1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Extent allocs and frees
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
32 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
33 #include <cluster/masklog.h>
40 #include "extent_map.h"
43 #include "localalloc.h"
50 #include "buffer_head_io.h"
54 * Operations for a specific extent tree type.
56 * To implement an on-disk btree (extent tree) type in ocfs2, add
57 * an ocfs2_extent_tree_operations structure and the matching
58 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
59 * for the allocation portion of the extent tree.
61 struct ocfs2_extent_tree_operations {
63 * last_eb_blk is the block number of the right most leaf extent
64 * block. Most on-disk structures containing an extent tree store
65 * this value for fast access. The ->eo_set_last_eb_blk() and
66 * ->eo_get_last_eb_blk() operations access this value. They are
69 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
71 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
74 * The on-disk structure usually keeps track of how many total
75 * clusters are stored in this extent tree. This function updates
76 * that value. new_clusters is the delta, and must be
77 * added to the total. Required.
79 void (*eo_update_clusters)(struct inode *inode,
80 struct ocfs2_extent_tree *et,
84 * If ->eo_insert_check() exists, it is called before rec is
85 * inserted into the extent tree. It is optional.
87 int (*eo_insert_check)(struct inode *inode,
88 struct ocfs2_extent_tree *et,
89 struct ocfs2_extent_rec *rec);
90 int (*eo_sanity_check)(struct inode *inode, struct ocfs2_extent_tree *et);
93 * --------------------------------------------------------------
94 * The remaining are internal to ocfs2_extent_tree and don't have
99 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
102 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
105 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
106 * it exists. If it does not, et->et_max_leaf_clusters is set
107 * to 0 (unlimited). Optional.
109 void (*eo_fill_max_leaf_clusters)(struct inode *inode,
110 struct ocfs2_extent_tree *et);
115 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
118 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
119 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
121 static void ocfs2_dinode_update_clusters(struct inode *inode,
122 struct ocfs2_extent_tree *et,
124 static int ocfs2_dinode_insert_check(struct inode *inode,
125 struct ocfs2_extent_tree *et,
126 struct ocfs2_extent_rec *rec);
127 static int ocfs2_dinode_sanity_check(struct inode *inode,
128 struct ocfs2_extent_tree *et);
129 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
130 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
131 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
132 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
133 .eo_update_clusters = ocfs2_dinode_update_clusters,
134 .eo_insert_check = ocfs2_dinode_insert_check,
135 .eo_sanity_check = ocfs2_dinode_sanity_check,
136 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
139 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
142 struct ocfs2_dinode *di = et->et_object;
144 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
145 di->i_last_eb_blk = cpu_to_le64(blkno);
148 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
150 struct ocfs2_dinode *di = et->et_object;
152 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
153 return le64_to_cpu(di->i_last_eb_blk);
156 static void ocfs2_dinode_update_clusters(struct inode *inode,
157 struct ocfs2_extent_tree *et,
160 struct ocfs2_dinode *di = et->et_object;
162 le32_add_cpu(&di->i_clusters, clusters);
163 spin_lock(&OCFS2_I(inode)->ip_lock);
164 OCFS2_I(inode)->ip_clusters = le32_to_cpu(di->i_clusters);
165 spin_unlock(&OCFS2_I(inode)->ip_lock);
168 static int ocfs2_dinode_insert_check(struct inode *inode,
169 struct ocfs2_extent_tree *et,
170 struct ocfs2_extent_rec *rec)
172 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
174 BUG_ON(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL);
175 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
176 (OCFS2_I(inode)->ip_clusters != rec->e_cpos),
177 "Device %s, asking for sparse allocation: inode %llu, "
178 "cpos %u, clusters %u\n",
180 (unsigned long long)OCFS2_I(inode)->ip_blkno,
182 OCFS2_I(inode)->ip_clusters);
187 static int ocfs2_dinode_sanity_check(struct inode *inode,
188 struct ocfs2_extent_tree *et)
191 struct ocfs2_dinode *di;
193 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
196 if (!OCFS2_IS_VALID_DINODE(di)) {
198 ocfs2_error(inode->i_sb,
199 "Inode %llu has invalid path root",
200 (unsigned long long)OCFS2_I(inode)->ip_blkno);
206 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
208 struct ocfs2_dinode *di = et->et_object;
210 et->et_root_el = &di->id2.i_list;
214 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
216 struct ocfs2_xattr_value_root *xv = et->et_object;
218 et->et_root_el = &xv->xr_list;
221 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
224 struct ocfs2_xattr_value_root *xv =
225 (struct ocfs2_xattr_value_root *)et->et_object;
227 xv->xr_last_eb_blk = cpu_to_le64(blkno);
230 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
232 struct ocfs2_xattr_value_root *xv =
233 (struct ocfs2_xattr_value_root *) et->et_object;
235 return le64_to_cpu(xv->xr_last_eb_blk);
238 static void ocfs2_xattr_value_update_clusters(struct inode *inode,
239 struct ocfs2_extent_tree *et,
242 struct ocfs2_xattr_value_root *xv =
243 (struct ocfs2_xattr_value_root *)et->et_object;
245 le32_add_cpu(&xv->xr_clusters, clusters);
248 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
249 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
250 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
251 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
252 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
255 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
257 struct ocfs2_xattr_block *xb = et->et_object;
259 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
262 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct inode *inode,
263 struct ocfs2_extent_tree *et)
265 et->et_max_leaf_clusters =
266 ocfs2_clusters_for_bytes(inode->i_sb,
267 OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
270 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
273 struct ocfs2_xattr_block *xb = et->et_object;
274 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
276 xt->xt_last_eb_blk = cpu_to_le64(blkno);
279 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
281 struct ocfs2_xattr_block *xb = et->et_object;
282 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
284 return le64_to_cpu(xt->xt_last_eb_blk);
287 static void ocfs2_xattr_tree_update_clusters(struct inode *inode,
288 struct ocfs2_extent_tree *et,
291 struct ocfs2_xattr_block *xb = et->et_object;
293 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
296 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
297 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
298 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
299 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
300 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
301 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
304 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
306 struct buffer_head *bh,
308 struct ocfs2_extent_tree_operations *ops)
313 obj = (void *)bh->b_data;
316 et->et_ops->eo_fill_root_el(et);
317 if (!et->et_ops->eo_fill_max_leaf_clusters)
318 et->et_max_leaf_clusters = 0;
320 et->et_ops->eo_fill_max_leaf_clusters(inode, et);
323 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
325 struct buffer_head *bh)
327 __ocfs2_init_extent_tree(et, inode, bh, NULL, &ocfs2_dinode_et_ops);
330 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
332 struct buffer_head *bh)
334 __ocfs2_init_extent_tree(et, inode, bh, NULL,
335 &ocfs2_xattr_tree_et_ops);
338 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
340 struct buffer_head *bh,
341 struct ocfs2_xattr_value_root *xv)
343 __ocfs2_init_extent_tree(et, inode, bh, xv,
344 &ocfs2_xattr_value_et_ops);
347 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
350 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
353 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
355 return et->et_ops->eo_get_last_eb_blk(et);
358 static inline void ocfs2_et_update_clusters(struct inode *inode,
359 struct ocfs2_extent_tree *et,
362 et->et_ops->eo_update_clusters(inode, et, clusters);
365 static inline int ocfs2_et_insert_check(struct inode *inode,
366 struct ocfs2_extent_tree *et,
367 struct ocfs2_extent_rec *rec)
371 if (et->et_ops->eo_insert_check)
372 ret = et->et_ops->eo_insert_check(inode, et, rec);
376 static inline int ocfs2_et_sanity_check(struct inode *inode,
377 struct ocfs2_extent_tree *et)
381 if (et->et_ops->eo_sanity_check)
382 ret = et->et_ops->eo_sanity_check(inode, et);
386 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
387 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
388 struct ocfs2_extent_block *eb);
391 * Structures which describe a path through a btree, and functions to
394 * The idea here is to be as generic as possible with the tree
397 struct ocfs2_path_item {
398 struct buffer_head *bh;
399 struct ocfs2_extent_list *el;
402 #define OCFS2_MAX_PATH_DEPTH 5
406 struct ocfs2_path_item p_node[OCFS2_MAX_PATH_DEPTH];
409 #define path_root_bh(_path) ((_path)->p_node[0].bh)
410 #define path_root_el(_path) ((_path)->p_node[0].el)
411 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
412 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
413 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
416 * Reset the actual path elements so that we can re-use the structure
417 * to build another path. Generally, this involves freeing the buffer
420 static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
422 int i, start = 0, depth = 0;
423 struct ocfs2_path_item *node;
428 for(i = start; i < path_num_items(path); i++) {
429 node = &path->p_node[i];
437 * Tree depth may change during truncate, or insert. If we're
438 * keeping the root extent list, then make sure that our path
439 * structure reflects the proper depth.
442 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
444 path->p_tree_depth = depth;
447 static void ocfs2_free_path(struct ocfs2_path *path)
450 ocfs2_reinit_path(path, 0);
456 * All the elements of src into dest. After this call, src could be freed
457 * without affecting dest.
459 * Both paths should have the same root. Any non-root elements of dest
462 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
466 BUG_ON(path_root_bh(dest) != path_root_bh(src));
467 BUG_ON(path_root_el(dest) != path_root_el(src));
469 ocfs2_reinit_path(dest, 1);
471 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
472 dest->p_node[i].bh = src->p_node[i].bh;
473 dest->p_node[i].el = src->p_node[i].el;
475 if (dest->p_node[i].bh)
476 get_bh(dest->p_node[i].bh);
481 * Make the *dest path the same as src and re-initialize src path to
484 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
488 BUG_ON(path_root_bh(dest) != path_root_bh(src));
490 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
491 brelse(dest->p_node[i].bh);
493 dest->p_node[i].bh = src->p_node[i].bh;
494 dest->p_node[i].el = src->p_node[i].el;
496 src->p_node[i].bh = NULL;
497 src->p_node[i].el = NULL;
502 * Insert an extent block at given index.
504 * This will not take an additional reference on eb_bh.
506 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
507 struct buffer_head *eb_bh)
509 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
512 * Right now, no root bh is an extent block, so this helps
513 * catch code errors with dinode trees. The assertion can be
514 * safely removed if we ever need to insert extent block
515 * structures at the root.
519 path->p_node[index].bh = eb_bh;
520 path->p_node[index].el = &eb->h_list;
523 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
524 struct ocfs2_extent_list *root_el)
526 struct ocfs2_path *path;
528 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
530 path = kzalloc(sizeof(*path), GFP_NOFS);
532 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
534 path_root_bh(path) = root_bh;
535 path_root_el(path) = root_el;
542 * Convenience function to journal all components in a path.
544 static int ocfs2_journal_access_path(struct inode *inode, handle_t *handle,
545 struct ocfs2_path *path)
552 for(i = 0; i < path_num_items(path); i++) {
553 ret = ocfs2_journal_access(handle, inode, path->p_node[i].bh,
554 OCFS2_JOURNAL_ACCESS_WRITE);
566 * Return the index of the extent record which contains cluster #v_cluster.
567 * -1 is returned if it was not found.
569 * Should work fine on interior and exterior nodes.
571 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
575 struct ocfs2_extent_rec *rec;
576 u32 rec_end, rec_start, clusters;
578 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
579 rec = &el->l_recs[i];
581 rec_start = le32_to_cpu(rec->e_cpos);
582 clusters = ocfs2_rec_clusters(el, rec);
584 rec_end = rec_start + clusters;
586 if (v_cluster >= rec_start && v_cluster < rec_end) {
595 enum ocfs2_contig_type {
604 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
605 * ocfs2_extent_contig only work properly against leaf nodes!
607 static int ocfs2_block_extent_contig(struct super_block *sb,
608 struct ocfs2_extent_rec *ext,
611 u64 blk_end = le64_to_cpu(ext->e_blkno);
613 blk_end += ocfs2_clusters_to_blocks(sb,
614 le16_to_cpu(ext->e_leaf_clusters));
616 return blkno == blk_end;
619 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
620 struct ocfs2_extent_rec *right)
624 left_range = le32_to_cpu(left->e_cpos) +
625 le16_to_cpu(left->e_leaf_clusters);
627 return (left_range == le32_to_cpu(right->e_cpos));
630 static enum ocfs2_contig_type
631 ocfs2_extent_contig(struct inode *inode,
632 struct ocfs2_extent_rec *ext,
633 struct ocfs2_extent_rec *insert_rec)
635 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
638 * Refuse to coalesce extent records with different flag
639 * fields - we don't want to mix unwritten extents with user
642 if (ext->e_flags != insert_rec->e_flags)
645 if (ocfs2_extents_adjacent(ext, insert_rec) &&
646 ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
649 blkno = le64_to_cpu(ext->e_blkno);
650 if (ocfs2_extents_adjacent(insert_rec, ext) &&
651 ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno))
658 * NOTE: We can have pretty much any combination of contiguousness and
661 * The usefulness of APPEND_TAIL is more in that it lets us know that
662 * we'll have to update the path to that leaf.
664 enum ocfs2_append_type {
669 enum ocfs2_split_type {
675 struct ocfs2_insert_type {
676 enum ocfs2_split_type ins_split;
677 enum ocfs2_append_type ins_appending;
678 enum ocfs2_contig_type ins_contig;
679 int ins_contig_index;
683 struct ocfs2_merge_ctxt {
684 enum ocfs2_contig_type c_contig_type;
685 int c_has_empty_extent;
686 int c_split_covers_rec;
690 * How many free extents have we got before we need more meta data?
692 int ocfs2_num_free_extents(struct ocfs2_super *osb,
694 struct ocfs2_extent_tree *et)
697 struct ocfs2_extent_list *el = NULL;
698 struct ocfs2_extent_block *eb;
699 struct buffer_head *eb_bh = NULL;
705 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
708 retval = ocfs2_read_block(osb, last_eb_blk,
709 &eb_bh, OCFS2_BH_CACHED, inode);
714 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
718 BUG_ON(el->l_tree_depth != 0);
720 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
728 /* expects array to already be allocated
730 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
733 static int ocfs2_create_new_meta_bhs(struct ocfs2_super *osb,
737 struct ocfs2_alloc_context *meta_ac,
738 struct buffer_head *bhs[])
740 int count, status, i;
741 u16 suballoc_bit_start;
744 struct ocfs2_extent_block *eb;
749 while (count < wanted) {
750 status = ocfs2_claim_metadata(osb,
762 for(i = count; i < (num_got + count); i++) {
763 bhs[i] = sb_getblk(osb->sb, first_blkno);
764 if (bhs[i] == NULL) {
769 ocfs2_set_new_buffer_uptodate(inode, bhs[i]);
771 status = ocfs2_journal_access(handle, inode, bhs[i],
772 OCFS2_JOURNAL_ACCESS_CREATE);
778 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
779 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
780 /* Ok, setup the minimal stuff here. */
781 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
782 eb->h_blkno = cpu_to_le64(first_blkno);
783 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
784 eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
785 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
787 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
789 suballoc_bit_start++;
792 /* We'll also be dirtied by the caller, so
793 * this isn't absolutely necessary. */
794 status = ocfs2_journal_dirty(handle, bhs[i]);
807 for(i = 0; i < wanted; i++) {
817 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
819 * Returns the sum of the rightmost extent rec logical offset and
822 * ocfs2_add_branch() uses this to determine what logical cluster
823 * value should be populated into the leftmost new branch records.
825 * ocfs2_shift_tree_depth() uses this to determine the # clusters
826 * value for the new topmost tree record.
828 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
832 i = le16_to_cpu(el->l_next_free_rec) - 1;
834 return le32_to_cpu(el->l_recs[i].e_cpos) +
835 ocfs2_rec_clusters(el, &el->l_recs[i]);
839 * Add an entire tree branch to our inode. eb_bh is the extent block
840 * to start at, if we don't want to start the branch at the dinode
843 * last_eb_bh is required as we have to update it's next_leaf pointer
844 * for the new last extent block.
846 * the new branch will be 'empty' in the sense that every block will
847 * contain a single record with cluster count == 0.
849 static int ocfs2_add_branch(struct ocfs2_super *osb,
852 struct ocfs2_extent_tree *et,
853 struct buffer_head *eb_bh,
854 struct buffer_head **last_eb_bh,
855 struct ocfs2_alloc_context *meta_ac)
857 int status, new_blocks, i;
858 u64 next_blkno, new_last_eb_blk;
859 struct buffer_head *bh;
860 struct buffer_head **new_eb_bhs = NULL;
861 struct ocfs2_extent_block *eb;
862 struct ocfs2_extent_list *eb_el;
863 struct ocfs2_extent_list *el;
868 BUG_ON(!last_eb_bh || !*last_eb_bh);
871 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
876 /* we never add a branch to a leaf. */
877 BUG_ON(!el->l_tree_depth);
879 new_blocks = le16_to_cpu(el->l_tree_depth);
881 /* allocate the number of new eb blocks we need */
882 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
890 status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
891 meta_ac, new_eb_bhs);
897 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
898 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
900 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
901 * linked with the rest of the tree.
902 * conversly, new_eb_bhs[0] is the new bottommost leaf.
904 * when we leave the loop, new_last_eb_blk will point to the
905 * newest leaf, and next_blkno will point to the topmost extent
907 next_blkno = new_last_eb_blk = 0;
908 for(i = 0; i < new_blocks; i++) {
910 eb = (struct ocfs2_extent_block *) bh->b_data;
911 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
912 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
918 status = ocfs2_journal_access(handle, inode, bh,
919 OCFS2_JOURNAL_ACCESS_CREATE);
925 eb->h_next_leaf_blk = 0;
926 eb_el->l_tree_depth = cpu_to_le16(i);
927 eb_el->l_next_free_rec = cpu_to_le16(1);
929 * This actually counts as an empty extent as
932 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
933 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
935 * eb_el isn't always an interior node, but even leaf
936 * nodes want a zero'd flags and reserved field so
937 * this gets the whole 32 bits regardless of use.
939 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
940 if (!eb_el->l_tree_depth)
941 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
943 status = ocfs2_journal_dirty(handle, bh);
949 next_blkno = le64_to_cpu(eb->h_blkno);
952 /* This is a bit hairy. We want to update up to three blocks
953 * here without leaving any of them in an inconsistent state
954 * in case of error. We don't have to worry about
955 * journal_dirty erroring as it won't unless we've aborted the
956 * handle (in which case we would never be here) so reserving
957 * the write with journal_access is all we need to do. */
958 status = ocfs2_journal_access(handle, inode, *last_eb_bh,
959 OCFS2_JOURNAL_ACCESS_WRITE);
964 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
965 OCFS2_JOURNAL_ACCESS_WRITE);
971 status = ocfs2_journal_access(handle, inode, eb_bh,
972 OCFS2_JOURNAL_ACCESS_WRITE);
979 /* Link the new branch into the rest of the tree (el will
980 * either be on the root_bh, or the extent block passed in. */
981 i = le16_to_cpu(el->l_next_free_rec);
982 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
983 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
984 el->l_recs[i].e_int_clusters = 0;
985 le16_add_cpu(&el->l_next_free_rec, 1);
987 /* fe needs a new last extent block pointer, as does the
988 * next_leaf on the previously last-extent-block. */
989 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
991 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
992 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
994 status = ocfs2_journal_dirty(handle, *last_eb_bh);
997 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1001 status = ocfs2_journal_dirty(handle, eb_bh);
1007 * Some callers want to track the rightmost leaf so pass it
1010 brelse(*last_eb_bh);
1011 get_bh(new_eb_bhs[0]);
1012 *last_eb_bh = new_eb_bhs[0];
1017 for (i = 0; i < new_blocks; i++)
1018 brelse(new_eb_bhs[i]);
1027 * adds another level to the allocation tree.
1028 * returns back the new extent block so you can add a branch to it
1031 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
1033 struct inode *inode,
1034 struct ocfs2_extent_tree *et,
1035 struct ocfs2_alloc_context *meta_ac,
1036 struct buffer_head **ret_new_eb_bh)
1040 struct buffer_head *new_eb_bh = NULL;
1041 struct ocfs2_extent_block *eb;
1042 struct ocfs2_extent_list *root_el;
1043 struct ocfs2_extent_list *eb_el;
1047 status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
1054 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1055 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1056 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1061 eb_el = &eb->h_list;
1062 root_el = et->et_root_el;
1064 status = ocfs2_journal_access(handle, inode, new_eb_bh,
1065 OCFS2_JOURNAL_ACCESS_CREATE);
1071 /* copy the root extent list data into the new extent block */
1072 eb_el->l_tree_depth = root_el->l_tree_depth;
1073 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1074 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1075 eb_el->l_recs[i] = root_el->l_recs[i];
1077 status = ocfs2_journal_dirty(handle, new_eb_bh);
1083 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
1084 OCFS2_JOURNAL_ACCESS_WRITE);
1090 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1092 /* update root_bh now */
1093 le16_add_cpu(&root_el->l_tree_depth, 1);
1094 root_el->l_recs[0].e_cpos = 0;
1095 root_el->l_recs[0].e_blkno = eb->h_blkno;
1096 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1097 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1098 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1099 root_el->l_next_free_rec = cpu_to_le16(1);
1101 /* If this is our 1st tree depth shift, then last_eb_blk
1102 * becomes the allocated extent block */
1103 if (root_el->l_tree_depth == cpu_to_le16(1))
1104 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1106 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1112 *ret_new_eb_bh = new_eb_bh;
1123 * Should only be called when there is no space left in any of the
1124 * leaf nodes. What we want to do is find the lowest tree depth
1125 * non-leaf extent block with room for new records. There are three
1126 * valid results of this search:
1128 * 1) a lowest extent block is found, then we pass it back in
1129 * *lowest_eb_bh and return '0'
1131 * 2) the search fails to find anything, but the root_el has room. We
1132 * pass NULL back in *lowest_eb_bh, but still return '0'
1134 * 3) the search fails to find anything AND the root_el is full, in
1135 * which case we return > 0
1137 * return status < 0 indicates an error.
1139 static int ocfs2_find_branch_target(struct ocfs2_super *osb,
1140 struct inode *inode,
1141 struct ocfs2_extent_tree *et,
1142 struct buffer_head **target_bh)
1146 struct ocfs2_extent_block *eb;
1147 struct ocfs2_extent_list *el;
1148 struct buffer_head *bh = NULL;
1149 struct buffer_head *lowest_bh = NULL;
1155 el = et->et_root_el;
1157 while(le16_to_cpu(el->l_tree_depth) > 1) {
1158 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1159 ocfs2_error(inode->i_sb, "Dinode %llu has empty "
1160 "extent list (next_free_rec == 0)",
1161 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1165 i = le16_to_cpu(el->l_next_free_rec) - 1;
1166 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1168 ocfs2_error(inode->i_sb, "Dinode %llu has extent "
1169 "list where extent # %d has no physical "
1171 (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
1179 status = ocfs2_read_block(osb, blkno, &bh, OCFS2_BH_CACHED,
1186 eb = (struct ocfs2_extent_block *) bh->b_data;
1187 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1188 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1194 if (le16_to_cpu(el->l_next_free_rec) <
1195 le16_to_cpu(el->l_count)) {
1202 /* If we didn't find one and the fe doesn't have any room,
1203 * then return '1' */
1204 el = et->et_root_el;
1205 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1208 *target_bh = lowest_bh;
1217 * Grow a b-tree so that it has more records.
1219 * We might shift the tree depth in which case existing paths should
1220 * be considered invalid.
1222 * Tree depth after the grow is returned via *final_depth.
1224 * *last_eb_bh will be updated by ocfs2_add_branch().
1226 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
1227 struct ocfs2_extent_tree *et, int *final_depth,
1228 struct buffer_head **last_eb_bh,
1229 struct ocfs2_alloc_context *meta_ac)
1232 struct ocfs2_extent_list *el = et->et_root_el;
1233 int depth = le16_to_cpu(el->l_tree_depth);
1234 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1235 struct buffer_head *bh = NULL;
1237 BUG_ON(meta_ac == NULL);
1239 shift = ocfs2_find_branch_target(osb, inode, et, &bh);
1246 /* We traveled all the way to the bottom of the allocation tree
1247 * and didn't find room for any more extents - we need to add
1248 * another tree level */
1251 mlog(0, "need to shift tree depth (current = %d)\n", depth);
1253 /* ocfs2_shift_tree_depth will return us a buffer with
1254 * the new extent block (so we can pass that to
1255 * ocfs2_add_branch). */
1256 ret = ocfs2_shift_tree_depth(osb, handle, inode, et,
1265 * Special case: we have room now if we shifted from
1266 * tree_depth 0, so no more work needs to be done.
1268 * We won't be calling add_branch, so pass
1269 * back *last_eb_bh as the new leaf. At depth
1270 * zero, it should always be null so there's
1271 * no reason to brelse.
1273 BUG_ON(*last_eb_bh);
1280 /* call ocfs2_add_branch to add the final part of the tree with
1282 mlog(0, "add branch. bh = %p\n", bh);
1283 ret = ocfs2_add_branch(osb, handle, inode, et, bh, last_eb_bh,
1292 *final_depth = depth;
1298 * This function will discard the rightmost extent record.
1300 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1302 int next_free = le16_to_cpu(el->l_next_free_rec);
1303 int count = le16_to_cpu(el->l_count);
1304 unsigned int num_bytes;
1307 /* This will cause us to go off the end of our extent list. */
1308 BUG_ON(next_free >= count);
1310 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1312 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1315 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1316 struct ocfs2_extent_rec *insert_rec)
1318 int i, insert_index, next_free, has_empty, num_bytes;
1319 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1320 struct ocfs2_extent_rec *rec;
1322 next_free = le16_to_cpu(el->l_next_free_rec);
1323 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1327 /* The tree code before us didn't allow enough room in the leaf. */
1328 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1331 * The easiest way to approach this is to just remove the
1332 * empty extent and temporarily decrement next_free.
1336 * If next_free was 1 (only an empty extent), this
1337 * loop won't execute, which is fine. We still want
1338 * the decrement above to happen.
1340 for(i = 0; i < (next_free - 1); i++)
1341 el->l_recs[i] = el->l_recs[i+1];
1347 * Figure out what the new record index should be.
1349 for(i = 0; i < next_free; i++) {
1350 rec = &el->l_recs[i];
1352 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1357 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1358 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
1360 BUG_ON(insert_index < 0);
1361 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1362 BUG_ON(insert_index > next_free);
1365 * No need to memmove if we're just adding to the tail.
1367 if (insert_index != next_free) {
1368 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1370 num_bytes = next_free - insert_index;
1371 num_bytes *= sizeof(struct ocfs2_extent_rec);
1372 memmove(&el->l_recs[insert_index + 1],
1373 &el->l_recs[insert_index],
1378 * Either we had an empty extent, and need to re-increment or
1379 * there was no empty extent on a non full rightmost leaf node,
1380 * in which case we still need to increment.
1383 el->l_next_free_rec = cpu_to_le16(next_free);
1385 * Make sure none of the math above just messed up our tree.
1387 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1389 el->l_recs[insert_index] = *insert_rec;
1393 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1395 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1397 BUG_ON(num_recs == 0);
1399 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1401 size = num_recs * sizeof(struct ocfs2_extent_rec);
1402 memmove(&el->l_recs[0], &el->l_recs[1], size);
1403 memset(&el->l_recs[num_recs], 0,
1404 sizeof(struct ocfs2_extent_rec));
1405 el->l_next_free_rec = cpu_to_le16(num_recs);
1410 * Create an empty extent record .
1412 * l_next_free_rec may be updated.
1414 * If an empty extent already exists do nothing.
1416 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1418 int next_free = le16_to_cpu(el->l_next_free_rec);
1420 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1425 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1428 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1429 "Asked to create an empty extent in a full list:\n"
1430 "count = %u, tree depth = %u",
1431 le16_to_cpu(el->l_count),
1432 le16_to_cpu(el->l_tree_depth));
1434 ocfs2_shift_records_right(el);
1437 le16_add_cpu(&el->l_next_free_rec, 1);
1438 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1442 * For a rotation which involves two leaf nodes, the "root node" is
1443 * the lowest level tree node which contains a path to both leafs. This
1444 * resulting set of information can be used to form a complete "subtree"
1446 * This function is passed two full paths from the dinode down to a
1447 * pair of adjacent leaves. It's task is to figure out which path
1448 * index contains the subtree root - this can be the root index itself
1449 * in a worst-case rotation.
1451 * The array index of the subtree root is passed back.
1453 static int ocfs2_find_subtree_root(struct inode *inode,
1454 struct ocfs2_path *left,
1455 struct ocfs2_path *right)
1460 * Check that the caller passed in two paths from the same tree.
1462 BUG_ON(path_root_bh(left) != path_root_bh(right));
1468 * The caller didn't pass two adjacent paths.
1470 mlog_bug_on_msg(i > left->p_tree_depth,
1471 "Inode %lu, left depth %u, right depth %u\n"
1472 "left leaf blk %llu, right leaf blk %llu\n",
1473 inode->i_ino, left->p_tree_depth,
1474 right->p_tree_depth,
1475 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1476 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1477 } while (left->p_node[i].bh->b_blocknr ==
1478 right->p_node[i].bh->b_blocknr);
1483 typedef void (path_insert_t)(void *, struct buffer_head *);
1486 * Traverse a btree path in search of cpos, starting at root_el.
1488 * This code can be called with a cpos larger than the tree, in which
1489 * case it will return the rightmost path.
1491 static int __ocfs2_find_path(struct inode *inode,
1492 struct ocfs2_extent_list *root_el, u32 cpos,
1493 path_insert_t *func, void *data)
1498 struct buffer_head *bh = NULL;
1499 struct ocfs2_extent_block *eb;
1500 struct ocfs2_extent_list *el;
1501 struct ocfs2_extent_rec *rec;
1502 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1505 while (el->l_tree_depth) {
1506 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1507 ocfs2_error(inode->i_sb,
1508 "Inode %llu has empty extent list at "
1510 (unsigned long long)oi->ip_blkno,
1511 le16_to_cpu(el->l_tree_depth));
1517 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1518 rec = &el->l_recs[i];
1521 * In the case that cpos is off the allocation
1522 * tree, this should just wind up returning the
1525 range = le32_to_cpu(rec->e_cpos) +
1526 ocfs2_rec_clusters(el, rec);
1527 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1531 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1533 ocfs2_error(inode->i_sb,
1534 "Inode %llu has bad blkno in extent list "
1535 "at depth %u (index %d)\n",
1536 (unsigned long long)oi->ip_blkno,
1537 le16_to_cpu(el->l_tree_depth), i);
1544 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno,
1545 &bh, OCFS2_BH_CACHED, inode);
1551 eb = (struct ocfs2_extent_block *) bh->b_data;
1553 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1554 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1559 if (le16_to_cpu(el->l_next_free_rec) >
1560 le16_to_cpu(el->l_count)) {
1561 ocfs2_error(inode->i_sb,
1562 "Inode %llu has bad count in extent list "
1563 "at block %llu (next free=%u, count=%u)\n",
1564 (unsigned long long)oi->ip_blkno,
1565 (unsigned long long)bh->b_blocknr,
1566 le16_to_cpu(el->l_next_free_rec),
1567 le16_to_cpu(el->l_count));
1578 * Catch any trailing bh that the loop didn't handle.
1586 * Given an initialized path (that is, it has a valid root extent
1587 * list), this function will traverse the btree in search of the path
1588 * which would contain cpos.
1590 * The path traveled is recorded in the path structure.
1592 * Note that this will not do any comparisons on leaf node extent
1593 * records, so it will work fine in the case that we just added a tree
1596 struct find_path_data {
1598 struct ocfs2_path *path;
1600 static void find_path_ins(void *data, struct buffer_head *bh)
1602 struct find_path_data *fp = data;
1605 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1608 static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
1611 struct find_path_data data;
1615 return __ocfs2_find_path(inode, path_root_el(path), cpos,
1616 find_path_ins, &data);
1619 static void find_leaf_ins(void *data, struct buffer_head *bh)
1621 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1622 struct ocfs2_extent_list *el = &eb->h_list;
1623 struct buffer_head **ret = data;
1625 /* We want to retain only the leaf block. */
1626 if (le16_to_cpu(el->l_tree_depth) == 0) {
1632 * Find the leaf block in the tree which would contain cpos. No
1633 * checking of the actual leaf is done.
1635 * Some paths want to call this instead of allocating a path structure
1636 * and calling ocfs2_find_path().
1638 * This function doesn't handle non btree extent lists.
1640 int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
1641 u32 cpos, struct buffer_head **leaf_bh)
1644 struct buffer_head *bh = NULL;
1646 ret = __ocfs2_find_path(inode, root_el, cpos, find_leaf_ins, &bh);
1658 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1660 * Basically, we've moved stuff around at the bottom of the tree and
1661 * we need to fix up the extent records above the changes to reflect
1664 * left_rec: the record on the left.
1665 * left_child_el: is the child list pointed to by left_rec
1666 * right_rec: the record to the right of left_rec
1667 * right_child_el: is the child list pointed to by right_rec
1669 * By definition, this only works on interior nodes.
1671 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1672 struct ocfs2_extent_list *left_child_el,
1673 struct ocfs2_extent_rec *right_rec,
1674 struct ocfs2_extent_list *right_child_el)
1676 u32 left_clusters, right_end;
1679 * Interior nodes never have holes. Their cpos is the cpos of
1680 * the leftmost record in their child list. Their cluster
1681 * count covers the full theoretical range of their child list
1682 * - the range between their cpos and the cpos of the record
1683 * immediately to their right.
1685 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1686 if (ocfs2_is_empty_extent(&right_child_el->l_recs[0])) {
1687 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1688 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1690 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1691 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1694 * Calculate the rightmost cluster count boundary before
1695 * moving cpos - we will need to adjust clusters after
1696 * updating e_cpos to keep the same highest cluster count.
1698 right_end = le32_to_cpu(right_rec->e_cpos);
1699 right_end += le32_to_cpu(right_rec->e_int_clusters);
1701 right_rec->e_cpos = left_rec->e_cpos;
1702 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1704 right_end -= le32_to_cpu(right_rec->e_cpos);
1705 right_rec->e_int_clusters = cpu_to_le32(right_end);
1709 * Adjust the adjacent root node records involved in a
1710 * rotation. left_el_blkno is passed in as a key so that we can easily
1711 * find it's index in the root list.
1713 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1714 struct ocfs2_extent_list *left_el,
1715 struct ocfs2_extent_list *right_el,
1720 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1721 le16_to_cpu(left_el->l_tree_depth));
1723 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1724 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1729 * The path walking code should have never returned a root and
1730 * two paths which are not adjacent.
1732 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1734 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1735 &root_el->l_recs[i + 1], right_el);
1739 * We've changed a leaf block (in right_path) and need to reflect that
1740 * change back up the subtree.
1742 * This happens in multiple places:
1743 * - When we've moved an extent record from the left path leaf to the right
1744 * path leaf to make room for an empty extent in the left path leaf.
1745 * - When our insert into the right path leaf is at the leftmost edge
1746 * and requires an update of the path immediately to it's left. This
1747 * can occur at the end of some types of rotation and appending inserts.
1748 * - When we've adjusted the last extent record in the left path leaf and the
1749 * 1st extent record in the right path leaf during cross extent block merge.
1751 static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1752 struct ocfs2_path *left_path,
1753 struct ocfs2_path *right_path,
1757 struct ocfs2_extent_list *el, *left_el, *right_el;
1758 struct ocfs2_extent_rec *left_rec, *right_rec;
1759 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1762 * Update the counts and position values within all the
1763 * interior nodes to reflect the leaf rotation we just did.
1765 * The root node is handled below the loop.
1767 * We begin the loop with right_el and left_el pointing to the
1768 * leaf lists and work our way up.
1770 * NOTE: within this loop, left_el and right_el always refer
1771 * to the *child* lists.
1773 left_el = path_leaf_el(left_path);
1774 right_el = path_leaf_el(right_path);
1775 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
1776 mlog(0, "Adjust records at index %u\n", i);
1779 * One nice property of knowing that all of these
1780 * nodes are below the root is that we only deal with
1781 * the leftmost right node record and the rightmost
1784 el = left_path->p_node[i].el;
1785 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
1786 left_rec = &el->l_recs[idx];
1788 el = right_path->p_node[i].el;
1789 right_rec = &el->l_recs[0];
1791 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
1794 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
1798 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
1803 * Setup our list pointers now so that the current
1804 * parents become children in the next iteration.
1806 left_el = left_path->p_node[i].el;
1807 right_el = right_path->p_node[i].el;
1811 * At the root node, adjust the two adjacent records which
1812 * begin our path to the leaves.
1815 el = left_path->p_node[subtree_index].el;
1816 left_el = left_path->p_node[subtree_index + 1].el;
1817 right_el = right_path->p_node[subtree_index + 1].el;
1819 ocfs2_adjust_root_records(el, left_el, right_el,
1820 left_path->p_node[subtree_index + 1].bh->b_blocknr);
1822 root_bh = left_path->p_node[subtree_index].bh;
1824 ret = ocfs2_journal_dirty(handle, root_bh);
1829 static int ocfs2_rotate_subtree_right(struct inode *inode,
1831 struct ocfs2_path *left_path,
1832 struct ocfs2_path *right_path,
1836 struct buffer_head *right_leaf_bh;
1837 struct buffer_head *left_leaf_bh = NULL;
1838 struct buffer_head *root_bh;
1839 struct ocfs2_extent_list *right_el, *left_el;
1840 struct ocfs2_extent_rec move_rec;
1842 left_leaf_bh = path_leaf_bh(left_path);
1843 left_el = path_leaf_el(left_path);
1845 if (left_el->l_next_free_rec != left_el->l_count) {
1846 ocfs2_error(inode->i_sb,
1847 "Inode %llu has non-full interior leaf node %llu"
1849 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1850 (unsigned long long)left_leaf_bh->b_blocknr,
1851 le16_to_cpu(left_el->l_next_free_rec));
1856 * This extent block may already have an empty record, so we
1857 * return early if so.
1859 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
1862 root_bh = left_path->p_node[subtree_index].bh;
1863 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
1865 ret = ocfs2_journal_access(handle, inode, root_bh,
1866 OCFS2_JOURNAL_ACCESS_WRITE);
1872 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
1873 ret = ocfs2_journal_access(handle, inode,
1874 right_path->p_node[i].bh,
1875 OCFS2_JOURNAL_ACCESS_WRITE);
1881 ret = ocfs2_journal_access(handle, inode,
1882 left_path->p_node[i].bh,
1883 OCFS2_JOURNAL_ACCESS_WRITE);
1890 right_leaf_bh = path_leaf_bh(right_path);
1891 right_el = path_leaf_el(right_path);
1893 /* This is a code error, not a disk corruption. */
1894 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
1895 "because rightmost leaf block %llu is empty\n",
1896 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1897 (unsigned long long)right_leaf_bh->b_blocknr);
1899 ocfs2_create_empty_extent(right_el);
1901 ret = ocfs2_journal_dirty(handle, right_leaf_bh);
1907 /* Do the copy now. */
1908 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
1909 move_rec = left_el->l_recs[i];
1910 right_el->l_recs[0] = move_rec;
1913 * Clear out the record we just copied and shift everything
1914 * over, leaving an empty extent in the left leaf.
1916 * We temporarily subtract from next_free_rec so that the
1917 * shift will lose the tail record (which is now defunct).
1919 le16_add_cpu(&left_el->l_next_free_rec, -1);
1920 ocfs2_shift_records_right(left_el);
1921 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1922 le16_add_cpu(&left_el->l_next_free_rec, 1);
1924 ret = ocfs2_journal_dirty(handle, left_leaf_bh);
1930 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
1938 * Given a full path, determine what cpos value would return us a path
1939 * containing the leaf immediately to the left of the current one.
1941 * Will return zero if the path passed in is already the leftmost path.
1943 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
1944 struct ocfs2_path *path, u32 *cpos)
1948 struct ocfs2_extent_list *el;
1950 BUG_ON(path->p_tree_depth == 0);
1954 blkno = path_leaf_bh(path)->b_blocknr;
1956 /* Start at the tree node just above the leaf and work our way up. */
1957 i = path->p_tree_depth - 1;
1959 el = path->p_node[i].el;
1962 * Find the extent record just before the one in our
1965 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
1966 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
1970 * We've determined that the
1971 * path specified is already
1972 * the leftmost one - return a
1978 * The leftmost record points to our
1979 * leaf - we need to travel up the
1985 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
1986 *cpos = *cpos + ocfs2_rec_clusters(el,
1987 &el->l_recs[j - 1]);
1994 * If we got here, we never found a valid node where
1995 * the tree indicated one should be.
1998 "Invalid extent tree at extent block %llu\n",
1999 (unsigned long long)blkno);
2004 blkno = path->p_node[i].bh->b_blocknr;
2013 * Extend the transaction by enough credits to complete the rotation,
2014 * and still leave at least the original number of credits allocated
2015 * to this transaction.
2017 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2019 struct ocfs2_path *path)
2021 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2023 if (handle->h_buffer_credits < credits)
2024 return ocfs2_extend_trans(handle, credits);
2030 * Trap the case where we're inserting into the theoretical range past
2031 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2032 * whose cpos is less than ours into the right leaf.
2034 * It's only necessary to look at the rightmost record of the left
2035 * leaf because the logic that calls us should ensure that the
2036 * theoretical ranges in the path components above the leaves are
2039 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2042 struct ocfs2_extent_list *left_el;
2043 struct ocfs2_extent_rec *rec;
2046 left_el = path_leaf_el(left_path);
2047 next_free = le16_to_cpu(left_el->l_next_free_rec);
2048 rec = &left_el->l_recs[next_free - 1];
2050 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2055 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2057 int next_free = le16_to_cpu(el->l_next_free_rec);
2059 struct ocfs2_extent_rec *rec;
2064 rec = &el->l_recs[0];
2065 if (ocfs2_is_empty_extent(rec)) {
2069 rec = &el->l_recs[1];
2072 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2073 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2079 * Rotate all the records in a btree right one record, starting at insert_cpos.
2081 * The path to the rightmost leaf should be passed in.
2083 * The array is assumed to be large enough to hold an entire path (tree depth).
2085 * Upon succesful return from this function:
2087 * - The 'right_path' array will contain a path to the leaf block
2088 * whose range contains e_cpos.
2089 * - That leaf block will have a single empty extent in list index 0.
2090 * - In the case that the rotation requires a post-insert update,
2091 * *ret_left_path will contain a valid path which can be passed to
2092 * ocfs2_insert_path().
2094 static int ocfs2_rotate_tree_right(struct inode *inode,
2096 enum ocfs2_split_type split,
2098 struct ocfs2_path *right_path,
2099 struct ocfs2_path **ret_left_path)
2101 int ret, start, orig_credits = handle->h_buffer_credits;
2103 struct ocfs2_path *left_path = NULL;
2105 *ret_left_path = NULL;
2107 left_path = ocfs2_new_path(path_root_bh(right_path),
2108 path_root_el(right_path));
2115 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
2121 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
2124 * What we want to do here is:
2126 * 1) Start with the rightmost path.
2128 * 2) Determine a path to the leaf block directly to the left
2131 * 3) Determine the 'subtree root' - the lowest level tree node
2132 * which contains a path to both leaves.
2134 * 4) Rotate the subtree.
2136 * 5) Find the next subtree by considering the left path to be
2137 * the new right path.
2139 * The check at the top of this while loop also accepts
2140 * insert_cpos == cpos because cpos is only a _theoretical_
2141 * value to get us the left path - insert_cpos might very well
2142 * be filling that hole.
2144 * Stop at a cpos of '0' because we either started at the
2145 * leftmost branch (i.e., a tree with one branch and a
2146 * rotation inside of it), or we've gone as far as we can in
2147 * rotating subtrees.
2149 while (cpos && insert_cpos <= cpos) {
2150 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2153 ret = ocfs2_find_path(inode, left_path, cpos);
2159 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2160 path_leaf_bh(right_path),
2161 "Inode %lu: error during insert of %u "
2162 "(left path cpos %u) results in two identical "
2163 "paths ending at %llu\n",
2164 inode->i_ino, insert_cpos, cpos,
2165 (unsigned long long)
2166 path_leaf_bh(left_path)->b_blocknr);
2168 if (split == SPLIT_NONE &&
2169 ocfs2_rotate_requires_path_adjustment(left_path,
2173 * We've rotated the tree as much as we
2174 * should. The rest is up to
2175 * ocfs2_insert_path() to complete, after the
2176 * record insertion. We indicate this
2177 * situation by returning the left path.
2179 * The reason we don't adjust the records here
2180 * before the record insert is that an error
2181 * later might break the rule where a parent
2182 * record e_cpos will reflect the actual
2183 * e_cpos of the 1st nonempty record of the
2186 *ret_left_path = left_path;
2190 start = ocfs2_find_subtree_root(inode, left_path, right_path);
2192 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2194 (unsigned long long) right_path->p_node[start].bh->b_blocknr,
2195 right_path->p_tree_depth);
2197 ret = ocfs2_extend_rotate_transaction(handle, start,
2198 orig_credits, right_path);
2204 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
2211 if (split != SPLIT_NONE &&
2212 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2215 * A rotate moves the rightmost left leaf
2216 * record over to the leftmost right leaf
2217 * slot. If we're doing an extent split
2218 * instead of a real insert, then we have to
2219 * check that the extent to be split wasn't
2220 * just moved over. If it was, then we can
2221 * exit here, passing left_path back -
2222 * ocfs2_split_extent() is smart enough to
2223 * search both leaves.
2225 *ret_left_path = left_path;
2230 * There is no need to re-read the next right path
2231 * as we know that it'll be our current left
2232 * path. Optimize by copying values instead.
2234 ocfs2_mv_path(right_path, left_path);
2236 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
2245 ocfs2_free_path(left_path);
2251 static void ocfs2_update_edge_lengths(struct inode *inode, handle_t *handle,
2252 struct ocfs2_path *path)
2255 struct ocfs2_extent_rec *rec;
2256 struct ocfs2_extent_list *el;
2257 struct ocfs2_extent_block *eb;
2260 /* Path should always be rightmost. */
2261 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2262 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2265 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2266 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2267 rec = &el->l_recs[idx];
2268 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2270 for (i = 0; i < path->p_tree_depth; i++) {
2271 el = path->p_node[i].el;
2272 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2273 rec = &el->l_recs[idx];
2275 rec->e_int_clusters = cpu_to_le32(range);
2276 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2278 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2282 static void ocfs2_unlink_path(struct inode *inode, handle_t *handle,
2283 struct ocfs2_cached_dealloc_ctxt *dealloc,
2284 struct ocfs2_path *path, int unlink_start)
2287 struct ocfs2_extent_block *eb;
2288 struct ocfs2_extent_list *el;
2289 struct buffer_head *bh;
2291 for(i = unlink_start; i < path_num_items(path); i++) {
2292 bh = path->p_node[i].bh;
2294 eb = (struct ocfs2_extent_block *)bh->b_data;
2296 * Not all nodes might have had their final count
2297 * decremented by the caller - handle this here.
2300 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2302 "Inode %llu, attempted to remove extent block "
2303 "%llu with %u records\n",
2304 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2305 (unsigned long long)le64_to_cpu(eb->h_blkno),
2306 le16_to_cpu(el->l_next_free_rec));
2308 ocfs2_journal_dirty(handle, bh);
2309 ocfs2_remove_from_cache(inode, bh);
2313 el->l_next_free_rec = 0;
2314 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2316 ocfs2_journal_dirty(handle, bh);
2318 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2322 ocfs2_remove_from_cache(inode, bh);
2326 static void ocfs2_unlink_subtree(struct inode *inode, handle_t *handle,
2327 struct ocfs2_path *left_path,
2328 struct ocfs2_path *right_path,
2330 struct ocfs2_cached_dealloc_ctxt *dealloc)
2333 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2334 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2335 struct ocfs2_extent_list *el;
2336 struct ocfs2_extent_block *eb;
2338 el = path_leaf_el(left_path);
2340 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2342 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2343 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2346 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2348 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2349 le16_add_cpu(&root_el->l_next_free_rec, -1);
2351 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2352 eb->h_next_leaf_blk = 0;
2354 ocfs2_journal_dirty(handle, root_bh);
2355 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2357 ocfs2_unlink_path(inode, handle, dealloc, right_path,
2361 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2362 struct ocfs2_path *left_path,
2363 struct ocfs2_path *right_path,
2365 struct ocfs2_cached_dealloc_ctxt *dealloc,
2367 struct ocfs2_extent_tree *et)
2369 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2370 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2371 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2372 struct ocfs2_extent_block *eb;
2376 right_leaf_el = path_leaf_el(right_path);
2377 left_leaf_el = path_leaf_el(left_path);
2378 root_bh = left_path->p_node[subtree_index].bh;
2379 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2381 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2384 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2385 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2387 * It's legal for us to proceed if the right leaf is
2388 * the rightmost one and it has an empty extent. There
2389 * are two cases to handle - whether the leaf will be
2390 * empty after removal or not. If the leaf isn't empty
2391 * then just remove the empty extent up front. The
2392 * next block will handle empty leaves by flagging
2395 * Non rightmost leaves will throw -EAGAIN and the
2396 * caller can manually move the subtree and retry.
2399 if (eb->h_next_leaf_blk != 0ULL)
2402 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2403 ret = ocfs2_journal_access(handle, inode,
2404 path_leaf_bh(right_path),
2405 OCFS2_JOURNAL_ACCESS_WRITE);
2411 ocfs2_remove_empty_extent(right_leaf_el);
2413 right_has_empty = 1;
2416 if (eb->h_next_leaf_blk == 0ULL &&
2417 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2419 * We have to update i_last_eb_blk during the meta
2422 ret = ocfs2_journal_access(handle, inode, et_root_bh,
2423 OCFS2_JOURNAL_ACCESS_WRITE);
2429 del_right_subtree = 1;
2433 * Getting here with an empty extent in the right path implies
2434 * that it's the rightmost path and will be deleted.
2436 BUG_ON(right_has_empty && !del_right_subtree);
2438 ret = ocfs2_journal_access(handle, inode, root_bh,
2439 OCFS2_JOURNAL_ACCESS_WRITE);
2445 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2446 ret = ocfs2_journal_access(handle, inode,
2447 right_path->p_node[i].bh,
2448 OCFS2_JOURNAL_ACCESS_WRITE);
2454 ret = ocfs2_journal_access(handle, inode,
2455 left_path->p_node[i].bh,
2456 OCFS2_JOURNAL_ACCESS_WRITE);
2463 if (!right_has_empty) {
2465 * Only do this if we're moving a real
2466 * record. Otherwise, the action is delayed until
2467 * after removal of the right path in which case we
2468 * can do a simple shift to remove the empty extent.
2470 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2471 memset(&right_leaf_el->l_recs[0], 0,
2472 sizeof(struct ocfs2_extent_rec));
2474 if (eb->h_next_leaf_blk == 0ULL) {
2476 * Move recs over to get rid of empty extent, decrease
2477 * next_free. This is allowed to remove the last
2478 * extent in our leaf (setting l_next_free_rec to
2479 * zero) - the delete code below won't care.
2481 ocfs2_remove_empty_extent(right_leaf_el);
2484 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2487 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2491 if (del_right_subtree) {
2492 ocfs2_unlink_subtree(inode, handle, left_path, right_path,
2493 subtree_index, dealloc);
2494 ocfs2_update_edge_lengths(inode, handle, left_path);
2496 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2497 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2500 * Removal of the extent in the left leaf was skipped
2501 * above so we could delete the right path
2504 if (right_has_empty)
2505 ocfs2_remove_empty_extent(left_leaf_el);
2507 ret = ocfs2_journal_dirty(handle, et_root_bh);
2513 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2521 * Given a full path, determine what cpos value would return us a path
2522 * containing the leaf immediately to the right of the current one.
2524 * Will return zero if the path passed in is already the rightmost path.
2526 * This looks similar, but is subtly different to
2527 * ocfs2_find_cpos_for_left_leaf().
2529 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2530 struct ocfs2_path *path, u32 *cpos)
2534 struct ocfs2_extent_list *el;
2538 if (path->p_tree_depth == 0)
2541 blkno = path_leaf_bh(path)->b_blocknr;
2543 /* Start at the tree node just above the leaf and work our way up. */
2544 i = path->p_tree_depth - 1;
2548 el = path->p_node[i].el;
2551 * Find the extent record just after the one in our
2554 next_free = le16_to_cpu(el->l_next_free_rec);
2555 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2556 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2557 if (j == (next_free - 1)) {
2560 * We've determined that the
2561 * path specified is already
2562 * the rightmost one - return a
2568 * The rightmost record points to our
2569 * leaf - we need to travel up the
2575 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2581 * If we got here, we never found a valid node where
2582 * the tree indicated one should be.
2585 "Invalid extent tree at extent block %llu\n",
2586 (unsigned long long)blkno);
2591 blkno = path->p_node[i].bh->b_blocknr;
2599 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2601 struct buffer_head *bh,
2602 struct ocfs2_extent_list *el)
2606 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2609 ret = ocfs2_journal_access(handle, inode, bh,
2610 OCFS2_JOURNAL_ACCESS_WRITE);
2616 ocfs2_remove_empty_extent(el);
2618 ret = ocfs2_journal_dirty(handle, bh);
2626 static int __ocfs2_rotate_tree_left(struct inode *inode,
2627 handle_t *handle, int orig_credits,
2628 struct ocfs2_path *path,
2629 struct ocfs2_cached_dealloc_ctxt *dealloc,
2630 struct ocfs2_path **empty_extent_path,
2631 struct ocfs2_extent_tree *et)
2633 int ret, subtree_root, deleted;
2635 struct ocfs2_path *left_path = NULL;
2636 struct ocfs2_path *right_path = NULL;
2638 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2640 *empty_extent_path = NULL;
2642 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2649 left_path = ocfs2_new_path(path_root_bh(path),
2650 path_root_el(path));
2657 ocfs2_cp_path(left_path, path);
2659 right_path = ocfs2_new_path(path_root_bh(path),
2660 path_root_el(path));
2667 while (right_cpos) {
2668 ret = ocfs2_find_path(inode, right_path, right_cpos);
2674 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2677 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2679 (unsigned long long)
2680 right_path->p_node[subtree_root].bh->b_blocknr,
2681 right_path->p_tree_depth);
2683 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2684 orig_credits, left_path);
2691 * Caller might still want to make changes to the
2692 * tree root, so re-add it to the journal here.
2694 ret = ocfs2_journal_access(handle, inode,
2695 path_root_bh(left_path),
2696 OCFS2_JOURNAL_ACCESS_WRITE);
2702 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2703 right_path, subtree_root,
2704 dealloc, &deleted, et);
2705 if (ret == -EAGAIN) {
2707 * The rotation has to temporarily stop due to
2708 * the right subtree having an empty
2709 * extent. Pass it back to the caller for a
2712 *empty_extent_path = right_path;
2722 * The subtree rotate might have removed records on
2723 * the rightmost edge. If so, then rotation is
2729 ocfs2_mv_path(left_path, right_path);
2731 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2740 ocfs2_free_path(right_path);
2741 ocfs2_free_path(left_path);
2746 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
2747 struct ocfs2_path *path,
2748 struct ocfs2_cached_dealloc_ctxt *dealloc,
2749 struct ocfs2_extent_tree *et)
2751 int ret, subtree_index;
2753 struct ocfs2_path *left_path = NULL;
2754 struct ocfs2_extent_block *eb;
2755 struct ocfs2_extent_list *el;
2758 ret = ocfs2_et_sanity_check(inode, et);
2762 * There's two ways we handle this depending on
2763 * whether path is the only existing one.
2765 ret = ocfs2_extend_rotate_transaction(handle, 0,
2766 handle->h_buffer_credits,
2773 ret = ocfs2_journal_access_path(inode, handle, path);
2779 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
2787 * We have a path to the left of this one - it needs
2790 left_path = ocfs2_new_path(path_root_bh(path),
2791 path_root_el(path));
2798 ret = ocfs2_find_path(inode, left_path, cpos);
2804 ret = ocfs2_journal_access_path(inode, handle, left_path);
2810 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
2812 ocfs2_unlink_subtree(inode, handle, left_path, path,
2813 subtree_index, dealloc);
2814 ocfs2_update_edge_lengths(inode, handle, left_path);
2816 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2817 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2820 * 'path' is also the leftmost path which
2821 * means it must be the only one. This gets
2822 * handled differently because we want to
2823 * revert the inode back to having extents
2826 ocfs2_unlink_path(inode, handle, dealloc, path, 1);
2828 el = et->et_root_el;
2829 el->l_tree_depth = 0;
2830 el->l_next_free_rec = 0;
2831 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2833 ocfs2_et_set_last_eb_blk(et, 0);
2836 ocfs2_journal_dirty(handle, path_root_bh(path));
2839 ocfs2_free_path(left_path);
2844 * Left rotation of btree records.
2846 * In many ways, this is (unsurprisingly) the opposite of right
2847 * rotation. We start at some non-rightmost path containing an empty
2848 * extent in the leaf block. The code works its way to the rightmost
2849 * path by rotating records to the left in every subtree.
2851 * This is used by any code which reduces the number of extent records
2852 * in a leaf. After removal, an empty record should be placed in the
2853 * leftmost list position.
2855 * This won't handle a length update of the rightmost path records if
2856 * the rightmost tree leaf record is removed so the caller is
2857 * responsible for detecting and correcting that.
2859 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
2860 struct ocfs2_path *path,
2861 struct ocfs2_cached_dealloc_ctxt *dealloc,
2862 struct ocfs2_extent_tree *et)
2864 int ret, orig_credits = handle->h_buffer_credits;
2865 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
2866 struct ocfs2_extent_block *eb;
2867 struct ocfs2_extent_list *el;
2869 el = path_leaf_el(path);
2870 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2873 if (path->p_tree_depth == 0) {
2874 rightmost_no_delete:
2876 * Inline extents. This is trivially handled, so do
2879 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
2881 path_leaf_el(path));
2888 * Handle rightmost branch now. There's several cases:
2889 * 1) simple rotation leaving records in there. That's trivial.
2890 * 2) rotation requiring a branch delete - there's no more
2891 * records left. Two cases of this:
2892 * a) There are branches to the left.
2893 * b) This is also the leftmost (the only) branch.
2895 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
2896 * 2a) we need the left branch so that we can update it with the unlink
2897 * 2b) we need to bring the inode back to inline extents.
2900 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2902 if (eb->h_next_leaf_blk == 0) {
2904 * This gets a bit tricky if we're going to delete the
2905 * rightmost path. Get the other cases out of the way
2908 if (le16_to_cpu(el->l_next_free_rec) > 1)
2909 goto rightmost_no_delete;
2911 if (le16_to_cpu(el->l_next_free_rec) == 0) {
2913 ocfs2_error(inode->i_sb,
2914 "Inode %llu has empty extent block at %llu",
2915 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2916 (unsigned long long)le64_to_cpu(eb->h_blkno));
2921 * XXX: The caller can not trust "path" any more after
2922 * this as it will have been deleted. What do we do?
2924 * In theory the rotate-for-merge code will never get
2925 * here because it'll always ask for a rotate in a
2929 ret = ocfs2_remove_rightmost_path(inode, handle, path,
2937 * Now we can loop, remembering the path we get from -EAGAIN
2938 * and restarting from there.
2941 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
2942 dealloc, &restart_path, et);
2943 if (ret && ret != -EAGAIN) {
2948 while (ret == -EAGAIN) {
2949 tmp_path = restart_path;
2950 restart_path = NULL;
2952 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
2955 if (ret && ret != -EAGAIN) {
2960 ocfs2_free_path(tmp_path);
2968 ocfs2_free_path(tmp_path);
2969 ocfs2_free_path(restart_path);
2973 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
2976 struct ocfs2_extent_rec *rec = &el->l_recs[index];
2979 if (rec->e_leaf_clusters == 0) {
2981 * We consumed all of the merged-from record. An empty
2982 * extent cannot exist anywhere but the 1st array
2983 * position, so move things over if the merged-from
2984 * record doesn't occupy that position.
2986 * This creates a new empty extent so the caller
2987 * should be smart enough to have removed any existing
2991 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
2992 size = index * sizeof(struct ocfs2_extent_rec);
2993 memmove(&el->l_recs[1], &el->l_recs[0], size);
2997 * Always memset - the caller doesn't check whether it
2998 * created an empty extent, so there could be junk in
3001 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3005 static int ocfs2_get_right_path(struct inode *inode,
3006 struct ocfs2_path *left_path,
3007 struct ocfs2_path **ret_right_path)
3011 struct ocfs2_path *right_path = NULL;
3012 struct ocfs2_extent_list *left_el;
3014 *ret_right_path = NULL;
3016 /* This function shouldn't be called for non-trees. */
3017 BUG_ON(left_path->p_tree_depth == 0);
3019 left_el = path_leaf_el(left_path);
3020 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3022 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
3029 /* This function shouldn't be called for the rightmost leaf. */
3030 BUG_ON(right_cpos == 0);
3032 right_path = ocfs2_new_path(path_root_bh(left_path),
3033 path_root_el(left_path));
3040 ret = ocfs2_find_path(inode, right_path, right_cpos);
3046 *ret_right_path = right_path;
3049 ocfs2_free_path(right_path);
3054 * Remove split_rec clusters from the record at index and merge them
3055 * onto the beginning of the record "next" to it.
3056 * For index < l_count - 1, the next means the extent rec at index + 1.
3057 * For index == l_count - 1, the "next" means the 1st extent rec of the
3058 * next extent block.
3060 static int ocfs2_merge_rec_right(struct inode *inode,
3061 struct ocfs2_path *left_path,
3063 struct ocfs2_extent_rec *split_rec,
3066 int ret, next_free, i;
3067 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3068 struct ocfs2_extent_rec *left_rec;
3069 struct ocfs2_extent_rec *right_rec;
3070 struct ocfs2_extent_list *right_el;
3071 struct ocfs2_path *right_path = NULL;
3072 int subtree_index = 0;
3073 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3074 struct buffer_head *bh = path_leaf_bh(left_path);
3075 struct buffer_head *root_bh = NULL;
3077 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3078 left_rec = &el->l_recs[index];
3080 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3081 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3082 /* we meet with a cross extent block merge. */
3083 ret = ocfs2_get_right_path(inode, left_path, &right_path);
3089 right_el = path_leaf_el(right_path);
3090 next_free = le16_to_cpu(right_el->l_next_free_rec);
3091 BUG_ON(next_free <= 0);
3092 right_rec = &right_el->l_recs[0];
3093 if (ocfs2_is_empty_extent(right_rec)) {
3094 BUG_ON(next_free <= 1);
3095 right_rec = &right_el->l_recs[1];
3098 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3099 le16_to_cpu(left_rec->e_leaf_clusters) !=
3100 le32_to_cpu(right_rec->e_cpos));
3102 subtree_index = ocfs2_find_subtree_root(inode,
3103 left_path, right_path);
3105 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3106 handle->h_buffer_credits,
3113 root_bh = left_path->p_node[subtree_index].bh;
3114 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3116 ret = ocfs2_journal_access(handle, inode, root_bh,
3117 OCFS2_JOURNAL_ACCESS_WRITE);
3123 for (i = subtree_index + 1;
3124 i < path_num_items(right_path); i++) {
3125 ret = ocfs2_journal_access(handle, inode,
3126 right_path->p_node[i].bh,
3127 OCFS2_JOURNAL_ACCESS_WRITE);
3133 ret = ocfs2_journal_access(handle, inode,
3134 left_path->p_node[i].bh,
3135 OCFS2_JOURNAL_ACCESS_WRITE);
3143 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3144 right_rec = &el->l_recs[index + 1];
3147 ret = ocfs2_journal_access(handle, inode, bh,
3148 OCFS2_JOURNAL_ACCESS_WRITE);
3154 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3156 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3157 le64_add_cpu(&right_rec->e_blkno,
3158 -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3159 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3161 ocfs2_cleanup_merge(el, index);
3163 ret = ocfs2_journal_dirty(handle, bh);
3168 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3172 ocfs2_complete_edge_insert(inode, handle, left_path,
3173 right_path, subtree_index);
3177 ocfs2_free_path(right_path);
3181 static int ocfs2_get_left_path(struct inode *inode,
3182 struct ocfs2_path *right_path,
3183 struct ocfs2_path **ret_left_path)
3187 struct ocfs2_path *left_path = NULL;
3189 *ret_left_path = NULL;
3191 /* This function shouldn't be called for non-trees. */
3192 BUG_ON(right_path->p_tree_depth == 0);
3194 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
3195 right_path, &left_cpos);
3201 /* This function shouldn't be called for the leftmost leaf. */
3202 BUG_ON(left_cpos == 0);
3204 left_path = ocfs2_new_path(path_root_bh(right_path),
3205 path_root_el(right_path));
3212 ret = ocfs2_find_path(inode, left_path, left_cpos);
3218 *ret_left_path = left_path;
3221 ocfs2_free_path(left_path);
3226 * Remove split_rec clusters from the record at index and merge them
3227 * onto the tail of the record "before" it.
3228 * For index > 0, the "before" means the extent rec at index - 1.
3230 * For index == 0, the "before" means the last record of the previous
3231 * extent block. And there is also a situation that we may need to
3232 * remove the rightmost leaf extent block in the right_path and change
3233 * the right path to indicate the new rightmost path.
3235 static int ocfs2_merge_rec_left(struct inode *inode,
3236 struct ocfs2_path *right_path,
3238 struct ocfs2_extent_rec *split_rec,
3239 struct ocfs2_cached_dealloc_ctxt *dealloc,
3240 struct ocfs2_extent_tree *et,
3243 int ret, i, subtree_index = 0, has_empty_extent = 0;
3244 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3245 struct ocfs2_extent_rec *left_rec;
3246 struct ocfs2_extent_rec *right_rec;
3247 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3248 struct buffer_head *bh = path_leaf_bh(right_path);
3249 struct buffer_head *root_bh = NULL;
3250 struct ocfs2_path *left_path = NULL;
3251 struct ocfs2_extent_list *left_el;
3255 right_rec = &el->l_recs[index];
3257 /* we meet with a cross extent block merge. */
3258 ret = ocfs2_get_left_path(inode, right_path, &left_path);
3264 left_el = path_leaf_el(left_path);
3265 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3266 le16_to_cpu(left_el->l_count));
3268 left_rec = &left_el->l_recs[
3269 le16_to_cpu(left_el->l_next_free_rec) - 1];
3270 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3271 le16_to_cpu(left_rec->e_leaf_clusters) !=
3272 le32_to_cpu(split_rec->e_cpos));
3274 subtree_index = ocfs2_find_subtree_root(inode,
3275 left_path, right_path);
3277 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3278 handle->h_buffer_credits,
3285 root_bh = left_path->p_node[subtree_index].bh;
3286 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3288 ret = ocfs2_journal_access(handle, inode, root_bh,
3289 OCFS2_JOURNAL_ACCESS_WRITE);
3295 for (i = subtree_index + 1;
3296 i < path_num_items(right_path); i++) {
3297 ret = ocfs2_journal_access(handle, inode,
3298 right_path->p_node[i].bh,
3299 OCFS2_JOURNAL_ACCESS_WRITE);
3305 ret = ocfs2_journal_access(handle, inode,
3306 left_path->p_node[i].bh,
3307 OCFS2_JOURNAL_ACCESS_WRITE);
3314 left_rec = &el->l_recs[index - 1];
3315 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3316 has_empty_extent = 1;
3319 ret = ocfs2_journal_access(handle, inode, bh,
3320 OCFS2_JOURNAL_ACCESS_WRITE);
3326 if (has_empty_extent && index == 1) {
3328 * The easy case - we can just plop the record right in.
3330 *left_rec = *split_rec;
3332 has_empty_extent = 0;
3334 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3336 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3337 le64_add_cpu(&right_rec->e_blkno,
3338 ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3339 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3341 ocfs2_cleanup_merge(el, index);
3343 ret = ocfs2_journal_dirty(handle, bh);
3348 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3353 * In the situation that the right_rec is empty and the extent
3354 * block is empty also, ocfs2_complete_edge_insert can't handle
3355 * it and we need to delete the right extent block.
3357 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3358 le16_to_cpu(el->l_next_free_rec) == 1) {
3360 ret = ocfs2_remove_rightmost_path(inode, handle,
3368 /* Now the rightmost extent block has been deleted.
3369 * So we use the new rightmost path.
3371 ocfs2_mv_path(right_path, left_path);
3374 ocfs2_complete_edge_insert(inode, handle, left_path,
3375 right_path, subtree_index);
3379 ocfs2_free_path(left_path);
3383 static int ocfs2_try_to_merge_extent(struct inode *inode,
3385 struct ocfs2_path *path,
3387 struct ocfs2_extent_rec *split_rec,
3388 struct ocfs2_cached_dealloc_ctxt *dealloc,
3389 struct ocfs2_merge_ctxt *ctxt,
3390 struct ocfs2_extent_tree *et)
3394 struct ocfs2_extent_list *el = path_leaf_el(path);
3395 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3397 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3399 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3401 * The merge code will need to create an empty
3402 * extent to take the place of the newly
3403 * emptied slot. Remove any pre-existing empty
3404 * extents - having more than one in a leaf is
3407 ret = ocfs2_rotate_tree_left(inode, handle, path,
3414 rec = &el->l_recs[split_index];
3417 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3419 * Left-right contig implies this.
3421 BUG_ON(!ctxt->c_split_covers_rec);
3424 * Since the leftright insert always covers the entire
3425 * extent, this call will delete the insert record
3426 * entirely, resulting in an empty extent record added to
3429 * Since the adding of an empty extent shifts
3430 * everything back to the right, there's no need to
3431 * update split_index here.
3433 * When the split_index is zero, we need to merge it to the
3434 * prevoius extent block. It is more efficient and easier
3435 * if we do merge_right first and merge_left later.
3437 ret = ocfs2_merge_rec_right(inode, path,
3446 * We can only get this from logic error above.
3448 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3450 /* The merge left us with an empty extent, remove it. */
3451 ret = ocfs2_rotate_tree_left(inode, handle, path,
3458 rec = &el->l_recs[split_index];
3461 * Note that we don't pass split_rec here on purpose -
3462 * we've merged it into the rec already.
3464 ret = ocfs2_merge_rec_left(inode, path,
3474 ret = ocfs2_rotate_tree_left(inode, handle, path,
3477 * Error from this last rotate is not critical, so
3478 * print but don't bubble it up.
3485 * Merge a record to the left or right.
3487 * 'contig_type' is relative to the existing record,
3488 * so for example, if we're "right contig", it's to
3489 * the record on the left (hence the left merge).
3491 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3492 ret = ocfs2_merge_rec_left(inode,
3502 ret = ocfs2_merge_rec_right(inode,
3512 if (ctxt->c_split_covers_rec) {
3514 * The merge may have left an empty extent in
3515 * our leaf. Try to rotate it away.
3517 ret = ocfs2_rotate_tree_left(inode, handle, path,
3529 static void ocfs2_subtract_from_rec(struct super_block *sb,
3530 enum ocfs2_split_type split,
3531 struct ocfs2_extent_rec *rec,
3532 struct ocfs2_extent_rec *split_rec)
3536 len_blocks = ocfs2_clusters_to_blocks(sb,
3537 le16_to_cpu(split_rec->e_leaf_clusters));
3539 if (split == SPLIT_LEFT) {
3541 * Region is on the left edge of the existing
3544 le32_add_cpu(&rec->e_cpos,
3545 le16_to_cpu(split_rec->e_leaf_clusters));
3546 le64_add_cpu(&rec->e_blkno, len_blocks);
3547 le16_add_cpu(&rec->e_leaf_clusters,
3548 -le16_to_cpu(split_rec->e_leaf_clusters));
3551 * Region is on the right edge of the existing
3554 le16_add_cpu(&rec->e_leaf_clusters,
3555 -le16_to_cpu(split_rec->e_leaf_clusters));
3560 * Do the final bits of extent record insertion at the target leaf
3561 * list. If this leaf is part of an allocation tree, it is assumed
3562 * that the tree above has been prepared.
3564 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
3565 struct ocfs2_extent_list *el,
3566 struct ocfs2_insert_type *insert,
3567 struct inode *inode)
3569 int i = insert->ins_contig_index;
3571 struct ocfs2_extent_rec *rec;
3573 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3575 if (insert->ins_split != SPLIT_NONE) {
3576 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3578 rec = &el->l_recs[i];
3579 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
3585 * Contiguous insert - either left or right.
3587 if (insert->ins_contig != CONTIG_NONE) {
3588 rec = &el->l_recs[i];
3589 if (insert->ins_contig == CONTIG_LEFT) {
3590 rec->e_blkno = insert_rec->e_blkno;
3591 rec->e_cpos = insert_rec->e_cpos;
3593 le16_add_cpu(&rec->e_leaf_clusters,
3594 le16_to_cpu(insert_rec->e_leaf_clusters));
3599 * Handle insert into an empty leaf.
3601 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3602 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3603 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3604 el->l_recs[0] = *insert_rec;
3605 el->l_next_free_rec = cpu_to_le16(1);
3612 if (insert->ins_appending == APPEND_TAIL) {
3613 i = le16_to_cpu(el->l_next_free_rec) - 1;
3614 rec = &el->l_recs[i];
3615 range = le32_to_cpu(rec->e_cpos)
3616 + le16_to_cpu(rec->e_leaf_clusters);
3617 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3619 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3620 le16_to_cpu(el->l_count),
3621 "inode %lu, depth %u, count %u, next free %u, "
3622 "rec.cpos %u, rec.clusters %u, "
3623 "insert.cpos %u, insert.clusters %u\n",
3625 le16_to_cpu(el->l_tree_depth),
3626 le16_to_cpu(el->l_count),
3627 le16_to_cpu(el->l_next_free_rec),
3628 le32_to_cpu(el->l_recs[i].e_cpos),
3629 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3630 le32_to_cpu(insert_rec->e_cpos),
3631 le16_to_cpu(insert_rec->e_leaf_clusters));
3633 el->l_recs[i] = *insert_rec;
3634 le16_add_cpu(&el->l_next_free_rec, 1);
3640 * Ok, we have to rotate.
3642 * At this point, it is safe to assume that inserting into an
3643 * empty leaf and appending to a leaf have both been handled
3646 * This leaf needs to have space, either by the empty 1st
3647 * extent record, or by virtue of an l_next_rec < l_count.
3649 ocfs2_rotate_leaf(el, insert_rec);
3652 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3654 struct ocfs2_path *path,
3655 struct ocfs2_extent_rec *insert_rec)
3657 int ret, i, next_free;
3658 struct buffer_head *bh;
3659 struct ocfs2_extent_list *el;
3660 struct ocfs2_extent_rec *rec;
3663 * Update everything except the leaf block.
3665 for (i = 0; i < path->p_tree_depth; i++) {
3666 bh = path->p_node[i].bh;
3667 el = path->p_node[i].el;
3669 next_free = le16_to_cpu(el->l_next_free_rec);
3670 if (next_free == 0) {
3671 ocfs2_error(inode->i_sb,
3672 "Dinode %llu has a bad extent list",
3673 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3678 rec = &el->l_recs[next_free - 1];
3680 rec->e_int_clusters = insert_rec->e_cpos;
3681 le32_add_cpu(&rec->e_int_clusters,
3682 le16_to_cpu(insert_rec->e_leaf_clusters));
3683 le32_add_cpu(&rec->e_int_clusters,
3684 -le32_to_cpu(rec->e_cpos));
3686 ret = ocfs2_journal_dirty(handle, bh);
3693 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3694 struct ocfs2_extent_rec *insert_rec,
3695 struct ocfs2_path *right_path,
3696 struct ocfs2_path **ret_left_path)
3699 struct ocfs2_extent_list *el;
3700 struct ocfs2_path *left_path = NULL;
3702 *ret_left_path = NULL;
3705 * This shouldn't happen for non-trees. The extent rec cluster
3706 * count manipulation below only works for interior nodes.
3708 BUG_ON(right_path->p_tree_depth == 0);
3711 * If our appending insert is at the leftmost edge of a leaf,
3712 * then we might need to update the rightmost records of the
3715 el = path_leaf_el(right_path);
3716 next_free = le16_to_cpu(el->l_next_free_rec);
3717 if (next_free == 0 ||
3718 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3721 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3728 mlog(0, "Append may need a left path update. cpos: %u, "
3729 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3733 * No need to worry if the append is already in the
3737 left_path = ocfs2_new_path(path_root_bh(right_path),
3738 path_root_el(right_path));
3745 ret = ocfs2_find_path(inode, left_path, left_cpos);
3752 * ocfs2_insert_path() will pass the left_path to the
3758 ret = ocfs2_journal_access_path(inode, handle, right_path);
3764 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
3766 *ret_left_path = left_path;
3770 ocfs2_free_path(left_path);
3775 static void ocfs2_split_record(struct inode *inode,
3776 struct ocfs2_path *left_path,
3777 struct ocfs2_path *right_path,
3778 struct ocfs2_extent_rec *split_rec,
3779 enum ocfs2_split_type split)
3782 u32 cpos = le32_to_cpu(split_rec->e_cpos);
3783 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
3784 struct ocfs2_extent_rec *rec, *tmprec;
3786 right_el = path_leaf_el(right_path);;
3788 left_el = path_leaf_el(left_path);
3791 insert_el = right_el;
3792 index = ocfs2_search_extent_list(el, cpos);
3794 if (index == 0 && left_path) {
3795 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3798 * This typically means that the record
3799 * started in the left path but moved to the
3800 * right as a result of rotation. We either
3801 * move the existing record to the left, or we
3802 * do the later insert there.
3804 * In this case, the left path should always
3805 * exist as the rotate code will have passed
3806 * it back for a post-insert update.
3809 if (split == SPLIT_LEFT) {
3811 * It's a left split. Since we know
3812 * that the rotate code gave us an
3813 * empty extent in the left path, we
3814 * can just do the insert there.
3816 insert_el = left_el;
3819 * Right split - we have to move the
3820 * existing record over to the left
3821 * leaf. The insert will be into the
3822 * newly created empty extent in the
3825 tmprec = &right_el->l_recs[index];
3826 ocfs2_rotate_leaf(left_el, tmprec);
3829 memset(tmprec, 0, sizeof(*tmprec));
3830 index = ocfs2_search_extent_list(left_el, cpos);
3831 BUG_ON(index == -1);
3836 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
3838 * Left path is easy - we can just allow the insert to
3842 insert_el = left_el;
3843 index = ocfs2_search_extent_list(el, cpos);
3844 BUG_ON(index == -1);
3847 rec = &el->l_recs[index];
3848 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
3849 ocfs2_rotate_leaf(insert_el, split_rec);
3853 * This function only does inserts on an allocation b-tree. For tree
3854 * depth = 0, ocfs2_insert_at_leaf() is called directly.
3856 * right_path is the path we want to do the actual insert
3857 * in. left_path should only be passed in if we need to update that
3858 * portion of the tree after an edge insert.
3860 static int ocfs2_insert_path(struct inode *inode,
3862 struct ocfs2_path *left_path,
3863 struct ocfs2_path *right_path,
3864 struct ocfs2_extent_rec *insert_rec,
3865 struct ocfs2_insert_type *insert)
3867 int ret, subtree_index;
3868 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
3871 int credits = handle->h_buffer_credits;
3874 * There's a chance that left_path got passed back to
3875 * us without being accounted for in the
3876 * journal. Extend our transaction here to be sure we
3877 * can change those blocks.
3879 credits += left_path->p_tree_depth;
3881 ret = ocfs2_extend_trans(handle, credits);
3887 ret = ocfs2_journal_access_path(inode, handle, left_path);
3895 * Pass both paths to the journal. The majority of inserts
3896 * will be touching all components anyway.
3898 ret = ocfs2_journal_access_path(inode, handle, right_path);
3904 if (insert->ins_split != SPLIT_NONE) {
3906 * We could call ocfs2_insert_at_leaf() for some types
3907 * of splits, but it's easier to just let one separate
3908 * function sort it all out.
3910 ocfs2_split_record(inode, left_path, right_path,
3911 insert_rec, insert->ins_split);
3914 * Split might have modified either leaf and we don't
3915 * have a guarantee that the later edge insert will
3916 * dirty this for us.
3919 ret = ocfs2_journal_dirty(handle,
3920 path_leaf_bh(left_path));
3924 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
3927 ret = ocfs2_journal_dirty(handle, leaf_bh);
3933 * The rotate code has indicated that we need to fix
3934 * up portions of the tree after the insert.
3936 * XXX: Should we extend the transaction here?
3938 subtree_index = ocfs2_find_subtree_root(inode, left_path,
3940 ocfs2_complete_edge_insert(inode, handle, left_path,
3941 right_path, subtree_index);
3949 static int ocfs2_do_insert_extent(struct inode *inode,
3951 struct ocfs2_extent_tree *et,
3952 struct ocfs2_extent_rec *insert_rec,
3953 struct ocfs2_insert_type *type)
3955 int ret, rotate = 0;
3957 struct ocfs2_path *right_path = NULL;
3958 struct ocfs2_path *left_path = NULL;
3959 struct ocfs2_extent_list *el;
3961 el = et->et_root_el;
3963 ret = ocfs2_journal_access(handle, inode, et->et_root_bh,
3964 OCFS2_JOURNAL_ACCESS_WRITE);
3970 if (le16_to_cpu(el->l_tree_depth) == 0) {
3971 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
3972 goto out_update_clusters;
3975 right_path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
3983 * Determine the path to start with. Rotations need the
3984 * rightmost path, everything else can go directly to the
3987 cpos = le32_to_cpu(insert_rec->e_cpos);
3988 if (type->ins_appending == APPEND_NONE &&
3989 type->ins_contig == CONTIG_NONE) {
3994 ret = ocfs2_find_path(inode, right_path, cpos);
4001 * Rotations and appends need special treatment - they modify
4002 * parts of the tree's above them.
4004 * Both might pass back a path immediate to the left of the
4005 * one being inserted to. This will be cause
4006 * ocfs2_insert_path() to modify the rightmost records of
4007 * left_path to account for an edge insert.
4009 * XXX: When modifying this code, keep in mind that an insert
4010 * can wind up skipping both of these two special cases...
4013 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
4014 le32_to_cpu(insert_rec->e_cpos),
4015 right_path, &left_path);
4022 * ocfs2_rotate_tree_right() might have extended the
4023 * transaction without re-journaling our tree root.
4025 ret = ocfs2_journal_access(handle, inode, et->et_root_bh,
4026 OCFS2_JOURNAL_ACCESS_WRITE);
4031 } else if (type->ins_appending == APPEND_TAIL
4032 && type->ins_contig != CONTIG_LEFT) {
4033 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
4034 right_path, &left_path);
4041 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
4048 out_update_clusters:
4049 if (type->ins_split == SPLIT_NONE)
4050 ocfs2_et_update_clusters(inode, et,
4051 le16_to_cpu(insert_rec->e_leaf_clusters));
4053 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
4058 ocfs2_free_path(left_path);
4059 ocfs2_free_path(right_path);
4064 static enum ocfs2_contig_type
4065 ocfs2_figure_merge_contig_type(struct inode *inode, struct ocfs2_path *path,
4066 struct ocfs2_extent_list *el, int index,
4067 struct ocfs2_extent_rec *split_rec)
4070 enum ocfs2_contig_type ret = CONTIG_NONE;
4071 u32 left_cpos, right_cpos;
4072 struct ocfs2_extent_rec *rec = NULL;
4073 struct ocfs2_extent_list *new_el;
4074 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4075 struct buffer_head *bh;
4076 struct ocfs2_extent_block *eb;
4079 rec = &el->l_recs[index - 1];
4080 } else if (path->p_tree_depth > 0) {
4081 status = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
4086 if (left_cpos != 0) {
4087 left_path = ocfs2_new_path(path_root_bh(path),
4088 path_root_el(path));
4092 status = ocfs2_find_path(inode, left_path, left_cpos);
4096 new_el = path_leaf_el(left_path);
4098 if (le16_to_cpu(new_el->l_next_free_rec) !=
4099 le16_to_cpu(new_el->l_count)) {
4100 bh = path_leaf_bh(left_path);
4101 eb = (struct ocfs2_extent_block *)bh->b_data;
4102 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
4106 rec = &new_el->l_recs[
4107 le16_to_cpu(new_el->l_next_free_rec) - 1];
4112 * We're careful to check for an empty extent record here -
4113 * the merge code will know what to do if it sees one.
4116 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4117 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4120 ret = ocfs2_extent_contig(inode, rec, split_rec);
4125 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4126 rec = &el->l_recs[index + 1];
4127 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4128 path->p_tree_depth > 0) {
4129 status = ocfs2_find_cpos_for_right_leaf(inode->i_sb,
4134 if (right_cpos == 0)
4137 right_path = ocfs2_new_path(path_root_bh(path),
4138 path_root_el(path));
4142 status = ocfs2_find_path(inode, right_path, right_cpos);
4146 new_el = path_leaf_el(right_path);
4147 rec = &new_el->l_recs[0];
4148 if (ocfs2_is_empty_extent(rec)) {
4149 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4150 bh = path_leaf_bh(right_path);
4151 eb = (struct ocfs2_extent_block *)bh->b_data;
4152 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
4156 rec = &new_el->l_recs[1];
4161 enum ocfs2_contig_type contig_type;
4163 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
4165 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4166 ret = CONTIG_LEFTRIGHT;
4167 else if (ret == CONTIG_NONE)
4173 ocfs2_free_path(left_path);
4175 ocfs2_free_path(right_path);
4180 static void ocfs2_figure_contig_type(struct inode *inode,
4181 struct ocfs2_insert_type *insert,
4182 struct ocfs2_extent_list *el,
4183 struct ocfs2_extent_rec *insert_rec,
4184 struct ocfs2_extent_tree *et)
4187 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4189 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4191 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4192 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
4194 if (contig_type != CONTIG_NONE) {
4195 insert->ins_contig_index = i;
4199 insert->ins_contig = contig_type;
4201 if (insert->ins_contig != CONTIG_NONE) {
4202 struct ocfs2_extent_rec *rec =
4203 &el->l_recs[insert->ins_contig_index];
4204 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4205 le16_to_cpu(insert_rec->e_leaf_clusters);
4208 * Caller might want us to limit the size of extents, don't
4209 * calculate contiguousness if we might exceed that limit.
4211 if (et->et_max_leaf_clusters &&
4212 (len > et->et_max_leaf_clusters))
4213 insert->ins_contig = CONTIG_NONE;
4218 * This should only be called against the righmost leaf extent list.
4220 * ocfs2_figure_appending_type() will figure out whether we'll have to
4221 * insert at the tail of the rightmost leaf.
4223 * This should also work against the root extent list for tree's with 0
4224 * depth. If we consider the root extent list to be the rightmost leaf node
4225 * then the logic here makes sense.
4227 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4228 struct ocfs2_extent_list *el,
4229 struct ocfs2_extent_rec *insert_rec)
4232 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4233 struct ocfs2_extent_rec *rec;
4235 insert->ins_appending = APPEND_NONE;
4237 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4239 if (!el->l_next_free_rec)
4240 goto set_tail_append;
4242 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4243 /* Were all records empty? */
4244 if (le16_to_cpu(el->l_next_free_rec) == 1)
4245 goto set_tail_append;
4248 i = le16_to_cpu(el->l_next_free_rec) - 1;
4249 rec = &el->l_recs[i];
4252 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4253 goto set_tail_append;
4258 insert->ins_appending = APPEND_TAIL;
4262 * Helper function called at the begining of an insert.
4264 * This computes a few things that are commonly used in the process of
4265 * inserting into the btree:
4266 * - Whether the new extent is contiguous with an existing one.
4267 * - The current tree depth.
4268 * - Whether the insert is an appending one.
4269 * - The total # of free records in the tree.
4271 * All of the information is stored on the ocfs2_insert_type
4274 static int ocfs2_figure_insert_type(struct inode *inode,
4275 struct ocfs2_extent_tree *et,
4276 struct buffer_head **last_eb_bh,
4277 struct ocfs2_extent_rec *insert_rec,
4279 struct ocfs2_insert_type *insert)
4282 struct ocfs2_extent_block *eb;
4283 struct ocfs2_extent_list *el;
4284 struct ocfs2_path *path = NULL;
4285 struct buffer_head *bh = NULL;
4287 insert->ins_split = SPLIT_NONE;
4289 el = et->et_root_el;
4290 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4292 if (el->l_tree_depth) {
4294 * If we have tree depth, we read in the
4295 * rightmost extent block ahead of time as
4296 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4297 * may want it later.
4299 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4300 ocfs2_et_get_last_eb_blk(et), &bh,
4301 OCFS2_BH_CACHED, inode);
4306 eb = (struct ocfs2_extent_block *) bh->b_data;
4311 * Unless we have a contiguous insert, we'll need to know if
4312 * there is room left in our allocation tree for another
4315 * XXX: This test is simplistic, we can search for empty
4316 * extent records too.
4318 *free_records = le16_to_cpu(el->l_count) -
4319 le16_to_cpu(el->l_next_free_rec);
4321 if (!insert->ins_tree_depth) {
4322 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4323 ocfs2_figure_appending_type(insert, el, insert_rec);
4327 path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
4335 * In the case that we're inserting past what the tree
4336 * currently accounts for, ocfs2_find_path() will return for
4337 * us the rightmost tree path. This is accounted for below in
4338 * the appending code.
4340 ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
4346 el = path_leaf_el(path);
4349 * Now that we have the path, there's two things we want to determine:
4350 * 1) Contiguousness (also set contig_index if this is so)
4352 * 2) Are we doing an append? We can trivially break this up
4353 * into two types of appends: simple record append, or a
4354 * rotate inside the tail leaf.
4356 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4359 * The insert code isn't quite ready to deal with all cases of
4360 * left contiguousness. Specifically, if it's an insert into
4361 * the 1st record in a leaf, it will require the adjustment of
4362 * cluster count on the last record of the path directly to it's
4363 * left. For now, just catch that case and fool the layers
4364 * above us. This works just fine for tree_depth == 0, which
4365 * is why we allow that above.
4367 if (insert->ins_contig == CONTIG_LEFT &&
4368 insert->ins_contig_index == 0)
4369 insert->ins_contig = CONTIG_NONE;
4372 * Ok, so we can simply compare against last_eb to figure out
4373 * whether the path doesn't exist. This will only happen in
4374 * the case that we're doing a tail append, so maybe we can
4375 * take advantage of that information somehow.
4377 if (ocfs2_et_get_last_eb_blk(et) ==
4378 path_leaf_bh(path)->b_blocknr) {
4380 * Ok, ocfs2_find_path() returned us the rightmost
4381 * tree path. This might be an appending insert. There are
4383 * 1) We're doing a true append at the tail:
4384 * -This might even be off the end of the leaf
4385 * 2) We're "appending" by rotating in the tail
4387 ocfs2_figure_appending_type(insert, el, insert_rec);
4391 ocfs2_free_path(path);
4401 * Insert an extent into an inode btree.
4403 * The caller needs to update fe->i_clusters
4405 int ocfs2_insert_extent(struct ocfs2_super *osb,
4407 struct inode *inode,
4408 struct ocfs2_extent_tree *et,
4413 struct ocfs2_alloc_context *meta_ac)
4416 int uninitialized_var(free_records);
4417 struct buffer_head *last_eb_bh = NULL;
4418 struct ocfs2_insert_type insert = {0, };
4419 struct ocfs2_extent_rec rec;
4421 mlog(0, "add %u clusters at position %u to inode %llu\n",
4422 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4424 memset(&rec, 0, sizeof(rec));
4425 rec.e_cpos = cpu_to_le32(cpos);
4426 rec.e_blkno = cpu_to_le64(start_blk);
4427 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4428 rec.e_flags = flags;
4429 status = ocfs2_et_insert_check(inode, et, &rec);
4435 status = ocfs2_figure_insert_type(inode, et, &last_eb_bh, &rec,
4436 &free_records, &insert);
4442 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4443 "Insert.contig_index: %d, Insert.free_records: %d, "
4444 "Insert.tree_depth: %d\n",
4445 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
4446 free_records, insert.ins_tree_depth);
4448 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4449 status = ocfs2_grow_tree(inode, handle, et,
4450 &insert.ins_tree_depth, &last_eb_bh,
4458 /* Finally, we can add clusters. This might rotate the tree for us. */
4459 status = ocfs2_do_insert_extent(inode, handle, et, &rec, &insert);
4462 else if (et->et_ops == &ocfs2_dinode_et_ops)
4463 ocfs2_extent_map_insert_rec(inode, &rec);
4473 * Allcate and add clusters into the extent b-tree.
4474 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4475 * The extent b-tree's root is specified by et, and
4476 * it is not limited to the file storage. Any extent tree can use this
4477 * function if it implements the proper ocfs2_extent_tree.
4479 int ocfs2_add_clusters_in_btree(struct ocfs2_super *osb,
4480 struct inode *inode,
4481 u32 *logical_offset,
4482 u32 clusters_to_add,
4484 struct ocfs2_extent_tree *et,
4486 struct ocfs2_alloc_context *data_ac,
4487 struct ocfs2_alloc_context *meta_ac,
4488 enum ocfs2_alloc_restarted *reason_ret)
4492 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4493 u32 bit_off, num_bits;
4497 BUG_ON(!clusters_to_add);
4500 flags = OCFS2_EXT_UNWRITTEN;
4502 free_extents = ocfs2_num_free_extents(osb, inode, et);
4503 if (free_extents < 0) {
4504 status = free_extents;
4509 /* there are two cases which could cause us to EAGAIN in the
4510 * we-need-more-metadata case:
4511 * 1) we haven't reserved *any*
4512 * 2) we are so fragmented, we've needed to add metadata too
4514 if (!free_extents && !meta_ac) {
4515 mlog(0, "we haven't reserved any metadata!\n");
4517 reason = RESTART_META;
4519 } else if ((!free_extents)
4520 && (ocfs2_alloc_context_bits_left(meta_ac)
4521 < ocfs2_extend_meta_needed(et->et_root_el))) {
4522 mlog(0, "filesystem is really fragmented...\n");
4524 reason = RESTART_META;
4528 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
4529 clusters_to_add, &bit_off, &num_bits);
4531 if (status != -ENOSPC)
4536 BUG_ON(num_bits > clusters_to_add);
4538 /* reserve our write early -- insert_extent may update the inode */
4539 status = ocfs2_journal_access(handle, inode, et->et_root_bh,
4540 OCFS2_JOURNAL_ACCESS_WRITE);
4546 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4547 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
4548 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4549 status = ocfs2_insert_extent(osb, handle, inode, et,
4550 *logical_offset, block,
4551 num_bits, flags, meta_ac);
4557 status = ocfs2_journal_dirty(handle, et->et_root_bh);
4563 clusters_to_add -= num_bits;
4564 *logical_offset += num_bits;
4566 if (clusters_to_add) {
4567 mlog(0, "need to alloc once more, wanted = %u\n",
4570 reason = RESTART_TRANS;
4576 *reason_ret = reason;
4580 static void ocfs2_make_right_split_rec(struct super_block *sb,
4581 struct ocfs2_extent_rec *split_rec,
4583 struct ocfs2_extent_rec *rec)
4585 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4586 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4588 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4590 split_rec->e_cpos = cpu_to_le32(cpos);
4591 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4593 split_rec->e_blkno = rec->e_blkno;
4594 le64_add_cpu(&split_rec->e_blkno,
4595 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4597 split_rec->e_flags = rec->e_flags;
4600 static int ocfs2_split_and_insert(struct inode *inode,
4602 struct ocfs2_path *path,
4603 struct ocfs2_extent_tree *et,
4604 struct buffer_head **last_eb_bh,
4606 struct ocfs2_extent_rec *orig_split_rec,
4607 struct ocfs2_alloc_context *meta_ac)
4610 unsigned int insert_range, rec_range, do_leftright = 0;
4611 struct ocfs2_extent_rec tmprec;
4612 struct ocfs2_extent_list *rightmost_el;
4613 struct ocfs2_extent_rec rec;
4614 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4615 struct ocfs2_insert_type insert;
4616 struct ocfs2_extent_block *eb;
4620 * Store a copy of the record on the stack - it might move
4621 * around as the tree is manipulated below.
4623 rec = path_leaf_el(path)->l_recs[split_index];
4625 rightmost_el = et->et_root_el;
4627 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4629 BUG_ON(!(*last_eb_bh));
4630 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4631 rightmost_el = &eb->h_list;
4634 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4635 le16_to_cpu(rightmost_el->l_count)) {
4636 ret = ocfs2_grow_tree(inode, handle, et,
4637 &depth, last_eb_bh, meta_ac);
4644 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4645 insert.ins_appending = APPEND_NONE;
4646 insert.ins_contig = CONTIG_NONE;
4647 insert.ins_tree_depth = depth;
4649 insert_range = le32_to_cpu(split_rec.e_cpos) +
4650 le16_to_cpu(split_rec.e_leaf_clusters);
4651 rec_range = le32_to_cpu(rec.e_cpos) +
4652 le16_to_cpu(rec.e_leaf_clusters);
4654 if (split_rec.e_cpos == rec.e_cpos) {
4655 insert.ins_split = SPLIT_LEFT;
4656 } else if (insert_range == rec_range) {
4657 insert.ins_split = SPLIT_RIGHT;
4660 * Left/right split. We fake this as a right split
4661 * first and then make a second pass as a left split.
4663 insert.ins_split = SPLIT_RIGHT;
4665 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
4670 BUG_ON(do_leftright);
4674 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4680 if (do_leftright == 1) {
4682 struct ocfs2_extent_list *el;
4685 split_rec = *orig_split_rec;
4687 ocfs2_reinit_path(path, 1);
4689 cpos = le32_to_cpu(split_rec.e_cpos);
4690 ret = ocfs2_find_path(inode, path, cpos);
4696 el = path_leaf_el(path);
4697 split_index = ocfs2_search_extent_list(el, cpos);
4706 * Mark part or all of the extent record at split_index in the leaf
4707 * pointed to by path as written. This removes the unwritten
4710 * Care is taken to handle contiguousness so as to not grow the tree.
4712 * meta_ac is not strictly necessary - we only truly need it if growth
4713 * of the tree is required. All other cases will degrade into a less
4714 * optimal tree layout.
4716 * last_eb_bh should be the rightmost leaf block for any extent
4717 * btree. Since a split may grow the tree or a merge might shrink it,
4718 * the caller cannot trust the contents of that buffer after this call.
4720 * This code is optimized for readability - several passes might be
4721 * made over certain portions of the tree. All of those blocks will
4722 * have been brought into cache (and pinned via the journal), so the
4723 * extra overhead is not expressed in terms of disk reads.
4725 static int __ocfs2_mark_extent_written(struct inode *inode,
4726 struct ocfs2_extent_tree *et,
4728 struct ocfs2_path *path,
4730 struct ocfs2_extent_rec *split_rec,
4731 struct ocfs2_alloc_context *meta_ac,
4732 struct ocfs2_cached_dealloc_ctxt *dealloc)
4735 struct ocfs2_extent_list *el = path_leaf_el(path);
4736 struct buffer_head *last_eb_bh = NULL;
4737 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
4738 struct ocfs2_merge_ctxt ctxt;
4739 struct ocfs2_extent_list *rightmost_el;
4741 if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) {
4747 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
4748 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
4749 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
4755 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, path, el,
4760 * The core merge / split code wants to know how much room is
4761 * left in this inodes allocation tree, so we pass the
4762 * rightmost extent list.
4764 if (path->p_tree_depth) {
4765 struct ocfs2_extent_block *eb;
4767 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4768 ocfs2_et_get_last_eb_blk(et),
4769 &last_eb_bh, OCFS2_BH_CACHED, inode);
4775 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4776 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
4777 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
4782 rightmost_el = &eb->h_list;
4784 rightmost_el = path_root_el(path);
4786 if (rec->e_cpos == split_rec->e_cpos &&
4787 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
4788 ctxt.c_split_covers_rec = 1;
4790 ctxt.c_split_covers_rec = 0;
4792 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
4794 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
4795 split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent,
4796 ctxt.c_split_covers_rec);
4798 if (ctxt.c_contig_type == CONTIG_NONE) {
4799 if (ctxt.c_split_covers_rec)
4800 el->l_recs[split_index] = *split_rec;
4802 ret = ocfs2_split_and_insert(inode, handle, path, et,
4803 &last_eb_bh, split_index,
4804 split_rec, meta_ac);
4808 ret = ocfs2_try_to_merge_extent(inode, handle, path,
4809 split_index, split_rec,
4810 dealloc, &ctxt, et);
4821 * Mark the already-existing extent at cpos as written for len clusters.
4823 * If the existing extent is larger than the request, initiate a
4824 * split. An attempt will be made at merging with adjacent extents.
4826 * The caller is responsible for passing down meta_ac if we'll need it.
4828 int ocfs2_mark_extent_written(struct inode *inode,
4829 struct ocfs2_extent_tree *et,
4830 handle_t *handle, u32 cpos, u32 len, u32 phys,
4831 struct ocfs2_alloc_context *meta_ac,
4832 struct ocfs2_cached_dealloc_ctxt *dealloc)
4835 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
4836 struct ocfs2_extent_rec split_rec;
4837 struct ocfs2_path *left_path = NULL;
4838 struct ocfs2_extent_list *el;
4840 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4841 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
4843 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
4844 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
4845 "that are being written to, but the feature bit "
4846 "is not set in the super block.",
4847 (unsigned long long)OCFS2_I(inode)->ip_blkno);
4853 * XXX: This should be fixed up so that we just re-insert the
4854 * next extent records.
4856 * XXX: This is a hack on the extent tree, maybe it should be
4859 if (et->et_ops == &ocfs2_dinode_et_ops)
4860 ocfs2_extent_map_trunc(inode, 0);
4862 left_path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
4869 ret = ocfs2_find_path(inode, left_path, cpos);
4874 el = path_leaf_el(left_path);
4876 index = ocfs2_search_extent_list(el, cpos);
4877 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4878 ocfs2_error(inode->i_sb,
4879 "Inode %llu has an extent at cpos %u which can no "
4880 "longer be found.\n",
4881 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
4886 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
4887 split_rec.e_cpos = cpu_to_le32(cpos);
4888 split_rec.e_leaf_clusters = cpu_to_le16(len);
4889 split_rec.e_blkno = cpu_to_le64(start_blkno);
4890 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
4891 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
4893 ret = __ocfs2_mark_extent_written(inode, et, handle, left_path,
4894 index, &split_rec, meta_ac,
4900 ocfs2_free_path(left_path);
4904 static int ocfs2_split_tree(struct inode *inode, struct ocfs2_extent_tree *et,
4905 handle_t *handle, struct ocfs2_path *path,
4906 int index, u32 new_range,
4907 struct ocfs2_alloc_context *meta_ac)
4909 int ret, depth, credits = handle->h_buffer_credits;
4910 struct buffer_head *last_eb_bh = NULL;
4911 struct ocfs2_extent_block *eb;
4912 struct ocfs2_extent_list *rightmost_el, *el;
4913 struct ocfs2_extent_rec split_rec;
4914 struct ocfs2_extent_rec *rec;
4915 struct ocfs2_insert_type insert;
4918 * Setup the record to split before we grow the tree.
4920 el = path_leaf_el(path);
4921 rec = &el->l_recs[index];
4922 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
4924 depth = path->p_tree_depth;
4926 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4927 ocfs2_et_get_last_eb_blk(et),
4928 &last_eb_bh, OCFS2_BH_CACHED, inode);
4934 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4935 rightmost_el = &eb->h_list;
4937 rightmost_el = path_leaf_el(path);
4939 credits += path->p_tree_depth +
4940 ocfs2_extend_meta_needed(et->et_root_el);
4941 ret = ocfs2_extend_trans(handle, credits);
4947 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4948 le16_to_cpu(rightmost_el->l_count)) {
4949 ret = ocfs2_grow_tree(inode, handle, et, &depth, &last_eb_bh,
4957 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4958 insert.ins_appending = APPEND_NONE;
4959 insert.ins_contig = CONTIG_NONE;
4960 insert.ins_split = SPLIT_RIGHT;
4961 insert.ins_tree_depth = depth;
4963 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4972 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
4973 struct ocfs2_path *path, int index,
4974 struct ocfs2_cached_dealloc_ctxt *dealloc,
4976 struct ocfs2_extent_tree *et)
4979 u32 left_cpos, rec_range, trunc_range;
4980 int wants_rotate = 0, is_rightmost_tree_rec = 0;
4981 struct super_block *sb = inode->i_sb;
4982 struct ocfs2_path *left_path = NULL;
4983 struct ocfs2_extent_list *el = path_leaf_el(path);
4984 struct ocfs2_extent_rec *rec;
4985 struct ocfs2_extent_block *eb;
4987 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
4988 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
4997 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
4998 path->p_tree_depth) {
5000 * Check whether this is the rightmost tree record. If
5001 * we remove all of this record or part of its right
5002 * edge then an update of the record lengths above it
5005 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5006 if (eb->h_next_leaf_blk == 0)
5007 is_rightmost_tree_rec = 1;
5010 rec = &el->l_recs[index];
5011 if (index == 0 && path->p_tree_depth &&
5012 le32_to_cpu(rec->e_cpos) == cpos) {
5014 * Changing the leftmost offset (via partial or whole
5015 * record truncate) of an interior (or rightmost) path
5016 * means we have to update the subtree that is formed
5017 * by this leaf and the one to it's left.
5019 * There are two cases we can skip:
5020 * 1) Path is the leftmost one in our inode tree.
5021 * 2) The leaf is rightmost and will be empty after
5022 * we remove the extent record - the rotate code
5023 * knows how to update the newly formed edge.
5026 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
5033 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5034 left_path = ocfs2_new_path(path_root_bh(path),
5035 path_root_el(path));
5042 ret = ocfs2_find_path(inode, left_path, left_cpos);
5050 ret = ocfs2_extend_rotate_transaction(handle, 0,
5051 handle->h_buffer_credits,
5058 ret = ocfs2_journal_access_path(inode, handle, path);
5064 ret = ocfs2_journal_access_path(inode, handle, left_path);
5070 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5071 trunc_range = cpos + len;
5073 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5076 memset(rec, 0, sizeof(*rec));
5077 ocfs2_cleanup_merge(el, index);
5080 next_free = le16_to_cpu(el->l_next_free_rec);
5081 if (is_rightmost_tree_rec && next_free > 1) {
5083 * We skip the edge update if this path will
5084 * be deleted by the rotate code.
5086 rec = &el->l_recs[next_free - 1];
5087 ocfs2_adjust_rightmost_records(inode, handle, path,
5090 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5091 /* Remove leftmost portion of the record. */
5092 le32_add_cpu(&rec->e_cpos, len);
5093 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5094 le16_add_cpu(&rec->e_leaf_clusters, -len);
5095 } else if (rec_range == trunc_range) {
5096 /* Remove rightmost portion of the record */
5097 le16_add_cpu(&rec->e_leaf_clusters, -len);
5098 if (is_rightmost_tree_rec)
5099 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
5101 /* Caller should have trapped this. */
5102 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
5103 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
5104 le32_to_cpu(rec->e_cpos),
5105 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5112 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
5113 ocfs2_complete_edge_insert(inode, handle, left_path, path,
5117 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5119 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5126 ocfs2_free_path(left_path);
5130 int ocfs2_remove_extent(struct inode *inode,
5131 struct ocfs2_extent_tree *et,
5132 u32 cpos, u32 len, handle_t *handle,
5133 struct ocfs2_alloc_context *meta_ac,
5134 struct ocfs2_cached_dealloc_ctxt *dealloc)
5137 u32 rec_range, trunc_range;
5138 struct ocfs2_extent_rec *rec;
5139 struct ocfs2_extent_list *el;
5140 struct ocfs2_path *path = NULL;
5142 ocfs2_extent_map_trunc(inode, 0);
5144 path = ocfs2_new_path(et->et_root_bh, et->et_root_el);
5151 ret = ocfs2_find_path(inode, path, cpos);
5157 el = path_leaf_el(path);
5158 index = ocfs2_search_extent_list(el, cpos);
5159 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5160 ocfs2_error(inode->i_sb,
5161 "Inode %llu has an extent at cpos %u which can no "
5162 "longer be found.\n",
5163 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5169 * We have 3 cases of extent removal:
5170 * 1) Range covers the entire extent rec
5171 * 2) Range begins or ends on one edge of the extent rec
5172 * 3) Range is in the middle of the extent rec (no shared edges)
5174 * For case 1 we remove the extent rec and left rotate to
5177 * For case 2 we just shrink the existing extent rec, with a
5178 * tree update if the shrinking edge is also the edge of an
5181 * For case 3 we do a right split to turn the extent rec into
5182 * something case 2 can handle.
5184 rec = &el->l_recs[index];
5185 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5186 trunc_range = cpos + len;
5188 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5190 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
5191 "(cpos %u, len %u)\n",
5192 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
5193 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
5195 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5196 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5203 ret = ocfs2_split_tree(inode, et, handle, path, index,
5204 trunc_range, meta_ac);
5211 * The split could have manipulated the tree enough to
5212 * move the record location, so we have to look for it again.
5214 ocfs2_reinit_path(path, 1);
5216 ret = ocfs2_find_path(inode, path, cpos);
5222 el = path_leaf_el(path);
5223 index = ocfs2_search_extent_list(el, cpos);
5224 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5225 ocfs2_error(inode->i_sb,
5226 "Inode %llu: split at cpos %u lost record.",
5227 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5234 * Double check our values here. If anything is fishy,
5235 * it's easier to catch it at the top level.
5237 rec = &el->l_recs[index];
5238 rec_range = le32_to_cpu(rec->e_cpos) +
5239 ocfs2_rec_clusters(el, rec);
5240 if (rec_range != trunc_range) {
5241 ocfs2_error(inode->i_sb,
5242 "Inode %llu: error after split at cpos %u"
5243 "trunc len %u, existing record is (%u,%u)",
5244 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5245 cpos, len, le32_to_cpu(rec->e_cpos),
5246 ocfs2_rec_clusters(el, rec));
5251 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5260 ocfs2_free_path(path);
5264 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5266 struct buffer_head *tl_bh = osb->osb_tl_bh;
5267 struct ocfs2_dinode *di;
5268 struct ocfs2_truncate_log *tl;
5270 di = (struct ocfs2_dinode *) tl_bh->b_data;
5271 tl = &di->id2.i_dealloc;
5273 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5274 "slot %d, invalid truncate log parameters: used = "
5275 "%u, count = %u\n", osb->slot_num,
5276 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5277 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5280 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5281 unsigned int new_start)
5283 unsigned int tail_index;
5284 unsigned int current_tail;
5286 /* No records, nothing to coalesce */
5287 if (!le16_to_cpu(tl->tl_used))
5290 tail_index = le16_to_cpu(tl->tl_used) - 1;
5291 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5292 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5294 return current_tail == new_start;
5297 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5300 unsigned int num_clusters)
5303 unsigned int start_cluster, tl_count;
5304 struct inode *tl_inode = osb->osb_tl_inode;
5305 struct buffer_head *tl_bh = osb->osb_tl_bh;
5306 struct ocfs2_dinode *di;
5307 struct ocfs2_truncate_log *tl;
5309 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5310 (unsigned long long)start_blk, num_clusters);
5312 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5314 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5316 di = (struct ocfs2_dinode *) tl_bh->b_data;
5317 tl = &di->id2.i_dealloc;
5318 if (!OCFS2_IS_VALID_DINODE(di)) {
5319 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
5324 tl_count = le16_to_cpu(tl->tl_count);
5325 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5327 "Truncate record count on #%llu invalid "
5328 "wanted %u, actual %u\n",
5329 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5330 ocfs2_truncate_recs_per_inode(osb->sb),
5331 le16_to_cpu(tl->tl_count));
5333 /* Caller should have known to flush before calling us. */
5334 index = le16_to_cpu(tl->tl_used);
5335 if (index >= tl_count) {
5341 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
5342 OCFS2_JOURNAL_ACCESS_WRITE);
5348 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5349 "%llu (index = %d)\n", num_clusters, start_cluster,
5350 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
5352 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5354 * Move index back to the record we are coalescing with.
5355 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5359 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5360 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5361 index, le32_to_cpu(tl->tl_recs[index].t_start),
5364 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5365 tl->tl_used = cpu_to_le16(index + 1);
5367 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5369 status = ocfs2_journal_dirty(handle, tl_bh);
5380 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5382 struct inode *data_alloc_inode,
5383 struct buffer_head *data_alloc_bh)
5387 unsigned int num_clusters;
5389 struct ocfs2_truncate_rec rec;
5390 struct ocfs2_dinode *di;
5391 struct ocfs2_truncate_log *tl;
5392 struct inode *tl_inode = osb->osb_tl_inode;
5393 struct buffer_head *tl_bh = osb->osb_tl_bh;
5397 di = (struct ocfs2_dinode *) tl_bh->b_data;
5398 tl = &di->id2.i_dealloc;
5399 i = le16_to_cpu(tl->tl_used) - 1;
5401 /* Caller has given us at least enough credits to
5402 * update the truncate log dinode */
5403 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
5404 OCFS2_JOURNAL_ACCESS_WRITE);
5410 tl->tl_used = cpu_to_le16(i);
5412 status = ocfs2_journal_dirty(handle, tl_bh);
5418 /* TODO: Perhaps we can calculate the bulk of the
5419 * credits up front rather than extending like
5421 status = ocfs2_extend_trans(handle,
5422 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5428 rec = tl->tl_recs[i];
5429 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5430 le32_to_cpu(rec.t_start));
5431 num_clusters = le32_to_cpu(rec.t_clusters);
5433 /* if start_blk is not set, we ignore the record as
5436 mlog(0, "free record %d, start = %u, clusters = %u\n",
5437 i, le32_to_cpu(rec.t_start), num_clusters);
5439 status = ocfs2_free_clusters(handle, data_alloc_inode,
5440 data_alloc_bh, start_blk,
5455 /* Expects you to already be holding tl_inode->i_mutex */
5456 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5459 unsigned int num_to_flush;
5461 struct inode *tl_inode = osb->osb_tl_inode;
5462 struct inode *data_alloc_inode = NULL;
5463 struct buffer_head *tl_bh = osb->osb_tl_bh;
5464 struct buffer_head *data_alloc_bh = NULL;
5465 struct ocfs2_dinode *di;
5466 struct ocfs2_truncate_log *tl;
5470 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5472 di = (struct ocfs2_dinode *) tl_bh->b_data;
5473 tl = &di->id2.i_dealloc;
5474 if (!OCFS2_IS_VALID_DINODE(di)) {
5475 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
5480 num_to_flush = le16_to_cpu(tl->tl_used);
5481 mlog(0, "Flush %u records from truncate log #%llu\n",
5482 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
5483 if (!num_to_flush) {
5488 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5489 GLOBAL_BITMAP_SYSTEM_INODE,
5490 OCFS2_INVALID_SLOT);
5491 if (!data_alloc_inode) {
5493 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5497 mutex_lock(&data_alloc_inode->i_mutex);
5499 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5505 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5506 if (IS_ERR(handle)) {
5507 status = PTR_ERR(handle);
5512 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
5517 ocfs2_commit_trans(osb, handle);
5520 brelse(data_alloc_bh);
5521 ocfs2_inode_unlock(data_alloc_inode, 1);
5524 mutex_unlock(&data_alloc_inode->i_mutex);
5525 iput(data_alloc_inode);
5532 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5535 struct inode *tl_inode = osb->osb_tl_inode;
5537 mutex_lock(&tl_inode->i_mutex);
5538 status = __ocfs2_flush_truncate_log(osb);
5539 mutex_unlock(&tl_inode->i_mutex);
5544 static void ocfs2_truncate_log_worker(struct work_struct *work)
5547 struct ocfs2_super *osb =
5548 container_of(work, struct ocfs2_super,
5549 osb_truncate_log_wq.work);
5553 status = ocfs2_flush_truncate_log(osb);
5557 ocfs2_init_inode_steal_slot(osb);
5562 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5563 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
5566 if (osb->osb_tl_inode) {
5567 /* We want to push off log flushes while truncates are
5570 cancel_delayed_work(&osb->osb_truncate_log_wq);
5572 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
5573 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
5577 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
5579 struct inode **tl_inode,
5580 struct buffer_head **tl_bh)
5583 struct inode *inode = NULL;
5584 struct buffer_head *bh = NULL;
5586 inode = ocfs2_get_system_file_inode(osb,
5587 TRUNCATE_LOG_SYSTEM_INODE,
5591 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
5595 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
5596 OCFS2_BH_CACHED, inode);
5610 /* called during the 1st stage of node recovery. we stamp a clean
5611 * truncate log and pass back a copy for processing later. if the
5612 * truncate log does not require processing, a *tl_copy is set to
5614 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
5616 struct ocfs2_dinode **tl_copy)
5619 struct inode *tl_inode = NULL;
5620 struct buffer_head *tl_bh = NULL;
5621 struct ocfs2_dinode *di;
5622 struct ocfs2_truncate_log *tl;
5626 mlog(0, "recover truncate log from slot %d\n", slot_num);
5628 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
5634 di = (struct ocfs2_dinode *) tl_bh->b_data;
5635 tl = &di->id2.i_dealloc;
5636 if (!OCFS2_IS_VALID_DINODE(di)) {
5637 OCFS2_RO_ON_INVALID_DINODE(tl_inode->i_sb, di);
5642 if (le16_to_cpu(tl->tl_used)) {
5643 mlog(0, "We'll have %u logs to recover\n",
5644 le16_to_cpu(tl->tl_used));
5646 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
5653 /* Assuming the write-out below goes well, this copy
5654 * will be passed back to recovery for processing. */
5655 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
5657 /* All we need to do to clear the truncate log is set
5661 status = ocfs2_write_block(osb, tl_bh, tl_inode);
5673 if (status < 0 && (*tl_copy)) {
5682 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
5683 struct ocfs2_dinode *tl_copy)
5687 unsigned int clusters, num_recs, start_cluster;
5690 struct inode *tl_inode = osb->osb_tl_inode;
5691 struct ocfs2_truncate_log *tl;
5695 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
5696 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
5700 tl = &tl_copy->id2.i_dealloc;
5701 num_recs = le16_to_cpu(tl->tl_used);
5702 mlog(0, "cleanup %u records from %llu\n", num_recs,
5703 (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
5705 mutex_lock(&tl_inode->i_mutex);
5706 for(i = 0; i < num_recs; i++) {
5707 if (ocfs2_truncate_log_needs_flush(osb)) {
5708 status = __ocfs2_flush_truncate_log(osb);
5715 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5716 if (IS_ERR(handle)) {
5717 status = PTR_ERR(handle);
5722 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
5723 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
5724 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
5726 status = ocfs2_truncate_log_append(osb, handle,
5727 start_blk, clusters);
5728 ocfs2_commit_trans(osb, handle);
5736 mutex_unlock(&tl_inode->i_mutex);
5742 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
5745 struct inode *tl_inode = osb->osb_tl_inode;
5750 cancel_delayed_work(&osb->osb_truncate_log_wq);
5751 flush_workqueue(ocfs2_wq);
5753 status = ocfs2_flush_truncate_log(osb);
5757 brelse(osb->osb_tl_bh);
5758 iput(osb->osb_tl_inode);
5764 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
5767 struct inode *tl_inode = NULL;
5768 struct buffer_head *tl_bh = NULL;
5772 status = ocfs2_get_truncate_log_info(osb,
5779 /* ocfs2_truncate_log_shutdown keys on the existence of
5780 * osb->osb_tl_inode so we don't set any of the osb variables
5781 * until we're sure all is well. */
5782 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
5783 ocfs2_truncate_log_worker);
5784 osb->osb_tl_bh = tl_bh;
5785 osb->osb_tl_inode = tl_inode;
5792 * Delayed de-allocation of suballocator blocks.
5794 * Some sets of block de-allocations might involve multiple suballocator inodes.
5796 * The locking for this can get extremely complicated, especially when
5797 * the suballocator inodes to delete from aren't known until deep
5798 * within an unrelated codepath.
5800 * ocfs2_extent_block structures are a good example of this - an inode
5801 * btree could have been grown by any number of nodes each allocating
5802 * out of their own suballoc inode.
5804 * These structures allow the delay of block de-allocation until a
5805 * later time, when locking of multiple cluster inodes won't cause
5810 * Describes a single block free from a suballocator
5812 struct ocfs2_cached_block_free {
5813 struct ocfs2_cached_block_free *free_next;
5815 unsigned int free_bit;
5818 struct ocfs2_per_slot_free_list {
5819 struct ocfs2_per_slot_free_list *f_next_suballocator;
5822 struct ocfs2_cached_block_free *f_first;
5825 static int ocfs2_free_cached_items(struct ocfs2_super *osb,
5828 struct ocfs2_cached_block_free *head)
5833 struct inode *inode;
5834 struct buffer_head *di_bh = NULL;
5835 struct ocfs2_cached_block_free *tmp;
5837 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
5844 mutex_lock(&inode->i_mutex);
5846 ret = ocfs2_inode_lock(inode, &di_bh, 1);
5852 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
5853 if (IS_ERR(handle)) {
5854 ret = PTR_ERR(handle);
5860 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
5862 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
5863 head->free_bit, (unsigned long long)head->free_blk);
5865 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
5866 head->free_bit, bg_blkno, 1);
5872 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
5879 head = head->free_next;
5884 ocfs2_commit_trans(osb, handle);
5887 ocfs2_inode_unlock(inode, 1);
5890 mutex_unlock(&inode->i_mutex);
5894 /* Premature exit may have left some dangling items. */
5896 head = head->free_next;
5903 int ocfs2_run_deallocs(struct ocfs2_super *osb,
5904 struct ocfs2_cached_dealloc_ctxt *ctxt)
5907 struct ocfs2_per_slot_free_list *fl;
5912 while (ctxt->c_first_suballocator) {
5913 fl = ctxt->c_first_suballocator;
5916 mlog(0, "Free items: (type %u, slot %d)\n",
5917 fl->f_inode_type, fl->f_slot);
5918 ret2 = ocfs2_free_cached_items(osb, fl->f_inode_type,
5919 fl->f_slot, fl->f_first);
5926 ctxt->c_first_suballocator = fl->f_next_suballocator;
5933 static struct ocfs2_per_slot_free_list *
5934 ocfs2_find_per_slot_free_list(int type,
5936 struct ocfs2_cached_dealloc_ctxt *ctxt)
5938 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
5941 if (fl->f_inode_type == type && fl->f_slot == slot)
5944 fl = fl->f_next_suballocator;
5947 fl = kmalloc(sizeof(*fl), GFP_NOFS);
5949 fl->f_inode_type = type;
5952 fl->f_next_suballocator = ctxt->c_first_suballocator;
5954 ctxt->c_first_suballocator = fl;
5959 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
5960 int type, int slot, u64 blkno,
5964 struct ocfs2_per_slot_free_list *fl;
5965 struct ocfs2_cached_block_free *item;
5967 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
5974 item = kmalloc(sizeof(*item), GFP_NOFS);
5981 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
5982 type, slot, bit, (unsigned long long)blkno);
5984 item->free_blk = blkno;
5985 item->free_bit = bit;
5986 item->free_next = fl->f_first;
5995 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
5996 struct ocfs2_extent_block *eb)
5998 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
5999 le16_to_cpu(eb->h_suballoc_slot),
6000 le64_to_cpu(eb->h_blkno),
6001 le16_to_cpu(eb->h_suballoc_bit));
6004 /* This function will figure out whether the currently last extent
6005 * block will be deleted, and if it will, what the new last extent
6006 * block will be so we can update his h_next_leaf_blk field, as well
6007 * as the dinodes i_last_eb_blk */
6008 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
6009 unsigned int clusters_to_del,
6010 struct ocfs2_path *path,
6011 struct buffer_head **new_last_eb)
6013 int next_free, ret = 0;
6015 struct ocfs2_extent_rec *rec;
6016 struct ocfs2_extent_block *eb;
6017 struct ocfs2_extent_list *el;
6018 struct buffer_head *bh = NULL;
6020 *new_last_eb = NULL;
6022 /* we have no tree, so of course, no last_eb. */
6023 if (!path->p_tree_depth)
6026 /* trunc to zero special case - this makes tree_depth = 0
6027 * regardless of what it is. */
6028 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
6031 el = path_leaf_el(path);
6032 BUG_ON(!el->l_next_free_rec);
6035 * Make sure that this extent list will actually be empty
6036 * after we clear away the data. We can shortcut out if
6037 * there's more than one non-empty extent in the
6038 * list. Otherwise, a check of the remaining extent is
6041 next_free = le16_to_cpu(el->l_next_free_rec);
6043 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6047 /* We may have a valid extent in index 1, check it. */
6049 rec = &el->l_recs[1];
6052 * Fall through - no more nonempty extents, so we want
6053 * to delete this leaf.
6059 rec = &el->l_recs[0];
6064 * Check it we'll only be trimming off the end of this
6067 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
6071 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
6077 ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
6083 eb = (struct ocfs2_extent_block *) bh->b_data;
6085 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
6086 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
6092 get_bh(*new_last_eb);
6093 mlog(0, "returning block %llu, (cpos: %u)\n",
6094 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
6102 * Trim some clusters off the rightmost edge of a tree. Only called
6105 * The caller needs to:
6106 * - start journaling of each path component.
6107 * - compute and fully set up any new last ext block
6109 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
6110 handle_t *handle, struct ocfs2_truncate_context *tc,
6111 u32 clusters_to_del, u64 *delete_start)
6113 int ret, i, index = path->p_tree_depth;
6116 struct buffer_head *bh;
6117 struct ocfs2_extent_list *el;
6118 struct ocfs2_extent_rec *rec;
6122 while (index >= 0) {
6123 bh = path->p_node[index].bh;
6124 el = path->p_node[index].el;
6126 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6127 index, (unsigned long long)bh->b_blocknr);
6129 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
6132 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
6133 ocfs2_error(inode->i_sb,
6134 "Inode %lu has invalid ext. block %llu",
6136 (unsigned long long)bh->b_blocknr);
6142 i = le16_to_cpu(el->l_next_free_rec) - 1;
6143 rec = &el->l_recs[i];
6145 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6146 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
6147 ocfs2_rec_clusters(el, rec),
6148 (unsigned long long)le64_to_cpu(rec->e_blkno),
6149 le16_to_cpu(el->l_next_free_rec));
6151 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
6153 if (le16_to_cpu(el->l_tree_depth) == 0) {
6155 * If the leaf block contains a single empty
6156 * extent and no records, we can just remove
6159 if (i == 0 && ocfs2_is_empty_extent(rec)) {
6161 sizeof(struct ocfs2_extent_rec));
6162 el->l_next_free_rec = cpu_to_le16(0);
6168 * Remove any empty extents by shifting things
6169 * left. That should make life much easier on
6170 * the code below. This condition is rare
6171 * enough that we shouldn't see a performance
6174 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6175 le16_add_cpu(&el->l_next_free_rec, -1);
6178 i < le16_to_cpu(el->l_next_free_rec); i++)
6179 el->l_recs[i] = el->l_recs[i + 1];
6181 memset(&el->l_recs[i], 0,
6182 sizeof(struct ocfs2_extent_rec));
6185 * We've modified our extent list. The
6186 * simplest way to handle this change
6187 * is to being the search from the
6190 goto find_tail_record;
6193 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
6196 * We'll use "new_edge" on our way back up the
6197 * tree to know what our rightmost cpos is.
6199 new_edge = le16_to_cpu(rec->e_leaf_clusters);
6200 new_edge += le32_to_cpu(rec->e_cpos);
6203 * The caller will use this to delete data blocks.
6205 *delete_start = le64_to_cpu(rec->e_blkno)
6206 + ocfs2_clusters_to_blocks(inode->i_sb,
6207 le16_to_cpu(rec->e_leaf_clusters));
6210 * If it's now empty, remove this record.
6212 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
6214 sizeof(struct ocfs2_extent_rec));
6215 le16_add_cpu(&el->l_next_free_rec, -1);
6218 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
6220 sizeof(struct ocfs2_extent_rec));
6221 le16_add_cpu(&el->l_next_free_rec, -1);
6226 /* Can this actually happen? */
6227 if (le16_to_cpu(el->l_next_free_rec) == 0)
6231 * We never actually deleted any clusters
6232 * because our leaf was empty. There's no
6233 * reason to adjust the rightmost edge then.
6238 rec->e_int_clusters = cpu_to_le32(new_edge);
6239 le32_add_cpu(&rec->e_int_clusters,
6240 -le32_to_cpu(rec->e_cpos));
6243 * A deleted child record should have been
6246 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
6250 ret = ocfs2_journal_dirty(handle, bh);
6256 mlog(0, "extent list container %llu, after: record %d: "
6257 "(%u, %u, %llu), next = %u.\n",
6258 (unsigned long long)bh->b_blocknr, i,
6259 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
6260 (unsigned long long)le64_to_cpu(rec->e_blkno),
6261 le16_to_cpu(el->l_next_free_rec));
6264 * We must be careful to only attempt delete of an
6265 * extent block (and not the root inode block).
6267 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
6268 struct ocfs2_extent_block *eb =
6269 (struct ocfs2_extent_block *)bh->b_data;
6272 * Save this for use when processing the
6275 deleted_eb = le64_to_cpu(eb->h_blkno);
6277 mlog(0, "deleting this extent block.\n");
6279 ocfs2_remove_from_cache(inode, bh);
6281 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
6282 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
6283 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
6285 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
6286 /* An error here is not fatal. */
6301 static int ocfs2_do_truncate(struct ocfs2_super *osb,
6302 unsigned int clusters_to_del,
6303 struct inode *inode,
6304 struct buffer_head *fe_bh,
6306 struct ocfs2_truncate_context *tc,
6307 struct ocfs2_path *path)
6310 struct ocfs2_dinode *fe;
6311 struct ocfs2_extent_block *last_eb = NULL;
6312 struct ocfs2_extent_list *el;
6313 struct buffer_head *last_eb_bh = NULL;
6316 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6318 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
6326 * Each component will be touched, so we might as well journal
6327 * here to avoid having to handle errors later.
6329 status = ocfs2_journal_access_path(inode, handle, path);
6336 status = ocfs2_journal_access(handle, inode, last_eb_bh,
6337 OCFS2_JOURNAL_ACCESS_WRITE);
6343 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6346 el = &(fe->id2.i_list);
6349 * Lower levels depend on this never happening, but it's best
6350 * to check it up here before changing the tree.
6352 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
6353 ocfs2_error(inode->i_sb,
6354 "Inode %lu has an empty extent record, depth %u\n",
6355 inode->i_ino, le16_to_cpu(el->l_tree_depth));
6360 spin_lock(&OCFS2_I(inode)->ip_lock);
6361 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
6363 spin_unlock(&OCFS2_I(inode)->ip_lock);
6364 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
6365 inode->i_blocks = ocfs2_inode_sector_count(inode);
6367 status = ocfs2_trim_tree(inode, path, handle, tc,
6368 clusters_to_del, &delete_blk);
6374 if (le32_to_cpu(fe->i_clusters) == 0) {
6375 /* trunc to zero is a special case. */
6376 el->l_tree_depth = 0;
6377 fe->i_last_eb_blk = 0;
6379 fe->i_last_eb_blk = last_eb->h_blkno;
6381 status = ocfs2_journal_dirty(handle, fe_bh);
6388 /* If there will be a new last extent block, then by
6389 * definition, there cannot be any leaves to the right of
6391 last_eb->h_next_leaf_blk = 0;
6392 status = ocfs2_journal_dirty(handle, last_eb_bh);
6400 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
6414 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6416 set_buffer_uptodate(bh);
6417 mark_buffer_dirty(bh);
6421 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6422 unsigned int from, unsigned int to,
6423 struct page *page, int zero, u64 *phys)
6425 int ret, partial = 0;
6427 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6432 zero_user_segment(page, from, to);
6435 * Need to set the buffers we zero'd into uptodate
6436 * here if they aren't - ocfs2_map_page_blocks()
6437 * might've skipped some
6439 ret = walk_page_buffers(handle, page_buffers(page),
6444 else if (ocfs2_should_order_data(inode)) {
6445 ret = ocfs2_jbd2_file_inode(handle, inode);
6446 #ifdef CONFIG_OCFS2_COMPAT_JBD
6447 ret = walk_page_buffers(handle, page_buffers(page),
6449 ocfs2_journal_dirty_data);
6456 SetPageUptodate(page);
6458 flush_dcache_page(page);
6461 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6462 loff_t end, struct page **pages,
6463 int numpages, u64 phys, handle_t *handle)
6467 unsigned int from, to = PAGE_CACHE_SIZE;
6468 struct super_block *sb = inode->i_sb;
6470 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6475 to = PAGE_CACHE_SIZE;
6476 for(i = 0; i < numpages; i++) {
6479 from = start & (PAGE_CACHE_SIZE - 1);
6480 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6481 to = end & (PAGE_CACHE_SIZE - 1);
6483 BUG_ON(from > PAGE_CACHE_SIZE);
6484 BUG_ON(to > PAGE_CACHE_SIZE);
6486 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6489 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6493 ocfs2_unlock_and_free_pages(pages, numpages);
6496 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6497 struct page **pages, int *num)
6499 int numpages, ret = 0;
6500 struct super_block *sb = inode->i_sb;
6501 struct address_space *mapping = inode->i_mapping;
6502 unsigned long index;
6503 loff_t last_page_bytes;
6505 BUG_ON(start > end);
6507 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6508 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6511 last_page_bytes = PAGE_ALIGN(end);
6512 index = start >> PAGE_CACHE_SHIFT;
6514 pages[numpages] = grab_cache_page(mapping, index);
6515 if (!pages[numpages]) {
6523 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6528 ocfs2_unlock_and_free_pages(pages, numpages);
6538 * Zero the area past i_size but still within an allocated
6539 * cluster. This avoids exposing nonzero data on subsequent file
6542 * We need to call this before i_size is updated on the inode because
6543 * otherwise block_write_full_page() will skip writeout of pages past
6544 * i_size. The new_i_size parameter is passed for this reason.
6546 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6547 u64 range_start, u64 range_end)
6549 int ret = 0, numpages;
6550 struct page **pages = NULL;
6552 unsigned int ext_flags;
6553 struct super_block *sb = inode->i_sb;
6556 * File systems which don't support sparse files zero on every
6559 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6562 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6563 sizeof(struct page *), GFP_NOFS);
6564 if (pages == NULL) {
6570 if (range_start == range_end)
6573 ret = ocfs2_extent_map_get_blocks(inode,
6574 range_start >> sb->s_blocksize_bits,
6575 &phys, NULL, &ext_flags);
6582 * Tail is a hole, or is marked unwritten. In either case, we
6583 * can count on read and write to return/push zero's.
6585 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6588 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6595 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6596 numpages, phys, handle);
6599 * Initiate writeout of the pages we zero'd here. We don't
6600 * wait on them - the truncate_inode_pages() call later will
6603 ret = do_sync_mapping_range(inode->i_mapping, range_start,
6604 range_end - 1, SYNC_FILE_RANGE_WRITE);
6615 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6616 struct ocfs2_dinode *di)
6618 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6619 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6621 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6622 memset(&di->id2, 0, blocksize -
6623 offsetof(struct ocfs2_dinode, id2) -
6626 memset(&di->id2, 0, blocksize -
6627 offsetof(struct ocfs2_dinode, id2));
6630 void ocfs2_dinode_new_extent_list(struct inode *inode,
6631 struct ocfs2_dinode *di)
6633 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6634 di->id2.i_list.l_tree_depth = 0;
6635 di->id2.i_list.l_next_free_rec = 0;
6636 di->id2.i_list.l_count = cpu_to_le16(
6637 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6640 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6642 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6643 struct ocfs2_inline_data *idata = &di->id2.i_data;
6645 spin_lock(&oi->ip_lock);
6646 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6647 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6648 spin_unlock(&oi->ip_lock);
6651 * We clear the entire i_data structure here so that all
6652 * fields can be properly initialized.
6654 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6656 idata->id_count = cpu_to_le16(
6657 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6660 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6661 struct buffer_head *di_bh)
6663 int ret, i, has_data, num_pages = 0;
6665 u64 uninitialized_var(block);
6666 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6667 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6668 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6669 struct ocfs2_alloc_context *data_ac = NULL;
6670 struct page **pages = NULL;
6671 loff_t end = osb->s_clustersize;
6672 struct ocfs2_extent_tree et;
6674 has_data = i_size_read(inode) ? 1 : 0;
6677 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6678 sizeof(struct page *), GFP_NOFS);
6679 if (pages == NULL) {
6685 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6692 handle = ocfs2_start_trans(osb, OCFS2_INLINE_TO_EXTENTS_CREDITS);
6693 if (IS_ERR(handle)) {
6694 ret = PTR_ERR(handle);
6699 ret = ocfs2_journal_access(handle, inode, di_bh,
6700 OCFS2_JOURNAL_ACCESS_WRITE);
6708 unsigned int page_end;
6711 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
6719 * Save two copies, one for insert, and one that can
6720 * be changed by ocfs2_map_and_dirty_page() below.
6722 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6725 * Non sparse file systems zero on extend, so no need
6728 if (!ocfs2_sparse_alloc(osb) &&
6729 PAGE_CACHE_SIZE < osb->s_clustersize)
6730 end = PAGE_CACHE_SIZE;
6732 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6739 * This should populate the 1st page for us and mark
6742 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6748 page_end = PAGE_CACHE_SIZE;
6749 if (PAGE_CACHE_SIZE > osb->s_clustersize)
6750 page_end = osb->s_clustersize;
6752 for (i = 0; i < num_pages; i++)
6753 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6754 pages[i], i > 0, &phys);
6757 spin_lock(&oi->ip_lock);
6758 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6759 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6760 spin_unlock(&oi->ip_lock);
6762 ocfs2_dinode_new_extent_list(inode, di);
6764 ocfs2_journal_dirty(handle, di_bh);
6768 * An error at this point should be extremely rare. If
6769 * this proves to be false, we could always re-build
6770 * the in-inode data from our pages.
6772 ocfs2_init_dinode_extent_tree(&et, inode, di_bh);
6773 ret = ocfs2_insert_extent(osb, handle, inode, &et,
6774 0, block, 1, 0, NULL);
6780 inode->i_blocks = ocfs2_inode_sector_count(inode);
6784 ocfs2_commit_trans(osb, handle);
6788 ocfs2_free_alloc_context(data_ac);
6792 ocfs2_unlock_and_free_pages(pages, num_pages);
6800 * It is expected, that by the time you call this function,
6801 * inode->i_size and fe->i_size have been adjusted.
6803 * WARNING: This will kfree the truncate context
6805 int ocfs2_commit_truncate(struct ocfs2_super *osb,
6806 struct inode *inode,
6807 struct buffer_head *fe_bh,
6808 struct ocfs2_truncate_context *tc)
6810 int status, i, credits, tl_sem = 0;
6811 u32 clusters_to_del, new_highest_cpos, range;
6812 struct ocfs2_extent_list *el;
6813 handle_t *handle = NULL;
6814 struct inode *tl_inode = osb->osb_tl_inode;
6815 struct ocfs2_path *path = NULL;
6816 struct ocfs2_dinode *di = (struct ocfs2_dinode *)fe_bh->b_data;
6820 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
6821 i_size_read(inode));
6823 path = ocfs2_new_path(fe_bh, &di->id2.i_list);
6830 ocfs2_extent_map_trunc(inode, new_highest_cpos);
6834 * Check that we still have allocation to delete.
6836 if (OCFS2_I(inode)->ip_clusters == 0) {
6842 * Truncate always works against the rightmost tree branch.
6844 status = ocfs2_find_path(inode, path, UINT_MAX);
6850 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
6851 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
6854 * By now, el will point to the extent list on the bottom most
6855 * portion of this tree. Only the tail record is considered in
6858 * We handle the following cases, in order:
6859 * - empty extent: delete the remaining branch
6860 * - remove the entire record
6861 * - remove a partial record
6862 * - no record needs to be removed (truncate has completed)
6864 el = path_leaf_el(path);
6865 if (le16_to_cpu(el->l_next_free_rec) == 0) {
6866 ocfs2_error(inode->i_sb,
6867 "Inode %llu has empty extent block at %llu\n",
6868 (unsigned long long)OCFS2_I(inode)->ip_blkno,
6869 (unsigned long long)path_leaf_bh(path)->b_blocknr);
6874 i = le16_to_cpu(el->l_next_free_rec) - 1;
6875 range = le32_to_cpu(el->l_recs[i].e_cpos) +
6876 ocfs2_rec_clusters(el, &el->l_recs[i]);
6877 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
6878 clusters_to_del = 0;
6879 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
6880 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
6881 } else if (range > new_highest_cpos) {
6882 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
6883 le32_to_cpu(el->l_recs[i].e_cpos)) -
6890 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
6891 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
6893 mutex_lock(&tl_inode->i_mutex);
6895 /* ocfs2_truncate_log_needs_flush guarantees us at least one
6896 * record is free for use. If there isn't any, we flush to get
6897 * an empty truncate log. */
6898 if (ocfs2_truncate_log_needs_flush(osb)) {
6899 status = __ocfs2_flush_truncate_log(osb);
6906 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
6907 (struct ocfs2_dinode *)fe_bh->b_data,
6909 handle = ocfs2_start_trans(osb, credits);
6910 if (IS_ERR(handle)) {
6911 status = PTR_ERR(handle);
6917 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
6924 mutex_unlock(&tl_inode->i_mutex);
6927 ocfs2_commit_trans(osb, handle);
6930 ocfs2_reinit_path(path, 1);
6933 * The check above will catch the case where we've truncated
6934 * away all allocation.
6940 ocfs2_schedule_truncate_log_flush(osb, 1);
6943 mutex_unlock(&tl_inode->i_mutex);
6946 ocfs2_commit_trans(osb, handle);
6948 ocfs2_run_deallocs(osb, &tc->tc_dealloc);
6950 ocfs2_free_path(path);
6952 /* This will drop the ext_alloc cluster lock for us */
6953 ocfs2_free_truncate_context(tc);
6960 * Expects the inode to already be locked.
6962 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
6963 struct inode *inode,
6964 struct buffer_head *fe_bh,
6965 struct ocfs2_truncate_context **tc)
6968 unsigned int new_i_clusters;
6969 struct ocfs2_dinode *fe;
6970 struct ocfs2_extent_block *eb;
6971 struct buffer_head *last_eb_bh = NULL;
6977 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
6978 i_size_read(inode));
6979 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6981 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
6982 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
6983 (unsigned long long)le64_to_cpu(fe->i_size));
6985 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
6991 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
6993 if (fe->id2.i_list.l_tree_depth) {
6994 status = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
6995 &last_eb_bh, OCFS2_BH_CACHED, inode);
7000 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
7001 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
7002 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
7010 (*tc)->tc_last_eb_bh = last_eb_bh;
7016 ocfs2_free_truncate_context(*tc);
7024 * 'start' is inclusive, 'end' is not.
7026 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7027 unsigned int start, unsigned int end, int trunc)
7030 unsigned int numbytes;
7032 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7033 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7034 struct ocfs2_inline_data *idata = &di->id2.i_data;
7036 if (end > i_size_read(inode))
7037 end = i_size_read(inode);
7039 BUG_ON(start >= end);
7041 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7042 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7043 !ocfs2_supports_inline_data(osb)) {
7044 ocfs2_error(inode->i_sb,
7045 "Inline data flags for inode %llu don't agree! "
7046 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7047 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7048 le16_to_cpu(di->i_dyn_features),
7049 OCFS2_I(inode)->ip_dyn_features,
7050 osb->s_feature_incompat);
7055 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7056 if (IS_ERR(handle)) {
7057 ret = PTR_ERR(handle);
7062 ret = ocfs2_journal_access(handle, inode, di_bh,
7063 OCFS2_JOURNAL_ACCESS_WRITE);
7069 numbytes = end - start;
7070 memset(idata->id_data + start, 0, numbytes);
7073 * No need to worry about the data page here - it's been
7074 * truncated already and inline data doesn't need it for
7075 * pushing zero's to disk, so we'll let readpage pick it up
7079 i_size_write(inode, start);
7080 di->i_size = cpu_to_le64(start);
7083 inode->i_blocks = ocfs2_inode_sector_count(inode);
7084 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7086 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7087 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7089 ocfs2_journal_dirty(handle, di_bh);
7092 ocfs2_commit_trans(osb, handle);
7098 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
7101 * The caller is responsible for completing deallocation
7102 * before freeing the context.
7104 if (tc->tc_dealloc.c_first_suballocator != NULL)
7106 "Truncate completion has non-empty dealloc context\n");
7108 brelse(tc->tc_last_eb_bh);