416 #define path_root_el(_path) ((_path)->p_node[0].el)
417 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
418 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
419 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
422 * Reset the actual path elements so that we can re-use the structure
423 * to build another path. Generally, this involves freeing the buffer
426 static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
428 int i, start = 0, depth = 0;
429 struct ocfs2_path_item *node;
434 for(i = start; i < path_num_items(path); i++) {
435 node = &path->p_node[i];
443 * Tree depth may change during truncate, or insert. If we're
444 * keeping the root extent list, then make sure that our path
445 * structure reflects the proper depth.
448 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
450 path->p_tree_depth = depth;
453 static void ocfs2_free_path(struct ocfs2_path *path)
456 ocfs2_reinit_path(path, 0);
462 * All the elements of src into dest. After this call, src could be freed
463 * without affecting dest.
465 * Both paths should have the same root. Any non-root elements of dest
468 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
472 BUG_ON(path_root_bh(dest) != path_root_bh(src));
473 BUG_ON(path_root_el(dest) != path_root_el(src));
475 ocfs2_reinit_path(dest, 1);
477 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
478 dest->p_node[i].bh = src->p_node[i].bh;
479 dest->p_node[i].el = src->p_node[i].el;
481 if (dest->p_node[i].bh)
482 get_bh(dest->p_node[i].bh);
487 * Make the *dest path the same as src and re-initialize src path to
490 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
494 BUG_ON(path_root_bh(dest) != path_root_bh(src));
496 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
497 brelse(dest->p_node[i].bh);
499 dest->p_node[i].bh = src->p_node[i].bh;
500 dest->p_node[i].el = src->p_node[i].el;
502 src->p_node[i].bh = NULL;
503 src->p_node[i].el = NULL;
508 * Insert an extent block at given index.
510 * This will not take an additional reference on eb_bh.
512 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
513 struct buffer_head *eb_bh)
515 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
518 * Right now, no root bh is an extent block, so this helps
519 * catch code errors with dinode trees. The assertion can be
520 * safely removed if we ever need to insert extent block
521 * structures at the root.
525 path->p_node[index].bh = eb_bh;
526 path->p_node[index].el = &eb->h_list;
529 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
530 struct ocfs2_extent_list *root_el)
532 struct ocfs2_path *path;
534 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
536 path = kzalloc(sizeof(*path), GFP_NOFS);
538 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
540 path_root_bh(path) = root_bh;
541 path_root_el(path) = root_el;
548 * Convenience function to journal all components in a path.
550 static int ocfs2_journal_access_path(struct inode *inode, handle_t *handle,
551 struct ocfs2_path *path)
558 for(i = 0; i < path_num_items(path); i++) {
559 ret = ocfs2_journal_access(handle, inode, path->p_node[i].bh,
560 OCFS2_JOURNAL_ACCESS_WRITE);
572 * Return the index of the extent record which contains cluster #v_cluster.
573 * -1 is returned if it was not found.
575 * Should work fine on interior and exterior nodes.
577 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
581 struct ocfs2_extent_rec *rec;
582 u32 rec_end, rec_start, clusters;
584 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
585 rec = &el->l_recs[i];
587 rec_start = le32_to_cpu(rec->e_cpos);
588 clusters = ocfs2_rec_clusters(el, rec);
590 rec_end = rec_start + clusters;
592 if (v_cluster >= rec_start && v_cluster < rec_end) {
601 enum ocfs2_contig_type {
610 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
611 * ocfs2_extent_contig only work properly against leaf nodes!
613 static int ocfs2_block_extent_contig(struct super_block *sb,
614 struct ocfs2_extent_rec *ext,
617 u64 blk_end = le64_to_cpu(ext->e_blkno);
619 blk_end += ocfs2_clusters_to_blocks(sb,
620 le16_to_cpu(ext->e_leaf_clusters));
622 return blkno == blk_end;
625 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
626 struct ocfs2_extent_rec *right)
630 left_range = le32_to_cpu(left->e_cpos) +
631 le16_to_cpu(left->e_leaf_clusters);
633 return (left_range == le32_to_cpu(right->e_cpos));
636 static enum ocfs2_contig_type
637 ocfs2_extent_contig(struct inode *inode,
638 struct ocfs2_extent_rec *ext,
639 struct ocfs2_extent_rec *insert_rec)
641 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
644 * Refuse to coalesce extent records with different flag
645 * fields - we don't want to mix unwritten extents with user
648 if (ext->e_flags != insert_rec->e_flags)
651 if (ocfs2_extents_adjacent(ext, insert_rec) &&
652 ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
655 blkno = le64_to_cpu(ext->e_blkno);
656 if (ocfs2_extents_adjacent(insert_rec, ext) &&
657 ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno))
664 * NOTE: We can have pretty much any combination of contiguousness and
667 * The usefulness of APPEND_TAIL is more in that it lets us know that
668 * we'll have to update the path to that leaf.
670 enum ocfs2_append_type {
675 enum ocfs2_split_type {
681 struct ocfs2_insert_type {
682 enum ocfs2_split_type ins_split;
683 enum ocfs2_append_type ins_appending;
684 enum ocfs2_contig_type ins_contig;
685 int ins_contig_index;
689 struct ocfs2_merge_ctxt {
690 enum ocfs2_contig_type c_contig_type;
691 int c_has_empty_extent;
692 int c_split_covers_rec;
696 * How many free extents have we got before we need more meta data?
698 int ocfs2_num_free_extents(struct ocfs2_super *osb,
700 struct ocfs2_extent_tree *et)
703 struct ocfs2_extent_list *el = NULL;
704 struct ocfs2_extent_block *eb;
705 struct buffer_head *eb_bh = NULL;
711 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
714 retval = ocfs2_read_block(osb, last_eb_blk,
715 &eb_bh, OCFS2_BH_CACHED, inode);
720 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
724 BUG_ON(el->l_tree_depth != 0);
726 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
735 /* expects array to already be allocated
737 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
740 static int ocfs2_create_new_meta_bhs(struct ocfs2_super *osb,
744 struct ocfs2_alloc_context *meta_ac,
745 struct buffer_head *bhs[])
747 int count, status, i;
748 u16 suballoc_bit_start;
751 struct ocfs2_extent_block *eb;
756 while (count < wanted) {
757 status = ocfs2_claim_metadata(osb,
769 for(i = count; i < (num_got + count); i++) {
770 bhs[i] = sb_getblk(osb->sb, first_blkno);
771 if (bhs[i] == NULL) {
776 ocfs2_set_new_buffer_uptodate(inode, bhs[i]);
778 status = ocfs2_journal_access(handle, inode, bhs[i],
779 OCFS2_JOURNAL_ACCESS_CREATE);
785 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
786 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
787 /* Ok, setup the minimal stuff here. */
788 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
789 eb->h_blkno = cpu_to_le64(first_blkno);
790 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
791 eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
792 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
794 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
796 suballoc_bit_start++;
799 /* We'll also be dirtied by the caller, so
800 * this isn't absolutely necessary. */
801 status = ocfs2_journal_dirty(handle, bhs[i]);
814 for(i = 0; i < wanted; i++) {
825 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
827 * Returns the sum of the rightmost extent rec logical offset and
830 * ocfs2_add_branch() uses this to determine what logical cluster
831 * value should be populated into the leftmost new branch records.
833 * ocfs2_shift_tree_depth() uses this to determine the # clusters
834 * value for the new topmost tree record.
836 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
840 i = le16_to_cpu(el->l_next_free_rec) - 1;
842 return le32_to_cpu(el->l_recs[i].e_cpos) +
843 ocfs2_rec_clusters(el, &el->l_recs[i]);
847 * Add an entire tree branch to our inode. eb_bh is the extent block
848 * to start at, if we don't want to start the branch at the dinode
851 * last_eb_bh is required as we have to update it's next_leaf pointer
852 * for the new last extent block.
854 * the new branch will be 'empty' in the sense that every block will
855 * contain a single record with cluster count == 0.
857 static int ocfs2_add_branch(struct ocfs2_super *osb,
860 struct ocfs2_extent_tree *et,
861 struct buffer_head *eb_bh,
862 struct buffer_head **last_eb_bh,
863 struct ocfs2_alloc_context *meta_ac)
865 int status, new_blocks, i;
866 u64 next_blkno, new_last_eb_blk;
867 struct buffer_head *bh;
868 struct buffer_head **new_eb_bhs = NULL;
869 struct ocfs2_extent_block *eb;
870 struct ocfs2_extent_list *eb_el;
871 struct ocfs2_extent_list *el;
876 BUG_ON(!last_eb_bh || !*last_eb_bh);
879 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
884 /* we never add a branch to a leaf. */
885 BUG_ON(!el->l_tree_depth);
887 new_blocks = le16_to_cpu(el->l_tree_depth);
889 /* allocate the number of new eb blocks we need */
890 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
898 status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
899 meta_ac, new_eb_bhs);
905 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
906 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
908 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
909 * linked with the rest of the tree.
910 * conversly, new_eb_bhs[0] is the new bottommost leaf.
912 * when we leave the loop, new_last_eb_blk will point to the
913 * newest leaf, and next_blkno will point to the topmost extent
915 next_blkno = new_last_eb_blk = 0;
916 for(i = 0; i < new_blocks; i++) {
918 eb = (struct ocfs2_extent_block *) bh->b_data;
919 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
920 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
926 status = ocfs2_journal_access(handle, inode, bh,
927 OCFS2_JOURNAL_ACCESS_CREATE);
933 eb->h_next_leaf_blk = 0;
934 eb_el->l_tree_depth = cpu_to_le16(i);
935 eb_el->l_next_free_rec = cpu_to_le16(1);
937 * This actually counts as an empty extent as
940 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
941 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
943 * eb_el isn't always an interior node, but even leaf
944 * nodes want a zero'd flags and reserved field so
945 * this gets the whole 32 bits regardless of use.
947 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
948 if (!eb_el->l_tree_depth)
949 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
951 status = ocfs2_journal_dirty(handle, bh);
957 next_blkno = le64_to_cpu(eb->h_blkno);
960 /* This is a bit hairy. We want to update up to three blocks
961 * here without leaving any of them in an inconsistent state
962 * in case of error. We don't have to worry about
963 * journal_dirty erroring as it won't unless we've aborted the
964 * handle (in which case we would never be here) so reserving
965 * the write with journal_access is all we need to do. */
966 status = ocfs2_journal_access(handle, inode, *last_eb_bh,
967 OCFS2_JOURNAL_ACCESS_WRITE);
972 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
973 OCFS2_JOURNAL_ACCESS_WRITE);
979 status = ocfs2_journal_access(handle, inode, eb_bh,
980 OCFS2_JOURNAL_ACCESS_WRITE);
987 /* Link the new branch into the rest of the tree (el will
988 * either be on the root_bh, or the extent block passed in. */
989 i = le16_to_cpu(el->l_next_free_rec);
990 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
991 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
992 el->l_recs[i].e_int_clusters = 0;
993 le16_add_cpu(&el->l_next_free_rec, 1);
995 /* fe needs a new last extent block pointer, as does the
996 * next_leaf on the previously last-extent-block. */
997 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
999 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1000 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1002 status = ocfs2_journal_dirty(handle, *last_eb_bh);
1005 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1009 status = ocfs2_journal_dirty(handle, eb_bh);
1015 * Some callers want to track the rightmost leaf so pass it
1018 brelse(*last_eb_bh);
1019 get_bh(new_eb_bhs[0]);
1020 *last_eb_bh = new_eb_bhs[0];
1025 for (i = 0; i < new_blocks; i++)
1027 brelse(new_eb_bhs[i]);
1036 * adds another level to the allocation tree.
1037 * returns back the new extent block so you can add a branch to it
1040 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
1042 struct inode *inode,
1043 struct ocfs2_extent_tree *et,
1044 struct ocfs2_alloc_context *meta_ac,
1045 struct buffer_head **ret_new_eb_bh)
1049 struct buffer_head *new_eb_bh = NULL;
1050 struct ocfs2_extent_block *eb;
1051 struct ocfs2_extent_list *root_el;
1052 struct ocfs2_extent_list *eb_el;
1056 status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
1063 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1064 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1065 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1070 eb_el = &eb->h_list;
1071 root_el = et->et_root_el;
1073 status = ocfs2_journal_access(handle, inode, new_eb_bh,
1074 OCFS2_JOURNAL_ACCESS_CREATE);
1080 /* copy the root extent list data into the new extent block */
1081 eb_el->l_tree_depth = root_el->l_tree_depth;
1082 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1083 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1084 eb_el->l_recs[i] = root_el->l_recs[i];
1086 status = ocfs2_journal_dirty(handle, new_eb_bh);
1092 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
1093 OCFS2_JOURNAL_ACCESS_WRITE);
1099 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1101 /* update root_bh now */
1102 le16_add_cpu(&root_el->l_tree_depth, 1);
1103 root_el->l_recs[0].e_cpos = 0;
1104 root_el->l_recs[0].e_blkno = eb->h_blkno;
1105 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1106 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1107 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1108 root_el->l_next_free_rec = cpu_to_le16(1);
1110 /* If this is our 1st tree depth shift, then last_eb_blk
1111 * becomes the allocated extent block */
1112 if (root_el->l_tree_depth == cpu_to_le16(1))
1113 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1115 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1121 *ret_new_eb_bh = new_eb_bh;
1133 * Should only be called when there is no space left in any of the
1134 * leaf nodes. What we want to do is find the lowest tree depth
1135 * non-leaf extent block with room for new records. There are three
1136 * valid results of this search:
1138 * 1) a lowest extent block is found, then we pass it back in
1139 * *lowest_eb_bh and return '0'
1141 * 2) the search fails to find anything, but the root_el has room. We
1142 * pass NULL back in *lowest_eb_bh, but still return '0'
1144 * 3) the search fails to find anything AND the root_el is full, in
1145 * which case we return > 0
1147 * return status < 0 indicates an error.
1149 static int ocfs2_find_branch_target(struct ocfs2_super *osb,
1150 struct inode *inode,
1151 struct ocfs2_extent_tree *et,
1152 struct buffer_head **target_bh)
1156 struct ocfs2_extent_block *eb;
1157 struct ocfs2_extent_list *el;
1158 struct buffer_head *bh = NULL;
1159 struct buffer_head *lowest_bh = NULL;
1165 el = et->et_root_el;
1167 while(le16_to_cpu(el->l_tree_depth) > 1) {
1168 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1169 ocfs2_error(inode->i_sb, "Dinode %llu has empty "
1170 "extent list (next_free_rec == 0)",
1171 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1175 i = le16_to_cpu(el->l_next_free_rec) - 1;
1176 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1178 ocfs2_error(inode->i_sb, "Dinode %llu has extent "
1179 "list where extent # %d has no physical "
1181 (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
1191 status = ocfs2_read_block(osb, blkno, &bh, OCFS2_BH_CACHED,
1198 eb = (struct ocfs2_extent_block *) bh->b_data;
1199 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1200 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1206 if (le16_to_cpu(el->l_next_free_rec) <
1207 le16_to_cpu(el->l_count)) {
1215 /* If we didn't find one and the fe doesn't have any room,
1216 * then return '1' */
1217 el = et->et_root_el;
1218 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1221 *target_bh = lowest_bh;
1231 * Grow a b-tree so that it has more records.
1233 * We might shift the tree depth in which case existing paths should
1234 * be considered invalid.
1236 * Tree depth after the grow is returned via *final_depth.
1238 * *last_eb_bh will be updated by ocfs2_add_branch().
1240 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
1241 struct ocfs2_extent_tree *et, int *final_depth,
1242 struct buffer_head **last_eb_bh,
1243 struct ocfs2_alloc_context *meta_ac)
1246 struct ocfs2_extent_list *el = et->et_root_el;
1247 int depth = le16_to_cpu(el->l_tree_depth);
1248 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1249 struct buffer_head *bh = NULL;
1251 BUG_ON(meta_ac == NULL);
1253 shift = ocfs2_find_branch_target(osb, inode, et, &bh);
1260 /* We traveled all the way to the bottom of the allocation tree
1261 * and didn't find room for any more extents - we need to add
1262 * another tree level */
1265 mlog(0, "need to shift tree depth (current = %d)\n", depth);
1267 /* ocfs2_shift_tree_depth will return us a buffer with
1268 * the new extent block (so we can pass that to
1269 * ocfs2_add_branch). */
1270 ret = ocfs2_shift_tree_depth(osb, handle, inode, et,
1279 * Special case: we have room now if we shifted from
1280 * tree_depth 0, so no more work needs to be done.
1282 * We won't be calling add_branch, so pass
1283 * back *last_eb_bh as the new leaf. At depth
1284 * zero, it should always be null so there's
1285 * no reason to brelse.
1287 BUG_ON(*last_eb_bh);
1294 /* call ocfs2_add_branch to add the final part of the tree with
1296 mlog(0, "add branch. bh = %p\n", bh);
1297 ret = ocfs2_add_branch(osb, handle, inode, et, bh, last_eb_bh,
1306 *final_depth = depth;
1312 * This function will discard the rightmost extent record.
1314 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1316 int next_free = le16_to_cpu(el->l_next_free_rec);
1317 int count = le16_to_cpu(el->l_count);
1318 unsigned int num_bytes;
1321 /* This will cause us to go off the end of our extent list. */
1322 BUG_ON(next_free >= count);
1324 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1326 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1329 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1330 struct ocfs2_extent_rec *insert_rec)
1332 int i, insert_index, next_free, has_empty, num_bytes;
1333 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1334 struct ocfs2_extent_rec *rec;
1336 next_free = le16_to_cpu(el->l_next_free_rec);
1337 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1341 /* The tree code before us didn't allow enough room in the leaf. */
1342 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1345 * The easiest way to approach this is to just remove the
1346 * empty extent and temporarily decrement next_free.
1350 * If next_free was 1 (only an empty extent), this
1351 * loop won't execute, which is fine. We still want
1352 * the decrement above to happen.
1354 for(i = 0; i < (next_free - 1); i++)
1355 el->l_recs[i] = el->l_recs[i+1];
1361 * Figure out what the new record index should be.
1363 for(i = 0; i < next_free; i++) {
1364 rec = &el->l_recs[i];
1366 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1371 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1372 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
1374 BUG_ON(insert_index < 0);
1375 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1376 BUG_ON(insert_index > next_free);
1379 * No need to memmove if we're just adding to the tail.
1381 if (insert_index != next_free) {
1382 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1384 num_bytes = next_free - insert_index;
1385 num_bytes *= sizeof(struct ocfs2_extent_rec);
1386 memmove(&el->l_recs[insert_index + 1],
1387 &el->l_recs[insert_index],
1392 * Either we had an empty extent, and need to re-increment or
1393 * there was no empty extent on a non full rightmost leaf node,
1394 * in which case we still need to increment.
1397 el->l_next_free_rec = cpu_to_le16(next_free);
1399 * Make sure none of the math above just messed up our tree.
1401 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1403 el->l_recs[insert_index] = *insert_rec;
1407 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1409 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1411 BUG_ON(num_recs == 0);
1413 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1415 size = num_recs * sizeof(struct ocfs2_extent_rec);
1416 memmove(&el->l_recs[0], &el->l_recs[1], size);
1417 memset(&el->l_recs[num_recs], 0,
1418 sizeof(struct ocfs2_extent_rec));
1419 el->l_next_free_rec = cpu_to_le16(num_recs);
1424 * Create an empty extent record .
1426 * l_next_free_rec may be updated.
1428 * If an empty extent already exists do nothing.
1430 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1432 int next_free = le16_to_cpu(el->l_next_free_rec);
1434 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1439 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1442 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1443 "Asked to create an empty extent in a full list:\n"
1444 "count = %u, tree depth = %u",
1445 le16_to_cpu(el->l_count),
1446 le16_to_cpu(el->l_tree_depth));
1448 ocfs2_shift_records_right(el);
1451 le16_add_cpu(&el->l_next_free_rec, 1);
1452 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1456 * For a rotation which involves two leaf nodes, the "root node" is
1457 * the lowest level tree node which contains a path to both leafs. This
1458 * resulting set of information can be used to form a complete "subtree"
1460 * This function is passed two full paths from the dinode down to a
1461 * pair of adjacent leaves. It's task is to figure out which path
1462 * index contains the subtree root - this can be the root index itself
1463 * in a worst-case rotation.
1465 * The array index of the subtree root is passed back.
1467 static int ocfs2_find_subtree_root(struct inode *inode,
1468 struct ocfs2_path *left,
1469 struct ocfs2_path *right)
1474 * Check that the caller passed in two paths from the same tree.
1476 BUG_ON(path_root_bh(left) != path_root_bh(right));
1482 * The caller didn't pass two adjacent paths.
1484 mlog_bug_on_msg(i > left->p_tree_depth,
1485 "Inode %lu, left depth %u, right depth %u\n"
1486 "left leaf blk %llu, right leaf blk %llu\n",
1487 inode->i_ino, left->p_tree_depth,
1488 right->p_tree_depth,
1489 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1490 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1491 } while (left->p_node[i].bh->b_blocknr ==
1492 right->p_node[i].bh->b_blocknr);
1497 typedef void (path_insert_t)(void *, struct buffer_head *);
1500 * Traverse a btree path in search of cpos, starting at root_el.
1502 * This code can be called with a cpos larger than the tree, in which
1503 * case it will return the rightmost path.
1505 static int __ocfs2_find_path(struct inode *inode,
1506 struct ocfs2_extent_list *root_el, u32 cpos,
1507 path_insert_t *func, void *data)
1512 struct buffer_head *bh = NULL;
1513 struct ocfs2_extent_block *eb;
1514 struct ocfs2_extent_list *el;
1515 struct ocfs2_extent_rec *rec;
1516 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1519 while (el->l_tree_depth) {
1520 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1521 ocfs2_error(inode->i_sb,
1522 "Inode %llu has empty extent list at "
1524 (unsigned long long)oi->ip_blkno,
1525 le16_to_cpu(el->l_tree_depth));
1531 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1532 rec = &el->l_recs[i];
1535 * In the case that cpos is off the allocation
1536 * tree, this should just wind up returning the
1539 range = le32_to_cpu(rec->e_cpos) +
1540 ocfs2_rec_clusters(el, rec);
1541 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1545 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1547 ocfs2_error(inode->i_sb,
1548 "Inode %llu has bad blkno in extent list "
1549 "at depth %u (index %d)\n",
1550 (unsigned long long)oi->ip_blkno,
1551 le16_to_cpu(el->l_tree_depth), i);
1558 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno,
1559 &bh, OCFS2_BH_CACHED, inode);
1565 eb = (struct ocfs2_extent_block *) bh->b_data;
1567 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1568 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1573 if (le16_to_cpu(el->l_next_free_rec) >
1574 le16_to_cpu(el->l_count)) {
1575 ocfs2_error(inode->i_sb,
1576 "Inode %llu has bad count in extent list "
1577 "at block %llu (next free=%u, count=%u)\n",
1578 (unsigned long long)oi->ip_blkno,
1579 (unsigned long long)bh->b_blocknr,
1580 le16_to_cpu(el->l_next_free_rec),
1581 le16_to_cpu(el->l_count));
1592 * Catch any trailing bh that the loop didn't handle.
1600 * Given an initialized path (that is, it has a valid root extent
1601 * list), this function will traverse the btree in search of the path
1602 * which would contain cpos.
1604 * The path traveled is recorded in the path structure.
1606 * Note that this will not do any comparisons on leaf node extent
1607 * records, so it will work fine in the case that we just added a tree
1610 struct find_path_data {
1612 struct ocfs2_path *path;
1614 static void find_path_ins(void *data, struct buffer_head *bh)
1616 struct find_path_data *fp = data;
1619 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1622 static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
1625 struct find_path_data data;
1629 return __ocfs2_find_path(inode, path_root_el(path), cpos,
1630 find_path_ins, &data);
1633 static void find_leaf_ins(void *data, struct buffer_head *bh)
1635 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1636 struct ocfs2_extent_list *el = &eb->h_list;
1637 struct buffer_head **ret = data;
1639 /* We want to retain only the leaf block. */
1640 if (le16_to_cpu(el->l_tree_depth) == 0) {
1646 * Find the leaf block in the tree which would contain cpos. No
1647 * checking of the actual leaf is done.
1649 * Some paths want to call this instead of allocating a path structure
1650 * and calling ocfs2_find_path().
1652 * This function doesn't handle non btree extent lists.
1654 int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
1655 u32 cpos, struct buffer_head **leaf_bh)
1658 struct buffer_head *bh = NULL;
1660 ret = __ocfs2_find_path(inode, root_el, cpos, find_leaf_ins, &bh);
1672 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1674 * Basically, we've moved stuff around at the bottom of the tree and
1675 * we need to fix up the extent records above the changes to reflect
1678 * left_rec: the record on the left.
1679 * left_child_el: is the child list pointed to by left_rec
1680 * right_rec: the record to the right of left_rec
1681 * right_child_el: is the child list pointed to by right_rec
1683 * By definition, this only works on interior nodes.
1685 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1686 struct ocfs2_extent_list *left_child_el,
1687 struct ocfs2_extent_rec *right_rec,
1688 struct ocfs2_extent_list *right_child_el)
1690 u32 left_clusters, right_end;
1693 * Interior nodes never have holes. Their cpos is the cpos of
1694 * the leftmost record in their child list. Their cluster
1695 * count covers the full theoretical range of their child list
1696 * - the range between their cpos and the cpos of the record
1697 * immediately to their right.
1699 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1700 if (ocfs2_is_empty_extent(&right_child_el->l_recs[0])) {
1701 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1702 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1704 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1705 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1708 * Calculate the rightmost cluster count boundary before
1709 * moving cpos - we will need to adjust clusters after
1710 * updating e_cpos to keep the same highest cluster count.
1712 right_end = le32_to_cpu(right_rec->e_cpos);
1713 right_end += le32_to_cpu(right_rec->e_int_clusters);
1715 right_rec->e_cpos = left_rec->e_cpos;
1716 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1718 right_end -= le32_to_cpu(right_rec->e_cpos);
1719 right_rec->e_int_clusters = cpu_to_le32(right_end);
1723 * Adjust the adjacent root node records involved in a
1724 * rotation. left_el_blkno is passed in as a key so that we can easily
1725 * find it's index in the root list.
1727 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1728 struct ocfs2_extent_list *left_el,
1729 struct ocfs2_extent_list *right_el,
1734 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1735 le16_to_cpu(left_el->l_tree_depth));
1737 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1738 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1743 * The path walking code should have never returned a root and
1744 * two paths which are not adjacent.
1746 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1748 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1749 &root_el->l_recs[i + 1], right_el);
1753 * We've changed a leaf block (in right_path) and need to reflect that
1754 * change back up the subtree.
1756 * This happens in multiple places:
1757 * - When we've moved an extent record from the left path leaf to the right
1758 * path leaf to make room for an empty extent in the left path leaf.
1759 * - When our insert into the right path leaf is at the leftmost edge
1760 * and requires an update of the path immediately to it's left. This
1761 * can occur at the end of some types of rotation and appending inserts.
1762 * - When we've adjusted the last extent record in the left path leaf and the
1763 * 1st extent record in the right path leaf during cross extent block merge.
1765 static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1766 struct ocfs2_path *left_path,
1767 struct ocfs2_path *right_path,
1771 struct ocfs2_extent_list *el, *left_el, *right_el;
1772 struct ocfs2_extent_rec *left_rec, *right_rec;
1773 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1776 * Update the counts and position values within all the
1777 * interior nodes to reflect the leaf rotation we just did.
1779 * The root node is handled below the loop.
1781 * We begin the loop with right_el and left_el pointing to the
1782 * leaf lists and work our way up.
1784 * NOTE: within this loop, left_el and right_el always refer
1785 * to the *child* lists.
1787 left_el = path_leaf_el(left_path);
1788 right_el = path_leaf_el(right_path);
1789 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
1790 mlog(0, "Adjust records at index %u\n", i);
1793 * One nice property of knowing that all of these
1794 * nodes are below the root is that we only deal with
1795 * the leftmost right node record and the rightmost
1798 el = left_path->p_node[i].el;
1799 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
1800 left_rec = &el->l_recs[idx];
1802 el = right_path->p_node[i].el;
1803 right_rec = &el->l_recs[0];
1805 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
1808 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
1812 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
1817 * Setup our list pointers now so that the current
1818 * parents become children in the next iteration.
1820 left_el = left_path->p_node[i].el;
1821 right_el = right_path->p_node[i].el;
1825 * At the root node, adjust the two adjacent records which
1826 * begin our path to the leaves.
1829 el = left_path->p_node[subtree_index].el;
1830 left_el = left_path->p_node[subtree_index + 1].el;
1831 right_el = right_path->p_node[subtree_index + 1].el;
1833 ocfs2_adjust_root_records(el, left_el, right_el,
1834 left_path->p_node[subtree_index + 1].bh->b_blocknr);
1836 root_bh = left_path->p_node[subtree_index].bh;
1838 ret = ocfs2_journal_dirty(handle, root_bh);
1843 static int ocfs2_rotate_subtree_right(struct inode *inode,
1845 struct ocfs2_path *left_path,
1846 struct ocfs2_path *right_path,
1850 struct buffer_head *right_leaf_bh;
1851 struct buffer_head *left_leaf_bh = NULL;
1852 struct buffer_head *root_bh;
1853 struct ocfs2_extent_list *right_el, *left_el;
1854 struct ocfs2_extent_rec move_rec;
1856 left_leaf_bh = path_leaf_bh(left_path);
1857 left_el = path_leaf_el(left_path);
1859 if (left_el->l_next_free_rec != left_el->l_count) {
1860 ocfs2_error(inode->i_sb,
1861 "Inode %llu has non-full interior leaf node %llu"
1863 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1864 (unsigned long long)left_leaf_bh->b_blocknr,
1865 le16_to_cpu(left_el->l_next_free_rec));
1870 * This extent block may already have an empty record, so we
1871 * return early if so.
1873 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
1876 root_bh = left_path->p_node[subtree_index].bh;
1877 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
1879 ret = ocfs2_journal_access(handle, inode, root_bh,
1880 OCFS2_JOURNAL_ACCESS_WRITE);
1886 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
1887 ret = ocfs2_journal_access(handle, inode,
1888 right_path->p_node[i].bh,
1889 OCFS2_JOURNAL_ACCESS_WRITE);
1895 ret = ocfs2_journal_access(handle, inode,
1896 left_path->p_node[i].bh,
1897 OCFS2_JOURNAL_ACCESS_WRITE);
1904 right_leaf_bh = path_leaf_bh(right_path);
1905 right_el = path_leaf_el(right_path);
1907 /* This is a code error, not a disk corruption. */
1908 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
1909 "because rightmost leaf block %llu is empty\n",
1910 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1911 (unsigned long long)right_leaf_bh->b_blocknr);
1913 ocfs2_create_empty_extent(right_el);
1915 ret = ocfs2_journal_dirty(handle, right_leaf_bh);
1921 /* Do the copy now. */
1922 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
1923 move_rec = left_el->l_recs[i];
1924 right_el->l_recs[0] = move_rec;
1927 * Clear out the record we just copied and shift everything
1928 * over, leaving an empty extent in the left leaf.
1930 * We temporarily subtract from next_free_rec so that the
1931 * shift will lose the tail record (which is now defunct).
1933 le16_add_cpu(&left_el->l_next_free_rec, -1);
1934 ocfs2_shift_records_right(left_el);
1935 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1936 le16_add_cpu(&left_el->l_next_free_rec, 1);
1938 ret = ocfs2_journal_dirty(handle, left_leaf_bh);
1944 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
1952 * Given a full path, determine what cpos value would return us a path
1953 * containing the leaf immediately to the left of the current one.
1955 * Will return zero if the path passed in is already the leftmost path.
1957 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
1958 struct ocfs2_path *path, u32 *cpos)
1962 struct ocfs2_extent_list *el;
1964 BUG_ON(path->p_tree_depth == 0);
1968 blkno = path_leaf_bh(path)->b_blocknr;
1970 /* Start at the tree node just above the leaf and work our way up. */
1971 i = path->p_tree_depth - 1;
1973 el = path->p_node[i].el;
1976 * Find the extent record just before the one in our
1979 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
1980 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
1984 * We've determined that the
1985 * path specified is already
1986 * the leftmost one - return a
1992 * The leftmost record points to our
1993 * leaf - we need to travel up the
1999 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2000 *cpos = *cpos + ocfs2_rec_clusters(el,
2001 &el->l_recs[j - 1]);
2008 * If we got here, we never found a valid node where
2009 * the tree indicated one should be.
2012 "Invalid extent tree at extent block %llu\n",
2013 (unsigned long long)blkno);
2018 blkno = path->p_node[i].bh->b_blocknr;
2027 * Extend the transaction by enough credits to complete the rotation,
2028 * and still leave at least the original number of credits allocated
2029 * to this transaction.
2031 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2033 struct ocfs2_path *path)
2035 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2037 if (handle->h_buffer_credits < credits)
2038 return ocfs2_extend_trans(handle, credits);
2044 * Trap the case where we're inserting into the theoretical range past
2045 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2046 * whose cpos is less than ours into the right leaf.
2048 * It's only necessary to look at the rightmost record of the left
2049 * leaf because the logic that calls us should ensure that the
2050 * theoretical ranges in the path components above the leaves are
2053 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2056 struct ocfs2_extent_list *left_el;
2057 struct ocfs2_extent_rec *rec;
2060 left_el = path_leaf_el(left_path);
2061 next_free = le16_to_cpu(left_el->l_next_free_rec);
2062 rec = &left_el->l_recs[next_free - 1];
2064 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2069 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2071 int next_free = le16_to_cpu(el->l_next_free_rec);
2073 struct ocfs2_extent_rec *rec;
2078 rec = &el->l_recs[0];
2079 if (ocfs2_is_empty_extent(rec)) {
2083 rec = &el->l_recs[1];
2086 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2087 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2093 * Rotate all the records in a btree right one record, starting at insert_cpos.
2095 * The path to the rightmost leaf should be passed in.
2097 * The array is assumed to be large enough to hold an entire path (tree depth).
2099 * Upon succesful return from this function:
2101 * - The 'right_path' array will contain a path to the leaf block
2102 * whose range contains e_cpos.
2103 * - That leaf block will have a single empty extent in list index 0.
2104 * - In the case that the rotation requires a post-insert update,
2105 * *ret_left_path will contain a valid path which can be passed to
2106 * ocfs2_insert_path().
2108 static int ocfs2_rotate_tree_right(struct inode *inode,
2110 enum ocfs2_split_type split,
2112 struct ocfs2_path *right_path,
2113 struct ocfs2_path **ret_left_path)
2115 int ret, start, orig_credits = handle->h_buffer_credits;
2117 struct ocfs2_path *left_path = NULL;
2119 *ret_left_path = NULL;
2121 left_path = ocfs2_new_path(path_root_bh(right_path),
2122 path_root_el(right_path));
2129 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
2135 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
2138 * What we want to do here is:
2140 * 1) Start with the rightmost path.
2142 * 2) Determine a path to the leaf block directly to the left
2145 * 3) Determine the 'subtree root' - the lowest level tree node
2146 * which contains a path to both leaves.
2148 * 4) Rotate the subtree.
2150 * 5) Find the next subtree by considering the left path to be
2151 * the new right path.
2153 * The check at the top of this while loop also accepts
2154 * insert_cpos == cpos because cpos is only a _theoretical_
2155 * value to get us the left path - insert_cpos might very well
2156 * be filling that hole.
2158 * Stop at a cpos of '0' because we either started at the
2159 * leftmost branch (i.e., a tree with one branch and a
2160 * rotation inside of it), or we've gone as far as we can in
2161 * rotating subtrees.
2163 while (cpos && insert_cpos <= cpos) {
2164 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2167 ret = ocfs2_find_path(inode, left_path, cpos);
2173 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2174 path_leaf_bh(right_path),
2175 "Inode %lu: error during insert of %u "
2176 "(left path cpos %u) results in two identical "
2177 "paths ending at %llu\n",
2178 inode->i_ino, insert_cpos, cpos,
2179 (unsigned long long)
2180 path_leaf_bh(left_path)->b_blocknr);
2182 if (split == SPLIT_NONE &&
2183 ocfs2_rotate_requires_path_adjustment(left_path,
2187 * We've rotated the tree as much as we
2188 * should. The rest is up to
2189 * ocfs2_insert_path() to complete, after the
2190 * record insertion. We indicate this
2191 * situation by returning the left path.
2193 * The reason we don't adjust the records here
2194 * before the record insert is that an error
2195 * later might break the rule where a parent
2196 * record e_cpos will reflect the actual
2197 * e_cpos of the 1st nonempty record of the
2200 *ret_left_path = left_path;
2204 start = ocfs2_find_subtree_root(inode, left_path, right_path);
2206 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2208 (unsigned long long) right_path->p_node[start].bh->b_blocknr,
2209 right_path->p_tree_depth);
2211 ret = ocfs2_extend_rotate_transaction(handle, start,
2212 orig_credits, right_path);
2218 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
2225 if (split != SPLIT_NONE &&
2226 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2229 * A rotate moves the rightmost left leaf
2230 * record over to the leftmost right leaf
2231 * slot. If we're doing an extent split
2232 * instead of a real insert, then we have to
2233 * check that the extent to be split wasn't
2234 * just moved over. If it was, then we can
2235 * exit here, passing left_path back -
2236 * ocfs2_split_extent() is smart enough to
2237 * search both leaves.
2239 *ret_left_path = left_path;
2244 * There is no need to re-read the next right path
2245 * as we know that it'll be our current left
2246 * path. Optimize by copying values instead.
2248 ocfs2_mv_path(right_path, left_path);
2250 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
2259 ocfs2_free_path(left_path);
2265 static void ocfs2_update_edge_lengths(struct inode *inode, handle_t *handle,
2266 struct ocfs2_path *path)
2269 struct ocfs2_extent_rec *rec;
2270 struct ocfs2_extent_list *el;
2271 struct ocfs2_extent_block *eb;
2274 /* Path should always be rightmost. */
2275 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2276 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2279 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2280 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2281 rec = &el->l_recs[idx];
2282 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2284 for (i = 0; i < path->p_tree_depth; i++) {
2285 el = path->p_node[i].el;
2286 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2287 rec = &el->l_recs[idx];
2289 rec->e_int_clusters = cpu_to_le32(range);
2290 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2292 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2296 static void ocfs2_unlink_path(struct inode *inode, handle_t *handle,
2297 struct ocfs2_cached_dealloc_ctxt *dealloc,
2298 struct ocfs2_path *path, int unlink_start)
2301 struct ocfs2_extent_block *eb;
2302 struct ocfs2_extent_list *el;
2303 struct buffer_head *bh;
2305 for(i = unlink_start; i < path_num_items(path); i++) {
2306 bh = path->p_node[i].bh;
2308 eb = (struct ocfs2_extent_block *)bh->b_data;
2310 * Not all nodes might have had their final count
2311 * decremented by the caller - handle this here.
2314 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2316 "Inode %llu, attempted to remove extent block "
2317 "%llu with %u records\n",
2318 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2319 (unsigned long long)le64_to_cpu(eb->h_blkno),
2320 le16_to_cpu(el->l_next_free_rec));
2322 ocfs2_journal_dirty(handle, bh);
2323 ocfs2_remove_from_cache(inode, bh);
2327 el->l_next_free_rec = 0;
2328 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2330 ocfs2_journal_dirty(handle, bh);
2332 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2336 ocfs2_remove_from_cache(inode, bh);
2340 static void ocfs2_unlink_subtree(struct inode *inode, handle_t *handle,
2341 struct ocfs2_path *left_path,
2342 struct ocfs2_path *right_path,
2344 struct ocfs2_cached_dealloc_ctxt *dealloc)
2347 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2348 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2349 struct ocfs2_extent_list *el;
2350 struct ocfs2_extent_block *eb;
2352 el = path_leaf_el(left_path);
2354 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2356 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2357 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2360 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2362 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2363 le16_add_cpu(&root_el->l_next_free_rec, -1);
2365 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2366 eb->h_next_leaf_blk = 0;
2368 ocfs2_journal_dirty(handle, root_bh);
2369 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2371 ocfs2_unlink_path(inode, handle, dealloc, right_path,
2375 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2376 struct ocfs2_path *left_path,
2377 struct ocfs2_path *right_path,
2379 struct ocfs2_cached_dealloc_ctxt *dealloc,
2381 struct ocfs2_extent_tree *et)
2383 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2384 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2385 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2386 struct ocfs2_extent_block *eb;
2390 right_leaf_el = path_leaf_el(right_path);
2391 left_leaf_el = path_leaf_el(left_path);
2392 root_bh = left_path->p_node[subtree_index].bh;
2393 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2395 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2398 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2399 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2401 * It's legal for us to proceed if the right leaf is
2402 * the rightmost one and it has an empty extent. There
2403 * are two cases to handle - whether the leaf will be
2404 * empty after removal or not. If the leaf isn't empty
2405 * then just remove the empty extent up front. The
2406 * next block will handle empty leaves by flagging
2409 * Non rightmost leaves will throw -EAGAIN and the
2410 * caller can manually move the subtree and retry.
2413 if (eb->h_next_leaf_blk != 0ULL)
2416 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2417 ret = ocfs2_journal_access(handle, inode,
2418 path_leaf_bh(right_path),
2419 OCFS2_JOURNAL_ACCESS_WRITE);
2425 ocfs2_remove_empty_extent(right_leaf_el);
2427 right_has_empty = 1;
2430 if (eb->h_next_leaf_blk == 0ULL &&
2431 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2433 * We have to update i_last_eb_blk during the meta
2436 ret = ocfs2_journal_access(handle, inode, et_root_bh,
2437 OCFS2_JOURNAL_ACCESS_WRITE);
2443 del_right_subtree = 1;
2447 * Getting here with an empty extent in the right path implies
2448 * that it's the rightmost path and will be deleted.
2450 BUG_ON(right_has_empty && !del_right_subtree);
2452 ret = ocfs2_journal_access(handle, inode, root_bh,
2453 OCFS2_JOURNAL_ACCESS_WRITE);
2459 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2460 ret = ocfs2_journal_access(handle, inode,
2461 right_path->p_node[i].bh,
2462 OCFS2_JOURNAL_ACCESS_WRITE);
2468 ret = ocfs2_journal_access(handle, inode,
2469 left_path->p_node[i].bh,
2470 OCFS2_JOURNAL_ACCESS_WRITE);
2477 if (!right_has_empty) {
2479 * Only do this if we're moving a real
2480 * record. Otherwise, the action is delayed until
2481 * after removal of the right path in which case we
2482 * can do a simple shift to remove the empty extent.
2484 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2485 memset(&right_leaf_el->l_recs[0], 0,
2486 sizeof(struct ocfs2_extent_rec));
2488 if (eb->h_next_leaf_blk == 0ULL) {
2490 * Move recs over to get rid of empty extent, decrease
2491 * next_free. This is allowed to remove the last
2492 * extent in our leaf (setting l_next_free_rec to
2493 * zero) - the delete code below won't care.
2495 ocfs2_remove_empty_extent(right_leaf_el);
2498 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2501 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2505 if (del_right_subtree) {
2506 ocfs2_unlink_subtree(inode, handle, left_path, right_path,
2507 subtree_index, dealloc);
2508 ocfs2_update_edge_lengths(inode, handle, left_path);
2510 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2511 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2514 * Removal of the extent in the left leaf was skipped
2515 * above so we could delete the right path
2518 if (right_has_empty)
2519 ocfs2_remove_empty_extent(left_leaf_el);
2521 ret = ocfs2_journal_dirty(handle, et_root_bh);
2527 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2535 * Given a full path, determine what cpos value would return us a path
2536 * containing the leaf immediately to the right of the current one.
2538 * Will return zero if the path passed in is already the rightmost path.
2540 * This looks similar, but is subtly different to
2541 * ocfs2_find_cpos_for_left_leaf().
2543 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2544 struct ocfs2_path *path, u32 *cpos)
2548 struct ocfs2_extent_list *el;
2552 if (path->p_tree_depth == 0)
2555 blkno = path_leaf_bh(path)->b_blocknr;
2557 /* Start at the tree node just above the leaf and work our way up. */
2558 i = path->p_tree_depth - 1;
2562 el = path->p_node[i].el;
2565 * Find the extent record just after the one in our
2568 next_free = le16_to_cpu(el->l_next_free_rec);
2569 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2570 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2571 if (j == (next_free - 1)) {
2574 * We've determined that the
2575 * path specified is already
2576 * the rightmost one - return a
2582 * The rightmost record points to our
2583 * leaf - we need to travel up the
2589 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2595 * If we got here, we never found a valid node where
2596 * the tree indicated one should be.
2599 "Invalid extent tree at extent block %llu\n",
2600 (unsigned long long)blkno);
2605 blkno = path->p_node[i].bh->b_blocknr;
2613 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2615 struct buffer_head *bh,
2616 struct ocfs2_extent_list *el)
2620 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2623 ret = ocfs2_journal_access(handle, inode, bh,
2624 OCFS2_JOURNAL_ACCESS_WRITE);
2630 ocfs2_remove_empty_extent(el);
2632 ret = ocfs2_journal_dirty(handle, bh);
2640 static int __ocfs2_rotate_tree_left(struct inode *inode,
2641 handle_t *handle, int orig_credits,
2642 struct ocfs2_path *path,
2643 struct ocfs2_cached_dealloc_ctxt *dealloc,
2644 struct ocfs2_path **empty_extent_path,
2645 struct ocfs2_extent_tree *et)
2647 int ret, subtree_root, deleted;
2649 struct ocfs2_path *left_path = NULL;
2650 struct ocfs2_path *right_path = NULL;
2652 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2654 *empty_extent_path = NULL;
2656 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2663 left_path = ocfs2_new_path(path_root_bh(path),
2664 path_root_el(path));
2671 ocfs2_cp_path(left_path, path);
2673 right_path = ocfs2_new_path(path_root_bh(path),
2674 path_root_el(path));
2681 while (right_cpos) {
2682 ret = ocfs2_find_path(inode, right_path, right_cpos);
2688 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2691 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2693 (unsigned long long)
2694 right_path->p_node[subtree_root].bh->b_blocknr,
2695 right_path->p_tree_depth);
2697 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2698 orig_credits, left_path);
2705 * Caller might still want to make changes to the
2706 * tree root, so re-add it to the journal here.
2708 ret = ocfs2_journal_access(handle, inode,
2709 path_root_bh(left_path),
2710 OCFS2_JOURNAL_ACCESS_WRITE);
2716 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2717 right_path, subtree_root,
2718 dealloc, &deleted, et);
2719 if (ret == -EAGAIN) {
2721 * The rotation has to temporarily stop due to
2722 * the right subtree having an empty
2723 * extent. Pass it back to the caller for a
2726 *empty_extent_path = right_path;
2736 * The subtree rotate might have removed records on
2737 * the rightmost edge. If so, then rotation is
2743 ocfs2_mv_path(left_path, right_path);
2745 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2754 ocfs2_free_path(right_path);
2755 ocfs2_free_path(left_path);
2760 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
2761 struct ocfs2_path *path,
2762 struct ocfs2_cached_dealloc_ctxt *dealloc,
2763 struct ocfs2_extent_tree *et)
2765 int ret, subtree_index;
2767 struct ocfs2_path *left_path = NULL;
2768 struct ocfs2_extent_block *eb;
2769 struct ocfs2_extent_list *el;
2772 ret = ocfs2_et_sanity_check(inode, et);
2776 * There's two ways we handle this depending on
2777 * whether path is the only existing one.
2779 ret = ocfs2_extend_rotate_transaction(handle, 0,
2780 handle->h_buffer_credits,
2787 ret = ocfs2_journal_access_path(inode, handle, path);
2793 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
2801 * We have a path to the left of this one - it needs
2804 left_path = ocfs2_new_path(path_root_bh(path),
2805 path_root_el(path));
2812 ret = ocfs2_find_path(inode, left_path, cpos);
2818 ret = ocfs2_journal_access_path(inode, handle, left_path);
2824 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
2826 ocfs2_unlink_subtree(inode, handle, left_path, path,
2827 subtree_index, dealloc);
2828 ocfs2_update_edge_lengths(inode, handle, left_path);
2830 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2831 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2834 * 'path' is also the leftmost path which
2835 * means it must be the only one. This gets
2836 * handled differently because we want to
2837 * revert the inode back to having extents
2840 ocfs2_unlink_path(inode, handle, dealloc, path, 1);
2842 el = et->et_root_el;
2843 el->l_tree_depth = 0;
2844 el->l_next_free_rec = 0;
2845 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2847 ocfs2_et_set_last_eb_blk(et, 0);
2850 ocfs2_journal_dirty(handle, path_root_bh(path));
2853 ocfs2_free_path(left_path);
2858 * Left rotation of btree records.
2860 * In many ways, this is (unsurprisingly) the opposite of right
2861 * rotation. We start at some non-rightmost path containing an empty
2862 * extent in the leaf block. The code works its way to the rightmost
2863 * path by rotating records to the left in every subtree.
2865 * This is used by any code which reduces the number of extent records
2866 * in a leaf. After removal, an empty record should be placed in the
2867 * leftmost list position.
2869 * This won't handle a length update of the rightmost path records if
2870 * the rightmost tree leaf record is removed so the caller is
2871 * responsible for detecting and correcting that.
2873 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
2874 struct ocfs2_path *path,
2875 struct ocfs2_cached_dealloc_ctxt *dealloc,
2876 struct ocfs2_extent_tree *et)
2878 int ret, orig_credits = handle->h_buffer_credits;
2879 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
2880 struct ocfs2_extent_block *eb;
2881 struct ocfs2_extent_list *el;
2883 el = path_leaf_el(path);
2884 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2887 if (path->p_tree_depth == 0) {
2888 rightmost_no_delete:
2890 * Inline extents. This is trivially handled, so do
2893 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
2895 path_leaf_el(path));
2902 * Handle rightmost branch now. There's several cases:
2903 * 1) simple rotation leaving records in there. That's trivial.
2904 * 2) rotation requiring a branch delete - there's no more
2905 * records left. Two cases of this:
2906 * a) There are branches to the left.
2907 * b) This is also the leftmost (the only) branch.
2909 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
2910 * 2a) we need the left branch so that we can update it with the unlink
2911 * 2b) we need to bring the inode back to inline extents.
2914 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2916 if (eb->h_next_leaf_blk == 0) {
2918 * This gets a bit tricky if we're going to delete the
2919 * rightmost path. Get the other cases out of the way
2922 if (le16_to_cpu(el->l_next_free_rec) > 1)
2923 goto rightmost_no_delete;
2925 if (le16_to_cpu(el->l_next_free_rec) == 0) {
2927 ocfs2_error(inode->i_sb,
2928 "Inode %llu has empty extent block at %llu",
2929 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2930 (unsigned long long)le64_to_cpu(eb->h_blkno));
2935 * XXX: The caller can not trust "path" any more after
2936 * this as it will have been deleted. What do we do?
2938 * In theory the rotate-for-merge code will never get
2939 * here because it'll always ask for a rotate in a
2943 ret = ocfs2_remove_rightmost_path(inode, handle, path,
2951 * Now we can loop, remembering the path we get from -EAGAIN
2952 * and restarting from there.
2955 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
2956 dealloc, &restart_path, et);
2957 if (ret && ret != -EAGAIN) {
2962 while (ret == -EAGAIN) {
2963 tmp_path = restart_path;
2964 restart_path = NULL;
2966 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
2969 if (ret && ret != -EAGAIN) {
2974 ocfs2_free_path(tmp_path);
2982 ocfs2_free_path(tmp_path);
2983 ocfs2_free_path(restart_path);
2987 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
2990 struct ocfs2_extent_rec *rec = &el->l_recs[index];
2993 if (rec->e_leaf_clusters == 0) {
2995 * We consumed all of the merged-from record. An empty
2996 * extent cannot exist anywhere but the 1st array
2997 * position, so move things over if the merged-from
2998 * record doesn't occupy that position.
3000 * This creates a new empty extent so the caller
3001 * should be smart enough to have removed any existing
3005 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3006 size = index * sizeof(struct ocfs2_extent_rec);
3007 memmove(&el->l_recs[1], &el->l_recs[0], size);
3011 * Always memset - the caller doesn't check whether it
3012 * created an empty extent, so there could be junk in
3015 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3019 static int ocfs2_get_right_path(struct inode *inode,
3020 struct ocfs2_path *left_path,
3021 struct ocfs2_path **ret_right_path)
3025 struct ocfs2_path *right_path = NULL;
3026 struct ocfs2_extent_list *left_el;
3028 *ret_right_path = NULL;
3030 /* This function shouldn't be called for non-trees. */
3031 BUG_ON(left_path->p_tree_depth == 0);
3033 left_el = path_leaf_el(left_path);
3034 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3036 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
3043 /* This function shouldn't be called for the rightmost leaf. */
3044 BUG_ON(right_cpos == 0);
3046 right_path = ocfs2_new_path(path_root_bh(left_path),
3047 path_root_el(left_path));
3054 ret = ocfs2_find_path(inode, right_path, right_cpos);
3060 *ret_right_path = right_path;
3063 ocfs2_free_path(right_path);
3068 * Remove split_rec clusters from the record at index and merge them
3069 * onto the beginning of the record "next" to it.
3070 * For index < l_count - 1, the next means the extent rec at index + 1.
3071 * For index == l_count - 1, the "next" means the 1st extent rec of the
3072 * next extent block.
3074 static int ocfs2_merge_rec_right(struct inode *inode,
3075 struct ocfs2_path *left_path,
3077 struct ocfs2_extent_rec *split_rec,
3080 int ret, next_free, i;
3081 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3082 struct ocfs2_extent_rec *left_rec;
3083 struct ocfs2_extent_rec *right_rec;
3084 struct ocfs2_extent_list *right_el;
3085 struct ocfs2_path *right_path = NULL;
3086 int subtree_index = 0;
3087 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3088 struct buffer_head *bh = path_leaf_bh(left_path);
3089 struct buffer_head *root_bh = NULL;
3091 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3092 left_rec = &el->l_recs[index];
3094 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3095 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3096 /* we meet with a cross extent block merge. */
3097 ret = ocfs2_get_right_path(inode, left_path, &right_path);
3103 right_el = path_leaf_el(right_path);
3104 next_free = le16_to_cpu(right_el->l_next_free_rec);
3105 BUG_ON(next_free <= 0);
3106 right_rec = &right_el->l_recs[0];
3107 if (ocfs2_is_empty_extent(right_rec)) {
3108 BUG_ON(next_free <= 1);
3109 right_rec = &right_el->l_recs[1];
3112 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3113 le16_to_cpu(left_rec->e_leaf_clusters) !=
3114 le32_to_cpu(right_rec->e_cpos));
3116 subtree_index = ocfs2_find_subtree_root(inode,
3117 left_path, right_path);
3119 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3120 handle->h_buffer_credits,
3127 root_bh = left_path->p_node[subtree_index].bh;
3128 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3130 ret = ocfs2_journal_access(handle, inode, root_bh,
3131 OCFS2_JOURNAL_ACCESS_WRITE);
3137 for (i = subtree_index + 1;
3138 i < path_num_items(right_path); i++) {
3139 ret = ocfs2_journal_access(handle, inode,
3140 right_path->p_node[i].bh,
3141 OCFS2_JOURNAL_ACCESS_WRITE);
3147 ret = ocfs2_journal_access(handle, inode,
3148 left_path->p_node[i].bh,
3149 OCFS2_JOURNAL_ACCESS_WRITE);
3157 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3158 right_rec = &el->l_recs[index + 1];
3161 ret = ocfs2_journal_access(handle, inode, bh,
3162 OCFS2_JOURNAL_ACCESS_WRITE);
3168 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3170 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3171 le64_add_cpu(&right_rec->e_blkno,
3172 -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3173 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3175 ocfs2_cleanup_merge(el, index);
3177 ret = ocfs2_journal_dirty(handle, bh);
3182 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3186 ocfs2_complete_edge_insert(inode, handle, left_path,
3187 right_path, subtree_index);
3191 ocfs2_free_path(right_path);
3195 static int ocfs2_get_left_path(struct inode *inode,
3196 struct ocfs2_path *right_path,
3197 struct ocfs2_path **ret_left_path)
3201 struct ocfs2_path *left_path = NULL;
3203 *ret_left_path = NULL;
3205 /* This function shouldn't be called for non-trees. */
3206 BUG_ON(right_path->p_tree_depth == 0);
3208 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
3209 right_path, &left_cpos);
3215 /* This function shouldn't be called for the leftmost leaf. */
3216 BUG_ON(left_cpos == 0);
3218 left_path = ocfs2_new_path(path_root_bh(right_path),
3219 path_root_el(right_path));
3226 ret = ocfs2_find_path(inode, left_path, left_cpos);
3232 *ret_left_path = left_path;
3235 ocfs2_free_path(left_path);
3240 * Remove split_rec clusters from the record at index and merge them
3241 * onto the tail of the record "before" it.
3242 * For index > 0, the "before" means the extent rec at index - 1.
3244 * For index == 0, the "before" means the last record of the previous
3245 * extent block. And there is also a situation that we may need to
3246 * remove the rightmost leaf extent block in the right_path and change
3247 * the right path to indicate the new rightmost path.
3249 static int ocfs2_merge_rec_left(struct inode *inode,
3250 struct ocfs2_path *right_path,
3252 struct ocfs2_extent_rec *split_rec,
3253 struct ocfs2_cached_dealloc_ctxt *dealloc,
3254 struct ocfs2_extent_tree *et,
3257 int ret, i, subtree_index = 0, has_empty_extent = 0;
3258 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3259 struct ocfs2_extent_rec *left_rec;
3260 struct ocfs2_extent_rec *right_rec;
3261 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3262 struct buffer_head *bh = path_leaf_bh(right_path);
3263 struct buffer_head *root_bh = NULL;
3264 struct ocfs2_path *left_path = NULL;
3265 struct ocfs2_extent_list *left_el;
3269 right_rec = &el->l_recs[index];
3271 /* we meet with a cross extent block merge. */
3272 ret = ocfs2_get_left_path(inode, right_path, &left_path);
3278 left_el = path_leaf_el(left_path);
3279 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3280 le16_to_cpu(left_el->l_count));
3282 left_rec = &left_el->l_recs[
3283 le16_to_cpu(left_el->l_next_free_rec) - 1];
3284 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3285 le16_to_cpu(left_rec->e_leaf_clusters) !=
3286 le32_to_cpu(split_rec->e_cpos));
3288 subtree_index = ocfs2_find_subtree_root(inode,
3289 left_path, right_path);
3291 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3292 handle->h_buffer_credits,
3299 root_bh = left_path->p_node[subtree_index].bh;
3300 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3302 ret = ocfs2_journal_access(handle, inode, root_bh,
3303 OCFS2_JOURNAL_ACCESS_WRITE);
3309 for (i = subtree_index + 1;
3310 i < path_num_items(right_path); i++) {
3311 ret = ocfs2_journal_access(handle, inode,
3312 right_path->p_node[i].bh,
3313 OCFS2_JOURNAL_ACCESS_WRITE);
3319 ret = ocfs2_journal_access(handle, inode,
3320 left_path->p_node[i].bh,
3321 OCFS2_JOURNAL_ACCESS_WRITE);
3328 left_rec = &el->l_recs[index - 1];
3329 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3330 has_empty_extent = 1;
3333 ret = ocfs2_journal_access(handle, inode, bh,
3334 OCFS2_JOURNAL_ACCESS_WRITE);
3340 if (has_empty_extent && index == 1) {
3342 * The easy case - we can just plop the record right in.
3344 *left_rec = *split_rec;
3346 has_empty_extent = 0;
3348 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3350 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3351 le64_add_cpu(&right_rec->e_blkno,
3352 ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3353 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3355 ocfs2_cleanup_merge(el, index);
3357 ret = ocfs2_journal_dirty(handle, bh);
3362 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3367 * In the situation that the right_rec is empty and the extent
3368 * block is empty also, ocfs2_complete_edge_insert can't handle
3369 * it and we need to delete the right extent block.
3371 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3372 le16_to_cpu(el->l_next_free_rec) == 1) {
3374 ret = ocfs2_remove_rightmost_path(inode, handle,
3382 /* Now the rightmost extent block has been deleted.
3383 * So we use the new rightmost path.
3385 ocfs2_mv_path(right_path, left_path);
3388 ocfs2_complete_edge_insert(inode, handle, left_path,
3389 right_path, subtree_index);
3393 ocfs2_free_path(left_path);
3397 static int ocfs2_try_to_merge_extent(struct inode *inode,
3399 struct ocfs2_path *path,
3401 struct ocfs2_extent_rec *split_rec,
3402 struct ocfs2_cached_dealloc_ctxt *dealloc,
3403 struct ocfs2_merge_ctxt *ctxt,
3404 struct ocfs2_extent_tree *et)
3408 struct ocfs2_extent_list *el = path_leaf_el(path);
3409 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3411 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3413 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3415 * The merge code will need to create an empty
3416 * extent to take the place of the newly
3417 * emptied slot. Remove any pre-existing empty
3418 * extents - having more than one in a leaf is
3421 ret = ocfs2_rotate_tree_left(inode, handle, path,
3428 rec = &el->l_recs[split_index];
3431 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3433 * Left-right contig implies this.
3435 BUG_ON(!ctxt->c_split_covers_rec);
3438 * Since the leftright insert always covers the entire
3439 * extent, this call will delete the insert record
3440 * entirely, resulting in an empty extent record added to
3443 * Since the adding of an empty extent shifts
3444 * everything back to the right, there's no need to
3445 * update split_index here.
3447 * When the split_index is zero, we need to merge it to the
3448 * prevoius extent block. It is more efficient and easier
3449 * if we do merge_right first and merge_left later.
3451 ret = ocfs2_merge_rec_right(inode, path,
3460 * We can only get this from logic error above.
3462 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3464 /* The merge left us with an empty extent, remove it. */
3465 ret = ocfs2_rotate_tree_left(inode, handle, path,
3472 rec = &el->l_recs[split_index];
3475 * Note that we don't pass split_rec here on purpose -
3476 * we've merged it into the rec already.
3478 ret = ocfs2_merge_rec_left(inode, path,
3488 ret = ocfs2_rotate_tree_left(inode, handle, path,
3491 * Error from this last rotate is not critical, so
3492 * print but don't bubble it up.
3499 * Merge a record to the left or right.
3501 * 'contig_type' is relative to the existing record,
3502 * so for example, if we're "right contig", it's to
3503 * the record on the left (hence the left merge).
3505 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3506 ret = ocfs2_merge_rec_left(inode,
3516 ret = ocfs2_merge_rec_right(inode,
3526 if (ctxt->c_split_covers_rec) {
3528 * The merge may have left an empty extent in
3529 * our leaf. Try to rotate it away.
3531 ret = ocfs2_rotate_tree_left(inode, handle, path,
3543 static void ocfs2_subtract_from_rec(struct super_block *sb,
3544 enum ocfs2_split_type split,
3545 struct ocfs2_extent_rec *rec,
3546 struct ocfs2_extent_rec *split_rec)
3550 len_blocks = ocfs2_clusters_to_blocks(sb,
3551 le16_to_cpu(split_rec->e_leaf_clusters));
3553 if (split == SPLIT_LEFT) {
3555 * Region is on the left edge of the existing
3558 le32_add_cpu(&rec->e_cpos,
3559 le16_to_cpu(split_rec->e_leaf_clusters));
3560 le64_add_cpu(&rec->e_blkno, len_blocks);
3561 le16_add_cpu(&rec->e_leaf_clusters,
3562 -le16_to_cpu(split_rec->e_leaf_clusters));
3565 * Region is on the right edge of the existing
3568 le16_add_cpu(&rec->e_leaf_clusters,
3569 -le16_to_cpu(split_rec->e_leaf_clusters));
3574 * Do the final bits of extent record insertion at the target leaf
3575 * list. If this leaf is part of an allocation tree, it is assumed
3576 * that the tree above has been prepared.
3578 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
3579 struct ocfs2_extent_list *el,
3580 struct ocfs2_insert_type *insert,
3581 struct inode *inode)
3583 int i = insert->ins_contig_index;
3585 struct ocfs2_extent_rec *rec;
3587 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3589 if (insert->ins_split != SPLIT_NONE) {
3590 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3592 rec = &el->l_recs[i];
3593 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
3599 * Contiguous insert - either left or right.
3601 if (insert->ins_contig != CONTIG_NONE) {
3602 rec = &el->l_recs[i];
3603 if (insert->ins_contig == CONTIG_LEFT) {
3604 rec->e_blkno = insert_rec->e_blkno;
3605 rec->e_cpos = insert_rec->e_cpos;
3607 le16_add_cpu(&rec->e_leaf_clusters,
3608 le16_to_cpu(insert_rec->e_leaf_clusters));
3613 * Handle insert into an empty leaf.
3615 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3616 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3617 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3618 el->l_recs[0] = *insert_rec;
3619 el->l_next_free_rec = cpu_to_le16(1);
3626 if (insert->ins_appending == APPEND_TAIL) {
3627 i = le16_to_cpu(el->l_next_free_rec) - 1;
3628 rec = &el->l_recs[i];
3629 range = le32_to_cpu(rec->e_cpos)
3630 + le16_to_cpu(rec->e_leaf_clusters);
3631 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3633 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3634 le16_to_cpu(el->l_count),
3635 "inode %lu, depth %u, count %u, next free %u, "
3636 "rec.cpos %u, rec.clusters %u, "
3637 "insert.cpos %u, insert.clusters %u\n",
3639 le16_to_cpu(el->l_tree_depth),
3640 le16_to_cpu(el->l_count),
3641 le16_to_cpu(el->l_next_free_rec),
3642 le32_to_cpu(el->l_recs[i].e_cpos),
3643 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3644 le32_to_cpu(insert_rec->e_cpos),
3645 le16_to_cpu(insert_rec->e_leaf_clusters));
3647 el->l_recs[i] = *insert_rec;
3648 le16_add_cpu(&el->l_next_free_rec, 1);
3654 * Ok, we have to rotate.
3656 * At this point, it is safe to assume that inserting into an
3657 * empty leaf and appending to a leaf have both been handled
3660 * This leaf needs to have space, either by the empty 1st
3661 * extent record, or by virtue of an l_next_rec < l_count.
3663 ocfs2_rotate_leaf(el, insert_rec);
3666 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3668 struct ocfs2_path *path,
3669 struct ocfs2_extent_rec *insert_rec)
3671 int ret, i, next_free;
3672 struct buffer_head *bh;
3673 struct ocfs2_extent_list *el;
3674 struct ocfs2_extent_rec *rec;
3677 * Update everything except the leaf block.
3679 for (i = 0; i < path->p_tree_depth; i++) {
3680 bh = path->p_node[i].bh;
3681 el = path->p_node[i].el;
3683 next_free = le16_to_cpu(el->l_next_free_rec);
3684 if (next_free == 0) {
3685 ocfs2_error(inode->i_sb,
3686 "Dinode %llu has a bad extent list",
3687 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3692 rec = &el->l_recs[next_free - 1];
3694 rec->e_int_clusters = insert_rec->e_cpos;
3695 le32_add_cpu(&rec->e_int_clusters,
3696 le16_to_cpu(insert_rec->e_leaf_clusters));
3697 le32_add_cpu(&rec->e_int_clusters,
3698 -le32_to_cpu(rec->e_cpos));
3700 ret = ocfs2_journal_dirty(handle, bh);
3707 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3708 struct ocfs2_extent_rec *insert_rec,
3709 struct ocfs2_path *right_path,
3710 struct ocfs2_path **ret_left_path)
3713 struct ocfs2_extent_list *el;
3714 struct ocfs2_path *left_path = NULL;
3716 *ret_left_path = NULL;
3719 * This shouldn't happen for non-trees. The extent rec cluster
3720 * count manipulation below only works for interior nodes.
3722 BUG_ON(right_path->p_tree_depth == 0);
3725 * If our appending insert is at the leftmost edge of a leaf,
3726 * then we might need to update the rightmost records of the
3729 el = path_leaf_el(right_path);
3730 next_free = le16_to_cpu(el->l_next_free_rec);
3731 if (next_free == 0 ||
3732 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3735 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3742 mlog(0, "Append may need a left path update. cpos: %u, "
3743 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3747 * No need to worry if the append is already in the
3751 left_path = ocfs2_new_path(path_root_bh(right_path),
3752 path_root_el(right_path));
3759 ret = ocfs2_find_path(inode, left_path, left_cpos);
3766 * ocfs2_insert_path() will pass the left_path to the
3772 ret = ocfs2_journal_access_path(inode, handle, right_path);
3778 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
3780 *ret_left_path = left_path;
3784 ocfs2_free_path(left_path);
3789 static void ocfs2_split_record(struct inode *inode,
3790 struct ocfs2_path *left_path,
3791 struct ocfs2_path *right_path,
3792 struct ocfs2_extent_rec *split_rec,
3793 enum ocfs2_split_type split)
3796 u32 cpos = le32_to_cpu(split_rec->e_cpos);
3797 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
3798 struct ocfs2_extent_rec *rec, *tmprec;
3800 right_el = path_leaf_el(right_path);;
3802 left_el = path_leaf_el(left_path);
3805 insert_el = right_el;
3806 index = ocfs2_search_extent_list(el, cpos);
3808 if (index == 0 && left_path) {
3809 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3812 * This typically means that the record
3813 * started in the left path but moved to the
3814 * right as a result of rotation. We either
3815 * move the existing record to the left, or we
3816 * do the later insert there.
3818 * In this case, the left path should always
3819 * exist as the rotate code will have passed
3820 * it back for a post-insert update.
3823 if (split == SPLIT_LEFT) {
3825 * It's a left split. Since we know
3826 * that the rotate code gave us an
3827 * empty extent in the left path, we
3828 * can just do the insert there.
3830 insert_el = left_el;
3833 * Right split - we have to move the
3834 * existing record over to the left
3835 * leaf. The insert will be into the
3836 * newly created empty extent in the
3839 tmprec = &right_el->l_recs[index];
3840 ocfs2_rotate_leaf(left_el, tmprec);
3843 memset(tmprec, 0, sizeof(*tmprec));
3844 index = ocfs2_search_extent_list(left_el, cpos);
3845 BUG_ON(index == -1);
3850 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
3852 * Left path is easy - we can just allow the insert to
3856 insert_el = left_el;
3857 index = ocfs2_search_extent_list(el, cpos);
3858 BUG_ON(index == -1);
3861 rec = &el->l_recs[index];
3862 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
3863 ocfs2_rotate_leaf(insert_el, split_rec);
3867 * This function only does inserts on an allocation b-tree. For tree
3868 * depth = 0, ocfs2_insert_at_leaf() is called directly.
3870 * right_path is the path we want to do the actual insert
3871 * in. left_path should only be passed in if we need to update that
3872 * portion of the tree after an edge insert.
3874 static int ocfs2_insert_path(struct inode *inode,
3876 struct ocfs2_path *left_path,
3877 struct ocfs2_path *right_path,
3878 struct ocfs2_extent_rec *insert_rec,
3879 struct ocfs2_insert_type *insert)
3881 int ret, subtree_index;
3882 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
3885 int credits = handle->h_buffer_credits;
3888 * There's a chance that left_path got passed back to
3889 * us without being accounted for in the
3890 * journal. Extend our transaction here to be sure we
3891 * can change those blocks.
3893 credits += left_path->p_tree_depth;
3895 ret = ocfs2_extend_trans(handle, credits);
3901 ret = ocfs2_journal_access_path(inode, handle, left_path);
3909 * Pass both paths to the journal. The majority of inserts
3910 * will be touching all components anyway.
3912 ret = ocfs2_journal_access_path(inode, handle, right_path);
3918 if (insert->ins_split != SPLIT_NONE) {
3920 * We could call ocfs2_insert_at_leaf() for some types
3921 * of splits, but it's easier to just let one separate
3922 * function sort it all out.
3924 ocfs2_split_record(inode, left_path, right_path,
3925 insert_rec, insert->ins_split);
3928 * Split might have modified either leaf and we don't
3929 * have a guarantee that the later edge insert will
3930 * dirty this for us.
3933 ret = ocfs2_journal_dirty(handle,
3934 path_leaf_bh(left_path));
3938 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
3941 ret = ocfs2_journal_dirty(handle, leaf_bh);
3947 * The rotate code has indicated that we need to fix
3948 * up portions of the tree after the insert.
3950 * XXX: Should we extend the transaction here?
3952 subtree_index = ocfs2_find_subtree_root(inode, left_path,
3954 ocfs2_complete_edge_insert(inode, handle, left_path,
3955 right_path, subtree_index);
3963 static int ocfs2_do_insert_extent(struct inode *inode,
3965 struct ocfs2_extent_tree *et,
3966 struct ocfs2_extent_rec *insert_rec,
3967 struct ocfs2_insert_type *type)
3969 int ret, rotate = 0;
3971 struct ocfs2_path *right_path = NULL;
3972 struct ocfs2_path *left_path = NULL;
3973 struct ocfs2_extent_list *el;
3975 el = et->et_root_el;
3977 ret = ocfs2_journal_access(handle, inode, et->et_root_bh,
3978 OCFS2_JOURNAL_ACCESS_WRITE);
3984 if (le16_to_cpu(el->l_tree_depth) == 0) {
3985 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
3986 goto out_update_clusters;
3989 right_path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
3997 * Determine the path to start with. Rotations need the
3998 * rightmost path, everything else can go directly to the
4001 cpos = le32_to_cpu(insert_rec->e_cpos);
4002 if (type->ins_appending == APPEND_NONE &&
4003 type->ins_contig == CONTIG_NONE) {
4008 ret = ocfs2_find_path(inode, right_path, cpos);
4015 * Rotations and appends need special treatment - they modify
4016 * parts of the tree's above them.
4018 * Both might pass back a path immediate to the left of the
4019 * one being inserted to. This will be cause
4020 * ocfs2_insert_path() to modify the rightmost records of
4021 * left_path to account for an edge insert.
4023 * XXX: When modifying this code, keep in mind that an insert
4024 * can wind up skipping both of these two special cases...
4027 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
4028 le32_to_cpu(insert_rec->e_cpos),
4029 right_path, &left_path);
4036 * ocfs2_rotate_tree_right() might have extended the
4037 * transaction without re-journaling our tree root.
4039 ret = ocfs2_journal_access(handle, inode, et->et_root_bh,
4040 OCFS2_JOURNAL_ACCESS_WRITE);
4045 } else if (type->ins_appending == APPEND_TAIL
4046 && type->ins_contig != CONTIG_LEFT) {
4047 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
4048 right_path, &left_path);
4055 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
4062 out_update_clusters:
4063 if (type->ins_split == SPLIT_NONE)
4064 ocfs2_et_update_clusters(inode, et,
4065 le16_to_cpu(insert_rec->e_leaf_clusters));
4067 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
4072 ocfs2_free_path(left_path);
4073 ocfs2_free_path(right_path);
4078 static enum ocfs2_contig_type
4079 ocfs2_figure_merge_contig_type(struct inode *inode, struct ocfs2_path *path,
4080 struct ocfs2_extent_list *el, int index,
4081 struct ocfs2_extent_rec *split_rec)
4084 enum ocfs2_contig_type ret = CONTIG_NONE;
4085 u32 left_cpos, right_cpos;
4086 struct ocfs2_extent_rec *rec = NULL;
4087 struct ocfs2_extent_list *new_el;
4088 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4089 struct buffer_head *bh;
4090 struct ocfs2_extent_block *eb;
4093 rec = &el->l_recs[index - 1];
4094 } else if (path->p_tree_depth > 0) {
4095 status = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
4100 if (left_cpos != 0) {
4101 left_path = ocfs2_new_path(path_root_bh(path),
4102 path_root_el(path));
4106 status = ocfs2_find_path(inode, left_path, left_cpos);
4110 new_el = path_leaf_el(left_path);
4112 if (le16_to_cpu(new_el->l_next_free_rec) !=
4113 le16_to_cpu(new_el->l_count)) {
4114 bh = path_leaf_bh(left_path);
4115 eb = (struct ocfs2_extent_block *)bh->b_data;
4116 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
4120 rec = &new_el->l_recs[
4121 le16_to_cpu(new_el->l_next_free_rec) - 1];
4126 * We're careful to check for an empty extent record here -
4127 * the merge code will know what to do if it sees one.
4130 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4131 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4134 ret = ocfs2_extent_contig(inode, rec, split_rec);
4139 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4140 rec = &el->l_recs[index + 1];
4141 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4142 path->p_tree_depth > 0) {
4143 status = ocfs2_find_cpos_for_right_leaf(inode->i_sb,
4148 if (right_cpos == 0)
4151 right_path = ocfs2_new_path(path_root_bh(path),
4152 path_root_el(path));
4156 status = ocfs2_find_path(inode, right_path, right_cpos);
4160 new_el = path_leaf_el(right_path);
4161 rec = &new_el->l_recs[0];
4162 if (ocfs2_is_empty_extent(rec)) {
4163 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4164 bh = path_leaf_bh(right_path);
4165 eb = (struct ocfs2_extent_block *)bh->b_data;
4166 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
4170 rec = &new_el->l_recs[1];
4175 enum ocfs2_contig_type contig_type;
4177 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
4179 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4180 ret = CONTIG_LEFTRIGHT;
4181 else if (ret == CONTIG_NONE)
4187 ocfs2_free_path(left_path);
4189 ocfs2_free_path(right_path);
4194 static void ocfs2_figure_contig_type(struct inode *inode,
4195 struct ocfs2_insert_type *insert,
4196 struct ocfs2_extent_list *el,
4197 struct ocfs2_extent_rec *insert_rec,
4198 struct ocfs2_extent_tree *et)
4201 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4203 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4205 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4206 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
4208 if (contig_type != CONTIG_NONE) {
4209 insert->ins_contig_index = i;
4213 insert->ins_contig = contig_type;
4215 if (insert->ins_contig != CONTIG_NONE) {
4216 struct ocfs2_extent_rec *rec =
4217 &el->l_recs[insert->ins_contig_index];
4218 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4219 le16_to_cpu(insert_rec->e_leaf_clusters);
4222 * Caller might want us to limit the size of extents, don't
4223 * calculate contiguousness if we might exceed that limit.
4225 if (et->et_max_leaf_clusters &&
4226 (len > et->et_max_leaf_clusters))
4227 insert->ins_contig = CONTIG_NONE;
4232 * This should only be called against the righmost leaf extent list.
4234 * ocfs2_figure_appending_type() will figure out whether we'll have to
4235 * insert at the tail of the rightmost leaf.
4237 * This should also work against the root extent list for tree's with 0
4238 * depth. If we consider the root extent list to be the rightmost leaf node
4239 * then the logic here makes sense.
4241 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4242 struct ocfs2_extent_list *el,
4243 struct ocfs2_extent_rec *insert_rec)
4246 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4247 struct ocfs2_extent_rec *rec;
4249 insert->ins_appending = APPEND_NONE;
4251 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4253 if (!el->l_next_free_rec)
4254 goto set_tail_append;
4256 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4257 /* Were all records empty? */
4258 if (le16_to_cpu(el->l_next_free_rec) == 1)
4259 goto set_tail_append;
4262 i = le16_to_cpu(el->l_next_free_rec) - 1;
4263 rec = &el->l_recs[i];
4266 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4267 goto set_tail_append;
4272 insert->ins_appending = APPEND_TAIL;
4276 * Helper function called at the begining of an insert.
4278 * This computes a few things that are commonly used in the process of
4279 * inserting into the btree:
4280 * - Whether the new extent is contiguous with an existing one.
4281 * - The current tree depth.
4282 * - Whether the insert is an appending one.
4283 * - The total # of free records in the tree.
4285 * All of the information is stored on the ocfs2_insert_type
4288 static int ocfs2_figure_insert_type(struct inode *inode,
4289 struct ocfs2_extent_tree *et,
4290 struct buffer_head **last_eb_bh,
4291 struct ocfs2_extent_rec *insert_rec,
4293 struct ocfs2_insert_type *insert)
4296 struct ocfs2_extent_block *eb;
4297 struct ocfs2_extent_list *el;
4298 struct ocfs2_path *path = NULL;
4299 struct buffer_head *bh = NULL;
4301 insert->ins_split = SPLIT_NONE;
4303 el = et->et_root_el;
4304 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4306 if (el->l_tree_depth) {
4308 * If we have tree depth, we read in the
4309 * rightmost extent block ahead of time as
4310 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4311 * may want it later.
4313 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4314 ocfs2_et_get_last_eb_blk(et), &bh,
4315 OCFS2_BH_CACHED, inode);
4320 eb = (struct ocfs2_extent_block *) bh->b_data;
4325 * Unless we have a contiguous insert, we'll need to know if
4326 * there is room left in our allocation tree for another
4329 * XXX: This test is simplistic, we can search for empty
4330 * extent records too.
4332 *free_records = le16_to_cpu(el->l_count) -
4333 le16_to_cpu(el->l_next_free_rec);
4335 if (!insert->ins_tree_depth) {
4336 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4337 ocfs2_figure_appending_type(insert, el, insert_rec);
4341 path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
4349 * In the case that we're inserting past what the tree
4350 * currently accounts for, ocfs2_find_path() will return for
4351 * us the rightmost tree path. This is accounted for below in
4352 * the appending code.
4354 ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
4360 el = path_leaf_el(path);
4363 * Now that we have the path, there's two things we want to determine:
4364 * 1) Contiguousness (also set contig_index if this is so)
4366 * 2) Are we doing an append? We can trivially break this up
4367 * into two types of appends: simple record append, or a
4368 * rotate inside the tail leaf.
4370 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4373 * The insert code isn't quite ready to deal with all cases of
4374 * left contiguousness. Specifically, if it's an insert into
4375 * the 1st record in a leaf, it will require the adjustment of
4376 * cluster count on the last record of the path directly to it's
4377 * left. For now, just catch that case and fool the layers
4378 * above us. This works just fine for tree_depth == 0, which
4379 * is why we allow that above.
4381 if (insert->ins_contig == CONTIG_LEFT &&
4382 insert->ins_contig_index == 0)
4383 insert->ins_contig = CONTIG_NONE;
4386 * Ok, so we can simply compare against last_eb to figure out
4387 * whether the path doesn't exist. This will only happen in
4388 * the case that we're doing a tail append, so maybe we can
4389 * take advantage of that information somehow.
4391 if (ocfs2_et_get_last_eb_blk(et) ==
4392 path_leaf_bh(path)->b_blocknr) {
4394 * Ok, ocfs2_find_path() returned us the rightmost
4395 * tree path. This might be an appending insert. There are
4397 * 1) We're doing a true append at the tail:
4398 * -This might even be off the end of the leaf
4399 * 2) We're "appending" by rotating in the tail
4401 ocfs2_figure_appending_type(insert, el, insert_rec);
4405 ocfs2_free_path(path);
4415 * Insert an extent into an inode btree.
4417 * The caller needs to update fe->i_clusters
4419 int ocfs2_insert_extent(struct ocfs2_super *osb,
4421 struct inode *inode,
4422 struct ocfs2_extent_tree *et,
4427 struct ocfs2_alloc_context *meta_ac)
4430 int uninitialized_var(free_records);
4431 struct buffer_head *last_eb_bh = NULL;
4432 struct ocfs2_insert_type insert = {0, };
4433 struct ocfs2_extent_rec rec;
4435 mlog(0, "add %u clusters at position %u to inode %llu\n",
4436 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4438 memset(&rec, 0, sizeof(rec));
4439 rec.e_cpos = cpu_to_le32(cpos);
4440 rec.e_blkno = cpu_to_le64(start_blk);
4441 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4442 rec.e_flags = flags;
4443 status = ocfs2_et_insert_check(inode, et, &rec);
4449 status = ocfs2_figure_insert_type(inode, et, &last_eb_bh, &rec,
4450 &free_records, &insert);
4456 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4457 "Insert.contig_index: %d, Insert.free_records: %d, "
4458 "Insert.tree_depth: %d\n",
4459 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
4460 free_records, insert.ins_tree_depth);
4462 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4463 status = ocfs2_grow_tree(inode, handle, et,
4464 &insert.ins_tree_depth, &last_eb_bh,
4472 /* Finally, we can add clusters. This might rotate the tree for us. */
4473 status = ocfs2_do_insert_extent(inode, handle, et, &rec, &insert);
4476 else if (et->et_ops == &ocfs2_dinode_et_ops)
4477 ocfs2_extent_map_insert_rec(inode, &rec);
4488 * Allcate and add clusters into the extent b-tree.
4489 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4490 * The extent b-tree's root is specified by et, and
4491 * it is not limited to the file storage. Any extent tree can use this
4492 * function if it implements the proper ocfs2_extent_tree.
4494 int ocfs2_add_clusters_in_btree(struct ocfs2_super *osb,
4495 struct inode *inode,
4496 u32 *logical_offset,
4497 u32 clusters_to_add,
4499 struct ocfs2_extent_tree *et,
4501 struct ocfs2_alloc_context *data_ac,
4502 struct ocfs2_alloc_context *meta_ac,
4503 enum ocfs2_alloc_restarted *reason_ret)
4507 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4508 u32 bit_off, num_bits;
4512 BUG_ON(!clusters_to_add);
4515 flags = OCFS2_EXT_UNWRITTEN;
4517 free_extents = ocfs2_num_free_extents(osb, inode, et);
4518 if (free_extents < 0) {
4519 status = free_extents;
4524 /* there are two cases which could cause us to EAGAIN in the
4525 * we-need-more-metadata case:
4526 * 1) we haven't reserved *any*
4527 * 2) we are so fragmented, we've needed to add metadata too
4529 if (!free_extents && !meta_ac) {
4530 mlog(0, "we haven't reserved any metadata!\n");
4532 reason = RESTART_META;
4534 } else if ((!free_extents)
4535 && (ocfs2_alloc_context_bits_left(meta_ac)
4536 < ocfs2_extend_meta_needed(et->et_root_el))) {
4537 mlog(0, "filesystem is really fragmented...\n");
4539 reason = RESTART_META;
4543 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
4544 clusters_to_add, &bit_off, &num_bits);
4546 if (status != -ENOSPC)
4551 BUG_ON(num_bits > clusters_to_add);
4553 /* reserve our write early -- insert_extent may update the inode */
4554 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
4555 OCFS2_JOURNAL_ACCESS_WRITE);
4561 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4562 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
4563 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4564 status = ocfs2_insert_extent(osb, handle, inode, et,
4565 *logical_offset, block,
4566 num_bits, flags, meta_ac);
4572 status = ocfs2_journal_dirty(handle, et->et_root_bh);
4578 clusters_to_add -= num_bits;
4579 *logical_offset += num_bits;
4581 if (clusters_to_add) {
4582 mlog(0, "need to alloc once more, wanted = %u\n",
4585 reason = RESTART_TRANS;
4591 *reason_ret = reason;
4595 static void ocfs2_make_right_split_rec(struct super_block *sb,
4596 struct ocfs2_extent_rec *split_rec,
4598 struct ocfs2_extent_rec *rec)
4600 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4601 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4603 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4605 split_rec->e_cpos = cpu_to_le32(cpos);
4606 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4608 split_rec->e_blkno = rec->e_blkno;
4609 le64_add_cpu(&split_rec->e_blkno,
4610 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4612 split_rec->e_flags = rec->e_flags;
4615 static int ocfs2_split_and_insert(struct inode *inode,
4617 struct ocfs2_path *path,
4618 struct ocfs2_extent_tree *et,
4619 struct buffer_head **last_eb_bh,
4621 struct ocfs2_extent_rec *orig_split_rec,
4622 struct ocfs2_alloc_context *meta_ac)
4625 unsigned int insert_range, rec_range, do_leftright = 0;
4626 struct ocfs2_extent_rec tmprec;
4627 struct ocfs2_extent_list *rightmost_el;
4628 struct ocfs2_extent_rec rec;
4629 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4630 struct ocfs2_insert_type insert;
4631 struct ocfs2_extent_block *eb;
4635 * Store a copy of the record on the stack - it might move
4636 * around as the tree is manipulated below.
4638 rec = path_leaf_el(path)->l_recs[split_index];
4640 rightmost_el = et->et_root_el;
4642 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4644 BUG_ON(!(*last_eb_bh));
4645 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4646 rightmost_el = &eb->h_list;
4649 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4650 le16_to_cpu(rightmost_el->l_count)) {
4651 ret = ocfs2_grow_tree(inode, handle, et,
4652 &depth, last_eb_bh, meta_ac);
4659 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4660 insert.ins_appending = APPEND_NONE;
4661 insert.ins_contig = CONTIG_NONE;
4662 insert.ins_tree_depth = depth;
4664 insert_range = le32_to_cpu(split_rec.e_cpos) +
4665 le16_to_cpu(split_rec.e_leaf_clusters);
4666 rec_range = le32_to_cpu(rec.e_cpos) +
4667 le16_to_cpu(rec.e_leaf_clusters);
4669 if (split_rec.e_cpos == rec.e_cpos) {
4670 insert.ins_split = SPLIT_LEFT;
4671 } else if (insert_range == rec_range) {
4672 insert.ins_split = SPLIT_RIGHT;
4675 * Left/right split. We fake this as a right split
4676 * first and then make a second pass as a left split.
4678 insert.ins_split = SPLIT_RIGHT;
4680 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
4685 BUG_ON(do_leftright);
4689 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4695 if (do_leftright == 1) {
4697 struct ocfs2_extent_list *el;
4700 split_rec = *orig_split_rec;
4702 ocfs2_reinit_path(path, 1);
4704 cpos = le32_to_cpu(split_rec.e_cpos);
4705 ret = ocfs2_find_path(inode, path, cpos);
4711 el = path_leaf_el(path);
4712 split_index = ocfs2_search_extent_list(el, cpos);
4721 * Mark part or all of the extent record at split_index in the leaf
4722 * pointed to by path as written. This removes the unwritten
4725 * Care is taken to handle contiguousness so as to not grow the tree.
4727 * meta_ac is not strictly necessary - we only truly need it if growth
4728 * of the tree is required. All other cases will degrade into a less
4729 * optimal tree layout.
4731 * last_eb_bh should be the rightmost leaf block for any extent
4732 * btree. Since a split may grow the tree or a merge might shrink it,
4733 * the caller cannot trust the contents of that buffer after this call.
4735 * This code is optimized for readability - several passes might be
4736 * made over certain portions of the tree. All of those blocks will
4737 * have been brought into cache (and pinned via the journal), so the
4738 * extra overhead is not expressed in terms of disk reads.
4740 static int __ocfs2_mark_extent_written(struct inode *inode,
4741 struct ocfs2_extent_tree *et,
4743 struct ocfs2_path *path,
4745 struct ocfs2_extent_rec *split_rec,
4746 struct ocfs2_alloc_context *meta_ac,
4747 struct ocfs2_cached_dealloc_ctxt *dealloc)
4750 struct ocfs2_extent_list *el = path_leaf_el(path);
4751 struct buffer_head *last_eb_bh = NULL;
4752 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
4753 struct ocfs2_merge_ctxt ctxt;
4754 struct ocfs2_extent_list *rightmost_el;
4756 if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) {
4762 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
4763 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
4764 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
4770 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, path, el,
4775 * The core merge / split code wants to know how much room is
4776 * left in this inodes allocation tree, so we pass the
4777 * rightmost extent list.
4779 if (path->p_tree_depth) {
4780 struct ocfs2_extent_block *eb;
4782 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4783 ocfs2_et_get_last_eb_blk(et),
4784 &last_eb_bh, OCFS2_BH_CACHED, inode);
4790 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4791 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
4792 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
4797 rightmost_el = &eb->h_list;
4799 rightmost_el = path_root_el(path);
4801 if (rec->e_cpos == split_rec->e_cpos &&
4802 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
4803 ctxt.c_split_covers_rec = 1;
4805 ctxt.c_split_covers_rec = 0;
4807 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
4809 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
4810 split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent,
4811 ctxt.c_split_covers_rec);
4813 if (ctxt.c_contig_type == CONTIG_NONE) {
4814 if (ctxt.c_split_covers_rec)
4815 el->l_recs[split_index] = *split_rec;
4817 ret = ocfs2_split_and_insert(inode, handle, path, et,
4818 &last_eb_bh, split_index,
4819 split_rec, meta_ac);
4823 ret = ocfs2_try_to_merge_extent(inode, handle, path,
4824 split_index, split_rec,
4825 dealloc, &ctxt, et);
4836 * Mark the already-existing extent at cpos as written for len clusters.
4838 * If the existing extent is larger than the request, initiate a
4839 * split. An attempt will be made at merging with adjacent extents.
4841 * The caller is responsible for passing down meta_ac if we'll need it.
4843 int ocfs2_mark_extent_written(struct inode *inode,
4844 struct ocfs2_extent_tree *et,
4845 handle_t *handle, u32 cpos, u32 len, u32 phys,
4846 struct ocfs2_alloc_context *meta_ac,
4847 struct ocfs2_cached_dealloc_ctxt *dealloc)
4850 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
4851 struct ocfs2_extent_rec split_rec;
4852 struct ocfs2_path *left_path = NULL;
4853 struct ocfs2_extent_list *el;
4855 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4856 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
4858 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
4859 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
4860 "that are being written to, but the feature bit "
4861 "is not set in the super block.",
4862 (unsigned long long)OCFS2_I(inode)->ip_blkno);
4868 * XXX: This should be fixed up so that we just re-insert the
4869 * next extent records.
4871 * XXX: This is a hack on the extent tree, maybe it should be
4874 if (et->et_ops == &ocfs2_dinode_et_ops)
4875 ocfs2_extent_map_trunc(inode, 0);
4877 left_path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
4884 ret = ocfs2_find_path(inode, left_path, cpos);
4889 el = path_leaf_el(left_path);
4891 index = ocfs2_search_extent_list(el, cpos);
4892 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4893 ocfs2_error(inode->i_sb,
4894 "Inode %llu has an extent at cpos %u which can no "
4895 "longer be found.\n",
4896 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
4901 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
4902 split_rec.e_cpos = cpu_to_le32(cpos);
4903 split_rec.e_leaf_clusters = cpu_to_le16(len);
4904 split_rec.e_blkno = cpu_to_le64(start_blkno);
4905 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
4906 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
4908 ret = __ocfs2_mark_extent_written(inode, et, handle, left_path,
4909 index, &split_rec, meta_ac,
4915 ocfs2_free_path(left_path);
4919 static int ocfs2_split_tree(struct inode *inode, struct ocfs2_extent_tree *et,
4920 handle_t *handle, struct ocfs2_path *path,
4921 int index, u32 new_range,
4922 struct ocfs2_alloc_context *meta_ac)
4924 int ret, depth, credits = handle->h_buffer_credits;
4925 struct buffer_head *last_eb_bh = NULL;
4926 struct ocfs2_extent_block *eb;
4927 struct ocfs2_extent_list *rightmost_el, *el;
4928 struct ocfs2_extent_rec split_rec;
4929 struct ocfs2_extent_rec *rec;
4930 struct ocfs2_insert_type insert;
4933 * Setup the record to split before we grow the tree.
4935 el = path_leaf_el(path);
4936 rec = &el->l_recs[index];
4937 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
4939 depth = path->p_tree_depth;
4941 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4942 ocfs2_et_get_last_eb_blk(et),
4943 &last_eb_bh, OCFS2_BH_CACHED, inode);
4949 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4950 rightmost_el = &eb->h_list;
4952 rightmost_el = path_leaf_el(path);
4954 credits += path->p_tree_depth +
4955 ocfs2_extend_meta_needed(et->et_root_el);
4956 ret = ocfs2_extend_trans(handle, credits);
4962 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4963 le16_to_cpu(rightmost_el->l_count)) {
4964 ret = ocfs2_grow_tree(inode, handle, et, &depth, &last_eb_bh,
4972 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4973 insert.ins_appending = APPEND_NONE;
4974 insert.ins_contig = CONTIG_NONE;
4975 insert.ins_split = SPLIT_RIGHT;
4976 insert.ins_tree_depth = depth;
4978 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4987 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
4988 struct ocfs2_path *path, int index,
4989 struct ocfs2_cached_dealloc_ctxt *dealloc,
4991 struct ocfs2_extent_tree *et)
4994 u32 left_cpos, rec_range, trunc_range;
4995 int wants_rotate = 0, is_rightmost_tree_rec = 0;
4996 struct super_block *sb = inode->i_sb;
4997 struct ocfs2_path *left_path = NULL;
4998 struct ocfs2_extent_list *el = path_leaf_el(path);
4999 struct ocfs2_extent_rec *rec;
5000 struct ocfs2_extent_block *eb;
5002 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5003 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5012 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5013 path->p_tree_depth) {
5015 * Check whether this is the rightmost tree record. If
5016 * we remove all of this record or part of its right
5017 * edge then an update of the record lengths above it
5020 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5021 if (eb->h_next_leaf_blk == 0)
5022 is_rightmost_tree_rec = 1;
5025 rec = &el->l_recs[index];
5026 if (index == 0 && path->p_tree_depth &&
5027 le32_to_cpu(rec->e_cpos) == cpos) {
5029 * Changing the leftmost offset (via partial or whole
5030 * record truncate) of an interior (or rightmost) path
5031 * means we have to update the subtree that is formed
5032 * by this leaf and the one to it's left.
5034 * There are two cases we can skip:
5035 * 1) Path is the leftmost one in our inode tree.
5036 * 2) The leaf is rightmost and will be empty after
5037 * we remove the extent record - the rotate code
5038 * knows how to update the newly formed edge.
5041 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
5048 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5049 left_path = ocfs2_new_path(path_root_bh(path),
5050 path_root_el(path));
5057 ret = ocfs2_find_path(inode, left_path, left_cpos);
5065 ret = ocfs2_extend_rotate_transaction(handle, 0,
5066 handle->h_buffer_credits,
5073 ret = ocfs2_journal_access_path(inode, handle, path);
5079 ret = ocfs2_journal_access_path(inode, handle, left_path);
5085 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5086 trunc_range = cpos + len;
5088 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5091 memset(rec, 0, sizeof(*rec));
5092 ocfs2_cleanup_merge(el, index);
5095 next_free = le16_to_cpu(el->l_next_free_rec);
5096 if (is_rightmost_tree_rec && next_free > 1) {
5098 * We skip the edge update if this path will
5099 * be deleted by the rotate code.
5101 rec = &el->l_recs[next_free - 1];
5102 ocfs2_adjust_rightmost_records(inode, handle, path,
5105 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5106 /* Remove leftmost portion of the record. */
5107 le32_add_cpu(&rec->e_cpos, len);
5108 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5109 le16_add_cpu(&rec->e_leaf_clusters, -len);
5110 } else if (rec_range == trunc_range) {
5111 /* Remove rightmost portion of the record */
5112 le16_add_cpu(&rec->e_leaf_clusters, -len);
5113 if (is_rightmost_tree_rec)
5114 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
5116 /* Caller should have trapped this. */
5117 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
5118 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
5119 le32_to_cpu(rec->e_cpos),
5120 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5127 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
5128 ocfs2_complete_edge_insert(inode, handle, left_path, path,
5132 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5134 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5141 ocfs2_free_path(left_path);
5145 int ocfs2_remove_extent(struct inode *inode,
5146 struct ocfs2_extent_tree *et,
5147 u32 cpos, u32 len, handle_t *handle,
5148 struct ocfs2_alloc_context *meta_ac,
5149 struct ocfs2_cached_dealloc_ctxt *dealloc)
5152 u32 rec_range, trunc_range;
5153 struct ocfs2_extent_rec *rec;
5154 struct ocfs2_extent_list *el;
5155 struct ocfs2_path *path = NULL;
5157 ocfs2_extent_map_trunc(inode, 0);
5159 path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
5166 ret = ocfs2_find_path(inode, path, cpos);
5172 el = path_leaf_el(path);
5173 index = ocfs2_search_extent_list(el, cpos);
5174 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5175 ocfs2_error(inode->i_sb,
5176 "Inode %llu has an extent at cpos %u which can no "
5177 "longer be found.\n",
5178 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5184 * We have 3 cases of extent removal:
5185 * 1) Range covers the entire extent rec
5186 * 2) Range begins or ends on one edge of the extent rec
5187 * 3) Range is in the middle of the extent rec (no shared edges)
5189 * For case 1 we remove the extent rec and left rotate to
5192 * For case 2 we just shrink the existing extent rec, with a
5193 * tree update if the shrinking edge is also the edge of an
5196 * For case 3 we do a right split to turn the extent rec into
5197 * something case 2 can handle.
5199 rec = &el->l_recs[index];
5200 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5201 trunc_range = cpos + len;
5203 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5205 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
5206 "(cpos %u, len %u)\n",
5207 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
5208 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
5210 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5211 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5218 ret = ocfs2_split_tree(inode, et, handle, path, index,
5219 trunc_range, meta_ac);
5226 * The split could have manipulated the tree enough to
5227 * move the record location, so we have to look for it again.
5229 ocfs2_reinit_path(path, 1);
5231 ret = ocfs2_find_path(inode, path, cpos);
5237 el = path_leaf_el(path);
5238 index = ocfs2_search_extent_list(el, cpos);
5239 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5240 ocfs2_error(inode->i_sb,
5241 "Inode %llu: split at cpos %u lost record.",
5242 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5249 * Double check our values here. If anything is fishy,
5250 * it's easier to catch it at the top level.
5252 rec = &el->l_recs[index];
5253 rec_range = le32_to_cpu(rec->e_cpos) +
5254 ocfs2_rec_clusters(el, rec);
5255 if (rec_range != trunc_range) {
5256 ocfs2_error(inode->i_sb,
5257 "Inode %llu: error after split at cpos %u"
5258 "trunc len %u, existing record is (%u,%u)",
5259 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5260 cpos, len, le32_to_cpu(rec->e_cpos),
5261 ocfs2_rec_clusters(el, rec));
5266 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5275 ocfs2_free_path(path);
5279 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5281 struct buffer_head *tl_bh = osb->osb_tl_bh;
5282 struct ocfs2_dinode *di;
5283 struct ocfs2_truncate_log *tl;
5285 di = (struct ocfs2_dinode *) tl_bh->b_data;
5286 tl = &di->id2.i_dealloc;
5288 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5289 "slot %d, invalid truncate log parameters: used = "
5290 "%u, count = %u\n", osb->slot_num,
5291 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5292 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5295 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5296 unsigned int new_start)
5298 unsigned int tail_index;
5299 unsigned int current_tail;
5301 /* No records, nothing to coalesce */
5302 if (!le16_to_cpu(tl->tl_used))
5305 tail_index = le16_to_cpu(tl->tl_used) - 1;
5306 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5307 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5309 return current_tail == new_start;
5312 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5315 unsigned int num_clusters)
5318 unsigned int start_cluster, tl_count;
5319 struct inode *tl_inode = osb->osb_tl_inode;
5320 struct buffer_head *tl_bh = osb->osb_tl_bh;
5321 struct ocfs2_dinode *di;
5322 struct ocfs2_truncate_log *tl;
5324 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5325 (unsigned long long)start_blk, num_clusters);
5327 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5329 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5331 di = (struct ocfs2_dinode *) tl_bh->b_data;
5332 tl = &di->id2.i_dealloc;
5333 if (!OCFS2_IS_VALID_DINODE(di)) {
5334 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
5339 tl_count = le16_to_cpu(tl->tl_count);
5340 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5342 "Truncate record count on #%llu invalid "
5343 "wanted %u, actual %u\n",
5344 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5345 ocfs2_truncate_recs_per_inode(osb->sb),
5346 le16_to_cpu(tl->tl_count));
5348 /* Caller should have known to flush before calling us. */
5349 index = le16_to_cpu(tl->tl_used);
5350 if (index >= tl_count) {
5356 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
5357 OCFS2_JOURNAL_ACCESS_WRITE);
5363 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5364 "%llu (index = %d)\n", num_clusters, start_cluster,
5365 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
5367 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5369 * Move index back to the record we are coalescing with.
5370 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5374 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5375 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5376 index, le32_to_cpu(tl->tl_recs[index].t_start),
5379 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5380 tl->tl_used = cpu_to_le16(index + 1);
5382 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5384 status = ocfs2_journal_dirty(handle, tl_bh);
5395 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5397 struct inode *data_alloc_inode,
5398 struct buffer_head *data_alloc_bh)
5402 unsigned int num_clusters;
5404 struct ocfs2_truncate_rec rec;
5405 struct ocfs2_dinode *di;
5406 struct ocfs2_truncate_log *tl;
5407 struct inode *tl_inode = osb->osb_tl_inode;
5408 struct buffer_head *tl_bh = osb->osb_tl_bh;
5412 di = (struct ocfs2_dinode *) tl_bh->b_data;
5413 tl = &di->id2.i_dealloc;
5414 i = le16_to_cpu(tl->tl_used) - 1;
5416 /* Caller has given us at least enough credits to
5417 * update the truncate log dinode */
5418 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
5419 OCFS2_JOURNAL_ACCESS_WRITE);
5425 tl->tl_used = cpu_to_le16(i);
5427 status = ocfs2_journal_dirty(handle, tl_bh);
5433 /* TODO: Perhaps we can calculate the bulk of the
5434 * credits up front rather than extending like
5436 status = ocfs2_extend_trans(handle,
5437 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5443 rec = tl->tl_recs[i];
5444 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5445 le32_to_cpu(rec.t_start));
5446 num_clusters = le32_to_cpu(rec.t_clusters);
5448 /* if start_blk is not set, we ignore the record as
5451 mlog(0, "free record %d, start = %u, clusters = %u\n",
5452 i, le32_to_cpu(rec.t_start), num_clusters);
5454 status = ocfs2_free_clusters(handle, data_alloc_inode,
5455 data_alloc_bh, start_blk,
5470 /* Expects you to already be holding tl_inode->i_mutex */
5471 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5474 unsigned int num_to_flush;
5476 struct inode *tl_inode = osb->osb_tl_inode;
5477 struct inode *data_alloc_inode = NULL;
5478 struct buffer_head *tl_bh = osb->osb_tl_bh;
5479 struct buffer_head *data_alloc_bh = NULL;
5480 struct ocfs2_dinode *di;
5481 struct ocfs2_truncate_log *tl;
5485 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5487 di = (struct ocfs2_dinode *) tl_bh->b_data;
5488 tl = &di->id2.i_dealloc;
5489 if (!OCFS2_IS_VALID_DINODE(di)) {
5490 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
5495 num_to_flush = le16_to_cpu(tl->tl_used);
5496 mlog(0, "Flush %u records from truncate log #%llu\n",
5497 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
5498 if (!num_to_flush) {
5503 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5504 GLOBAL_BITMAP_SYSTEM_INODE,
5505 OCFS2_INVALID_SLOT);
5506 if (!data_alloc_inode) {
5508 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5512 mutex_lock(&data_alloc_inode->i_mutex);
5514 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5520 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5521 if (IS_ERR(handle)) {
5522 status = PTR_ERR(handle);
5527 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
5532 ocfs2_commit_trans(osb, handle);
5535 brelse(data_alloc_bh);
5536 ocfs2_inode_unlock(data_alloc_inode, 1);
5539 mutex_unlock(&data_alloc_inode->i_mutex);
5540 iput(data_alloc_inode);
5547 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5550 struct inode *tl_inode = osb->osb_tl_inode;
5552 mutex_lock(&tl_inode->i_mutex);
5553 status = __ocfs2_flush_truncate_log(osb);
5554 mutex_unlock(&tl_inode->i_mutex);
5559 static void ocfs2_truncate_log_worker(struct work_struct *work)
5562 struct ocfs2_super *osb =
5563 container_of(work, struct ocfs2_super,
5564 osb_truncate_log_wq.work);
5568 status = ocfs2_flush_truncate_log(osb);
5572 ocfs2_init_inode_steal_slot(osb);
5577 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5578 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
5581 if (osb->osb_tl_inode) {
5582 /* We want to push off log flushes while truncates are
5585 cancel_delayed_work(&osb->osb_truncate_log_wq);
5587 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
5588 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
5592 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
5594 struct inode **tl_inode,
5595 struct buffer_head **tl_bh)
5598 struct inode *inode = NULL;
5599 struct buffer_head *bh = NULL;
5601 inode = ocfs2_get_system_file_inode(osb,
5602 TRUNCATE_LOG_SYSTEM_INODE,
5606 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
5610 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
5611 OCFS2_BH_CACHED, inode);
5625 /* called during the 1st stage of node recovery. we stamp a clean
5626 * truncate log and pass back a copy for processing later. if the
5627 * truncate log does not require processing, a *tl_copy is set to
5629 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
5631 struct ocfs2_dinode **tl_copy)
5634 struct inode *tl_inode = NULL;
5635 struct buffer_head *tl_bh = NULL;
5636 struct ocfs2_dinode *di;
5637 struct ocfs2_truncate_log *tl;
5641 mlog(0, "recover truncate log from slot %d\n", slot_num);
5643 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
5649 di = (struct ocfs2_dinode *) tl_bh->b_data;
5650 tl = &di->id2.i_dealloc;
5651 if (!OCFS2_IS_VALID_DINODE(di)) {
5652 OCFS2_RO_ON_INVALID_DINODE(tl_inode->i_sb, di);
5657 if (le16_to_cpu(tl->tl_used)) {
5658 mlog(0, "We'll have %u logs to recover\n",
5659 le16_to_cpu(tl->tl_used));
5661 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
5668 /* Assuming the write-out below goes well, this copy
5669 * will be passed back to recovery for processing. */
5670 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
5672 /* All we need to do to clear the truncate log is set
5676 status = ocfs2_write_block(osb, tl_bh, tl_inode);
5689 if (status < 0 && (*tl_copy)) {
5698 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
5699 struct ocfs2_dinode *tl_copy)
5703 unsigned int clusters, num_recs, start_cluster;
5706 struct inode *tl_inode = osb->osb_tl_inode;
5707 struct ocfs2_truncate_log *tl;
5711 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
5712 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
5716 tl = &tl_copy->id2.i_dealloc;
5717 num_recs = le16_to_cpu(tl->tl_used);
5718 mlog(0, "cleanup %u records from %llu\n", num_recs,
5719 (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
5721 mutex_lock(&tl_inode->i_mutex);
5722 for(i = 0; i < num_recs; i++) {
5723 if (ocfs2_truncate_log_needs_flush(osb)) {
5724 status = __ocfs2_flush_truncate_log(osb);
5731 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5732 if (IS_ERR(handle)) {
5733 status = PTR_ERR(handle);
5738 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
5739 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
5740 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
5742 status = ocfs2_truncate_log_append(osb, handle,
5743 start_blk, clusters);
5744 ocfs2_commit_trans(osb, handle);
5752 mutex_unlock(&tl_inode->i_mutex);
5758 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
5761 struct inode *tl_inode = osb->osb_tl_inode;
5766 cancel_delayed_work(&osb->osb_truncate_log_wq);
5767 flush_workqueue(ocfs2_wq);
5769 status = ocfs2_flush_truncate_log(osb);
5773 brelse(osb->osb_tl_bh);
5774 iput(osb->osb_tl_inode);
5780 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
5783 struct inode *tl_inode = NULL;
5784 struct buffer_head *tl_bh = NULL;
5788 status = ocfs2_get_truncate_log_info(osb,
5795 /* ocfs2_truncate_log_shutdown keys on the existence of
5796 * osb->osb_tl_inode so we don't set any of the osb variables
5797 * until we're sure all is well. */
5798 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
5799 ocfs2_truncate_log_worker);
5800 osb->osb_tl_bh = tl_bh;
5801 osb->osb_tl_inode = tl_inode;
5808 * Delayed de-allocation of suballocator blocks.
5810 * Some sets of block de-allocations might involve multiple suballocator inodes.
5812 * The locking for this can get extremely complicated, especially when
5813 * the suballocator inodes to delete from aren't known until deep
5814 * within an unrelated codepath.
5816 * ocfs2_extent_block structures are a good example of this - an inode
5817 * btree could have been grown by any number of nodes each allocating
5818 * out of their own suballoc inode.
5820 * These structures allow the delay of block de-allocation until a
5821 * later time, when locking of multiple cluster inodes won't cause
5826 * Describes a single block free from a suballocator
5828 struct ocfs2_cached_block_free {
5829 struct ocfs2_cached_block_free *free_next;
5831 unsigned int free_bit;
5834 struct ocfs2_per_slot_free_list {
5835 struct ocfs2_per_slot_free_list *f_next_suballocator;
5838 struct ocfs2_cached_block_free *f_first;
5841 static int ocfs2_free_cached_items(struct ocfs2_super *osb,
5844 struct ocfs2_cached_block_free *head)
5849 struct inode *inode;
5850 struct buffer_head *di_bh = NULL;
5851 struct ocfs2_cached_block_free *tmp;
5853 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
5860 mutex_lock(&inode->i_mutex);
5862 ret = ocfs2_inode_lock(inode, &di_bh, 1);
5868 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
5869 if (IS_ERR(handle)) {
5870 ret = PTR_ERR(handle);
5876 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
5878 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
5879 head->free_bit, (unsigned long long)head->free_blk);
5881 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
5882 head->free_bit, bg_blkno, 1);
5888 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
5895 head = head->free_next;
5900 ocfs2_commit_trans(osb, handle);
5903 ocfs2_inode_unlock(inode, 1);
5906 mutex_unlock(&inode->i_mutex);
5910 /* Premature exit may have left some dangling items. */
5912 head = head->free_next;
5919 int ocfs2_run_deallocs(struct ocfs2_super *osb,
5920 struct ocfs2_cached_dealloc_ctxt *ctxt)
5923 struct ocfs2_per_slot_free_list *fl;
5928 while (ctxt->c_first_suballocator) {
5929 fl = ctxt->c_first_suballocator;
5932 mlog(0, "Free items: (type %u, slot %d)\n",
5933 fl->f_inode_type, fl->f_slot);
5934 ret2 = ocfs2_free_cached_items(osb, fl->f_inode_type,
5935 fl->f_slot, fl->f_first);
5942 ctxt->c_first_suballocator = fl->f_next_suballocator;
5949 static struct ocfs2_per_slot_free_list *
5950 ocfs2_find_per_slot_free_list(int type,
5952 struct ocfs2_cached_dealloc_ctxt *ctxt)
5954 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
5957 if (fl->f_inode_type == type && fl->f_slot == slot)
5960 fl = fl->f_next_suballocator;
5963 fl = kmalloc(sizeof(*fl), GFP_NOFS);
5965 fl->f_inode_type = type;
5968 fl->f_next_suballocator = ctxt->c_first_suballocator;
5970 ctxt->c_first_suballocator = fl;
5975 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
5976 int type, int slot, u64 blkno,
5980 struct ocfs2_per_slot_free_list *fl;
5981 struct ocfs2_cached_block_free *item;
5983 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
5990 item = kmalloc(sizeof(*item), GFP_NOFS);
5997 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
5998 type, slot, bit, (unsigned long long)blkno);
6000 item->free_blk = blkno;
6001 item->free_bit = bit;
6002 item->free_next = fl->f_first;
6011 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6012 struct ocfs2_extent_block *eb)
6014 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6015 le16_to_cpu(eb->h_suballoc_slot),
6016 le64_to_cpu(eb->h_blkno),
6017 le16_to_cpu(eb->h_suballoc_bit));
6020 /* This function will figure out whether the currently last extent
6021 * block will be deleted, and if it will, what the new last extent
6022 * block will be so we can update his h_next_leaf_blk field, as well
6023 * as the dinodes i_last_eb_blk */
6024 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
6025 unsigned int clusters_to_del,
6026 struct ocfs2_path *path,
6027 struct buffer_head **new_last_eb)
6029 int next_free, ret = 0;
6031 struct ocfs2_extent_rec *rec;
6032 struct ocfs2_extent_block *eb;
6033 struct ocfs2_extent_list *el;
6034 struct buffer_head *bh = NULL;
6036 *new_last_eb = NULL;
6038 /* we have no tree, so of course, no last_eb. */
6039 if (!path->p_tree_depth)
6042 /* trunc to zero special case - this makes tree_depth = 0
6043 * regardless of what it is. */
6044 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
6047 el = path_leaf_el(path);
6048 BUG_ON(!el->l_next_free_rec);
6051 * Make sure that this extent list will actually be empty
6052 * after we clear away the data. We can shortcut out if
6053 * there's more than one non-empty extent in the
6054 * list. Otherwise, a check of the remaining extent is
6057 next_free = le16_to_cpu(el->l_next_free_rec);
6059 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6063 /* We may have a valid extent in index 1, check it. */
6065 rec = &el->l_recs[1];
6068 * Fall through - no more nonempty extents, so we want
6069 * to delete this leaf.
6075 rec = &el->l_recs[0];
6080 * Check it we'll only be trimming off the end of this
6083 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
6087 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
6093 ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
6099 eb = (struct ocfs2_extent_block *) bh->b_data;
6101 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
6102 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
6108 get_bh(*new_last_eb);
6109 mlog(0, "returning block %llu, (cpos: %u)\n",
6110 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
6118 * Trim some clusters off the rightmost edge of a tree. Only called
6121 * The caller needs to:
6122 * - start journaling of each path component.
6123 * - compute and fully set up any new last ext block
6125 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
6126 handle_t *handle, struct ocfs2_truncate_context *tc,
6127 u32 clusters_to_del, u64 *delete_start)
6129 int ret, i, index = path->p_tree_depth;
6132 struct buffer_head *bh;
6133 struct ocfs2_extent_list *el;
6134 struct ocfs2_extent_rec *rec;
6138 while (index >= 0) {
6139 bh = path->p_node[index].bh;
6140 el = path->p_node[index].el;
6142 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6143 index, (unsigned long long)bh->b_blocknr);
6145 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
6148 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
6149 ocfs2_error(inode->i_sb,
6150 "Inode %lu has invalid ext. block %llu",
6152 (unsigned long long)bh->b_blocknr);
6158 i = le16_to_cpu(el->l_next_free_rec) - 1;
6159 rec = &el->l_recs[i];
6161 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6162 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
6163 ocfs2_rec_clusters(el, rec),
6164 (unsigned long long)le64_to_cpu(rec->e_blkno),
6165 le16_to_cpu(el->l_next_free_rec));
6167 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
6169 if (le16_to_cpu(el->l_tree_depth) == 0) {
6171 * If the leaf block contains a single empty
6172 * extent and no records, we can just remove
6175 if (i == 0 && ocfs2_is_empty_extent(rec)) {
6177 sizeof(struct ocfs2_extent_rec));
6178 el->l_next_free_rec = cpu_to_le16(0);
6184 * Remove any empty extents by shifting things
6185 * left. That should make life much easier on
6186 * the code below. This condition is rare
6187 * enough that we shouldn't see a performance
6190 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6191 le16_add_cpu(&el->l_next_free_rec, -1);
6194 i < le16_to_cpu(el->l_next_free_rec); i++)
6195 el->l_recs[i] = el->l_recs[i + 1];
6197 memset(&el->l_recs[i], 0,
6198 sizeof(struct ocfs2_extent_rec));
6201 * We've modified our extent list. The
6202 * simplest way to handle this change
6203 * is to being the search from the
6206 goto find_tail_record;
6209 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
6212 * We'll use "new_edge" on our way back up the
6213 * tree to know what our rightmost cpos is.
6215 new_edge = le16_to_cpu(rec->e_leaf_clusters);
6216 new_edge += le32_to_cpu(rec->e_cpos);
6219 * The caller will use this to delete data blocks.
6221 *delete_start = le64_to_cpu(rec->e_blkno)
6222 + ocfs2_clusters_to_blocks(inode->i_sb,
6223 le16_to_cpu(rec->e_leaf_clusters));
6226 * If it's now empty, remove this record.
6228 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
6230 sizeof(struct ocfs2_extent_rec));
6231 le16_add_cpu(&el->l_next_free_rec, -1);
6234 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
6236 sizeof(struct ocfs2_extent_rec));
6237 le16_add_cpu(&el->l_next_free_rec, -1);
6242 /* Can this actually happen? */
6243 if (le16_to_cpu(el->l_next_free_rec) == 0)
6247 * We never actually deleted any clusters
6248 * because our leaf was empty. There's no
6249 * reason to adjust the rightmost edge then.
6254 rec->e_int_clusters = cpu_to_le32(new_edge);
6255 le32_add_cpu(&rec->e_int_clusters,
6256 -le32_to_cpu(rec->e_cpos));
6259 * A deleted child record should have been
6262 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
6266 ret = ocfs2_journal_dirty(handle, bh);
6272 mlog(0, "extent list container %llu, after: record %d: "
6273 "(%u, %u, %llu), next = %u.\n",
6274 (unsigned long long)bh->b_blocknr, i,
6275 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
6276 (unsigned long long)le64_to_cpu(rec->e_blkno),
6277 le16_to_cpu(el->l_next_free_rec));
6280 * We must be careful to only attempt delete of an
6281 * extent block (and not the root inode block).
6283 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
6284 struct ocfs2_extent_block *eb =
6285 (struct ocfs2_extent_block *)bh->b_data;
6288 * Save this for use when processing the
6291 deleted_eb = le64_to_cpu(eb->h_blkno);
6293 mlog(0, "deleting this extent block.\n");
6295 ocfs2_remove_from_cache(inode, bh);
6297 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
6298 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
6299 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
6301 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
6302 /* An error here is not fatal. */
6317 static int ocfs2_do_truncate(struct ocfs2_super *osb,
6318 unsigned int clusters_to_del,
6319 struct inode *inode,
6320 struct buffer_head *fe_bh,
6322 struct ocfs2_truncate_context *tc,
6323 struct ocfs2_path *path)
6326 struct ocfs2_dinode *fe;
6327 struct ocfs2_extent_block *last_eb = NULL;
6328 struct ocfs2_extent_list *el;
6329 struct buffer_head *last_eb_bh = NULL;
6332 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6334 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
6342 * Each component will be touched, so we might as well journal
6343 * here to avoid having to handle errors later.
6345 status = ocfs2_journal_access_path(inode, handle, path);
6352 status = ocfs2_journal_access(handle, inode, last_eb_bh,
6353 OCFS2_JOURNAL_ACCESS_WRITE);
6359 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6362 el = &(fe->id2.i_list);
6365 * Lower levels depend on this never happening, but it's best
6366 * to check it up here before changing the tree.
6368 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
6369 ocfs2_error(inode->i_sb,
6370 "Inode %lu has an empty extent record, depth %u\n",
6371 inode->i_ino, le16_to_cpu(el->l_tree_depth));
6376 spin_lock(&OCFS2_I(inode)->ip_lock);
6377 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
6379 spin_unlock(&OCFS2_I(inode)->ip_lock);
6380 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
6381 inode->i_blocks = ocfs2_inode_sector_count(inode);
6383 status = ocfs2_trim_tree(inode, path, handle, tc,
6384 clusters_to_del, &delete_blk);
6390 if (le32_to_cpu(fe->i_clusters) == 0) {
6391 /* trunc to zero is a special case. */
6392 el->l_tree_depth = 0;
6393 fe->i_last_eb_blk = 0;
6395 fe->i_last_eb_blk = last_eb->h_blkno;
6397 status = ocfs2_journal_dirty(handle, fe_bh);
6404 /* If there will be a new last extent block, then by
6405 * definition, there cannot be any leaves to the right of
6407 last_eb->h_next_leaf_blk = 0;
6408 status = ocfs2_journal_dirty(handle, last_eb_bh);
6416 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
6430 static int ocfs2_writeback_zero_func(handle_t *handle, struct buffer_head *bh)
6432 set_buffer_uptodate(bh);
6433 mark_buffer_dirty(bh);
6437 static int ocfs2_ordered_zero_func(handle_t *handle, struct buffer_head *bh)
6439 set_buffer_uptodate(bh);
6440 mark_buffer_dirty(bh);
6441 return ocfs2_journal_dirty_data(handle, bh);
6444 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6445 unsigned int from, unsigned int to,
6446 struct page *page, int zero, u64 *phys)
6448 int ret, partial = 0;
6450 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6455 zero_user_segment(page, from, to);
6458 * Need to set the buffers we zero'd into uptodate
6459 * here if they aren't - ocfs2_map_page_blocks()
6460 * might've skipped some
6462 if (ocfs2_should_order_data(inode)) {
6463 ret = walk_page_buffers(handle,
6466 ocfs2_ordered_zero_func);
6470 ret = walk_page_buffers(handle, page_buffers(page),
6472 ocfs2_writeback_zero_func);
6478 SetPageUptodate(page);
6480 flush_dcache_page(page);
6483 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6484 loff_t end, struct page **pages,
6485 int numpages, u64 phys, handle_t *handle)
6489 unsigned int from, to = PAGE_CACHE_SIZE;
6490 struct super_block *sb = inode->i_sb;
6492 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6497 to = PAGE_CACHE_SIZE;
6498 for(i = 0; i < numpages; i++) {
6501 from = start & (PAGE_CACHE_SIZE - 1);
6502 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6503 to = end & (PAGE_CACHE_SIZE - 1);
6505 BUG_ON(from > PAGE_CACHE_SIZE);
6506 BUG_ON(to > PAGE_CACHE_SIZE);
6508 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6511 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6515 ocfs2_unlock_and_free_pages(pages, numpages);
6518 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6519 struct page **pages, int *num)
6521 int numpages, ret = 0;
6522 struct super_block *sb = inode->i_sb;
6523 struct address_space *mapping = inode->i_mapping;
6524 unsigned long index;
6525 loff_t last_page_bytes;
6527 BUG_ON(start > end);
6529 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6530 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6533 last_page_bytes = PAGE_ALIGN(end);
6534 index = start >> PAGE_CACHE_SHIFT;
6536 pages[numpages] = grab_cache_page(mapping, index);
6537 if (!pages[numpages]) {
6545 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6550 ocfs2_unlock_and_free_pages(pages, numpages);
6560 * Zero the area past i_size but still within an allocated
6561 * cluster. This avoids exposing nonzero data on subsequent file
6564 * We need to call this before i_size is updated on the inode because
6565 * otherwise block_write_full_page() will skip writeout of pages past
6566 * i_size. The new_i_size parameter is passed for this reason.
6568 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6569 u64 range_start, u64 range_end)
6571 int ret = 0, numpages;
6572 struct page **pages = NULL;
6574 unsigned int ext_flags;
6575 struct super_block *sb = inode->i_sb;
6578 * File systems which don't support sparse files zero on every
6581 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6584 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6585 sizeof(struct page *), GFP_NOFS);
6586 if (pages == NULL) {
6592 if (range_start == range_end)
6595 ret = ocfs2_extent_map_get_blocks(inode,
6596 range_start >> sb->s_blocksize_bits,
6597 &phys, NULL, &ext_flags);
6604 * Tail is a hole, or is marked unwritten. In either case, we
6605 * can count on read and write to return/push zero's.
6607 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6610 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6617 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6618 numpages, phys, handle);
6621 * Initiate writeout of the pages we zero'd here. We don't
6622 * wait on them - the truncate_inode_pages() call later will
6625 ret = do_sync_mapping_range(inode->i_mapping, range_start,
6626 range_end - 1, SYNC_FILE_RANGE_WRITE);
6637 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6638 struct ocfs2_dinode *di)
6640 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6641 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6643 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6644 memset(&di->id2, 0, blocksize -
6645 offsetof(struct ocfs2_dinode, id2) -
6648 memset(&di->id2, 0, blocksize -
6649 offsetof(struct ocfs2_dinode, id2));
6652 void ocfs2_dinode_new_extent_list(struct inode *inode,
6653 struct ocfs2_dinode *di)
6655 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6656 di->id2.i_list.l_tree_depth = 0;
6657 di->id2.i_list.l_next_free_rec = 0;
6658 di->id2.i_list.l_count = cpu_to_le16(
6659 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6662 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6664 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6665 struct ocfs2_inline_data *idata = &di->id2.i_data;
6667 spin_lock(&oi->ip_lock);
6668 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6669 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6670 spin_unlock(&oi->ip_lock);
6673 * We clear the entire i_data structure here so that all
6674 * fields can be properly initialized.
6676 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6678 idata->id_count = cpu_to_le16(
6679 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6682 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6683 struct buffer_head *di_bh)
6685 int ret, i, has_data, num_pages = 0;
6687 u64 uninitialized_var(block);
6688 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6689 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6690 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6691 struct ocfs2_alloc_context *data_ac = NULL;
6692 struct page **pages = NULL;
6693 loff_t end = osb->s_clustersize;
6694 struct ocfs2_extent_tree et;
6696 has_data = i_size_read(inode) ? 1 : 0;
6699 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6700 sizeof(struct page *), GFP_NOFS);
6701 if (pages == NULL) {
6707 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6714 handle = ocfs2_start_trans(osb, OCFS2_INLINE_TO_EXTENTS_CREDITS);
6715 if (IS_ERR(handle)) {
6716 ret = PTR_ERR(handle);
6721 ret = ocfs2_journal_access(handle, inode, di_bh,
6722 OCFS2_JOURNAL_ACCESS_WRITE);
6730 unsigned int page_end;
6733 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
6741 * Save two copies, one for insert, and one that can
6742 * be changed by ocfs2_map_and_dirty_page() below.
6744 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6747 * Non sparse file systems zero on extend, so no need
6750 if (!ocfs2_sparse_alloc(osb) &&
6751 PAGE_CACHE_SIZE < osb->s_clustersize)
6752 end = PAGE_CACHE_SIZE;
6754 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6761 * This should populate the 1st page for us and mark
6764 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6770 page_end = PAGE_CACHE_SIZE;
6771 if (PAGE_CACHE_SIZE > osb->s_clustersize)
6772 page_end = osb->s_clustersize;
6774 for (i = 0; i < num_pages; i++)
6775 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6776 pages[i], i > 0, &phys);
6779 spin_lock(&oi->ip_lock);
6780 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6781 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6782 spin_unlock(&oi->ip_lock);
6784 ocfs2_dinode_new_extent_list(inode, di);
6786 ocfs2_journal_dirty(handle, di_bh);
6790 * An error at this point should be extremely rare. If
6791 * this proves to be false, we could always re-build
6792 * the in-inode data from our pages.
6794 ocfs2_get_dinode_extent_tree(&et, inode, di_bh);
6795 ret = ocfs2_insert_extent(osb, handle, inode, &et,
6796 0, block, 1, 0, NULL);
6797 ocfs2_put_extent_tree(&et);
6803 inode->i_blocks = ocfs2_inode_sector_count(inode);
6807 ocfs2_commit_trans(osb, handle);
6811 ocfs2_free_alloc_context(data_ac);
6815 ocfs2_unlock_and_free_pages(pages, num_pages);
6823 * It is expected, that by the time you call this function,
6824 * inode->i_size and fe->i_size have been adjusted.
6826 * WARNING: This will kfree the truncate context
6828 int ocfs2_commit_truncate(struct ocfs2_super *osb,
6829 struct inode *inode,
6830 struct buffer_head *fe_bh,
6831 struct ocfs2_truncate_context *tc)
6833 int status, i, credits, tl_sem = 0;
6834 u32 clusters_to_del, new_highest_cpos, range;
6835 struct ocfs2_extent_list *el;
6836 handle_t *handle = NULL;
6837 struct inode *tl_inode = osb->osb_tl_inode;
6838 struct ocfs2_path *path = NULL;
6839 struct ocfs2_dinode *di = (struct ocfs2_dinode *)fe_bh->b_data;
6843 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
6844 i_size_read(inode));
6846 path = ocfs2_new_path(fe_bh, &di->id2.i_list);
6853 ocfs2_extent_map_trunc(inode, new_highest_cpos);
6857 * Check that we still have allocation to delete.
6859 if (OCFS2_I(inode)->ip_clusters == 0) {
6865 * Truncate always works against the rightmost tree branch.
6867 status = ocfs2_find_path(inode, path, UINT_MAX);
6873 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
6874 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
6877 * By now, el will point to the extent list on the bottom most
6878 * portion of this tree. Only the tail record is considered in
6881 * We handle the following cases, in order:
6882 * - empty extent: delete the remaining branch
6883 * - remove the entire record
6884 * - remove a partial record
6885 * - no record needs to be removed (truncate has completed)
6887 el = path_leaf_el(path);
6888 if (le16_to_cpu(el->l_next_free_rec) == 0) {
6889 ocfs2_error(inode->i_sb,
6890 "Inode %llu has empty extent block at %llu\n",
6891 (unsigned long long)OCFS2_I(inode)->ip_blkno,
6892 (unsigned long long)path_leaf_bh(path)->b_blocknr);
6897 i = le16_to_cpu(el->l_next_free_rec) - 1;
6898 range = le32_to_cpu(el->l_recs[i].e_cpos) +
6899 ocfs2_rec_clusters(el, &el->l_recs[i]);
6900 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
6901 clusters_to_del = 0;
6902 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
6903 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
6904 } else if (range > new_highest_cpos) {
6905 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
6906 le32_to_cpu(el->l_recs[i].e_cpos)) -
6913 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
6914 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
6916 mutex_lock(&tl_inode->i_mutex);
6918 /* ocfs2_truncate_log_needs_flush guarantees us at least one
6919 * record is free for use. If there isn't any, we flush to get
6920 * an empty truncate log. */
6921 if (ocfs2_truncate_log_needs_flush(osb)) {
6922 status = __ocfs2_flush_truncate_log(osb);
6929 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
6930 (struct ocfs2_dinode *)fe_bh->b_data,
6932 handle = ocfs2_start_trans(osb, credits);
6933 if (IS_ERR(handle)) {
6934 status = PTR_ERR(handle);
6940 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
6947 mutex_unlock(&tl_inode->i_mutex);
6950 ocfs2_commit_trans(osb, handle);
6953 ocfs2_reinit_path(path, 1);
6956 * The check above will catch the case where we've truncated
6957 * away all allocation.
6963 ocfs2_schedule_truncate_log_flush(osb, 1);
6966 mutex_unlock(&tl_inode->i_mutex);
6969 ocfs2_commit_trans(osb, handle);
6971 ocfs2_run_deallocs(osb, &tc->tc_dealloc);
6973 ocfs2_free_path(path);
6975 /* This will drop the ext_alloc cluster lock for us */
6976 ocfs2_free_truncate_context(tc);
6983 * Expects the inode to already be locked.
6985 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
6986 struct inode *inode,
6987 struct buffer_head *fe_bh,
6988 struct ocfs2_truncate_context **tc)
6991 unsigned int new_i_clusters;
6992 struct ocfs2_dinode *fe;
6993 struct ocfs2_extent_block *eb;
6994 struct buffer_head *last_eb_bh = NULL;
7000 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
7001 i_size_read(inode));
7002 fe = (struct ocfs2_dinode *) fe_bh->b_data;
7004 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
7005 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
7006 (unsigned long long)le64_to_cpu(fe->i_size));
7008 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
7014 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
7016 if (fe->id2.i_list.l_tree_depth) {
7017 status = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
7018 &last_eb_bh, OCFS2_BH_CACHED, inode);
7023 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
7024 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
7025 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
7033 (*tc)->tc_last_eb_bh = last_eb_bh;
7039 ocfs2_free_truncate_context(*tc);
7047 * 'start' is inclusive, 'end' is not.
7049 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7050 unsigned int start, unsigned int end, int trunc)
7053 unsigned int numbytes;
7055 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7056 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7057 struct ocfs2_inline_data *idata = &di->id2.i_data;
7059 if (end > i_size_read(inode))
7060 end = i_size_read(inode);
7062 BUG_ON(start >= end);
7064 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7065 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7066 !ocfs2_supports_inline_data(osb)) {
7067 ocfs2_error(inode->i_sb,
7068 "Inline data flags for inode %llu don't agree! "
7069 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7070 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7071 le16_to_cpu(di->i_dyn_features),
7072 OCFS2_I(inode)->ip_dyn_features,
7073 osb->s_feature_incompat);
7078 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7079 if (IS_ERR(handle)) {
7080 ret = PTR_ERR(handle);
7085 ret = ocfs2_journal_access(handle, inode, di_bh,
7086 OCFS2_JOURNAL_ACCESS_WRITE);
7092 numbytes = end - start;
7093 memset(idata->id_data + start, 0, numbytes);
7096 * No need to worry about the data page here - it's been
7097 * truncated already and inline data doesn't need it for
7098 * pushing zero's to disk, so we'll let readpage pick it up
7102 i_size_write(inode, start);
7103 di->i_size = cpu_to_le64(start);
7106 inode->i_blocks = ocfs2_inode_sector_count(inode);
7107 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7109 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7110 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7112 ocfs2_journal_dirty(handle, di_bh);
7115 ocfs2_commit_trans(osb, handle);
7121 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
7124 * The caller is responsible for completing deallocation
7125 * before freeing the context.
7127 if (tc->tc_dealloc.c_first_suballocator != NULL)
7129 "Truncate completion has non-empty dealloc context\n");
7131 if (tc->tc_last_eb_bh)
7132 brelse(tc->tc_last_eb_bh);