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"
53 * ocfs2_extent_tree and ocfs2_extent_tree_operations are used to abstract
54 * the b-tree operations in ocfs2. Now all the b-tree operations are not
55 * limited to ocfs2_dinode only. Any data which need to allocate clusters
56 * to store can use b-tree. And it only needs to implement its ocfs2_extent_tree
59 * ocfs2_extent_tree contains info for the root of the b-tree, it must have a
60 * root ocfs2_extent_list and a root_bh so that they can be used in the b-tree
62 * ocfs2_extent_tree_operations abstract the normal operations we do for
63 * the root of extent b-tree.
65 struct ocfs2_extent_tree;
67 struct ocfs2_extent_tree_operations {
68 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
70 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
71 void (*eo_update_clusters)(struct inode *inode,
72 struct ocfs2_extent_tree *et,
74 int (*eo_sanity_check)(struct inode *inode, struct ocfs2_extent_tree *et);
76 /* These are internal to ocfs2_extent_tree and don't have
77 * accessor functions */
78 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
81 struct ocfs2_extent_tree {
82 enum ocfs2_extent_tree_type et_type;
83 struct ocfs2_extent_tree_operations *et_ops;
84 struct buffer_head *et_root_bh;
85 struct ocfs2_extent_list *et_root_el;
87 unsigned int et_max_leaf_clusters;
90 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
92 struct ocfs2_dinode *di = et->et_object;
94 et->et_root_el = &di->id2.i_list;
97 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
100 struct ocfs2_dinode *di = et->et_object;
102 BUG_ON(et->et_type != OCFS2_DINODE_EXTENT);
103 di->i_last_eb_blk = cpu_to_le64(blkno);
106 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
108 struct ocfs2_dinode *di = et->et_object;
110 BUG_ON(et->et_type != OCFS2_DINODE_EXTENT);
111 return le64_to_cpu(di->i_last_eb_blk);
114 static void ocfs2_dinode_update_clusters(struct inode *inode,
115 struct ocfs2_extent_tree *et,
118 struct ocfs2_dinode *di = et->et_object;
120 le32_add_cpu(&di->i_clusters, clusters);
121 spin_lock(&OCFS2_I(inode)->ip_lock);
122 OCFS2_I(inode)->ip_clusters = le32_to_cpu(di->i_clusters);
123 spin_unlock(&OCFS2_I(inode)->ip_lock);
126 static int ocfs2_dinode_sanity_check(struct inode *inode,
127 struct ocfs2_extent_tree *et)
130 struct ocfs2_dinode *di;
132 BUG_ON(et->et_type != OCFS2_DINODE_EXTENT);
135 if (!OCFS2_IS_VALID_DINODE(di)) {
137 ocfs2_error(inode->i_sb,
138 "Inode %llu has invalid path root",
139 (unsigned long long)OCFS2_I(inode)->ip_blkno);
145 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
146 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
147 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
148 .eo_update_clusters = ocfs2_dinode_update_clusters,
149 .eo_sanity_check = ocfs2_dinode_sanity_check,
150 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
153 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
155 struct ocfs2_xattr_value_root *xv = et->et_object;
157 et->et_root_el = &xv->xr_list;
160 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
163 struct ocfs2_xattr_value_root *xv =
164 (struct ocfs2_xattr_value_root *)et->et_object;
166 xv->xr_last_eb_blk = cpu_to_le64(blkno);
169 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
171 struct ocfs2_xattr_value_root *xv =
172 (struct ocfs2_xattr_value_root *) et->et_object;
174 return le64_to_cpu(xv->xr_last_eb_blk);
177 static void ocfs2_xattr_value_update_clusters(struct inode *inode,
178 struct ocfs2_extent_tree *et,
181 struct ocfs2_xattr_value_root *xv =
182 (struct ocfs2_xattr_value_root *)et->et_object;
184 le32_add_cpu(&xv->xr_clusters, clusters);
187 static int ocfs2_xattr_value_sanity_check(struct inode *inode,
188 struct ocfs2_extent_tree *et)
193 static struct ocfs2_extent_tree_operations ocfs2_xattr_et_ops = {
194 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
195 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
196 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
197 .eo_sanity_check = ocfs2_xattr_value_sanity_check,
198 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
201 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
203 struct ocfs2_xattr_block *xb = et->et_object;
205 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
208 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
211 struct ocfs2_xattr_block *xb = et->et_object;
212 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
214 xt->xt_last_eb_blk = cpu_to_le64(blkno);
217 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
219 struct ocfs2_xattr_block *xb = et->et_object;
220 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
222 return le64_to_cpu(xt->xt_last_eb_blk);
225 static void ocfs2_xattr_tree_update_clusters(struct inode *inode,
226 struct ocfs2_extent_tree *et,
229 struct ocfs2_xattr_block *xb = et->et_object;
231 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
234 static int ocfs2_xattr_tree_sanity_check(struct inode *inode,
235 struct ocfs2_extent_tree *et)
240 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
241 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
242 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
243 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
244 .eo_sanity_check = ocfs2_xattr_tree_sanity_check,
245 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
248 static void ocfs2_get_extent_tree(struct ocfs2_extent_tree *et,
250 struct buffer_head *bh,
251 enum ocfs2_extent_tree_type et_type,
254 et->et_type = et_type;
257 et->et_max_leaf_clusters = 0;
259 obj = (void *)bh->b_data;
262 if (et_type == OCFS2_DINODE_EXTENT) {
263 et->et_ops = &ocfs2_dinode_et_ops;
264 } else if (et_type == OCFS2_XATTR_VALUE_EXTENT) {
265 et->et_ops = &ocfs2_xattr_et_ops;
266 } else if (et_type == OCFS2_XATTR_TREE_EXTENT) {
267 et->et_ops = &ocfs2_xattr_tree_et_ops;
268 et->et_max_leaf_clusters = ocfs2_clusters_for_bytes(inode->i_sb,
269 OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
272 et->et_ops->eo_fill_root_el(et);
275 static void ocfs2_put_extent_tree(struct ocfs2_extent_tree *et)
277 brelse(et->et_root_bh);
280 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
283 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
286 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
288 return et->et_ops->eo_get_last_eb_blk(et);
291 static inline void ocfs2_et_update_clusters(struct inode *inode,
292 struct ocfs2_extent_tree *et,
295 et->et_ops->eo_update_clusters(inode, et, clusters);
298 static inline int ocfs2_et_sanity_check(struct inode *inode,
299 struct ocfs2_extent_tree *et)
301 return et->et_ops->eo_sanity_check(inode, et);
304 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
305 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
306 struct ocfs2_extent_block *eb);
309 * Structures which describe a path through a btree, and functions to
312 * The idea here is to be as generic as possible with the tree
315 struct ocfs2_path_item {
316 struct buffer_head *bh;
317 struct ocfs2_extent_list *el;
320 #define OCFS2_MAX_PATH_DEPTH 5
324 struct ocfs2_path_item p_node[OCFS2_MAX_PATH_DEPTH];
327 #define path_root_bh(_path) ((_path)->p_node[0].bh)
328 #define path_root_el(_path) ((_path)->p_node[0].el)
329 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
330 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
331 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
334 * Reset the actual path elements so that we can re-use the structure
335 * to build another path. Generally, this involves freeing the buffer
338 static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
340 int i, start = 0, depth = 0;
341 struct ocfs2_path_item *node;
346 for(i = start; i < path_num_items(path); i++) {
347 node = &path->p_node[i];
355 * Tree depth may change during truncate, or insert. If we're
356 * keeping the root extent list, then make sure that our path
357 * structure reflects the proper depth.
360 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
362 path->p_tree_depth = depth;
365 static void ocfs2_free_path(struct ocfs2_path *path)
368 ocfs2_reinit_path(path, 0);
374 * All the elements of src into dest. After this call, src could be freed
375 * without affecting dest.
377 * Both paths should have the same root. Any non-root elements of dest
380 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
384 BUG_ON(path_root_bh(dest) != path_root_bh(src));
385 BUG_ON(path_root_el(dest) != path_root_el(src));
387 ocfs2_reinit_path(dest, 1);
389 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
390 dest->p_node[i].bh = src->p_node[i].bh;
391 dest->p_node[i].el = src->p_node[i].el;
393 if (dest->p_node[i].bh)
394 get_bh(dest->p_node[i].bh);
399 * Make the *dest path the same as src and re-initialize src path to
402 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
406 BUG_ON(path_root_bh(dest) != path_root_bh(src));
408 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
409 brelse(dest->p_node[i].bh);
411 dest->p_node[i].bh = src->p_node[i].bh;
412 dest->p_node[i].el = src->p_node[i].el;
414 src->p_node[i].bh = NULL;
415 src->p_node[i].el = NULL;
420 * Insert an extent block at given index.
422 * This will not take an additional reference on eb_bh.
424 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
425 struct buffer_head *eb_bh)
427 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
430 * Right now, no root bh is an extent block, so this helps
431 * catch code errors with dinode trees. The assertion can be
432 * safely removed if we ever need to insert extent block
433 * structures at the root.
437 path->p_node[index].bh = eb_bh;
438 path->p_node[index].el = &eb->h_list;
441 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
442 struct ocfs2_extent_list *root_el)
444 struct ocfs2_path *path;
446 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
448 path = kzalloc(sizeof(*path), GFP_NOFS);
450 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
452 path_root_bh(path) = root_bh;
453 path_root_el(path) = root_el;
460 * Convenience function to journal all components in a path.
462 static int ocfs2_journal_access_path(struct inode *inode, handle_t *handle,
463 struct ocfs2_path *path)
470 for(i = 0; i < path_num_items(path); i++) {
471 ret = ocfs2_journal_access(handle, inode, path->p_node[i].bh,
472 OCFS2_JOURNAL_ACCESS_WRITE);
484 * Return the index of the extent record which contains cluster #v_cluster.
485 * -1 is returned if it was not found.
487 * Should work fine on interior and exterior nodes.
489 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
493 struct ocfs2_extent_rec *rec;
494 u32 rec_end, rec_start, clusters;
496 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
497 rec = &el->l_recs[i];
499 rec_start = le32_to_cpu(rec->e_cpos);
500 clusters = ocfs2_rec_clusters(el, rec);
502 rec_end = rec_start + clusters;
504 if (v_cluster >= rec_start && v_cluster < rec_end) {
513 enum ocfs2_contig_type {
522 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
523 * ocfs2_extent_contig only work properly against leaf nodes!
525 static int ocfs2_block_extent_contig(struct super_block *sb,
526 struct ocfs2_extent_rec *ext,
529 u64 blk_end = le64_to_cpu(ext->e_blkno);
531 blk_end += ocfs2_clusters_to_blocks(sb,
532 le16_to_cpu(ext->e_leaf_clusters));
534 return blkno == blk_end;
537 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
538 struct ocfs2_extent_rec *right)
542 left_range = le32_to_cpu(left->e_cpos) +
543 le16_to_cpu(left->e_leaf_clusters);
545 return (left_range == le32_to_cpu(right->e_cpos));
548 static enum ocfs2_contig_type
549 ocfs2_extent_contig(struct inode *inode,
550 struct ocfs2_extent_rec *ext,
551 struct ocfs2_extent_rec *insert_rec)
553 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
556 * Refuse to coalesce extent records with different flag
557 * fields - we don't want to mix unwritten extents with user
560 if (ext->e_flags != insert_rec->e_flags)
563 if (ocfs2_extents_adjacent(ext, insert_rec) &&
564 ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
567 blkno = le64_to_cpu(ext->e_blkno);
568 if (ocfs2_extents_adjacent(insert_rec, ext) &&
569 ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno))
576 * NOTE: We can have pretty much any combination of contiguousness and
579 * The usefulness of APPEND_TAIL is more in that it lets us know that
580 * we'll have to update the path to that leaf.
582 enum ocfs2_append_type {
587 enum ocfs2_split_type {
593 struct ocfs2_insert_type {
594 enum ocfs2_split_type ins_split;
595 enum ocfs2_append_type ins_appending;
596 enum ocfs2_contig_type ins_contig;
597 int ins_contig_index;
601 struct ocfs2_merge_ctxt {
602 enum ocfs2_contig_type c_contig_type;
603 int c_has_empty_extent;
604 int c_split_covers_rec;
608 * How many free extents have we got before we need more meta data?
610 int ocfs2_num_free_extents(struct ocfs2_super *osb,
612 struct buffer_head *root_bh,
613 enum ocfs2_extent_tree_type type,
617 struct ocfs2_extent_list *el = NULL;
618 struct ocfs2_extent_block *eb;
619 struct buffer_head *eb_bh = NULL;
624 if (type == OCFS2_DINODE_EXTENT) {
625 struct ocfs2_dinode *fe =
626 (struct ocfs2_dinode *)root_bh->b_data;
627 if (!OCFS2_IS_VALID_DINODE(fe)) {
628 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
633 if (fe->i_last_eb_blk)
634 last_eb_blk = le64_to_cpu(fe->i_last_eb_blk);
635 el = &fe->id2.i_list;
636 } else if (type == OCFS2_XATTR_VALUE_EXTENT) {
637 struct ocfs2_xattr_value_root *xv =
638 (struct ocfs2_xattr_value_root *) obj;
640 last_eb_blk = le64_to_cpu(xv->xr_last_eb_blk);
642 } else if (type == OCFS2_XATTR_TREE_EXTENT) {
643 struct ocfs2_xattr_block *xb =
644 (struct ocfs2_xattr_block *)root_bh->b_data;
646 last_eb_blk = le64_to_cpu(xb->xb_attrs.xb_root.xt_last_eb_blk);
647 el = &xb->xb_attrs.xb_root.xt_list;
651 retval = ocfs2_read_block(osb, last_eb_blk,
652 &eb_bh, OCFS2_BH_CACHED, inode);
657 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
661 BUG_ON(el->l_tree_depth != 0);
663 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
672 /* expects array to already be allocated
674 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
677 static int ocfs2_create_new_meta_bhs(struct ocfs2_super *osb,
681 struct ocfs2_alloc_context *meta_ac,
682 struct buffer_head *bhs[])
684 int count, status, i;
685 u16 suballoc_bit_start;
688 struct ocfs2_extent_block *eb;
693 while (count < wanted) {
694 status = ocfs2_claim_metadata(osb,
706 for(i = count; i < (num_got + count); i++) {
707 bhs[i] = sb_getblk(osb->sb, first_blkno);
708 if (bhs[i] == NULL) {
713 ocfs2_set_new_buffer_uptodate(inode, bhs[i]);
715 status = ocfs2_journal_access(handle, inode, bhs[i],
716 OCFS2_JOURNAL_ACCESS_CREATE);
722 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
723 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
724 /* Ok, setup the minimal stuff here. */
725 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
726 eb->h_blkno = cpu_to_le64(first_blkno);
727 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
728 eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
729 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
731 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
733 suballoc_bit_start++;
736 /* We'll also be dirtied by the caller, so
737 * this isn't absolutely necessary. */
738 status = ocfs2_journal_dirty(handle, bhs[i]);
751 for(i = 0; i < wanted; i++) {
762 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
764 * Returns the sum of the rightmost extent rec logical offset and
767 * ocfs2_add_branch() uses this to determine what logical cluster
768 * value should be populated into the leftmost new branch records.
770 * ocfs2_shift_tree_depth() uses this to determine the # clusters
771 * value for the new topmost tree record.
773 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
777 i = le16_to_cpu(el->l_next_free_rec) - 1;
779 return le32_to_cpu(el->l_recs[i].e_cpos) +
780 ocfs2_rec_clusters(el, &el->l_recs[i]);
784 * Add an entire tree branch to our inode. eb_bh is the extent block
785 * to start at, if we don't want to start the branch at the dinode
788 * last_eb_bh is required as we have to update it's next_leaf pointer
789 * for the new last extent block.
791 * the new branch will be 'empty' in the sense that every block will
792 * contain a single record with cluster count == 0.
794 static int ocfs2_add_branch(struct ocfs2_super *osb,
797 struct ocfs2_extent_tree *et,
798 struct buffer_head *eb_bh,
799 struct buffer_head **last_eb_bh,
800 struct ocfs2_alloc_context *meta_ac)
802 int status, new_blocks, i;
803 u64 next_blkno, new_last_eb_blk;
804 struct buffer_head *bh;
805 struct buffer_head **new_eb_bhs = NULL;
806 struct ocfs2_extent_block *eb;
807 struct ocfs2_extent_list *eb_el;
808 struct ocfs2_extent_list *el;
813 BUG_ON(!last_eb_bh || !*last_eb_bh);
816 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
821 /* we never add a branch to a leaf. */
822 BUG_ON(!el->l_tree_depth);
824 new_blocks = le16_to_cpu(el->l_tree_depth);
826 /* allocate the number of new eb blocks we need */
827 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
835 status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
836 meta_ac, new_eb_bhs);
842 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
843 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
845 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
846 * linked with the rest of the tree.
847 * conversly, new_eb_bhs[0] is the new bottommost leaf.
849 * when we leave the loop, new_last_eb_blk will point to the
850 * newest leaf, and next_blkno will point to the topmost extent
852 next_blkno = new_last_eb_blk = 0;
853 for(i = 0; i < new_blocks; i++) {
855 eb = (struct ocfs2_extent_block *) bh->b_data;
856 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
857 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
863 status = ocfs2_journal_access(handle, inode, bh,
864 OCFS2_JOURNAL_ACCESS_CREATE);
870 eb->h_next_leaf_blk = 0;
871 eb_el->l_tree_depth = cpu_to_le16(i);
872 eb_el->l_next_free_rec = cpu_to_le16(1);
874 * This actually counts as an empty extent as
877 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
878 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
880 * eb_el isn't always an interior node, but even leaf
881 * nodes want a zero'd flags and reserved field so
882 * this gets the whole 32 bits regardless of use.
884 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
885 if (!eb_el->l_tree_depth)
886 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
888 status = ocfs2_journal_dirty(handle, bh);
894 next_blkno = le64_to_cpu(eb->h_blkno);
897 /* This is a bit hairy. We want to update up to three blocks
898 * here without leaving any of them in an inconsistent state
899 * in case of error. We don't have to worry about
900 * journal_dirty erroring as it won't unless we've aborted the
901 * handle (in which case we would never be here) so reserving
902 * the write with journal_access is all we need to do. */
903 status = ocfs2_journal_access(handle, inode, *last_eb_bh,
904 OCFS2_JOURNAL_ACCESS_WRITE);
909 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
910 OCFS2_JOURNAL_ACCESS_WRITE);
916 status = ocfs2_journal_access(handle, inode, eb_bh,
917 OCFS2_JOURNAL_ACCESS_WRITE);
924 /* Link the new branch into the rest of the tree (el will
925 * either be on the root_bh, or the extent block passed in. */
926 i = le16_to_cpu(el->l_next_free_rec);
927 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
928 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
929 el->l_recs[i].e_int_clusters = 0;
930 le16_add_cpu(&el->l_next_free_rec, 1);
932 /* fe needs a new last extent block pointer, as does the
933 * next_leaf on the previously last-extent-block. */
934 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
936 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
937 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
939 status = ocfs2_journal_dirty(handle, *last_eb_bh);
942 status = ocfs2_journal_dirty(handle, et->et_root_bh);
946 status = ocfs2_journal_dirty(handle, eb_bh);
952 * Some callers want to track the rightmost leaf so pass it
956 get_bh(new_eb_bhs[0]);
957 *last_eb_bh = new_eb_bhs[0];
962 for (i = 0; i < new_blocks; i++)
964 brelse(new_eb_bhs[i]);
973 * adds another level to the allocation tree.
974 * returns back the new extent block so you can add a branch to it
977 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
980 struct ocfs2_extent_tree *et,
981 struct ocfs2_alloc_context *meta_ac,
982 struct buffer_head **ret_new_eb_bh)
986 struct buffer_head *new_eb_bh = NULL;
987 struct ocfs2_extent_block *eb;
988 struct ocfs2_extent_list *root_el;
989 struct ocfs2_extent_list *eb_el;
993 status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
1000 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1001 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1002 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1007 eb_el = &eb->h_list;
1008 root_el = et->et_root_el;
1010 status = ocfs2_journal_access(handle, inode, new_eb_bh,
1011 OCFS2_JOURNAL_ACCESS_CREATE);
1017 /* copy the root extent list data into the new extent block */
1018 eb_el->l_tree_depth = root_el->l_tree_depth;
1019 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1020 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1021 eb_el->l_recs[i] = root_el->l_recs[i];
1023 status = ocfs2_journal_dirty(handle, new_eb_bh);
1029 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
1030 OCFS2_JOURNAL_ACCESS_WRITE);
1036 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1038 /* update root_bh now */
1039 le16_add_cpu(&root_el->l_tree_depth, 1);
1040 root_el->l_recs[0].e_cpos = 0;
1041 root_el->l_recs[0].e_blkno = eb->h_blkno;
1042 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1043 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1044 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1045 root_el->l_next_free_rec = cpu_to_le16(1);
1047 /* If this is our 1st tree depth shift, then last_eb_blk
1048 * becomes the allocated extent block */
1049 if (root_el->l_tree_depth == cpu_to_le16(1))
1050 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1052 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1058 *ret_new_eb_bh = new_eb_bh;
1070 * Should only be called when there is no space left in any of the
1071 * leaf nodes. What we want to do is find the lowest tree depth
1072 * non-leaf extent block with room for new records. There are three
1073 * valid results of this search:
1075 * 1) a lowest extent block is found, then we pass it back in
1076 * *lowest_eb_bh and return '0'
1078 * 2) the search fails to find anything, but the root_el has room. We
1079 * pass NULL back in *lowest_eb_bh, but still return '0'
1081 * 3) the search fails to find anything AND the root_el is full, in
1082 * which case we return > 0
1084 * return status < 0 indicates an error.
1086 static int ocfs2_find_branch_target(struct ocfs2_super *osb,
1087 struct inode *inode,
1088 struct ocfs2_extent_tree *et,
1089 struct buffer_head **target_bh)
1093 struct ocfs2_extent_block *eb;
1094 struct ocfs2_extent_list *el;
1095 struct buffer_head *bh = NULL;
1096 struct buffer_head *lowest_bh = NULL;
1102 el = et->et_root_el;
1104 while(le16_to_cpu(el->l_tree_depth) > 1) {
1105 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1106 ocfs2_error(inode->i_sb, "Dinode %llu has empty "
1107 "extent list (next_free_rec == 0)",
1108 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1112 i = le16_to_cpu(el->l_next_free_rec) - 1;
1113 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1115 ocfs2_error(inode->i_sb, "Dinode %llu has extent "
1116 "list where extent # %d has no physical "
1118 (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
1128 status = ocfs2_read_block(osb, blkno, &bh, OCFS2_BH_CACHED,
1135 eb = (struct ocfs2_extent_block *) bh->b_data;
1136 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1137 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1143 if (le16_to_cpu(el->l_next_free_rec) <
1144 le16_to_cpu(el->l_count)) {
1152 /* If we didn't find one and the fe doesn't have any room,
1153 * then return '1' */
1154 el = et->et_root_el;
1155 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1158 *target_bh = lowest_bh;
1168 * Grow a b-tree so that it has more records.
1170 * We might shift the tree depth in which case existing paths should
1171 * be considered invalid.
1173 * Tree depth after the grow is returned via *final_depth.
1175 * *last_eb_bh will be updated by ocfs2_add_branch().
1177 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
1178 struct ocfs2_extent_tree *et, int *final_depth,
1179 struct buffer_head **last_eb_bh,
1180 struct ocfs2_alloc_context *meta_ac)
1183 struct ocfs2_extent_list *el = et->et_root_el;
1184 int depth = le16_to_cpu(el->l_tree_depth);
1185 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1186 struct buffer_head *bh = NULL;
1188 BUG_ON(meta_ac == NULL);
1190 shift = ocfs2_find_branch_target(osb, inode, et, &bh);
1197 /* We traveled all the way to the bottom of the allocation tree
1198 * and didn't find room for any more extents - we need to add
1199 * another tree level */
1202 mlog(0, "need to shift tree depth (current = %d)\n", depth);
1204 /* ocfs2_shift_tree_depth will return us a buffer with
1205 * the new extent block (so we can pass that to
1206 * ocfs2_add_branch). */
1207 ret = ocfs2_shift_tree_depth(osb, handle, inode, et,
1216 * Special case: we have room now if we shifted from
1217 * tree_depth 0, so no more work needs to be done.
1219 * We won't be calling add_branch, so pass
1220 * back *last_eb_bh as the new leaf. At depth
1221 * zero, it should always be null so there's
1222 * no reason to brelse.
1224 BUG_ON(*last_eb_bh);
1231 /* call ocfs2_add_branch to add the final part of the tree with
1233 mlog(0, "add branch. bh = %p\n", bh);
1234 ret = ocfs2_add_branch(osb, handle, inode, et, bh, last_eb_bh,
1243 *final_depth = depth;
1249 * This function will discard the rightmost extent record.
1251 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1253 int next_free = le16_to_cpu(el->l_next_free_rec);
1254 int count = le16_to_cpu(el->l_count);
1255 unsigned int num_bytes;
1258 /* This will cause us to go off the end of our extent list. */
1259 BUG_ON(next_free >= count);
1261 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1263 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1266 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1267 struct ocfs2_extent_rec *insert_rec)
1269 int i, insert_index, next_free, has_empty, num_bytes;
1270 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1271 struct ocfs2_extent_rec *rec;
1273 next_free = le16_to_cpu(el->l_next_free_rec);
1274 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1278 /* The tree code before us didn't allow enough room in the leaf. */
1279 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1282 * The easiest way to approach this is to just remove the
1283 * empty extent and temporarily decrement next_free.
1287 * If next_free was 1 (only an empty extent), this
1288 * loop won't execute, which is fine. We still want
1289 * the decrement above to happen.
1291 for(i = 0; i < (next_free - 1); i++)
1292 el->l_recs[i] = el->l_recs[i+1];
1298 * Figure out what the new record index should be.
1300 for(i = 0; i < next_free; i++) {
1301 rec = &el->l_recs[i];
1303 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1308 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1309 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
1311 BUG_ON(insert_index < 0);
1312 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1313 BUG_ON(insert_index > next_free);
1316 * No need to memmove if we're just adding to the tail.
1318 if (insert_index != next_free) {
1319 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1321 num_bytes = next_free - insert_index;
1322 num_bytes *= sizeof(struct ocfs2_extent_rec);
1323 memmove(&el->l_recs[insert_index + 1],
1324 &el->l_recs[insert_index],
1329 * Either we had an empty extent, and need to re-increment or
1330 * there was no empty extent on a non full rightmost leaf node,
1331 * in which case we still need to increment.
1334 el->l_next_free_rec = cpu_to_le16(next_free);
1336 * Make sure none of the math above just messed up our tree.
1338 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1340 el->l_recs[insert_index] = *insert_rec;
1344 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1346 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1348 BUG_ON(num_recs == 0);
1350 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1352 size = num_recs * sizeof(struct ocfs2_extent_rec);
1353 memmove(&el->l_recs[0], &el->l_recs[1], size);
1354 memset(&el->l_recs[num_recs], 0,
1355 sizeof(struct ocfs2_extent_rec));
1356 el->l_next_free_rec = cpu_to_le16(num_recs);
1361 * Create an empty extent record .
1363 * l_next_free_rec may be updated.
1365 * If an empty extent already exists do nothing.
1367 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1369 int next_free = le16_to_cpu(el->l_next_free_rec);
1371 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1376 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1379 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1380 "Asked to create an empty extent in a full list:\n"
1381 "count = %u, tree depth = %u",
1382 le16_to_cpu(el->l_count),
1383 le16_to_cpu(el->l_tree_depth));
1385 ocfs2_shift_records_right(el);
1388 le16_add_cpu(&el->l_next_free_rec, 1);
1389 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1393 * For a rotation which involves two leaf nodes, the "root node" is
1394 * the lowest level tree node which contains a path to both leafs. This
1395 * resulting set of information can be used to form a complete "subtree"
1397 * This function is passed two full paths from the dinode down to a
1398 * pair of adjacent leaves. It's task is to figure out which path
1399 * index contains the subtree root - this can be the root index itself
1400 * in a worst-case rotation.
1402 * The array index of the subtree root is passed back.
1404 static int ocfs2_find_subtree_root(struct inode *inode,
1405 struct ocfs2_path *left,
1406 struct ocfs2_path *right)
1411 * Check that the caller passed in two paths from the same tree.
1413 BUG_ON(path_root_bh(left) != path_root_bh(right));
1419 * The caller didn't pass two adjacent paths.
1421 mlog_bug_on_msg(i > left->p_tree_depth,
1422 "Inode %lu, left depth %u, right depth %u\n"
1423 "left leaf blk %llu, right leaf blk %llu\n",
1424 inode->i_ino, left->p_tree_depth,
1425 right->p_tree_depth,
1426 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1427 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1428 } while (left->p_node[i].bh->b_blocknr ==
1429 right->p_node[i].bh->b_blocknr);
1434 typedef void (path_insert_t)(void *, struct buffer_head *);
1437 * Traverse a btree path in search of cpos, starting at root_el.
1439 * This code can be called with a cpos larger than the tree, in which
1440 * case it will return the rightmost path.
1442 static int __ocfs2_find_path(struct inode *inode,
1443 struct ocfs2_extent_list *root_el, u32 cpos,
1444 path_insert_t *func, void *data)
1449 struct buffer_head *bh = NULL;
1450 struct ocfs2_extent_block *eb;
1451 struct ocfs2_extent_list *el;
1452 struct ocfs2_extent_rec *rec;
1453 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1456 while (el->l_tree_depth) {
1457 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1458 ocfs2_error(inode->i_sb,
1459 "Inode %llu has empty extent list at "
1461 (unsigned long long)oi->ip_blkno,
1462 le16_to_cpu(el->l_tree_depth));
1468 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1469 rec = &el->l_recs[i];
1472 * In the case that cpos is off the allocation
1473 * tree, this should just wind up returning the
1476 range = le32_to_cpu(rec->e_cpos) +
1477 ocfs2_rec_clusters(el, rec);
1478 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1482 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1484 ocfs2_error(inode->i_sb,
1485 "Inode %llu has bad blkno in extent list "
1486 "at depth %u (index %d)\n",
1487 (unsigned long long)oi->ip_blkno,
1488 le16_to_cpu(el->l_tree_depth), i);
1495 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno,
1496 &bh, OCFS2_BH_CACHED, inode);
1502 eb = (struct ocfs2_extent_block *) bh->b_data;
1504 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1505 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1510 if (le16_to_cpu(el->l_next_free_rec) >
1511 le16_to_cpu(el->l_count)) {
1512 ocfs2_error(inode->i_sb,
1513 "Inode %llu has bad count in extent list "
1514 "at block %llu (next free=%u, count=%u)\n",
1515 (unsigned long long)oi->ip_blkno,
1516 (unsigned long long)bh->b_blocknr,
1517 le16_to_cpu(el->l_next_free_rec),
1518 le16_to_cpu(el->l_count));
1529 * Catch any trailing bh that the loop didn't handle.
1537 * Given an initialized path (that is, it has a valid root extent
1538 * list), this function will traverse the btree in search of the path
1539 * which would contain cpos.
1541 * The path traveled is recorded in the path structure.
1543 * Note that this will not do any comparisons on leaf node extent
1544 * records, so it will work fine in the case that we just added a tree
1547 struct find_path_data {
1549 struct ocfs2_path *path;
1551 static void find_path_ins(void *data, struct buffer_head *bh)
1553 struct find_path_data *fp = data;
1556 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1559 static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
1562 struct find_path_data data;
1566 return __ocfs2_find_path(inode, path_root_el(path), cpos,
1567 find_path_ins, &data);
1570 static void find_leaf_ins(void *data, struct buffer_head *bh)
1572 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1573 struct ocfs2_extent_list *el = &eb->h_list;
1574 struct buffer_head **ret = data;
1576 /* We want to retain only the leaf block. */
1577 if (le16_to_cpu(el->l_tree_depth) == 0) {
1583 * Find the leaf block in the tree which would contain cpos. No
1584 * checking of the actual leaf is done.
1586 * Some paths want to call this instead of allocating a path structure
1587 * and calling ocfs2_find_path().
1589 * This function doesn't handle non btree extent lists.
1591 int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
1592 u32 cpos, struct buffer_head **leaf_bh)
1595 struct buffer_head *bh = NULL;
1597 ret = __ocfs2_find_path(inode, root_el, cpos, find_leaf_ins, &bh);
1609 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1611 * Basically, we've moved stuff around at the bottom of the tree and
1612 * we need to fix up the extent records above the changes to reflect
1615 * left_rec: the record on the left.
1616 * left_child_el: is the child list pointed to by left_rec
1617 * right_rec: the record to the right of left_rec
1618 * right_child_el: is the child list pointed to by right_rec
1620 * By definition, this only works on interior nodes.
1622 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1623 struct ocfs2_extent_list *left_child_el,
1624 struct ocfs2_extent_rec *right_rec,
1625 struct ocfs2_extent_list *right_child_el)
1627 u32 left_clusters, right_end;
1630 * Interior nodes never have holes. Their cpos is the cpos of
1631 * the leftmost record in their child list. Their cluster
1632 * count covers the full theoretical range of their child list
1633 * - the range between their cpos and the cpos of the record
1634 * immediately to their right.
1636 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1637 if (ocfs2_is_empty_extent(&right_child_el->l_recs[0])) {
1638 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1639 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1641 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1642 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1645 * Calculate the rightmost cluster count boundary before
1646 * moving cpos - we will need to adjust clusters after
1647 * updating e_cpos to keep the same highest cluster count.
1649 right_end = le32_to_cpu(right_rec->e_cpos);
1650 right_end += le32_to_cpu(right_rec->e_int_clusters);
1652 right_rec->e_cpos = left_rec->e_cpos;
1653 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1655 right_end -= le32_to_cpu(right_rec->e_cpos);
1656 right_rec->e_int_clusters = cpu_to_le32(right_end);
1660 * Adjust the adjacent root node records involved in a
1661 * rotation. left_el_blkno is passed in as a key so that we can easily
1662 * find it's index in the root list.
1664 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1665 struct ocfs2_extent_list *left_el,
1666 struct ocfs2_extent_list *right_el,
1671 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1672 le16_to_cpu(left_el->l_tree_depth));
1674 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1675 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1680 * The path walking code should have never returned a root and
1681 * two paths which are not adjacent.
1683 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1685 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1686 &root_el->l_recs[i + 1], right_el);
1690 * We've changed a leaf block (in right_path) and need to reflect that
1691 * change back up the subtree.
1693 * This happens in multiple places:
1694 * - When we've moved an extent record from the left path leaf to the right
1695 * path leaf to make room for an empty extent in the left path leaf.
1696 * - When our insert into the right path leaf is at the leftmost edge
1697 * and requires an update of the path immediately to it's left. This
1698 * can occur at the end of some types of rotation and appending inserts.
1699 * - When we've adjusted the last extent record in the left path leaf and the
1700 * 1st extent record in the right path leaf during cross extent block merge.
1702 static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1703 struct ocfs2_path *left_path,
1704 struct ocfs2_path *right_path,
1708 struct ocfs2_extent_list *el, *left_el, *right_el;
1709 struct ocfs2_extent_rec *left_rec, *right_rec;
1710 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1713 * Update the counts and position values within all the
1714 * interior nodes to reflect the leaf rotation we just did.
1716 * The root node is handled below the loop.
1718 * We begin the loop with right_el and left_el pointing to the
1719 * leaf lists and work our way up.
1721 * NOTE: within this loop, left_el and right_el always refer
1722 * to the *child* lists.
1724 left_el = path_leaf_el(left_path);
1725 right_el = path_leaf_el(right_path);
1726 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
1727 mlog(0, "Adjust records at index %u\n", i);
1730 * One nice property of knowing that all of these
1731 * nodes are below the root is that we only deal with
1732 * the leftmost right node record and the rightmost
1735 el = left_path->p_node[i].el;
1736 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
1737 left_rec = &el->l_recs[idx];
1739 el = right_path->p_node[i].el;
1740 right_rec = &el->l_recs[0];
1742 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
1745 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
1749 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
1754 * Setup our list pointers now so that the current
1755 * parents become children in the next iteration.
1757 left_el = left_path->p_node[i].el;
1758 right_el = right_path->p_node[i].el;
1762 * At the root node, adjust the two adjacent records which
1763 * begin our path to the leaves.
1766 el = left_path->p_node[subtree_index].el;
1767 left_el = left_path->p_node[subtree_index + 1].el;
1768 right_el = right_path->p_node[subtree_index + 1].el;
1770 ocfs2_adjust_root_records(el, left_el, right_el,
1771 left_path->p_node[subtree_index + 1].bh->b_blocknr);
1773 root_bh = left_path->p_node[subtree_index].bh;
1775 ret = ocfs2_journal_dirty(handle, root_bh);
1780 static int ocfs2_rotate_subtree_right(struct inode *inode,
1782 struct ocfs2_path *left_path,
1783 struct ocfs2_path *right_path,
1787 struct buffer_head *right_leaf_bh;
1788 struct buffer_head *left_leaf_bh = NULL;
1789 struct buffer_head *root_bh;
1790 struct ocfs2_extent_list *right_el, *left_el;
1791 struct ocfs2_extent_rec move_rec;
1793 left_leaf_bh = path_leaf_bh(left_path);
1794 left_el = path_leaf_el(left_path);
1796 if (left_el->l_next_free_rec != left_el->l_count) {
1797 ocfs2_error(inode->i_sb,
1798 "Inode %llu has non-full interior leaf node %llu"
1800 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1801 (unsigned long long)left_leaf_bh->b_blocknr,
1802 le16_to_cpu(left_el->l_next_free_rec));
1807 * This extent block may already have an empty record, so we
1808 * return early if so.
1810 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
1813 root_bh = left_path->p_node[subtree_index].bh;
1814 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
1816 ret = ocfs2_journal_access(handle, inode, root_bh,
1817 OCFS2_JOURNAL_ACCESS_WRITE);
1823 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
1824 ret = ocfs2_journal_access(handle, inode,
1825 right_path->p_node[i].bh,
1826 OCFS2_JOURNAL_ACCESS_WRITE);
1832 ret = ocfs2_journal_access(handle, inode,
1833 left_path->p_node[i].bh,
1834 OCFS2_JOURNAL_ACCESS_WRITE);
1841 right_leaf_bh = path_leaf_bh(right_path);
1842 right_el = path_leaf_el(right_path);
1844 /* This is a code error, not a disk corruption. */
1845 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
1846 "because rightmost leaf block %llu is empty\n",
1847 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1848 (unsigned long long)right_leaf_bh->b_blocknr);
1850 ocfs2_create_empty_extent(right_el);
1852 ret = ocfs2_journal_dirty(handle, right_leaf_bh);
1858 /* Do the copy now. */
1859 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
1860 move_rec = left_el->l_recs[i];
1861 right_el->l_recs[0] = move_rec;
1864 * Clear out the record we just copied and shift everything
1865 * over, leaving an empty extent in the left leaf.
1867 * We temporarily subtract from next_free_rec so that the
1868 * shift will lose the tail record (which is now defunct).
1870 le16_add_cpu(&left_el->l_next_free_rec, -1);
1871 ocfs2_shift_records_right(left_el);
1872 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1873 le16_add_cpu(&left_el->l_next_free_rec, 1);
1875 ret = ocfs2_journal_dirty(handle, left_leaf_bh);
1881 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
1889 * Given a full path, determine what cpos value would return us a path
1890 * containing the leaf immediately to the left of the current one.
1892 * Will return zero if the path passed in is already the leftmost path.
1894 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
1895 struct ocfs2_path *path, u32 *cpos)
1899 struct ocfs2_extent_list *el;
1901 BUG_ON(path->p_tree_depth == 0);
1905 blkno = path_leaf_bh(path)->b_blocknr;
1907 /* Start at the tree node just above the leaf and work our way up. */
1908 i = path->p_tree_depth - 1;
1910 el = path->p_node[i].el;
1913 * Find the extent record just before the one in our
1916 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
1917 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
1921 * We've determined that the
1922 * path specified is already
1923 * the leftmost one - return a
1929 * The leftmost record points to our
1930 * leaf - we need to travel up the
1936 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
1937 *cpos = *cpos + ocfs2_rec_clusters(el,
1938 &el->l_recs[j - 1]);
1945 * If we got here, we never found a valid node where
1946 * the tree indicated one should be.
1949 "Invalid extent tree at extent block %llu\n",
1950 (unsigned long long)blkno);
1955 blkno = path->p_node[i].bh->b_blocknr;
1964 * Extend the transaction by enough credits to complete the rotation,
1965 * and still leave at least the original number of credits allocated
1966 * to this transaction.
1968 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
1970 struct ocfs2_path *path)
1972 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
1974 if (handle->h_buffer_credits < credits)
1975 return ocfs2_extend_trans(handle, credits);
1981 * Trap the case where we're inserting into the theoretical range past
1982 * the _actual_ left leaf range. Otherwise, we'll rotate a record
1983 * whose cpos is less than ours into the right leaf.
1985 * It's only necessary to look at the rightmost record of the left
1986 * leaf because the logic that calls us should ensure that the
1987 * theoretical ranges in the path components above the leaves are
1990 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
1993 struct ocfs2_extent_list *left_el;
1994 struct ocfs2_extent_rec *rec;
1997 left_el = path_leaf_el(left_path);
1998 next_free = le16_to_cpu(left_el->l_next_free_rec);
1999 rec = &left_el->l_recs[next_free - 1];
2001 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2006 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2008 int next_free = le16_to_cpu(el->l_next_free_rec);
2010 struct ocfs2_extent_rec *rec;
2015 rec = &el->l_recs[0];
2016 if (ocfs2_is_empty_extent(rec)) {
2020 rec = &el->l_recs[1];
2023 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2024 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2030 * Rotate all the records in a btree right one record, starting at insert_cpos.
2032 * The path to the rightmost leaf should be passed in.
2034 * The array is assumed to be large enough to hold an entire path (tree depth).
2036 * Upon succesful return from this function:
2038 * - The 'right_path' array will contain a path to the leaf block
2039 * whose range contains e_cpos.
2040 * - That leaf block will have a single empty extent in list index 0.
2041 * - In the case that the rotation requires a post-insert update,
2042 * *ret_left_path will contain a valid path which can be passed to
2043 * ocfs2_insert_path().
2045 static int ocfs2_rotate_tree_right(struct inode *inode,
2047 enum ocfs2_split_type split,
2049 struct ocfs2_path *right_path,
2050 struct ocfs2_path **ret_left_path)
2052 int ret, start, orig_credits = handle->h_buffer_credits;
2054 struct ocfs2_path *left_path = NULL;
2056 *ret_left_path = NULL;
2058 left_path = ocfs2_new_path(path_root_bh(right_path),
2059 path_root_el(right_path));
2066 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
2072 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
2075 * What we want to do here is:
2077 * 1) Start with the rightmost path.
2079 * 2) Determine a path to the leaf block directly to the left
2082 * 3) Determine the 'subtree root' - the lowest level tree node
2083 * which contains a path to both leaves.
2085 * 4) Rotate the subtree.
2087 * 5) Find the next subtree by considering the left path to be
2088 * the new right path.
2090 * The check at the top of this while loop also accepts
2091 * insert_cpos == cpos because cpos is only a _theoretical_
2092 * value to get us the left path - insert_cpos might very well
2093 * be filling that hole.
2095 * Stop at a cpos of '0' because we either started at the
2096 * leftmost branch (i.e., a tree with one branch and a
2097 * rotation inside of it), or we've gone as far as we can in
2098 * rotating subtrees.
2100 while (cpos && insert_cpos <= cpos) {
2101 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2104 ret = ocfs2_find_path(inode, left_path, cpos);
2110 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2111 path_leaf_bh(right_path),
2112 "Inode %lu: error during insert of %u "
2113 "(left path cpos %u) results in two identical "
2114 "paths ending at %llu\n",
2115 inode->i_ino, insert_cpos, cpos,
2116 (unsigned long long)
2117 path_leaf_bh(left_path)->b_blocknr);
2119 if (split == SPLIT_NONE &&
2120 ocfs2_rotate_requires_path_adjustment(left_path,
2124 * We've rotated the tree as much as we
2125 * should. The rest is up to
2126 * ocfs2_insert_path() to complete, after the
2127 * record insertion. We indicate this
2128 * situation by returning the left path.
2130 * The reason we don't adjust the records here
2131 * before the record insert is that an error
2132 * later might break the rule where a parent
2133 * record e_cpos will reflect the actual
2134 * e_cpos of the 1st nonempty record of the
2137 *ret_left_path = left_path;
2141 start = ocfs2_find_subtree_root(inode, left_path, right_path);
2143 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2145 (unsigned long long) right_path->p_node[start].bh->b_blocknr,
2146 right_path->p_tree_depth);
2148 ret = ocfs2_extend_rotate_transaction(handle, start,
2149 orig_credits, right_path);
2155 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
2162 if (split != SPLIT_NONE &&
2163 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2166 * A rotate moves the rightmost left leaf
2167 * record over to the leftmost right leaf
2168 * slot. If we're doing an extent split
2169 * instead of a real insert, then we have to
2170 * check that the extent to be split wasn't
2171 * just moved over. If it was, then we can
2172 * exit here, passing left_path back -
2173 * ocfs2_split_extent() is smart enough to
2174 * search both leaves.
2176 *ret_left_path = left_path;
2181 * There is no need to re-read the next right path
2182 * as we know that it'll be our current left
2183 * path. Optimize by copying values instead.
2185 ocfs2_mv_path(right_path, left_path);
2187 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
2196 ocfs2_free_path(left_path);
2202 static void ocfs2_update_edge_lengths(struct inode *inode, handle_t *handle,
2203 struct ocfs2_path *path)
2206 struct ocfs2_extent_rec *rec;
2207 struct ocfs2_extent_list *el;
2208 struct ocfs2_extent_block *eb;
2211 /* Path should always be rightmost. */
2212 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2213 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2216 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2217 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2218 rec = &el->l_recs[idx];
2219 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2221 for (i = 0; i < path->p_tree_depth; i++) {
2222 el = path->p_node[i].el;
2223 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2224 rec = &el->l_recs[idx];
2226 rec->e_int_clusters = cpu_to_le32(range);
2227 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2229 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2233 static void ocfs2_unlink_path(struct inode *inode, handle_t *handle,
2234 struct ocfs2_cached_dealloc_ctxt *dealloc,
2235 struct ocfs2_path *path, int unlink_start)
2238 struct ocfs2_extent_block *eb;
2239 struct ocfs2_extent_list *el;
2240 struct buffer_head *bh;
2242 for(i = unlink_start; i < path_num_items(path); i++) {
2243 bh = path->p_node[i].bh;
2245 eb = (struct ocfs2_extent_block *)bh->b_data;
2247 * Not all nodes might have had their final count
2248 * decremented by the caller - handle this here.
2251 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2253 "Inode %llu, attempted to remove extent block "
2254 "%llu with %u records\n",
2255 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2256 (unsigned long long)le64_to_cpu(eb->h_blkno),
2257 le16_to_cpu(el->l_next_free_rec));
2259 ocfs2_journal_dirty(handle, bh);
2260 ocfs2_remove_from_cache(inode, bh);
2264 el->l_next_free_rec = 0;
2265 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2267 ocfs2_journal_dirty(handle, bh);
2269 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2273 ocfs2_remove_from_cache(inode, bh);
2277 static void ocfs2_unlink_subtree(struct inode *inode, handle_t *handle,
2278 struct ocfs2_path *left_path,
2279 struct ocfs2_path *right_path,
2281 struct ocfs2_cached_dealloc_ctxt *dealloc)
2284 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2285 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2286 struct ocfs2_extent_list *el;
2287 struct ocfs2_extent_block *eb;
2289 el = path_leaf_el(left_path);
2291 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2293 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2294 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2297 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2299 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2300 le16_add_cpu(&root_el->l_next_free_rec, -1);
2302 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2303 eb->h_next_leaf_blk = 0;
2305 ocfs2_journal_dirty(handle, root_bh);
2306 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2308 ocfs2_unlink_path(inode, handle, dealloc, right_path,
2312 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2313 struct ocfs2_path *left_path,
2314 struct ocfs2_path *right_path,
2316 struct ocfs2_cached_dealloc_ctxt *dealloc,
2318 struct ocfs2_extent_tree *et)
2320 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2321 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2322 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2323 struct ocfs2_extent_block *eb;
2327 right_leaf_el = path_leaf_el(right_path);
2328 left_leaf_el = path_leaf_el(left_path);
2329 root_bh = left_path->p_node[subtree_index].bh;
2330 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2332 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2335 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2336 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2338 * It's legal for us to proceed if the right leaf is
2339 * the rightmost one and it has an empty extent. There
2340 * are two cases to handle - whether the leaf will be
2341 * empty after removal or not. If the leaf isn't empty
2342 * then just remove the empty extent up front. The
2343 * next block will handle empty leaves by flagging
2346 * Non rightmost leaves will throw -EAGAIN and the
2347 * caller can manually move the subtree and retry.
2350 if (eb->h_next_leaf_blk != 0ULL)
2353 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2354 ret = ocfs2_journal_access(handle, inode,
2355 path_leaf_bh(right_path),
2356 OCFS2_JOURNAL_ACCESS_WRITE);
2362 ocfs2_remove_empty_extent(right_leaf_el);
2364 right_has_empty = 1;
2367 if (eb->h_next_leaf_blk == 0ULL &&
2368 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2370 * We have to update i_last_eb_blk during the meta
2373 ret = ocfs2_journal_access(handle, inode, et_root_bh,
2374 OCFS2_JOURNAL_ACCESS_WRITE);
2380 del_right_subtree = 1;
2384 * Getting here with an empty extent in the right path implies
2385 * that it's the rightmost path and will be deleted.
2387 BUG_ON(right_has_empty && !del_right_subtree);
2389 ret = ocfs2_journal_access(handle, inode, root_bh,
2390 OCFS2_JOURNAL_ACCESS_WRITE);
2396 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2397 ret = ocfs2_journal_access(handle, inode,
2398 right_path->p_node[i].bh,
2399 OCFS2_JOURNAL_ACCESS_WRITE);
2405 ret = ocfs2_journal_access(handle, inode,
2406 left_path->p_node[i].bh,
2407 OCFS2_JOURNAL_ACCESS_WRITE);
2414 if (!right_has_empty) {
2416 * Only do this if we're moving a real
2417 * record. Otherwise, the action is delayed until
2418 * after removal of the right path in which case we
2419 * can do a simple shift to remove the empty extent.
2421 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2422 memset(&right_leaf_el->l_recs[0], 0,
2423 sizeof(struct ocfs2_extent_rec));
2425 if (eb->h_next_leaf_blk == 0ULL) {
2427 * Move recs over to get rid of empty extent, decrease
2428 * next_free. This is allowed to remove the last
2429 * extent in our leaf (setting l_next_free_rec to
2430 * zero) - the delete code below won't care.
2432 ocfs2_remove_empty_extent(right_leaf_el);
2435 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2438 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2442 if (del_right_subtree) {
2443 ocfs2_unlink_subtree(inode, handle, left_path, right_path,
2444 subtree_index, dealloc);
2445 ocfs2_update_edge_lengths(inode, handle, left_path);
2447 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2448 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2451 * Removal of the extent in the left leaf was skipped
2452 * above so we could delete the right path
2455 if (right_has_empty)
2456 ocfs2_remove_empty_extent(left_leaf_el);
2458 ret = ocfs2_journal_dirty(handle, et_root_bh);
2464 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2472 * Given a full path, determine what cpos value would return us a path
2473 * containing the leaf immediately to the right of the current one.
2475 * Will return zero if the path passed in is already the rightmost path.
2477 * This looks similar, but is subtly different to
2478 * ocfs2_find_cpos_for_left_leaf().
2480 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2481 struct ocfs2_path *path, u32 *cpos)
2485 struct ocfs2_extent_list *el;
2489 if (path->p_tree_depth == 0)
2492 blkno = path_leaf_bh(path)->b_blocknr;
2494 /* Start at the tree node just above the leaf and work our way up. */
2495 i = path->p_tree_depth - 1;
2499 el = path->p_node[i].el;
2502 * Find the extent record just after the one in our
2505 next_free = le16_to_cpu(el->l_next_free_rec);
2506 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2507 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2508 if (j == (next_free - 1)) {
2511 * We've determined that the
2512 * path specified is already
2513 * the rightmost one - return a
2519 * The rightmost record points to our
2520 * leaf - we need to travel up the
2526 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2532 * If we got here, we never found a valid node where
2533 * the tree indicated one should be.
2536 "Invalid extent tree at extent block %llu\n",
2537 (unsigned long long)blkno);
2542 blkno = path->p_node[i].bh->b_blocknr;
2550 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2552 struct buffer_head *bh,
2553 struct ocfs2_extent_list *el)
2557 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2560 ret = ocfs2_journal_access(handle, inode, bh,
2561 OCFS2_JOURNAL_ACCESS_WRITE);
2567 ocfs2_remove_empty_extent(el);
2569 ret = ocfs2_journal_dirty(handle, bh);
2577 static int __ocfs2_rotate_tree_left(struct inode *inode,
2578 handle_t *handle, int orig_credits,
2579 struct ocfs2_path *path,
2580 struct ocfs2_cached_dealloc_ctxt *dealloc,
2581 struct ocfs2_path **empty_extent_path,
2582 struct ocfs2_extent_tree *et)
2584 int ret, subtree_root, deleted;
2586 struct ocfs2_path *left_path = NULL;
2587 struct ocfs2_path *right_path = NULL;
2589 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2591 *empty_extent_path = NULL;
2593 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2600 left_path = ocfs2_new_path(path_root_bh(path),
2601 path_root_el(path));
2608 ocfs2_cp_path(left_path, path);
2610 right_path = ocfs2_new_path(path_root_bh(path),
2611 path_root_el(path));
2618 while (right_cpos) {
2619 ret = ocfs2_find_path(inode, right_path, right_cpos);
2625 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2628 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2630 (unsigned long long)
2631 right_path->p_node[subtree_root].bh->b_blocknr,
2632 right_path->p_tree_depth);
2634 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2635 orig_credits, left_path);
2642 * Caller might still want to make changes to the
2643 * tree root, so re-add it to the journal here.
2645 ret = ocfs2_journal_access(handle, inode,
2646 path_root_bh(left_path),
2647 OCFS2_JOURNAL_ACCESS_WRITE);
2653 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2654 right_path, subtree_root,
2655 dealloc, &deleted, et);
2656 if (ret == -EAGAIN) {
2658 * The rotation has to temporarily stop due to
2659 * the right subtree having an empty
2660 * extent. Pass it back to the caller for a
2663 *empty_extent_path = right_path;
2673 * The subtree rotate might have removed records on
2674 * the rightmost edge. If so, then rotation is
2680 ocfs2_mv_path(left_path, right_path);
2682 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2691 ocfs2_free_path(right_path);
2692 ocfs2_free_path(left_path);
2697 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
2698 struct ocfs2_path *path,
2699 struct ocfs2_cached_dealloc_ctxt *dealloc,
2700 struct ocfs2_extent_tree *et)
2702 int ret, subtree_index;
2704 struct ocfs2_path *left_path = NULL;
2705 struct ocfs2_extent_block *eb;
2706 struct ocfs2_extent_list *el;
2709 ret = ocfs2_et_sanity_check(inode, et);
2713 * There's two ways we handle this depending on
2714 * whether path is the only existing one.
2716 ret = ocfs2_extend_rotate_transaction(handle, 0,
2717 handle->h_buffer_credits,
2724 ret = ocfs2_journal_access_path(inode, handle, path);
2730 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
2738 * We have a path to the left of this one - it needs
2741 left_path = ocfs2_new_path(path_root_bh(path),
2742 path_root_el(path));
2749 ret = ocfs2_find_path(inode, left_path, cpos);
2755 ret = ocfs2_journal_access_path(inode, handle, left_path);
2761 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
2763 ocfs2_unlink_subtree(inode, handle, left_path, path,
2764 subtree_index, dealloc);
2765 ocfs2_update_edge_lengths(inode, handle, left_path);
2767 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2768 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2771 * 'path' is also the leftmost path which
2772 * means it must be the only one. This gets
2773 * handled differently because we want to
2774 * revert the inode back to having extents
2777 ocfs2_unlink_path(inode, handle, dealloc, path, 1);
2779 el = et->et_root_el;
2780 el->l_tree_depth = 0;
2781 el->l_next_free_rec = 0;
2782 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2784 ocfs2_et_set_last_eb_blk(et, 0);
2787 ocfs2_journal_dirty(handle, path_root_bh(path));
2790 ocfs2_free_path(left_path);
2795 * Left rotation of btree records.
2797 * In many ways, this is (unsurprisingly) the opposite of right
2798 * rotation. We start at some non-rightmost path containing an empty
2799 * extent in the leaf block. The code works its way to the rightmost
2800 * path by rotating records to the left in every subtree.
2802 * This is used by any code which reduces the number of extent records
2803 * in a leaf. After removal, an empty record should be placed in the
2804 * leftmost list position.
2806 * This won't handle a length update of the rightmost path records if
2807 * the rightmost tree leaf record is removed so the caller is
2808 * responsible for detecting and correcting that.
2810 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
2811 struct ocfs2_path *path,
2812 struct ocfs2_cached_dealloc_ctxt *dealloc,
2813 struct ocfs2_extent_tree *et)
2815 int ret, orig_credits = handle->h_buffer_credits;
2816 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
2817 struct ocfs2_extent_block *eb;
2818 struct ocfs2_extent_list *el;
2820 el = path_leaf_el(path);
2821 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2824 if (path->p_tree_depth == 0) {
2825 rightmost_no_delete:
2827 * Inline extents. This is trivially handled, so do
2830 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
2832 path_leaf_el(path));
2839 * Handle rightmost branch now. There's several cases:
2840 * 1) simple rotation leaving records in there. That's trivial.
2841 * 2) rotation requiring a branch delete - there's no more
2842 * records left. Two cases of this:
2843 * a) There are branches to the left.
2844 * b) This is also the leftmost (the only) branch.
2846 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
2847 * 2a) we need the left branch so that we can update it with the unlink
2848 * 2b) we need to bring the inode back to inline extents.
2851 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2853 if (eb->h_next_leaf_blk == 0) {
2855 * This gets a bit tricky if we're going to delete the
2856 * rightmost path. Get the other cases out of the way
2859 if (le16_to_cpu(el->l_next_free_rec) > 1)
2860 goto rightmost_no_delete;
2862 if (le16_to_cpu(el->l_next_free_rec) == 0) {
2864 ocfs2_error(inode->i_sb,
2865 "Inode %llu has empty extent block at %llu",
2866 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2867 (unsigned long long)le64_to_cpu(eb->h_blkno));
2872 * XXX: The caller can not trust "path" any more after
2873 * this as it will have been deleted. What do we do?
2875 * In theory the rotate-for-merge code will never get
2876 * here because it'll always ask for a rotate in a
2880 ret = ocfs2_remove_rightmost_path(inode, handle, path,
2888 * Now we can loop, remembering the path we get from -EAGAIN
2889 * and restarting from there.
2892 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
2893 dealloc, &restart_path, et);
2894 if (ret && ret != -EAGAIN) {
2899 while (ret == -EAGAIN) {
2900 tmp_path = restart_path;
2901 restart_path = NULL;
2903 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
2906 if (ret && ret != -EAGAIN) {
2911 ocfs2_free_path(tmp_path);
2919 ocfs2_free_path(tmp_path);
2920 ocfs2_free_path(restart_path);
2924 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
2927 struct ocfs2_extent_rec *rec = &el->l_recs[index];
2930 if (rec->e_leaf_clusters == 0) {
2932 * We consumed all of the merged-from record. An empty
2933 * extent cannot exist anywhere but the 1st array
2934 * position, so move things over if the merged-from
2935 * record doesn't occupy that position.
2937 * This creates a new empty extent so the caller
2938 * should be smart enough to have removed any existing
2942 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
2943 size = index * sizeof(struct ocfs2_extent_rec);
2944 memmove(&el->l_recs[1], &el->l_recs[0], size);
2948 * Always memset - the caller doesn't check whether it
2949 * created an empty extent, so there could be junk in
2952 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2956 static int ocfs2_get_right_path(struct inode *inode,
2957 struct ocfs2_path *left_path,
2958 struct ocfs2_path **ret_right_path)
2962 struct ocfs2_path *right_path = NULL;
2963 struct ocfs2_extent_list *left_el;
2965 *ret_right_path = NULL;
2967 /* This function shouldn't be called for non-trees. */
2968 BUG_ON(left_path->p_tree_depth == 0);
2970 left_el = path_leaf_el(left_path);
2971 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
2973 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2980 /* This function shouldn't be called for the rightmost leaf. */
2981 BUG_ON(right_cpos == 0);
2983 right_path = ocfs2_new_path(path_root_bh(left_path),
2984 path_root_el(left_path));
2991 ret = ocfs2_find_path(inode, right_path, right_cpos);
2997 *ret_right_path = right_path;
3000 ocfs2_free_path(right_path);
3005 * Remove split_rec clusters from the record at index and merge them
3006 * onto the beginning of the record "next" to it.
3007 * For index < l_count - 1, the next means the extent rec at index + 1.
3008 * For index == l_count - 1, the "next" means the 1st extent rec of the
3009 * next extent block.
3011 static int ocfs2_merge_rec_right(struct inode *inode,
3012 struct ocfs2_path *left_path,
3014 struct ocfs2_extent_rec *split_rec,
3017 int ret, next_free, i;
3018 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3019 struct ocfs2_extent_rec *left_rec;
3020 struct ocfs2_extent_rec *right_rec;
3021 struct ocfs2_extent_list *right_el;
3022 struct ocfs2_path *right_path = NULL;
3023 int subtree_index = 0;
3024 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3025 struct buffer_head *bh = path_leaf_bh(left_path);
3026 struct buffer_head *root_bh = NULL;
3028 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3029 left_rec = &el->l_recs[index];
3031 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3032 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3033 /* we meet with a cross extent block merge. */
3034 ret = ocfs2_get_right_path(inode, left_path, &right_path);
3040 right_el = path_leaf_el(right_path);
3041 next_free = le16_to_cpu(right_el->l_next_free_rec);
3042 BUG_ON(next_free <= 0);
3043 right_rec = &right_el->l_recs[0];
3044 if (ocfs2_is_empty_extent(right_rec)) {
3045 BUG_ON(next_free <= 1);
3046 right_rec = &right_el->l_recs[1];
3049 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3050 le16_to_cpu(left_rec->e_leaf_clusters) !=
3051 le32_to_cpu(right_rec->e_cpos));
3053 subtree_index = ocfs2_find_subtree_root(inode,
3054 left_path, right_path);
3056 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3057 handle->h_buffer_credits,
3064 root_bh = left_path->p_node[subtree_index].bh;
3065 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3067 ret = ocfs2_journal_access(handle, inode, root_bh,
3068 OCFS2_JOURNAL_ACCESS_WRITE);
3074 for (i = subtree_index + 1;
3075 i < path_num_items(right_path); i++) {
3076 ret = ocfs2_journal_access(handle, inode,
3077 right_path->p_node[i].bh,
3078 OCFS2_JOURNAL_ACCESS_WRITE);
3084 ret = ocfs2_journal_access(handle, inode,
3085 left_path->p_node[i].bh,
3086 OCFS2_JOURNAL_ACCESS_WRITE);
3094 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3095 right_rec = &el->l_recs[index + 1];
3098 ret = ocfs2_journal_access(handle, inode, bh,
3099 OCFS2_JOURNAL_ACCESS_WRITE);
3105 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3107 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3108 le64_add_cpu(&right_rec->e_blkno,
3109 -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3110 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3112 ocfs2_cleanup_merge(el, index);
3114 ret = ocfs2_journal_dirty(handle, bh);
3119 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3123 ocfs2_complete_edge_insert(inode, handle, left_path,
3124 right_path, subtree_index);
3128 ocfs2_free_path(right_path);
3132 static int ocfs2_get_left_path(struct inode *inode,
3133 struct ocfs2_path *right_path,
3134 struct ocfs2_path **ret_left_path)
3138 struct ocfs2_path *left_path = NULL;
3140 *ret_left_path = NULL;
3142 /* This function shouldn't be called for non-trees. */
3143 BUG_ON(right_path->p_tree_depth == 0);
3145 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
3146 right_path, &left_cpos);
3152 /* This function shouldn't be called for the leftmost leaf. */
3153 BUG_ON(left_cpos == 0);
3155 left_path = ocfs2_new_path(path_root_bh(right_path),
3156 path_root_el(right_path));
3163 ret = ocfs2_find_path(inode, left_path, left_cpos);
3169 *ret_left_path = left_path;
3172 ocfs2_free_path(left_path);
3177 * Remove split_rec clusters from the record at index and merge them
3178 * onto the tail of the record "before" it.
3179 * For index > 0, the "before" means the extent rec at index - 1.
3181 * For index == 0, the "before" means the last record of the previous
3182 * extent block. And there is also a situation that we may need to
3183 * remove the rightmost leaf extent block in the right_path and change
3184 * the right path to indicate the new rightmost path.
3186 static int ocfs2_merge_rec_left(struct inode *inode,
3187 struct ocfs2_path *right_path,
3189 struct ocfs2_extent_rec *split_rec,
3190 struct ocfs2_cached_dealloc_ctxt *dealloc,
3191 struct ocfs2_extent_tree *et,
3194 int ret, i, subtree_index = 0, has_empty_extent = 0;
3195 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3196 struct ocfs2_extent_rec *left_rec;
3197 struct ocfs2_extent_rec *right_rec;
3198 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3199 struct buffer_head *bh = path_leaf_bh(right_path);
3200 struct buffer_head *root_bh = NULL;
3201 struct ocfs2_path *left_path = NULL;
3202 struct ocfs2_extent_list *left_el;
3206 right_rec = &el->l_recs[index];
3208 /* we meet with a cross extent block merge. */
3209 ret = ocfs2_get_left_path(inode, right_path, &left_path);
3215 left_el = path_leaf_el(left_path);
3216 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3217 le16_to_cpu(left_el->l_count));
3219 left_rec = &left_el->l_recs[
3220 le16_to_cpu(left_el->l_next_free_rec) - 1];
3221 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3222 le16_to_cpu(left_rec->e_leaf_clusters) !=
3223 le32_to_cpu(split_rec->e_cpos));
3225 subtree_index = ocfs2_find_subtree_root(inode,
3226 left_path, right_path);
3228 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3229 handle->h_buffer_credits,
3236 root_bh = left_path->p_node[subtree_index].bh;
3237 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3239 ret = ocfs2_journal_access(handle, inode, root_bh,
3240 OCFS2_JOURNAL_ACCESS_WRITE);
3246 for (i = subtree_index + 1;
3247 i < path_num_items(right_path); i++) {
3248 ret = ocfs2_journal_access(handle, inode,
3249 right_path->p_node[i].bh,
3250 OCFS2_JOURNAL_ACCESS_WRITE);
3256 ret = ocfs2_journal_access(handle, inode,
3257 left_path->p_node[i].bh,
3258 OCFS2_JOURNAL_ACCESS_WRITE);
3265 left_rec = &el->l_recs[index - 1];
3266 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3267 has_empty_extent = 1;
3270 ret = ocfs2_journal_access(handle, inode, bh,
3271 OCFS2_JOURNAL_ACCESS_WRITE);
3277 if (has_empty_extent && index == 1) {
3279 * The easy case - we can just plop the record right in.
3281 *left_rec = *split_rec;
3283 has_empty_extent = 0;
3285 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3287 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3288 le64_add_cpu(&right_rec->e_blkno,
3289 ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3290 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3292 ocfs2_cleanup_merge(el, index);
3294 ret = ocfs2_journal_dirty(handle, bh);
3299 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3304 * In the situation that the right_rec is empty and the extent
3305 * block is empty also, ocfs2_complete_edge_insert can't handle
3306 * it and we need to delete the right extent block.
3308 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3309 le16_to_cpu(el->l_next_free_rec) == 1) {
3311 ret = ocfs2_remove_rightmost_path(inode, handle,
3319 /* Now the rightmost extent block has been deleted.
3320 * So we use the new rightmost path.
3322 ocfs2_mv_path(right_path, left_path);
3325 ocfs2_complete_edge_insert(inode, handle, left_path,
3326 right_path, subtree_index);
3330 ocfs2_free_path(left_path);
3334 static int ocfs2_try_to_merge_extent(struct inode *inode,
3336 struct ocfs2_path *path,
3338 struct ocfs2_extent_rec *split_rec,
3339 struct ocfs2_cached_dealloc_ctxt *dealloc,
3340 struct ocfs2_merge_ctxt *ctxt,
3341 struct ocfs2_extent_tree *et)
3345 struct ocfs2_extent_list *el = path_leaf_el(path);
3346 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3348 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3350 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3352 * The merge code will need to create an empty
3353 * extent to take the place of the newly
3354 * emptied slot. Remove any pre-existing empty
3355 * extents - having more than one in a leaf is
3358 ret = ocfs2_rotate_tree_left(inode, handle, path,
3365 rec = &el->l_recs[split_index];
3368 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3370 * Left-right contig implies this.
3372 BUG_ON(!ctxt->c_split_covers_rec);
3375 * Since the leftright insert always covers the entire
3376 * extent, this call will delete the insert record
3377 * entirely, resulting in an empty extent record added to
3380 * Since the adding of an empty extent shifts
3381 * everything back to the right, there's no need to
3382 * update split_index here.
3384 * When the split_index is zero, we need to merge it to the
3385 * prevoius extent block. It is more efficient and easier
3386 * if we do merge_right first and merge_left later.
3388 ret = ocfs2_merge_rec_right(inode, path,
3397 * We can only get this from logic error above.
3399 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3401 /* The merge left us with an empty extent, remove it. */
3402 ret = ocfs2_rotate_tree_left(inode, handle, path,
3409 rec = &el->l_recs[split_index];
3412 * Note that we don't pass split_rec here on purpose -
3413 * we've merged it into the rec already.
3415 ret = ocfs2_merge_rec_left(inode, path,
3425 ret = ocfs2_rotate_tree_left(inode, handle, path,
3428 * Error from this last rotate is not critical, so
3429 * print but don't bubble it up.
3436 * Merge a record to the left or right.
3438 * 'contig_type' is relative to the existing record,
3439 * so for example, if we're "right contig", it's to
3440 * the record on the left (hence the left merge).
3442 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3443 ret = ocfs2_merge_rec_left(inode,
3453 ret = ocfs2_merge_rec_right(inode,
3463 if (ctxt->c_split_covers_rec) {
3465 * The merge may have left an empty extent in
3466 * our leaf. Try to rotate it away.
3468 ret = ocfs2_rotate_tree_left(inode, handle, path,
3480 static void ocfs2_subtract_from_rec(struct super_block *sb,
3481 enum ocfs2_split_type split,
3482 struct ocfs2_extent_rec *rec,
3483 struct ocfs2_extent_rec *split_rec)
3487 len_blocks = ocfs2_clusters_to_blocks(sb,
3488 le16_to_cpu(split_rec->e_leaf_clusters));
3490 if (split == SPLIT_LEFT) {
3492 * Region is on the left edge of the existing
3495 le32_add_cpu(&rec->e_cpos,
3496 le16_to_cpu(split_rec->e_leaf_clusters));
3497 le64_add_cpu(&rec->e_blkno, len_blocks);
3498 le16_add_cpu(&rec->e_leaf_clusters,
3499 -le16_to_cpu(split_rec->e_leaf_clusters));
3502 * Region is on the right edge of the existing
3505 le16_add_cpu(&rec->e_leaf_clusters,
3506 -le16_to_cpu(split_rec->e_leaf_clusters));
3511 * Do the final bits of extent record insertion at the target leaf
3512 * list. If this leaf is part of an allocation tree, it is assumed
3513 * that the tree above has been prepared.
3515 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
3516 struct ocfs2_extent_list *el,
3517 struct ocfs2_insert_type *insert,
3518 struct inode *inode)
3520 int i = insert->ins_contig_index;
3522 struct ocfs2_extent_rec *rec;
3524 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3526 if (insert->ins_split != SPLIT_NONE) {
3527 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3529 rec = &el->l_recs[i];
3530 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
3536 * Contiguous insert - either left or right.
3538 if (insert->ins_contig != CONTIG_NONE) {
3539 rec = &el->l_recs[i];
3540 if (insert->ins_contig == CONTIG_LEFT) {
3541 rec->e_blkno = insert_rec->e_blkno;
3542 rec->e_cpos = insert_rec->e_cpos;
3544 le16_add_cpu(&rec->e_leaf_clusters,
3545 le16_to_cpu(insert_rec->e_leaf_clusters));
3550 * Handle insert into an empty leaf.
3552 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3553 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3554 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3555 el->l_recs[0] = *insert_rec;
3556 el->l_next_free_rec = cpu_to_le16(1);
3563 if (insert->ins_appending == APPEND_TAIL) {
3564 i = le16_to_cpu(el->l_next_free_rec) - 1;
3565 rec = &el->l_recs[i];
3566 range = le32_to_cpu(rec->e_cpos)
3567 + le16_to_cpu(rec->e_leaf_clusters);
3568 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3570 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3571 le16_to_cpu(el->l_count),
3572 "inode %lu, depth %u, count %u, next free %u, "
3573 "rec.cpos %u, rec.clusters %u, "
3574 "insert.cpos %u, insert.clusters %u\n",
3576 le16_to_cpu(el->l_tree_depth),
3577 le16_to_cpu(el->l_count),
3578 le16_to_cpu(el->l_next_free_rec),
3579 le32_to_cpu(el->l_recs[i].e_cpos),
3580 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3581 le32_to_cpu(insert_rec->e_cpos),
3582 le16_to_cpu(insert_rec->e_leaf_clusters));
3584 el->l_recs[i] = *insert_rec;
3585 le16_add_cpu(&el->l_next_free_rec, 1);
3591 * Ok, we have to rotate.
3593 * At this point, it is safe to assume that inserting into an
3594 * empty leaf and appending to a leaf have both been handled
3597 * This leaf needs to have space, either by the empty 1st
3598 * extent record, or by virtue of an l_next_rec < l_count.
3600 ocfs2_rotate_leaf(el, insert_rec);
3603 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3605 struct ocfs2_path *path,
3606 struct ocfs2_extent_rec *insert_rec)
3608 int ret, i, next_free;
3609 struct buffer_head *bh;
3610 struct ocfs2_extent_list *el;
3611 struct ocfs2_extent_rec *rec;
3614 * Update everything except the leaf block.
3616 for (i = 0; i < path->p_tree_depth; i++) {
3617 bh = path->p_node[i].bh;
3618 el = path->p_node[i].el;
3620 next_free = le16_to_cpu(el->l_next_free_rec);
3621 if (next_free == 0) {
3622 ocfs2_error(inode->i_sb,
3623 "Dinode %llu has a bad extent list",
3624 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3629 rec = &el->l_recs[next_free - 1];
3631 rec->e_int_clusters = insert_rec->e_cpos;
3632 le32_add_cpu(&rec->e_int_clusters,
3633 le16_to_cpu(insert_rec->e_leaf_clusters));
3634 le32_add_cpu(&rec->e_int_clusters,
3635 -le32_to_cpu(rec->e_cpos));
3637 ret = ocfs2_journal_dirty(handle, bh);
3644 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3645 struct ocfs2_extent_rec *insert_rec,
3646 struct ocfs2_path *right_path,
3647 struct ocfs2_path **ret_left_path)
3650 struct ocfs2_extent_list *el;
3651 struct ocfs2_path *left_path = NULL;
3653 *ret_left_path = NULL;
3656 * This shouldn't happen for non-trees. The extent rec cluster
3657 * count manipulation below only works for interior nodes.
3659 BUG_ON(right_path->p_tree_depth == 0);
3662 * If our appending insert is at the leftmost edge of a leaf,
3663 * then we might need to update the rightmost records of the
3666 el = path_leaf_el(right_path);
3667 next_free = le16_to_cpu(el->l_next_free_rec);
3668 if (next_free == 0 ||
3669 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3672 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3679 mlog(0, "Append may need a left path update. cpos: %u, "
3680 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3684 * No need to worry if the append is already in the
3688 left_path = ocfs2_new_path(path_root_bh(right_path),
3689 path_root_el(right_path));
3696 ret = ocfs2_find_path(inode, left_path, left_cpos);
3703 * ocfs2_insert_path() will pass the left_path to the
3709 ret = ocfs2_journal_access_path(inode, handle, right_path);
3715 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
3717 *ret_left_path = left_path;
3721 ocfs2_free_path(left_path);
3726 static void ocfs2_split_record(struct inode *inode,
3727 struct ocfs2_path *left_path,
3728 struct ocfs2_path *right_path,
3729 struct ocfs2_extent_rec *split_rec,
3730 enum ocfs2_split_type split)
3733 u32 cpos = le32_to_cpu(split_rec->e_cpos);
3734 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
3735 struct ocfs2_extent_rec *rec, *tmprec;
3737 right_el = path_leaf_el(right_path);;
3739 left_el = path_leaf_el(left_path);
3742 insert_el = right_el;
3743 index = ocfs2_search_extent_list(el, cpos);
3745 if (index == 0 && left_path) {
3746 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3749 * This typically means that the record
3750 * started in the left path but moved to the
3751 * right as a result of rotation. We either
3752 * move the existing record to the left, or we
3753 * do the later insert there.
3755 * In this case, the left path should always
3756 * exist as the rotate code will have passed
3757 * it back for a post-insert update.
3760 if (split == SPLIT_LEFT) {
3762 * It's a left split. Since we know
3763 * that the rotate code gave us an
3764 * empty extent in the left path, we
3765 * can just do the insert there.
3767 insert_el = left_el;
3770 * Right split - we have to move the
3771 * existing record over to the left
3772 * leaf. The insert will be into the
3773 * newly created empty extent in the
3776 tmprec = &right_el->l_recs[index];
3777 ocfs2_rotate_leaf(left_el, tmprec);
3780 memset(tmprec, 0, sizeof(*tmprec));
3781 index = ocfs2_search_extent_list(left_el, cpos);
3782 BUG_ON(index == -1);
3787 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
3789 * Left path is easy - we can just allow the insert to
3793 insert_el = left_el;
3794 index = ocfs2_search_extent_list(el, cpos);
3795 BUG_ON(index == -1);
3798 rec = &el->l_recs[index];
3799 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
3800 ocfs2_rotate_leaf(insert_el, split_rec);
3804 * This function only does inserts on an allocation b-tree. For tree
3805 * depth = 0, ocfs2_insert_at_leaf() is called directly.
3807 * right_path is the path we want to do the actual insert
3808 * in. left_path should only be passed in if we need to update that
3809 * portion of the tree after an edge insert.
3811 static int ocfs2_insert_path(struct inode *inode,
3813 struct ocfs2_path *left_path,
3814 struct ocfs2_path *right_path,
3815 struct ocfs2_extent_rec *insert_rec,
3816 struct ocfs2_insert_type *insert)
3818 int ret, subtree_index;
3819 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
3822 int credits = handle->h_buffer_credits;
3825 * There's a chance that left_path got passed back to
3826 * us without being accounted for in the
3827 * journal. Extend our transaction here to be sure we
3828 * can change those blocks.
3830 credits += left_path->p_tree_depth;
3832 ret = ocfs2_extend_trans(handle, credits);
3838 ret = ocfs2_journal_access_path(inode, handle, left_path);
3846 * Pass both paths to the journal. The majority of inserts
3847 * will be touching all components anyway.
3849 ret = ocfs2_journal_access_path(inode, handle, right_path);
3855 if (insert->ins_split != SPLIT_NONE) {
3857 * We could call ocfs2_insert_at_leaf() for some types
3858 * of splits, but it's easier to just let one separate
3859 * function sort it all out.
3861 ocfs2_split_record(inode, left_path, right_path,
3862 insert_rec, insert->ins_split);
3865 * Split might have modified either leaf and we don't
3866 * have a guarantee that the later edge insert will
3867 * dirty this for us.
3870 ret = ocfs2_journal_dirty(handle,
3871 path_leaf_bh(left_path));
3875 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
3878 ret = ocfs2_journal_dirty(handle, leaf_bh);
3884 * The rotate code has indicated that we need to fix
3885 * up portions of the tree after the insert.
3887 * XXX: Should we extend the transaction here?
3889 subtree_index = ocfs2_find_subtree_root(inode, left_path,
3891 ocfs2_complete_edge_insert(inode, handle, left_path,
3892 right_path, subtree_index);
3900 static int ocfs2_do_insert_extent(struct inode *inode,
3902 struct ocfs2_extent_tree *et,
3903 struct ocfs2_extent_rec *insert_rec,
3904 struct ocfs2_insert_type *type)
3906 int ret, rotate = 0;
3908 struct ocfs2_path *right_path = NULL;
3909 struct ocfs2_path *left_path = NULL;
3910 struct ocfs2_extent_list *el;
3912 el = et->et_root_el;
3914 ret = ocfs2_journal_access(handle, inode, et->et_root_bh,
3915 OCFS2_JOURNAL_ACCESS_WRITE);
3921 if (le16_to_cpu(el->l_tree_depth) == 0) {
3922 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
3923 goto out_update_clusters;
3926 right_path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
3934 * Determine the path to start with. Rotations need the
3935 * rightmost path, everything else can go directly to the
3938 cpos = le32_to_cpu(insert_rec->e_cpos);
3939 if (type->ins_appending == APPEND_NONE &&
3940 type->ins_contig == CONTIG_NONE) {
3945 ret = ocfs2_find_path(inode, right_path, cpos);
3952 * Rotations and appends need special treatment - they modify
3953 * parts of the tree's above them.
3955 * Both might pass back a path immediate to the left of the
3956 * one being inserted to. This will be cause
3957 * ocfs2_insert_path() to modify the rightmost records of
3958 * left_path to account for an edge insert.
3960 * XXX: When modifying this code, keep in mind that an insert
3961 * can wind up skipping both of these two special cases...
3964 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
3965 le32_to_cpu(insert_rec->e_cpos),
3966 right_path, &left_path);
3973 * ocfs2_rotate_tree_right() might have extended the
3974 * transaction without re-journaling our tree root.
3976 ret = ocfs2_journal_access(handle, inode, et->et_root_bh,
3977 OCFS2_JOURNAL_ACCESS_WRITE);
3982 } else if (type->ins_appending == APPEND_TAIL
3983 && type->ins_contig != CONTIG_LEFT) {
3984 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
3985 right_path, &left_path);
3992 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
3999 out_update_clusters:
4000 if (type->ins_split == SPLIT_NONE)
4001 ocfs2_et_update_clusters(inode, et,
4002 le16_to_cpu(insert_rec->e_leaf_clusters));
4004 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
4009 ocfs2_free_path(left_path);
4010 ocfs2_free_path(right_path);
4015 static enum ocfs2_contig_type
4016 ocfs2_figure_merge_contig_type(struct inode *inode, struct ocfs2_path *path,
4017 struct ocfs2_extent_list *el, int index,
4018 struct ocfs2_extent_rec *split_rec)
4021 enum ocfs2_contig_type ret = CONTIG_NONE;
4022 u32 left_cpos, right_cpos;
4023 struct ocfs2_extent_rec *rec = NULL;
4024 struct ocfs2_extent_list *new_el;
4025 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4026 struct buffer_head *bh;
4027 struct ocfs2_extent_block *eb;
4030 rec = &el->l_recs[index - 1];
4031 } else if (path->p_tree_depth > 0) {
4032 status = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
4037 if (left_cpos != 0) {
4038 left_path = ocfs2_new_path(path_root_bh(path),
4039 path_root_el(path));
4043 status = ocfs2_find_path(inode, left_path, left_cpos);
4047 new_el = path_leaf_el(left_path);
4049 if (le16_to_cpu(new_el->l_next_free_rec) !=
4050 le16_to_cpu(new_el->l_count)) {
4051 bh = path_leaf_bh(left_path);
4052 eb = (struct ocfs2_extent_block *)bh->b_data;
4053 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
4057 rec = &new_el->l_recs[
4058 le16_to_cpu(new_el->l_next_free_rec) - 1];
4063 * We're careful to check for an empty extent record here -
4064 * the merge code will know what to do if it sees one.
4067 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4068 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4071 ret = ocfs2_extent_contig(inode, rec, split_rec);
4076 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4077 rec = &el->l_recs[index + 1];
4078 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4079 path->p_tree_depth > 0) {
4080 status = ocfs2_find_cpos_for_right_leaf(inode->i_sb,
4085 if (right_cpos == 0)
4088 right_path = ocfs2_new_path(path_root_bh(path),
4089 path_root_el(path));
4093 status = ocfs2_find_path(inode, right_path, right_cpos);
4097 new_el = path_leaf_el(right_path);
4098 rec = &new_el->l_recs[0];
4099 if (ocfs2_is_empty_extent(rec)) {
4100 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4101 bh = path_leaf_bh(right_path);
4102 eb = (struct ocfs2_extent_block *)bh->b_data;
4103 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
4107 rec = &new_el->l_recs[1];
4112 enum ocfs2_contig_type contig_type;
4114 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
4116 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4117 ret = CONTIG_LEFTRIGHT;
4118 else if (ret == CONTIG_NONE)
4124 ocfs2_free_path(left_path);
4126 ocfs2_free_path(right_path);
4131 static void ocfs2_figure_contig_type(struct inode *inode,
4132 struct ocfs2_insert_type *insert,
4133 struct ocfs2_extent_list *el,
4134 struct ocfs2_extent_rec *insert_rec,
4135 struct ocfs2_extent_tree *et)
4138 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4140 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4142 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4143 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
4145 if (contig_type != CONTIG_NONE) {
4146 insert->ins_contig_index = i;
4150 insert->ins_contig = contig_type;
4152 if (insert->ins_contig != CONTIG_NONE) {
4153 struct ocfs2_extent_rec *rec =
4154 &el->l_recs[insert->ins_contig_index];
4155 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4156 le16_to_cpu(insert_rec->e_leaf_clusters);
4159 * Caller might want us to limit the size of extents, don't
4160 * calculate contiguousness if we might exceed that limit.
4162 if (et->et_max_leaf_clusters &&
4163 (len > et->et_max_leaf_clusters))
4164 insert->ins_contig = CONTIG_NONE;
4169 * This should only be called against the righmost leaf extent list.
4171 * ocfs2_figure_appending_type() will figure out whether we'll have to
4172 * insert at the tail of the rightmost leaf.
4174 * This should also work against the root extent list for tree's with 0
4175 * depth. If we consider the root extent list to be the rightmost leaf node
4176 * then the logic here makes sense.
4178 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4179 struct ocfs2_extent_list *el,
4180 struct ocfs2_extent_rec *insert_rec)
4183 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4184 struct ocfs2_extent_rec *rec;
4186 insert->ins_appending = APPEND_NONE;
4188 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4190 if (!el->l_next_free_rec)
4191 goto set_tail_append;
4193 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4194 /* Were all records empty? */
4195 if (le16_to_cpu(el->l_next_free_rec) == 1)
4196 goto set_tail_append;
4199 i = le16_to_cpu(el->l_next_free_rec) - 1;
4200 rec = &el->l_recs[i];
4203 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4204 goto set_tail_append;
4209 insert->ins_appending = APPEND_TAIL;
4213 * Helper function called at the begining of an insert.
4215 * This computes a few things that are commonly used in the process of
4216 * inserting into the btree:
4217 * - Whether the new extent is contiguous with an existing one.
4218 * - The current tree depth.
4219 * - Whether the insert is an appending one.
4220 * - The total # of free records in the tree.
4222 * All of the information is stored on the ocfs2_insert_type
4225 static int ocfs2_figure_insert_type(struct inode *inode,
4226 struct ocfs2_extent_tree *et,
4227 struct buffer_head **last_eb_bh,
4228 struct ocfs2_extent_rec *insert_rec,
4230 struct ocfs2_insert_type *insert)
4233 struct ocfs2_extent_block *eb;
4234 struct ocfs2_extent_list *el;
4235 struct ocfs2_path *path = NULL;
4236 struct buffer_head *bh = NULL;
4238 insert->ins_split = SPLIT_NONE;
4240 el = et->et_root_el;
4241 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4243 if (el->l_tree_depth) {
4245 * If we have tree depth, we read in the
4246 * rightmost extent block ahead of time as
4247 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4248 * may want it later.
4250 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4251 ocfs2_et_get_last_eb_blk(et), &bh,
4252 OCFS2_BH_CACHED, inode);
4257 eb = (struct ocfs2_extent_block *) bh->b_data;
4262 * Unless we have a contiguous insert, we'll need to know if
4263 * there is room left in our allocation tree for another
4266 * XXX: This test is simplistic, we can search for empty
4267 * extent records too.
4269 *free_records = le16_to_cpu(el->l_count) -
4270 le16_to_cpu(el->l_next_free_rec);
4272 if (!insert->ins_tree_depth) {
4273 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4274 ocfs2_figure_appending_type(insert, el, insert_rec);
4278 path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
4286 * In the case that we're inserting past what the tree
4287 * currently accounts for, ocfs2_find_path() will return for
4288 * us the rightmost tree path. This is accounted for below in
4289 * the appending code.
4291 ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
4297 el = path_leaf_el(path);
4300 * Now that we have the path, there's two things we want to determine:
4301 * 1) Contiguousness (also set contig_index if this is so)
4303 * 2) Are we doing an append? We can trivially break this up
4304 * into two types of appends: simple record append, or a
4305 * rotate inside the tail leaf.
4307 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4310 * The insert code isn't quite ready to deal with all cases of
4311 * left contiguousness. Specifically, if it's an insert into
4312 * the 1st record in a leaf, it will require the adjustment of
4313 * cluster count on the last record of the path directly to it's
4314 * left. For now, just catch that case and fool the layers
4315 * above us. This works just fine for tree_depth == 0, which
4316 * is why we allow that above.
4318 if (insert->ins_contig == CONTIG_LEFT &&
4319 insert->ins_contig_index == 0)
4320 insert->ins_contig = CONTIG_NONE;
4323 * Ok, so we can simply compare against last_eb to figure out
4324 * whether the path doesn't exist. This will only happen in
4325 * the case that we're doing a tail append, so maybe we can
4326 * take advantage of that information somehow.
4328 if (ocfs2_et_get_last_eb_blk(et) ==
4329 path_leaf_bh(path)->b_blocknr) {
4331 * Ok, ocfs2_find_path() returned us the rightmost
4332 * tree path. This might be an appending insert. There are
4334 * 1) We're doing a true append at the tail:
4335 * -This might even be off the end of the leaf
4336 * 2) We're "appending" by rotating in the tail
4338 ocfs2_figure_appending_type(insert, el, insert_rec);
4342 ocfs2_free_path(path);
4352 * Insert an extent into an inode btree.
4354 * The caller needs to update fe->i_clusters
4356 static int ocfs2_insert_extent(struct ocfs2_super *osb,
4358 struct inode *inode,
4359 struct buffer_head *root_bh,
4364 struct ocfs2_alloc_context *meta_ac,
4365 struct ocfs2_extent_tree *et)
4368 int uninitialized_var(free_records);
4369 struct buffer_head *last_eb_bh = NULL;
4370 struct ocfs2_insert_type insert = {0, };
4371 struct ocfs2_extent_rec rec;
4373 BUG_ON(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL);
4375 mlog(0, "add %u clusters at position %u to inode %llu\n",
4376 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4378 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
4379 (OCFS2_I(inode)->ip_clusters != cpos),
4380 "Device %s, asking for sparse allocation: inode %llu, "
4381 "cpos %u, clusters %u\n",
4383 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos,
4384 OCFS2_I(inode)->ip_clusters);
4386 memset(&rec, 0, sizeof(rec));
4387 rec.e_cpos = cpu_to_le32(cpos);
4388 rec.e_blkno = cpu_to_le64(start_blk);
4389 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4390 rec.e_flags = flags;
4392 status = ocfs2_figure_insert_type(inode, et, &last_eb_bh, &rec,
4393 &free_records, &insert);
4399 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4400 "Insert.contig_index: %d, Insert.free_records: %d, "
4401 "Insert.tree_depth: %d\n",
4402 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
4403 free_records, insert.ins_tree_depth);
4405 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4406 status = ocfs2_grow_tree(inode, handle, et,
4407 &insert.ins_tree_depth, &last_eb_bh,
4415 /* Finally, we can add clusters. This might rotate the tree for us. */
4416 status = ocfs2_do_insert_extent(inode, handle, et, &rec, &insert);
4419 else if (et->et_type == OCFS2_DINODE_EXTENT)
4420 ocfs2_extent_map_insert_rec(inode, &rec);
4430 int ocfs2_dinode_insert_extent(struct ocfs2_super *osb,
4432 struct inode *inode,
4433 struct buffer_head *root_bh,
4438 struct ocfs2_alloc_context *meta_ac)
4441 struct ocfs2_extent_tree et;
4443 ocfs2_get_extent_tree(&et, inode, root_bh, OCFS2_DINODE_EXTENT,
4445 status = ocfs2_insert_extent(osb, handle, inode, root_bh,
4446 cpos, start_blk, new_clusters,
4447 flags, meta_ac, &et);
4448 ocfs2_put_extent_tree(&et);
4453 int ocfs2_xattr_value_insert_extent(struct ocfs2_super *osb,
4455 struct inode *inode,
4456 struct buffer_head *root_bh,
4461 struct ocfs2_alloc_context *meta_ac,
4465 struct ocfs2_extent_tree et;
4467 ocfs2_get_extent_tree(&et, inode, root_bh,
4468 OCFS2_XATTR_VALUE_EXTENT, obj);
4469 status = ocfs2_insert_extent(osb, handle, inode, root_bh,
4470 cpos, start_blk, new_clusters,
4471 flags, meta_ac, &et);
4472 ocfs2_put_extent_tree(&et);
4477 int ocfs2_xattr_tree_insert_extent(struct ocfs2_super *osb,
4479 struct inode *inode,
4480 struct buffer_head *root_bh,
4485 struct ocfs2_alloc_context *meta_ac)
4488 struct ocfs2_extent_tree et;
4490 ocfs2_get_extent_tree(&et, inode, root_bh, OCFS2_XATTR_TREE_EXTENT,
4492 status = ocfs2_insert_extent(osb, handle, inode, root_bh,
4493 cpos, start_blk, new_clusters,
4494 flags, meta_ac, &et);
4495 ocfs2_put_extent_tree(&et);
4501 * Allcate and add clusters into the extent b-tree.
4502 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4503 * The extent b-tree's root is root_el and it should be in root_bh, and
4504 * it is not limited to the file storage. Any extent tree can use this
4505 * function if it implements the proper ocfs2_extent_tree.
4507 int ocfs2_add_clusters_in_btree(struct ocfs2_super *osb,
4508 struct inode *inode,
4509 u32 *logical_offset,
4510 u32 clusters_to_add,
4512 struct buffer_head *root_bh,
4513 struct ocfs2_extent_list *root_el,
4515 struct ocfs2_alloc_context *data_ac,
4516 struct ocfs2_alloc_context *meta_ac,
4517 enum ocfs2_alloc_restarted *reason_ret,
4518 enum ocfs2_extent_tree_type type,
4523 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4524 u32 bit_off, num_bits;
4528 BUG_ON(!clusters_to_add);
4531 flags = OCFS2_EXT_UNWRITTEN;
4533 free_extents = ocfs2_num_free_extents(osb, inode, root_bh, type,
4535 if (free_extents < 0) {
4536 status = free_extents;
4541 /* there are two cases which could cause us to EAGAIN in the
4542 * we-need-more-metadata case:
4543 * 1) we haven't reserved *any*
4544 * 2) we are so fragmented, we've needed to add metadata too
4546 if (!free_extents && !meta_ac) {
4547 mlog(0, "we haven't reserved any metadata!\n");
4549 reason = RESTART_META;
4551 } else if ((!free_extents)
4552 && (ocfs2_alloc_context_bits_left(meta_ac)
4553 < ocfs2_extend_meta_needed(root_el))) {
4554 mlog(0, "filesystem is really fragmented...\n");
4556 reason = RESTART_META;
4560 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
4561 clusters_to_add, &bit_off, &num_bits);
4563 if (status != -ENOSPC)
4568 BUG_ON(num_bits > clusters_to_add);
4570 /* reserve our write early -- insert_extent may update the inode */
4571 status = ocfs2_journal_access(handle, inode, root_bh,
4572 OCFS2_JOURNAL_ACCESS_WRITE);
4578 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4579 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
4580 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4581 if (type == OCFS2_DINODE_EXTENT)
4582 status = ocfs2_dinode_insert_extent(osb, handle, inode, root_bh,
4583 *logical_offset, block,
4584 num_bits, flags, meta_ac);
4585 else if (type == OCFS2_XATTR_TREE_EXTENT)
4586 status = ocfs2_xattr_tree_insert_extent(osb, handle,
4589 block, num_bits, flags,
4592 status = ocfs2_xattr_value_insert_extent(osb, handle,
4595 block, num_bits, flags,
4602 status = ocfs2_journal_dirty(handle, root_bh);
4608 clusters_to_add -= num_bits;
4609 *logical_offset += num_bits;
4611 if (clusters_to_add) {
4612 mlog(0, "need to alloc once more, wanted = %u\n",
4615 reason = RESTART_TRANS;
4621 *reason_ret = reason;
4625 static void ocfs2_make_right_split_rec(struct super_block *sb,
4626 struct ocfs2_extent_rec *split_rec,
4628 struct ocfs2_extent_rec *rec)
4630 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4631 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4633 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4635 split_rec->e_cpos = cpu_to_le32(cpos);
4636 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4638 split_rec->e_blkno = rec->e_blkno;
4639 le64_add_cpu(&split_rec->e_blkno,
4640 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4642 split_rec->e_flags = rec->e_flags;
4645 static int ocfs2_split_and_insert(struct inode *inode,
4647 struct ocfs2_path *path,
4648 struct ocfs2_extent_tree *et,
4649 struct buffer_head **last_eb_bh,
4651 struct ocfs2_extent_rec *orig_split_rec,
4652 struct ocfs2_alloc_context *meta_ac)
4655 unsigned int insert_range, rec_range, do_leftright = 0;
4656 struct ocfs2_extent_rec tmprec;
4657 struct ocfs2_extent_list *rightmost_el;
4658 struct ocfs2_extent_rec rec;
4659 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4660 struct ocfs2_insert_type insert;
4661 struct ocfs2_extent_block *eb;
4665 * Store a copy of the record on the stack - it might move
4666 * around as the tree is manipulated below.
4668 rec = path_leaf_el(path)->l_recs[split_index];
4670 rightmost_el = et->et_root_el;
4672 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4674 BUG_ON(!(*last_eb_bh));
4675 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4676 rightmost_el = &eb->h_list;
4679 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4680 le16_to_cpu(rightmost_el->l_count)) {
4681 ret = ocfs2_grow_tree(inode, handle, et,
4682 &depth, last_eb_bh, meta_ac);
4689 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4690 insert.ins_appending = APPEND_NONE;
4691 insert.ins_contig = CONTIG_NONE;
4692 insert.ins_tree_depth = depth;
4694 insert_range = le32_to_cpu(split_rec.e_cpos) +
4695 le16_to_cpu(split_rec.e_leaf_clusters);
4696 rec_range = le32_to_cpu(rec.e_cpos) +
4697 le16_to_cpu(rec.e_leaf_clusters);
4699 if (split_rec.e_cpos == rec.e_cpos) {
4700 insert.ins_split = SPLIT_LEFT;
4701 } else if (insert_range == rec_range) {
4702 insert.ins_split = SPLIT_RIGHT;
4705 * Left/right split. We fake this as a right split
4706 * first and then make a second pass as a left split.
4708 insert.ins_split = SPLIT_RIGHT;
4710 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
4715 BUG_ON(do_leftright);
4719 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4725 if (do_leftright == 1) {
4727 struct ocfs2_extent_list *el;
4730 split_rec = *orig_split_rec;
4732 ocfs2_reinit_path(path, 1);
4734 cpos = le32_to_cpu(split_rec.e_cpos);
4735 ret = ocfs2_find_path(inode, path, cpos);
4741 el = path_leaf_el(path);
4742 split_index = ocfs2_search_extent_list(el, cpos);
4751 * Mark part or all of the extent record at split_index in the leaf
4752 * pointed to by path as written. This removes the unwritten
4755 * Care is taken to handle contiguousness so as to not grow the tree.
4757 * meta_ac is not strictly necessary - we only truly need it if growth
4758 * of the tree is required. All other cases will degrade into a less
4759 * optimal tree layout.
4761 * last_eb_bh should be the rightmost leaf block for any extent
4762 * btree. Since a split may grow the tree or a merge might shrink it,
4763 * the caller cannot trust the contents of that buffer after this call.
4765 * This code is optimized for readability - several passes might be
4766 * made over certain portions of the tree. All of those blocks will
4767 * have been brought into cache (and pinned via the journal), so the
4768 * extra overhead is not expressed in terms of disk reads.
4770 static int __ocfs2_mark_extent_written(struct inode *inode,
4771 struct ocfs2_extent_tree *et,
4773 struct ocfs2_path *path,
4775 struct ocfs2_extent_rec *split_rec,
4776 struct ocfs2_alloc_context *meta_ac,
4777 struct ocfs2_cached_dealloc_ctxt *dealloc)
4780 struct ocfs2_extent_list *el = path_leaf_el(path);
4781 struct buffer_head *last_eb_bh = NULL;
4782 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
4783 struct ocfs2_merge_ctxt ctxt;
4784 struct ocfs2_extent_list *rightmost_el;
4786 if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) {
4792 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
4793 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
4794 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
4800 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, path, el,
4805 * The core merge / split code wants to know how much room is
4806 * left in this inodes allocation tree, so we pass the
4807 * rightmost extent list.
4809 if (path->p_tree_depth) {
4810 struct ocfs2_extent_block *eb;
4812 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4813 ocfs2_et_get_last_eb_blk(et),
4814 &last_eb_bh, OCFS2_BH_CACHED, inode);
4820 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4821 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
4822 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
4827 rightmost_el = &eb->h_list;
4829 rightmost_el = path_root_el(path);
4831 if (rec->e_cpos == split_rec->e_cpos &&
4832 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
4833 ctxt.c_split_covers_rec = 1;
4835 ctxt.c_split_covers_rec = 0;
4837 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
4839 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
4840 split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent,
4841 ctxt.c_split_covers_rec);
4843 if (ctxt.c_contig_type == CONTIG_NONE) {
4844 if (ctxt.c_split_covers_rec)
4845 el->l_recs[split_index] = *split_rec;
4847 ret = ocfs2_split_and_insert(inode, handle, path, et,
4848 &last_eb_bh, split_index,
4849 split_rec, meta_ac);
4853 ret = ocfs2_try_to_merge_extent(inode, handle, path,
4854 split_index, split_rec,
4855 dealloc, &ctxt, et);
4866 * Mark the already-existing extent at cpos as written for len clusters.
4868 * If the existing extent is larger than the request, initiate a
4869 * split. An attempt will be made at merging with adjacent extents.
4871 * The caller is responsible for passing down meta_ac if we'll need it.
4873 int ocfs2_mark_extent_written(struct inode *inode, struct buffer_head *root_bh,
4874 handle_t *handle, u32 cpos, u32 len, u32 phys,
4875 struct ocfs2_alloc_context *meta_ac,
4876 struct ocfs2_cached_dealloc_ctxt *dealloc,
4877 enum ocfs2_extent_tree_type et_type,
4881 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
4882 struct ocfs2_extent_rec split_rec;
4883 struct ocfs2_path *left_path = NULL;
4884 struct ocfs2_extent_list *el;
4885 struct ocfs2_extent_tree et;
4887 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4888 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
4890 ocfs2_get_extent_tree(&et, inode, root_bh, et_type, obj);
4892 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
4893 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
4894 "that are being written to, but the feature bit "
4895 "is not set in the super block.",
4896 (unsigned long long)OCFS2_I(inode)->ip_blkno);
4902 * XXX: This should be fixed up so that we just re-insert the
4903 * next extent records.
4905 if (et_type == OCFS2_DINODE_EXTENT)
4906 ocfs2_extent_map_trunc(inode, 0);
4908 left_path = ocfs2_new_path(et.et_root_bh, et.et_root_el);
4915 ret = ocfs2_find_path(inode, left_path, cpos);
4920 el = path_leaf_el(left_path);
4922 index = ocfs2_search_extent_list(el, cpos);
4923 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4924 ocfs2_error(inode->i_sb,
4925 "Inode %llu has an extent at cpos %u which can no "
4926 "longer be found.\n",
4927 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
4932 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
4933 split_rec.e_cpos = cpu_to_le32(cpos);
4934 split_rec.e_leaf_clusters = cpu_to_le16(len);
4935 split_rec.e_blkno = cpu_to_le64(start_blkno);
4936 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
4937 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
4939 ret = __ocfs2_mark_extent_written(inode, &et, handle, left_path,
4940 index, &split_rec, meta_ac,
4946 ocfs2_free_path(left_path);
4947 ocfs2_put_extent_tree(&et);
4951 static int ocfs2_split_tree(struct inode *inode, struct ocfs2_extent_tree *et,
4952 handle_t *handle, struct ocfs2_path *path,
4953 int index, u32 new_range,
4954 struct ocfs2_alloc_context *meta_ac)
4956 int ret, depth, credits = handle->h_buffer_credits;
4957 struct buffer_head *last_eb_bh = NULL;
4958 struct ocfs2_extent_block *eb;
4959 struct ocfs2_extent_list *rightmost_el, *el;
4960 struct ocfs2_extent_rec split_rec;
4961 struct ocfs2_extent_rec *rec;
4962 struct ocfs2_insert_type insert;
4965 * Setup the record to split before we grow the tree.
4967 el = path_leaf_el(path);
4968 rec = &el->l_recs[index];
4969 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
4971 depth = path->p_tree_depth;
4973 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4974 ocfs2_et_get_last_eb_blk(et),
4975 &last_eb_bh, OCFS2_BH_CACHED, inode);
4981 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4982 rightmost_el = &eb->h_list;
4984 rightmost_el = path_leaf_el(path);
4986 credits += path->p_tree_depth +
4987 ocfs2_extend_meta_needed(et->et_root_el);
4988 ret = ocfs2_extend_trans(handle, credits);
4994 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4995 le16_to_cpu(rightmost_el->l_count)) {
4996 ret = ocfs2_grow_tree(inode, handle, et, &depth, &last_eb_bh,
5004 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5005 insert.ins_appending = APPEND_NONE;
5006 insert.ins_contig = CONTIG_NONE;
5007 insert.ins_split = SPLIT_RIGHT;
5008 insert.ins_tree_depth = depth;
5010 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
5019 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
5020 struct ocfs2_path *path, int index,
5021 struct ocfs2_cached_dealloc_ctxt *dealloc,
5023 struct ocfs2_extent_tree *et)
5026 u32 left_cpos, rec_range, trunc_range;
5027 int wants_rotate = 0, is_rightmost_tree_rec = 0;
5028 struct super_block *sb = inode->i_sb;
5029 struct ocfs2_path *left_path = NULL;
5030 struct ocfs2_extent_list *el = path_leaf_el(path);
5031 struct ocfs2_extent_rec *rec;
5032 struct ocfs2_extent_block *eb;
5034 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5035 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5044 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5045 path->p_tree_depth) {
5047 * Check whether this is the rightmost tree record. If
5048 * we remove all of this record or part of its right
5049 * edge then an update of the record lengths above it
5052 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5053 if (eb->h_next_leaf_blk == 0)
5054 is_rightmost_tree_rec = 1;
5057 rec = &el->l_recs[index];
5058 if (index == 0 && path->p_tree_depth &&
5059 le32_to_cpu(rec->e_cpos) == cpos) {
5061 * Changing the leftmost offset (via partial or whole
5062 * record truncate) of an interior (or rightmost) path
5063 * means we have to update the subtree that is formed
5064 * by this leaf and the one to it's left.
5066 * There are two cases we can skip:
5067 * 1) Path is the leftmost one in our inode tree.
5068 * 2) The leaf is rightmost and will be empty after
5069 * we remove the extent record - the rotate code
5070 * knows how to update the newly formed edge.
5073 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
5080 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5081 left_path = ocfs2_new_path(path_root_bh(path),
5082 path_root_el(path));
5089 ret = ocfs2_find_path(inode, left_path, left_cpos);
5097 ret = ocfs2_extend_rotate_transaction(handle, 0,
5098 handle->h_buffer_credits,
5105 ret = ocfs2_journal_access_path(inode, handle, path);
5111 ret = ocfs2_journal_access_path(inode, handle, left_path);
5117 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5118 trunc_range = cpos + len;
5120 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5123 memset(rec, 0, sizeof(*rec));
5124 ocfs2_cleanup_merge(el, index);
5127 next_free = le16_to_cpu(el->l_next_free_rec);
5128 if (is_rightmost_tree_rec && next_free > 1) {
5130 * We skip the edge update if this path will
5131 * be deleted by the rotate code.
5133 rec = &el->l_recs[next_free - 1];
5134 ocfs2_adjust_rightmost_records(inode, handle, path,
5137 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5138 /* Remove leftmost portion of the record. */
5139 le32_add_cpu(&rec->e_cpos, len);
5140 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5141 le16_add_cpu(&rec->e_leaf_clusters, -len);
5142 } else if (rec_range == trunc_range) {
5143 /* Remove rightmost portion of the record */
5144 le16_add_cpu(&rec->e_leaf_clusters, -len);
5145 if (is_rightmost_tree_rec)
5146 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
5148 /* Caller should have trapped this. */
5149 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
5150 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
5151 le32_to_cpu(rec->e_cpos),
5152 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5159 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
5160 ocfs2_complete_edge_insert(inode, handle, left_path, path,
5164 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5166 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5173 ocfs2_free_path(left_path);
5177 int ocfs2_remove_extent(struct inode *inode, struct buffer_head *root_bh,
5178 u32 cpos, u32 len, handle_t *handle,
5179 struct ocfs2_alloc_context *meta_ac,
5180 struct ocfs2_cached_dealloc_ctxt *dealloc,
5181 enum ocfs2_extent_tree_type et_type,
5185 u32 rec_range, trunc_range;
5186 struct ocfs2_extent_rec *rec;
5187 struct ocfs2_extent_list *el;
5188 struct ocfs2_path *path = NULL;
5189 struct ocfs2_extent_tree et;
5191 ocfs2_get_extent_tree(&et, inode, root_bh, et_type, obj);
5193 ocfs2_extent_map_trunc(inode, 0);
5195 path = ocfs2_new_path(et.et_root_bh, et.et_root_el);
5202 ret = ocfs2_find_path(inode, path, cpos);
5208 el = path_leaf_el(path);
5209 index = ocfs2_search_extent_list(el, cpos);
5210 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5211 ocfs2_error(inode->i_sb,
5212 "Inode %llu has an extent at cpos %u which can no "
5213 "longer be found.\n",
5214 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5220 * We have 3 cases of extent removal:
5221 * 1) Range covers the entire extent rec
5222 * 2) Range begins or ends on one edge of the extent rec
5223 * 3) Range is in the middle of the extent rec (no shared edges)
5225 * For case 1 we remove the extent rec and left rotate to
5228 * For case 2 we just shrink the existing extent rec, with a
5229 * tree update if the shrinking edge is also the edge of an
5232 * For case 3 we do a right split to turn the extent rec into
5233 * something case 2 can handle.
5235 rec = &el->l_recs[index];
5236 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5237 trunc_range = cpos + len;
5239 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5241 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
5242 "(cpos %u, len %u)\n",
5243 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
5244 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
5246 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5247 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5254 ret = ocfs2_split_tree(inode, &et, handle, path, index,
5255 trunc_range, meta_ac);
5262 * The split could have manipulated the tree enough to
5263 * move the record location, so we have to look for it again.
5265 ocfs2_reinit_path(path, 1);
5267 ret = ocfs2_find_path(inode, path, cpos);
5273 el = path_leaf_el(path);
5274 index = ocfs2_search_extent_list(el, cpos);
5275 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5276 ocfs2_error(inode->i_sb,
5277 "Inode %llu: split at cpos %u lost record.",
5278 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5285 * Double check our values here. If anything is fishy,
5286 * it's easier to catch it at the top level.
5288 rec = &el->l_recs[index];
5289 rec_range = le32_to_cpu(rec->e_cpos) +
5290 ocfs2_rec_clusters(el, rec);
5291 if (rec_range != trunc_range) {
5292 ocfs2_error(inode->i_sb,
5293 "Inode %llu: error after split at cpos %u"
5294 "trunc len %u, existing record is (%u,%u)",
5295 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5296 cpos, len, le32_to_cpu(rec->e_cpos),
5297 ocfs2_rec_clusters(el, rec));
5302 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5311 ocfs2_free_path(path);
5312 ocfs2_put_extent_tree(&et);
5316 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5318 struct buffer_head *tl_bh = osb->osb_tl_bh;
5319 struct ocfs2_dinode *di;
5320 struct ocfs2_truncate_log *tl;
5322 di = (struct ocfs2_dinode *) tl_bh->b_data;
5323 tl = &di->id2.i_dealloc;
5325 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5326 "slot %d, invalid truncate log parameters: used = "
5327 "%u, count = %u\n", osb->slot_num,
5328 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5329 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5332 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5333 unsigned int new_start)
5335 unsigned int tail_index;
5336 unsigned int current_tail;
5338 /* No records, nothing to coalesce */
5339 if (!le16_to_cpu(tl->tl_used))
5342 tail_index = le16_to_cpu(tl->tl_used) - 1;
5343 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5344 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5346 return current_tail == new_start;
5349 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5352 unsigned int num_clusters)
5355 unsigned int start_cluster, tl_count;
5356 struct inode *tl_inode = osb->osb_tl_inode;
5357 struct buffer_head *tl_bh = osb->osb_tl_bh;
5358 struct ocfs2_dinode *di;
5359 struct ocfs2_truncate_log *tl;
5361 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5362 (unsigned long long)start_blk, num_clusters);
5364 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5366 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5368 di = (struct ocfs2_dinode *) tl_bh->b_data;
5369 tl = &di->id2.i_dealloc;
5370 if (!OCFS2_IS_VALID_DINODE(di)) {
5371 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
5376 tl_count = le16_to_cpu(tl->tl_count);
5377 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5379 "Truncate record count on #%llu invalid "
5380 "wanted %u, actual %u\n",
5381 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5382 ocfs2_truncate_recs_per_inode(osb->sb),
5383 le16_to_cpu(tl->tl_count));
5385 /* Caller should have known to flush before calling us. */
5386 index = le16_to_cpu(tl->tl_used);
5387 if (index >= tl_count) {
5393 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
5394 OCFS2_JOURNAL_ACCESS_WRITE);
5400 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5401 "%llu (index = %d)\n", num_clusters, start_cluster,
5402 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
5404 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5406 * Move index back to the record we are coalescing with.
5407 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5411 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5412 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5413 index, le32_to_cpu(tl->tl_recs[index].t_start),
5416 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5417 tl->tl_used = cpu_to_le16(index + 1);
5419 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5421 status = ocfs2_journal_dirty(handle, tl_bh);
5432 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5434 struct inode *data_alloc_inode,
5435 struct buffer_head *data_alloc_bh)
5439 unsigned int num_clusters;
5441 struct ocfs2_truncate_rec rec;
5442 struct ocfs2_dinode *di;
5443 struct ocfs2_truncate_log *tl;
5444 struct inode *tl_inode = osb->osb_tl_inode;
5445 struct buffer_head *tl_bh = osb->osb_tl_bh;
5449 di = (struct ocfs2_dinode *) tl_bh->b_data;
5450 tl = &di->id2.i_dealloc;
5451 i = le16_to_cpu(tl->tl_used) - 1;
5453 /* Caller has given us at least enough credits to
5454 * update the truncate log dinode */
5455 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
5456 OCFS2_JOURNAL_ACCESS_WRITE);
5462 tl->tl_used = cpu_to_le16(i);
5464 status = ocfs2_journal_dirty(handle, tl_bh);
5470 /* TODO: Perhaps we can calculate the bulk of the
5471 * credits up front rather than extending like
5473 status = ocfs2_extend_trans(handle,
5474 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5480 rec = tl->tl_recs[i];
5481 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5482 le32_to_cpu(rec.t_start));
5483 num_clusters = le32_to_cpu(rec.t_clusters);
5485 /* if start_blk is not set, we ignore the record as
5488 mlog(0, "free record %d, start = %u, clusters = %u\n",
5489 i, le32_to_cpu(rec.t_start), num_clusters);
5491 status = ocfs2_free_clusters(handle, data_alloc_inode,
5492 data_alloc_bh, start_blk,
5507 /* Expects you to already be holding tl_inode->i_mutex */
5508 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5511 unsigned int num_to_flush;
5513 struct inode *tl_inode = osb->osb_tl_inode;
5514 struct inode *data_alloc_inode = NULL;
5515 struct buffer_head *tl_bh = osb->osb_tl_bh;
5516 struct buffer_head *data_alloc_bh = NULL;
5517 struct ocfs2_dinode *di;
5518 struct ocfs2_truncate_log *tl;
5522 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5524 di = (struct ocfs2_dinode *) tl_bh->b_data;
5525 tl = &di->id2.i_dealloc;
5526 if (!OCFS2_IS_VALID_DINODE(di)) {
5527 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
5532 num_to_flush = le16_to_cpu(tl->tl_used);
5533 mlog(0, "Flush %u records from truncate log #%llu\n",
5534 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
5535 if (!num_to_flush) {
5540 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5541 GLOBAL_BITMAP_SYSTEM_INODE,
5542 OCFS2_INVALID_SLOT);
5543 if (!data_alloc_inode) {
5545 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5549 mutex_lock(&data_alloc_inode->i_mutex);
5551 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5557 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5558 if (IS_ERR(handle)) {
5559 status = PTR_ERR(handle);
5564 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
5569 ocfs2_commit_trans(osb, handle);
5572 brelse(data_alloc_bh);
5573 ocfs2_inode_unlock(data_alloc_inode, 1);
5576 mutex_unlock(&data_alloc_inode->i_mutex);
5577 iput(data_alloc_inode);
5584 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5587 struct inode *tl_inode = osb->osb_tl_inode;
5589 mutex_lock(&tl_inode->i_mutex);
5590 status = __ocfs2_flush_truncate_log(osb);
5591 mutex_unlock(&tl_inode->i_mutex);
5596 static void ocfs2_truncate_log_worker(struct work_struct *work)
5599 struct ocfs2_super *osb =
5600 container_of(work, struct ocfs2_super,
5601 osb_truncate_log_wq.work);
5605 status = ocfs2_flush_truncate_log(osb);
5609 ocfs2_init_inode_steal_slot(osb);
5614 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5615 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
5618 if (osb->osb_tl_inode) {
5619 /* We want to push off log flushes while truncates are
5622 cancel_delayed_work(&osb->osb_truncate_log_wq);
5624 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
5625 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
5629 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
5631 struct inode **tl_inode,
5632 struct buffer_head **tl_bh)
5635 struct inode *inode = NULL;
5636 struct buffer_head *bh = NULL;
5638 inode = ocfs2_get_system_file_inode(osb,
5639 TRUNCATE_LOG_SYSTEM_INODE,
5643 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
5647 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
5648 OCFS2_BH_CACHED, inode);
5662 /* called during the 1st stage of node recovery. we stamp a clean
5663 * truncate log and pass back a copy for processing later. if the
5664 * truncate log does not require processing, a *tl_copy is set to
5666 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
5668 struct ocfs2_dinode **tl_copy)
5671 struct inode *tl_inode = NULL;
5672 struct buffer_head *tl_bh = NULL;
5673 struct ocfs2_dinode *di;
5674 struct ocfs2_truncate_log *tl;
5678 mlog(0, "recover truncate log from slot %d\n", slot_num);
5680 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
5686 di = (struct ocfs2_dinode *) tl_bh->b_data;
5687 tl = &di->id2.i_dealloc;
5688 if (!OCFS2_IS_VALID_DINODE(di)) {
5689 OCFS2_RO_ON_INVALID_DINODE(tl_inode->i_sb, di);
5694 if (le16_to_cpu(tl->tl_used)) {
5695 mlog(0, "We'll have %u logs to recover\n",
5696 le16_to_cpu(tl->tl_used));
5698 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
5705 /* Assuming the write-out below goes well, this copy
5706 * will be passed back to recovery for processing. */
5707 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
5709 /* All we need to do to clear the truncate log is set
5713 status = ocfs2_write_block(osb, tl_bh, tl_inode);
5726 if (status < 0 && (*tl_copy)) {
5735 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
5736 struct ocfs2_dinode *tl_copy)
5740 unsigned int clusters, num_recs, start_cluster;
5743 struct inode *tl_inode = osb->osb_tl_inode;
5744 struct ocfs2_truncate_log *tl;
5748 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
5749 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
5753 tl = &tl_copy->id2.i_dealloc;
5754 num_recs = le16_to_cpu(tl->tl_used);
5755 mlog(0, "cleanup %u records from %llu\n", num_recs,
5756 (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
5758 mutex_lock(&tl_inode->i_mutex);
5759 for(i = 0; i < num_recs; i++) {
5760 if (ocfs2_truncate_log_needs_flush(osb)) {
5761 status = __ocfs2_flush_truncate_log(osb);
5768 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5769 if (IS_ERR(handle)) {
5770 status = PTR_ERR(handle);
5775 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
5776 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
5777 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
5779 status = ocfs2_truncate_log_append(osb, handle,
5780 start_blk, clusters);
5781 ocfs2_commit_trans(osb, handle);
5789 mutex_unlock(&tl_inode->i_mutex);
5795 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
5798 struct inode *tl_inode = osb->osb_tl_inode;
5803 cancel_delayed_work(&osb->osb_truncate_log_wq);
5804 flush_workqueue(ocfs2_wq);
5806 status = ocfs2_flush_truncate_log(osb);
5810 brelse(osb->osb_tl_bh);
5811 iput(osb->osb_tl_inode);
5817 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
5820 struct inode *tl_inode = NULL;
5821 struct buffer_head *tl_bh = NULL;
5825 status = ocfs2_get_truncate_log_info(osb,
5832 /* ocfs2_truncate_log_shutdown keys on the existence of
5833 * osb->osb_tl_inode so we don't set any of the osb variables
5834 * until we're sure all is well. */
5835 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
5836 ocfs2_truncate_log_worker);
5837 osb->osb_tl_bh = tl_bh;
5838 osb->osb_tl_inode = tl_inode;
5845 * Delayed de-allocation of suballocator blocks.
5847 * Some sets of block de-allocations might involve multiple suballocator inodes.
5849 * The locking for this can get extremely complicated, especially when
5850 * the suballocator inodes to delete from aren't known until deep
5851 * within an unrelated codepath.
5853 * ocfs2_extent_block structures are a good example of this - an inode
5854 * btree could have been grown by any number of nodes each allocating
5855 * out of their own suballoc inode.
5857 * These structures allow the delay of block de-allocation until a
5858 * later time, when locking of multiple cluster inodes won't cause
5863 * Describes a single block free from a suballocator
5865 struct ocfs2_cached_block_free {
5866 struct ocfs2_cached_block_free *free_next;
5868 unsigned int free_bit;
5871 struct ocfs2_per_slot_free_list {
5872 struct ocfs2_per_slot_free_list *f_next_suballocator;
5875 struct ocfs2_cached_block_free *f_first;
5878 static int ocfs2_free_cached_items(struct ocfs2_super *osb,
5881 struct ocfs2_cached_block_free *head)
5886 struct inode *inode;
5887 struct buffer_head *di_bh = NULL;
5888 struct ocfs2_cached_block_free *tmp;
5890 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
5897 mutex_lock(&inode->i_mutex);
5899 ret = ocfs2_inode_lock(inode, &di_bh, 1);
5905 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
5906 if (IS_ERR(handle)) {
5907 ret = PTR_ERR(handle);
5913 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
5915 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
5916 head->free_bit, (unsigned long long)head->free_blk);
5918 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
5919 head->free_bit, bg_blkno, 1);
5925 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
5932 head = head->free_next;
5937 ocfs2_commit_trans(osb, handle);
5940 ocfs2_inode_unlock(inode, 1);
5943 mutex_unlock(&inode->i_mutex);
5947 /* Premature exit may have left some dangling items. */
5949 head = head->free_next;
5956 int ocfs2_run_deallocs(struct ocfs2_super *osb,
5957 struct ocfs2_cached_dealloc_ctxt *ctxt)
5960 struct ocfs2_per_slot_free_list *fl;
5965 while (ctxt->c_first_suballocator) {
5966 fl = ctxt->c_first_suballocator;
5969 mlog(0, "Free items: (type %u, slot %d)\n",
5970 fl->f_inode_type, fl->f_slot);
5971 ret2 = ocfs2_free_cached_items(osb, fl->f_inode_type,
5972 fl->f_slot, fl->f_first);
5979 ctxt->c_first_suballocator = fl->f_next_suballocator;
5986 static struct ocfs2_per_slot_free_list *
5987 ocfs2_find_per_slot_free_list(int type,
5989 struct ocfs2_cached_dealloc_ctxt *ctxt)
5991 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
5994 if (fl->f_inode_type == type && fl->f_slot == slot)
5997 fl = fl->f_next_suballocator;
6000 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6002 fl->f_inode_type = type;
6005 fl->f_next_suballocator = ctxt->c_first_suballocator;
6007 ctxt->c_first_suballocator = fl;
6012 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6013 int type, int slot, u64 blkno,
6017 struct ocfs2_per_slot_free_list *fl;
6018 struct ocfs2_cached_block_free *item;
6020 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6027 item = kmalloc(sizeof(*item), GFP_NOFS);
6034 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
6035 type, slot, bit, (unsigned long long)blkno);
6037 item->free_blk = blkno;
6038 item->free_bit = bit;
6039 item->free_next = fl->f_first;
6048 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6049 struct ocfs2_extent_block *eb)
6051 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6052 le16_to_cpu(eb->h_suballoc_slot),
6053 le64_to_cpu(eb->h_blkno),
6054 le16_to_cpu(eb->h_suballoc_bit));
6057 /* This function will figure out whether the currently last extent
6058 * block will be deleted, and if it will, what the new last extent
6059 * block will be so we can update his h_next_leaf_blk field, as well
6060 * as the dinodes i_last_eb_blk */
6061 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
6062 unsigned int clusters_to_del,
6063 struct ocfs2_path *path,
6064 struct buffer_head **new_last_eb)
6066 int next_free, ret = 0;
6068 struct ocfs2_extent_rec *rec;
6069 struct ocfs2_extent_block *eb;
6070 struct ocfs2_extent_list *el;
6071 struct buffer_head *bh = NULL;
6073 *new_last_eb = NULL;
6075 /* we have no tree, so of course, no last_eb. */
6076 if (!path->p_tree_depth)
6079 /* trunc to zero special case - this makes tree_depth = 0
6080 * regardless of what it is. */
6081 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
6084 el = path_leaf_el(path);
6085 BUG_ON(!el->l_next_free_rec);
6088 * Make sure that this extent list will actually be empty
6089 * after we clear away the data. We can shortcut out if
6090 * there's more than one non-empty extent in the
6091 * list. Otherwise, a check of the remaining extent is
6094 next_free = le16_to_cpu(el->l_next_free_rec);
6096 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6100 /* We may have a valid extent in index 1, check it. */
6102 rec = &el->l_recs[1];
6105 * Fall through - no more nonempty extents, so we want
6106 * to delete this leaf.
6112 rec = &el->l_recs[0];
6117 * Check it we'll only be trimming off the end of this
6120 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
6124 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
6130 ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
6136 eb = (struct ocfs2_extent_block *) bh->b_data;
6138 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
6139 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
6145 get_bh(*new_last_eb);
6146 mlog(0, "returning block %llu, (cpos: %u)\n",
6147 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
6155 * Trim some clusters off the rightmost edge of a tree. Only called
6158 * The caller needs to:
6159 * - start journaling of each path component.
6160 * - compute and fully set up any new last ext block
6162 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
6163 handle_t *handle, struct ocfs2_truncate_context *tc,
6164 u32 clusters_to_del, u64 *delete_start)
6166 int ret, i, index = path->p_tree_depth;
6169 struct buffer_head *bh;
6170 struct ocfs2_extent_list *el;
6171 struct ocfs2_extent_rec *rec;
6175 while (index >= 0) {
6176 bh = path->p_node[index].bh;
6177 el = path->p_node[index].el;
6179 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6180 index, (unsigned long long)bh->b_blocknr);
6182 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
6185 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
6186 ocfs2_error(inode->i_sb,
6187 "Inode %lu has invalid ext. block %llu",
6189 (unsigned long long)bh->b_blocknr);
6195 i = le16_to_cpu(el->l_next_free_rec) - 1;
6196 rec = &el->l_recs[i];
6198 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6199 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
6200 ocfs2_rec_clusters(el, rec),
6201 (unsigned long long)le64_to_cpu(rec->e_blkno),
6202 le16_to_cpu(el->l_next_free_rec));
6204 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
6206 if (le16_to_cpu(el->l_tree_depth) == 0) {
6208 * If the leaf block contains a single empty
6209 * extent and no records, we can just remove
6212 if (i == 0 && ocfs2_is_empty_extent(rec)) {
6214 sizeof(struct ocfs2_extent_rec));
6215 el->l_next_free_rec = cpu_to_le16(0);
6221 * Remove any empty extents by shifting things
6222 * left. That should make life much easier on
6223 * the code below. This condition is rare
6224 * enough that we shouldn't see a performance
6227 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6228 le16_add_cpu(&el->l_next_free_rec, -1);
6231 i < le16_to_cpu(el->l_next_free_rec); i++)
6232 el->l_recs[i] = el->l_recs[i + 1];
6234 memset(&el->l_recs[i], 0,
6235 sizeof(struct ocfs2_extent_rec));
6238 * We've modified our extent list. The
6239 * simplest way to handle this change
6240 * is to being the search from the
6243 goto find_tail_record;
6246 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
6249 * We'll use "new_edge" on our way back up the
6250 * tree to know what our rightmost cpos is.
6252 new_edge = le16_to_cpu(rec->e_leaf_clusters);
6253 new_edge += le32_to_cpu(rec->e_cpos);
6256 * The caller will use this to delete data blocks.
6258 *delete_start = le64_to_cpu(rec->e_blkno)
6259 + ocfs2_clusters_to_blocks(inode->i_sb,
6260 le16_to_cpu(rec->e_leaf_clusters));
6263 * If it's now empty, remove this record.
6265 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
6267 sizeof(struct ocfs2_extent_rec));
6268 le16_add_cpu(&el->l_next_free_rec, -1);
6271 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
6273 sizeof(struct ocfs2_extent_rec));
6274 le16_add_cpu(&el->l_next_free_rec, -1);
6279 /* Can this actually happen? */
6280 if (le16_to_cpu(el->l_next_free_rec) == 0)
6284 * We never actually deleted any clusters
6285 * because our leaf was empty. There's no
6286 * reason to adjust the rightmost edge then.
6291 rec->e_int_clusters = cpu_to_le32(new_edge);
6292 le32_add_cpu(&rec->e_int_clusters,
6293 -le32_to_cpu(rec->e_cpos));
6296 * A deleted child record should have been
6299 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
6303 ret = ocfs2_journal_dirty(handle, bh);
6309 mlog(0, "extent list container %llu, after: record %d: "
6310 "(%u, %u, %llu), next = %u.\n",
6311 (unsigned long long)bh->b_blocknr, i,
6312 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
6313 (unsigned long long)le64_to_cpu(rec->e_blkno),
6314 le16_to_cpu(el->l_next_free_rec));
6317 * We must be careful to only attempt delete of an
6318 * extent block (and not the root inode block).
6320 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
6321 struct ocfs2_extent_block *eb =
6322 (struct ocfs2_extent_block *)bh->b_data;
6325 * Save this for use when processing the
6328 deleted_eb = le64_to_cpu(eb->h_blkno);
6330 mlog(0, "deleting this extent block.\n");
6332 ocfs2_remove_from_cache(inode, bh);
6334 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
6335 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
6336 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
6338 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
6339 /* An error here is not fatal. */
6354 static int ocfs2_do_truncate(struct ocfs2_super *osb,
6355 unsigned int clusters_to_del,
6356 struct inode *inode,
6357 struct buffer_head *fe_bh,
6359 struct ocfs2_truncate_context *tc,
6360 struct ocfs2_path *path)
6363 struct ocfs2_dinode *fe;
6364 struct ocfs2_extent_block *last_eb = NULL;
6365 struct ocfs2_extent_list *el;
6366 struct buffer_head *last_eb_bh = NULL;
6369 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6371 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
6379 * Each component will be touched, so we might as well journal
6380 * here to avoid having to handle errors later.
6382 status = ocfs2_journal_access_path(inode, handle, path);
6389 status = ocfs2_journal_access(handle, inode, last_eb_bh,
6390 OCFS2_JOURNAL_ACCESS_WRITE);
6396 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6399 el = &(fe->id2.i_list);
6402 * Lower levels depend on this never happening, but it's best
6403 * to check it up here before changing the tree.
6405 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
6406 ocfs2_error(inode->i_sb,
6407 "Inode %lu has an empty extent record, depth %u\n",
6408 inode->i_ino, le16_to_cpu(el->l_tree_depth));
6413 spin_lock(&OCFS2_I(inode)->ip_lock);
6414 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
6416 spin_unlock(&OCFS2_I(inode)->ip_lock);
6417 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
6418 inode->i_blocks = ocfs2_inode_sector_count(inode);
6420 status = ocfs2_trim_tree(inode, path, handle, tc,
6421 clusters_to_del, &delete_blk);
6427 if (le32_to_cpu(fe->i_clusters) == 0) {
6428 /* trunc to zero is a special case. */
6429 el->l_tree_depth = 0;
6430 fe->i_last_eb_blk = 0;
6432 fe->i_last_eb_blk = last_eb->h_blkno;
6434 status = ocfs2_journal_dirty(handle, fe_bh);
6441 /* If there will be a new last extent block, then by
6442 * definition, there cannot be any leaves to the right of
6444 last_eb->h_next_leaf_blk = 0;
6445 status = ocfs2_journal_dirty(handle, last_eb_bh);
6453 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
6467 static int ocfs2_writeback_zero_func(handle_t *handle, struct buffer_head *bh)
6469 set_buffer_uptodate(bh);
6470 mark_buffer_dirty(bh);
6474 static int ocfs2_ordered_zero_func(handle_t *handle, struct buffer_head *bh)
6476 set_buffer_uptodate(bh);
6477 mark_buffer_dirty(bh);
6478 return ocfs2_journal_dirty_data(handle, bh);
6481 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6482 unsigned int from, unsigned int to,
6483 struct page *page, int zero, u64 *phys)
6485 int ret, partial = 0;
6487 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6492 zero_user_segment(page, from, to);
6495 * Need to set the buffers we zero'd into uptodate
6496 * here if they aren't - ocfs2_map_page_blocks()
6497 * might've skipped some
6499 if (ocfs2_should_order_data(inode)) {
6500 ret = walk_page_buffers(handle,
6503 ocfs2_ordered_zero_func);
6507 ret = walk_page_buffers(handle, page_buffers(page),
6509 ocfs2_writeback_zero_func);
6515 SetPageUptodate(page);
6517 flush_dcache_page(page);
6520 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6521 loff_t end, struct page **pages,
6522 int numpages, u64 phys, handle_t *handle)
6526 unsigned int from, to = PAGE_CACHE_SIZE;
6527 struct super_block *sb = inode->i_sb;
6529 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6534 to = PAGE_CACHE_SIZE;
6535 for(i = 0; i < numpages; i++) {
6538 from = start & (PAGE_CACHE_SIZE - 1);
6539 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6540 to = end & (PAGE_CACHE_SIZE - 1);
6542 BUG_ON(from > PAGE_CACHE_SIZE);
6543 BUG_ON(to > PAGE_CACHE_SIZE);
6545 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6548 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6552 ocfs2_unlock_and_free_pages(pages, numpages);
6555 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6556 struct page **pages, int *num)
6558 int numpages, ret = 0;
6559 struct super_block *sb = inode->i_sb;
6560 struct address_space *mapping = inode->i_mapping;
6561 unsigned long index;
6562 loff_t last_page_bytes;
6564 BUG_ON(start > end);
6566 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6567 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6570 last_page_bytes = PAGE_ALIGN(end);
6571 index = start >> PAGE_CACHE_SHIFT;
6573 pages[numpages] = grab_cache_page(mapping, index);
6574 if (!pages[numpages]) {
6582 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6587 ocfs2_unlock_and_free_pages(pages, numpages);
6597 * Zero the area past i_size but still within an allocated
6598 * cluster. This avoids exposing nonzero data on subsequent file
6601 * We need to call this before i_size is updated on the inode because
6602 * otherwise block_write_full_page() will skip writeout of pages past
6603 * i_size. The new_i_size parameter is passed for this reason.
6605 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6606 u64 range_start, u64 range_end)
6608 int ret = 0, numpages;
6609 struct page **pages = NULL;
6611 unsigned int ext_flags;
6612 struct super_block *sb = inode->i_sb;
6615 * File systems which don't support sparse files zero on every
6618 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6621 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6622 sizeof(struct page *), GFP_NOFS);
6623 if (pages == NULL) {
6629 if (range_start == range_end)
6632 ret = ocfs2_extent_map_get_blocks(inode,
6633 range_start >> sb->s_blocksize_bits,
6634 &phys, NULL, &ext_flags);
6641 * Tail is a hole, or is marked unwritten. In either case, we
6642 * can count on read and write to return/push zero's.
6644 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6647 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6654 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6655 numpages, phys, handle);
6658 * Initiate writeout of the pages we zero'd here. We don't
6659 * wait on them - the truncate_inode_pages() call later will
6662 ret = do_sync_mapping_range(inode->i_mapping, range_start,
6663 range_end - 1, SYNC_FILE_RANGE_WRITE);
6674 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6675 struct ocfs2_dinode *di)
6677 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6678 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6680 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6681 memset(&di->id2, 0, blocksize -
6682 offsetof(struct ocfs2_dinode, id2) -
6685 memset(&di->id2, 0, blocksize -
6686 offsetof(struct ocfs2_dinode, id2));
6689 void ocfs2_dinode_new_extent_list(struct inode *inode,
6690 struct ocfs2_dinode *di)
6692 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6693 di->id2.i_list.l_tree_depth = 0;
6694 di->id2.i_list.l_next_free_rec = 0;
6695 di->id2.i_list.l_count = cpu_to_le16(
6696 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6699 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6701 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6702 struct ocfs2_inline_data *idata = &di->id2.i_data;
6704 spin_lock(&oi->ip_lock);
6705 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6706 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6707 spin_unlock(&oi->ip_lock);
6710 * We clear the entire i_data structure here so that all
6711 * fields can be properly initialized.
6713 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6715 idata->id_count = cpu_to_le16(
6716 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6719 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6720 struct buffer_head *di_bh)
6722 int ret, i, has_data, num_pages = 0;
6724 u64 uninitialized_var(block);
6725 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6726 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6727 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6728 struct ocfs2_alloc_context *data_ac = NULL;
6729 struct page **pages = NULL;
6730 loff_t end = osb->s_clustersize;
6732 has_data = i_size_read(inode) ? 1 : 0;
6735 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6736 sizeof(struct page *), GFP_NOFS);
6737 if (pages == NULL) {
6743 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6750 handle = ocfs2_start_trans(osb, OCFS2_INLINE_TO_EXTENTS_CREDITS);
6751 if (IS_ERR(handle)) {
6752 ret = PTR_ERR(handle);
6757 ret = ocfs2_journal_access(handle, inode, di_bh,
6758 OCFS2_JOURNAL_ACCESS_WRITE);
6766 unsigned int page_end;
6769 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
6777 * Save two copies, one for insert, and one that can
6778 * be changed by ocfs2_map_and_dirty_page() below.
6780 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6783 * Non sparse file systems zero on extend, so no need
6786 if (!ocfs2_sparse_alloc(osb) &&
6787 PAGE_CACHE_SIZE < osb->s_clustersize)
6788 end = PAGE_CACHE_SIZE;
6790 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6797 * This should populate the 1st page for us and mark
6800 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6806 page_end = PAGE_CACHE_SIZE;
6807 if (PAGE_CACHE_SIZE > osb->s_clustersize)
6808 page_end = osb->s_clustersize;
6810 for (i = 0; i < num_pages; i++)
6811 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6812 pages[i], i > 0, &phys);
6815 spin_lock(&oi->ip_lock);
6816 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6817 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6818 spin_unlock(&oi->ip_lock);
6820 ocfs2_dinode_new_extent_list(inode, di);
6822 ocfs2_journal_dirty(handle, di_bh);
6826 * An error at this point should be extremely rare. If
6827 * this proves to be false, we could always re-build
6828 * the in-inode data from our pages.
6830 ret = ocfs2_dinode_insert_extent(osb, handle, inode, di_bh,
6831 0, block, 1, 0, NULL);
6837 inode->i_blocks = ocfs2_inode_sector_count(inode);
6841 ocfs2_commit_trans(osb, handle);
6845 ocfs2_free_alloc_context(data_ac);
6849 ocfs2_unlock_and_free_pages(pages, num_pages);
6857 * It is expected, that by the time you call this function,
6858 * inode->i_size and fe->i_size have been adjusted.
6860 * WARNING: This will kfree the truncate context
6862 int ocfs2_commit_truncate(struct ocfs2_super *osb,
6863 struct inode *inode,
6864 struct buffer_head *fe_bh,
6865 struct ocfs2_truncate_context *tc)
6867 int status, i, credits, tl_sem = 0;
6868 u32 clusters_to_del, new_highest_cpos, range;
6869 struct ocfs2_extent_list *el;
6870 handle_t *handle = NULL;
6871 struct inode *tl_inode = osb->osb_tl_inode;
6872 struct ocfs2_path *path = NULL;
6873 struct ocfs2_dinode *di = (struct ocfs2_dinode *)fe_bh->b_data;
6877 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
6878 i_size_read(inode));
6880 path = ocfs2_new_path(fe_bh, &di->id2.i_list);
6887 ocfs2_extent_map_trunc(inode, new_highest_cpos);
6891 * Check that we still have allocation to delete.
6893 if (OCFS2_I(inode)->ip_clusters == 0) {
6899 * Truncate always works against the rightmost tree branch.
6901 status = ocfs2_find_path(inode, path, UINT_MAX);
6907 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
6908 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
6911 * By now, el will point to the extent list on the bottom most
6912 * portion of this tree. Only the tail record is considered in
6915 * We handle the following cases, in order:
6916 * - empty extent: delete the remaining branch
6917 * - remove the entire record
6918 * - remove a partial record
6919 * - no record needs to be removed (truncate has completed)
6921 el = path_leaf_el(path);
6922 if (le16_to_cpu(el->l_next_free_rec) == 0) {
6923 ocfs2_error(inode->i_sb,
6924 "Inode %llu has empty extent block at %llu\n",
6925 (unsigned long long)OCFS2_I(inode)->ip_blkno,
6926 (unsigned long long)path_leaf_bh(path)->b_blocknr);
6931 i = le16_to_cpu(el->l_next_free_rec) - 1;
6932 range = le32_to_cpu(el->l_recs[i].e_cpos) +
6933 ocfs2_rec_clusters(el, &el->l_recs[i]);
6934 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
6935 clusters_to_del = 0;
6936 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
6937 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
6938 } else if (range > new_highest_cpos) {
6939 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
6940 le32_to_cpu(el->l_recs[i].e_cpos)) -
6947 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
6948 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
6950 mutex_lock(&tl_inode->i_mutex);
6952 /* ocfs2_truncate_log_needs_flush guarantees us at least one
6953 * record is free for use. If there isn't any, we flush to get
6954 * an empty truncate log. */
6955 if (ocfs2_truncate_log_needs_flush(osb)) {
6956 status = __ocfs2_flush_truncate_log(osb);
6963 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
6964 (struct ocfs2_dinode *)fe_bh->b_data,
6966 handle = ocfs2_start_trans(osb, credits);
6967 if (IS_ERR(handle)) {
6968 status = PTR_ERR(handle);
6974 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
6981 mutex_unlock(&tl_inode->i_mutex);
6984 ocfs2_commit_trans(osb, handle);
6987 ocfs2_reinit_path(path, 1);
6990 * The check above will catch the case where we've truncated
6991 * away all allocation.
6997 ocfs2_schedule_truncate_log_flush(osb, 1);
7000 mutex_unlock(&tl_inode->i_mutex);
7003 ocfs2_commit_trans(osb, handle);
7005 ocfs2_run_deallocs(osb, &tc->tc_dealloc);
7007 ocfs2_free_path(path);
7009 /* This will drop the ext_alloc cluster lock for us */
7010 ocfs2_free_truncate_context(tc);
7017 * Expects the inode to already be locked.
7019 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
7020 struct inode *inode,
7021 struct buffer_head *fe_bh,
7022 struct ocfs2_truncate_context **tc)
7025 unsigned int new_i_clusters;
7026 struct ocfs2_dinode *fe;
7027 struct ocfs2_extent_block *eb;
7028 struct buffer_head *last_eb_bh = NULL;
7034 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
7035 i_size_read(inode));
7036 fe = (struct ocfs2_dinode *) fe_bh->b_data;
7038 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
7039 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
7040 (unsigned long long)le64_to_cpu(fe->i_size));
7042 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
7048 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
7050 if (fe->id2.i_list.l_tree_depth) {
7051 status = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
7052 &last_eb_bh, OCFS2_BH_CACHED, inode);
7057 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
7058 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
7059 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
7067 (*tc)->tc_last_eb_bh = last_eb_bh;
7073 ocfs2_free_truncate_context(*tc);
7081 * 'start' is inclusive, 'end' is not.
7083 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7084 unsigned int start, unsigned int end, int trunc)
7087 unsigned int numbytes;
7089 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7090 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7091 struct ocfs2_inline_data *idata = &di->id2.i_data;
7093 if (end > i_size_read(inode))
7094 end = i_size_read(inode);
7096 BUG_ON(start >= end);
7098 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7099 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7100 !ocfs2_supports_inline_data(osb)) {
7101 ocfs2_error(inode->i_sb,
7102 "Inline data flags for inode %llu don't agree! "
7103 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7104 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7105 le16_to_cpu(di->i_dyn_features),
7106 OCFS2_I(inode)->ip_dyn_features,
7107 osb->s_feature_incompat);
7112 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7113 if (IS_ERR(handle)) {
7114 ret = PTR_ERR(handle);
7119 ret = ocfs2_journal_access(handle, inode, di_bh,
7120 OCFS2_JOURNAL_ACCESS_WRITE);
7126 numbytes = end - start;
7127 memset(idata->id_data + start, 0, numbytes);
7130 * No need to worry about the data page here - it's been
7131 * truncated already and inline data doesn't need it for
7132 * pushing zero's to disk, so we'll let readpage pick it up
7136 i_size_write(inode, start);
7137 di->i_size = cpu_to_le64(start);
7140 inode->i_blocks = ocfs2_inode_sector_count(inode);
7141 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7143 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7144 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7146 ocfs2_journal_dirty(handle, di_bh);
7149 ocfs2_commit_trans(osb, handle);
7155 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
7158 * The caller is responsible for completing deallocation
7159 * before freeing the context.
7161 if (tc->tc_dealloc.c_first_suballocator != NULL)
7163 "Truncate completion has non-empty dealloc context\n");
7165 if (tc->tc_last_eb_bh)
7166 brelse(tc->tc_last_eb_bh);