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>
31 #include <linux/quotaops.h>
33 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
34 #include <cluster/masklog.h>
40 #include "blockcheck.h"
42 #include "extent_map.h"
45 #include "localalloc.h"
53 #include "buffer_head_io.h"
57 * Operations for a specific extent tree type.
59 * To implement an on-disk btree (extent tree) type in ocfs2, add
60 * an ocfs2_extent_tree_operations structure and the matching
61 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
62 * for the allocation portion of the extent tree.
64 struct ocfs2_extent_tree_operations {
66 * last_eb_blk is the block number of the right most leaf extent
67 * block. Most on-disk structures containing an extent tree store
68 * this value for fast access. The ->eo_set_last_eb_blk() and
69 * ->eo_get_last_eb_blk() operations access this value. They are
72 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
74 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
77 * The on-disk structure usually keeps track of how many total
78 * clusters are stored in this extent tree. This function updates
79 * that value. new_clusters is the delta, and must be
80 * added to the total. Required.
82 void (*eo_update_clusters)(struct inode *inode,
83 struct ocfs2_extent_tree *et,
87 * If ->eo_insert_check() exists, it is called before rec is
88 * inserted into the extent tree. It is optional.
90 int (*eo_insert_check)(struct inode *inode,
91 struct ocfs2_extent_tree *et,
92 struct ocfs2_extent_rec *rec);
93 int (*eo_sanity_check)(struct inode *inode, struct ocfs2_extent_tree *et);
96 * --------------------------------------------------------------
97 * The remaining are internal to ocfs2_extent_tree and don't have
102 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
105 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
108 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
109 * it exists. If it does not, et->et_max_leaf_clusters is set
110 * to 0 (unlimited). Optional.
112 void (*eo_fill_max_leaf_clusters)(struct inode *inode,
113 struct ocfs2_extent_tree *et);
118 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
121 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
122 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
124 static void ocfs2_dinode_update_clusters(struct inode *inode,
125 struct ocfs2_extent_tree *et,
127 static int ocfs2_dinode_insert_check(struct inode *inode,
128 struct ocfs2_extent_tree *et,
129 struct ocfs2_extent_rec *rec);
130 static int ocfs2_dinode_sanity_check(struct inode *inode,
131 struct ocfs2_extent_tree *et);
132 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
133 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
134 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
135 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
136 .eo_update_clusters = ocfs2_dinode_update_clusters,
137 .eo_insert_check = ocfs2_dinode_insert_check,
138 .eo_sanity_check = ocfs2_dinode_sanity_check,
139 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
142 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
145 struct ocfs2_dinode *di = et->et_object;
147 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
148 di->i_last_eb_blk = cpu_to_le64(blkno);
151 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
153 struct ocfs2_dinode *di = et->et_object;
155 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
156 return le64_to_cpu(di->i_last_eb_blk);
159 static void ocfs2_dinode_update_clusters(struct inode *inode,
160 struct ocfs2_extent_tree *et,
163 struct ocfs2_dinode *di = et->et_object;
165 le32_add_cpu(&di->i_clusters, clusters);
166 spin_lock(&OCFS2_I(inode)->ip_lock);
167 OCFS2_I(inode)->ip_clusters = le32_to_cpu(di->i_clusters);
168 spin_unlock(&OCFS2_I(inode)->ip_lock);
171 static int ocfs2_dinode_insert_check(struct inode *inode,
172 struct ocfs2_extent_tree *et,
173 struct ocfs2_extent_rec *rec)
175 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
177 BUG_ON(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL);
178 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
179 (OCFS2_I(inode)->ip_clusters !=
180 le32_to_cpu(rec->e_cpos)),
181 "Device %s, asking for sparse allocation: inode %llu, "
182 "cpos %u, clusters %u\n",
184 (unsigned long long)OCFS2_I(inode)->ip_blkno,
186 OCFS2_I(inode)->ip_clusters);
191 static int ocfs2_dinode_sanity_check(struct inode *inode,
192 struct ocfs2_extent_tree *et)
194 struct ocfs2_dinode *di = et->et_object;
196 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
197 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
202 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
204 struct ocfs2_dinode *di = et->et_object;
206 et->et_root_el = &di->id2.i_list;
210 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
212 struct ocfs2_xattr_value_buf *vb = et->et_object;
214 et->et_root_el = &vb->vb_xv->xr_list;
217 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
220 struct ocfs2_xattr_value_buf *vb = et->et_object;
222 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
225 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
227 struct ocfs2_xattr_value_buf *vb = et->et_object;
229 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
232 static void ocfs2_xattr_value_update_clusters(struct inode *inode,
233 struct ocfs2_extent_tree *et,
236 struct ocfs2_xattr_value_buf *vb = et->et_object;
238 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
241 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
242 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
243 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
244 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
245 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
248 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
250 struct ocfs2_xattr_block *xb = et->et_object;
252 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
255 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct inode *inode,
256 struct ocfs2_extent_tree *et)
258 et->et_max_leaf_clusters =
259 ocfs2_clusters_for_bytes(inode->i_sb,
260 OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
263 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
266 struct ocfs2_xattr_block *xb = et->et_object;
267 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
269 xt->xt_last_eb_blk = cpu_to_le64(blkno);
272 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
274 struct ocfs2_xattr_block *xb = et->et_object;
275 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
277 return le64_to_cpu(xt->xt_last_eb_blk);
280 static void ocfs2_xattr_tree_update_clusters(struct inode *inode,
281 struct ocfs2_extent_tree *et,
284 struct ocfs2_xattr_block *xb = et->et_object;
286 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
289 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
290 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
291 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
292 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
293 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
294 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
297 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
300 struct ocfs2_dx_root_block *dx_root = et->et_object;
302 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
305 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
307 struct ocfs2_dx_root_block *dx_root = et->et_object;
309 return le64_to_cpu(dx_root->dr_last_eb_blk);
312 static void ocfs2_dx_root_update_clusters(struct inode *inode,
313 struct ocfs2_extent_tree *et,
316 struct ocfs2_dx_root_block *dx_root = et->et_object;
318 le32_add_cpu(&dx_root->dr_clusters, clusters);
321 static int ocfs2_dx_root_sanity_check(struct inode *inode,
322 struct ocfs2_extent_tree *et)
324 struct ocfs2_dx_root_block *dx_root = et->et_object;
326 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
331 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
333 struct ocfs2_dx_root_block *dx_root = et->et_object;
335 et->et_root_el = &dx_root->dr_list;
338 static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
339 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
340 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
341 .eo_update_clusters = ocfs2_dx_root_update_clusters,
342 .eo_sanity_check = ocfs2_dx_root_sanity_check,
343 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
346 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
348 struct buffer_head *bh,
349 ocfs2_journal_access_func access,
351 struct ocfs2_extent_tree_operations *ops)
355 et->et_ci = INODE_CACHE(inode);
356 et->et_root_journal_access = access;
358 obj = (void *)bh->b_data;
361 et->et_ops->eo_fill_root_el(et);
362 if (!et->et_ops->eo_fill_max_leaf_clusters)
363 et->et_max_leaf_clusters = 0;
365 et->et_ops->eo_fill_max_leaf_clusters(inode, et);
368 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
370 struct buffer_head *bh)
372 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_di,
373 NULL, &ocfs2_dinode_et_ops);
376 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
378 struct buffer_head *bh)
380 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_xb,
381 NULL, &ocfs2_xattr_tree_et_ops);
384 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
386 struct ocfs2_xattr_value_buf *vb)
388 __ocfs2_init_extent_tree(et, inode, vb->vb_bh, vb->vb_access, vb,
389 &ocfs2_xattr_value_et_ops);
392 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
394 struct buffer_head *bh)
396 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_dr,
397 NULL, &ocfs2_dx_root_et_ops);
400 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
403 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
406 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
408 return et->et_ops->eo_get_last_eb_blk(et);
411 static inline void ocfs2_et_update_clusters(struct inode *inode,
412 struct ocfs2_extent_tree *et,
415 et->et_ops->eo_update_clusters(inode, et, clusters);
418 static inline int ocfs2_et_root_journal_access(handle_t *handle,
419 struct ocfs2_extent_tree *et,
422 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
426 static inline int ocfs2_et_insert_check(struct inode *inode,
427 struct ocfs2_extent_tree *et,
428 struct ocfs2_extent_rec *rec)
432 if (et->et_ops->eo_insert_check)
433 ret = et->et_ops->eo_insert_check(inode, et, rec);
437 static inline int ocfs2_et_sanity_check(struct inode *inode,
438 struct ocfs2_extent_tree *et)
442 if (et->et_ops->eo_sanity_check)
443 ret = et->et_ops->eo_sanity_check(inode, et);
447 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
448 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
449 struct ocfs2_extent_block *eb);
452 * Structures which describe a path through a btree, and functions to
455 * The idea here is to be as generic as possible with the tree
458 struct ocfs2_path_item {
459 struct buffer_head *bh;
460 struct ocfs2_extent_list *el;
463 #define OCFS2_MAX_PATH_DEPTH 5
467 ocfs2_journal_access_func p_root_access;
468 struct ocfs2_path_item p_node[OCFS2_MAX_PATH_DEPTH];
471 #define path_root_bh(_path) ((_path)->p_node[0].bh)
472 #define path_root_el(_path) ((_path)->p_node[0].el)
473 #define path_root_access(_path)((_path)->p_root_access)
474 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
475 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
476 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
478 static int ocfs2_find_path(struct ocfs2_caching_info *ci,
479 struct ocfs2_path *path, u32 cpos);
480 static void ocfs2_adjust_rightmost_records(struct inode *inode,
482 struct ocfs2_path *path,
483 struct ocfs2_extent_rec *insert_rec);
485 * Reset the actual path elements so that we can re-use the structure
486 * to build another path. Generally, this involves freeing the buffer
489 static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
491 int i, start = 0, depth = 0;
492 struct ocfs2_path_item *node;
497 for(i = start; i < path_num_items(path); i++) {
498 node = &path->p_node[i];
506 * Tree depth may change during truncate, or insert. If we're
507 * keeping the root extent list, then make sure that our path
508 * structure reflects the proper depth.
511 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
513 path_root_access(path) = NULL;
515 path->p_tree_depth = depth;
518 static void ocfs2_free_path(struct ocfs2_path *path)
521 ocfs2_reinit_path(path, 0);
527 * All the elements of src into dest. After this call, src could be freed
528 * without affecting dest.
530 * Both paths should have the same root. Any non-root elements of dest
533 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
537 BUG_ON(path_root_bh(dest) != path_root_bh(src));
538 BUG_ON(path_root_el(dest) != path_root_el(src));
539 BUG_ON(path_root_access(dest) != path_root_access(src));
541 ocfs2_reinit_path(dest, 1);
543 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
544 dest->p_node[i].bh = src->p_node[i].bh;
545 dest->p_node[i].el = src->p_node[i].el;
547 if (dest->p_node[i].bh)
548 get_bh(dest->p_node[i].bh);
553 * Make the *dest path the same as src and re-initialize src path to
556 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
560 BUG_ON(path_root_bh(dest) != path_root_bh(src));
561 BUG_ON(path_root_access(dest) != path_root_access(src));
563 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
564 brelse(dest->p_node[i].bh);
566 dest->p_node[i].bh = src->p_node[i].bh;
567 dest->p_node[i].el = src->p_node[i].el;
569 src->p_node[i].bh = NULL;
570 src->p_node[i].el = NULL;
575 * Insert an extent block at given index.
577 * This will not take an additional reference on eb_bh.
579 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
580 struct buffer_head *eb_bh)
582 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
585 * Right now, no root bh is an extent block, so this helps
586 * catch code errors with dinode trees. The assertion can be
587 * safely removed if we ever need to insert extent block
588 * structures at the root.
592 path->p_node[index].bh = eb_bh;
593 path->p_node[index].el = &eb->h_list;
596 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
597 struct ocfs2_extent_list *root_el,
598 ocfs2_journal_access_func access)
600 struct ocfs2_path *path;
602 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
604 path = kzalloc(sizeof(*path), GFP_NOFS);
606 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
608 path_root_bh(path) = root_bh;
609 path_root_el(path) = root_el;
610 path_root_access(path) = access;
616 static struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
618 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
619 path_root_access(path));
622 static struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
624 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
625 et->et_root_journal_access);
629 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
630 * otherwise it's the root_access function.
632 * I don't like the way this function's name looks next to
633 * ocfs2_journal_access_path(), but I don't have a better one.
635 static int ocfs2_path_bh_journal_access(handle_t *handle,
636 struct ocfs2_caching_info *ci,
637 struct ocfs2_path *path,
640 ocfs2_journal_access_func access = path_root_access(path);
643 access = ocfs2_journal_access;
646 access = ocfs2_journal_access_eb;
648 return access(handle, ci, path->p_node[idx].bh,
649 OCFS2_JOURNAL_ACCESS_WRITE);
653 * Convenience function to journal all components in a path.
655 static int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
657 struct ocfs2_path *path)
664 for(i = 0; i < path_num_items(path); i++) {
665 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
677 * Return the index of the extent record which contains cluster #v_cluster.
678 * -1 is returned if it was not found.
680 * Should work fine on interior and exterior nodes.
682 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
686 struct ocfs2_extent_rec *rec;
687 u32 rec_end, rec_start, clusters;
689 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
690 rec = &el->l_recs[i];
692 rec_start = le32_to_cpu(rec->e_cpos);
693 clusters = ocfs2_rec_clusters(el, rec);
695 rec_end = rec_start + clusters;
697 if (v_cluster >= rec_start && v_cluster < rec_end) {
706 enum ocfs2_contig_type {
715 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
716 * ocfs2_extent_contig only work properly against leaf nodes!
718 static int ocfs2_block_extent_contig(struct super_block *sb,
719 struct ocfs2_extent_rec *ext,
722 u64 blk_end = le64_to_cpu(ext->e_blkno);
724 blk_end += ocfs2_clusters_to_blocks(sb,
725 le16_to_cpu(ext->e_leaf_clusters));
727 return blkno == blk_end;
730 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
731 struct ocfs2_extent_rec *right)
735 left_range = le32_to_cpu(left->e_cpos) +
736 le16_to_cpu(left->e_leaf_clusters);
738 return (left_range == le32_to_cpu(right->e_cpos));
741 static enum ocfs2_contig_type
742 ocfs2_extent_contig(struct inode *inode,
743 struct ocfs2_extent_rec *ext,
744 struct ocfs2_extent_rec *insert_rec)
746 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
749 * Refuse to coalesce extent records with different flag
750 * fields - we don't want to mix unwritten extents with user
753 if (ext->e_flags != insert_rec->e_flags)
756 if (ocfs2_extents_adjacent(ext, insert_rec) &&
757 ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
760 blkno = le64_to_cpu(ext->e_blkno);
761 if (ocfs2_extents_adjacent(insert_rec, ext) &&
762 ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno))
769 * NOTE: We can have pretty much any combination of contiguousness and
772 * The usefulness of APPEND_TAIL is more in that it lets us know that
773 * we'll have to update the path to that leaf.
775 enum ocfs2_append_type {
780 enum ocfs2_split_type {
786 struct ocfs2_insert_type {
787 enum ocfs2_split_type ins_split;
788 enum ocfs2_append_type ins_appending;
789 enum ocfs2_contig_type ins_contig;
790 int ins_contig_index;
794 struct ocfs2_merge_ctxt {
795 enum ocfs2_contig_type c_contig_type;
796 int c_has_empty_extent;
797 int c_split_covers_rec;
800 static int ocfs2_validate_extent_block(struct super_block *sb,
801 struct buffer_head *bh)
804 struct ocfs2_extent_block *eb =
805 (struct ocfs2_extent_block *)bh->b_data;
807 mlog(0, "Validating extent block %llu\n",
808 (unsigned long long)bh->b_blocknr);
810 BUG_ON(!buffer_uptodate(bh));
813 * If the ecc fails, we return the error but otherwise
814 * leave the filesystem running. We know any error is
815 * local to this block.
817 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
819 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
820 (unsigned long long)bh->b_blocknr);
825 * Errors after here are fatal.
828 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
830 "Extent block #%llu has bad signature %.*s",
831 (unsigned long long)bh->b_blocknr, 7,
836 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
838 "Extent block #%llu has an invalid h_blkno "
840 (unsigned long long)bh->b_blocknr,
841 (unsigned long long)le64_to_cpu(eb->h_blkno));
845 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
847 "Extent block #%llu has an invalid "
848 "h_fs_generation of #%u",
849 (unsigned long long)bh->b_blocknr,
850 le32_to_cpu(eb->h_fs_generation));
857 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
858 struct buffer_head **bh)
861 struct buffer_head *tmp = *bh;
863 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
864 ocfs2_validate_extent_block);
866 /* If ocfs2_read_block() got us a new bh, pass it up. */
875 * How many free extents have we got before we need more meta data?
877 int ocfs2_num_free_extents(struct ocfs2_super *osb,
878 struct ocfs2_extent_tree *et)
881 struct ocfs2_extent_list *el = NULL;
882 struct ocfs2_extent_block *eb;
883 struct buffer_head *eb_bh = NULL;
889 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
892 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
898 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
902 BUG_ON(el->l_tree_depth != 0);
904 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
912 /* expects array to already be allocated
914 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
917 static int ocfs2_create_new_meta_bhs(handle_t *handle,
918 struct ocfs2_extent_tree *et,
920 struct ocfs2_alloc_context *meta_ac,
921 struct buffer_head *bhs[])
923 int count, status, i;
924 u16 suballoc_bit_start;
927 struct ocfs2_super *osb =
928 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
929 struct ocfs2_extent_block *eb;
934 while (count < wanted) {
935 status = ocfs2_claim_metadata(osb,
947 for(i = count; i < (num_got + count); i++) {
948 bhs[i] = sb_getblk(osb->sb, first_blkno);
949 if (bhs[i] == NULL) {
954 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
956 status = ocfs2_journal_access_eb(handle, et->et_ci,
958 OCFS2_JOURNAL_ACCESS_CREATE);
964 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
965 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
966 /* Ok, setup the minimal stuff here. */
967 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
968 eb->h_blkno = cpu_to_le64(first_blkno);
969 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
970 eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
971 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
973 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
975 suballoc_bit_start++;
978 /* We'll also be dirtied by the caller, so
979 * this isn't absolutely necessary. */
980 status = ocfs2_journal_dirty(handle, bhs[i]);
993 for(i = 0; i < wanted; i++) {
1003 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1005 * Returns the sum of the rightmost extent rec logical offset and
1008 * ocfs2_add_branch() uses this to determine what logical cluster
1009 * value should be populated into the leftmost new branch records.
1011 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1012 * value for the new topmost tree record.
1014 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1018 i = le16_to_cpu(el->l_next_free_rec) - 1;
1020 return le32_to_cpu(el->l_recs[i].e_cpos) +
1021 ocfs2_rec_clusters(el, &el->l_recs[i]);
1025 * Change range of the branches in the right most path according to the leaf
1026 * extent block's rightmost record.
1028 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1029 struct inode *inode,
1030 struct ocfs2_extent_tree *et)
1033 struct ocfs2_path *path = NULL;
1034 struct ocfs2_extent_list *el;
1035 struct ocfs2_extent_rec *rec;
1037 path = ocfs2_new_path_from_et(et);
1043 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1049 status = ocfs2_extend_trans(handle, path_num_items(path) +
1050 handle->h_buffer_credits);
1056 status = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
1062 el = path_leaf_el(path);
1063 rec = &el->l_recs[le32_to_cpu(el->l_next_free_rec) - 1];
1065 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
1068 ocfs2_free_path(path);
1073 * Add an entire tree branch to our inode. eb_bh is the extent block
1074 * to start at, if we don't want to start the branch at the dinode
1077 * last_eb_bh is required as we have to update it's next_leaf pointer
1078 * for the new last extent block.
1080 * the new branch will be 'empty' in the sense that every block will
1081 * contain a single record with cluster count == 0.
1083 static int ocfs2_add_branch(struct ocfs2_super *osb,
1085 struct inode *inode,
1086 struct ocfs2_extent_tree *et,
1087 struct buffer_head *eb_bh,
1088 struct buffer_head **last_eb_bh,
1089 struct ocfs2_alloc_context *meta_ac)
1091 int status, new_blocks, i;
1092 u64 next_blkno, new_last_eb_blk;
1093 struct buffer_head *bh;
1094 struct buffer_head **new_eb_bhs = NULL;
1095 struct ocfs2_extent_block *eb;
1096 struct ocfs2_extent_list *eb_el;
1097 struct ocfs2_extent_list *el;
1098 u32 new_cpos, root_end;
1102 BUG_ON(!last_eb_bh || !*last_eb_bh);
1105 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1108 el = et->et_root_el;
1110 /* we never add a branch to a leaf. */
1111 BUG_ON(!el->l_tree_depth);
1113 new_blocks = le16_to_cpu(el->l_tree_depth);
1115 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1116 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1117 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1120 * If there is a gap before the root end and the real end
1121 * of the righmost leaf block, we need to remove the gap
1122 * between new_cpos and root_end first so that the tree
1123 * is consistent after we add a new branch(it will start
1126 if (root_end > new_cpos) {
1127 mlog(0, "adjust the cluster end from %u to %u\n",
1128 root_end, new_cpos);
1129 status = ocfs2_adjust_rightmost_branch(handle, inode, et);
1136 /* allocate the number of new eb blocks we need */
1137 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1145 status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1146 meta_ac, new_eb_bhs);
1152 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1153 * linked with the rest of the tree.
1154 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1156 * when we leave the loop, new_last_eb_blk will point to the
1157 * newest leaf, and next_blkno will point to the topmost extent
1159 next_blkno = new_last_eb_blk = 0;
1160 for(i = 0; i < new_blocks; i++) {
1162 eb = (struct ocfs2_extent_block *) bh->b_data;
1163 /* ocfs2_create_new_meta_bhs() should create it right! */
1164 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1165 eb_el = &eb->h_list;
1167 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), bh,
1168 OCFS2_JOURNAL_ACCESS_CREATE);
1174 eb->h_next_leaf_blk = 0;
1175 eb_el->l_tree_depth = cpu_to_le16(i);
1176 eb_el->l_next_free_rec = cpu_to_le16(1);
1178 * This actually counts as an empty extent as
1181 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1182 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1184 * eb_el isn't always an interior node, but even leaf
1185 * nodes want a zero'd flags and reserved field so
1186 * this gets the whole 32 bits regardless of use.
1188 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1189 if (!eb_el->l_tree_depth)
1190 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1192 status = ocfs2_journal_dirty(handle, bh);
1198 next_blkno = le64_to_cpu(eb->h_blkno);
1201 /* This is a bit hairy. We want to update up to three blocks
1202 * here without leaving any of them in an inconsistent state
1203 * in case of error. We don't have to worry about
1204 * journal_dirty erroring as it won't unless we've aborted the
1205 * handle (in which case we would never be here) so reserving
1206 * the write with journal_access is all we need to do. */
1207 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), *last_eb_bh,
1208 OCFS2_JOURNAL_ACCESS_WRITE);
1213 status = ocfs2_et_root_journal_access(handle, et,
1214 OCFS2_JOURNAL_ACCESS_WRITE);
1220 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), eb_bh,
1221 OCFS2_JOURNAL_ACCESS_WRITE);
1228 /* Link the new branch into the rest of the tree (el will
1229 * either be on the root_bh, or the extent block passed in. */
1230 i = le16_to_cpu(el->l_next_free_rec);
1231 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1232 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1233 el->l_recs[i].e_int_clusters = 0;
1234 le16_add_cpu(&el->l_next_free_rec, 1);
1236 /* fe needs a new last extent block pointer, as does the
1237 * next_leaf on the previously last-extent-block. */
1238 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1240 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1241 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1243 status = ocfs2_journal_dirty(handle, *last_eb_bh);
1246 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1250 status = ocfs2_journal_dirty(handle, eb_bh);
1256 * Some callers want to track the rightmost leaf so pass it
1259 brelse(*last_eb_bh);
1260 get_bh(new_eb_bhs[0]);
1261 *last_eb_bh = new_eb_bhs[0];
1266 for (i = 0; i < new_blocks; i++)
1267 brelse(new_eb_bhs[i]);
1276 * adds another level to the allocation tree.
1277 * returns back the new extent block so you can add a branch to it
1280 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
1282 struct inode *inode,
1283 struct ocfs2_extent_tree *et,
1284 struct ocfs2_alloc_context *meta_ac,
1285 struct buffer_head **ret_new_eb_bh)
1289 struct buffer_head *new_eb_bh = NULL;
1290 struct ocfs2_extent_block *eb;
1291 struct ocfs2_extent_list *root_el;
1292 struct ocfs2_extent_list *eb_el;
1296 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1303 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1304 /* ocfs2_create_new_meta_bhs() should create it right! */
1305 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1307 eb_el = &eb->h_list;
1308 root_el = et->et_root_el;
1310 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), new_eb_bh,
1311 OCFS2_JOURNAL_ACCESS_CREATE);
1317 /* copy the root extent list data into the new extent block */
1318 eb_el->l_tree_depth = root_el->l_tree_depth;
1319 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1320 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1321 eb_el->l_recs[i] = root_el->l_recs[i];
1323 status = ocfs2_journal_dirty(handle, new_eb_bh);
1329 status = ocfs2_et_root_journal_access(handle, et,
1330 OCFS2_JOURNAL_ACCESS_WRITE);
1336 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1338 /* update root_bh now */
1339 le16_add_cpu(&root_el->l_tree_depth, 1);
1340 root_el->l_recs[0].e_cpos = 0;
1341 root_el->l_recs[0].e_blkno = eb->h_blkno;
1342 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1343 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1344 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1345 root_el->l_next_free_rec = cpu_to_le16(1);
1347 /* If this is our 1st tree depth shift, then last_eb_blk
1348 * becomes the allocated extent block */
1349 if (root_el->l_tree_depth == cpu_to_le16(1))
1350 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1352 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1358 *ret_new_eb_bh = new_eb_bh;
1369 * Should only be called when there is no space left in any of the
1370 * leaf nodes. What we want to do is find the lowest tree depth
1371 * non-leaf extent block with room for new records. There are three
1372 * valid results of this search:
1374 * 1) a lowest extent block is found, then we pass it back in
1375 * *lowest_eb_bh and return '0'
1377 * 2) the search fails to find anything, but the root_el has room. We
1378 * pass NULL back in *lowest_eb_bh, but still return '0'
1380 * 3) the search fails to find anything AND the root_el is full, in
1381 * which case we return > 0
1383 * return status < 0 indicates an error.
1385 static int ocfs2_find_branch_target(struct ocfs2_super *osb,
1386 struct ocfs2_extent_tree *et,
1387 struct buffer_head **target_bh)
1391 struct ocfs2_extent_block *eb;
1392 struct ocfs2_extent_list *el;
1393 struct buffer_head *bh = NULL;
1394 struct buffer_head *lowest_bh = NULL;
1400 el = et->et_root_el;
1402 while(le16_to_cpu(el->l_tree_depth) > 1) {
1403 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1404 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1405 "Owner %llu has empty "
1406 "extent list (next_free_rec == 0)",
1407 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1411 i = le16_to_cpu(el->l_next_free_rec) - 1;
1412 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1414 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1415 "Owner %llu has extent "
1416 "list where extent # %d has no physical "
1418 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1426 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1432 eb = (struct ocfs2_extent_block *) bh->b_data;
1435 if (le16_to_cpu(el->l_next_free_rec) <
1436 le16_to_cpu(el->l_count)) {
1443 /* If we didn't find one and the fe doesn't have any room,
1444 * then return '1' */
1445 el = et->et_root_el;
1446 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1449 *target_bh = lowest_bh;
1458 * Grow a b-tree so that it has more records.
1460 * We might shift the tree depth in which case existing paths should
1461 * be considered invalid.
1463 * Tree depth after the grow is returned via *final_depth.
1465 * *last_eb_bh will be updated by ocfs2_add_branch().
1467 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
1468 struct ocfs2_extent_tree *et, int *final_depth,
1469 struct buffer_head **last_eb_bh,
1470 struct ocfs2_alloc_context *meta_ac)
1473 struct ocfs2_extent_list *el = et->et_root_el;
1474 int depth = le16_to_cpu(el->l_tree_depth);
1475 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1476 struct buffer_head *bh = NULL;
1478 BUG_ON(meta_ac == NULL);
1480 shift = ocfs2_find_branch_target(osb, et, &bh);
1487 /* We traveled all the way to the bottom of the allocation tree
1488 * and didn't find room for any more extents - we need to add
1489 * another tree level */
1492 mlog(0, "need to shift tree depth (current = %d)\n", depth);
1494 /* ocfs2_shift_tree_depth will return us a buffer with
1495 * the new extent block (so we can pass that to
1496 * ocfs2_add_branch). */
1497 ret = ocfs2_shift_tree_depth(osb, handle, inode, et,
1506 * Special case: we have room now if we shifted from
1507 * tree_depth 0, so no more work needs to be done.
1509 * We won't be calling add_branch, so pass
1510 * back *last_eb_bh as the new leaf. At depth
1511 * zero, it should always be null so there's
1512 * no reason to brelse.
1514 BUG_ON(*last_eb_bh);
1521 /* call ocfs2_add_branch to add the final part of the tree with
1523 mlog(0, "add branch. bh = %p\n", bh);
1524 ret = ocfs2_add_branch(osb, handle, inode, et, bh, last_eb_bh,
1533 *final_depth = depth;
1539 * This function will discard the rightmost extent record.
1541 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1543 int next_free = le16_to_cpu(el->l_next_free_rec);
1544 int count = le16_to_cpu(el->l_count);
1545 unsigned int num_bytes;
1548 /* This will cause us to go off the end of our extent list. */
1549 BUG_ON(next_free >= count);
1551 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1553 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1556 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1557 struct ocfs2_extent_rec *insert_rec)
1559 int i, insert_index, next_free, has_empty, num_bytes;
1560 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1561 struct ocfs2_extent_rec *rec;
1563 next_free = le16_to_cpu(el->l_next_free_rec);
1564 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1568 /* The tree code before us didn't allow enough room in the leaf. */
1569 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1572 * The easiest way to approach this is to just remove the
1573 * empty extent and temporarily decrement next_free.
1577 * If next_free was 1 (only an empty extent), this
1578 * loop won't execute, which is fine. We still want
1579 * the decrement above to happen.
1581 for(i = 0; i < (next_free - 1); i++)
1582 el->l_recs[i] = el->l_recs[i+1];
1588 * Figure out what the new record index should be.
1590 for(i = 0; i < next_free; i++) {
1591 rec = &el->l_recs[i];
1593 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1598 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1599 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
1601 BUG_ON(insert_index < 0);
1602 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1603 BUG_ON(insert_index > next_free);
1606 * No need to memmove if we're just adding to the tail.
1608 if (insert_index != next_free) {
1609 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1611 num_bytes = next_free - insert_index;
1612 num_bytes *= sizeof(struct ocfs2_extent_rec);
1613 memmove(&el->l_recs[insert_index + 1],
1614 &el->l_recs[insert_index],
1619 * Either we had an empty extent, and need to re-increment or
1620 * there was no empty extent on a non full rightmost leaf node,
1621 * in which case we still need to increment.
1624 el->l_next_free_rec = cpu_to_le16(next_free);
1626 * Make sure none of the math above just messed up our tree.
1628 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1630 el->l_recs[insert_index] = *insert_rec;
1634 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1636 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1638 BUG_ON(num_recs == 0);
1640 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1642 size = num_recs * sizeof(struct ocfs2_extent_rec);
1643 memmove(&el->l_recs[0], &el->l_recs[1], size);
1644 memset(&el->l_recs[num_recs], 0,
1645 sizeof(struct ocfs2_extent_rec));
1646 el->l_next_free_rec = cpu_to_le16(num_recs);
1651 * Create an empty extent record .
1653 * l_next_free_rec may be updated.
1655 * If an empty extent already exists do nothing.
1657 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1659 int next_free = le16_to_cpu(el->l_next_free_rec);
1661 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1666 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1669 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1670 "Asked to create an empty extent in a full list:\n"
1671 "count = %u, tree depth = %u",
1672 le16_to_cpu(el->l_count),
1673 le16_to_cpu(el->l_tree_depth));
1675 ocfs2_shift_records_right(el);
1678 le16_add_cpu(&el->l_next_free_rec, 1);
1679 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1683 * For a rotation which involves two leaf nodes, the "root node" is
1684 * the lowest level tree node which contains a path to both leafs. This
1685 * resulting set of information can be used to form a complete "subtree"
1687 * This function is passed two full paths from the dinode down to a
1688 * pair of adjacent leaves. It's task is to figure out which path
1689 * index contains the subtree root - this can be the root index itself
1690 * in a worst-case rotation.
1692 * The array index of the subtree root is passed back.
1694 static int ocfs2_find_subtree_root(struct inode *inode,
1695 struct ocfs2_path *left,
1696 struct ocfs2_path *right)
1701 * Check that the caller passed in two paths from the same tree.
1703 BUG_ON(path_root_bh(left) != path_root_bh(right));
1709 * The caller didn't pass two adjacent paths.
1711 mlog_bug_on_msg(i > left->p_tree_depth,
1712 "Inode %lu, left depth %u, right depth %u\n"
1713 "left leaf blk %llu, right leaf blk %llu\n",
1714 inode->i_ino, left->p_tree_depth,
1715 right->p_tree_depth,
1716 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1717 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1718 } while (left->p_node[i].bh->b_blocknr ==
1719 right->p_node[i].bh->b_blocknr);
1724 typedef void (path_insert_t)(void *, struct buffer_head *);
1727 * Traverse a btree path in search of cpos, starting at root_el.
1729 * This code can be called with a cpos larger than the tree, in which
1730 * case it will return the rightmost path.
1732 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1733 struct ocfs2_extent_list *root_el, u32 cpos,
1734 path_insert_t *func, void *data)
1739 struct buffer_head *bh = NULL;
1740 struct ocfs2_extent_block *eb;
1741 struct ocfs2_extent_list *el;
1742 struct ocfs2_extent_rec *rec;
1745 while (el->l_tree_depth) {
1746 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1747 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1748 "Owner %llu has empty extent list at "
1750 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1751 le16_to_cpu(el->l_tree_depth));
1757 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1758 rec = &el->l_recs[i];
1761 * In the case that cpos is off the allocation
1762 * tree, this should just wind up returning the
1765 range = le32_to_cpu(rec->e_cpos) +
1766 ocfs2_rec_clusters(el, rec);
1767 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1771 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1773 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1774 "Owner %llu has bad blkno in extent list "
1775 "at depth %u (index %d)\n",
1776 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1777 le16_to_cpu(el->l_tree_depth), i);
1784 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1790 eb = (struct ocfs2_extent_block *) bh->b_data;
1793 if (le16_to_cpu(el->l_next_free_rec) >
1794 le16_to_cpu(el->l_count)) {
1795 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1796 "Owner %llu has bad count in extent list "
1797 "at block %llu (next free=%u, count=%u)\n",
1798 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1799 (unsigned long long)bh->b_blocknr,
1800 le16_to_cpu(el->l_next_free_rec),
1801 le16_to_cpu(el->l_count));
1812 * Catch any trailing bh that the loop didn't handle.
1820 * Given an initialized path (that is, it has a valid root extent
1821 * list), this function will traverse the btree in search of the path
1822 * which would contain cpos.
1824 * The path traveled is recorded in the path structure.
1826 * Note that this will not do any comparisons on leaf node extent
1827 * records, so it will work fine in the case that we just added a tree
1830 struct find_path_data {
1832 struct ocfs2_path *path;
1834 static void find_path_ins(void *data, struct buffer_head *bh)
1836 struct find_path_data *fp = data;
1839 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1842 static int ocfs2_find_path(struct ocfs2_caching_info *ci,
1843 struct ocfs2_path *path, u32 cpos)
1845 struct find_path_data data;
1849 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1850 find_path_ins, &data);
1853 static void find_leaf_ins(void *data, struct buffer_head *bh)
1855 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1856 struct ocfs2_extent_list *el = &eb->h_list;
1857 struct buffer_head **ret = data;
1859 /* We want to retain only the leaf block. */
1860 if (le16_to_cpu(el->l_tree_depth) == 0) {
1866 * Find the leaf block in the tree which would contain cpos. No
1867 * checking of the actual leaf is done.
1869 * Some paths want to call this instead of allocating a path structure
1870 * and calling ocfs2_find_path().
1872 * This function doesn't handle non btree extent lists.
1874 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1875 struct ocfs2_extent_list *root_el, u32 cpos,
1876 struct buffer_head **leaf_bh)
1879 struct buffer_head *bh = NULL;
1881 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1893 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1895 * Basically, we've moved stuff around at the bottom of the tree and
1896 * we need to fix up the extent records above the changes to reflect
1899 * left_rec: the record on the left.
1900 * left_child_el: is the child list pointed to by left_rec
1901 * right_rec: the record to the right of left_rec
1902 * right_child_el: is the child list pointed to by right_rec
1904 * By definition, this only works on interior nodes.
1906 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1907 struct ocfs2_extent_list *left_child_el,
1908 struct ocfs2_extent_rec *right_rec,
1909 struct ocfs2_extent_list *right_child_el)
1911 u32 left_clusters, right_end;
1914 * Interior nodes never have holes. Their cpos is the cpos of
1915 * the leftmost record in their child list. Their cluster
1916 * count covers the full theoretical range of their child list
1917 * - the range between their cpos and the cpos of the record
1918 * immediately to their right.
1920 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1921 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1922 BUG_ON(right_child_el->l_tree_depth);
1923 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1924 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1926 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1927 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1930 * Calculate the rightmost cluster count boundary before
1931 * moving cpos - we will need to adjust clusters after
1932 * updating e_cpos to keep the same highest cluster count.
1934 right_end = le32_to_cpu(right_rec->e_cpos);
1935 right_end += le32_to_cpu(right_rec->e_int_clusters);
1937 right_rec->e_cpos = left_rec->e_cpos;
1938 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1940 right_end -= le32_to_cpu(right_rec->e_cpos);
1941 right_rec->e_int_clusters = cpu_to_le32(right_end);
1945 * Adjust the adjacent root node records involved in a
1946 * rotation. left_el_blkno is passed in as a key so that we can easily
1947 * find it's index in the root list.
1949 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1950 struct ocfs2_extent_list *left_el,
1951 struct ocfs2_extent_list *right_el,
1956 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1957 le16_to_cpu(left_el->l_tree_depth));
1959 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1960 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1965 * The path walking code should have never returned a root and
1966 * two paths which are not adjacent.
1968 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1970 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1971 &root_el->l_recs[i + 1], right_el);
1975 * We've changed a leaf block (in right_path) and need to reflect that
1976 * change back up the subtree.
1978 * This happens in multiple places:
1979 * - When we've moved an extent record from the left path leaf to the right
1980 * path leaf to make room for an empty extent in the left path leaf.
1981 * - When our insert into the right path leaf is at the leftmost edge
1982 * and requires an update of the path immediately to it's left. This
1983 * can occur at the end of some types of rotation and appending inserts.
1984 * - When we've adjusted the last extent record in the left path leaf and the
1985 * 1st extent record in the right path leaf during cross extent block merge.
1987 static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1988 struct ocfs2_path *left_path,
1989 struct ocfs2_path *right_path,
1993 struct ocfs2_extent_list *el, *left_el, *right_el;
1994 struct ocfs2_extent_rec *left_rec, *right_rec;
1995 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1998 * Update the counts and position values within all the
1999 * interior nodes to reflect the leaf rotation we just did.
2001 * The root node is handled below the loop.
2003 * We begin the loop with right_el and left_el pointing to the
2004 * leaf lists and work our way up.
2006 * NOTE: within this loop, left_el and right_el always refer
2007 * to the *child* lists.
2009 left_el = path_leaf_el(left_path);
2010 right_el = path_leaf_el(right_path);
2011 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2012 mlog(0, "Adjust records at index %u\n", i);
2015 * One nice property of knowing that all of these
2016 * nodes are below the root is that we only deal with
2017 * the leftmost right node record and the rightmost
2020 el = left_path->p_node[i].el;
2021 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2022 left_rec = &el->l_recs[idx];
2024 el = right_path->p_node[i].el;
2025 right_rec = &el->l_recs[0];
2027 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2030 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2034 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2039 * Setup our list pointers now so that the current
2040 * parents become children in the next iteration.
2042 left_el = left_path->p_node[i].el;
2043 right_el = right_path->p_node[i].el;
2047 * At the root node, adjust the two adjacent records which
2048 * begin our path to the leaves.
2051 el = left_path->p_node[subtree_index].el;
2052 left_el = left_path->p_node[subtree_index + 1].el;
2053 right_el = right_path->p_node[subtree_index + 1].el;
2055 ocfs2_adjust_root_records(el, left_el, right_el,
2056 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2058 root_bh = left_path->p_node[subtree_index].bh;
2060 ret = ocfs2_journal_dirty(handle, root_bh);
2065 static int ocfs2_rotate_subtree_right(struct inode *inode,
2067 struct ocfs2_path *left_path,
2068 struct ocfs2_path *right_path,
2072 struct buffer_head *right_leaf_bh;
2073 struct buffer_head *left_leaf_bh = NULL;
2074 struct buffer_head *root_bh;
2075 struct ocfs2_extent_list *right_el, *left_el;
2076 struct ocfs2_extent_rec move_rec;
2078 left_leaf_bh = path_leaf_bh(left_path);
2079 left_el = path_leaf_el(left_path);
2081 if (left_el->l_next_free_rec != left_el->l_count) {
2082 ocfs2_error(inode->i_sb,
2083 "Inode %llu has non-full interior leaf node %llu"
2085 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2086 (unsigned long long)left_leaf_bh->b_blocknr,
2087 le16_to_cpu(left_el->l_next_free_rec));
2092 * This extent block may already have an empty record, so we
2093 * return early if so.
2095 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2098 root_bh = left_path->p_node[subtree_index].bh;
2099 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2101 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
2108 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2109 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2116 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2124 right_leaf_bh = path_leaf_bh(right_path);
2125 right_el = path_leaf_el(right_path);
2127 /* This is a code error, not a disk corruption. */
2128 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2129 "because rightmost leaf block %llu is empty\n",
2130 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2131 (unsigned long long)right_leaf_bh->b_blocknr);
2133 ocfs2_create_empty_extent(right_el);
2135 ret = ocfs2_journal_dirty(handle, right_leaf_bh);
2141 /* Do the copy now. */
2142 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2143 move_rec = left_el->l_recs[i];
2144 right_el->l_recs[0] = move_rec;
2147 * Clear out the record we just copied and shift everything
2148 * over, leaving an empty extent in the left leaf.
2150 * We temporarily subtract from next_free_rec so that the
2151 * shift will lose the tail record (which is now defunct).
2153 le16_add_cpu(&left_el->l_next_free_rec, -1);
2154 ocfs2_shift_records_right(left_el);
2155 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2156 le16_add_cpu(&left_el->l_next_free_rec, 1);
2158 ret = ocfs2_journal_dirty(handle, left_leaf_bh);
2164 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2172 * Given a full path, determine what cpos value would return us a path
2173 * containing the leaf immediately to the left of the current one.
2175 * Will return zero if the path passed in is already the leftmost path.
2177 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2178 struct ocfs2_path *path, u32 *cpos)
2182 struct ocfs2_extent_list *el;
2184 BUG_ON(path->p_tree_depth == 0);
2188 blkno = path_leaf_bh(path)->b_blocknr;
2190 /* Start at the tree node just above the leaf and work our way up. */
2191 i = path->p_tree_depth - 1;
2193 el = path->p_node[i].el;
2196 * Find the extent record just before the one in our
2199 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2200 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2204 * We've determined that the
2205 * path specified is already
2206 * the leftmost one - return a
2212 * The leftmost record points to our
2213 * leaf - we need to travel up the
2219 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2220 *cpos = *cpos + ocfs2_rec_clusters(el,
2221 &el->l_recs[j - 1]);
2228 * If we got here, we never found a valid node where
2229 * the tree indicated one should be.
2232 "Invalid extent tree at extent block %llu\n",
2233 (unsigned long long)blkno);
2238 blkno = path->p_node[i].bh->b_blocknr;
2247 * Extend the transaction by enough credits to complete the rotation,
2248 * and still leave at least the original number of credits allocated
2249 * to this transaction.
2251 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2253 struct ocfs2_path *path)
2255 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2257 if (handle->h_buffer_credits < credits)
2258 return ocfs2_extend_trans(handle, credits);
2264 * Trap the case where we're inserting into the theoretical range past
2265 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2266 * whose cpos is less than ours into the right leaf.
2268 * It's only necessary to look at the rightmost record of the left
2269 * leaf because the logic that calls us should ensure that the
2270 * theoretical ranges in the path components above the leaves are
2273 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2276 struct ocfs2_extent_list *left_el;
2277 struct ocfs2_extent_rec *rec;
2280 left_el = path_leaf_el(left_path);
2281 next_free = le16_to_cpu(left_el->l_next_free_rec);
2282 rec = &left_el->l_recs[next_free - 1];
2284 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2289 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2291 int next_free = le16_to_cpu(el->l_next_free_rec);
2293 struct ocfs2_extent_rec *rec;
2298 rec = &el->l_recs[0];
2299 if (ocfs2_is_empty_extent(rec)) {
2303 rec = &el->l_recs[1];
2306 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2307 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2313 * Rotate all the records in a btree right one record, starting at insert_cpos.
2315 * The path to the rightmost leaf should be passed in.
2317 * The array is assumed to be large enough to hold an entire path (tree depth).
2319 * Upon succesful return from this function:
2321 * - The 'right_path' array will contain a path to the leaf block
2322 * whose range contains e_cpos.
2323 * - That leaf block will have a single empty extent in list index 0.
2324 * - In the case that the rotation requires a post-insert update,
2325 * *ret_left_path will contain a valid path which can be passed to
2326 * ocfs2_insert_path().
2328 static int ocfs2_rotate_tree_right(struct inode *inode,
2330 enum ocfs2_split_type split,
2332 struct ocfs2_path *right_path,
2333 struct ocfs2_path **ret_left_path)
2335 int ret, start, orig_credits = handle->h_buffer_credits;
2337 struct ocfs2_path *left_path = NULL;
2339 *ret_left_path = NULL;
2341 left_path = ocfs2_new_path_from_path(right_path);
2348 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
2354 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
2357 * What we want to do here is:
2359 * 1) Start with the rightmost path.
2361 * 2) Determine a path to the leaf block directly to the left
2364 * 3) Determine the 'subtree root' - the lowest level tree node
2365 * which contains a path to both leaves.
2367 * 4) Rotate the subtree.
2369 * 5) Find the next subtree by considering the left path to be
2370 * the new right path.
2372 * The check at the top of this while loop also accepts
2373 * insert_cpos == cpos because cpos is only a _theoretical_
2374 * value to get us the left path - insert_cpos might very well
2375 * be filling that hole.
2377 * Stop at a cpos of '0' because we either started at the
2378 * leftmost branch (i.e., a tree with one branch and a
2379 * rotation inside of it), or we've gone as far as we can in
2380 * rotating subtrees.
2382 while (cpos && insert_cpos <= cpos) {
2383 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2386 ret = ocfs2_find_path(INODE_CACHE(inode), left_path, cpos);
2392 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2393 path_leaf_bh(right_path),
2394 "Inode %lu: error during insert of %u "
2395 "(left path cpos %u) results in two identical "
2396 "paths ending at %llu\n",
2397 inode->i_ino, insert_cpos, cpos,
2398 (unsigned long long)
2399 path_leaf_bh(left_path)->b_blocknr);
2401 if (split == SPLIT_NONE &&
2402 ocfs2_rotate_requires_path_adjustment(left_path,
2406 * We've rotated the tree as much as we
2407 * should. The rest is up to
2408 * ocfs2_insert_path() to complete, after the
2409 * record insertion. We indicate this
2410 * situation by returning the left path.
2412 * The reason we don't adjust the records here
2413 * before the record insert is that an error
2414 * later might break the rule where a parent
2415 * record e_cpos will reflect the actual
2416 * e_cpos of the 1st nonempty record of the
2419 *ret_left_path = left_path;
2423 start = ocfs2_find_subtree_root(inode, left_path, right_path);
2425 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2427 (unsigned long long) right_path->p_node[start].bh->b_blocknr,
2428 right_path->p_tree_depth);
2430 ret = ocfs2_extend_rotate_transaction(handle, start,
2431 orig_credits, right_path);
2437 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
2444 if (split != SPLIT_NONE &&
2445 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2448 * A rotate moves the rightmost left leaf
2449 * record over to the leftmost right leaf
2450 * slot. If we're doing an extent split
2451 * instead of a real insert, then we have to
2452 * check that the extent to be split wasn't
2453 * just moved over. If it was, then we can
2454 * exit here, passing left_path back -
2455 * ocfs2_split_extent() is smart enough to
2456 * search both leaves.
2458 *ret_left_path = left_path;
2463 * There is no need to re-read the next right path
2464 * as we know that it'll be our current left
2465 * path. Optimize by copying values instead.
2467 ocfs2_mv_path(right_path, left_path);
2469 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
2478 ocfs2_free_path(left_path);
2484 static int ocfs2_update_edge_lengths(struct inode *inode, handle_t *handle,
2485 int subtree_index, struct ocfs2_path *path)
2488 struct ocfs2_extent_rec *rec;
2489 struct ocfs2_extent_list *el;
2490 struct ocfs2_extent_block *eb;
2494 * In normal tree rotation process, we will never touch the
2495 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2496 * doesn't reserve the credits for them either.
2498 * But we do have a special case here which will update the rightmost
2499 * records for all the bh in the path.
2500 * So we have to allocate extra credits and access them.
2502 ret = ocfs2_extend_trans(handle,
2503 handle->h_buffer_credits + subtree_index);
2509 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
2515 /* Path should always be rightmost. */
2516 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2517 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2520 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2521 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2522 rec = &el->l_recs[idx];
2523 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2525 for (i = 0; i < path->p_tree_depth; i++) {
2526 el = path->p_node[i].el;
2527 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2528 rec = &el->l_recs[idx];
2530 rec->e_int_clusters = cpu_to_le32(range);
2531 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2533 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2539 static void ocfs2_unlink_path(handle_t *handle,
2540 struct ocfs2_extent_tree *et,
2541 struct ocfs2_cached_dealloc_ctxt *dealloc,
2542 struct ocfs2_path *path, int unlink_start)
2545 struct ocfs2_extent_block *eb;
2546 struct ocfs2_extent_list *el;
2547 struct buffer_head *bh;
2549 for(i = unlink_start; i < path_num_items(path); i++) {
2550 bh = path->p_node[i].bh;
2552 eb = (struct ocfs2_extent_block *)bh->b_data;
2554 * Not all nodes might have had their final count
2555 * decremented by the caller - handle this here.
2558 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2560 "Inode %llu, attempted to remove extent block "
2561 "%llu with %u records\n",
2562 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2563 (unsigned long long)le64_to_cpu(eb->h_blkno),
2564 le16_to_cpu(el->l_next_free_rec));
2566 ocfs2_journal_dirty(handle, bh);
2567 ocfs2_remove_from_cache(et->et_ci, bh);
2571 el->l_next_free_rec = 0;
2572 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2574 ocfs2_journal_dirty(handle, bh);
2576 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2580 ocfs2_remove_from_cache(et->et_ci, bh);
2584 static void ocfs2_unlink_subtree(handle_t *handle,
2585 struct ocfs2_extent_tree *et,
2586 struct ocfs2_path *left_path,
2587 struct ocfs2_path *right_path,
2589 struct ocfs2_cached_dealloc_ctxt *dealloc)
2592 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2593 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2594 struct ocfs2_extent_list *el;
2595 struct ocfs2_extent_block *eb;
2597 el = path_leaf_el(left_path);
2599 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2601 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2602 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2605 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2607 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2608 le16_add_cpu(&root_el->l_next_free_rec, -1);
2610 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2611 eb->h_next_leaf_blk = 0;
2613 ocfs2_journal_dirty(handle, root_bh);
2614 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2616 ocfs2_unlink_path(handle, et, dealloc, right_path,
2620 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2621 struct ocfs2_path *left_path,
2622 struct ocfs2_path *right_path,
2624 struct ocfs2_cached_dealloc_ctxt *dealloc,
2626 struct ocfs2_extent_tree *et)
2628 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2629 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2630 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2631 struct ocfs2_extent_block *eb;
2635 right_leaf_el = path_leaf_el(right_path);
2636 left_leaf_el = path_leaf_el(left_path);
2637 root_bh = left_path->p_node[subtree_index].bh;
2638 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2640 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2643 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2644 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2646 * It's legal for us to proceed if the right leaf is
2647 * the rightmost one and it has an empty extent. There
2648 * are two cases to handle - whether the leaf will be
2649 * empty after removal or not. If the leaf isn't empty
2650 * then just remove the empty extent up front. The
2651 * next block will handle empty leaves by flagging
2654 * Non rightmost leaves will throw -EAGAIN and the
2655 * caller can manually move the subtree and retry.
2658 if (eb->h_next_leaf_blk != 0ULL)
2661 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2662 ret = ocfs2_journal_access_eb(handle, INODE_CACHE(inode),
2663 path_leaf_bh(right_path),
2664 OCFS2_JOURNAL_ACCESS_WRITE);
2670 ocfs2_remove_empty_extent(right_leaf_el);
2672 right_has_empty = 1;
2675 if (eb->h_next_leaf_blk == 0ULL &&
2676 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2678 * We have to update i_last_eb_blk during the meta
2681 ret = ocfs2_et_root_journal_access(handle, et,
2682 OCFS2_JOURNAL_ACCESS_WRITE);
2688 del_right_subtree = 1;
2692 * Getting here with an empty extent in the right path implies
2693 * that it's the rightmost path and will be deleted.
2695 BUG_ON(right_has_empty && !del_right_subtree);
2697 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
2704 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2705 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2712 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2720 if (!right_has_empty) {
2722 * Only do this if we're moving a real
2723 * record. Otherwise, the action is delayed until
2724 * after removal of the right path in which case we
2725 * can do a simple shift to remove the empty extent.
2727 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2728 memset(&right_leaf_el->l_recs[0], 0,
2729 sizeof(struct ocfs2_extent_rec));
2731 if (eb->h_next_leaf_blk == 0ULL) {
2733 * Move recs over to get rid of empty extent, decrease
2734 * next_free. This is allowed to remove the last
2735 * extent in our leaf (setting l_next_free_rec to
2736 * zero) - the delete code below won't care.
2738 ocfs2_remove_empty_extent(right_leaf_el);
2741 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2744 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2748 if (del_right_subtree) {
2749 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2750 subtree_index, dealloc);
2751 ret = ocfs2_update_edge_lengths(inode, handle, subtree_index,
2758 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2759 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2762 * Removal of the extent in the left leaf was skipped
2763 * above so we could delete the right path
2766 if (right_has_empty)
2767 ocfs2_remove_empty_extent(left_leaf_el);
2769 ret = ocfs2_journal_dirty(handle, et_root_bh);
2775 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2783 * Given a full path, determine what cpos value would return us a path
2784 * containing the leaf immediately to the right of the current one.
2786 * Will return zero if the path passed in is already the rightmost path.
2788 * This looks similar, but is subtly different to
2789 * ocfs2_find_cpos_for_left_leaf().
2791 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2792 struct ocfs2_path *path, u32 *cpos)
2796 struct ocfs2_extent_list *el;
2800 if (path->p_tree_depth == 0)
2803 blkno = path_leaf_bh(path)->b_blocknr;
2805 /* Start at the tree node just above the leaf and work our way up. */
2806 i = path->p_tree_depth - 1;
2810 el = path->p_node[i].el;
2813 * Find the extent record just after the one in our
2816 next_free = le16_to_cpu(el->l_next_free_rec);
2817 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2818 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2819 if (j == (next_free - 1)) {
2822 * We've determined that the
2823 * path specified is already
2824 * the rightmost one - return a
2830 * The rightmost record points to our
2831 * leaf - we need to travel up the
2837 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2843 * If we got here, we never found a valid node where
2844 * the tree indicated one should be.
2847 "Invalid extent tree at extent block %llu\n",
2848 (unsigned long long)blkno);
2853 blkno = path->p_node[i].bh->b_blocknr;
2861 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2863 struct ocfs2_path *path)
2866 struct buffer_head *bh = path_leaf_bh(path);
2867 struct ocfs2_extent_list *el = path_leaf_el(path);
2869 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2872 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), path,
2873 path_num_items(path) - 1);
2879 ocfs2_remove_empty_extent(el);
2881 ret = ocfs2_journal_dirty(handle, bh);
2889 static int __ocfs2_rotate_tree_left(struct inode *inode,
2890 handle_t *handle, int orig_credits,
2891 struct ocfs2_path *path,
2892 struct ocfs2_cached_dealloc_ctxt *dealloc,
2893 struct ocfs2_path **empty_extent_path,
2894 struct ocfs2_extent_tree *et)
2896 int ret, subtree_root, deleted;
2898 struct ocfs2_path *left_path = NULL;
2899 struct ocfs2_path *right_path = NULL;
2901 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2903 *empty_extent_path = NULL;
2905 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2912 left_path = ocfs2_new_path_from_path(path);
2919 ocfs2_cp_path(left_path, path);
2921 right_path = ocfs2_new_path_from_path(path);
2928 while (right_cpos) {
2929 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2935 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2938 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2940 (unsigned long long)
2941 right_path->p_node[subtree_root].bh->b_blocknr,
2942 right_path->p_tree_depth);
2944 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2945 orig_credits, left_path);
2952 * Caller might still want to make changes to the
2953 * tree root, so re-add it to the journal here.
2955 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2962 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2963 right_path, subtree_root,
2964 dealloc, &deleted, et);
2965 if (ret == -EAGAIN) {
2967 * The rotation has to temporarily stop due to
2968 * the right subtree having an empty
2969 * extent. Pass it back to the caller for a
2972 *empty_extent_path = right_path;
2982 * The subtree rotate might have removed records on
2983 * the rightmost edge. If so, then rotation is
2989 ocfs2_mv_path(left_path, right_path);
2991 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
3000 ocfs2_free_path(right_path);
3001 ocfs2_free_path(left_path);
3006 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
3007 struct ocfs2_path *path,
3008 struct ocfs2_cached_dealloc_ctxt *dealloc,
3009 struct ocfs2_extent_tree *et)
3011 int ret, subtree_index;
3013 struct ocfs2_path *left_path = NULL;
3014 struct ocfs2_extent_block *eb;
3015 struct ocfs2_extent_list *el;
3018 ret = ocfs2_et_sanity_check(inode, et);
3022 * There's two ways we handle this depending on
3023 * whether path is the only existing one.
3025 ret = ocfs2_extend_rotate_transaction(handle, 0,
3026 handle->h_buffer_credits,
3033 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3039 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3048 * We have a path to the left of this one - it needs
3051 left_path = ocfs2_new_path_from_path(path);
3058 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3064 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3070 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
3072 ocfs2_unlink_subtree(handle, et, left_path, path,
3073 subtree_index, dealloc);
3074 ret = ocfs2_update_edge_lengths(inode, handle, subtree_index,
3081 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3082 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3085 * 'path' is also the leftmost path which
3086 * means it must be the only one. This gets
3087 * handled differently because we want to
3088 * revert the inode back to having extents
3091 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3093 el = et->et_root_el;
3094 el->l_tree_depth = 0;
3095 el->l_next_free_rec = 0;
3096 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3098 ocfs2_et_set_last_eb_blk(et, 0);
3101 ocfs2_journal_dirty(handle, path_root_bh(path));
3104 ocfs2_free_path(left_path);
3109 * Left rotation of btree records.
3111 * In many ways, this is (unsurprisingly) the opposite of right
3112 * rotation. We start at some non-rightmost path containing an empty
3113 * extent in the leaf block. The code works its way to the rightmost
3114 * path by rotating records to the left in every subtree.
3116 * This is used by any code which reduces the number of extent records
3117 * in a leaf. After removal, an empty record should be placed in the
3118 * leftmost list position.
3120 * This won't handle a length update of the rightmost path records if
3121 * the rightmost tree leaf record is removed so the caller is
3122 * responsible for detecting and correcting that.
3124 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
3125 struct ocfs2_path *path,
3126 struct ocfs2_cached_dealloc_ctxt *dealloc,
3127 struct ocfs2_extent_tree *et)
3129 int ret, orig_credits = handle->h_buffer_credits;
3130 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3131 struct ocfs2_extent_block *eb;
3132 struct ocfs2_extent_list *el;
3134 el = path_leaf_el(path);
3135 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3138 if (path->p_tree_depth == 0) {
3139 rightmost_no_delete:
3141 * Inline extents. This is trivially handled, so do
3144 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
3152 * Handle rightmost branch now. There's several cases:
3153 * 1) simple rotation leaving records in there. That's trivial.
3154 * 2) rotation requiring a branch delete - there's no more
3155 * records left. Two cases of this:
3156 * a) There are branches to the left.
3157 * b) This is also the leftmost (the only) branch.
3159 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3160 * 2a) we need the left branch so that we can update it with the unlink
3161 * 2b) we need to bring the inode back to inline extents.
3164 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3166 if (eb->h_next_leaf_blk == 0) {
3168 * This gets a bit tricky if we're going to delete the
3169 * rightmost path. Get the other cases out of the way
3172 if (le16_to_cpu(el->l_next_free_rec) > 1)
3173 goto rightmost_no_delete;
3175 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3177 ocfs2_error(inode->i_sb,
3178 "Inode %llu has empty extent block at %llu",
3179 (unsigned long long)OCFS2_I(inode)->ip_blkno,
3180 (unsigned long long)le64_to_cpu(eb->h_blkno));
3185 * XXX: The caller can not trust "path" any more after
3186 * this as it will have been deleted. What do we do?
3188 * In theory the rotate-for-merge code will never get
3189 * here because it'll always ask for a rotate in a
3193 ret = ocfs2_remove_rightmost_path(inode, handle, path,
3201 * Now we can loop, remembering the path we get from -EAGAIN
3202 * and restarting from there.
3205 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
3206 dealloc, &restart_path, et);
3207 if (ret && ret != -EAGAIN) {
3212 while (ret == -EAGAIN) {
3213 tmp_path = restart_path;
3214 restart_path = NULL;
3216 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
3219 if (ret && ret != -EAGAIN) {
3224 ocfs2_free_path(tmp_path);
3232 ocfs2_free_path(tmp_path);
3233 ocfs2_free_path(restart_path);
3237 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3240 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3243 if (rec->e_leaf_clusters == 0) {
3245 * We consumed all of the merged-from record. An empty
3246 * extent cannot exist anywhere but the 1st array
3247 * position, so move things over if the merged-from
3248 * record doesn't occupy that position.
3250 * This creates a new empty extent so the caller
3251 * should be smart enough to have removed any existing
3255 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3256 size = index * sizeof(struct ocfs2_extent_rec);
3257 memmove(&el->l_recs[1], &el->l_recs[0], size);
3261 * Always memset - the caller doesn't check whether it
3262 * created an empty extent, so there could be junk in
3265 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3269 static int ocfs2_get_right_path(struct inode *inode,
3270 struct ocfs2_path *left_path,
3271 struct ocfs2_path **ret_right_path)
3275 struct ocfs2_path *right_path = NULL;
3276 struct ocfs2_extent_list *left_el;
3278 *ret_right_path = NULL;
3280 /* This function shouldn't be called for non-trees. */
3281 BUG_ON(left_path->p_tree_depth == 0);
3283 left_el = path_leaf_el(left_path);
3284 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3286 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
3293 /* This function shouldn't be called for the rightmost leaf. */
3294 BUG_ON(right_cpos == 0);
3296 right_path = ocfs2_new_path_from_path(left_path);
3303 ret = ocfs2_find_path(INODE_CACHE(inode), right_path, right_cpos);
3309 *ret_right_path = right_path;
3312 ocfs2_free_path(right_path);
3317 * Remove split_rec clusters from the record at index and merge them
3318 * onto the beginning of the record "next" to it.
3319 * For index < l_count - 1, the next means the extent rec at index + 1.
3320 * For index == l_count - 1, the "next" means the 1st extent rec of the
3321 * next extent block.
3323 static int ocfs2_merge_rec_right(struct inode *inode,
3324 struct ocfs2_path *left_path,
3326 struct ocfs2_extent_rec *split_rec,
3329 int ret, next_free, i;
3330 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3331 struct ocfs2_extent_rec *left_rec;
3332 struct ocfs2_extent_rec *right_rec;
3333 struct ocfs2_extent_list *right_el;
3334 struct ocfs2_path *right_path = NULL;
3335 int subtree_index = 0;
3336 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3337 struct buffer_head *bh = path_leaf_bh(left_path);
3338 struct buffer_head *root_bh = NULL;
3340 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3341 left_rec = &el->l_recs[index];
3343 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3344 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3345 /* we meet with a cross extent block merge. */
3346 ret = ocfs2_get_right_path(inode, left_path, &right_path);
3352 right_el = path_leaf_el(right_path);
3353 next_free = le16_to_cpu(right_el->l_next_free_rec);
3354 BUG_ON(next_free <= 0);
3355 right_rec = &right_el->l_recs[0];
3356 if (ocfs2_is_empty_extent(right_rec)) {
3357 BUG_ON(next_free <= 1);
3358 right_rec = &right_el->l_recs[1];
3361 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3362 le16_to_cpu(left_rec->e_leaf_clusters) !=
3363 le32_to_cpu(right_rec->e_cpos));
3365 subtree_index = ocfs2_find_subtree_root(inode,
3366 left_path, right_path);
3368 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3369 handle->h_buffer_credits,
3376 root_bh = left_path->p_node[subtree_index].bh;
3377 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3379 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3386 for (i = subtree_index + 1;
3387 i < path_num_items(right_path); i++) {
3388 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3395 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3404 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3405 right_rec = &el->l_recs[index + 1];
3408 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), left_path,
3409 path_num_items(left_path) - 1);
3415 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3417 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3418 le64_add_cpu(&right_rec->e_blkno,
3419 -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3420 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3422 ocfs2_cleanup_merge(el, index);
3424 ret = ocfs2_journal_dirty(handle, bh);
3429 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3433 ocfs2_complete_edge_insert(inode, handle, left_path,
3434 right_path, subtree_index);
3438 ocfs2_free_path(right_path);
3442 static int ocfs2_get_left_path(struct inode *inode,
3443 struct ocfs2_path *right_path,
3444 struct ocfs2_path **ret_left_path)
3448 struct ocfs2_path *left_path = NULL;
3450 *ret_left_path = NULL;
3452 /* This function shouldn't be called for non-trees. */
3453 BUG_ON(right_path->p_tree_depth == 0);
3455 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
3456 right_path, &left_cpos);
3462 /* This function shouldn't be called for the leftmost leaf. */
3463 BUG_ON(left_cpos == 0);
3465 left_path = ocfs2_new_path_from_path(right_path);
3472 ret = ocfs2_find_path(INODE_CACHE(inode), left_path, left_cpos);
3478 *ret_left_path = left_path;
3481 ocfs2_free_path(left_path);
3486 * Remove split_rec clusters from the record at index and merge them
3487 * onto the tail of the record "before" it.
3488 * For index > 0, the "before" means the extent rec at index - 1.
3490 * For index == 0, the "before" means the last record of the previous
3491 * extent block. And there is also a situation that we may need to
3492 * remove the rightmost leaf extent block in the right_path and change
3493 * the right path to indicate the new rightmost path.
3495 static int ocfs2_merge_rec_left(struct inode *inode,
3496 struct ocfs2_path *right_path,
3498 struct ocfs2_extent_rec *split_rec,
3499 struct ocfs2_cached_dealloc_ctxt *dealloc,
3500 struct ocfs2_extent_tree *et,
3503 int ret, i, subtree_index = 0, has_empty_extent = 0;
3504 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3505 struct ocfs2_extent_rec *left_rec;
3506 struct ocfs2_extent_rec *right_rec;
3507 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3508 struct buffer_head *bh = path_leaf_bh(right_path);
3509 struct buffer_head *root_bh = NULL;
3510 struct ocfs2_path *left_path = NULL;
3511 struct ocfs2_extent_list *left_el;
3515 right_rec = &el->l_recs[index];
3517 /* we meet with a cross extent block merge. */
3518 ret = ocfs2_get_left_path(inode, right_path, &left_path);
3524 left_el = path_leaf_el(left_path);
3525 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3526 le16_to_cpu(left_el->l_count));
3528 left_rec = &left_el->l_recs[
3529 le16_to_cpu(left_el->l_next_free_rec) - 1];
3530 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3531 le16_to_cpu(left_rec->e_leaf_clusters) !=
3532 le32_to_cpu(split_rec->e_cpos));
3534 subtree_index = ocfs2_find_subtree_root(inode,
3535 left_path, right_path);
3537 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3538 handle->h_buffer_credits,
3545 root_bh = left_path->p_node[subtree_index].bh;
3546 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3548 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3555 for (i = subtree_index + 1;
3556 i < path_num_items(right_path); i++) {
3557 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3564 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3572 left_rec = &el->l_recs[index - 1];
3573 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3574 has_empty_extent = 1;
3577 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3578 path_num_items(right_path) - 1);
3584 if (has_empty_extent && index == 1) {
3586 * The easy case - we can just plop the record right in.
3588 *left_rec = *split_rec;
3590 has_empty_extent = 0;
3592 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3594 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3595 le64_add_cpu(&right_rec->e_blkno,
3596 ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3597 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3599 ocfs2_cleanup_merge(el, index);
3601 ret = ocfs2_journal_dirty(handle, bh);
3606 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3611 * In the situation that the right_rec is empty and the extent
3612 * block is empty also, ocfs2_complete_edge_insert can't handle
3613 * it and we need to delete the right extent block.
3615 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3616 le16_to_cpu(el->l_next_free_rec) == 1) {
3618 ret = ocfs2_remove_rightmost_path(inode, handle,
3626 /* Now the rightmost extent block has been deleted.
3627 * So we use the new rightmost path.
3629 ocfs2_mv_path(right_path, left_path);
3632 ocfs2_complete_edge_insert(inode, handle, left_path,
3633 right_path, subtree_index);
3637 ocfs2_free_path(left_path);
3641 static int ocfs2_try_to_merge_extent(struct inode *inode,
3643 struct ocfs2_path *path,
3645 struct ocfs2_extent_rec *split_rec,
3646 struct ocfs2_cached_dealloc_ctxt *dealloc,
3647 struct ocfs2_merge_ctxt *ctxt,
3648 struct ocfs2_extent_tree *et)
3652 struct ocfs2_extent_list *el = path_leaf_el(path);
3653 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3655 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3657 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3659 * The merge code will need to create an empty
3660 * extent to take the place of the newly
3661 * emptied slot. Remove any pre-existing empty
3662 * extents - having more than one in a leaf is
3665 ret = ocfs2_rotate_tree_left(inode, handle, path,
3672 rec = &el->l_recs[split_index];
3675 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3677 * Left-right contig implies this.
3679 BUG_ON(!ctxt->c_split_covers_rec);
3682 * Since the leftright insert always covers the entire
3683 * extent, this call will delete the insert record
3684 * entirely, resulting in an empty extent record added to
3687 * Since the adding of an empty extent shifts
3688 * everything back to the right, there's no need to
3689 * update split_index here.
3691 * When the split_index is zero, we need to merge it to the
3692 * prevoius extent block. It is more efficient and easier
3693 * if we do merge_right first and merge_left later.
3695 ret = ocfs2_merge_rec_right(inode, path,
3704 * We can only get this from logic error above.
3706 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3708 /* The merge left us with an empty extent, remove it. */
3709 ret = ocfs2_rotate_tree_left(inode, handle, path,
3716 rec = &el->l_recs[split_index];
3719 * Note that we don't pass split_rec here on purpose -
3720 * we've merged it into the rec already.
3722 ret = ocfs2_merge_rec_left(inode, path,
3732 ret = ocfs2_rotate_tree_left(inode, handle, path,
3735 * Error from this last rotate is not critical, so
3736 * print but don't bubble it up.
3743 * Merge a record to the left or right.
3745 * 'contig_type' is relative to the existing record,
3746 * so for example, if we're "right contig", it's to
3747 * the record on the left (hence the left merge).
3749 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3750 ret = ocfs2_merge_rec_left(inode,
3760 ret = ocfs2_merge_rec_right(inode,
3770 if (ctxt->c_split_covers_rec) {
3772 * The merge may have left an empty extent in
3773 * our leaf. Try to rotate it away.
3775 ret = ocfs2_rotate_tree_left(inode, handle, path,
3787 static void ocfs2_subtract_from_rec(struct super_block *sb,
3788 enum ocfs2_split_type split,
3789 struct ocfs2_extent_rec *rec,
3790 struct ocfs2_extent_rec *split_rec)
3794 len_blocks = ocfs2_clusters_to_blocks(sb,
3795 le16_to_cpu(split_rec->e_leaf_clusters));
3797 if (split == SPLIT_LEFT) {
3799 * Region is on the left edge of the existing
3802 le32_add_cpu(&rec->e_cpos,
3803 le16_to_cpu(split_rec->e_leaf_clusters));
3804 le64_add_cpu(&rec->e_blkno, len_blocks);
3805 le16_add_cpu(&rec->e_leaf_clusters,
3806 -le16_to_cpu(split_rec->e_leaf_clusters));
3809 * Region is on the right edge of the existing
3812 le16_add_cpu(&rec->e_leaf_clusters,
3813 -le16_to_cpu(split_rec->e_leaf_clusters));
3818 * Do the final bits of extent record insertion at the target leaf
3819 * list. If this leaf is part of an allocation tree, it is assumed
3820 * that the tree above has been prepared.
3822 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
3823 struct ocfs2_extent_list *el,
3824 struct ocfs2_insert_type *insert,
3825 struct inode *inode)
3827 int i = insert->ins_contig_index;
3829 struct ocfs2_extent_rec *rec;
3831 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3833 if (insert->ins_split != SPLIT_NONE) {
3834 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3836 rec = &el->l_recs[i];
3837 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
3843 * Contiguous insert - either left or right.
3845 if (insert->ins_contig != CONTIG_NONE) {
3846 rec = &el->l_recs[i];
3847 if (insert->ins_contig == CONTIG_LEFT) {
3848 rec->e_blkno = insert_rec->e_blkno;
3849 rec->e_cpos = insert_rec->e_cpos;
3851 le16_add_cpu(&rec->e_leaf_clusters,
3852 le16_to_cpu(insert_rec->e_leaf_clusters));
3857 * Handle insert into an empty leaf.
3859 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3860 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3861 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3862 el->l_recs[0] = *insert_rec;
3863 el->l_next_free_rec = cpu_to_le16(1);
3870 if (insert->ins_appending == APPEND_TAIL) {
3871 i = le16_to_cpu(el->l_next_free_rec) - 1;
3872 rec = &el->l_recs[i];
3873 range = le32_to_cpu(rec->e_cpos)
3874 + le16_to_cpu(rec->e_leaf_clusters);
3875 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3877 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3878 le16_to_cpu(el->l_count),
3879 "inode %lu, depth %u, count %u, next free %u, "
3880 "rec.cpos %u, rec.clusters %u, "
3881 "insert.cpos %u, insert.clusters %u\n",
3883 le16_to_cpu(el->l_tree_depth),
3884 le16_to_cpu(el->l_count),
3885 le16_to_cpu(el->l_next_free_rec),
3886 le32_to_cpu(el->l_recs[i].e_cpos),
3887 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3888 le32_to_cpu(insert_rec->e_cpos),
3889 le16_to_cpu(insert_rec->e_leaf_clusters));
3891 el->l_recs[i] = *insert_rec;
3892 le16_add_cpu(&el->l_next_free_rec, 1);
3898 * Ok, we have to rotate.
3900 * At this point, it is safe to assume that inserting into an
3901 * empty leaf and appending to a leaf have both been handled
3904 * This leaf needs to have space, either by the empty 1st
3905 * extent record, or by virtue of an l_next_rec < l_count.
3907 ocfs2_rotate_leaf(el, insert_rec);
3910 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3912 struct ocfs2_path *path,
3913 struct ocfs2_extent_rec *insert_rec)
3915 int ret, i, next_free;
3916 struct buffer_head *bh;
3917 struct ocfs2_extent_list *el;
3918 struct ocfs2_extent_rec *rec;
3921 * Update everything except the leaf block.
3923 for (i = 0; i < path->p_tree_depth; i++) {
3924 bh = path->p_node[i].bh;
3925 el = path->p_node[i].el;
3927 next_free = le16_to_cpu(el->l_next_free_rec);
3928 if (next_free == 0) {
3929 ocfs2_error(inode->i_sb,
3930 "Dinode %llu has a bad extent list",
3931 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3936 rec = &el->l_recs[next_free - 1];
3938 rec->e_int_clusters = insert_rec->e_cpos;
3939 le32_add_cpu(&rec->e_int_clusters,
3940 le16_to_cpu(insert_rec->e_leaf_clusters));
3941 le32_add_cpu(&rec->e_int_clusters,
3942 -le32_to_cpu(rec->e_cpos));
3944 ret = ocfs2_journal_dirty(handle, bh);
3951 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3952 struct ocfs2_extent_rec *insert_rec,
3953 struct ocfs2_path *right_path,
3954 struct ocfs2_path **ret_left_path)
3957 struct ocfs2_extent_list *el;
3958 struct ocfs2_path *left_path = NULL;
3960 *ret_left_path = NULL;
3963 * This shouldn't happen for non-trees. The extent rec cluster
3964 * count manipulation below only works for interior nodes.
3966 BUG_ON(right_path->p_tree_depth == 0);
3969 * If our appending insert is at the leftmost edge of a leaf,
3970 * then we might need to update the rightmost records of the
3973 el = path_leaf_el(right_path);
3974 next_free = le16_to_cpu(el->l_next_free_rec);
3975 if (next_free == 0 ||
3976 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3979 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3986 mlog(0, "Append may need a left path update. cpos: %u, "
3987 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3991 * No need to worry if the append is already in the
3995 left_path = ocfs2_new_path_from_path(right_path);
4002 ret = ocfs2_find_path(INODE_CACHE(inode), left_path,
4010 * ocfs2_insert_path() will pass the left_path to the
4016 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, right_path);
4022 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
4024 *ret_left_path = left_path;
4028 ocfs2_free_path(left_path);
4033 static void ocfs2_split_record(struct inode *inode,
4034 struct ocfs2_path *left_path,
4035 struct ocfs2_path *right_path,
4036 struct ocfs2_extent_rec *split_rec,
4037 enum ocfs2_split_type split)
4040 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4041 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4042 struct ocfs2_extent_rec *rec, *tmprec;
4044 right_el = path_leaf_el(right_path);
4046 left_el = path_leaf_el(left_path);
4049 insert_el = right_el;
4050 index = ocfs2_search_extent_list(el, cpos);
4052 if (index == 0 && left_path) {
4053 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4056 * This typically means that the record
4057 * started in the left path but moved to the
4058 * right as a result of rotation. We either
4059 * move the existing record to the left, or we
4060 * do the later insert there.
4062 * In this case, the left path should always
4063 * exist as the rotate code will have passed
4064 * it back for a post-insert update.
4067 if (split == SPLIT_LEFT) {
4069 * It's a left split. Since we know
4070 * that the rotate code gave us an
4071 * empty extent in the left path, we
4072 * can just do the insert there.
4074 insert_el = left_el;
4077 * Right split - we have to move the
4078 * existing record over to the left
4079 * leaf. The insert will be into the
4080 * newly created empty extent in the
4083 tmprec = &right_el->l_recs[index];
4084 ocfs2_rotate_leaf(left_el, tmprec);
4087 memset(tmprec, 0, sizeof(*tmprec));
4088 index = ocfs2_search_extent_list(left_el, cpos);
4089 BUG_ON(index == -1);
4094 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4096 * Left path is easy - we can just allow the insert to
4100 insert_el = left_el;
4101 index = ocfs2_search_extent_list(el, cpos);
4102 BUG_ON(index == -1);
4105 rec = &el->l_recs[index];
4106 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
4107 ocfs2_rotate_leaf(insert_el, split_rec);
4111 * This function only does inserts on an allocation b-tree. For tree
4112 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4114 * right_path is the path we want to do the actual insert
4115 * in. left_path should only be passed in if we need to update that
4116 * portion of the tree after an edge insert.
4118 static int ocfs2_insert_path(struct inode *inode,
4120 struct ocfs2_path *left_path,
4121 struct ocfs2_path *right_path,
4122 struct ocfs2_extent_rec *insert_rec,
4123 struct ocfs2_insert_type *insert)
4125 int ret, subtree_index;
4126 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4129 int credits = handle->h_buffer_credits;
4132 * There's a chance that left_path got passed back to
4133 * us without being accounted for in the
4134 * journal. Extend our transaction here to be sure we
4135 * can change those blocks.
4137 credits += left_path->p_tree_depth;
4139 ret = ocfs2_extend_trans(handle, credits);
4145 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, left_path);
4153 * Pass both paths to the journal. The majority of inserts
4154 * will be touching all components anyway.
4156 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, right_path);
4162 if (insert->ins_split != SPLIT_NONE) {
4164 * We could call ocfs2_insert_at_leaf() for some types
4165 * of splits, but it's easier to just let one separate
4166 * function sort it all out.
4168 ocfs2_split_record(inode, left_path, right_path,
4169 insert_rec, insert->ins_split);
4172 * Split might have modified either leaf and we don't
4173 * have a guarantee that the later edge insert will
4174 * dirty this for us.
4177 ret = ocfs2_journal_dirty(handle,
4178 path_leaf_bh(left_path));
4182 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
4185 ret = ocfs2_journal_dirty(handle, leaf_bh);
4191 * The rotate code has indicated that we need to fix
4192 * up portions of the tree after the insert.
4194 * XXX: Should we extend the transaction here?
4196 subtree_index = ocfs2_find_subtree_root(inode, left_path,
4198 ocfs2_complete_edge_insert(inode, handle, left_path,
4199 right_path, subtree_index);
4207 static int ocfs2_do_insert_extent(struct inode *inode,
4209 struct ocfs2_extent_tree *et,
4210 struct ocfs2_extent_rec *insert_rec,
4211 struct ocfs2_insert_type *type)
4213 int ret, rotate = 0;
4215 struct ocfs2_path *right_path = NULL;
4216 struct ocfs2_path *left_path = NULL;
4217 struct ocfs2_extent_list *el;
4219 el = et->et_root_el;
4221 ret = ocfs2_et_root_journal_access(handle, et,
4222 OCFS2_JOURNAL_ACCESS_WRITE);
4228 if (le16_to_cpu(el->l_tree_depth) == 0) {
4229 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
4230 goto out_update_clusters;
4233 right_path = ocfs2_new_path_from_et(et);
4241 * Determine the path to start with. Rotations need the
4242 * rightmost path, everything else can go directly to the
4245 cpos = le32_to_cpu(insert_rec->e_cpos);
4246 if (type->ins_appending == APPEND_NONE &&
4247 type->ins_contig == CONTIG_NONE) {
4252 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4259 * Rotations and appends need special treatment - they modify
4260 * parts of the tree's above them.
4262 * Both might pass back a path immediate to the left of the
4263 * one being inserted to. This will be cause
4264 * ocfs2_insert_path() to modify the rightmost records of
4265 * left_path to account for an edge insert.
4267 * XXX: When modifying this code, keep in mind that an insert
4268 * can wind up skipping both of these two special cases...
4271 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
4272 le32_to_cpu(insert_rec->e_cpos),
4273 right_path, &left_path);
4280 * ocfs2_rotate_tree_right() might have extended the
4281 * transaction without re-journaling our tree root.
4283 ret = ocfs2_et_root_journal_access(handle, et,
4284 OCFS2_JOURNAL_ACCESS_WRITE);
4289 } else if (type->ins_appending == APPEND_TAIL
4290 && type->ins_contig != CONTIG_LEFT) {
4291 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
4292 right_path, &left_path);
4299 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
4306 out_update_clusters:
4307 if (type->ins_split == SPLIT_NONE)
4308 ocfs2_et_update_clusters(inode, et,
4309 le16_to_cpu(insert_rec->e_leaf_clusters));
4311 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
4316 ocfs2_free_path(left_path);
4317 ocfs2_free_path(right_path);
4322 static enum ocfs2_contig_type
4323 ocfs2_figure_merge_contig_type(struct inode *inode, struct ocfs2_path *path,
4324 struct ocfs2_extent_list *el, int index,
4325 struct ocfs2_extent_rec *split_rec)
4328 enum ocfs2_contig_type ret = CONTIG_NONE;
4329 u32 left_cpos, right_cpos;
4330 struct ocfs2_extent_rec *rec = NULL;
4331 struct ocfs2_extent_list *new_el;
4332 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4333 struct buffer_head *bh;
4334 struct ocfs2_extent_block *eb;
4337 rec = &el->l_recs[index - 1];
4338 } else if (path->p_tree_depth > 0) {
4339 status = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
4344 if (left_cpos != 0) {
4345 left_path = ocfs2_new_path_from_path(path);
4349 status = ocfs2_find_path(INODE_CACHE(inode),
4350 left_path, left_cpos);
4354 new_el = path_leaf_el(left_path);
4356 if (le16_to_cpu(new_el->l_next_free_rec) !=
4357 le16_to_cpu(new_el->l_count)) {
4358 bh = path_leaf_bh(left_path);
4359 eb = (struct ocfs2_extent_block *)bh->b_data;
4360 ocfs2_error(inode->i_sb,
4361 "Extent block #%llu has an "
4362 "invalid l_next_free_rec of "
4363 "%d. It should have "
4364 "matched the l_count of %d",
4365 (unsigned long long)le64_to_cpu(eb->h_blkno),
4366 le16_to_cpu(new_el->l_next_free_rec),
4367 le16_to_cpu(new_el->l_count));
4371 rec = &new_el->l_recs[
4372 le16_to_cpu(new_el->l_next_free_rec) - 1];
4377 * We're careful to check for an empty extent record here -
4378 * the merge code will know what to do if it sees one.
4381 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4382 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4385 ret = ocfs2_extent_contig(inode, rec, split_rec);
4390 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4391 rec = &el->l_recs[index + 1];
4392 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4393 path->p_tree_depth > 0) {
4394 status = ocfs2_find_cpos_for_right_leaf(inode->i_sb,
4399 if (right_cpos == 0)
4402 right_path = ocfs2_new_path_from_path(path);
4406 status = ocfs2_find_path(INODE_CACHE(inode), right_path, right_cpos);
4410 new_el = path_leaf_el(right_path);
4411 rec = &new_el->l_recs[0];
4412 if (ocfs2_is_empty_extent(rec)) {
4413 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4414 bh = path_leaf_bh(right_path);
4415 eb = (struct ocfs2_extent_block *)bh->b_data;
4416 ocfs2_error(inode->i_sb,
4417 "Extent block #%llu has an "
4418 "invalid l_next_free_rec of %d",
4419 (unsigned long long)le64_to_cpu(eb->h_blkno),
4420 le16_to_cpu(new_el->l_next_free_rec));
4424 rec = &new_el->l_recs[1];
4429 enum ocfs2_contig_type contig_type;
4431 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
4433 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4434 ret = CONTIG_LEFTRIGHT;
4435 else if (ret == CONTIG_NONE)
4441 ocfs2_free_path(left_path);
4443 ocfs2_free_path(right_path);
4448 static void ocfs2_figure_contig_type(struct inode *inode,
4449 struct ocfs2_insert_type *insert,
4450 struct ocfs2_extent_list *el,
4451 struct ocfs2_extent_rec *insert_rec,
4452 struct ocfs2_extent_tree *et)
4455 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4457 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4459 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4460 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
4462 if (contig_type != CONTIG_NONE) {
4463 insert->ins_contig_index = i;
4467 insert->ins_contig = contig_type;
4469 if (insert->ins_contig != CONTIG_NONE) {
4470 struct ocfs2_extent_rec *rec =
4471 &el->l_recs[insert->ins_contig_index];
4472 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4473 le16_to_cpu(insert_rec->e_leaf_clusters);
4476 * Caller might want us to limit the size of extents, don't
4477 * calculate contiguousness if we might exceed that limit.
4479 if (et->et_max_leaf_clusters &&
4480 (len > et->et_max_leaf_clusters))
4481 insert->ins_contig = CONTIG_NONE;
4486 * This should only be called against the righmost leaf extent list.
4488 * ocfs2_figure_appending_type() will figure out whether we'll have to
4489 * insert at the tail of the rightmost leaf.
4491 * This should also work against the root extent list for tree's with 0
4492 * depth. If we consider the root extent list to be the rightmost leaf node
4493 * then the logic here makes sense.
4495 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4496 struct ocfs2_extent_list *el,
4497 struct ocfs2_extent_rec *insert_rec)
4500 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4501 struct ocfs2_extent_rec *rec;
4503 insert->ins_appending = APPEND_NONE;
4505 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4507 if (!el->l_next_free_rec)
4508 goto set_tail_append;
4510 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4511 /* Were all records empty? */
4512 if (le16_to_cpu(el->l_next_free_rec) == 1)
4513 goto set_tail_append;
4516 i = le16_to_cpu(el->l_next_free_rec) - 1;
4517 rec = &el->l_recs[i];
4520 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4521 goto set_tail_append;
4526 insert->ins_appending = APPEND_TAIL;
4530 * Helper function called at the begining of an insert.
4532 * This computes a few things that are commonly used in the process of
4533 * inserting into the btree:
4534 * - Whether the new extent is contiguous with an existing one.
4535 * - The current tree depth.
4536 * - Whether the insert is an appending one.
4537 * - The total # of free records in the tree.
4539 * All of the information is stored on the ocfs2_insert_type
4542 static int ocfs2_figure_insert_type(struct inode *inode,
4543 struct ocfs2_extent_tree *et,
4544 struct buffer_head **last_eb_bh,
4545 struct ocfs2_extent_rec *insert_rec,
4547 struct ocfs2_insert_type *insert)
4550 struct ocfs2_extent_block *eb;
4551 struct ocfs2_extent_list *el;
4552 struct ocfs2_path *path = NULL;
4553 struct buffer_head *bh = NULL;
4555 insert->ins_split = SPLIT_NONE;
4557 el = et->et_root_el;
4558 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4560 if (el->l_tree_depth) {
4562 * If we have tree depth, we read in the
4563 * rightmost extent block ahead of time as
4564 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4565 * may want it later.
4567 ret = ocfs2_read_extent_block(et->et_ci,
4568 ocfs2_et_get_last_eb_blk(et),
4574 eb = (struct ocfs2_extent_block *) bh->b_data;
4579 * Unless we have a contiguous insert, we'll need to know if
4580 * there is room left in our allocation tree for another
4583 * XXX: This test is simplistic, we can search for empty
4584 * extent records too.
4586 *free_records = le16_to_cpu(el->l_count) -
4587 le16_to_cpu(el->l_next_free_rec);
4589 if (!insert->ins_tree_depth) {
4590 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4591 ocfs2_figure_appending_type(insert, el, insert_rec);
4595 path = ocfs2_new_path_from_et(et);
4603 * In the case that we're inserting past what the tree
4604 * currently accounts for, ocfs2_find_path() will return for
4605 * us the rightmost tree path. This is accounted for below in
4606 * the appending code.
4608 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4614 el = path_leaf_el(path);
4617 * Now that we have the path, there's two things we want to determine:
4618 * 1) Contiguousness (also set contig_index if this is so)
4620 * 2) Are we doing an append? We can trivially break this up
4621 * into two types of appends: simple record append, or a
4622 * rotate inside the tail leaf.
4624 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4627 * The insert code isn't quite ready to deal with all cases of
4628 * left contiguousness. Specifically, if it's an insert into
4629 * the 1st record in a leaf, it will require the adjustment of
4630 * cluster count on the last record of the path directly to it's
4631 * left. For now, just catch that case and fool the layers
4632 * above us. This works just fine for tree_depth == 0, which
4633 * is why we allow that above.
4635 if (insert->ins_contig == CONTIG_LEFT &&
4636 insert->ins_contig_index == 0)
4637 insert->ins_contig = CONTIG_NONE;
4640 * Ok, so we can simply compare against last_eb to figure out
4641 * whether the path doesn't exist. This will only happen in
4642 * the case that we're doing a tail append, so maybe we can
4643 * take advantage of that information somehow.
4645 if (ocfs2_et_get_last_eb_blk(et) ==
4646 path_leaf_bh(path)->b_blocknr) {
4648 * Ok, ocfs2_find_path() returned us the rightmost
4649 * tree path. This might be an appending insert. There are
4651 * 1) We're doing a true append at the tail:
4652 * -This might even be off the end of the leaf
4653 * 2) We're "appending" by rotating in the tail
4655 ocfs2_figure_appending_type(insert, el, insert_rec);
4659 ocfs2_free_path(path);
4669 * Insert an extent into an inode btree.
4671 * The caller needs to update fe->i_clusters
4673 int ocfs2_insert_extent(struct ocfs2_super *osb,
4675 struct inode *inode,
4676 struct ocfs2_extent_tree *et,
4681 struct ocfs2_alloc_context *meta_ac)
4684 int uninitialized_var(free_records);
4685 struct buffer_head *last_eb_bh = NULL;
4686 struct ocfs2_insert_type insert = {0, };
4687 struct ocfs2_extent_rec rec;
4689 mlog(0, "add %u clusters at position %u to inode %llu\n",
4690 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4692 memset(&rec, 0, sizeof(rec));
4693 rec.e_cpos = cpu_to_le32(cpos);
4694 rec.e_blkno = cpu_to_le64(start_blk);
4695 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4696 rec.e_flags = flags;
4697 status = ocfs2_et_insert_check(inode, et, &rec);
4703 status = ocfs2_figure_insert_type(inode, et, &last_eb_bh, &rec,
4704 &free_records, &insert);
4710 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4711 "Insert.contig_index: %d, Insert.free_records: %d, "
4712 "Insert.tree_depth: %d\n",
4713 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
4714 free_records, insert.ins_tree_depth);
4716 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4717 status = ocfs2_grow_tree(inode, handle, et,
4718 &insert.ins_tree_depth, &last_eb_bh,
4726 /* Finally, we can add clusters. This might rotate the tree for us. */
4727 status = ocfs2_do_insert_extent(inode, handle, et, &rec, &insert);
4730 else if (et->et_ops == &ocfs2_dinode_et_ops)
4731 ocfs2_extent_map_insert_rec(inode, &rec);
4741 * Allcate and add clusters into the extent b-tree.
4742 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4743 * The extent b-tree's root is specified by et, and
4744 * it is not limited to the file storage. Any extent tree can use this
4745 * function if it implements the proper ocfs2_extent_tree.
4747 int ocfs2_add_clusters_in_btree(struct ocfs2_super *osb,
4748 struct inode *inode,
4749 u32 *logical_offset,
4750 u32 clusters_to_add,
4752 struct ocfs2_extent_tree *et,
4754 struct ocfs2_alloc_context *data_ac,
4755 struct ocfs2_alloc_context *meta_ac,
4756 enum ocfs2_alloc_restarted *reason_ret)
4760 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4761 u32 bit_off, num_bits;
4765 BUG_ON(!clusters_to_add);
4768 flags = OCFS2_EXT_UNWRITTEN;
4770 free_extents = ocfs2_num_free_extents(osb, et);
4771 if (free_extents < 0) {
4772 status = free_extents;
4777 /* there are two cases which could cause us to EAGAIN in the
4778 * we-need-more-metadata case:
4779 * 1) we haven't reserved *any*
4780 * 2) we are so fragmented, we've needed to add metadata too
4782 if (!free_extents && !meta_ac) {
4783 mlog(0, "we haven't reserved any metadata!\n");
4785 reason = RESTART_META;
4787 } else if ((!free_extents)
4788 && (ocfs2_alloc_context_bits_left(meta_ac)
4789 < ocfs2_extend_meta_needed(et->et_root_el))) {
4790 mlog(0, "filesystem is really fragmented...\n");
4792 reason = RESTART_META;
4796 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
4797 clusters_to_add, &bit_off, &num_bits);
4799 if (status != -ENOSPC)
4804 BUG_ON(num_bits > clusters_to_add);
4806 /* reserve our write early -- insert_extent may update the tree root */
4807 status = ocfs2_et_root_journal_access(handle, et,
4808 OCFS2_JOURNAL_ACCESS_WRITE);
4814 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4815 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
4816 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4817 status = ocfs2_insert_extent(osb, handle, inode, et,
4818 *logical_offset, block,
4819 num_bits, flags, meta_ac);
4825 status = ocfs2_journal_dirty(handle, et->et_root_bh);
4831 clusters_to_add -= num_bits;
4832 *logical_offset += num_bits;
4834 if (clusters_to_add) {
4835 mlog(0, "need to alloc once more, wanted = %u\n",
4838 reason = RESTART_TRANS;
4844 *reason_ret = reason;
4848 static void ocfs2_make_right_split_rec(struct super_block *sb,
4849 struct ocfs2_extent_rec *split_rec,
4851 struct ocfs2_extent_rec *rec)
4853 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4854 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4856 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4858 split_rec->e_cpos = cpu_to_le32(cpos);
4859 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4861 split_rec->e_blkno = rec->e_blkno;
4862 le64_add_cpu(&split_rec->e_blkno,
4863 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4865 split_rec->e_flags = rec->e_flags;
4868 static int ocfs2_split_and_insert(struct inode *inode,
4870 struct ocfs2_path *path,
4871 struct ocfs2_extent_tree *et,
4872 struct buffer_head **last_eb_bh,
4874 struct ocfs2_extent_rec *orig_split_rec,
4875 struct ocfs2_alloc_context *meta_ac)
4878 unsigned int insert_range, rec_range, do_leftright = 0;
4879 struct ocfs2_extent_rec tmprec;
4880 struct ocfs2_extent_list *rightmost_el;
4881 struct ocfs2_extent_rec rec;
4882 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4883 struct ocfs2_insert_type insert;
4884 struct ocfs2_extent_block *eb;
4888 * Store a copy of the record on the stack - it might move
4889 * around as the tree is manipulated below.
4891 rec = path_leaf_el(path)->l_recs[split_index];
4893 rightmost_el = et->et_root_el;
4895 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4897 BUG_ON(!(*last_eb_bh));
4898 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4899 rightmost_el = &eb->h_list;
4902 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4903 le16_to_cpu(rightmost_el->l_count)) {
4904 ret = ocfs2_grow_tree(inode, handle, et,
4905 &depth, last_eb_bh, meta_ac);
4912 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4913 insert.ins_appending = APPEND_NONE;
4914 insert.ins_contig = CONTIG_NONE;
4915 insert.ins_tree_depth = depth;
4917 insert_range = le32_to_cpu(split_rec.e_cpos) +
4918 le16_to_cpu(split_rec.e_leaf_clusters);
4919 rec_range = le32_to_cpu(rec.e_cpos) +
4920 le16_to_cpu(rec.e_leaf_clusters);
4922 if (split_rec.e_cpos == rec.e_cpos) {
4923 insert.ins_split = SPLIT_LEFT;
4924 } else if (insert_range == rec_range) {
4925 insert.ins_split = SPLIT_RIGHT;
4928 * Left/right split. We fake this as a right split
4929 * first and then make a second pass as a left split.
4931 insert.ins_split = SPLIT_RIGHT;
4933 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
4938 BUG_ON(do_leftright);
4942 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4948 if (do_leftright == 1) {
4950 struct ocfs2_extent_list *el;
4953 split_rec = *orig_split_rec;
4955 ocfs2_reinit_path(path, 1);
4957 cpos = le32_to_cpu(split_rec.e_cpos);
4958 ret = ocfs2_find_path(et->et_ci, path, cpos);
4964 el = path_leaf_el(path);
4965 split_index = ocfs2_search_extent_list(el, cpos);
4973 static int ocfs2_replace_extent_rec(struct inode *inode,
4975 struct ocfs2_path *path,
4976 struct ocfs2_extent_list *el,
4978 struct ocfs2_extent_rec *split_rec)
4982 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), path,
4983 path_num_items(path) - 1);
4989 el->l_recs[split_index] = *split_rec;
4991 ocfs2_journal_dirty(handle, path_leaf_bh(path));
4997 * Mark part or all of the extent record at split_index in the leaf
4998 * pointed to by path as written. This removes the unwritten
5001 * Care is taken to handle contiguousness so as to not grow the tree.
5003 * meta_ac is not strictly necessary - we only truly need it if growth
5004 * of the tree is required. All other cases will degrade into a less
5005 * optimal tree layout.
5007 * last_eb_bh should be the rightmost leaf block for any extent
5008 * btree. Since a split may grow the tree or a merge might shrink it,
5009 * the caller cannot trust the contents of that buffer after this call.
5011 * This code is optimized for readability - several passes might be
5012 * made over certain portions of the tree. All of those blocks will
5013 * have been brought into cache (and pinned via the journal), so the
5014 * extra overhead is not expressed in terms of disk reads.
5016 static int __ocfs2_mark_extent_written(struct inode *inode,
5017 struct ocfs2_extent_tree *et,
5019 struct ocfs2_path *path,
5021 struct ocfs2_extent_rec *split_rec,
5022 struct ocfs2_alloc_context *meta_ac,
5023 struct ocfs2_cached_dealloc_ctxt *dealloc)
5026 struct ocfs2_extent_list *el = path_leaf_el(path);
5027 struct buffer_head *last_eb_bh = NULL;
5028 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5029 struct ocfs2_merge_ctxt ctxt;
5030 struct ocfs2_extent_list *rightmost_el;
5032 if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) {
5038 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5039 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5040 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5046 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, path, el,
5051 * The core merge / split code wants to know how much room is
5052 * left in this inodes allocation tree, so we pass the
5053 * rightmost extent list.
5055 if (path->p_tree_depth) {
5056 struct ocfs2_extent_block *eb;
5058 ret = ocfs2_read_extent_block(et->et_ci,
5059 ocfs2_et_get_last_eb_blk(et),
5066 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5067 rightmost_el = &eb->h_list;
5069 rightmost_el = path_root_el(path);
5071 if (rec->e_cpos == split_rec->e_cpos &&
5072 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5073 ctxt.c_split_covers_rec = 1;
5075 ctxt.c_split_covers_rec = 0;
5077 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5079 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
5080 split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent,
5081 ctxt.c_split_covers_rec);
5083 if (ctxt.c_contig_type == CONTIG_NONE) {
5084 if (ctxt.c_split_covers_rec)
5085 ret = ocfs2_replace_extent_rec(inode, handle,
5087 split_index, split_rec);
5089 ret = ocfs2_split_and_insert(inode, handle, path, et,
5090 &last_eb_bh, split_index,
5091 split_rec, meta_ac);
5095 ret = ocfs2_try_to_merge_extent(inode, handle, path,
5096 split_index, split_rec,
5097 dealloc, &ctxt, et);
5108 * Mark the already-existing extent at cpos as written for len clusters.
5110 * If the existing extent is larger than the request, initiate a
5111 * split. An attempt will be made at merging with adjacent extents.
5113 * The caller is responsible for passing down meta_ac if we'll need it.
5115 int ocfs2_mark_extent_written(struct inode *inode,
5116 struct ocfs2_extent_tree *et,
5117 handle_t *handle, u32 cpos, u32 len, u32 phys,
5118 struct ocfs2_alloc_context *meta_ac,
5119 struct ocfs2_cached_dealloc_ctxt *dealloc)
5122 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
5123 struct ocfs2_extent_rec split_rec;
5124 struct ocfs2_path *left_path = NULL;
5125 struct ocfs2_extent_list *el;
5127 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
5128 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
5130 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5131 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
5132 "that are being written to, but the feature bit "
5133 "is not set in the super block.",
5134 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5140 * XXX: This should be fixed up so that we just re-insert the
5141 * next extent records.
5143 * XXX: This is a hack on the extent tree, maybe it should be
5146 if (et->et_ops == &ocfs2_dinode_et_ops)
5147 ocfs2_extent_map_trunc(inode, 0);
5149 left_path = ocfs2_new_path_from_et(et);
5156 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5161 el = path_leaf_el(left_path);
5163 index = ocfs2_search_extent_list(el, cpos);
5164 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5165 ocfs2_error(inode->i_sb,
5166 "Inode %llu has an extent at cpos %u which can no "
5167 "longer be found.\n",
5168 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5173 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5174 split_rec.e_cpos = cpu_to_le32(cpos);
5175 split_rec.e_leaf_clusters = cpu_to_le16(len);
5176 split_rec.e_blkno = cpu_to_le64(start_blkno);
5177 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
5178 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
5180 ret = __ocfs2_mark_extent_written(inode, et, handle, left_path,
5181 index, &split_rec, meta_ac,
5187 ocfs2_free_path(left_path);
5191 static int ocfs2_split_tree(struct inode *inode, struct ocfs2_extent_tree *et,
5192 handle_t *handle, struct ocfs2_path *path,
5193 int index, u32 new_range,
5194 struct ocfs2_alloc_context *meta_ac)
5196 int ret, depth, credits = handle->h_buffer_credits;
5197 struct buffer_head *last_eb_bh = NULL;
5198 struct ocfs2_extent_block *eb;
5199 struct ocfs2_extent_list *rightmost_el, *el;
5200 struct ocfs2_extent_rec split_rec;
5201 struct ocfs2_extent_rec *rec;
5202 struct ocfs2_insert_type insert;
5205 * Setup the record to split before we grow the tree.
5207 el = path_leaf_el(path);
5208 rec = &el->l_recs[index];
5209 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
5211 depth = path->p_tree_depth;
5213 ret = ocfs2_read_extent_block(et->et_ci,
5214 ocfs2_et_get_last_eb_blk(et),
5221 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5222 rightmost_el = &eb->h_list;
5224 rightmost_el = path_leaf_el(path);
5226 credits += path->p_tree_depth +
5227 ocfs2_extend_meta_needed(et->et_root_el);
5228 ret = ocfs2_extend_trans(handle, credits);
5234 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5235 le16_to_cpu(rightmost_el->l_count)) {
5236 ret = ocfs2_grow_tree(inode, handle, et, &depth, &last_eb_bh,
5244 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5245 insert.ins_appending = APPEND_NONE;
5246 insert.ins_contig = CONTIG_NONE;
5247 insert.ins_split = SPLIT_RIGHT;
5248 insert.ins_tree_depth = depth;
5250 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
5259 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
5260 struct ocfs2_path *path, int index,
5261 struct ocfs2_cached_dealloc_ctxt *dealloc,
5263 struct ocfs2_extent_tree *et)
5266 u32 left_cpos, rec_range, trunc_range;
5267 int wants_rotate = 0, is_rightmost_tree_rec = 0;
5268 struct super_block *sb = inode->i_sb;
5269 struct ocfs2_path *left_path = NULL;
5270 struct ocfs2_extent_list *el = path_leaf_el(path);
5271 struct ocfs2_extent_rec *rec;
5272 struct ocfs2_extent_block *eb;
5274 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5275 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5284 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5285 path->p_tree_depth) {
5287 * Check whether this is the rightmost tree record. If
5288 * we remove all of this record or part of its right
5289 * edge then an update of the record lengths above it
5292 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5293 if (eb->h_next_leaf_blk == 0)
5294 is_rightmost_tree_rec = 1;
5297 rec = &el->l_recs[index];
5298 if (index == 0 && path->p_tree_depth &&
5299 le32_to_cpu(rec->e_cpos) == cpos) {
5301 * Changing the leftmost offset (via partial or whole
5302 * record truncate) of an interior (or rightmost) path
5303 * means we have to update the subtree that is formed
5304 * by this leaf and the one to it's left.
5306 * There are two cases we can skip:
5307 * 1) Path is the leftmost one in our inode tree.
5308 * 2) The leaf is rightmost and will be empty after
5309 * we remove the extent record - the rotate code
5310 * knows how to update the newly formed edge.
5313 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
5320 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5321 left_path = ocfs2_new_path_from_path(path);
5328 ret = ocfs2_find_path(et->et_ci, left_path,
5337 ret = ocfs2_extend_rotate_transaction(handle, 0,
5338 handle->h_buffer_credits,
5345 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5351 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5357 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5358 trunc_range = cpos + len;
5360 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5363 memset(rec, 0, sizeof(*rec));
5364 ocfs2_cleanup_merge(el, index);
5367 next_free = le16_to_cpu(el->l_next_free_rec);
5368 if (is_rightmost_tree_rec && next_free > 1) {
5370 * We skip the edge update if this path will
5371 * be deleted by the rotate code.
5373 rec = &el->l_recs[next_free - 1];
5374 ocfs2_adjust_rightmost_records(inode, handle, path,
5377 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5378 /* Remove leftmost portion of the record. */
5379 le32_add_cpu(&rec->e_cpos, len);
5380 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5381 le16_add_cpu(&rec->e_leaf_clusters, -len);
5382 } else if (rec_range == trunc_range) {
5383 /* Remove rightmost portion of the record */
5384 le16_add_cpu(&rec->e_leaf_clusters, -len);
5385 if (is_rightmost_tree_rec)
5386 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
5388 /* Caller should have trapped this. */
5389 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
5390 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
5391 le32_to_cpu(rec->e_cpos),
5392 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5399 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
5400 ocfs2_complete_edge_insert(inode, handle, left_path, path,
5404 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5406 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5413 ocfs2_free_path(left_path);
5417 int ocfs2_remove_extent(struct inode *inode,
5418 struct ocfs2_extent_tree *et,
5419 u32 cpos, u32 len, handle_t *handle,
5420 struct ocfs2_alloc_context *meta_ac,
5421 struct ocfs2_cached_dealloc_ctxt *dealloc)
5424 u32 rec_range, trunc_range;
5425 struct ocfs2_extent_rec *rec;
5426 struct ocfs2_extent_list *el;
5427 struct ocfs2_path *path = NULL;
5429 ocfs2_extent_map_trunc(inode, 0);
5431 path = ocfs2_new_path_from_et(et);
5438 ret = ocfs2_find_path(et->et_ci, path, cpos);
5444 el = path_leaf_el(path);
5445 index = ocfs2_search_extent_list(el, cpos);
5446 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5447 ocfs2_error(inode->i_sb,
5448 "Inode %llu has an extent at cpos %u which can no "
5449 "longer be found.\n",
5450 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5456 * We have 3 cases of extent removal:
5457 * 1) Range covers the entire extent rec
5458 * 2) Range begins or ends on one edge of the extent rec
5459 * 3) Range is in the middle of the extent rec (no shared edges)
5461 * For case 1 we remove the extent rec and left rotate to
5464 * For case 2 we just shrink the existing extent rec, with a
5465 * tree update if the shrinking edge is also the edge of an
5468 * For case 3 we do a right split to turn the extent rec into
5469 * something case 2 can handle.
5471 rec = &el->l_recs[index];
5472 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5473 trunc_range = cpos + len;
5475 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5477 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
5478 "(cpos %u, len %u)\n",
5479 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
5480 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
5482 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5483 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5490 ret = ocfs2_split_tree(inode, et, handle, path, index,
5491 trunc_range, meta_ac);
5498 * The split could have manipulated the tree enough to
5499 * move the record location, so we have to look for it again.
5501 ocfs2_reinit_path(path, 1);
5503 ret = ocfs2_find_path(et->et_ci, path, cpos);
5509 el = path_leaf_el(path);
5510 index = ocfs2_search_extent_list(el, cpos);
5511 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5512 ocfs2_error(inode->i_sb,
5513 "Inode %llu: split at cpos %u lost record.",
5514 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5521 * Double check our values here. If anything is fishy,
5522 * it's easier to catch it at the top level.
5524 rec = &el->l_recs[index];
5525 rec_range = le32_to_cpu(rec->e_cpos) +
5526 ocfs2_rec_clusters(el, rec);
5527 if (rec_range != trunc_range) {
5528 ocfs2_error(inode->i_sb,
5529 "Inode %llu: error after split at cpos %u"
5530 "trunc len %u, existing record is (%u,%u)",
5531 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5532 cpos, len, le32_to_cpu(rec->e_cpos),
5533 ocfs2_rec_clusters(el, rec));
5538 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5547 ocfs2_free_path(path);
5551 int ocfs2_remove_btree_range(struct inode *inode,
5552 struct ocfs2_extent_tree *et,
5553 u32 cpos, u32 phys_cpos, u32 len,
5554 struct ocfs2_cached_dealloc_ctxt *dealloc)
5557 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5558 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5559 struct inode *tl_inode = osb->osb_tl_inode;
5561 struct ocfs2_alloc_context *meta_ac = NULL;
5563 ret = ocfs2_lock_allocators(inode, et, 0, 1, NULL, &meta_ac);
5569 mutex_lock(&tl_inode->i_mutex);
5571 if (ocfs2_truncate_log_needs_flush(osb)) {
5572 ret = __ocfs2_flush_truncate_log(osb);
5579 handle = ocfs2_start_trans(osb, ocfs2_remove_extent_credits(osb->sb));
5580 if (IS_ERR(handle)) {
5581 ret = PTR_ERR(handle);
5586 ret = ocfs2_et_root_journal_access(handle, et,
5587 OCFS2_JOURNAL_ACCESS_WRITE);
5593 vfs_dq_free_space_nodirty(inode,
5594 ocfs2_clusters_to_bytes(inode->i_sb, len));
5596 ret = ocfs2_remove_extent(inode, et, cpos, len, handle, meta_ac,
5603 ocfs2_et_update_clusters(inode, et, -len);
5605 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
5611 ret = ocfs2_truncate_log_append(osb, handle, phys_blkno, len);
5616 ocfs2_commit_trans(osb, handle);
5618 mutex_unlock(&tl_inode->i_mutex);
5621 ocfs2_free_alloc_context(meta_ac);
5626 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5628 struct buffer_head *tl_bh = osb->osb_tl_bh;
5629 struct ocfs2_dinode *di;
5630 struct ocfs2_truncate_log *tl;
5632 di = (struct ocfs2_dinode *) tl_bh->b_data;
5633 tl = &di->id2.i_dealloc;
5635 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5636 "slot %d, invalid truncate log parameters: used = "
5637 "%u, count = %u\n", osb->slot_num,
5638 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5639 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5642 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5643 unsigned int new_start)
5645 unsigned int tail_index;
5646 unsigned int current_tail;
5648 /* No records, nothing to coalesce */
5649 if (!le16_to_cpu(tl->tl_used))
5652 tail_index = le16_to_cpu(tl->tl_used) - 1;
5653 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5654 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5656 return current_tail == new_start;
5659 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5662 unsigned int num_clusters)
5665 unsigned int start_cluster, tl_count;
5666 struct inode *tl_inode = osb->osb_tl_inode;
5667 struct buffer_head *tl_bh = osb->osb_tl_bh;
5668 struct ocfs2_dinode *di;
5669 struct ocfs2_truncate_log *tl;
5671 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5672 (unsigned long long)start_blk, num_clusters);
5674 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5676 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5678 di = (struct ocfs2_dinode *) tl_bh->b_data;
5680 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5681 * by the underlying call to ocfs2_read_inode_block(), so any
5682 * corruption is a code bug */
5683 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5685 tl = &di->id2.i_dealloc;
5686 tl_count = le16_to_cpu(tl->tl_count);
5687 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5689 "Truncate record count on #%llu invalid "
5690 "wanted %u, actual %u\n",
5691 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5692 ocfs2_truncate_recs_per_inode(osb->sb),
5693 le16_to_cpu(tl->tl_count));
5695 /* Caller should have known to flush before calling us. */
5696 index = le16_to_cpu(tl->tl_used);
5697 if (index >= tl_count) {
5703 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5704 OCFS2_JOURNAL_ACCESS_WRITE);
5710 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5711 "%llu (index = %d)\n", num_clusters, start_cluster,
5712 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
5714 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5716 * Move index back to the record we are coalescing with.
5717 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5721 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5722 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5723 index, le32_to_cpu(tl->tl_recs[index].t_start),
5726 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5727 tl->tl_used = cpu_to_le16(index + 1);
5729 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5731 status = ocfs2_journal_dirty(handle, tl_bh);
5742 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5744 struct inode *data_alloc_inode,
5745 struct buffer_head *data_alloc_bh)
5749 unsigned int num_clusters;
5751 struct ocfs2_truncate_rec rec;
5752 struct ocfs2_dinode *di;
5753 struct ocfs2_truncate_log *tl;
5754 struct inode *tl_inode = osb->osb_tl_inode;
5755 struct buffer_head *tl_bh = osb->osb_tl_bh;
5759 di = (struct ocfs2_dinode *) tl_bh->b_data;
5760 tl = &di->id2.i_dealloc;
5761 i = le16_to_cpu(tl->tl_used) - 1;
5763 /* Caller has given us at least enough credits to
5764 * update the truncate log dinode */
5765 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5766 OCFS2_JOURNAL_ACCESS_WRITE);
5772 tl->tl_used = cpu_to_le16(i);
5774 status = ocfs2_journal_dirty(handle, tl_bh);
5780 /* TODO: Perhaps we can calculate the bulk of the
5781 * credits up front rather than extending like
5783 status = ocfs2_extend_trans(handle,
5784 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5790 rec = tl->tl_recs[i];
5791 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5792 le32_to_cpu(rec.t_start));
5793 num_clusters = le32_to_cpu(rec.t_clusters);
5795 /* if start_blk is not set, we ignore the record as
5798 mlog(0, "free record %d, start = %u, clusters = %u\n",
5799 i, le32_to_cpu(rec.t_start), num_clusters);
5801 status = ocfs2_free_clusters(handle, data_alloc_inode,
5802 data_alloc_bh, start_blk,
5817 /* Expects you to already be holding tl_inode->i_mutex */
5818 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5821 unsigned int num_to_flush;
5823 struct inode *tl_inode = osb->osb_tl_inode;
5824 struct inode *data_alloc_inode = NULL;
5825 struct buffer_head *tl_bh = osb->osb_tl_bh;
5826 struct buffer_head *data_alloc_bh = NULL;
5827 struct ocfs2_dinode *di;
5828 struct ocfs2_truncate_log *tl;
5832 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5834 di = (struct ocfs2_dinode *) tl_bh->b_data;
5836 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5837 * by the underlying call to ocfs2_read_inode_block(), so any
5838 * corruption is a code bug */
5839 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5841 tl = &di->id2.i_dealloc;
5842 num_to_flush = le16_to_cpu(tl->tl_used);
5843 mlog(0, "Flush %u records from truncate log #%llu\n",
5844 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
5845 if (!num_to_flush) {
5850 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5851 GLOBAL_BITMAP_SYSTEM_INODE,
5852 OCFS2_INVALID_SLOT);
5853 if (!data_alloc_inode) {
5855 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5859 mutex_lock(&data_alloc_inode->i_mutex);
5861 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5867 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5868 if (IS_ERR(handle)) {
5869 status = PTR_ERR(handle);
5874 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
5879 ocfs2_commit_trans(osb, handle);
5882 brelse(data_alloc_bh);
5883 ocfs2_inode_unlock(data_alloc_inode, 1);
5886 mutex_unlock(&data_alloc_inode->i_mutex);
5887 iput(data_alloc_inode);
5894 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5897 struct inode *tl_inode = osb->osb_tl_inode;
5899 mutex_lock(&tl_inode->i_mutex);
5900 status = __ocfs2_flush_truncate_log(osb);
5901 mutex_unlock(&tl_inode->i_mutex);
5906 static void ocfs2_truncate_log_worker(struct work_struct *work)
5909 struct ocfs2_super *osb =
5910 container_of(work, struct ocfs2_super,
5911 osb_truncate_log_wq.work);
5915 status = ocfs2_flush_truncate_log(osb);
5919 ocfs2_init_inode_steal_slot(osb);
5924 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5925 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
5928 if (osb->osb_tl_inode) {
5929 /* We want to push off log flushes while truncates are
5932 cancel_delayed_work(&osb->osb_truncate_log_wq);
5934 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
5935 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
5939 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
5941 struct inode **tl_inode,
5942 struct buffer_head **tl_bh)
5945 struct inode *inode = NULL;
5946 struct buffer_head *bh = NULL;
5948 inode = ocfs2_get_system_file_inode(osb,
5949 TRUNCATE_LOG_SYSTEM_INODE,
5953 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
5957 status = ocfs2_read_inode_block(inode, &bh);
5971 /* called during the 1st stage of node recovery. we stamp a clean
5972 * truncate log and pass back a copy for processing later. if the
5973 * truncate log does not require processing, a *tl_copy is set to
5975 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
5977 struct ocfs2_dinode **tl_copy)
5980 struct inode *tl_inode = NULL;
5981 struct buffer_head *tl_bh = NULL;
5982 struct ocfs2_dinode *di;
5983 struct ocfs2_truncate_log *tl;
5987 mlog(0, "recover truncate log from slot %d\n", slot_num);
5989 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
5995 di = (struct ocfs2_dinode *) tl_bh->b_data;
5997 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
5998 * validated by the underlying call to ocfs2_read_inode_block(),
5999 * so any corruption is a code bug */
6000 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6002 tl = &di->id2.i_dealloc;
6003 if (le16_to_cpu(tl->tl_used)) {
6004 mlog(0, "We'll have %u logs to recover\n",
6005 le16_to_cpu(tl->tl_used));
6007 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6014 /* Assuming the write-out below goes well, this copy
6015 * will be passed back to recovery for processing. */
6016 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6018 /* All we need to do to clear the truncate log is set
6022 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6023 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6035 if (status < 0 && (*tl_copy)) {
6044 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6045 struct ocfs2_dinode *tl_copy)
6049 unsigned int clusters, num_recs, start_cluster;
6052 struct inode *tl_inode = osb->osb_tl_inode;
6053 struct ocfs2_truncate_log *tl;
6057 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6058 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6062 tl = &tl_copy->id2.i_dealloc;
6063 num_recs = le16_to_cpu(tl->tl_used);
6064 mlog(0, "cleanup %u records from %llu\n", num_recs,
6065 (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
6067 mutex_lock(&tl_inode->i_mutex);
6068 for(i = 0; i < num_recs; i++) {
6069 if (ocfs2_truncate_log_needs_flush(osb)) {
6070 status = __ocfs2_flush_truncate_log(osb);
6077 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6078 if (IS_ERR(handle)) {
6079 status = PTR_ERR(handle);
6084 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6085 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6086 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6088 status = ocfs2_truncate_log_append(osb, handle,
6089 start_blk, clusters);
6090 ocfs2_commit_trans(osb, handle);
6098 mutex_unlock(&tl_inode->i_mutex);
6104 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6107 struct inode *tl_inode = osb->osb_tl_inode;
6112 cancel_delayed_work(&osb->osb_truncate_log_wq);
6113 flush_workqueue(ocfs2_wq);
6115 status = ocfs2_flush_truncate_log(osb);
6119 brelse(osb->osb_tl_bh);
6120 iput(osb->osb_tl_inode);
6126 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6129 struct inode *tl_inode = NULL;
6130 struct buffer_head *tl_bh = NULL;
6134 status = ocfs2_get_truncate_log_info(osb,
6141 /* ocfs2_truncate_log_shutdown keys on the existence of
6142 * osb->osb_tl_inode so we don't set any of the osb variables
6143 * until we're sure all is well. */
6144 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6145 ocfs2_truncate_log_worker);
6146 osb->osb_tl_bh = tl_bh;
6147 osb->osb_tl_inode = tl_inode;
6154 * Delayed de-allocation of suballocator blocks.
6156 * Some sets of block de-allocations might involve multiple suballocator inodes.
6158 * The locking for this can get extremely complicated, especially when
6159 * the suballocator inodes to delete from aren't known until deep
6160 * within an unrelated codepath.
6162 * ocfs2_extent_block structures are a good example of this - an inode
6163 * btree could have been grown by any number of nodes each allocating
6164 * out of their own suballoc inode.
6166 * These structures allow the delay of block de-allocation until a
6167 * later time, when locking of multiple cluster inodes won't cause
6172 * Describe a single bit freed from a suballocator. For the block
6173 * suballocators, it represents one block. For the global cluster
6174 * allocator, it represents some clusters and free_bit indicates
6177 struct ocfs2_cached_block_free {
6178 struct ocfs2_cached_block_free *free_next;
6180 unsigned int free_bit;
6183 struct ocfs2_per_slot_free_list {
6184 struct ocfs2_per_slot_free_list *f_next_suballocator;
6187 struct ocfs2_cached_block_free *f_first;
6190 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6193 struct ocfs2_cached_block_free *head)
6198 struct inode *inode;
6199 struct buffer_head *di_bh = NULL;
6200 struct ocfs2_cached_block_free *tmp;
6202 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6209 mutex_lock(&inode->i_mutex);
6211 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6217 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6218 if (IS_ERR(handle)) {
6219 ret = PTR_ERR(handle);
6225 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6227 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
6228 head->free_bit, (unsigned long long)head->free_blk);
6230 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6231 head->free_bit, bg_blkno, 1);
6237 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6244 head = head->free_next;
6249 ocfs2_commit_trans(osb, handle);
6252 ocfs2_inode_unlock(inode, 1);
6255 mutex_unlock(&inode->i_mutex);
6259 /* Premature exit may have left some dangling items. */
6261 head = head->free_next;
6268 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6269 u64 blkno, unsigned int bit)
6272 struct ocfs2_cached_block_free *item;
6274 item = kmalloc(sizeof(*item), GFP_NOFS);
6281 mlog(0, "Insert clusters: (bit %u, blk %llu)\n",
6282 bit, (unsigned long long)blkno);
6284 item->free_blk = blkno;
6285 item->free_bit = bit;
6286 item->free_next = ctxt->c_global_allocator;
6288 ctxt->c_global_allocator = item;
6292 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6293 struct ocfs2_cached_block_free *head)
6295 struct ocfs2_cached_block_free *tmp;
6296 struct inode *tl_inode = osb->osb_tl_inode;
6300 mutex_lock(&tl_inode->i_mutex);
6303 if (ocfs2_truncate_log_needs_flush(osb)) {
6304 ret = __ocfs2_flush_truncate_log(osb);
6311 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6312 if (IS_ERR(handle)) {
6313 ret = PTR_ERR(handle);
6318 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6321 ocfs2_commit_trans(osb, handle);
6323 head = head->free_next;
6332 mutex_unlock(&tl_inode->i_mutex);
6335 /* Premature exit may have left some dangling items. */
6337 head = head->free_next;
6344 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6345 struct ocfs2_cached_dealloc_ctxt *ctxt)
6348 struct ocfs2_per_slot_free_list *fl;
6353 while (ctxt->c_first_suballocator) {
6354 fl = ctxt->c_first_suballocator;
6357 mlog(0, "Free items: (type %u, slot %d)\n",
6358 fl->f_inode_type, fl->f_slot);
6359 ret2 = ocfs2_free_cached_blocks(osb,
6369 ctxt->c_first_suballocator = fl->f_next_suballocator;
6373 if (ctxt->c_global_allocator) {
6374 ret2 = ocfs2_free_cached_clusters(osb,
6375 ctxt->c_global_allocator);
6381 ctxt->c_global_allocator = NULL;
6387 static struct ocfs2_per_slot_free_list *
6388 ocfs2_find_per_slot_free_list(int type,
6390 struct ocfs2_cached_dealloc_ctxt *ctxt)
6392 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6395 if (fl->f_inode_type == type && fl->f_slot == slot)
6398 fl = fl->f_next_suballocator;
6401 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6403 fl->f_inode_type = type;
6406 fl->f_next_suballocator = ctxt->c_first_suballocator;
6408 ctxt->c_first_suballocator = fl;
6413 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6414 int type, int slot, u64 blkno,
6418 struct ocfs2_per_slot_free_list *fl;
6419 struct ocfs2_cached_block_free *item;
6421 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6428 item = kmalloc(sizeof(*item), GFP_NOFS);
6435 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
6436 type, slot, bit, (unsigned long long)blkno);
6438 item->free_blk = blkno;
6439 item->free_bit = bit;
6440 item->free_next = fl->f_first;
6449 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6450 struct ocfs2_extent_block *eb)
6452 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6453 le16_to_cpu(eb->h_suballoc_slot),
6454 le64_to_cpu(eb->h_blkno),
6455 le16_to_cpu(eb->h_suballoc_bit));
6458 /* This function will figure out whether the currently last extent
6459 * block will be deleted, and if it will, what the new last extent
6460 * block will be so we can update his h_next_leaf_blk field, as well
6461 * as the dinodes i_last_eb_blk */
6462 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
6463 unsigned int clusters_to_del,
6464 struct ocfs2_path *path,
6465 struct buffer_head **new_last_eb)
6467 int next_free, ret = 0;
6469 struct ocfs2_extent_rec *rec;
6470 struct ocfs2_extent_block *eb;
6471 struct ocfs2_extent_list *el;
6472 struct buffer_head *bh = NULL;
6474 *new_last_eb = NULL;
6476 /* we have no tree, so of course, no last_eb. */
6477 if (!path->p_tree_depth)
6480 /* trunc to zero special case - this makes tree_depth = 0
6481 * regardless of what it is. */
6482 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
6485 el = path_leaf_el(path);
6486 BUG_ON(!el->l_next_free_rec);
6489 * Make sure that this extent list will actually be empty
6490 * after we clear away the data. We can shortcut out if
6491 * there's more than one non-empty extent in the
6492 * list. Otherwise, a check of the remaining extent is
6495 next_free = le16_to_cpu(el->l_next_free_rec);
6497 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6501 /* We may have a valid extent in index 1, check it. */
6503 rec = &el->l_recs[1];
6506 * Fall through - no more nonempty extents, so we want
6507 * to delete this leaf.
6513 rec = &el->l_recs[0];
6518 * Check it we'll only be trimming off the end of this
6521 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
6525 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
6531 ret = ocfs2_find_leaf(INODE_CACHE(inode), path_root_el(path), cpos, &bh);
6537 eb = (struct ocfs2_extent_block *) bh->b_data;
6540 /* ocfs2_find_leaf() gets the eb from ocfs2_read_extent_block().
6541 * Any corruption is a code bug. */
6542 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
6545 get_bh(*new_last_eb);
6546 mlog(0, "returning block %llu, (cpos: %u)\n",
6547 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
6555 * Trim some clusters off the rightmost edge of a tree. Only called
6558 * The caller needs to:
6559 * - start journaling of each path component.
6560 * - compute and fully set up any new last ext block
6562 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
6563 handle_t *handle, struct ocfs2_truncate_context *tc,
6564 u32 clusters_to_del, u64 *delete_start)
6566 int ret, i, index = path->p_tree_depth;
6569 struct buffer_head *bh;
6570 struct ocfs2_extent_list *el;
6571 struct ocfs2_extent_rec *rec;
6575 while (index >= 0) {
6576 bh = path->p_node[index].bh;
6577 el = path->p_node[index].el;
6579 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6580 index, (unsigned long long)bh->b_blocknr);
6582 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
6585 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
6586 ocfs2_error(inode->i_sb,
6587 "Inode %lu has invalid ext. block %llu",
6589 (unsigned long long)bh->b_blocknr);
6595 i = le16_to_cpu(el->l_next_free_rec) - 1;
6596 rec = &el->l_recs[i];
6598 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6599 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
6600 ocfs2_rec_clusters(el, rec),
6601 (unsigned long long)le64_to_cpu(rec->e_blkno),
6602 le16_to_cpu(el->l_next_free_rec));
6604 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
6606 if (le16_to_cpu(el->l_tree_depth) == 0) {
6608 * If the leaf block contains a single empty
6609 * extent and no records, we can just remove
6612 if (i == 0 && ocfs2_is_empty_extent(rec)) {
6614 sizeof(struct ocfs2_extent_rec));
6615 el->l_next_free_rec = cpu_to_le16(0);
6621 * Remove any empty extents by shifting things
6622 * left. That should make life much easier on
6623 * the code below. This condition is rare
6624 * enough that we shouldn't see a performance
6627 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6628 le16_add_cpu(&el->l_next_free_rec, -1);
6631 i < le16_to_cpu(el->l_next_free_rec); i++)
6632 el->l_recs[i] = el->l_recs[i + 1];
6634 memset(&el->l_recs[i], 0,
6635 sizeof(struct ocfs2_extent_rec));
6638 * We've modified our extent list. The
6639 * simplest way to handle this change
6640 * is to being the search from the
6643 goto find_tail_record;
6646 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
6649 * We'll use "new_edge" on our way back up the
6650 * tree to know what our rightmost cpos is.
6652 new_edge = le16_to_cpu(rec->e_leaf_clusters);
6653 new_edge += le32_to_cpu(rec->e_cpos);
6656 * The caller will use this to delete data blocks.
6658 *delete_start = le64_to_cpu(rec->e_blkno)
6659 + ocfs2_clusters_to_blocks(inode->i_sb,
6660 le16_to_cpu(rec->e_leaf_clusters));
6663 * If it's now empty, remove this record.
6665 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
6667 sizeof(struct ocfs2_extent_rec));
6668 le16_add_cpu(&el->l_next_free_rec, -1);
6671 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
6673 sizeof(struct ocfs2_extent_rec));
6674 le16_add_cpu(&el->l_next_free_rec, -1);
6679 /* Can this actually happen? */
6680 if (le16_to_cpu(el->l_next_free_rec) == 0)
6684 * We never actually deleted any clusters
6685 * because our leaf was empty. There's no
6686 * reason to adjust the rightmost edge then.
6691 rec->e_int_clusters = cpu_to_le32(new_edge);
6692 le32_add_cpu(&rec->e_int_clusters,
6693 -le32_to_cpu(rec->e_cpos));
6696 * A deleted child record should have been
6699 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
6703 ret = ocfs2_journal_dirty(handle, bh);
6709 mlog(0, "extent list container %llu, after: record %d: "
6710 "(%u, %u, %llu), next = %u.\n",
6711 (unsigned long long)bh->b_blocknr, i,
6712 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
6713 (unsigned long long)le64_to_cpu(rec->e_blkno),
6714 le16_to_cpu(el->l_next_free_rec));
6717 * We must be careful to only attempt delete of an
6718 * extent block (and not the root inode block).
6720 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
6721 struct ocfs2_extent_block *eb =
6722 (struct ocfs2_extent_block *)bh->b_data;
6725 * Save this for use when processing the
6728 deleted_eb = le64_to_cpu(eb->h_blkno);
6730 mlog(0, "deleting this extent block.\n");
6732 ocfs2_remove_from_cache(INODE_CACHE(inode), bh);
6734 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
6735 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
6736 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
6738 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
6739 /* An error here is not fatal. */
6754 static int ocfs2_do_truncate(struct ocfs2_super *osb,
6755 unsigned int clusters_to_del,
6756 struct inode *inode,
6757 struct buffer_head *fe_bh,
6759 struct ocfs2_truncate_context *tc,
6760 struct ocfs2_path *path)
6763 struct ocfs2_dinode *fe;
6764 struct ocfs2_extent_block *last_eb = NULL;
6765 struct ocfs2_extent_list *el;
6766 struct buffer_head *last_eb_bh = NULL;
6769 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6771 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
6779 * Each component will be touched, so we might as well journal
6780 * here to avoid having to handle errors later.
6782 status = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
6789 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), last_eb_bh,
6790 OCFS2_JOURNAL_ACCESS_WRITE);
6796 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6799 el = &(fe->id2.i_list);
6802 * Lower levels depend on this never happening, but it's best
6803 * to check it up here before changing the tree.
6805 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
6806 ocfs2_error(inode->i_sb,
6807 "Inode %lu has an empty extent record, depth %u\n",
6808 inode->i_ino, le16_to_cpu(el->l_tree_depth));
6813 vfs_dq_free_space_nodirty(inode,
6814 ocfs2_clusters_to_bytes(osb->sb, clusters_to_del));
6815 spin_lock(&OCFS2_I(inode)->ip_lock);
6816 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
6818 spin_unlock(&OCFS2_I(inode)->ip_lock);
6819 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
6820 inode->i_blocks = ocfs2_inode_sector_count(inode);
6822 status = ocfs2_trim_tree(inode, path, handle, tc,
6823 clusters_to_del, &delete_blk);
6829 if (le32_to_cpu(fe->i_clusters) == 0) {
6830 /* trunc to zero is a special case. */
6831 el->l_tree_depth = 0;
6832 fe->i_last_eb_blk = 0;
6834 fe->i_last_eb_blk = last_eb->h_blkno;
6836 status = ocfs2_journal_dirty(handle, fe_bh);
6843 /* If there will be a new last extent block, then by
6844 * definition, there cannot be any leaves to the right of
6846 last_eb->h_next_leaf_blk = 0;
6847 status = ocfs2_journal_dirty(handle, last_eb_bh);
6855 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
6869 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6871 set_buffer_uptodate(bh);
6872 mark_buffer_dirty(bh);
6876 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6877 unsigned int from, unsigned int to,
6878 struct page *page, int zero, u64 *phys)
6880 int ret, partial = 0;
6882 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6887 zero_user_segment(page, from, to);
6890 * Need to set the buffers we zero'd into uptodate
6891 * here if they aren't - ocfs2_map_page_blocks()
6892 * might've skipped some
6894 ret = walk_page_buffers(handle, page_buffers(page),
6899 else if (ocfs2_should_order_data(inode)) {
6900 ret = ocfs2_jbd2_file_inode(handle, inode);
6906 SetPageUptodate(page);
6908 flush_dcache_page(page);
6911 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6912 loff_t end, struct page **pages,
6913 int numpages, u64 phys, handle_t *handle)
6917 unsigned int from, to = PAGE_CACHE_SIZE;
6918 struct super_block *sb = inode->i_sb;
6920 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6925 to = PAGE_CACHE_SIZE;
6926 for(i = 0; i < numpages; i++) {
6929 from = start & (PAGE_CACHE_SIZE - 1);
6930 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6931 to = end & (PAGE_CACHE_SIZE - 1);
6933 BUG_ON(from > PAGE_CACHE_SIZE);
6934 BUG_ON(to > PAGE_CACHE_SIZE);
6936 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6939 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6943 ocfs2_unlock_and_free_pages(pages, numpages);
6946 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6947 struct page **pages, int *num)
6949 int numpages, ret = 0;
6950 struct super_block *sb = inode->i_sb;
6951 struct address_space *mapping = inode->i_mapping;
6952 unsigned long index;
6953 loff_t last_page_bytes;
6955 BUG_ON(start > end);
6957 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6958 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6961 last_page_bytes = PAGE_ALIGN(end);
6962 index = start >> PAGE_CACHE_SHIFT;
6964 pages[numpages] = grab_cache_page(mapping, index);
6965 if (!pages[numpages]) {
6973 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6978 ocfs2_unlock_and_free_pages(pages, numpages);
6988 * Zero the area past i_size but still within an allocated
6989 * cluster. This avoids exposing nonzero data on subsequent file
6992 * We need to call this before i_size is updated on the inode because
6993 * otherwise block_write_full_page() will skip writeout of pages past
6994 * i_size. The new_i_size parameter is passed for this reason.
6996 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6997 u64 range_start, u64 range_end)
6999 int ret = 0, numpages;
7000 struct page **pages = NULL;
7002 unsigned int ext_flags;
7003 struct super_block *sb = inode->i_sb;
7006 * File systems which don't support sparse files zero on every
7009 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
7012 pages = kcalloc(ocfs2_pages_per_cluster(sb),
7013 sizeof(struct page *), GFP_NOFS);
7014 if (pages == NULL) {
7020 if (range_start == range_end)
7023 ret = ocfs2_extent_map_get_blocks(inode,
7024 range_start >> sb->s_blocksize_bits,
7025 &phys, NULL, &ext_flags);
7032 * Tail is a hole, or is marked unwritten. In either case, we
7033 * can count on read and write to return/push zero's.
7035 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
7038 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
7045 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
7046 numpages, phys, handle);
7049 * Initiate writeout of the pages we zero'd here. We don't
7050 * wait on them - the truncate_inode_pages() call later will
7053 ret = do_sync_mapping_range(inode->i_mapping, range_start,
7054 range_end - 1, SYNC_FILE_RANGE_WRITE);
7065 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7066 struct ocfs2_dinode *di)
7068 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7069 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7071 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7072 memset(&di->id2, 0, blocksize -
7073 offsetof(struct ocfs2_dinode, id2) -
7076 memset(&di->id2, 0, blocksize -
7077 offsetof(struct ocfs2_dinode, id2));
7080 void ocfs2_dinode_new_extent_list(struct inode *inode,
7081 struct ocfs2_dinode *di)
7083 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7084 di->id2.i_list.l_tree_depth = 0;
7085 di->id2.i_list.l_next_free_rec = 0;
7086 di->id2.i_list.l_count = cpu_to_le16(
7087 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7090 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7092 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7093 struct ocfs2_inline_data *idata = &di->id2.i_data;
7095 spin_lock(&oi->ip_lock);
7096 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7097 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7098 spin_unlock(&oi->ip_lock);
7101 * We clear the entire i_data structure here so that all
7102 * fields can be properly initialized.
7104 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7106 idata->id_count = cpu_to_le16(
7107 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7110 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7111 struct buffer_head *di_bh)
7113 int ret, i, has_data, num_pages = 0;
7115 u64 uninitialized_var(block);
7116 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7117 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7118 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7119 struct ocfs2_alloc_context *data_ac = NULL;
7120 struct page **pages = NULL;
7121 loff_t end = osb->s_clustersize;
7122 struct ocfs2_extent_tree et;
7125 has_data = i_size_read(inode) ? 1 : 0;
7128 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
7129 sizeof(struct page *), GFP_NOFS);
7130 if (pages == NULL) {
7136 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7143 handle = ocfs2_start_trans(osb,
7144 ocfs2_inline_to_extents_credits(osb->sb));
7145 if (IS_ERR(handle)) {
7146 ret = PTR_ERR(handle);
7151 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7152 OCFS2_JOURNAL_ACCESS_WRITE);
7160 unsigned int page_end;
7163 if (vfs_dq_alloc_space_nodirty(inode,
7164 ocfs2_clusters_to_bytes(osb->sb, 1))) {
7170 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
7178 * Save two copies, one for insert, and one that can
7179 * be changed by ocfs2_map_and_dirty_page() below.
7181 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7184 * Non sparse file systems zero on extend, so no need
7187 if (!ocfs2_sparse_alloc(osb) &&
7188 PAGE_CACHE_SIZE < osb->s_clustersize)
7189 end = PAGE_CACHE_SIZE;
7191 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
7198 * This should populate the 1st page for us and mark
7201 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
7207 page_end = PAGE_CACHE_SIZE;
7208 if (PAGE_CACHE_SIZE > osb->s_clustersize)
7209 page_end = osb->s_clustersize;
7211 for (i = 0; i < num_pages; i++)
7212 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
7213 pages[i], i > 0, &phys);
7216 spin_lock(&oi->ip_lock);
7217 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7218 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7219 spin_unlock(&oi->ip_lock);
7221 ocfs2_dinode_new_extent_list(inode, di);
7223 ocfs2_journal_dirty(handle, di_bh);
7227 * An error at this point should be extremely rare. If
7228 * this proves to be false, we could always re-build
7229 * the in-inode data from our pages.
7231 ocfs2_init_dinode_extent_tree(&et, inode, di_bh);
7232 ret = ocfs2_insert_extent(osb, handle, inode, &et,
7233 0, block, 1, 0, NULL);
7239 inode->i_blocks = ocfs2_inode_sector_count(inode);
7243 if (ret < 0 && did_quota)
7244 vfs_dq_free_space_nodirty(inode,
7245 ocfs2_clusters_to_bytes(osb->sb, 1));
7247 ocfs2_commit_trans(osb, handle);
7251 ocfs2_free_alloc_context(data_ac);
7255 ocfs2_unlock_and_free_pages(pages, num_pages);
7263 * It is expected, that by the time you call this function,
7264 * inode->i_size and fe->i_size have been adjusted.
7266 * WARNING: This will kfree the truncate context
7268 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7269 struct inode *inode,
7270 struct buffer_head *fe_bh,
7271 struct ocfs2_truncate_context *tc)
7273 int status, i, credits, tl_sem = 0;
7274 u32 clusters_to_del, new_highest_cpos, range;
7275 struct ocfs2_extent_list *el;
7276 handle_t *handle = NULL;
7277 struct inode *tl_inode = osb->osb_tl_inode;
7278 struct ocfs2_path *path = NULL;
7279 struct ocfs2_dinode *di = (struct ocfs2_dinode *)fe_bh->b_data;
7283 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7284 i_size_read(inode));
7286 path = ocfs2_new_path(fe_bh, &di->id2.i_list,
7287 ocfs2_journal_access_di);
7294 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7298 * Check that we still have allocation to delete.
7300 if (OCFS2_I(inode)->ip_clusters == 0) {
7306 * Truncate always works against the rightmost tree branch.
7308 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7314 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
7315 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
7318 * By now, el will point to the extent list on the bottom most
7319 * portion of this tree. Only the tail record is considered in
7322 * We handle the following cases, in order:
7323 * - empty extent: delete the remaining branch
7324 * - remove the entire record
7325 * - remove a partial record
7326 * - no record needs to be removed (truncate has completed)
7328 el = path_leaf_el(path);
7329 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7330 ocfs2_error(inode->i_sb,
7331 "Inode %llu has empty extent block at %llu\n",
7332 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7333 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7338 i = le16_to_cpu(el->l_next_free_rec) - 1;
7339 range = le32_to_cpu(el->l_recs[i].e_cpos) +
7340 ocfs2_rec_clusters(el, &el->l_recs[i]);
7341 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
7342 clusters_to_del = 0;
7343 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
7344 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
7345 } else if (range > new_highest_cpos) {
7346 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
7347 le32_to_cpu(el->l_recs[i].e_cpos)) -
7354 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
7355 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
7357 mutex_lock(&tl_inode->i_mutex);
7359 /* ocfs2_truncate_log_needs_flush guarantees us at least one
7360 * record is free for use. If there isn't any, we flush to get
7361 * an empty truncate log. */
7362 if (ocfs2_truncate_log_needs_flush(osb)) {
7363 status = __ocfs2_flush_truncate_log(osb);
7370 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
7371 (struct ocfs2_dinode *)fe_bh->b_data,
7373 handle = ocfs2_start_trans(osb, credits);
7374 if (IS_ERR(handle)) {
7375 status = PTR_ERR(handle);
7381 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
7388 mutex_unlock(&tl_inode->i_mutex);
7391 ocfs2_commit_trans(osb, handle);
7394 ocfs2_reinit_path(path, 1);
7397 * The check above will catch the case where we've truncated
7398 * away all allocation.
7404 ocfs2_schedule_truncate_log_flush(osb, 1);
7407 mutex_unlock(&tl_inode->i_mutex);
7410 ocfs2_commit_trans(osb, handle);
7412 ocfs2_run_deallocs(osb, &tc->tc_dealloc);
7414 ocfs2_free_path(path);
7416 /* This will drop the ext_alloc cluster lock for us */
7417 ocfs2_free_truncate_context(tc);
7424 * Expects the inode to already be locked.
7426 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
7427 struct inode *inode,
7428 struct buffer_head *fe_bh,
7429 struct ocfs2_truncate_context **tc)
7432 unsigned int new_i_clusters;
7433 struct ocfs2_dinode *fe;
7434 struct ocfs2_extent_block *eb;
7435 struct buffer_head *last_eb_bh = NULL;
7441 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
7442 i_size_read(inode));
7443 fe = (struct ocfs2_dinode *) fe_bh->b_data;
7445 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
7446 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
7447 (unsigned long long)le64_to_cpu(fe->i_size));
7449 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
7455 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
7457 if (fe->id2.i_list.l_tree_depth) {
7458 status = ocfs2_read_extent_block(INODE_CACHE(inode),
7459 le64_to_cpu(fe->i_last_eb_blk),
7465 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
7468 (*tc)->tc_last_eb_bh = last_eb_bh;
7474 ocfs2_free_truncate_context(*tc);
7482 * 'start' is inclusive, 'end' is not.
7484 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7485 unsigned int start, unsigned int end, int trunc)
7488 unsigned int numbytes;
7490 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7491 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7492 struct ocfs2_inline_data *idata = &di->id2.i_data;
7494 if (end > i_size_read(inode))
7495 end = i_size_read(inode);
7497 BUG_ON(start >= end);
7499 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7500 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7501 !ocfs2_supports_inline_data(osb)) {
7502 ocfs2_error(inode->i_sb,
7503 "Inline data flags for inode %llu don't agree! "
7504 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7505 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7506 le16_to_cpu(di->i_dyn_features),
7507 OCFS2_I(inode)->ip_dyn_features,
7508 osb->s_feature_incompat);
7513 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7514 if (IS_ERR(handle)) {
7515 ret = PTR_ERR(handle);
7520 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7521 OCFS2_JOURNAL_ACCESS_WRITE);
7527 numbytes = end - start;
7528 memset(idata->id_data + start, 0, numbytes);
7531 * No need to worry about the data page here - it's been
7532 * truncated already and inline data doesn't need it for
7533 * pushing zero's to disk, so we'll let readpage pick it up
7537 i_size_write(inode, start);
7538 di->i_size = cpu_to_le64(start);
7541 inode->i_blocks = ocfs2_inode_sector_count(inode);
7542 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7544 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7545 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7547 ocfs2_journal_dirty(handle, di_bh);
7550 ocfs2_commit_trans(osb, handle);
7556 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
7559 * The caller is responsible for completing deallocation
7560 * before freeing the context.
7562 if (tc->tc_dealloc.c_first_suballocator != NULL)
7564 "Truncate completion has non-empty dealloc context\n");
7566 brelse(tc->tc_last_eb_bh);