1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Extent allocs and frees
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
32 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
33 #include <cluster/masklog.h>
40 #include "extent_map.h"
43 #include "localalloc.h"
50 #include "buffer_head_io.h"
52 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
53 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
54 struct ocfs2_extent_block *eb);
57 * Structures which describe a path through a btree, and functions to
60 * The idea here is to be as generic as possible with the tree
63 struct ocfs2_path_item {
64 struct buffer_head *bh;
65 struct ocfs2_extent_list *el;
68 #define OCFS2_MAX_PATH_DEPTH 5
72 struct ocfs2_path_item p_node[OCFS2_MAX_PATH_DEPTH];
75 #define path_root_bh(_path) ((_path)->p_node[0].bh)
76 #define path_root_el(_path) ((_path)->p_node[0].el)
77 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
78 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
79 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
82 * Reset the actual path elements so that we can re-use the structure
83 * to build another path. Generally, this involves freeing the buffer
86 static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
88 int i, start = 0, depth = 0;
89 struct ocfs2_path_item *node;
94 for(i = start; i < path_num_items(path); i++) {
95 node = &path->p_node[i];
103 * Tree depth may change during truncate, or insert. If we're
104 * keeping the root extent list, then make sure that our path
105 * structure reflects the proper depth.
108 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
110 path->p_tree_depth = depth;
113 static void ocfs2_free_path(struct ocfs2_path *path)
116 ocfs2_reinit_path(path, 0);
122 * All the elements of src into dest. After this call, src could be freed
123 * without affecting dest.
125 * Both paths should have the same root. Any non-root elements of dest
128 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
132 BUG_ON(path_root_bh(dest) != path_root_bh(src));
133 BUG_ON(path_root_el(dest) != path_root_el(src));
135 ocfs2_reinit_path(dest, 1);
137 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
138 dest->p_node[i].bh = src->p_node[i].bh;
139 dest->p_node[i].el = src->p_node[i].el;
141 if (dest->p_node[i].bh)
142 get_bh(dest->p_node[i].bh);
147 * Make the *dest path the same as src and re-initialize src path to
150 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
154 BUG_ON(path_root_bh(dest) != path_root_bh(src));
156 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
157 brelse(dest->p_node[i].bh);
159 dest->p_node[i].bh = src->p_node[i].bh;
160 dest->p_node[i].el = src->p_node[i].el;
162 src->p_node[i].bh = NULL;
163 src->p_node[i].el = NULL;
168 * Insert an extent block at given index.
170 * This will not take an additional reference on eb_bh.
172 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
173 struct buffer_head *eb_bh)
175 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
178 * Right now, no root bh is an extent block, so this helps
179 * catch code errors with dinode trees. The assertion can be
180 * safely removed if we ever need to insert extent block
181 * structures at the root.
185 path->p_node[index].bh = eb_bh;
186 path->p_node[index].el = &eb->h_list;
189 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
190 struct ocfs2_extent_list *root_el)
192 struct ocfs2_path *path;
194 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
196 path = kzalloc(sizeof(*path), GFP_NOFS);
198 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
200 path_root_bh(path) = root_bh;
201 path_root_el(path) = root_el;
208 * Allocate and initialize a new path based on a disk inode tree.
210 static struct ocfs2_path *ocfs2_new_inode_path(struct buffer_head *di_bh)
212 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
213 struct ocfs2_extent_list *el = &di->id2.i_list;
215 return ocfs2_new_path(di_bh, el);
219 * Convenience function to journal all components in a path.
221 static int ocfs2_journal_access_path(struct inode *inode, handle_t *handle,
222 struct ocfs2_path *path)
229 for(i = 0; i < path_num_items(path); i++) {
230 ret = ocfs2_journal_access(handle, inode, path->p_node[i].bh,
231 OCFS2_JOURNAL_ACCESS_WRITE);
243 * Return the index of the extent record which contains cluster #v_cluster.
244 * -1 is returned if it was not found.
246 * Should work fine on interior and exterior nodes.
248 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
252 struct ocfs2_extent_rec *rec;
253 u32 rec_end, rec_start, clusters;
255 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
256 rec = &el->l_recs[i];
258 rec_start = le32_to_cpu(rec->e_cpos);
259 clusters = ocfs2_rec_clusters(el, rec);
261 rec_end = rec_start + clusters;
263 if (v_cluster >= rec_start && v_cluster < rec_end) {
272 enum ocfs2_contig_type {
281 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
282 * ocfs2_extent_contig only work properly against leaf nodes!
284 static int ocfs2_block_extent_contig(struct super_block *sb,
285 struct ocfs2_extent_rec *ext,
288 u64 blk_end = le64_to_cpu(ext->e_blkno);
290 blk_end += ocfs2_clusters_to_blocks(sb,
291 le16_to_cpu(ext->e_leaf_clusters));
293 return blkno == blk_end;
296 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
297 struct ocfs2_extent_rec *right)
301 left_range = le32_to_cpu(left->e_cpos) +
302 le16_to_cpu(left->e_leaf_clusters);
304 return (left_range == le32_to_cpu(right->e_cpos));
307 static enum ocfs2_contig_type
308 ocfs2_extent_contig(struct inode *inode,
309 struct ocfs2_extent_rec *ext,
310 struct ocfs2_extent_rec *insert_rec)
312 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
315 * Refuse to coalesce extent records with different flag
316 * fields - we don't want to mix unwritten extents with user
319 if (ext->e_flags != insert_rec->e_flags)
322 if (ocfs2_extents_adjacent(ext, insert_rec) &&
323 ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
326 blkno = le64_to_cpu(ext->e_blkno);
327 if (ocfs2_extents_adjacent(insert_rec, ext) &&
328 ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno))
335 * NOTE: We can have pretty much any combination of contiguousness and
338 * The usefulness of APPEND_TAIL is more in that it lets us know that
339 * we'll have to update the path to that leaf.
341 enum ocfs2_append_type {
346 enum ocfs2_split_type {
352 struct ocfs2_insert_type {
353 enum ocfs2_split_type ins_split;
354 enum ocfs2_append_type ins_appending;
355 enum ocfs2_contig_type ins_contig;
356 int ins_contig_index;
360 struct ocfs2_merge_ctxt {
361 enum ocfs2_contig_type c_contig_type;
362 int c_has_empty_extent;
363 int c_split_covers_rec;
364 int c_used_tail_recs;
368 * How many free extents have we got before we need more meta data?
370 int ocfs2_num_free_extents(struct ocfs2_super *osb,
372 struct ocfs2_dinode *fe)
375 struct ocfs2_extent_list *el;
376 struct ocfs2_extent_block *eb;
377 struct buffer_head *eb_bh = NULL;
381 if (!OCFS2_IS_VALID_DINODE(fe)) {
382 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
387 if (fe->i_last_eb_blk) {
388 retval = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
389 &eb_bh, OCFS2_BH_CACHED, inode);
394 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
397 el = &fe->id2.i_list;
399 BUG_ON(el->l_tree_depth != 0);
401 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
410 /* expects array to already be allocated
412 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
415 static int ocfs2_create_new_meta_bhs(struct ocfs2_super *osb,
419 struct ocfs2_alloc_context *meta_ac,
420 struct buffer_head *bhs[])
422 int count, status, i;
423 u16 suballoc_bit_start;
426 struct ocfs2_extent_block *eb;
431 while (count < wanted) {
432 status = ocfs2_claim_metadata(osb,
444 for(i = count; i < (num_got + count); i++) {
445 bhs[i] = sb_getblk(osb->sb, first_blkno);
446 if (bhs[i] == NULL) {
451 ocfs2_set_new_buffer_uptodate(inode, bhs[i]);
453 status = ocfs2_journal_access(handle, inode, bhs[i],
454 OCFS2_JOURNAL_ACCESS_CREATE);
460 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
461 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
462 /* Ok, setup the minimal stuff here. */
463 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
464 eb->h_blkno = cpu_to_le64(first_blkno);
465 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
466 eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
467 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
469 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
471 suballoc_bit_start++;
474 /* We'll also be dirtied by the caller, so
475 * this isn't absolutely necessary. */
476 status = ocfs2_journal_dirty(handle, bhs[i]);
489 for(i = 0; i < wanted; i++) {
500 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
502 * Returns the sum of the rightmost extent rec logical offset and
505 * ocfs2_add_branch() uses this to determine what logical cluster
506 * value should be populated into the leftmost new branch records.
508 * ocfs2_shift_tree_depth() uses this to determine the # clusters
509 * value for the new topmost tree record.
511 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
515 i = le16_to_cpu(el->l_next_free_rec) - 1;
517 return le32_to_cpu(el->l_recs[i].e_cpos) +
518 ocfs2_rec_clusters(el, &el->l_recs[i]);
522 * Add an entire tree branch to our inode. eb_bh is the extent block
523 * to start at, if we don't want to start the branch at the dinode
526 * last_eb_bh is required as we have to update it's next_leaf pointer
527 * for the new last extent block.
529 * the new branch will be 'empty' in the sense that every block will
530 * contain a single record with cluster count == 0.
532 static int ocfs2_add_branch(struct ocfs2_super *osb,
535 struct buffer_head *fe_bh,
536 struct buffer_head *eb_bh,
537 struct buffer_head **last_eb_bh,
538 struct ocfs2_alloc_context *meta_ac)
540 int status, new_blocks, i;
541 u64 next_blkno, new_last_eb_blk;
542 struct buffer_head *bh;
543 struct buffer_head **new_eb_bhs = NULL;
544 struct ocfs2_dinode *fe;
545 struct ocfs2_extent_block *eb;
546 struct ocfs2_extent_list *eb_el;
547 struct ocfs2_extent_list *el;
552 BUG_ON(!last_eb_bh || !*last_eb_bh);
554 fe = (struct ocfs2_dinode *) fe_bh->b_data;
557 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
560 el = &fe->id2.i_list;
562 /* we never add a branch to a leaf. */
563 BUG_ON(!el->l_tree_depth);
565 new_blocks = le16_to_cpu(el->l_tree_depth);
567 /* allocate the number of new eb blocks we need */
568 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
576 status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
577 meta_ac, new_eb_bhs);
583 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
584 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
586 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
587 * linked with the rest of the tree.
588 * conversly, new_eb_bhs[0] is the new bottommost leaf.
590 * when we leave the loop, new_last_eb_blk will point to the
591 * newest leaf, and next_blkno will point to the topmost extent
593 next_blkno = new_last_eb_blk = 0;
594 for(i = 0; i < new_blocks; i++) {
596 eb = (struct ocfs2_extent_block *) bh->b_data;
597 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
598 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
604 status = ocfs2_journal_access(handle, inode, bh,
605 OCFS2_JOURNAL_ACCESS_CREATE);
611 eb->h_next_leaf_blk = 0;
612 eb_el->l_tree_depth = cpu_to_le16(i);
613 eb_el->l_next_free_rec = cpu_to_le16(1);
615 * This actually counts as an empty extent as
618 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
619 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
621 * eb_el isn't always an interior node, but even leaf
622 * nodes want a zero'd flags and reserved field so
623 * this gets the whole 32 bits regardless of use.
625 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
626 if (!eb_el->l_tree_depth)
627 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
629 status = ocfs2_journal_dirty(handle, bh);
635 next_blkno = le64_to_cpu(eb->h_blkno);
638 /* This is a bit hairy. We want to update up to three blocks
639 * here without leaving any of them in an inconsistent state
640 * in case of error. We don't have to worry about
641 * journal_dirty erroring as it won't unless we've aborted the
642 * handle (in which case we would never be here) so reserving
643 * the write with journal_access is all we need to do. */
644 status = ocfs2_journal_access(handle, inode, *last_eb_bh,
645 OCFS2_JOURNAL_ACCESS_WRITE);
650 status = ocfs2_journal_access(handle, inode, fe_bh,
651 OCFS2_JOURNAL_ACCESS_WRITE);
657 status = ocfs2_journal_access(handle, inode, eb_bh,
658 OCFS2_JOURNAL_ACCESS_WRITE);
665 /* Link the new branch into the rest of the tree (el will
666 * either be on the fe, or the extent block passed in. */
667 i = le16_to_cpu(el->l_next_free_rec);
668 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
669 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
670 el->l_recs[i].e_int_clusters = 0;
671 le16_add_cpu(&el->l_next_free_rec, 1);
673 /* fe needs a new last extent block pointer, as does the
674 * next_leaf on the previously last-extent-block. */
675 fe->i_last_eb_blk = cpu_to_le64(new_last_eb_blk);
677 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
678 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
680 status = ocfs2_journal_dirty(handle, *last_eb_bh);
683 status = ocfs2_journal_dirty(handle, fe_bh);
687 status = ocfs2_journal_dirty(handle, eb_bh);
693 * Some callers want to track the rightmost leaf so pass it
697 get_bh(new_eb_bhs[0]);
698 *last_eb_bh = new_eb_bhs[0];
703 for (i = 0; i < new_blocks; i++)
705 brelse(new_eb_bhs[i]);
714 * adds another level to the allocation tree.
715 * returns back the new extent block so you can add a branch to it
718 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
721 struct buffer_head *fe_bh,
722 struct ocfs2_alloc_context *meta_ac,
723 struct buffer_head **ret_new_eb_bh)
727 struct buffer_head *new_eb_bh = NULL;
728 struct ocfs2_dinode *fe;
729 struct ocfs2_extent_block *eb;
730 struct ocfs2_extent_list *fe_el;
731 struct ocfs2_extent_list *eb_el;
735 status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
742 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
743 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
744 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
750 fe = (struct ocfs2_dinode *) fe_bh->b_data;
751 fe_el = &fe->id2.i_list;
753 status = ocfs2_journal_access(handle, inode, new_eb_bh,
754 OCFS2_JOURNAL_ACCESS_CREATE);
760 /* copy the fe data into the new extent block */
761 eb_el->l_tree_depth = fe_el->l_tree_depth;
762 eb_el->l_next_free_rec = fe_el->l_next_free_rec;
763 for(i = 0; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
764 eb_el->l_recs[i] = fe_el->l_recs[i];
766 status = ocfs2_journal_dirty(handle, new_eb_bh);
772 status = ocfs2_journal_access(handle, inode, fe_bh,
773 OCFS2_JOURNAL_ACCESS_WRITE);
779 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
782 le16_add_cpu(&fe_el->l_tree_depth, 1);
783 fe_el->l_recs[0].e_cpos = 0;
784 fe_el->l_recs[0].e_blkno = eb->h_blkno;
785 fe_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
786 for(i = 1; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
787 memset(&fe_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
788 fe_el->l_next_free_rec = cpu_to_le16(1);
790 /* If this is our 1st tree depth shift, then last_eb_blk
791 * becomes the allocated extent block */
792 if (fe_el->l_tree_depth == cpu_to_le16(1))
793 fe->i_last_eb_blk = eb->h_blkno;
795 status = ocfs2_journal_dirty(handle, fe_bh);
801 *ret_new_eb_bh = new_eb_bh;
813 * Should only be called when there is no space left in any of the
814 * leaf nodes. What we want to do is find the lowest tree depth
815 * non-leaf extent block with room for new records. There are three
816 * valid results of this search:
818 * 1) a lowest extent block is found, then we pass it back in
819 * *lowest_eb_bh and return '0'
821 * 2) the search fails to find anything, but the dinode has room. We
822 * pass NULL back in *lowest_eb_bh, but still return '0'
824 * 3) the search fails to find anything AND the dinode is full, in
825 * which case we return > 0
827 * return status < 0 indicates an error.
829 static int ocfs2_find_branch_target(struct ocfs2_super *osb,
831 struct buffer_head *fe_bh,
832 struct buffer_head **target_bh)
836 struct ocfs2_dinode *fe;
837 struct ocfs2_extent_block *eb;
838 struct ocfs2_extent_list *el;
839 struct buffer_head *bh = NULL;
840 struct buffer_head *lowest_bh = NULL;
846 fe = (struct ocfs2_dinode *) fe_bh->b_data;
847 el = &fe->id2.i_list;
849 while(le16_to_cpu(el->l_tree_depth) > 1) {
850 if (le16_to_cpu(el->l_next_free_rec) == 0) {
851 ocfs2_error(inode->i_sb, "Dinode %llu has empty "
852 "extent list (next_free_rec == 0)",
853 (unsigned long long)OCFS2_I(inode)->ip_blkno);
857 i = le16_to_cpu(el->l_next_free_rec) - 1;
858 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
860 ocfs2_error(inode->i_sb, "Dinode %llu has extent "
861 "list where extent # %d has no physical "
863 (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
873 status = ocfs2_read_block(osb, blkno, &bh, OCFS2_BH_CACHED,
880 eb = (struct ocfs2_extent_block *) bh->b_data;
881 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
882 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
888 if (le16_to_cpu(el->l_next_free_rec) <
889 le16_to_cpu(el->l_count)) {
897 /* If we didn't find one and the fe doesn't have any room,
900 && (fe->id2.i_list.l_next_free_rec == fe->id2.i_list.l_count))
903 *target_bh = lowest_bh;
913 * Grow a b-tree so that it has more records.
915 * We might shift the tree depth in which case existing paths should
916 * be considered invalid.
918 * Tree depth after the grow is returned via *final_depth.
920 * *last_eb_bh will be updated by ocfs2_add_branch().
922 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
923 struct buffer_head *di_bh, int *final_depth,
924 struct buffer_head **last_eb_bh,
925 struct ocfs2_alloc_context *meta_ac)
928 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
929 int depth = le16_to_cpu(di->id2.i_list.l_tree_depth);
930 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
931 struct buffer_head *bh = NULL;
933 BUG_ON(meta_ac == NULL);
935 shift = ocfs2_find_branch_target(osb, inode, di_bh, &bh);
942 /* We traveled all the way to the bottom of the allocation tree
943 * and didn't find room for any more extents - we need to add
944 * another tree level */
947 mlog(0, "need to shift tree depth (current = %d)\n", depth);
949 /* ocfs2_shift_tree_depth will return us a buffer with
950 * the new extent block (so we can pass that to
951 * ocfs2_add_branch). */
952 ret = ocfs2_shift_tree_depth(osb, handle, inode, di_bh,
961 * Special case: we have room now if we shifted from
962 * tree_depth 0, so no more work needs to be done.
964 * We won't be calling add_branch, so pass
965 * back *last_eb_bh as the new leaf. At depth
966 * zero, it should always be null so there's
967 * no reason to brelse.
976 /* call ocfs2_add_branch to add the final part of the tree with
978 mlog(0, "add branch. bh = %p\n", bh);
979 ret = ocfs2_add_branch(osb, handle, inode, di_bh, bh, last_eb_bh,
988 *final_depth = depth;
994 * This is only valid for leaf nodes, which are the only ones that can
995 * have empty extents anyway.
997 static inline int ocfs2_is_empty_extent(struct ocfs2_extent_rec *rec)
999 return !rec->e_leaf_clusters;
1003 * This function will discard the rightmost extent record.
1005 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1007 int next_free = le16_to_cpu(el->l_next_free_rec);
1008 int count = le16_to_cpu(el->l_count);
1009 unsigned int num_bytes;
1012 /* This will cause us to go off the end of our extent list. */
1013 BUG_ON(next_free >= count);
1015 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1017 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1020 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1021 struct ocfs2_extent_rec *insert_rec)
1023 int i, insert_index, next_free, has_empty, num_bytes;
1024 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1025 struct ocfs2_extent_rec *rec;
1027 next_free = le16_to_cpu(el->l_next_free_rec);
1028 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1032 /* The tree code before us didn't allow enough room in the leaf. */
1033 if (el->l_next_free_rec == el->l_count && !has_empty)
1037 * The easiest way to approach this is to just remove the
1038 * empty extent and temporarily decrement next_free.
1042 * If next_free was 1 (only an empty extent), this
1043 * loop won't execute, which is fine. We still want
1044 * the decrement above to happen.
1046 for(i = 0; i < (next_free - 1); i++)
1047 el->l_recs[i] = el->l_recs[i+1];
1053 * Figure out what the new record index should be.
1055 for(i = 0; i < next_free; i++) {
1056 rec = &el->l_recs[i];
1058 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1063 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1064 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
1066 BUG_ON(insert_index < 0);
1067 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1068 BUG_ON(insert_index > next_free);
1071 * No need to memmove if we're just adding to the tail.
1073 if (insert_index != next_free) {
1074 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1076 num_bytes = next_free - insert_index;
1077 num_bytes *= sizeof(struct ocfs2_extent_rec);
1078 memmove(&el->l_recs[insert_index + 1],
1079 &el->l_recs[insert_index],
1084 * Either we had an empty extent, and need to re-increment or
1085 * there was no empty extent on a non full rightmost leaf node,
1086 * in which case we still need to increment.
1089 el->l_next_free_rec = cpu_to_le16(next_free);
1091 * Make sure none of the math above just messed up our tree.
1093 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1095 el->l_recs[insert_index] = *insert_rec;
1099 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1101 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1103 BUG_ON(num_recs == 0);
1105 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1107 size = num_recs * sizeof(struct ocfs2_extent_rec);
1108 memmove(&el->l_recs[0], &el->l_recs[1], size);
1109 memset(&el->l_recs[num_recs], 0,
1110 sizeof(struct ocfs2_extent_rec));
1111 el->l_next_free_rec = cpu_to_le16(num_recs);
1116 * Create an empty extent record .
1118 * l_next_free_rec may be updated.
1120 * If an empty extent already exists do nothing.
1122 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1124 int next_free = le16_to_cpu(el->l_next_free_rec);
1126 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1131 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1134 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1135 "Asked to create an empty extent in a full list:\n"
1136 "count = %u, tree depth = %u",
1137 le16_to_cpu(el->l_count),
1138 le16_to_cpu(el->l_tree_depth));
1140 ocfs2_shift_records_right(el);
1143 le16_add_cpu(&el->l_next_free_rec, 1);
1144 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1148 * For a rotation which involves two leaf nodes, the "root node" is
1149 * the lowest level tree node which contains a path to both leafs. This
1150 * resulting set of information can be used to form a complete "subtree"
1152 * This function is passed two full paths from the dinode down to a
1153 * pair of adjacent leaves. It's task is to figure out which path
1154 * index contains the subtree root - this can be the root index itself
1155 * in a worst-case rotation.
1157 * The array index of the subtree root is passed back.
1159 static int ocfs2_find_subtree_root(struct inode *inode,
1160 struct ocfs2_path *left,
1161 struct ocfs2_path *right)
1166 * Check that the caller passed in two paths from the same tree.
1168 BUG_ON(path_root_bh(left) != path_root_bh(right));
1174 * The caller didn't pass two adjacent paths.
1176 mlog_bug_on_msg(i > left->p_tree_depth,
1177 "Inode %lu, left depth %u, right depth %u\n"
1178 "left leaf blk %llu, right leaf blk %llu\n",
1179 inode->i_ino, left->p_tree_depth,
1180 right->p_tree_depth,
1181 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1182 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1183 } while (left->p_node[i].bh->b_blocknr ==
1184 right->p_node[i].bh->b_blocknr);
1189 typedef void (path_insert_t)(void *, struct buffer_head *);
1192 * Traverse a btree path in search of cpos, starting at root_el.
1194 * This code can be called with a cpos larger than the tree, in which
1195 * case it will return the rightmost path.
1197 static int __ocfs2_find_path(struct inode *inode,
1198 struct ocfs2_extent_list *root_el, u32 cpos,
1199 path_insert_t *func, void *data)
1204 struct buffer_head *bh = NULL;
1205 struct ocfs2_extent_block *eb;
1206 struct ocfs2_extent_list *el;
1207 struct ocfs2_extent_rec *rec;
1208 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1211 while (el->l_tree_depth) {
1212 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1213 ocfs2_error(inode->i_sb,
1214 "Inode %llu has empty extent list at "
1216 (unsigned long long)oi->ip_blkno,
1217 le16_to_cpu(el->l_tree_depth));
1223 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1224 rec = &el->l_recs[i];
1227 * In the case that cpos is off the allocation
1228 * tree, this should just wind up returning the
1231 range = le32_to_cpu(rec->e_cpos) +
1232 ocfs2_rec_clusters(el, rec);
1233 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1237 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1239 ocfs2_error(inode->i_sb,
1240 "Inode %llu has bad blkno in extent list "
1241 "at depth %u (index %d)\n",
1242 (unsigned long long)oi->ip_blkno,
1243 le16_to_cpu(el->l_tree_depth), i);
1250 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno,
1251 &bh, OCFS2_BH_CACHED, inode);
1257 eb = (struct ocfs2_extent_block *) bh->b_data;
1259 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1260 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1265 if (le16_to_cpu(el->l_next_free_rec) >
1266 le16_to_cpu(el->l_count)) {
1267 ocfs2_error(inode->i_sb,
1268 "Inode %llu has bad count in extent list "
1269 "at block %llu (next free=%u, count=%u)\n",
1270 (unsigned long long)oi->ip_blkno,
1271 (unsigned long long)bh->b_blocknr,
1272 le16_to_cpu(el->l_next_free_rec),
1273 le16_to_cpu(el->l_count));
1284 * Catch any trailing bh that the loop didn't handle.
1292 * Given an initialized path (that is, it has a valid root extent
1293 * list), this function will traverse the btree in search of the path
1294 * which would contain cpos.
1296 * The path traveled is recorded in the path structure.
1298 * Note that this will not do any comparisons on leaf node extent
1299 * records, so it will work fine in the case that we just added a tree
1302 struct find_path_data {
1304 struct ocfs2_path *path;
1306 static void find_path_ins(void *data, struct buffer_head *bh)
1308 struct find_path_data *fp = data;
1311 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1314 static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
1317 struct find_path_data data;
1321 return __ocfs2_find_path(inode, path_root_el(path), cpos,
1322 find_path_ins, &data);
1325 static void find_leaf_ins(void *data, struct buffer_head *bh)
1327 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1328 struct ocfs2_extent_list *el = &eb->h_list;
1329 struct buffer_head **ret = data;
1331 /* We want to retain only the leaf block. */
1332 if (le16_to_cpu(el->l_tree_depth) == 0) {
1338 * Find the leaf block in the tree which would contain cpos. No
1339 * checking of the actual leaf is done.
1341 * Some paths want to call this instead of allocating a path structure
1342 * and calling ocfs2_find_path().
1344 * This function doesn't handle non btree extent lists.
1346 int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
1347 u32 cpos, struct buffer_head **leaf_bh)
1350 struct buffer_head *bh = NULL;
1352 ret = __ocfs2_find_path(inode, root_el, cpos, find_leaf_ins, &bh);
1364 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1366 * Basically, we've moved stuff around at the bottom of the tree and
1367 * we need to fix up the extent records above the changes to reflect
1370 * left_rec: the record on the left.
1371 * left_child_el: is the child list pointed to by left_rec
1372 * right_rec: the record to the right of left_rec
1373 * right_child_el: is the child list pointed to by right_rec
1375 * By definition, this only works on interior nodes.
1377 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1378 struct ocfs2_extent_list *left_child_el,
1379 struct ocfs2_extent_rec *right_rec,
1380 struct ocfs2_extent_list *right_child_el)
1382 u32 left_clusters, right_end;
1385 * Interior nodes never have holes. Their cpos is the cpos of
1386 * the leftmost record in their child list. Their cluster
1387 * count covers the full theoretical range of their child list
1388 * - the range between their cpos and the cpos of the record
1389 * immediately to their right.
1391 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1392 if (ocfs2_is_empty_extent(&right_child_el->l_recs[0])) {
1393 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1394 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1396 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1397 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1400 * Calculate the rightmost cluster count boundary before
1401 * moving cpos - we will need to adjust clusters after
1402 * updating e_cpos to keep the same highest cluster count.
1404 right_end = le32_to_cpu(right_rec->e_cpos);
1405 right_end += le32_to_cpu(right_rec->e_int_clusters);
1407 right_rec->e_cpos = left_rec->e_cpos;
1408 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1410 right_end -= le32_to_cpu(right_rec->e_cpos);
1411 right_rec->e_int_clusters = cpu_to_le32(right_end);
1415 * Adjust the adjacent root node records involved in a
1416 * rotation. left_el_blkno is passed in as a key so that we can easily
1417 * find it's index in the root list.
1419 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1420 struct ocfs2_extent_list *left_el,
1421 struct ocfs2_extent_list *right_el,
1426 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1427 le16_to_cpu(left_el->l_tree_depth));
1429 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1430 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1435 * The path walking code should have never returned a root and
1436 * two paths which are not adjacent.
1438 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1440 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1441 &root_el->l_recs[i + 1], right_el);
1445 * We've changed a leaf block (in right_path) and need to reflect that
1446 * change back up the subtree.
1448 * This happens in multiple places:
1449 * - When we've moved an extent record from the left path leaf to the right
1450 * path leaf to make room for an empty extent in the left path leaf.
1451 * - When our insert into the right path leaf is at the leftmost edge
1452 * and requires an update of the path immediately to it's left. This
1453 * can occur at the end of some types of rotation and appending inserts.
1455 static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1456 struct ocfs2_path *left_path,
1457 struct ocfs2_path *right_path,
1461 struct ocfs2_extent_list *el, *left_el, *right_el;
1462 struct ocfs2_extent_rec *left_rec, *right_rec;
1463 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1466 * Update the counts and position values within all the
1467 * interior nodes to reflect the leaf rotation we just did.
1469 * The root node is handled below the loop.
1471 * We begin the loop with right_el and left_el pointing to the
1472 * leaf lists and work our way up.
1474 * NOTE: within this loop, left_el and right_el always refer
1475 * to the *child* lists.
1477 left_el = path_leaf_el(left_path);
1478 right_el = path_leaf_el(right_path);
1479 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
1480 mlog(0, "Adjust records at index %u\n", i);
1483 * One nice property of knowing that all of these
1484 * nodes are below the root is that we only deal with
1485 * the leftmost right node record and the rightmost
1488 el = left_path->p_node[i].el;
1489 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
1490 left_rec = &el->l_recs[idx];
1492 el = right_path->p_node[i].el;
1493 right_rec = &el->l_recs[0];
1495 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
1498 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
1502 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
1507 * Setup our list pointers now so that the current
1508 * parents become children in the next iteration.
1510 left_el = left_path->p_node[i].el;
1511 right_el = right_path->p_node[i].el;
1515 * At the root node, adjust the two adjacent records which
1516 * begin our path to the leaves.
1519 el = left_path->p_node[subtree_index].el;
1520 left_el = left_path->p_node[subtree_index + 1].el;
1521 right_el = right_path->p_node[subtree_index + 1].el;
1523 ocfs2_adjust_root_records(el, left_el, right_el,
1524 left_path->p_node[subtree_index + 1].bh->b_blocknr);
1526 root_bh = left_path->p_node[subtree_index].bh;
1528 ret = ocfs2_journal_dirty(handle, root_bh);
1533 static int ocfs2_rotate_subtree_right(struct inode *inode,
1535 struct ocfs2_path *left_path,
1536 struct ocfs2_path *right_path,
1540 struct buffer_head *right_leaf_bh;
1541 struct buffer_head *left_leaf_bh = NULL;
1542 struct buffer_head *root_bh;
1543 struct ocfs2_extent_list *right_el, *left_el;
1544 struct ocfs2_extent_rec move_rec;
1546 left_leaf_bh = path_leaf_bh(left_path);
1547 left_el = path_leaf_el(left_path);
1549 if (left_el->l_next_free_rec != left_el->l_count) {
1550 ocfs2_error(inode->i_sb,
1551 "Inode %llu has non-full interior leaf node %llu"
1553 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1554 (unsigned long long)left_leaf_bh->b_blocknr,
1555 le16_to_cpu(left_el->l_next_free_rec));
1560 * This extent block may already have an empty record, so we
1561 * return early if so.
1563 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
1566 root_bh = left_path->p_node[subtree_index].bh;
1567 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
1569 ret = ocfs2_journal_access(handle, inode, root_bh,
1570 OCFS2_JOURNAL_ACCESS_WRITE);
1576 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
1577 ret = ocfs2_journal_access(handle, inode,
1578 right_path->p_node[i].bh,
1579 OCFS2_JOURNAL_ACCESS_WRITE);
1585 ret = ocfs2_journal_access(handle, inode,
1586 left_path->p_node[i].bh,
1587 OCFS2_JOURNAL_ACCESS_WRITE);
1594 right_leaf_bh = path_leaf_bh(right_path);
1595 right_el = path_leaf_el(right_path);
1597 /* This is a code error, not a disk corruption. */
1598 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
1599 "because rightmost leaf block %llu is empty\n",
1600 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1601 (unsigned long long)right_leaf_bh->b_blocknr);
1603 ocfs2_create_empty_extent(right_el);
1605 ret = ocfs2_journal_dirty(handle, right_leaf_bh);
1611 /* Do the copy now. */
1612 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
1613 move_rec = left_el->l_recs[i];
1614 right_el->l_recs[0] = move_rec;
1617 * Clear out the record we just copied and shift everything
1618 * over, leaving an empty extent in the left leaf.
1620 * We temporarily subtract from next_free_rec so that the
1621 * shift will lose the tail record (which is now defunct).
1623 le16_add_cpu(&left_el->l_next_free_rec, -1);
1624 ocfs2_shift_records_right(left_el);
1625 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1626 le16_add_cpu(&left_el->l_next_free_rec, 1);
1628 ret = ocfs2_journal_dirty(handle, left_leaf_bh);
1634 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
1642 * Given a full path, determine what cpos value would return us a path
1643 * containing the leaf immediately to the left of the current one.
1645 * Will return zero if the path passed in is already the leftmost path.
1647 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
1648 struct ocfs2_path *path, u32 *cpos)
1652 struct ocfs2_extent_list *el;
1654 BUG_ON(path->p_tree_depth == 0);
1658 blkno = path_leaf_bh(path)->b_blocknr;
1660 /* Start at the tree node just above the leaf and work our way up. */
1661 i = path->p_tree_depth - 1;
1663 el = path->p_node[i].el;
1666 * Find the extent record just before the one in our
1669 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
1670 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
1674 * We've determined that the
1675 * path specified is already
1676 * the leftmost one - return a
1682 * The leftmost record points to our
1683 * leaf - we need to travel up the
1689 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
1690 *cpos = *cpos + ocfs2_rec_clusters(el,
1691 &el->l_recs[j - 1]);
1698 * If we got here, we never found a valid node where
1699 * the tree indicated one should be.
1702 "Invalid extent tree at extent block %llu\n",
1703 (unsigned long long)blkno);
1708 blkno = path->p_node[i].bh->b_blocknr;
1717 * Extend the transaction by enough credits to complete the rotation,
1718 * and still leave at least the original number of credits allocated
1719 * to this transaction.
1721 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
1723 struct ocfs2_path *path)
1725 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
1727 if (handle->h_buffer_credits < credits)
1728 return ocfs2_extend_trans(handle, credits);
1734 * Trap the case where we're inserting into the theoretical range past
1735 * the _actual_ left leaf range. Otherwise, we'll rotate a record
1736 * whose cpos is less than ours into the right leaf.
1738 * It's only necessary to look at the rightmost record of the left
1739 * leaf because the logic that calls us should ensure that the
1740 * theoretical ranges in the path components above the leaves are
1743 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
1746 struct ocfs2_extent_list *left_el;
1747 struct ocfs2_extent_rec *rec;
1750 left_el = path_leaf_el(left_path);
1751 next_free = le16_to_cpu(left_el->l_next_free_rec);
1752 rec = &left_el->l_recs[next_free - 1];
1754 if (insert_cpos > le32_to_cpu(rec->e_cpos))
1759 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
1761 int next_free = le16_to_cpu(el->l_next_free_rec);
1763 struct ocfs2_extent_rec *rec;
1768 rec = &el->l_recs[0];
1769 if (ocfs2_is_empty_extent(rec)) {
1773 rec = &el->l_recs[1];
1776 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1777 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1783 * Rotate all the records in a btree right one record, starting at insert_cpos.
1785 * The path to the rightmost leaf should be passed in.
1787 * The array is assumed to be large enough to hold an entire path (tree depth).
1789 * Upon succesful return from this function:
1791 * - The 'right_path' array will contain a path to the leaf block
1792 * whose range contains e_cpos.
1793 * - That leaf block will have a single empty extent in list index 0.
1794 * - In the case that the rotation requires a post-insert update,
1795 * *ret_left_path will contain a valid path which can be passed to
1796 * ocfs2_insert_path().
1798 static int ocfs2_rotate_tree_right(struct inode *inode,
1800 enum ocfs2_split_type split,
1802 struct ocfs2_path *right_path,
1803 struct ocfs2_path **ret_left_path)
1805 int ret, start, orig_credits = handle->h_buffer_credits;
1807 struct ocfs2_path *left_path = NULL;
1809 *ret_left_path = NULL;
1811 left_path = ocfs2_new_path(path_root_bh(right_path),
1812 path_root_el(right_path));
1819 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
1825 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
1828 * What we want to do here is:
1830 * 1) Start with the rightmost path.
1832 * 2) Determine a path to the leaf block directly to the left
1835 * 3) Determine the 'subtree root' - the lowest level tree node
1836 * which contains a path to both leaves.
1838 * 4) Rotate the subtree.
1840 * 5) Find the next subtree by considering the left path to be
1841 * the new right path.
1843 * The check at the top of this while loop also accepts
1844 * insert_cpos == cpos because cpos is only a _theoretical_
1845 * value to get us the left path - insert_cpos might very well
1846 * be filling that hole.
1848 * Stop at a cpos of '0' because we either started at the
1849 * leftmost branch (i.e., a tree with one branch and a
1850 * rotation inside of it), or we've gone as far as we can in
1851 * rotating subtrees.
1853 while (cpos && insert_cpos <= cpos) {
1854 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
1857 ret = ocfs2_find_path(inode, left_path, cpos);
1863 mlog_bug_on_msg(path_leaf_bh(left_path) ==
1864 path_leaf_bh(right_path),
1865 "Inode %lu: error during insert of %u "
1866 "(left path cpos %u) results in two identical "
1867 "paths ending at %llu\n",
1868 inode->i_ino, insert_cpos, cpos,
1869 (unsigned long long)
1870 path_leaf_bh(left_path)->b_blocknr);
1872 if (split == SPLIT_NONE &&
1873 ocfs2_rotate_requires_path_adjustment(left_path,
1877 * We've rotated the tree as much as we
1878 * should. The rest is up to
1879 * ocfs2_insert_path() to complete, after the
1880 * record insertion. We indicate this
1881 * situation by returning the left path.
1883 * The reason we don't adjust the records here
1884 * before the record insert is that an error
1885 * later might break the rule where a parent
1886 * record e_cpos will reflect the actual
1887 * e_cpos of the 1st nonempty record of the
1890 *ret_left_path = left_path;
1894 start = ocfs2_find_subtree_root(inode, left_path, right_path);
1896 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
1898 (unsigned long long) right_path->p_node[start].bh->b_blocknr,
1899 right_path->p_tree_depth);
1901 ret = ocfs2_extend_rotate_transaction(handle, start,
1902 orig_credits, right_path);
1908 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
1915 if (split != SPLIT_NONE &&
1916 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
1919 * A rotate moves the rightmost left leaf
1920 * record over to the leftmost right leaf
1921 * slot. If we're doing an extent split
1922 * instead of a real insert, then we have to
1923 * check that the extent to be split wasn't
1924 * just moved over. If it was, then we can
1925 * exit here, passing left_path back -
1926 * ocfs2_split_extent() is smart enough to
1927 * search both leaves.
1929 *ret_left_path = left_path;
1934 * There is no need to re-read the next right path
1935 * as we know that it'll be our current left
1936 * path. Optimize by copying values instead.
1938 ocfs2_mv_path(right_path, left_path);
1940 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
1949 ocfs2_free_path(left_path);
1955 static void ocfs2_update_edge_lengths(struct inode *inode, handle_t *handle,
1956 struct ocfs2_path *path)
1959 struct ocfs2_extent_rec *rec;
1960 struct ocfs2_extent_list *el;
1961 struct ocfs2_extent_block *eb;
1964 /* Path should always be rightmost. */
1965 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
1966 BUG_ON(eb->h_next_leaf_blk != 0ULL);
1969 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
1970 idx = le16_to_cpu(el->l_next_free_rec) - 1;
1971 rec = &el->l_recs[idx];
1972 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1974 for (i = 0; i < path->p_tree_depth; i++) {
1975 el = path->p_node[i].el;
1976 idx = le16_to_cpu(el->l_next_free_rec) - 1;
1977 rec = &el->l_recs[idx];
1979 rec->e_int_clusters = cpu_to_le32(range);
1980 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
1982 ocfs2_journal_dirty(handle, path->p_node[i].bh);
1986 static void ocfs2_unlink_path(struct inode *inode, handle_t *handle,
1987 struct ocfs2_cached_dealloc_ctxt *dealloc,
1988 struct ocfs2_path *path, int unlink_start)
1991 struct ocfs2_extent_block *eb;
1992 struct ocfs2_extent_list *el;
1993 struct buffer_head *bh;
1995 for(i = unlink_start; i < path_num_items(path); i++) {
1996 bh = path->p_node[i].bh;
1998 eb = (struct ocfs2_extent_block *)bh->b_data;
2000 * Not all nodes might have had their final count
2001 * decremented by the caller - handle this here.
2004 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2006 "Inode %llu, attempted to remove extent block "
2007 "%llu with %u records\n",
2008 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2009 (unsigned long long)le64_to_cpu(eb->h_blkno),
2010 le16_to_cpu(el->l_next_free_rec));
2012 ocfs2_journal_dirty(handle, bh);
2013 ocfs2_remove_from_cache(inode, bh);
2017 el->l_next_free_rec = 0;
2018 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2020 ocfs2_journal_dirty(handle, bh);
2022 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2026 ocfs2_remove_from_cache(inode, bh);
2030 static void ocfs2_unlink_subtree(struct inode *inode, handle_t *handle,
2031 struct ocfs2_path *left_path,
2032 struct ocfs2_path *right_path,
2034 struct ocfs2_cached_dealloc_ctxt *dealloc)
2037 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2038 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2039 struct ocfs2_extent_list *el;
2040 struct ocfs2_extent_block *eb;
2042 el = path_leaf_el(left_path);
2044 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2046 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2047 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2050 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2052 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2053 le16_add_cpu(&root_el->l_next_free_rec, -1);
2055 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2056 eb->h_next_leaf_blk = 0;
2058 ocfs2_journal_dirty(handle, root_bh);
2059 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2061 ocfs2_unlink_path(inode, handle, dealloc, right_path,
2065 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2066 struct ocfs2_path *left_path,
2067 struct ocfs2_path *right_path,
2069 struct ocfs2_cached_dealloc_ctxt *dealloc,
2072 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2073 struct buffer_head *root_bh, *di_bh = path_root_bh(right_path);
2074 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2075 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2076 struct ocfs2_extent_block *eb;
2080 right_leaf_el = path_leaf_el(right_path);
2081 left_leaf_el = path_leaf_el(left_path);
2082 root_bh = left_path->p_node[subtree_index].bh;
2083 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2085 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2088 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2089 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2091 * It's legal for us to proceed if the right leaf is
2092 * the rightmost one and it has an empty extent. There
2093 * are two cases to handle - whether the leaf will be
2094 * empty after removal or not. If the leaf isn't empty
2095 * then just remove the empty extent up front. The
2096 * next block will handle empty leaves by flagging
2099 * Non rightmost leaves will throw -EAGAIN and the
2100 * caller can manually move the subtree and retry.
2103 if (eb->h_next_leaf_blk != 0ULL)
2106 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2107 ret = ocfs2_journal_access(handle, inode,
2108 path_leaf_bh(right_path),
2109 OCFS2_JOURNAL_ACCESS_WRITE);
2115 ocfs2_remove_empty_extent(right_leaf_el);
2117 right_has_empty = 1;
2120 if (eb->h_next_leaf_blk == 0ULL &&
2121 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2123 * We have to update i_last_eb_blk during the meta
2126 ret = ocfs2_journal_access(handle, inode, di_bh,
2127 OCFS2_JOURNAL_ACCESS_WRITE);
2133 del_right_subtree = 1;
2137 * Getting here with an empty extent in the right path implies
2138 * that it's the rightmost path and will be deleted.
2140 BUG_ON(right_has_empty && !del_right_subtree);
2142 ret = ocfs2_journal_access(handle, inode, root_bh,
2143 OCFS2_JOURNAL_ACCESS_WRITE);
2149 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2150 ret = ocfs2_journal_access(handle, inode,
2151 right_path->p_node[i].bh,
2152 OCFS2_JOURNAL_ACCESS_WRITE);
2158 ret = ocfs2_journal_access(handle, inode,
2159 left_path->p_node[i].bh,
2160 OCFS2_JOURNAL_ACCESS_WRITE);
2167 if (!right_has_empty) {
2169 * Only do this if we're moving a real
2170 * record. Otherwise, the action is delayed until
2171 * after removal of the right path in which case we
2172 * can do a simple shift to remove the empty extent.
2174 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2175 memset(&right_leaf_el->l_recs[0], 0,
2176 sizeof(struct ocfs2_extent_rec));
2178 if (eb->h_next_leaf_blk == 0ULL) {
2180 * Move recs over to get rid of empty extent, decrease
2181 * next_free. This is allowed to remove the last
2182 * extent in our leaf (setting l_next_free_rec to
2183 * zero) - the delete code below won't care.
2185 ocfs2_remove_empty_extent(right_leaf_el);
2188 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2191 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2195 if (del_right_subtree) {
2196 ocfs2_unlink_subtree(inode, handle, left_path, right_path,
2197 subtree_index, dealloc);
2198 ocfs2_update_edge_lengths(inode, handle, left_path);
2200 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2201 di->i_last_eb_blk = eb->h_blkno;
2204 * Removal of the extent in the left leaf was skipped
2205 * above so we could delete the right path
2208 if (right_has_empty)
2209 ocfs2_remove_empty_extent(left_leaf_el);
2211 ret = ocfs2_journal_dirty(handle, di_bh);
2217 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2225 * Given a full path, determine what cpos value would return us a path
2226 * containing the leaf immediately to the right of the current one.
2228 * Will return zero if the path passed in is already the rightmost path.
2230 * This looks similar, but is subtly different to
2231 * ocfs2_find_cpos_for_left_leaf().
2233 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2234 struct ocfs2_path *path, u32 *cpos)
2238 struct ocfs2_extent_list *el;
2242 if (path->p_tree_depth == 0)
2245 blkno = path_leaf_bh(path)->b_blocknr;
2247 /* Start at the tree node just above the leaf and work our way up. */
2248 i = path->p_tree_depth - 1;
2252 el = path->p_node[i].el;
2255 * Find the extent record just after the one in our
2258 next_free = le16_to_cpu(el->l_next_free_rec);
2259 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2260 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2261 if (j == (next_free - 1)) {
2264 * We've determined that the
2265 * path specified is already
2266 * the rightmost one - return a
2272 * The rightmost record points to our
2273 * leaf - we need to travel up the
2279 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2285 * If we got here, we never found a valid node where
2286 * the tree indicated one should be.
2289 "Invalid extent tree at extent block %llu\n",
2290 (unsigned long long)blkno);
2295 blkno = path->p_node[i].bh->b_blocknr;
2303 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2305 struct buffer_head *bh,
2306 struct ocfs2_extent_list *el)
2310 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2313 ret = ocfs2_journal_access(handle, inode, bh,
2314 OCFS2_JOURNAL_ACCESS_WRITE);
2320 ocfs2_remove_empty_extent(el);
2322 ret = ocfs2_journal_dirty(handle, bh);
2330 static int __ocfs2_rotate_tree_left(struct inode *inode,
2331 handle_t *handle, int orig_credits,
2332 struct ocfs2_path *path,
2333 struct ocfs2_cached_dealloc_ctxt *dealloc,
2334 struct ocfs2_path **empty_extent_path)
2336 int ret, subtree_root, deleted;
2338 struct ocfs2_path *left_path = NULL;
2339 struct ocfs2_path *right_path = NULL;
2341 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2343 *empty_extent_path = NULL;
2345 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2352 left_path = ocfs2_new_path(path_root_bh(path),
2353 path_root_el(path));
2360 ocfs2_cp_path(left_path, path);
2362 right_path = ocfs2_new_path(path_root_bh(path),
2363 path_root_el(path));
2370 while (right_cpos) {
2371 ret = ocfs2_find_path(inode, right_path, right_cpos);
2377 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2380 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2382 (unsigned long long)
2383 right_path->p_node[subtree_root].bh->b_blocknr,
2384 right_path->p_tree_depth);
2386 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2387 orig_credits, left_path);
2393 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2394 right_path, subtree_root,
2396 if (ret == -EAGAIN) {
2398 * The rotation has to temporarily stop due to
2399 * the right subtree having an empty
2400 * extent. Pass it back to the caller for a
2403 *empty_extent_path = right_path;
2413 * The subtree rotate might have removed records on
2414 * the rightmost edge. If so, then rotation is
2420 ocfs2_mv_path(left_path, right_path);
2422 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2431 ocfs2_free_path(right_path);
2432 ocfs2_free_path(left_path);
2437 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
2438 struct ocfs2_path *path,
2439 struct ocfs2_cached_dealloc_ctxt *dealloc)
2441 int ret, subtree_index;
2443 struct ocfs2_path *left_path = NULL;
2444 struct ocfs2_dinode *di;
2445 struct ocfs2_extent_block *eb;
2446 struct ocfs2_extent_list *el;
2449 * XXX: This code assumes that the root is an inode, which is
2450 * true for now but may change as tree code gets generic.
2452 di = (struct ocfs2_dinode *)path_root_bh(path)->b_data;
2453 if (!OCFS2_IS_VALID_DINODE(di)) {
2455 ocfs2_error(inode->i_sb,
2456 "Inode %llu has invalid path root",
2457 (unsigned long long)OCFS2_I(inode)->ip_blkno);
2462 * There's two ways we handle this depending on
2463 * whether path is the only existing one.
2465 ret = ocfs2_extend_rotate_transaction(handle, 0,
2466 handle->h_buffer_credits,
2473 ret = ocfs2_journal_access_path(inode, handle, path);
2479 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
2487 * We have a path to the left of this one - it needs
2490 left_path = ocfs2_new_path(path_root_bh(path),
2491 path_root_el(path));
2498 ret = ocfs2_find_path(inode, left_path, cpos);
2504 ret = ocfs2_journal_access_path(inode, handle, left_path);
2510 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
2512 ocfs2_unlink_subtree(inode, handle, left_path, path,
2513 subtree_index, dealloc);
2514 ocfs2_update_edge_lengths(inode, handle, left_path);
2516 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2517 di->i_last_eb_blk = eb->h_blkno;
2520 * 'path' is also the leftmost path which
2521 * means it must be the only one. This gets
2522 * handled differently because we want to
2523 * revert the inode back to having extents
2526 ocfs2_unlink_path(inode, handle, dealloc, path, 1);
2528 el = &di->id2.i_list;
2529 el->l_tree_depth = 0;
2530 el->l_next_free_rec = 0;
2531 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2533 di->i_last_eb_blk = 0;
2536 ocfs2_journal_dirty(handle, path_root_bh(path));
2539 ocfs2_free_path(left_path);
2544 * Left rotation of btree records.
2546 * In many ways, this is (unsurprisingly) the opposite of right
2547 * rotation. We start at some non-rightmost path containing an empty
2548 * extent in the leaf block. The code works its way to the rightmost
2549 * path by rotating records to the left in every subtree.
2551 * This is used by any code which reduces the number of extent records
2552 * in a leaf. After removal, an empty record should be placed in the
2553 * leftmost list position.
2555 * This won't handle a length update of the rightmost path records if
2556 * the rightmost tree leaf record is removed so the caller is
2557 * responsible for detecting and correcting that.
2559 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
2560 struct ocfs2_path *path,
2561 struct ocfs2_cached_dealloc_ctxt *dealloc)
2563 int ret, orig_credits = handle->h_buffer_credits;
2564 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
2565 struct ocfs2_extent_block *eb;
2566 struct ocfs2_extent_list *el;
2568 el = path_leaf_el(path);
2569 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2572 if (path->p_tree_depth == 0) {
2573 rightmost_no_delete:
2575 * In-inode extents. This is trivially handled, so do
2578 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
2580 path_leaf_el(path));
2587 * Handle rightmost branch now. There's several cases:
2588 * 1) simple rotation leaving records in there. That's trivial.
2589 * 2) rotation requiring a branch delete - there's no more
2590 * records left. Two cases of this:
2591 * a) There are branches to the left.
2592 * b) This is also the leftmost (the only) branch.
2594 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
2595 * 2a) we need the left branch so that we can update it with the unlink
2596 * 2b) we need to bring the inode back to inline extents.
2599 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2601 if (eb->h_next_leaf_blk == 0) {
2603 * This gets a bit tricky if we're going to delete the
2604 * rightmost path. Get the other cases out of the way
2607 if (le16_to_cpu(el->l_next_free_rec) > 1)
2608 goto rightmost_no_delete;
2610 if (le16_to_cpu(el->l_next_free_rec) == 0) {
2612 ocfs2_error(inode->i_sb,
2613 "Inode %llu has empty extent block at %llu",
2614 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2615 (unsigned long long)le64_to_cpu(eb->h_blkno));
2620 * XXX: The caller can not trust "path" any more after
2621 * this as it will have been deleted. What do we do?
2623 * In theory the rotate-for-merge code will never get
2624 * here because it'll always ask for a rotate in a
2628 ret = ocfs2_remove_rightmost_path(inode, handle, path,
2636 * Now we can loop, remembering the path we get from -EAGAIN
2637 * and restarting from there.
2640 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
2641 dealloc, &restart_path);
2642 if (ret && ret != -EAGAIN) {
2647 while (ret == -EAGAIN) {
2648 tmp_path = restart_path;
2649 restart_path = NULL;
2651 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
2654 if (ret && ret != -EAGAIN) {
2659 ocfs2_free_path(tmp_path);
2667 ocfs2_free_path(tmp_path);
2668 ocfs2_free_path(restart_path);
2672 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
2675 struct ocfs2_extent_rec *rec = &el->l_recs[index];
2678 if (rec->e_leaf_clusters == 0) {
2680 * We consumed all of the merged-from record. An empty
2681 * extent cannot exist anywhere but the 1st array
2682 * position, so move things over if the merged-from
2683 * record doesn't occupy that position.
2685 * This creates a new empty extent so the caller
2686 * should be smart enough to have removed any existing
2690 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
2691 size = index * sizeof(struct ocfs2_extent_rec);
2692 memmove(&el->l_recs[1], &el->l_recs[0], size);
2696 * Always memset - the caller doesn't check whether it
2697 * created an empty extent, so there could be junk in
2700 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2705 * Remove split_rec clusters from the record at index and merge them
2706 * onto the beginning of the record at index + 1.
2708 static int ocfs2_merge_rec_right(struct inode *inode, struct buffer_head *bh,
2710 struct ocfs2_extent_rec *split_rec,
2711 struct ocfs2_extent_list *el, int index)
2714 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
2715 struct ocfs2_extent_rec *left_rec;
2716 struct ocfs2_extent_rec *right_rec;
2718 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
2720 left_rec = &el->l_recs[index];
2721 right_rec = &el->l_recs[index + 1];
2723 ret = ocfs2_journal_access(handle, inode, bh,
2724 OCFS2_JOURNAL_ACCESS_WRITE);
2730 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
2732 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
2733 le64_add_cpu(&right_rec->e_blkno,
2734 -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
2735 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
2737 ocfs2_cleanup_merge(el, index);
2739 ret = ocfs2_journal_dirty(handle, bh);
2748 * Remove split_rec clusters from the record at index and merge them
2749 * onto the tail of the record at index - 1.
2751 static int ocfs2_merge_rec_left(struct inode *inode, struct buffer_head *bh,
2753 struct ocfs2_extent_rec *split_rec,
2754 struct ocfs2_extent_list *el, int index)
2756 int ret, has_empty_extent = 0;
2757 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
2758 struct ocfs2_extent_rec *left_rec;
2759 struct ocfs2_extent_rec *right_rec;
2763 left_rec = &el->l_recs[index - 1];
2764 right_rec = &el->l_recs[index];
2765 if (ocfs2_is_empty_extent(&el->l_recs[0]))
2766 has_empty_extent = 1;
2768 ret = ocfs2_journal_access(handle, inode, bh,
2769 OCFS2_JOURNAL_ACCESS_WRITE);
2775 if (has_empty_extent && index == 1) {
2777 * The easy case - we can just plop the record right in.
2779 *left_rec = *split_rec;
2781 has_empty_extent = 0;
2783 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
2786 le32_add_cpu(&right_rec->e_cpos, split_clusters);
2787 le64_add_cpu(&right_rec->e_blkno,
2788 ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
2789 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
2791 ocfs2_cleanup_merge(el, index);
2793 ret = ocfs2_journal_dirty(handle, bh);
2801 static int ocfs2_try_to_merge_extent(struct inode *inode,
2803 struct ocfs2_path *left_path,
2805 struct ocfs2_extent_rec *split_rec,
2806 struct ocfs2_cached_dealloc_ctxt *dealloc,
2807 struct ocfs2_merge_ctxt *ctxt)
2810 int ret = 0, delete_tail_recs = 0;
2811 struct ocfs2_extent_list *el = path_leaf_el(left_path);
2812 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
2814 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
2816 if (ctxt->c_split_covers_rec) {
2819 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT ||
2820 ctxt->c_has_empty_extent)
2823 if (ctxt->c_has_empty_extent) {
2825 * The merge code will need to create an empty
2826 * extent to take the place of the newly
2827 * emptied slot. Remove any pre-existing empty
2828 * extents - having more than one in a leaf is
2831 ret = ocfs2_rotate_tree_left(inode, handle, left_path,
2838 rec = &el->l_recs[split_index];
2842 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
2844 * Left-right contig implies this.
2846 BUG_ON(!ctxt->c_split_covers_rec);
2847 BUG_ON(split_index == 0);
2850 * Since the leftright insert always covers the entire
2851 * extent, this call will delete the insert record
2852 * entirely, resulting in an empty extent record added to
2855 * Since the adding of an empty extent shifts
2856 * everything back to the right, there's no need to
2857 * update split_index here.
2859 ret = ocfs2_merge_rec_left(inode, path_leaf_bh(left_path),
2860 handle, split_rec, el, split_index);
2867 * We can only get this from logic error above.
2869 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
2872 * The left merge left us with an empty extent, remove
2875 ret = ocfs2_rotate_tree_left(inode, handle, left_path, dealloc);
2881 rec = &el->l_recs[split_index];
2884 * Note that we don't pass split_rec here on purpose -
2885 * we've merged it into the left side.
2887 ret = ocfs2_merge_rec_right(inode, path_leaf_bh(left_path),
2888 handle, rec, el, split_index);
2894 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
2896 ret = ocfs2_rotate_tree_left(inode, handle, left_path,
2899 * Error from this last rotate is not critical, so
2900 * print but don't bubble it up.
2907 * Merge a record to the left or right.
2909 * 'contig_type' is relative to the existing record,
2910 * so for example, if we're "right contig", it's to
2911 * the record on the left (hence the left merge).
2913 if (ctxt->c_contig_type == CONTIG_RIGHT) {
2914 ret = ocfs2_merge_rec_left(inode,
2915 path_leaf_bh(left_path),
2916 handle, split_rec, el,
2923 ret = ocfs2_merge_rec_right(inode,
2924 path_leaf_bh(left_path),
2925 handle, split_rec, el,
2933 if (ctxt->c_split_covers_rec) {
2935 * The merge may have left an empty extent in
2936 * our leaf. Try to rotate it away.
2938 ret = ocfs2_rotate_tree_left(inode, handle, left_path,
2950 static void ocfs2_subtract_from_rec(struct super_block *sb,
2951 enum ocfs2_split_type split,
2952 struct ocfs2_extent_rec *rec,
2953 struct ocfs2_extent_rec *split_rec)
2957 len_blocks = ocfs2_clusters_to_blocks(sb,
2958 le16_to_cpu(split_rec->e_leaf_clusters));
2960 if (split == SPLIT_LEFT) {
2962 * Region is on the left edge of the existing
2965 le32_add_cpu(&rec->e_cpos,
2966 le16_to_cpu(split_rec->e_leaf_clusters));
2967 le64_add_cpu(&rec->e_blkno, len_blocks);
2968 le16_add_cpu(&rec->e_leaf_clusters,
2969 -le16_to_cpu(split_rec->e_leaf_clusters));
2972 * Region is on the right edge of the existing
2975 le16_add_cpu(&rec->e_leaf_clusters,
2976 -le16_to_cpu(split_rec->e_leaf_clusters));
2981 * Do the final bits of extent record insertion at the target leaf
2982 * list. If this leaf is part of an allocation tree, it is assumed
2983 * that the tree above has been prepared.
2985 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
2986 struct ocfs2_extent_list *el,
2987 struct ocfs2_insert_type *insert,
2988 struct inode *inode)
2990 int i = insert->ins_contig_index;
2992 struct ocfs2_extent_rec *rec;
2994 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
2996 if (insert->ins_split != SPLIT_NONE) {
2997 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
2999 rec = &el->l_recs[i];
3000 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
3006 * Contiguous insert - either left or right.
3008 if (insert->ins_contig != CONTIG_NONE) {
3009 rec = &el->l_recs[i];
3010 if (insert->ins_contig == CONTIG_LEFT) {
3011 rec->e_blkno = insert_rec->e_blkno;
3012 rec->e_cpos = insert_rec->e_cpos;
3014 le16_add_cpu(&rec->e_leaf_clusters,
3015 le16_to_cpu(insert_rec->e_leaf_clusters));
3020 * Handle insert into an empty leaf.
3022 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3023 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3024 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3025 el->l_recs[0] = *insert_rec;
3026 el->l_next_free_rec = cpu_to_le16(1);
3033 if (insert->ins_appending == APPEND_TAIL) {
3034 i = le16_to_cpu(el->l_next_free_rec) - 1;
3035 rec = &el->l_recs[i];
3036 range = le32_to_cpu(rec->e_cpos)
3037 + le16_to_cpu(rec->e_leaf_clusters);
3038 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3040 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3041 le16_to_cpu(el->l_count),
3042 "inode %lu, depth %u, count %u, next free %u, "
3043 "rec.cpos %u, rec.clusters %u, "
3044 "insert.cpos %u, insert.clusters %u\n",
3046 le16_to_cpu(el->l_tree_depth),
3047 le16_to_cpu(el->l_count),
3048 le16_to_cpu(el->l_next_free_rec),
3049 le32_to_cpu(el->l_recs[i].e_cpos),
3050 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3051 le32_to_cpu(insert_rec->e_cpos),
3052 le16_to_cpu(insert_rec->e_leaf_clusters));
3054 el->l_recs[i] = *insert_rec;
3055 le16_add_cpu(&el->l_next_free_rec, 1);
3061 * Ok, we have to rotate.
3063 * At this point, it is safe to assume that inserting into an
3064 * empty leaf and appending to a leaf have both been handled
3067 * This leaf needs to have space, either by the empty 1st
3068 * extent record, or by virtue of an l_next_rec < l_count.
3070 ocfs2_rotate_leaf(el, insert_rec);
3073 static inline void ocfs2_update_dinode_clusters(struct inode *inode,
3074 struct ocfs2_dinode *di,
3077 le32_add_cpu(&di->i_clusters, clusters);
3078 spin_lock(&OCFS2_I(inode)->ip_lock);
3079 OCFS2_I(inode)->ip_clusters = le32_to_cpu(di->i_clusters);
3080 spin_unlock(&OCFS2_I(inode)->ip_lock);
3083 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3085 struct ocfs2_path *path,
3086 struct ocfs2_extent_rec *insert_rec)
3088 int ret, i, next_free;
3089 struct buffer_head *bh;
3090 struct ocfs2_extent_list *el;
3091 struct ocfs2_extent_rec *rec;
3094 * Update everything except the leaf block.
3096 for (i = 0; i < path->p_tree_depth; i++) {
3097 bh = path->p_node[i].bh;
3098 el = path->p_node[i].el;
3100 next_free = le16_to_cpu(el->l_next_free_rec);
3101 if (next_free == 0) {
3102 ocfs2_error(inode->i_sb,
3103 "Dinode %llu has a bad extent list",
3104 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3109 rec = &el->l_recs[next_free - 1];
3111 rec->e_int_clusters = insert_rec->e_cpos;
3112 le32_add_cpu(&rec->e_int_clusters,
3113 le16_to_cpu(insert_rec->e_leaf_clusters));
3114 le32_add_cpu(&rec->e_int_clusters,
3115 -le32_to_cpu(rec->e_cpos));
3117 ret = ocfs2_journal_dirty(handle, bh);
3124 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3125 struct ocfs2_extent_rec *insert_rec,
3126 struct ocfs2_path *right_path,
3127 struct ocfs2_path **ret_left_path)
3130 struct ocfs2_extent_list *el;
3131 struct ocfs2_path *left_path = NULL;
3133 *ret_left_path = NULL;
3136 * This shouldn't happen for non-trees. The extent rec cluster
3137 * count manipulation below only works for interior nodes.
3139 BUG_ON(right_path->p_tree_depth == 0);
3142 * If our appending insert is at the leftmost edge of a leaf,
3143 * then we might need to update the rightmost records of the
3146 el = path_leaf_el(right_path);
3147 next_free = le16_to_cpu(el->l_next_free_rec);
3148 if (next_free == 0 ||
3149 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3152 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3159 mlog(0, "Append may need a left path update. cpos: %u, "
3160 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3164 * No need to worry if the append is already in the
3168 left_path = ocfs2_new_path(path_root_bh(right_path),
3169 path_root_el(right_path));
3176 ret = ocfs2_find_path(inode, left_path, left_cpos);
3183 * ocfs2_insert_path() will pass the left_path to the
3189 ret = ocfs2_journal_access_path(inode, handle, right_path);
3195 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
3197 *ret_left_path = left_path;
3201 ocfs2_free_path(left_path);
3206 static void ocfs2_split_record(struct inode *inode,
3207 struct ocfs2_path *left_path,
3208 struct ocfs2_path *right_path,
3209 struct ocfs2_extent_rec *split_rec,
3210 enum ocfs2_split_type split)
3213 u32 cpos = le32_to_cpu(split_rec->e_cpos);
3214 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
3215 struct ocfs2_extent_rec *rec, *tmprec;
3217 right_el = path_leaf_el(right_path);;
3219 left_el = path_leaf_el(left_path);
3222 insert_el = right_el;
3223 index = ocfs2_search_extent_list(el, cpos);
3225 if (index == 0 && left_path) {
3226 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3229 * This typically means that the record
3230 * started in the left path but moved to the
3231 * right as a result of rotation. We either
3232 * move the existing record to the left, or we
3233 * do the later insert there.
3235 * In this case, the left path should always
3236 * exist as the rotate code will have passed
3237 * it back for a post-insert update.
3240 if (split == SPLIT_LEFT) {
3242 * It's a left split. Since we know
3243 * that the rotate code gave us an
3244 * empty extent in the left path, we
3245 * can just do the insert there.
3247 insert_el = left_el;
3250 * Right split - we have to move the
3251 * existing record over to the left
3252 * leaf. The insert will be into the
3253 * newly created empty extent in the
3256 tmprec = &right_el->l_recs[index];
3257 ocfs2_rotate_leaf(left_el, tmprec);
3260 memset(tmprec, 0, sizeof(*tmprec));
3261 index = ocfs2_search_extent_list(left_el, cpos);
3262 BUG_ON(index == -1);
3267 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
3269 * Left path is easy - we can just allow the insert to
3273 insert_el = left_el;
3274 index = ocfs2_search_extent_list(el, cpos);
3275 BUG_ON(index == -1);
3278 rec = &el->l_recs[index];
3279 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
3280 ocfs2_rotate_leaf(insert_el, split_rec);
3284 * This function only does inserts on an allocation b-tree. For dinode
3285 * lists, ocfs2_insert_at_leaf() is called directly.
3287 * right_path is the path we want to do the actual insert
3288 * in. left_path should only be passed in if we need to update that
3289 * portion of the tree after an edge insert.
3291 static int ocfs2_insert_path(struct inode *inode,
3293 struct ocfs2_path *left_path,
3294 struct ocfs2_path *right_path,
3295 struct ocfs2_extent_rec *insert_rec,
3296 struct ocfs2_insert_type *insert)
3298 int ret, subtree_index;
3299 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
3302 * Pass both paths to the journal. The majority of inserts
3303 * will be touching all components anyway.
3305 ret = ocfs2_journal_access_path(inode, handle, right_path);
3312 int credits = handle->h_buffer_credits;
3315 * There's a chance that left_path got passed back to
3316 * us without being accounted for in the
3317 * journal. Extend our transaction here to be sure we
3318 * can change those blocks.
3320 credits += left_path->p_tree_depth;
3322 ret = ocfs2_extend_trans(handle, credits);
3328 ret = ocfs2_journal_access_path(inode, handle, left_path);
3335 if (insert->ins_split != SPLIT_NONE) {
3337 * We could call ocfs2_insert_at_leaf() for some types
3338 * of splits, but it's easier to just let one seperate
3339 * function sort it all out.
3341 ocfs2_split_record(inode, left_path, right_path,
3342 insert_rec, insert->ins_split);
3344 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
3347 ret = ocfs2_journal_dirty(handle, leaf_bh);
3353 * The rotate code has indicated that we need to fix
3354 * up portions of the tree after the insert.
3356 * XXX: Should we extend the transaction here?
3358 subtree_index = ocfs2_find_subtree_root(inode, left_path,
3360 ocfs2_complete_edge_insert(inode, handle, left_path,
3361 right_path, subtree_index);
3369 static int ocfs2_do_insert_extent(struct inode *inode,
3371 struct buffer_head *di_bh,
3372 struct ocfs2_extent_rec *insert_rec,
3373 struct ocfs2_insert_type *type)
3375 int ret, rotate = 0;
3377 struct ocfs2_path *right_path = NULL;
3378 struct ocfs2_path *left_path = NULL;
3379 struct ocfs2_dinode *di;
3380 struct ocfs2_extent_list *el;
3382 di = (struct ocfs2_dinode *) di_bh->b_data;
3383 el = &di->id2.i_list;
3385 ret = ocfs2_journal_access(handle, inode, di_bh,
3386 OCFS2_JOURNAL_ACCESS_WRITE);
3392 if (le16_to_cpu(el->l_tree_depth) == 0) {
3393 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
3394 goto out_update_clusters;
3397 right_path = ocfs2_new_inode_path(di_bh);
3405 * Determine the path to start with. Rotations need the
3406 * rightmost path, everything else can go directly to the
3409 cpos = le32_to_cpu(insert_rec->e_cpos);
3410 if (type->ins_appending == APPEND_NONE &&
3411 type->ins_contig == CONTIG_NONE) {
3416 ret = ocfs2_find_path(inode, right_path, cpos);
3423 * Rotations and appends need special treatment - they modify
3424 * parts of the tree's above them.
3426 * Both might pass back a path immediate to the left of the
3427 * one being inserted to. This will be cause
3428 * ocfs2_insert_path() to modify the rightmost records of
3429 * left_path to account for an edge insert.
3431 * XXX: When modifying this code, keep in mind that an insert
3432 * can wind up skipping both of these two special cases...
3435 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
3436 le32_to_cpu(insert_rec->e_cpos),
3437 right_path, &left_path);
3442 } else if (type->ins_appending == APPEND_TAIL
3443 && type->ins_contig != CONTIG_LEFT) {
3444 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
3445 right_path, &left_path);
3452 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
3459 out_update_clusters:
3460 if (type->ins_split == SPLIT_NONE)
3461 ocfs2_update_dinode_clusters(inode, di,
3462 le16_to_cpu(insert_rec->e_leaf_clusters));
3464 ret = ocfs2_journal_dirty(handle, di_bh);
3469 ocfs2_free_path(left_path);
3470 ocfs2_free_path(right_path);
3475 static enum ocfs2_contig_type
3476 ocfs2_figure_merge_contig_type(struct inode *inode,
3477 struct ocfs2_extent_list *el, int index,
3478 struct ocfs2_extent_rec *split_rec)
3480 struct ocfs2_extent_rec *rec;
3481 enum ocfs2_contig_type ret = CONTIG_NONE;
3484 * We're careful to check for an empty extent record here -
3485 * the merge code will know what to do if it sees one.
3489 rec = &el->l_recs[index - 1];
3490 if (index == 1 && ocfs2_is_empty_extent(rec)) {
3491 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
3494 ret = ocfs2_extent_contig(inode, rec, split_rec);
3498 if (index < (le16_to_cpu(el->l_next_free_rec) - 1)) {
3499 enum ocfs2_contig_type contig_type;
3501 rec = &el->l_recs[index + 1];
3502 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
3504 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
3505 ret = CONTIG_LEFTRIGHT;
3506 else if (ret == CONTIG_NONE)
3513 static void ocfs2_figure_contig_type(struct inode *inode,
3514 struct ocfs2_insert_type *insert,
3515 struct ocfs2_extent_list *el,
3516 struct ocfs2_extent_rec *insert_rec)
3519 enum ocfs2_contig_type contig_type = CONTIG_NONE;
3521 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3523 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
3524 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
3526 if (contig_type != CONTIG_NONE) {
3527 insert->ins_contig_index = i;
3531 insert->ins_contig = contig_type;
3535 * This should only be called against the righmost leaf extent list.
3537 * ocfs2_figure_appending_type() will figure out whether we'll have to
3538 * insert at the tail of the rightmost leaf.
3540 * This should also work against the dinode list for tree's with 0
3541 * depth. If we consider the dinode list to be the rightmost leaf node
3542 * then the logic here makes sense.
3544 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
3545 struct ocfs2_extent_list *el,
3546 struct ocfs2_extent_rec *insert_rec)
3549 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
3550 struct ocfs2_extent_rec *rec;
3552 insert->ins_appending = APPEND_NONE;
3554 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3556 if (!el->l_next_free_rec)
3557 goto set_tail_append;
3559 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
3560 /* Were all records empty? */
3561 if (le16_to_cpu(el->l_next_free_rec) == 1)
3562 goto set_tail_append;
3565 i = le16_to_cpu(el->l_next_free_rec) - 1;
3566 rec = &el->l_recs[i];
3569 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
3570 goto set_tail_append;
3575 insert->ins_appending = APPEND_TAIL;
3579 * Helper function called at the begining of an insert.
3581 * This computes a few things that are commonly used in the process of
3582 * inserting into the btree:
3583 * - Whether the new extent is contiguous with an existing one.
3584 * - The current tree depth.
3585 * - Whether the insert is an appending one.
3586 * - The total # of free records in the tree.
3588 * All of the information is stored on the ocfs2_insert_type
3591 static int ocfs2_figure_insert_type(struct inode *inode,
3592 struct buffer_head *di_bh,
3593 struct buffer_head **last_eb_bh,
3594 struct ocfs2_extent_rec *insert_rec,
3596 struct ocfs2_insert_type *insert)
3599 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3600 struct ocfs2_extent_block *eb;
3601 struct ocfs2_extent_list *el;
3602 struct ocfs2_path *path = NULL;
3603 struct buffer_head *bh = NULL;
3605 insert->ins_split = SPLIT_NONE;
3607 el = &di->id2.i_list;
3608 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
3610 if (el->l_tree_depth) {
3612 * If we have tree depth, we read in the
3613 * rightmost extent block ahead of time as
3614 * ocfs2_figure_insert_type() and ocfs2_add_branch()
3615 * may want it later.
3617 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
3618 le64_to_cpu(di->i_last_eb_blk), &bh,
3619 OCFS2_BH_CACHED, inode);
3624 eb = (struct ocfs2_extent_block *) bh->b_data;
3629 * Unless we have a contiguous insert, we'll need to know if
3630 * there is room left in our allocation tree for another
3633 * XXX: This test is simplistic, we can search for empty
3634 * extent records too.
3636 *free_records = le16_to_cpu(el->l_count) -
3637 le16_to_cpu(el->l_next_free_rec);
3639 if (!insert->ins_tree_depth) {
3640 ocfs2_figure_contig_type(inode, insert, el, insert_rec);
3641 ocfs2_figure_appending_type(insert, el, insert_rec);
3645 path = ocfs2_new_inode_path(di_bh);
3653 * In the case that we're inserting past what the tree
3654 * currently accounts for, ocfs2_find_path() will return for
3655 * us the rightmost tree path. This is accounted for below in
3656 * the appending code.
3658 ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
3664 el = path_leaf_el(path);
3667 * Now that we have the path, there's two things we want to determine:
3668 * 1) Contiguousness (also set contig_index if this is so)
3670 * 2) Are we doing an append? We can trivially break this up
3671 * into two types of appends: simple record append, or a
3672 * rotate inside the tail leaf.
3674 ocfs2_figure_contig_type(inode, insert, el, insert_rec);
3677 * The insert code isn't quite ready to deal with all cases of
3678 * left contiguousness. Specifically, if it's an insert into
3679 * the 1st record in a leaf, it will require the adjustment of
3680 * cluster count on the last record of the path directly to it's
3681 * left. For now, just catch that case and fool the layers
3682 * above us. This works just fine for tree_depth == 0, which
3683 * is why we allow that above.
3685 if (insert->ins_contig == CONTIG_LEFT &&
3686 insert->ins_contig_index == 0)
3687 insert->ins_contig = CONTIG_NONE;
3690 * Ok, so we can simply compare against last_eb to figure out
3691 * whether the path doesn't exist. This will only happen in
3692 * the case that we're doing a tail append, so maybe we can
3693 * take advantage of that information somehow.
3695 if (le64_to_cpu(di->i_last_eb_blk) == path_leaf_bh(path)->b_blocknr) {
3697 * Ok, ocfs2_find_path() returned us the rightmost
3698 * tree path. This might be an appending insert. There are
3700 * 1) We're doing a true append at the tail:
3701 * -This might even be off the end of the leaf
3702 * 2) We're "appending" by rotating in the tail
3704 ocfs2_figure_appending_type(insert, el, insert_rec);
3708 ocfs2_free_path(path);
3718 * Insert an extent into an inode btree.
3720 * The caller needs to update fe->i_clusters
3722 int ocfs2_insert_extent(struct ocfs2_super *osb,
3724 struct inode *inode,
3725 struct buffer_head *fe_bh,
3730 struct ocfs2_alloc_context *meta_ac)
3733 int uninitialized_var(free_records);
3734 struct buffer_head *last_eb_bh = NULL;
3735 struct ocfs2_insert_type insert = {0, };
3736 struct ocfs2_extent_rec rec;
3738 mlog(0, "add %u clusters at position %u to inode %llu\n",
3739 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
3741 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
3742 (OCFS2_I(inode)->ip_clusters != cpos),
3743 "Device %s, asking for sparse allocation: inode %llu, "
3744 "cpos %u, clusters %u\n",
3746 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos,
3747 OCFS2_I(inode)->ip_clusters);
3749 memset(&rec, 0, sizeof(rec));
3750 rec.e_cpos = cpu_to_le32(cpos);
3751 rec.e_blkno = cpu_to_le64(start_blk);
3752 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
3753 rec.e_flags = flags;
3755 status = ocfs2_figure_insert_type(inode, fe_bh, &last_eb_bh, &rec,
3756 &free_records, &insert);
3762 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
3763 "Insert.contig_index: %d, Insert.free_records: %d, "
3764 "Insert.tree_depth: %d\n",
3765 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
3766 free_records, insert.ins_tree_depth);
3768 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
3769 status = ocfs2_grow_tree(inode, handle, fe_bh,
3770 &insert.ins_tree_depth, &last_eb_bh,
3778 /* Finally, we can add clusters. This might rotate the tree for us. */
3779 status = ocfs2_do_insert_extent(inode, handle, fe_bh, &rec, &insert);
3783 ocfs2_extent_map_insert_rec(inode, &rec);
3793 static void ocfs2_make_right_split_rec(struct super_block *sb,
3794 struct ocfs2_extent_rec *split_rec,
3796 struct ocfs2_extent_rec *rec)
3798 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
3799 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
3801 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
3803 split_rec->e_cpos = cpu_to_le32(cpos);
3804 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
3806 split_rec->e_blkno = rec->e_blkno;
3807 le64_add_cpu(&split_rec->e_blkno,
3808 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
3810 split_rec->e_flags = rec->e_flags;
3813 static int ocfs2_split_and_insert(struct inode *inode,
3815 struct ocfs2_path *path,
3816 struct buffer_head *di_bh,
3817 struct buffer_head **last_eb_bh,
3819 struct ocfs2_extent_rec *orig_split_rec,
3820 struct ocfs2_alloc_context *meta_ac)
3823 unsigned int insert_range, rec_range, do_leftright = 0;
3824 struct ocfs2_extent_rec tmprec;
3825 struct ocfs2_extent_list *rightmost_el;
3826 struct ocfs2_extent_rec rec;
3827 struct ocfs2_extent_rec split_rec = *orig_split_rec;
3828 struct ocfs2_insert_type insert;
3829 struct ocfs2_extent_block *eb;
3830 struct ocfs2_dinode *di;
3834 * Store a copy of the record on the stack - it might move
3835 * around as the tree is manipulated below.
3837 rec = path_leaf_el(path)->l_recs[split_index];
3839 di = (struct ocfs2_dinode *)di_bh->b_data;
3840 rightmost_el = &di->id2.i_list;
3842 depth = le16_to_cpu(rightmost_el->l_tree_depth);
3844 BUG_ON(!(*last_eb_bh));
3845 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
3846 rightmost_el = &eb->h_list;
3849 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
3850 le16_to_cpu(rightmost_el->l_count)) {
3851 ret = ocfs2_grow_tree(inode, handle, di_bh, &depth, last_eb_bh,
3859 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
3860 insert.ins_appending = APPEND_NONE;
3861 insert.ins_contig = CONTIG_NONE;
3862 insert.ins_tree_depth = depth;
3864 insert_range = le32_to_cpu(split_rec.e_cpos) +
3865 le16_to_cpu(split_rec.e_leaf_clusters);
3866 rec_range = le32_to_cpu(rec.e_cpos) +
3867 le16_to_cpu(rec.e_leaf_clusters);
3869 if (split_rec.e_cpos == rec.e_cpos) {
3870 insert.ins_split = SPLIT_LEFT;
3871 } else if (insert_range == rec_range) {
3872 insert.ins_split = SPLIT_RIGHT;
3875 * Left/right split. We fake this as a right split
3876 * first and then make a second pass as a left split.
3878 insert.ins_split = SPLIT_RIGHT;
3880 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
3885 BUG_ON(do_leftright);
3889 ret = ocfs2_do_insert_extent(inode, handle, di_bh, &split_rec,
3896 if (do_leftright == 1) {
3898 struct ocfs2_extent_list *el;
3901 split_rec = *orig_split_rec;
3903 ocfs2_reinit_path(path, 1);
3905 cpos = le32_to_cpu(split_rec.e_cpos);
3906 ret = ocfs2_find_path(inode, path, cpos);
3912 el = path_leaf_el(path);
3913 split_index = ocfs2_search_extent_list(el, cpos);
3922 * Mark part or all of the extent record at split_index in the leaf
3923 * pointed to by path as written. This removes the unwritten
3926 * Care is taken to handle contiguousness so as to not grow the tree.
3928 * meta_ac is not strictly necessary - we only truly need it if growth
3929 * of the tree is required. All other cases will degrade into a less
3930 * optimal tree layout.
3932 * last_eb_bh should be the rightmost leaf block for any inode with a
3933 * btree. Since a split may grow the tree or a merge might shrink it, the caller cannot trust the contents of that buffer after this call.
3935 * This code is optimized for readability - several passes might be
3936 * made over certain portions of the tree. All of those blocks will
3937 * have been brought into cache (and pinned via the journal), so the
3938 * extra overhead is not expressed in terms of disk reads.
3940 static int __ocfs2_mark_extent_written(struct inode *inode,
3941 struct buffer_head *di_bh,
3943 struct ocfs2_path *path,
3945 struct ocfs2_extent_rec *split_rec,
3946 struct ocfs2_alloc_context *meta_ac,
3947 struct ocfs2_cached_dealloc_ctxt *dealloc)
3950 struct ocfs2_extent_list *el = path_leaf_el(path);
3951 struct buffer_head *eb_bh, *last_eb_bh = NULL;
3952 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3953 struct ocfs2_merge_ctxt ctxt;
3954 struct ocfs2_extent_list *rightmost_el;
3956 if (!rec->e_flags & OCFS2_EXT_UNWRITTEN) {
3962 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
3963 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
3964 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
3970 eb_bh = path_leaf_bh(path);
3971 ret = ocfs2_journal_access(handle, inode, eb_bh,
3972 OCFS2_JOURNAL_ACCESS_WRITE);
3978 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, el,
3983 * The core merge / split code wants to know how much room is
3984 * left in this inodes allocation tree, so we pass the
3985 * rightmost extent list.
3987 if (path->p_tree_depth) {
3988 struct ocfs2_extent_block *eb;
3989 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3991 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
3992 le64_to_cpu(di->i_last_eb_blk),
3993 &last_eb_bh, OCFS2_BH_CACHED, inode);
3999 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4000 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
4001 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
4006 rightmost_el = &eb->h_list;
4008 rightmost_el = path_root_el(path);
4010 ctxt.c_used_tail_recs = le16_to_cpu(rightmost_el->l_next_free_rec);
4011 if (ctxt.c_used_tail_recs > 0 &&
4012 ocfs2_is_empty_extent(&rightmost_el->l_recs[0]))
4013 ctxt.c_used_tail_recs--;
4015 if (rec->e_cpos == split_rec->e_cpos &&
4016 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
4017 ctxt.c_split_covers_rec = 1;
4019 ctxt.c_split_covers_rec = 0;
4021 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
4023 mlog(0, "index: %d, contig: %u, used_tail_recs: %u, "
4024 "has_empty: %u, split_covers: %u\n", split_index,
4025 ctxt.c_contig_type, ctxt.c_used_tail_recs,
4026 ctxt.c_has_empty_extent, ctxt.c_split_covers_rec);
4028 if (ctxt.c_contig_type == CONTIG_NONE) {
4029 if (ctxt.c_split_covers_rec)
4030 el->l_recs[split_index] = *split_rec;
4032 ret = ocfs2_split_and_insert(inode, handle, path, di_bh,
4033 &last_eb_bh, split_index,
4034 split_rec, meta_ac);
4038 ret = ocfs2_try_to_merge_extent(inode, handle, path,
4039 split_index, split_rec,
4045 ocfs2_journal_dirty(handle, eb_bh);
4053 * Mark the already-existing extent at cpos as written for len clusters.
4055 * If the existing extent is larger than the request, initiate a
4056 * split. An attempt will be made at merging with adjacent extents.
4058 * The caller is responsible for passing down meta_ac if we'll need it.
4060 int ocfs2_mark_extent_written(struct inode *inode, struct buffer_head *di_bh,
4061 handle_t *handle, u32 cpos, u32 len, u32 phys,
4062 struct ocfs2_alloc_context *meta_ac,
4063 struct ocfs2_cached_dealloc_ctxt *dealloc)
4066 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
4067 struct ocfs2_extent_rec split_rec;
4068 struct ocfs2_path *left_path = NULL;
4069 struct ocfs2_extent_list *el;
4071 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4072 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
4074 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
4075 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
4076 "that are being written to, but the feature bit "
4077 "is not set in the super block.",
4078 (unsigned long long)OCFS2_I(inode)->ip_blkno);
4084 * XXX: This should be fixed up so that we just re-insert the
4085 * next extent records.
4087 ocfs2_extent_map_trunc(inode, 0);
4089 left_path = ocfs2_new_inode_path(di_bh);
4096 ret = ocfs2_find_path(inode, left_path, cpos);
4101 el = path_leaf_el(left_path);
4103 index = ocfs2_search_extent_list(el, cpos);
4104 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4105 ocfs2_error(inode->i_sb,
4106 "Inode %llu has an extent at cpos %u which can no "
4107 "longer be found.\n",
4108 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
4113 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
4114 split_rec.e_cpos = cpu_to_le32(cpos);
4115 split_rec.e_leaf_clusters = cpu_to_le16(len);
4116 split_rec.e_blkno = cpu_to_le64(start_blkno);
4117 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
4118 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
4120 ret = __ocfs2_mark_extent_written(inode, di_bh, handle, left_path,
4121 index, &split_rec, meta_ac, dealloc);
4126 ocfs2_free_path(left_path);
4130 static int ocfs2_split_tree(struct inode *inode, struct buffer_head *di_bh,
4131 handle_t *handle, struct ocfs2_path *path,
4132 int index, u32 new_range,
4133 struct ocfs2_alloc_context *meta_ac)
4135 int ret, depth, credits = handle->h_buffer_credits;
4136 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4137 struct buffer_head *last_eb_bh = NULL;
4138 struct ocfs2_extent_block *eb;
4139 struct ocfs2_extent_list *rightmost_el, *el;
4140 struct ocfs2_extent_rec split_rec;
4141 struct ocfs2_extent_rec *rec;
4142 struct ocfs2_insert_type insert;
4145 * Setup the record to split before we grow the tree.
4147 el = path_leaf_el(path);
4148 rec = &el->l_recs[index];
4149 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
4151 depth = path->p_tree_depth;
4153 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4154 le64_to_cpu(di->i_last_eb_blk),
4155 &last_eb_bh, OCFS2_BH_CACHED, inode);
4161 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4162 rightmost_el = &eb->h_list;
4164 rightmost_el = path_leaf_el(path);
4166 credits += path->p_tree_depth + ocfs2_extend_meta_needed(di);
4167 ret = ocfs2_extend_trans(handle, credits);
4173 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4174 le16_to_cpu(rightmost_el->l_count)) {
4175 ret = ocfs2_grow_tree(inode, handle, di_bh, &depth, &last_eb_bh,
4183 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4184 insert.ins_appending = APPEND_NONE;
4185 insert.ins_contig = CONTIG_NONE;
4186 insert.ins_split = SPLIT_RIGHT;
4187 insert.ins_tree_depth = depth;
4189 ret = ocfs2_do_insert_extent(inode, handle, di_bh, &split_rec, &insert);
4198 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
4199 struct ocfs2_path *path, int index,
4200 struct ocfs2_cached_dealloc_ctxt *dealloc,
4204 u32 left_cpos, rec_range, trunc_range;
4205 int wants_rotate = 0, is_rightmost_tree_rec = 0;
4206 struct super_block *sb = inode->i_sb;
4207 struct ocfs2_path *left_path = NULL;
4208 struct ocfs2_extent_list *el = path_leaf_el(path);
4209 struct ocfs2_extent_rec *rec;
4210 struct ocfs2_extent_block *eb;
4212 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
4213 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc);
4222 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
4223 path->p_tree_depth) {
4225 * Check whether this is the rightmost tree record. If
4226 * we remove all of this record or part of its right
4227 * edge then an update of the record lengths above it
4230 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
4231 if (eb->h_next_leaf_blk == 0)
4232 is_rightmost_tree_rec = 1;
4235 rec = &el->l_recs[index];
4236 if (index == 0 && path->p_tree_depth &&
4237 le32_to_cpu(rec->e_cpos) == cpos) {
4239 * Changing the leftmost offset (via partial or whole
4240 * record truncate) of an interior (or rightmost) path
4241 * means we have to update the subtree that is formed
4242 * by this leaf and the one to it's left.
4244 * There are two cases we can skip:
4245 * 1) Path is the leftmost one in our inode tree.
4246 * 2) The leaf is rightmost and will be empty after
4247 * we remove the extent record - the rotate code
4248 * knows how to update the newly formed edge.
4251 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
4258 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
4259 left_path = ocfs2_new_path(path_root_bh(path),
4260 path_root_el(path));
4267 ret = ocfs2_find_path(inode, left_path, left_cpos);
4275 ret = ocfs2_extend_rotate_transaction(handle, 0,
4276 handle->h_buffer_credits,
4283 ret = ocfs2_journal_access_path(inode, handle, path);
4289 ret = ocfs2_journal_access_path(inode, handle, left_path);
4295 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
4296 trunc_range = cpos + len;
4298 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
4301 memset(rec, 0, sizeof(*rec));
4302 ocfs2_cleanup_merge(el, index);
4305 next_free = le16_to_cpu(el->l_next_free_rec);
4306 if (is_rightmost_tree_rec && next_free > 1) {
4308 * We skip the edge update if this path will
4309 * be deleted by the rotate code.
4311 rec = &el->l_recs[next_free - 1];
4312 ocfs2_adjust_rightmost_records(inode, handle, path,
4315 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
4316 /* Remove leftmost portion of the record. */
4317 le32_add_cpu(&rec->e_cpos, len);
4318 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
4319 le16_add_cpu(&rec->e_leaf_clusters, -len);
4320 } else if (rec_range == trunc_range) {
4321 /* Remove rightmost portion of the record */
4322 le16_add_cpu(&rec->e_leaf_clusters, -len);
4323 if (is_rightmost_tree_rec)
4324 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
4326 /* Caller should have trapped this. */
4327 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
4328 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
4329 le32_to_cpu(rec->e_cpos),
4330 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
4337 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
4338 ocfs2_complete_edge_insert(inode, handle, left_path, path,
4342 ocfs2_journal_dirty(handle, path_leaf_bh(path));
4344 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc);
4351 ocfs2_free_path(left_path);
4355 int ocfs2_remove_extent(struct inode *inode, struct buffer_head *di_bh,
4356 u32 cpos, u32 len, handle_t *handle,
4357 struct ocfs2_alloc_context *meta_ac,
4358 struct ocfs2_cached_dealloc_ctxt *dealloc)
4361 u32 rec_range, trunc_range;
4362 struct ocfs2_extent_rec *rec;
4363 struct ocfs2_extent_list *el;
4364 struct ocfs2_path *path;
4366 ocfs2_extent_map_trunc(inode, 0);
4368 path = ocfs2_new_inode_path(di_bh);
4375 ret = ocfs2_find_path(inode, path, cpos);
4381 el = path_leaf_el(path);
4382 index = ocfs2_search_extent_list(el, cpos);
4383 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4384 ocfs2_error(inode->i_sb,
4385 "Inode %llu has an extent at cpos %u which can no "
4386 "longer be found.\n",
4387 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
4393 * We have 3 cases of extent removal:
4394 * 1) Range covers the entire extent rec
4395 * 2) Range begins or ends on one edge of the extent rec
4396 * 3) Range is in the middle of the extent rec (no shared edges)
4398 * For case 1 we remove the extent rec and left rotate to
4401 * For case 2 we just shrink the existing extent rec, with a
4402 * tree update if the shrinking edge is also the edge of an
4405 * For case 3 we do a right split to turn the extent rec into
4406 * something case 2 can handle.
4408 rec = &el->l_recs[index];
4409 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
4410 trunc_range = cpos + len;
4412 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
4414 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
4415 "(cpos %u, len %u)\n",
4416 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
4417 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
4419 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
4420 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
4427 ret = ocfs2_split_tree(inode, di_bh, handle, path, index,
4428 trunc_range, meta_ac);
4435 * The split could have manipulated the tree enough to
4436 * move the record location, so we have to look for it again.
4438 ocfs2_reinit_path(path, 1);
4440 ret = ocfs2_find_path(inode, path, cpos);
4446 el = path_leaf_el(path);
4447 index = ocfs2_search_extent_list(el, cpos);
4448 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4449 ocfs2_error(inode->i_sb,
4450 "Inode %llu: split at cpos %u lost record.",
4451 (unsigned long long)OCFS2_I(inode)->ip_blkno,
4458 * Double check our values here. If anything is fishy,
4459 * it's easier to catch it at the top level.
4461 rec = &el->l_recs[index];
4462 rec_range = le32_to_cpu(rec->e_cpos) +
4463 ocfs2_rec_clusters(el, rec);
4464 if (rec_range != trunc_range) {
4465 ocfs2_error(inode->i_sb,
4466 "Inode %llu: error after split at cpos %u"
4467 "trunc len %u, existing record is (%u,%u)",
4468 (unsigned long long)OCFS2_I(inode)->ip_blkno,
4469 cpos, len, le32_to_cpu(rec->e_cpos),
4470 ocfs2_rec_clusters(el, rec));
4475 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
4484 ocfs2_free_path(path);
4488 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
4490 struct buffer_head *tl_bh = osb->osb_tl_bh;
4491 struct ocfs2_dinode *di;
4492 struct ocfs2_truncate_log *tl;
4494 di = (struct ocfs2_dinode *) tl_bh->b_data;
4495 tl = &di->id2.i_dealloc;
4497 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
4498 "slot %d, invalid truncate log parameters: used = "
4499 "%u, count = %u\n", osb->slot_num,
4500 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
4501 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
4504 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
4505 unsigned int new_start)
4507 unsigned int tail_index;
4508 unsigned int current_tail;
4510 /* No records, nothing to coalesce */
4511 if (!le16_to_cpu(tl->tl_used))
4514 tail_index = le16_to_cpu(tl->tl_used) - 1;
4515 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
4516 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
4518 return current_tail == new_start;
4521 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
4524 unsigned int num_clusters)
4527 unsigned int start_cluster, tl_count;
4528 struct inode *tl_inode = osb->osb_tl_inode;
4529 struct buffer_head *tl_bh = osb->osb_tl_bh;
4530 struct ocfs2_dinode *di;
4531 struct ocfs2_truncate_log *tl;
4533 mlog_entry("start_blk = %llu, num_clusters = %u\n",
4534 (unsigned long long)start_blk, num_clusters);
4536 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
4538 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
4540 di = (struct ocfs2_dinode *) tl_bh->b_data;
4541 tl = &di->id2.i_dealloc;
4542 if (!OCFS2_IS_VALID_DINODE(di)) {
4543 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
4548 tl_count = le16_to_cpu(tl->tl_count);
4549 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
4551 "Truncate record count on #%llu invalid "
4552 "wanted %u, actual %u\n",
4553 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
4554 ocfs2_truncate_recs_per_inode(osb->sb),
4555 le16_to_cpu(tl->tl_count));
4557 /* Caller should have known to flush before calling us. */
4558 index = le16_to_cpu(tl->tl_used);
4559 if (index >= tl_count) {
4565 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
4566 OCFS2_JOURNAL_ACCESS_WRITE);
4572 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
4573 "%llu (index = %d)\n", num_clusters, start_cluster,
4574 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
4576 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
4578 * Move index back to the record we are coalescing with.
4579 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
4583 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
4584 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
4585 index, le32_to_cpu(tl->tl_recs[index].t_start),
4588 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
4589 tl->tl_used = cpu_to_le16(index + 1);
4591 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
4593 status = ocfs2_journal_dirty(handle, tl_bh);
4604 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
4606 struct inode *data_alloc_inode,
4607 struct buffer_head *data_alloc_bh)
4611 unsigned int num_clusters;
4613 struct ocfs2_truncate_rec rec;
4614 struct ocfs2_dinode *di;
4615 struct ocfs2_truncate_log *tl;
4616 struct inode *tl_inode = osb->osb_tl_inode;
4617 struct buffer_head *tl_bh = osb->osb_tl_bh;
4621 di = (struct ocfs2_dinode *) tl_bh->b_data;
4622 tl = &di->id2.i_dealloc;
4623 i = le16_to_cpu(tl->tl_used) - 1;
4625 /* Caller has given us at least enough credits to
4626 * update the truncate log dinode */
4627 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
4628 OCFS2_JOURNAL_ACCESS_WRITE);
4634 tl->tl_used = cpu_to_le16(i);
4636 status = ocfs2_journal_dirty(handle, tl_bh);
4642 /* TODO: Perhaps we can calculate the bulk of the
4643 * credits up front rather than extending like
4645 status = ocfs2_extend_trans(handle,
4646 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
4652 rec = tl->tl_recs[i];
4653 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
4654 le32_to_cpu(rec.t_start));
4655 num_clusters = le32_to_cpu(rec.t_clusters);
4657 /* if start_blk is not set, we ignore the record as
4660 mlog(0, "free record %d, start = %u, clusters = %u\n",
4661 i, le32_to_cpu(rec.t_start), num_clusters);
4663 status = ocfs2_free_clusters(handle, data_alloc_inode,
4664 data_alloc_bh, start_blk,
4679 /* Expects you to already be holding tl_inode->i_mutex */
4680 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
4683 unsigned int num_to_flush;
4685 struct inode *tl_inode = osb->osb_tl_inode;
4686 struct inode *data_alloc_inode = NULL;
4687 struct buffer_head *tl_bh = osb->osb_tl_bh;
4688 struct buffer_head *data_alloc_bh = NULL;
4689 struct ocfs2_dinode *di;
4690 struct ocfs2_truncate_log *tl;
4694 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
4696 di = (struct ocfs2_dinode *) tl_bh->b_data;
4697 tl = &di->id2.i_dealloc;
4698 if (!OCFS2_IS_VALID_DINODE(di)) {
4699 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
4704 num_to_flush = le16_to_cpu(tl->tl_used);
4705 mlog(0, "Flush %u records from truncate log #%llu\n",
4706 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
4707 if (!num_to_flush) {
4712 data_alloc_inode = ocfs2_get_system_file_inode(osb,
4713 GLOBAL_BITMAP_SYSTEM_INODE,
4714 OCFS2_INVALID_SLOT);
4715 if (!data_alloc_inode) {
4717 mlog(ML_ERROR, "Could not get bitmap inode!\n");
4721 mutex_lock(&data_alloc_inode->i_mutex);
4723 status = ocfs2_meta_lock(data_alloc_inode, &data_alloc_bh, 1);
4729 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
4730 if (IS_ERR(handle)) {
4731 status = PTR_ERR(handle);
4736 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
4741 ocfs2_commit_trans(osb, handle);
4744 brelse(data_alloc_bh);
4745 ocfs2_meta_unlock(data_alloc_inode, 1);
4748 mutex_unlock(&data_alloc_inode->i_mutex);
4749 iput(data_alloc_inode);
4756 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
4759 struct inode *tl_inode = osb->osb_tl_inode;
4761 mutex_lock(&tl_inode->i_mutex);
4762 status = __ocfs2_flush_truncate_log(osb);
4763 mutex_unlock(&tl_inode->i_mutex);
4768 static void ocfs2_truncate_log_worker(struct work_struct *work)
4771 struct ocfs2_super *osb =
4772 container_of(work, struct ocfs2_super,
4773 osb_truncate_log_wq.work);
4777 status = ocfs2_flush_truncate_log(osb);
4784 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
4785 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
4788 if (osb->osb_tl_inode) {
4789 /* We want to push off log flushes while truncates are
4792 cancel_delayed_work(&osb->osb_truncate_log_wq);
4794 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
4795 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
4799 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
4801 struct inode **tl_inode,
4802 struct buffer_head **tl_bh)
4805 struct inode *inode = NULL;
4806 struct buffer_head *bh = NULL;
4808 inode = ocfs2_get_system_file_inode(osb,
4809 TRUNCATE_LOG_SYSTEM_INODE,
4813 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
4817 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
4818 OCFS2_BH_CACHED, inode);
4832 /* called during the 1st stage of node recovery. we stamp a clean
4833 * truncate log and pass back a copy for processing later. if the
4834 * truncate log does not require processing, a *tl_copy is set to
4836 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
4838 struct ocfs2_dinode **tl_copy)
4841 struct inode *tl_inode = NULL;
4842 struct buffer_head *tl_bh = NULL;
4843 struct ocfs2_dinode *di;
4844 struct ocfs2_truncate_log *tl;
4848 mlog(0, "recover truncate log from slot %d\n", slot_num);
4850 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
4856 di = (struct ocfs2_dinode *) tl_bh->b_data;
4857 tl = &di->id2.i_dealloc;
4858 if (!OCFS2_IS_VALID_DINODE(di)) {
4859 OCFS2_RO_ON_INVALID_DINODE(tl_inode->i_sb, di);
4864 if (le16_to_cpu(tl->tl_used)) {
4865 mlog(0, "We'll have %u logs to recover\n",
4866 le16_to_cpu(tl->tl_used));
4868 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
4875 /* Assuming the write-out below goes well, this copy
4876 * will be passed back to recovery for processing. */
4877 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
4879 /* All we need to do to clear the truncate log is set
4883 status = ocfs2_write_block(osb, tl_bh, tl_inode);
4896 if (status < 0 && (*tl_copy)) {
4905 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
4906 struct ocfs2_dinode *tl_copy)
4910 unsigned int clusters, num_recs, start_cluster;
4913 struct inode *tl_inode = osb->osb_tl_inode;
4914 struct ocfs2_truncate_log *tl;
4918 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
4919 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
4923 tl = &tl_copy->id2.i_dealloc;
4924 num_recs = le16_to_cpu(tl->tl_used);
4925 mlog(0, "cleanup %u records from %llu\n", num_recs,
4926 (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
4928 mutex_lock(&tl_inode->i_mutex);
4929 for(i = 0; i < num_recs; i++) {
4930 if (ocfs2_truncate_log_needs_flush(osb)) {
4931 status = __ocfs2_flush_truncate_log(osb);
4938 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
4939 if (IS_ERR(handle)) {
4940 status = PTR_ERR(handle);
4945 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
4946 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
4947 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
4949 status = ocfs2_truncate_log_append(osb, handle,
4950 start_blk, clusters);
4951 ocfs2_commit_trans(osb, handle);
4959 mutex_unlock(&tl_inode->i_mutex);
4965 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
4968 struct inode *tl_inode = osb->osb_tl_inode;
4973 cancel_delayed_work(&osb->osb_truncate_log_wq);
4974 flush_workqueue(ocfs2_wq);
4976 status = ocfs2_flush_truncate_log(osb);
4980 brelse(osb->osb_tl_bh);
4981 iput(osb->osb_tl_inode);
4987 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
4990 struct inode *tl_inode = NULL;
4991 struct buffer_head *tl_bh = NULL;
4995 status = ocfs2_get_truncate_log_info(osb,
5002 /* ocfs2_truncate_log_shutdown keys on the existence of
5003 * osb->osb_tl_inode so we don't set any of the osb variables
5004 * until we're sure all is well. */
5005 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
5006 ocfs2_truncate_log_worker);
5007 osb->osb_tl_bh = tl_bh;
5008 osb->osb_tl_inode = tl_inode;
5015 * Delayed de-allocation of suballocator blocks.
5017 * Some sets of block de-allocations might involve multiple suballocator inodes.
5019 * The locking for this can get extremely complicated, especially when
5020 * the suballocator inodes to delete from aren't known until deep
5021 * within an unrelated codepath.
5023 * ocfs2_extent_block structures are a good example of this - an inode
5024 * btree could have been grown by any number of nodes each allocating
5025 * out of their own suballoc inode.
5027 * These structures allow the delay of block de-allocation until a
5028 * later time, when locking of multiple cluster inodes won't cause
5033 * Describes a single block free from a suballocator
5035 struct ocfs2_cached_block_free {
5036 struct ocfs2_cached_block_free *free_next;
5038 unsigned int free_bit;
5041 struct ocfs2_per_slot_free_list {
5042 struct ocfs2_per_slot_free_list *f_next_suballocator;
5045 struct ocfs2_cached_block_free *f_first;
5048 static int ocfs2_free_cached_items(struct ocfs2_super *osb,
5051 struct ocfs2_cached_block_free *head)
5056 struct inode *inode;
5057 struct buffer_head *di_bh = NULL;
5058 struct ocfs2_cached_block_free *tmp;
5060 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
5067 mutex_lock(&inode->i_mutex);
5069 ret = ocfs2_meta_lock(inode, &di_bh, 1);
5075 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
5076 if (IS_ERR(handle)) {
5077 ret = PTR_ERR(handle);
5083 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
5085 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
5086 head->free_bit, (unsigned long long)head->free_blk);
5088 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
5089 head->free_bit, bg_blkno, 1);
5095 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
5102 head = head->free_next;
5107 ocfs2_commit_trans(osb, handle);
5110 ocfs2_meta_unlock(inode, 1);
5113 mutex_unlock(&inode->i_mutex);
5117 /* Premature exit may have left some dangling items. */
5119 head = head->free_next;
5126 int ocfs2_run_deallocs(struct ocfs2_super *osb,
5127 struct ocfs2_cached_dealloc_ctxt *ctxt)
5130 struct ocfs2_per_slot_free_list *fl;
5135 while (ctxt->c_first_suballocator) {
5136 fl = ctxt->c_first_suballocator;
5139 mlog(0, "Free items: (type %u, slot %d)\n",
5140 fl->f_inode_type, fl->f_slot);
5141 ret2 = ocfs2_free_cached_items(osb, fl->f_inode_type,
5142 fl->f_slot, fl->f_first);
5149 ctxt->c_first_suballocator = fl->f_next_suballocator;
5156 static struct ocfs2_per_slot_free_list *
5157 ocfs2_find_per_slot_free_list(int type,
5159 struct ocfs2_cached_dealloc_ctxt *ctxt)
5161 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
5164 if (fl->f_inode_type == type && fl->f_slot == slot)
5167 fl = fl->f_next_suballocator;
5170 fl = kmalloc(sizeof(*fl), GFP_NOFS);
5172 fl->f_inode_type = type;
5175 fl->f_next_suballocator = ctxt->c_first_suballocator;
5177 ctxt->c_first_suballocator = fl;
5182 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
5183 int type, int slot, u64 blkno,
5187 struct ocfs2_per_slot_free_list *fl;
5188 struct ocfs2_cached_block_free *item;
5190 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
5197 item = kmalloc(sizeof(*item), GFP_NOFS);
5204 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
5205 type, slot, bit, (unsigned long long)blkno);
5207 item->free_blk = blkno;
5208 item->free_bit = bit;
5209 item->free_next = fl->f_first;
5218 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
5219 struct ocfs2_extent_block *eb)
5221 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
5222 le16_to_cpu(eb->h_suballoc_slot),
5223 le64_to_cpu(eb->h_blkno),
5224 le16_to_cpu(eb->h_suballoc_bit));
5227 /* This function will figure out whether the currently last extent
5228 * block will be deleted, and if it will, what the new last extent
5229 * block will be so we can update his h_next_leaf_blk field, as well
5230 * as the dinodes i_last_eb_blk */
5231 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
5232 unsigned int clusters_to_del,
5233 struct ocfs2_path *path,
5234 struct buffer_head **new_last_eb)
5236 int next_free, ret = 0;
5238 struct ocfs2_extent_rec *rec;
5239 struct ocfs2_extent_block *eb;
5240 struct ocfs2_extent_list *el;
5241 struct buffer_head *bh = NULL;
5243 *new_last_eb = NULL;
5245 /* we have no tree, so of course, no last_eb. */
5246 if (!path->p_tree_depth)
5249 /* trunc to zero special case - this makes tree_depth = 0
5250 * regardless of what it is. */
5251 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
5254 el = path_leaf_el(path);
5255 BUG_ON(!el->l_next_free_rec);
5258 * Make sure that this extent list will actually be empty
5259 * after we clear away the data. We can shortcut out if
5260 * there's more than one non-empty extent in the
5261 * list. Otherwise, a check of the remaining extent is
5264 next_free = le16_to_cpu(el->l_next_free_rec);
5266 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
5270 /* We may have a valid extent in index 1, check it. */
5272 rec = &el->l_recs[1];
5275 * Fall through - no more nonempty extents, so we want
5276 * to delete this leaf.
5282 rec = &el->l_recs[0];
5287 * Check it we'll only be trimming off the end of this
5290 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
5294 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
5300 ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
5306 eb = (struct ocfs2_extent_block *) bh->b_data;
5308 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
5309 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
5315 get_bh(*new_last_eb);
5316 mlog(0, "returning block %llu, (cpos: %u)\n",
5317 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
5325 * Trim some clusters off the rightmost edge of a tree. Only called
5328 * The caller needs to:
5329 * - start journaling of each path component.
5330 * - compute and fully set up any new last ext block
5332 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
5333 handle_t *handle, struct ocfs2_truncate_context *tc,
5334 u32 clusters_to_del, u64 *delete_start)
5336 int ret, i, index = path->p_tree_depth;
5339 struct buffer_head *bh;
5340 struct ocfs2_extent_list *el;
5341 struct ocfs2_extent_rec *rec;
5345 while (index >= 0) {
5346 bh = path->p_node[index].bh;
5347 el = path->p_node[index].el;
5349 mlog(0, "traveling tree (index = %d, block = %llu)\n",
5350 index, (unsigned long long)bh->b_blocknr);
5352 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
5355 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
5356 ocfs2_error(inode->i_sb,
5357 "Inode %lu has invalid ext. block %llu",
5359 (unsigned long long)bh->b_blocknr);
5365 i = le16_to_cpu(el->l_next_free_rec) - 1;
5366 rec = &el->l_recs[i];
5368 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
5369 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
5370 ocfs2_rec_clusters(el, rec),
5371 (unsigned long long)le64_to_cpu(rec->e_blkno),
5372 le16_to_cpu(el->l_next_free_rec));
5374 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
5376 if (le16_to_cpu(el->l_tree_depth) == 0) {
5378 * If the leaf block contains a single empty
5379 * extent and no records, we can just remove
5382 if (i == 0 && ocfs2_is_empty_extent(rec)) {
5384 sizeof(struct ocfs2_extent_rec));
5385 el->l_next_free_rec = cpu_to_le16(0);
5391 * Remove any empty extents by shifting things
5392 * left. That should make life much easier on
5393 * the code below. This condition is rare
5394 * enough that we shouldn't see a performance
5397 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
5398 le16_add_cpu(&el->l_next_free_rec, -1);
5401 i < le16_to_cpu(el->l_next_free_rec); i++)
5402 el->l_recs[i] = el->l_recs[i + 1];
5404 memset(&el->l_recs[i], 0,
5405 sizeof(struct ocfs2_extent_rec));
5408 * We've modified our extent list. The
5409 * simplest way to handle this change
5410 * is to being the search from the
5413 goto find_tail_record;
5416 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
5419 * We'll use "new_edge" on our way back up the
5420 * tree to know what our rightmost cpos is.
5422 new_edge = le16_to_cpu(rec->e_leaf_clusters);
5423 new_edge += le32_to_cpu(rec->e_cpos);
5426 * The caller will use this to delete data blocks.
5428 *delete_start = le64_to_cpu(rec->e_blkno)
5429 + ocfs2_clusters_to_blocks(inode->i_sb,
5430 le16_to_cpu(rec->e_leaf_clusters));
5433 * If it's now empty, remove this record.
5435 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
5437 sizeof(struct ocfs2_extent_rec));
5438 le16_add_cpu(&el->l_next_free_rec, -1);
5441 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
5443 sizeof(struct ocfs2_extent_rec));
5444 le16_add_cpu(&el->l_next_free_rec, -1);
5449 /* Can this actually happen? */
5450 if (le16_to_cpu(el->l_next_free_rec) == 0)
5454 * We never actually deleted any clusters
5455 * because our leaf was empty. There's no
5456 * reason to adjust the rightmost edge then.
5461 rec->e_int_clusters = cpu_to_le32(new_edge);
5462 le32_add_cpu(&rec->e_int_clusters,
5463 -le32_to_cpu(rec->e_cpos));
5466 * A deleted child record should have been
5469 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
5473 ret = ocfs2_journal_dirty(handle, bh);
5479 mlog(0, "extent list container %llu, after: record %d: "
5480 "(%u, %u, %llu), next = %u.\n",
5481 (unsigned long long)bh->b_blocknr, i,
5482 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
5483 (unsigned long long)le64_to_cpu(rec->e_blkno),
5484 le16_to_cpu(el->l_next_free_rec));
5487 * We must be careful to only attempt delete of an
5488 * extent block (and not the root inode block).
5490 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
5491 struct ocfs2_extent_block *eb =
5492 (struct ocfs2_extent_block *)bh->b_data;
5495 * Save this for use when processing the
5498 deleted_eb = le64_to_cpu(eb->h_blkno);
5500 mlog(0, "deleting this extent block.\n");
5502 ocfs2_remove_from_cache(inode, bh);
5504 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
5505 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
5506 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
5508 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
5509 /* An error here is not fatal. */
5524 static int ocfs2_do_truncate(struct ocfs2_super *osb,
5525 unsigned int clusters_to_del,
5526 struct inode *inode,
5527 struct buffer_head *fe_bh,
5529 struct ocfs2_truncate_context *tc,
5530 struct ocfs2_path *path)
5533 struct ocfs2_dinode *fe;
5534 struct ocfs2_extent_block *last_eb = NULL;
5535 struct ocfs2_extent_list *el;
5536 struct buffer_head *last_eb_bh = NULL;
5539 fe = (struct ocfs2_dinode *) fe_bh->b_data;
5541 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
5549 * Each component will be touched, so we might as well journal
5550 * here to avoid having to handle errors later.
5552 status = ocfs2_journal_access_path(inode, handle, path);
5559 status = ocfs2_journal_access(handle, inode, last_eb_bh,
5560 OCFS2_JOURNAL_ACCESS_WRITE);
5566 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5569 el = &(fe->id2.i_list);
5572 * Lower levels depend on this never happening, but it's best
5573 * to check it up here before changing the tree.
5575 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
5576 ocfs2_error(inode->i_sb,
5577 "Inode %lu has an empty extent record, depth %u\n",
5578 inode->i_ino, le16_to_cpu(el->l_tree_depth));
5583 spin_lock(&OCFS2_I(inode)->ip_lock);
5584 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
5586 spin_unlock(&OCFS2_I(inode)->ip_lock);
5587 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
5588 inode->i_blocks = ocfs2_inode_sector_count(inode);
5590 status = ocfs2_trim_tree(inode, path, handle, tc,
5591 clusters_to_del, &delete_blk);
5597 if (le32_to_cpu(fe->i_clusters) == 0) {
5598 /* trunc to zero is a special case. */
5599 el->l_tree_depth = 0;
5600 fe->i_last_eb_blk = 0;
5602 fe->i_last_eb_blk = last_eb->h_blkno;
5604 status = ocfs2_journal_dirty(handle, fe_bh);
5611 /* If there will be a new last extent block, then by
5612 * definition, there cannot be any leaves to the right of
5614 last_eb->h_next_leaf_blk = 0;
5615 status = ocfs2_journal_dirty(handle, last_eb_bh);
5623 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
5637 static int ocfs2_writeback_zero_func(handle_t *handle, struct buffer_head *bh)
5639 set_buffer_uptodate(bh);
5640 mark_buffer_dirty(bh);
5644 static int ocfs2_ordered_zero_func(handle_t *handle, struct buffer_head *bh)
5646 set_buffer_uptodate(bh);
5647 mark_buffer_dirty(bh);
5648 return ocfs2_journal_dirty_data(handle, bh);
5651 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
5652 loff_t end, struct page **pages,
5653 int numpages, u64 phys, handle_t *handle)
5655 int i, ret, partial = 0;
5658 unsigned int from, to = PAGE_CACHE_SIZE;
5659 struct super_block *sb = inode->i_sb;
5661 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
5666 to = PAGE_CACHE_SIZE;
5667 for(i = 0; i < numpages; i++) {
5670 from = start & (PAGE_CACHE_SIZE - 1);
5671 if ((end >> PAGE_CACHE_SHIFT) == page->index)
5672 to = end & (PAGE_CACHE_SIZE - 1);
5674 BUG_ON(from > PAGE_CACHE_SIZE);
5675 BUG_ON(to > PAGE_CACHE_SIZE);
5677 ret = ocfs2_map_page_blocks(page, &phys, inode, from, to, 0);
5681 kaddr = kmap_atomic(page, KM_USER0);
5682 memset(kaddr + from, 0, to - from);
5683 kunmap_atomic(kaddr, KM_USER0);
5686 * Need to set the buffers we zero'd into uptodate
5687 * here if they aren't - ocfs2_map_page_blocks()
5688 * might've skipped some
5690 if (ocfs2_should_order_data(inode)) {
5691 ret = walk_page_buffers(handle,
5694 ocfs2_ordered_zero_func);
5698 ret = walk_page_buffers(handle, page_buffers(page),
5700 ocfs2_writeback_zero_func);
5706 SetPageUptodate(page);
5708 flush_dcache_page(page);
5710 start = (page->index + 1) << PAGE_CACHE_SHIFT;
5714 for (i = 0; i < numpages; i++) {
5717 mark_page_accessed(page);
5718 page_cache_release(page);
5723 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
5724 struct page **pages, int *num, u64 *phys)
5726 int i, numpages = 0, ret = 0;
5727 unsigned int ext_flags;
5728 struct super_block *sb = inode->i_sb;
5729 struct address_space *mapping = inode->i_mapping;
5730 unsigned long index;
5731 loff_t last_page_bytes;
5733 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
5734 BUG_ON(start > end);
5739 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
5740 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
5742 ret = ocfs2_extent_map_get_blocks(inode, start >> sb->s_blocksize_bits,
5743 phys, NULL, &ext_flags);
5749 /* Tail is a hole. */
5753 /* Tail is marked as unwritten, we can count on write to zero
5755 if (ext_flags & OCFS2_EXT_UNWRITTEN)
5758 last_page_bytes = PAGE_ALIGN(end);
5759 index = start >> PAGE_CACHE_SHIFT;
5761 pages[numpages] = grab_cache_page(mapping, index);
5762 if (!pages[numpages]) {
5770 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
5775 for (i = 0; i < numpages; i++) {
5777 unlock_page(pages[i]);
5778 page_cache_release(pages[i]);
5791 * Zero the area past i_size but still within an allocated
5792 * cluster. This avoids exposing nonzero data on subsequent file
5795 * We need to call this before i_size is updated on the inode because
5796 * otherwise block_write_full_page() will skip writeout of pages past
5797 * i_size. The new_i_size parameter is passed for this reason.
5799 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
5800 u64 range_start, u64 range_end)
5803 struct page **pages = NULL;
5807 * File systems which don't support sparse files zero on every
5810 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
5813 pages = kcalloc(ocfs2_pages_per_cluster(inode->i_sb),
5814 sizeof(struct page *), GFP_NOFS);
5815 if (pages == NULL) {
5821 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
5831 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
5832 numpages, phys, handle);
5835 * Initiate writeout of the pages we zero'd here. We don't
5836 * wait on them - the truncate_inode_pages() call later will
5839 ret = do_sync_mapping_range(inode->i_mapping, range_start,
5840 range_end - 1, SYNC_FILE_RANGE_WRITE);
5852 * It is expected, that by the time you call this function,
5853 * inode->i_size and fe->i_size have been adjusted.
5855 * WARNING: This will kfree the truncate context
5857 int ocfs2_commit_truncate(struct ocfs2_super *osb,
5858 struct inode *inode,
5859 struct buffer_head *fe_bh,
5860 struct ocfs2_truncate_context *tc)
5862 int status, i, credits, tl_sem = 0;
5863 u32 clusters_to_del, new_highest_cpos, range;
5864 struct ocfs2_extent_list *el;
5865 handle_t *handle = NULL;
5866 struct inode *tl_inode = osb->osb_tl_inode;
5867 struct ocfs2_path *path = NULL;
5871 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
5872 i_size_read(inode));
5874 path = ocfs2_new_inode_path(fe_bh);
5881 ocfs2_extent_map_trunc(inode, new_highest_cpos);
5885 * Check that we still have allocation to delete.
5887 if (OCFS2_I(inode)->ip_clusters == 0) {
5893 * Truncate always works against the rightmost tree branch.
5895 status = ocfs2_find_path(inode, path, UINT_MAX);
5901 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
5902 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
5905 * By now, el will point to the extent list on the bottom most
5906 * portion of this tree. Only the tail record is considered in
5909 * We handle the following cases, in order:
5910 * - empty extent: delete the remaining branch
5911 * - remove the entire record
5912 * - remove a partial record
5913 * - no record needs to be removed (truncate has completed)
5915 el = path_leaf_el(path);
5916 if (le16_to_cpu(el->l_next_free_rec) == 0) {
5917 ocfs2_error(inode->i_sb,
5918 "Inode %llu has empty extent block at %llu\n",
5919 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5920 (unsigned long long)path_leaf_bh(path)->b_blocknr);
5925 i = le16_to_cpu(el->l_next_free_rec) - 1;
5926 range = le32_to_cpu(el->l_recs[i].e_cpos) +
5927 ocfs2_rec_clusters(el, &el->l_recs[i]);
5928 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
5929 clusters_to_del = 0;
5930 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
5931 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
5932 } else if (range > new_highest_cpos) {
5933 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
5934 le32_to_cpu(el->l_recs[i].e_cpos)) -
5941 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
5942 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
5944 BUG_ON(clusters_to_del == 0);
5946 mutex_lock(&tl_inode->i_mutex);
5948 /* ocfs2_truncate_log_needs_flush guarantees us at least one
5949 * record is free for use. If there isn't any, we flush to get
5950 * an empty truncate log. */
5951 if (ocfs2_truncate_log_needs_flush(osb)) {
5952 status = __ocfs2_flush_truncate_log(osb);
5959 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
5960 (struct ocfs2_dinode *)fe_bh->b_data,
5962 handle = ocfs2_start_trans(osb, credits);
5963 if (IS_ERR(handle)) {
5964 status = PTR_ERR(handle);
5970 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
5977 mutex_unlock(&tl_inode->i_mutex);
5980 ocfs2_commit_trans(osb, handle);
5983 ocfs2_reinit_path(path, 1);
5986 * The check above will catch the case where we've truncated
5987 * away all allocation.
5993 ocfs2_schedule_truncate_log_flush(osb, 1);
5996 mutex_unlock(&tl_inode->i_mutex);
5999 ocfs2_commit_trans(osb, handle);
6001 ocfs2_run_deallocs(osb, &tc->tc_dealloc);
6003 ocfs2_free_path(path);
6005 /* This will drop the ext_alloc cluster lock for us */
6006 ocfs2_free_truncate_context(tc);
6013 * Expects the inode to already be locked.
6015 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
6016 struct inode *inode,
6017 struct buffer_head *fe_bh,
6018 struct ocfs2_truncate_context **tc)
6021 unsigned int new_i_clusters;
6022 struct ocfs2_dinode *fe;
6023 struct ocfs2_extent_block *eb;
6024 struct buffer_head *last_eb_bh = NULL;
6030 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
6031 i_size_read(inode));
6032 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6034 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
6035 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
6036 (unsigned long long)le64_to_cpu(fe->i_size));
6038 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
6044 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
6046 if (fe->id2.i_list.l_tree_depth) {
6047 status = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
6048 &last_eb_bh, OCFS2_BH_CACHED, inode);
6053 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6054 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
6055 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
6063 (*tc)->tc_last_eb_bh = last_eb_bh;
6069 ocfs2_free_truncate_context(*tc);
6076 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
6079 * The caller is responsible for completing deallocation
6080 * before freeing the context.
6082 if (tc->tc_dealloc.c_first_suballocator != NULL)
6084 "Truncate completion has non-empty dealloc context\n");
6086 if (tc->tc_last_eb_bh)
6087 brelse(tc->tc_last_eb_bh);