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(struct inode *inode, handle_t *handle,
2540 struct ocfs2_cached_dealloc_ctxt *dealloc,
2541 struct ocfs2_path *path, int unlink_start)
2544 struct ocfs2_extent_block *eb;
2545 struct ocfs2_extent_list *el;
2546 struct buffer_head *bh;
2548 for(i = unlink_start; i < path_num_items(path); i++) {
2549 bh = path->p_node[i].bh;
2551 eb = (struct ocfs2_extent_block *)bh->b_data;
2553 * Not all nodes might have had their final count
2554 * decremented by the caller - handle this here.
2557 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2559 "Inode %llu, attempted to remove extent block "
2560 "%llu with %u records\n",
2561 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2562 (unsigned long long)le64_to_cpu(eb->h_blkno),
2563 le16_to_cpu(el->l_next_free_rec));
2565 ocfs2_journal_dirty(handle, bh);
2566 ocfs2_remove_from_cache(INODE_CACHE(inode), bh);
2570 el->l_next_free_rec = 0;
2571 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2573 ocfs2_journal_dirty(handle, bh);
2575 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2579 ocfs2_remove_from_cache(INODE_CACHE(inode), bh);
2583 static void ocfs2_unlink_subtree(struct inode *inode, handle_t *handle,
2584 struct ocfs2_path *left_path,
2585 struct ocfs2_path *right_path,
2587 struct ocfs2_cached_dealloc_ctxt *dealloc)
2590 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2591 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2592 struct ocfs2_extent_list *el;
2593 struct ocfs2_extent_block *eb;
2595 el = path_leaf_el(left_path);
2597 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2599 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2600 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2603 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2605 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2606 le16_add_cpu(&root_el->l_next_free_rec, -1);
2608 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2609 eb->h_next_leaf_blk = 0;
2611 ocfs2_journal_dirty(handle, root_bh);
2612 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2614 ocfs2_unlink_path(inode, handle, dealloc, right_path,
2618 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2619 struct ocfs2_path *left_path,
2620 struct ocfs2_path *right_path,
2622 struct ocfs2_cached_dealloc_ctxt *dealloc,
2624 struct ocfs2_extent_tree *et)
2626 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2627 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2628 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2629 struct ocfs2_extent_block *eb;
2633 right_leaf_el = path_leaf_el(right_path);
2634 left_leaf_el = path_leaf_el(left_path);
2635 root_bh = left_path->p_node[subtree_index].bh;
2636 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2638 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2641 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2642 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2644 * It's legal for us to proceed if the right leaf is
2645 * the rightmost one and it has an empty extent. There
2646 * are two cases to handle - whether the leaf will be
2647 * empty after removal or not. If the leaf isn't empty
2648 * then just remove the empty extent up front. The
2649 * next block will handle empty leaves by flagging
2652 * Non rightmost leaves will throw -EAGAIN and the
2653 * caller can manually move the subtree and retry.
2656 if (eb->h_next_leaf_blk != 0ULL)
2659 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2660 ret = ocfs2_journal_access_eb(handle, INODE_CACHE(inode),
2661 path_leaf_bh(right_path),
2662 OCFS2_JOURNAL_ACCESS_WRITE);
2668 ocfs2_remove_empty_extent(right_leaf_el);
2670 right_has_empty = 1;
2673 if (eb->h_next_leaf_blk == 0ULL &&
2674 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2676 * We have to update i_last_eb_blk during the meta
2679 ret = ocfs2_et_root_journal_access(handle, et,
2680 OCFS2_JOURNAL_ACCESS_WRITE);
2686 del_right_subtree = 1;
2690 * Getting here with an empty extent in the right path implies
2691 * that it's the rightmost path and will be deleted.
2693 BUG_ON(right_has_empty && !del_right_subtree);
2695 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
2702 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2703 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2710 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2718 if (!right_has_empty) {
2720 * Only do this if we're moving a real
2721 * record. Otherwise, the action is delayed until
2722 * after removal of the right path in which case we
2723 * can do a simple shift to remove the empty extent.
2725 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2726 memset(&right_leaf_el->l_recs[0], 0,
2727 sizeof(struct ocfs2_extent_rec));
2729 if (eb->h_next_leaf_blk == 0ULL) {
2731 * Move recs over to get rid of empty extent, decrease
2732 * next_free. This is allowed to remove the last
2733 * extent in our leaf (setting l_next_free_rec to
2734 * zero) - the delete code below won't care.
2736 ocfs2_remove_empty_extent(right_leaf_el);
2739 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2742 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2746 if (del_right_subtree) {
2747 ocfs2_unlink_subtree(inode, handle, left_path, right_path,
2748 subtree_index, dealloc);
2749 ret = ocfs2_update_edge_lengths(inode, handle, subtree_index,
2756 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2757 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2760 * Removal of the extent in the left leaf was skipped
2761 * above so we could delete the right path
2764 if (right_has_empty)
2765 ocfs2_remove_empty_extent(left_leaf_el);
2767 ret = ocfs2_journal_dirty(handle, et_root_bh);
2773 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2781 * Given a full path, determine what cpos value would return us a path
2782 * containing the leaf immediately to the right of the current one.
2784 * Will return zero if the path passed in is already the rightmost path.
2786 * This looks similar, but is subtly different to
2787 * ocfs2_find_cpos_for_left_leaf().
2789 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2790 struct ocfs2_path *path, u32 *cpos)
2794 struct ocfs2_extent_list *el;
2798 if (path->p_tree_depth == 0)
2801 blkno = path_leaf_bh(path)->b_blocknr;
2803 /* Start at the tree node just above the leaf and work our way up. */
2804 i = path->p_tree_depth - 1;
2808 el = path->p_node[i].el;
2811 * Find the extent record just after the one in our
2814 next_free = le16_to_cpu(el->l_next_free_rec);
2815 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2816 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2817 if (j == (next_free - 1)) {
2820 * We've determined that the
2821 * path specified is already
2822 * the rightmost one - return a
2828 * The rightmost record points to our
2829 * leaf - we need to travel up the
2835 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2841 * If we got here, we never found a valid node where
2842 * the tree indicated one should be.
2845 "Invalid extent tree at extent block %llu\n",
2846 (unsigned long long)blkno);
2851 blkno = path->p_node[i].bh->b_blocknr;
2859 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2861 struct ocfs2_path *path)
2864 struct buffer_head *bh = path_leaf_bh(path);
2865 struct ocfs2_extent_list *el = path_leaf_el(path);
2867 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2870 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), path,
2871 path_num_items(path) - 1);
2877 ocfs2_remove_empty_extent(el);
2879 ret = ocfs2_journal_dirty(handle, bh);
2887 static int __ocfs2_rotate_tree_left(struct inode *inode,
2888 handle_t *handle, int orig_credits,
2889 struct ocfs2_path *path,
2890 struct ocfs2_cached_dealloc_ctxt *dealloc,
2891 struct ocfs2_path **empty_extent_path,
2892 struct ocfs2_extent_tree *et)
2894 int ret, subtree_root, deleted;
2896 struct ocfs2_path *left_path = NULL;
2897 struct ocfs2_path *right_path = NULL;
2899 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2901 *empty_extent_path = NULL;
2903 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2910 left_path = ocfs2_new_path_from_path(path);
2917 ocfs2_cp_path(left_path, path);
2919 right_path = ocfs2_new_path_from_path(path);
2926 while (right_cpos) {
2927 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2933 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2936 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2938 (unsigned long long)
2939 right_path->p_node[subtree_root].bh->b_blocknr,
2940 right_path->p_tree_depth);
2942 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2943 orig_credits, left_path);
2950 * Caller might still want to make changes to the
2951 * tree root, so re-add it to the journal here.
2953 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2960 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2961 right_path, subtree_root,
2962 dealloc, &deleted, et);
2963 if (ret == -EAGAIN) {
2965 * The rotation has to temporarily stop due to
2966 * the right subtree having an empty
2967 * extent. Pass it back to the caller for a
2970 *empty_extent_path = right_path;
2980 * The subtree rotate might have removed records on
2981 * the rightmost edge. If so, then rotation is
2987 ocfs2_mv_path(left_path, right_path);
2989 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2998 ocfs2_free_path(right_path);
2999 ocfs2_free_path(left_path);
3004 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
3005 struct ocfs2_path *path,
3006 struct ocfs2_cached_dealloc_ctxt *dealloc,
3007 struct ocfs2_extent_tree *et)
3009 int ret, subtree_index;
3011 struct ocfs2_path *left_path = NULL;
3012 struct ocfs2_extent_block *eb;
3013 struct ocfs2_extent_list *el;
3016 ret = ocfs2_et_sanity_check(inode, et);
3020 * There's two ways we handle this depending on
3021 * whether path is the only existing one.
3023 ret = ocfs2_extend_rotate_transaction(handle, 0,
3024 handle->h_buffer_credits,
3031 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3037 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3046 * We have a path to the left of this one - it needs
3049 left_path = ocfs2_new_path_from_path(path);
3056 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3062 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3068 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
3070 ocfs2_unlink_subtree(inode, handle, left_path, path,
3071 subtree_index, dealloc);
3072 ret = ocfs2_update_edge_lengths(inode, handle, subtree_index,
3079 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3080 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3083 * 'path' is also the leftmost path which
3084 * means it must be the only one. This gets
3085 * handled differently because we want to
3086 * revert the inode back to having extents
3089 ocfs2_unlink_path(inode, handle, dealloc, path, 1);
3091 el = et->et_root_el;
3092 el->l_tree_depth = 0;
3093 el->l_next_free_rec = 0;
3094 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3096 ocfs2_et_set_last_eb_blk(et, 0);
3099 ocfs2_journal_dirty(handle, path_root_bh(path));
3102 ocfs2_free_path(left_path);
3107 * Left rotation of btree records.
3109 * In many ways, this is (unsurprisingly) the opposite of right
3110 * rotation. We start at some non-rightmost path containing an empty
3111 * extent in the leaf block. The code works its way to the rightmost
3112 * path by rotating records to the left in every subtree.
3114 * This is used by any code which reduces the number of extent records
3115 * in a leaf. After removal, an empty record should be placed in the
3116 * leftmost list position.
3118 * This won't handle a length update of the rightmost path records if
3119 * the rightmost tree leaf record is removed so the caller is
3120 * responsible for detecting and correcting that.
3122 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
3123 struct ocfs2_path *path,
3124 struct ocfs2_cached_dealloc_ctxt *dealloc,
3125 struct ocfs2_extent_tree *et)
3127 int ret, orig_credits = handle->h_buffer_credits;
3128 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3129 struct ocfs2_extent_block *eb;
3130 struct ocfs2_extent_list *el;
3132 el = path_leaf_el(path);
3133 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3136 if (path->p_tree_depth == 0) {
3137 rightmost_no_delete:
3139 * Inline extents. This is trivially handled, so do
3142 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
3150 * Handle rightmost branch now. There's several cases:
3151 * 1) simple rotation leaving records in there. That's trivial.
3152 * 2) rotation requiring a branch delete - there's no more
3153 * records left. Two cases of this:
3154 * a) There are branches to the left.
3155 * b) This is also the leftmost (the only) branch.
3157 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3158 * 2a) we need the left branch so that we can update it with the unlink
3159 * 2b) we need to bring the inode back to inline extents.
3162 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3164 if (eb->h_next_leaf_blk == 0) {
3166 * This gets a bit tricky if we're going to delete the
3167 * rightmost path. Get the other cases out of the way
3170 if (le16_to_cpu(el->l_next_free_rec) > 1)
3171 goto rightmost_no_delete;
3173 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3175 ocfs2_error(inode->i_sb,
3176 "Inode %llu has empty extent block at %llu",
3177 (unsigned long long)OCFS2_I(inode)->ip_blkno,
3178 (unsigned long long)le64_to_cpu(eb->h_blkno));
3183 * XXX: The caller can not trust "path" any more after
3184 * this as it will have been deleted. What do we do?
3186 * In theory the rotate-for-merge code will never get
3187 * here because it'll always ask for a rotate in a
3191 ret = ocfs2_remove_rightmost_path(inode, handle, path,
3199 * Now we can loop, remembering the path we get from -EAGAIN
3200 * and restarting from there.
3203 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
3204 dealloc, &restart_path, et);
3205 if (ret && ret != -EAGAIN) {
3210 while (ret == -EAGAIN) {
3211 tmp_path = restart_path;
3212 restart_path = NULL;
3214 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
3217 if (ret && ret != -EAGAIN) {
3222 ocfs2_free_path(tmp_path);
3230 ocfs2_free_path(tmp_path);
3231 ocfs2_free_path(restart_path);
3235 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3238 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3241 if (rec->e_leaf_clusters == 0) {
3243 * We consumed all of the merged-from record. An empty
3244 * extent cannot exist anywhere but the 1st array
3245 * position, so move things over if the merged-from
3246 * record doesn't occupy that position.
3248 * This creates a new empty extent so the caller
3249 * should be smart enough to have removed any existing
3253 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3254 size = index * sizeof(struct ocfs2_extent_rec);
3255 memmove(&el->l_recs[1], &el->l_recs[0], size);
3259 * Always memset - the caller doesn't check whether it
3260 * created an empty extent, so there could be junk in
3263 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3267 static int ocfs2_get_right_path(struct inode *inode,
3268 struct ocfs2_path *left_path,
3269 struct ocfs2_path **ret_right_path)
3273 struct ocfs2_path *right_path = NULL;
3274 struct ocfs2_extent_list *left_el;
3276 *ret_right_path = NULL;
3278 /* This function shouldn't be called for non-trees. */
3279 BUG_ON(left_path->p_tree_depth == 0);
3281 left_el = path_leaf_el(left_path);
3282 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3284 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
3291 /* This function shouldn't be called for the rightmost leaf. */
3292 BUG_ON(right_cpos == 0);
3294 right_path = ocfs2_new_path_from_path(left_path);
3301 ret = ocfs2_find_path(INODE_CACHE(inode), right_path, right_cpos);
3307 *ret_right_path = right_path;
3310 ocfs2_free_path(right_path);
3315 * Remove split_rec clusters from the record at index and merge them
3316 * onto the beginning of the record "next" to it.
3317 * For index < l_count - 1, the next means the extent rec at index + 1.
3318 * For index == l_count - 1, the "next" means the 1st extent rec of the
3319 * next extent block.
3321 static int ocfs2_merge_rec_right(struct inode *inode,
3322 struct ocfs2_path *left_path,
3324 struct ocfs2_extent_rec *split_rec,
3327 int ret, next_free, i;
3328 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3329 struct ocfs2_extent_rec *left_rec;
3330 struct ocfs2_extent_rec *right_rec;
3331 struct ocfs2_extent_list *right_el;
3332 struct ocfs2_path *right_path = NULL;
3333 int subtree_index = 0;
3334 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3335 struct buffer_head *bh = path_leaf_bh(left_path);
3336 struct buffer_head *root_bh = NULL;
3338 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3339 left_rec = &el->l_recs[index];
3341 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3342 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3343 /* we meet with a cross extent block merge. */
3344 ret = ocfs2_get_right_path(inode, left_path, &right_path);
3350 right_el = path_leaf_el(right_path);
3351 next_free = le16_to_cpu(right_el->l_next_free_rec);
3352 BUG_ON(next_free <= 0);
3353 right_rec = &right_el->l_recs[0];
3354 if (ocfs2_is_empty_extent(right_rec)) {
3355 BUG_ON(next_free <= 1);
3356 right_rec = &right_el->l_recs[1];
3359 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3360 le16_to_cpu(left_rec->e_leaf_clusters) !=
3361 le32_to_cpu(right_rec->e_cpos));
3363 subtree_index = ocfs2_find_subtree_root(inode,
3364 left_path, right_path);
3366 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3367 handle->h_buffer_credits,
3374 root_bh = left_path->p_node[subtree_index].bh;
3375 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3377 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3384 for (i = subtree_index + 1;
3385 i < path_num_items(right_path); i++) {
3386 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3393 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3402 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3403 right_rec = &el->l_recs[index + 1];
3406 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), left_path,
3407 path_num_items(left_path) - 1);
3413 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3415 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3416 le64_add_cpu(&right_rec->e_blkno,
3417 -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3418 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3420 ocfs2_cleanup_merge(el, index);
3422 ret = ocfs2_journal_dirty(handle, bh);
3427 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3431 ocfs2_complete_edge_insert(inode, handle, left_path,
3432 right_path, subtree_index);
3436 ocfs2_free_path(right_path);
3440 static int ocfs2_get_left_path(struct inode *inode,
3441 struct ocfs2_path *right_path,
3442 struct ocfs2_path **ret_left_path)
3446 struct ocfs2_path *left_path = NULL;
3448 *ret_left_path = NULL;
3450 /* This function shouldn't be called for non-trees. */
3451 BUG_ON(right_path->p_tree_depth == 0);
3453 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
3454 right_path, &left_cpos);
3460 /* This function shouldn't be called for the leftmost leaf. */
3461 BUG_ON(left_cpos == 0);
3463 left_path = ocfs2_new_path_from_path(right_path);
3470 ret = ocfs2_find_path(INODE_CACHE(inode), left_path, left_cpos);
3476 *ret_left_path = left_path;
3479 ocfs2_free_path(left_path);
3484 * Remove split_rec clusters from the record at index and merge them
3485 * onto the tail of the record "before" it.
3486 * For index > 0, the "before" means the extent rec at index - 1.
3488 * For index == 0, the "before" means the last record of the previous
3489 * extent block. And there is also a situation that we may need to
3490 * remove the rightmost leaf extent block in the right_path and change
3491 * the right path to indicate the new rightmost path.
3493 static int ocfs2_merge_rec_left(struct inode *inode,
3494 struct ocfs2_path *right_path,
3496 struct ocfs2_extent_rec *split_rec,
3497 struct ocfs2_cached_dealloc_ctxt *dealloc,
3498 struct ocfs2_extent_tree *et,
3501 int ret, i, subtree_index = 0, has_empty_extent = 0;
3502 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3503 struct ocfs2_extent_rec *left_rec;
3504 struct ocfs2_extent_rec *right_rec;
3505 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3506 struct buffer_head *bh = path_leaf_bh(right_path);
3507 struct buffer_head *root_bh = NULL;
3508 struct ocfs2_path *left_path = NULL;
3509 struct ocfs2_extent_list *left_el;
3513 right_rec = &el->l_recs[index];
3515 /* we meet with a cross extent block merge. */
3516 ret = ocfs2_get_left_path(inode, right_path, &left_path);
3522 left_el = path_leaf_el(left_path);
3523 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3524 le16_to_cpu(left_el->l_count));
3526 left_rec = &left_el->l_recs[
3527 le16_to_cpu(left_el->l_next_free_rec) - 1];
3528 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3529 le16_to_cpu(left_rec->e_leaf_clusters) !=
3530 le32_to_cpu(split_rec->e_cpos));
3532 subtree_index = ocfs2_find_subtree_root(inode,
3533 left_path, right_path);
3535 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3536 handle->h_buffer_credits,
3543 root_bh = left_path->p_node[subtree_index].bh;
3544 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3546 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3553 for (i = subtree_index + 1;
3554 i < path_num_items(right_path); i++) {
3555 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3562 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3570 left_rec = &el->l_recs[index - 1];
3571 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3572 has_empty_extent = 1;
3575 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3576 path_num_items(right_path) - 1);
3582 if (has_empty_extent && index == 1) {
3584 * The easy case - we can just plop the record right in.
3586 *left_rec = *split_rec;
3588 has_empty_extent = 0;
3590 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3592 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3593 le64_add_cpu(&right_rec->e_blkno,
3594 ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3595 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3597 ocfs2_cleanup_merge(el, index);
3599 ret = ocfs2_journal_dirty(handle, bh);
3604 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3609 * In the situation that the right_rec is empty and the extent
3610 * block is empty also, ocfs2_complete_edge_insert can't handle
3611 * it and we need to delete the right extent block.
3613 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3614 le16_to_cpu(el->l_next_free_rec) == 1) {
3616 ret = ocfs2_remove_rightmost_path(inode, handle,
3624 /* Now the rightmost extent block has been deleted.
3625 * So we use the new rightmost path.
3627 ocfs2_mv_path(right_path, left_path);
3630 ocfs2_complete_edge_insert(inode, handle, left_path,
3631 right_path, subtree_index);
3635 ocfs2_free_path(left_path);
3639 static int ocfs2_try_to_merge_extent(struct inode *inode,
3641 struct ocfs2_path *path,
3643 struct ocfs2_extent_rec *split_rec,
3644 struct ocfs2_cached_dealloc_ctxt *dealloc,
3645 struct ocfs2_merge_ctxt *ctxt,
3646 struct ocfs2_extent_tree *et)
3650 struct ocfs2_extent_list *el = path_leaf_el(path);
3651 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3653 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3655 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3657 * The merge code will need to create an empty
3658 * extent to take the place of the newly
3659 * emptied slot. Remove any pre-existing empty
3660 * extents - having more than one in a leaf is
3663 ret = ocfs2_rotate_tree_left(inode, handle, path,
3670 rec = &el->l_recs[split_index];
3673 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3675 * Left-right contig implies this.
3677 BUG_ON(!ctxt->c_split_covers_rec);
3680 * Since the leftright insert always covers the entire
3681 * extent, this call will delete the insert record
3682 * entirely, resulting in an empty extent record added to
3685 * Since the adding of an empty extent shifts
3686 * everything back to the right, there's no need to
3687 * update split_index here.
3689 * When the split_index is zero, we need to merge it to the
3690 * prevoius extent block. It is more efficient and easier
3691 * if we do merge_right first and merge_left later.
3693 ret = ocfs2_merge_rec_right(inode, path,
3702 * We can only get this from logic error above.
3704 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3706 /* The merge left us with an empty extent, remove it. */
3707 ret = ocfs2_rotate_tree_left(inode, handle, path,
3714 rec = &el->l_recs[split_index];
3717 * Note that we don't pass split_rec here on purpose -
3718 * we've merged it into the rec already.
3720 ret = ocfs2_merge_rec_left(inode, path,
3730 ret = ocfs2_rotate_tree_left(inode, handle, path,
3733 * Error from this last rotate is not critical, so
3734 * print but don't bubble it up.
3741 * Merge a record to the left or right.
3743 * 'contig_type' is relative to the existing record,
3744 * so for example, if we're "right contig", it's to
3745 * the record on the left (hence the left merge).
3747 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3748 ret = ocfs2_merge_rec_left(inode,
3758 ret = ocfs2_merge_rec_right(inode,
3768 if (ctxt->c_split_covers_rec) {
3770 * The merge may have left an empty extent in
3771 * our leaf. Try to rotate it away.
3773 ret = ocfs2_rotate_tree_left(inode, handle, path,
3785 static void ocfs2_subtract_from_rec(struct super_block *sb,
3786 enum ocfs2_split_type split,
3787 struct ocfs2_extent_rec *rec,
3788 struct ocfs2_extent_rec *split_rec)
3792 len_blocks = ocfs2_clusters_to_blocks(sb,
3793 le16_to_cpu(split_rec->e_leaf_clusters));
3795 if (split == SPLIT_LEFT) {
3797 * Region is on the left edge of the existing
3800 le32_add_cpu(&rec->e_cpos,
3801 le16_to_cpu(split_rec->e_leaf_clusters));
3802 le64_add_cpu(&rec->e_blkno, len_blocks);
3803 le16_add_cpu(&rec->e_leaf_clusters,
3804 -le16_to_cpu(split_rec->e_leaf_clusters));
3807 * Region is on the right edge of the existing
3810 le16_add_cpu(&rec->e_leaf_clusters,
3811 -le16_to_cpu(split_rec->e_leaf_clusters));
3816 * Do the final bits of extent record insertion at the target leaf
3817 * list. If this leaf is part of an allocation tree, it is assumed
3818 * that the tree above has been prepared.
3820 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
3821 struct ocfs2_extent_list *el,
3822 struct ocfs2_insert_type *insert,
3823 struct inode *inode)
3825 int i = insert->ins_contig_index;
3827 struct ocfs2_extent_rec *rec;
3829 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3831 if (insert->ins_split != SPLIT_NONE) {
3832 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3834 rec = &el->l_recs[i];
3835 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
3841 * Contiguous insert - either left or right.
3843 if (insert->ins_contig != CONTIG_NONE) {
3844 rec = &el->l_recs[i];
3845 if (insert->ins_contig == CONTIG_LEFT) {
3846 rec->e_blkno = insert_rec->e_blkno;
3847 rec->e_cpos = insert_rec->e_cpos;
3849 le16_add_cpu(&rec->e_leaf_clusters,
3850 le16_to_cpu(insert_rec->e_leaf_clusters));
3855 * Handle insert into an empty leaf.
3857 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3858 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3859 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3860 el->l_recs[0] = *insert_rec;
3861 el->l_next_free_rec = cpu_to_le16(1);
3868 if (insert->ins_appending == APPEND_TAIL) {
3869 i = le16_to_cpu(el->l_next_free_rec) - 1;
3870 rec = &el->l_recs[i];
3871 range = le32_to_cpu(rec->e_cpos)
3872 + le16_to_cpu(rec->e_leaf_clusters);
3873 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3875 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3876 le16_to_cpu(el->l_count),
3877 "inode %lu, depth %u, count %u, next free %u, "
3878 "rec.cpos %u, rec.clusters %u, "
3879 "insert.cpos %u, insert.clusters %u\n",
3881 le16_to_cpu(el->l_tree_depth),
3882 le16_to_cpu(el->l_count),
3883 le16_to_cpu(el->l_next_free_rec),
3884 le32_to_cpu(el->l_recs[i].e_cpos),
3885 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3886 le32_to_cpu(insert_rec->e_cpos),
3887 le16_to_cpu(insert_rec->e_leaf_clusters));
3889 el->l_recs[i] = *insert_rec;
3890 le16_add_cpu(&el->l_next_free_rec, 1);
3896 * Ok, we have to rotate.
3898 * At this point, it is safe to assume that inserting into an
3899 * empty leaf and appending to a leaf have both been handled
3902 * This leaf needs to have space, either by the empty 1st
3903 * extent record, or by virtue of an l_next_rec < l_count.
3905 ocfs2_rotate_leaf(el, insert_rec);
3908 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3910 struct ocfs2_path *path,
3911 struct ocfs2_extent_rec *insert_rec)
3913 int ret, i, next_free;
3914 struct buffer_head *bh;
3915 struct ocfs2_extent_list *el;
3916 struct ocfs2_extent_rec *rec;
3919 * Update everything except the leaf block.
3921 for (i = 0; i < path->p_tree_depth; i++) {
3922 bh = path->p_node[i].bh;
3923 el = path->p_node[i].el;
3925 next_free = le16_to_cpu(el->l_next_free_rec);
3926 if (next_free == 0) {
3927 ocfs2_error(inode->i_sb,
3928 "Dinode %llu has a bad extent list",
3929 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3934 rec = &el->l_recs[next_free - 1];
3936 rec->e_int_clusters = insert_rec->e_cpos;
3937 le32_add_cpu(&rec->e_int_clusters,
3938 le16_to_cpu(insert_rec->e_leaf_clusters));
3939 le32_add_cpu(&rec->e_int_clusters,
3940 -le32_to_cpu(rec->e_cpos));
3942 ret = ocfs2_journal_dirty(handle, bh);
3949 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3950 struct ocfs2_extent_rec *insert_rec,
3951 struct ocfs2_path *right_path,
3952 struct ocfs2_path **ret_left_path)
3955 struct ocfs2_extent_list *el;
3956 struct ocfs2_path *left_path = NULL;
3958 *ret_left_path = NULL;
3961 * This shouldn't happen for non-trees. The extent rec cluster
3962 * count manipulation below only works for interior nodes.
3964 BUG_ON(right_path->p_tree_depth == 0);
3967 * If our appending insert is at the leftmost edge of a leaf,
3968 * then we might need to update the rightmost records of the
3971 el = path_leaf_el(right_path);
3972 next_free = le16_to_cpu(el->l_next_free_rec);
3973 if (next_free == 0 ||
3974 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3977 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3984 mlog(0, "Append may need a left path update. cpos: %u, "
3985 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3989 * No need to worry if the append is already in the
3993 left_path = ocfs2_new_path_from_path(right_path);
4000 ret = ocfs2_find_path(INODE_CACHE(inode), left_path,
4008 * ocfs2_insert_path() will pass the left_path to the
4014 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, right_path);
4020 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
4022 *ret_left_path = left_path;
4026 ocfs2_free_path(left_path);
4031 static void ocfs2_split_record(struct inode *inode,
4032 struct ocfs2_path *left_path,
4033 struct ocfs2_path *right_path,
4034 struct ocfs2_extent_rec *split_rec,
4035 enum ocfs2_split_type split)
4038 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4039 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4040 struct ocfs2_extent_rec *rec, *tmprec;
4042 right_el = path_leaf_el(right_path);
4044 left_el = path_leaf_el(left_path);
4047 insert_el = right_el;
4048 index = ocfs2_search_extent_list(el, cpos);
4050 if (index == 0 && left_path) {
4051 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4054 * This typically means that the record
4055 * started in the left path but moved to the
4056 * right as a result of rotation. We either
4057 * move the existing record to the left, or we
4058 * do the later insert there.
4060 * In this case, the left path should always
4061 * exist as the rotate code will have passed
4062 * it back for a post-insert update.
4065 if (split == SPLIT_LEFT) {
4067 * It's a left split. Since we know
4068 * that the rotate code gave us an
4069 * empty extent in the left path, we
4070 * can just do the insert there.
4072 insert_el = left_el;
4075 * Right split - we have to move the
4076 * existing record over to the left
4077 * leaf. The insert will be into the
4078 * newly created empty extent in the
4081 tmprec = &right_el->l_recs[index];
4082 ocfs2_rotate_leaf(left_el, tmprec);
4085 memset(tmprec, 0, sizeof(*tmprec));
4086 index = ocfs2_search_extent_list(left_el, cpos);
4087 BUG_ON(index == -1);
4092 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4094 * Left path is easy - we can just allow the insert to
4098 insert_el = left_el;
4099 index = ocfs2_search_extent_list(el, cpos);
4100 BUG_ON(index == -1);
4103 rec = &el->l_recs[index];
4104 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
4105 ocfs2_rotate_leaf(insert_el, split_rec);
4109 * This function only does inserts on an allocation b-tree. For tree
4110 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4112 * right_path is the path we want to do the actual insert
4113 * in. left_path should only be passed in if we need to update that
4114 * portion of the tree after an edge insert.
4116 static int ocfs2_insert_path(struct inode *inode,
4118 struct ocfs2_path *left_path,
4119 struct ocfs2_path *right_path,
4120 struct ocfs2_extent_rec *insert_rec,
4121 struct ocfs2_insert_type *insert)
4123 int ret, subtree_index;
4124 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4127 int credits = handle->h_buffer_credits;
4130 * There's a chance that left_path got passed back to
4131 * us without being accounted for in the
4132 * journal. Extend our transaction here to be sure we
4133 * can change those blocks.
4135 credits += left_path->p_tree_depth;
4137 ret = ocfs2_extend_trans(handle, credits);
4143 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, left_path);
4151 * Pass both paths to the journal. The majority of inserts
4152 * will be touching all components anyway.
4154 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, right_path);
4160 if (insert->ins_split != SPLIT_NONE) {
4162 * We could call ocfs2_insert_at_leaf() for some types
4163 * of splits, but it's easier to just let one separate
4164 * function sort it all out.
4166 ocfs2_split_record(inode, left_path, right_path,
4167 insert_rec, insert->ins_split);
4170 * Split might have modified either leaf and we don't
4171 * have a guarantee that the later edge insert will
4172 * dirty this for us.
4175 ret = ocfs2_journal_dirty(handle,
4176 path_leaf_bh(left_path));
4180 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
4183 ret = ocfs2_journal_dirty(handle, leaf_bh);
4189 * The rotate code has indicated that we need to fix
4190 * up portions of the tree after the insert.
4192 * XXX: Should we extend the transaction here?
4194 subtree_index = ocfs2_find_subtree_root(inode, left_path,
4196 ocfs2_complete_edge_insert(inode, handle, left_path,
4197 right_path, subtree_index);
4205 static int ocfs2_do_insert_extent(struct inode *inode,
4207 struct ocfs2_extent_tree *et,
4208 struct ocfs2_extent_rec *insert_rec,
4209 struct ocfs2_insert_type *type)
4211 int ret, rotate = 0;
4213 struct ocfs2_path *right_path = NULL;
4214 struct ocfs2_path *left_path = NULL;
4215 struct ocfs2_extent_list *el;
4217 el = et->et_root_el;
4219 ret = ocfs2_et_root_journal_access(handle, et,
4220 OCFS2_JOURNAL_ACCESS_WRITE);
4226 if (le16_to_cpu(el->l_tree_depth) == 0) {
4227 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
4228 goto out_update_clusters;
4231 right_path = ocfs2_new_path_from_et(et);
4239 * Determine the path to start with. Rotations need the
4240 * rightmost path, everything else can go directly to the
4243 cpos = le32_to_cpu(insert_rec->e_cpos);
4244 if (type->ins_appending == APPEND_NONE &&
4245 type->ins_contig == CONTIG_NONE) {
4250 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4257 * Rotations and appends need special treatment - they modify
4258 * parts of the tree's above them.
4260 * Both might pass back a path immediate to the left of the
4261 * one being inserted to. This will be cause
4262 * ocfs2_insert_path() to modify the rightmost records of
4263 * left_path to account for an edge insert.
4265 * XXX: When modifying this code, keep in mind that an insert
4266 * can wind up skipping both of these two special cases...
4269 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
4270 le32_to_cpu(insert_rec->e_cpos),
4271 right_path, &left_path);
4278 * ocfs2_rotate_tree_right() might have extended the
4279 * transaction without re-journaling our tree root.
4281 ret = ocfs2_et_root_journal_access(handle, et,
4282 OCFS2_JOURNAL_ACCESS_WRITE);
4287 } else if (type->ins_appending == APPEND_TAIL
4288 && type->ins_contig != CONTIG_LEFT) {
4289 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
4290 right_path, &left_path);
4297 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
4304 out_update_clusters:
4305 if (type->ins_split == SPLIT_NONE)
4306 ocfs2_et_update_clusters(inode, et,
4307 le16_to_cpu(insert_rec->e_leaf_clusters));
4309 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
4314 ocfs2_free_path(left_path);
4315 ocfs2_free_path(right_path);
4320 static enum ocfs2_contig_type
4321 ocfs2_figure_merge_contig_type(struct inode *inode, struct ocfs2_path *path,
4322 struct ocfs2_extent_list *el, int index,
4323 struct ocfs2_extent_rec *split_rec)
4326 enum ocfs2_contig_type ret = CONTIG_NONE;
4327 u32 left_cpos, right_cpos;
4328 struct ocfs2_extent_rec *rec = NULL;
4329 struct ocfs2_extent_list *new_el;
4330 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4331 struct buffer_head *bh;
4332 struct ocfs2_extent_block *eb;
4335 rec = &el->l_recs[index - 1];
4336 } else if (path->p_tree_depth > 0) {
4337 status = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
4342 if (left_cpos != 0) {
4343 left_path = ocfs2_new_path_from_path(path);
4347 status = ocfs2_find_path(INODE_CACHE(inode),
4348 left_path, left_cpos);
4352 new_el = path_leaf_el(left_path);
4354 if (le16_to_cpu(new_el->l_next_free_rec) !=
4355 le16_to_cpu(new_el->l_count)) {
4356 bh = path_leaf_bh(left_path);
4357 eb = (struct ocfs2_extent_block *)bh->b_data;
4358 ocfs2_error(inode->i_sb,
4359 "Extent block #%llu has an "
4360 "invalid l_next_free_rec of "
4361 "%d. It should have "
4362 "matched the l_count of %d",
4363 (unsigned long long)le64_to_cpu(eb->h_blkno),
4364 le16_to_cpu(new_el->l_next_free_rec),
4365 le16_to_cpu(new_el->l_count));
4369 rec = &new_el->l_recs[
4370 le16_to_cpu(new_el->l_next_free_rec) - 1];
4375 * We're careful to check for an empty extent record here -
4376 * the merge code will know what to do if it sees one.
4379 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4380 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4383 ret = ocfs2_extent_contig(inode, rec, split_rec);
4388 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4389 rec = &el->l_recs[index + 1];
4390 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4391 path->p_tree_depth > 0) {
4392 status = ocfs2_find_cpos_for_right_leaf(inode->i_sb,
4397 if (right_cpos == 0)
4400 right_path = ocfs2_new_path_from_path(path);
4404 status = ocfs2_find_path(INODE_CACHE(inode), right_path, right_cpos);
4408 new_el = path_leaf_el(right_path);
4409 rec = &new_el->l_recs[0];
4410 if (ocfs2_is_empty_extent(rec)) {
4411 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4412 bh = path_leaf_bh(right_path);
4413 eb = (struct ocfs2_extent_block *)bh->b_data;
4414 ocfs2_error(inode->i_sb,
4415 "Extent block #%llu has an "
4416 "invalid l_next_free_rec of %d",
4417 (unsigned long long)le64_to_cpu(eb->h_blkno),
4418 le16_to_cpu(new_el->l_next_free_rec));
4422 rec = &new_el->l_recs[1];
4427 enum ocfs2_contig_type contig_type;
4429 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
4431 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4432 ret = CONTIG_LEFTRIGHT;
4433 else if (ret == CONTIG_NONE)
4439 ocfs2_free_path(left_path);
4441 ocfs2_free_path(right_path);
4446 static void ocfs2_figure_contig_type(struct inode *inode,
4447 struct ocfs2_insert_type *insert,
4448 struct ocfs2_extent_list *el,
4449 struct ocfs2_extent_rec *insert_rec,
4450 struct ocfs2_extent_tree *et)
4453 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4455 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4457 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4458 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
4460 if (contig_type != CONTIG_NONE) {
4461 insert->ins_contig_index = i;
4465 insert->ins_contig = contig_type;
4467 if (insert->ins_contig != CONTIG_NONE) {
4468 struct ocfs2_extent_rec *rec =
4469 &el->l_recs[insert->ins_contig_index];
4470 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4471 le16_to_cpu(insert_rec->e_leaf_clusters);
4474 * Caller might want us to limit the size of extents, don't
4475 * calculate contiguousness if we might exceed that limit.
4477 if (et->et_max_leaf_clusters &&
4478 (len > et->et_max_leaf_clusters))
4479 insert->ins_contig = CONTIG_NONE;
4484 * This should only be called against the righmost leaf extent list.
4486 * ocfs2_figure_appending_type() will figure out whether we'll have to
4487 * insert at the tail of the rightmost leaf.
4489 * This should also work against the root extent list for tree's with 0
4490 * depth. If we consider the root extent list to be the rightmost leaf node
4491 * then the logic here makes sense.
4493 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4494 struct ocfs2_extent_list *el,
4495 struct ocfs2_extent_rec *insert_rec)
4498 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4499 struct ocfs2_extent_rec *rec;
4501 insert->ins_appending = APPEND_NONE;
4503 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4505 if (!el->l_next_free_rec)
4506 goto set_tail_append;
4508 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4509 /* Were all records empty? */
4510 if (le16_to_cpu(el->l_next_free_rec) == 1)
4511 goto set_tail_append;
4514 i = le16_to_cpu(el->l_next_free_rec) - 1;
4515 rec = &el->l_recs[i];
4518 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4519 goto set_tail_append;
4524 insert->ins_appending = APPEND_TAIL;
4528 * Helper function called at the begining of an insert.
4530 * This computes a few things that are commonly used in the process of
4531 * inserting into the btree:
4532 * - Whether the new extent is contiguous with an existing one.
4533 * - The current tree depth.
4534 * - Whether the insert is an appending one.
4535 * - The total # of free records in the tree.
4537 * All of the information is stored on the ocfs2_insert_type
4540 static int ocfs2_figure_insert_type(struct inode *inode,
4541 struct ocfs2_extent_tree *et,
4542 struct buffer_head **last_eb_bh,
4543 struct ocfs2_extent_rec *insert_rec,
4545 struct ocfs2_insert_type *insert)
4548 struct ocfs2_extent_block *eb;
4549 struct ocfs2_extent_list *el;
4550 struct ocfs2_path *path = NULL;
4551 struct buffer_head *bh = NULL;
4553 insert->ins_split = SPLIT_NONE;
4555 el = et->et_root_el;
4556 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4558 if (el->l_tree_depth) {
4560 * If we have tree depth, we read in the
4561 * rightmost extent block ahead of time as
4562 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4563 * may want it later.
4565 ret = ocfs2_read_extent_block(et->et_ci,
4566 ocfs2_et_get_last_eb_blk(et),
4572 eb = (struct ocfs2_extent_block *) bh->b_data;
4577 * Unless we have a contiguous insert, we'll need to know if
4578 * there is room left in our allocation tree for another
4581 * XXX: This test is simplistic, we can search for empty
4582 * extent records too.
4584 *free_records = le16_to_cpu(el->l_count) -
4585 le16_to_cpu(el->l_next_free_rec);
4587 if (!insert->ins_tree_depth) {
4588 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4589 ocfs2_figure_appending_type(insert, el, insert_rec);
4593 path = ocfs2_new_path_from_et(et);
4601 * In the case that we're inserting past what the tree
4602 * currently accounts for, ocfs2_find_path() will return for
4603 * us the rightmost tree path. This is accounted for below in
4604 * the appending code.
4606 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4612 el = path_leaf_el(path);
4615 * Now that we have the path, there's two things we want to determine:
4616 * 1) Contiguousness (also set contig_index if this is so)
4618 * 2) Are we doing an append? We can trivially break this up
4619 * into two types of appends: simple record append, or a
4620 * rotate inside the tail leaf.
4622 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4625 * The insert code isn't quite ready to deal with all cases of
4626 * left contiguousness. Specifically, if it's an insert into
4627 * the 1st record in a leaf, it will require the adjustment of
4628 * cluster count on the last record of the path directly to it's
4629 * left. For now, just catch that case and fool the layers
4630 * above us. This works just fine for tree_depth == 0, which
4631 * is why we allow that above.
4633 if (insert->ins_contig == CONTIG_LEFT &&
4634 insert->ins_contig_index == 0)
4635 insert->ins_contig = CONTIG_NONE;
4638 * Ok, so we can simply compare against last_eb to figure out
4639 * whether the path doesn't exist. This will only happen in
4640 * the case that we're doing a tail append, so maybe we can
4641 * take advantage of that information somehow.
4643 if (ocfs2_et_get_last_eb_blk(et) ==
4644 path_leaf_bh(path)->b_blocknr) {
4646 * Ok, ocfs2_find_path() returned us the rightmost
4647 * tree path. This might be an appending insert. There are
4649 * 1) We're doing a true append at the tail:
4650 * -This might even be off the end of the leaf
4651 * 2) We're "appending" by rotating in the tail
4653 ocfs2_figure_appending_type(insert, el, insert_rec);
4657 ocfs2_free_path(path);
4667 * Insert an extent into an inode btree.
4669 * The caller needs to update fe->i_clusters
4671 int ocfs2_insert_extent(struct ocfs2_super *osb,
4673 struct inode *inode,
4674 struct ocfs2_extent_tree *et,
4679 struct ocfs2_alloc_context *meta_ac)
4682 int uninitialized_var(free_records);
4683 struct buffer_head *last_eb_bh = NULL;
4684 struct ocfs2_insert_type insert = {0, };
4685 struct ocfs2_extent_rec rec;
4687 mlog(0, "add %u clusters at position %u to inode %llu\n",
4688 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4690 memset(&rec, 0, sizeof(rec));
4691 rec.e_cpos = cpu_to_le32(cpos);
4692 rec.e_blkno = cpu_to_le64(start_blk);
4693 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4694 rec.e_flags = flags;
4695 status = ocfs2_et_insert_check(inode, et, &rec);
4701 status = ocfs2_figure_insert_type(inode, et, &last_eb_bh, &rec,
4702 &free_records, &insert);
4708 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4709 "Insert.contig_index: %d, Insert.free_records: %d, "
4710 "Insert.tree_depth: %d\n",
4711 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
4712 free_records, insert.ins_tree_depth);
4714 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4715 status = ocfs2_grow_tree(inode, handle, et,
4716 &insert.ins_tree_depth, &last_eb_bh,
4724 /* Finally, we can add clusters. This might rotate the tree for us. */
4725 status = ocfs2_do_insert_extent(inode, handle, et, &rec, &insert);
4728 else if (et->et_ops == &ocfs2_dinode_et_ops)
4729 ocfs2_extent_map_insert_rec(inode, &rec);
4739 * Allcate and add clusters into the extent b-tree.
4740 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4741 * The extent b-tree's root is specified by et, and
4742 * it is not limited to the file storage. Any extent tree can use this
4743 * function if it implements the proper ocfs2_extent_tree.
4745 int ocfs2_add_clusters_in_btree(struct ocfs2_super *osb,
4746 struct inode *inode,
4747 u32 *logical_offset,
4748 u32 clusters_to_add,
4750 struct ocfs2_extent_tree *et,
4752 struct ocfs2_alloc_context *data_ac,
4753 struct ocfs2_alloc_context *meta_ac,
4754 enum ocfs2_alloc_restarted *reason_ret)
4758 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4759 u32 bit_off, num_bits;
4763 BUG_ON(!clusters_to_add);
4766 flags = OCFS2_EXT_UNWRITTEN;
4768 free_extents = ocfs2_num_free_extents(osb, et);
4769 if (free_extents < 0) {
4770 status = free_extents;
4775 /* there are two cases which could cause us to EAGAIN in the
4776 * we-need-more-metadata case:
4777 * 1) we haven't reserved *any*
4778 * 2) we are so fragmented, we've needed to add metadata too
4780 if (!free_extents && !meta_ac) {
4781 mlog(0, "we haven't reserved any metadata!\n");
4783 reason = RESTART_META;
4785 } else if ((!free_extents)
4786 && (ocfs2_alloc_context_bits_left(meta_ac)
4787 < ocfs2_extend_meta_needed(et->et_root_el))) {
4788 mlog(0, "filesystem is really fragmented...\n");
4790 reason = RESTART_META;
4794 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
4795 clusters_to_add, &bit_off, &num_bits);
4797 if (status != -ENOSPC)
4802 BUG_ON(num_bits > clusters_to_add);
4804 /* reserve our write early -- insert_extent may update the tree root */
4805 status = ocfs2_et_root_journal_access(handle, et,
4806 OCFS2_JOURNAL_ACCESS_WRITE);
4812 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4813 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
4814 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4815 status = ocfs2_insert_extent(osb, handle, inode, et,
4816 *logical_offset, block,
4817 num_bits, flags, meta_ac);
4823 status = ocfs2_journal_dirty(handle, et->et_root_bh);
4829 clusters_to_add -= num_bits;
4830 *logical_offset += num_bits;
4832 if (clusters_to_add) {
4833 mlog(0, "need to alloc once more, wanted = %u\n",
4836 reason = RESTART_TRANS;
4842 *reason_ret = reason;
4846 static void ocfs2_make_right_split_rec(struct super_block *sb,
4847 struct ocfs2_extent_rec *split_rec,
4849 struct ocfs2_extent_rec *rec)
4851 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4852 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4854 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4856 split_rec->e_cpos = cpu_to_le32(cpos);
4857 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4859 split_rec->e_blkno = rec->e_blkno;
4860 le64_add_cpu(&split_rec->e_blkno,
4861 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4863 split_rec->e_flags = rec->e_flags;
4866 static int ocfs2_split_and_insert(struct inode *inode,
4868 struct ocfs2_path *path,
4869 struct ocfs2_extent_tree *et,
4870 struct buffer_head **last_eb_bh,
4872 struct ocfs2_extent_rec *orig_split_rec,
4873 struct ocfs2_alloc_context *meta_ac)
4876 unsigned int insert_range, rec_range, do_leftright = 0;
4877 struct ocfs2_extent_rec tmprec;
4878 struct ocfs2_extent_list *rightmost_el;
4879 struct ocfs2_extent_rec rec;
4880 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4881 struct ocfs2_insert_type insert;
4882 struct ocfs2_extent_block *eb;
4886 * Store a copy of the record on the stack - it might move
4887 * around as the tree is manipulated below.
4889 rec = path_leaf_el(path)->l_recs[split_index];
4891 rightmost_el = et->et_root_el;
4893 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4895 BUG_ON(!(*last_eb_bh));
4896 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4897 rightmost_el = &eb->h_list;
4900 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4901 le16_to_cpu(rightmost_el->l_count)) {
4902 ret = ocfs2_grow_tree(inode, handle, et,
4903 &depth, last_eb_bh, meta_ac);
4910 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4911 insert.ins_appending = APPEND_NONE;
4912 insert.ins_contig = CONTIG_NONE;
4913 insert.ins_tree_depth = depth;
4915 insert_range = le32_to_cpu(split_rec.e_cpos) +
4916 le16_to_cpu(split_rec.e_leaf_clusters);
4917 rec_range = le32_to_cpu(rec.e_cpos) +
4918 le16_to_cpu(rec.e_leaf_clusters);
4920 if (split_rec.e_cpos == rec.e_cpos) {
4921 insert.ins_split = SPLIT_LEFT;
4922 } else if (insert_range == rec_range) {
4923 insert.ins_split = SPLIT_RIGHT;
4926 * Left/right split. We fake this as a right split
4927 * first and then make a second pass as a left split.
4929 insert.ins_split = SPLIT_RIGHT;
4931 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
4936 BUG_ON(do_leftright);
4940 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4946 if (do_leftright == 1) {
4948 struct ocfs2_extent_list *el;
4951 split_rec = *orig_split_rec;
4953 ocfs2_reinit_path(path, 1);
4955 cpos = le32_to_cpu(split_rec.e_cpos);
4956 ret = ocfs2_find_path(et->et_ci, path, cpos);
4962 el = path_leaf_el(path);
4963 split_index = ocfs2_search_extent_list(el, cpos);
4971 static int ocfs2_replace_extent_rec(struct inode *inode,
4973 struct ocfs2_path *path,
4974 struct ocfs2_extent_list *el,
4976 struct ocfs2_extent_rec *split_rec)
4980 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), path,
4981 path_num_items(path) - 1);
4987 el->l_recs[split_index] = *split_rec;
4989 ocfs2_journal_dirty(handle, path_leaf_bh(path));
4995 * Mark part or all of the extent record at split_index in the leaf
4996 * pointed to by path as written. This removes the unwritten
4999 * Care is taken to handle contiguousness so as to not grow the tree.
5001 * meta_ac is not strictly necessary - we only truly need it if growth
5002 * of the tree is required. All other cases will degrade into a less
5003 * optimal tree layout.
5005 * last_eb_bh should be the rightmost leaf block for any extent
5006 * btree. Since a split may grow the tree or a merge might shrink it,
5007 * the caller cannot trust the contents of that buffer after this call.
5009 * This code is optimized for readability - several passes might be
5010 * made over certain portions of the tree. All of those blocks will
5011 * have been brought into cache (and pinned via the journal), so the
5012 * extra overhead is not expressed in terms of disk reads.
5014 static int __ocfs2_mark_extent_written(struct inode *inode,
5015 struct ocfs2_extent_tree *et,
5017 struct ocfs2_path *path,
5019 struct ocfs2_extent_rec *split_rec,
5020 struct ocfs2_alloc_context *meta_ac,
5021 struct ocfs2_cached_dealloc_ctxt *dealloc)
5024 struct ocfs2_extent_list *el = path_leaf_el(path);
5025 struct buffer_head *last_eb_bh = NULL;
5026 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5027 struct ocfs2_merge_ctxt ctxt;
5028 struct ocfs2_extent_list *rightmost_el;
5030 if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) {
5036 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5037 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5038 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5044 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, path, el,
5049 * The core merge / split code wants to know how much room is
5050 * left in this inodes allocation tree, so we pass the
5051 * rightmost extent list.
5053 if (path->p_tree_depth) {
5054 struct ocfs2_extent_block *eb;
5056 ret = ocfs2_read_extent_block(et->et_ci,
5057 ocfs2_et_get_last_eb_blk(et),
5064 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5065 rightmost_el = &eb->h_list;
5067 rightmost_el = path_root_el(path);
5069 if (rec->e_cpos == split_rec->e_cpos &&
5070 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5071 ctxt.c_split_covers_rec = 1;
5073 ctxt.c_split_covers_rec = 0;
5075 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5077 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
5078 split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent,
5079 ctxt.c_split_covers_rec);
5081 if (ctxt.c_contig_type == CONTIG_NONE) {
5082 if (ctxt.c_split_covers_rec)
5083 ret = ocfs2_replace_extent_rec(inode, handle,
5085 split_index, split_rec);
5087 ret = ocfs2_split_and_insert(inode, handle, path, et,
5088 &last_eb_bh, split_index,
5089 split_rec, meta_ac);
5093 ret = ocfs2_try_to_merge_extent(inode, handle, path,
5094 split_index, split_rec,
5095 dealloc, &ctxt, et);
5106 * Mark the already-existing extent at cpos as written for len clusters.
5108 * If the existing extent is larger than the request, initiate a
5109 * split. An attempt will be made at merging with adjacent extents.
5111 * The caller is responsible for passing down meta_ac if we'll need it.
5113 int ocfs2_mark_extent_written(struct inode *inode,
5114 struct ocfs2_extent_tree *et,
5115 handle_t *handle, u32 cpos, u32 len, u32 phys,
5116 struct ocfs2_alloc_context *meta_ac,
5117 struct ocfs2_cached_dealloc_ctxt *dealloc)
5120 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
5121 struct ocfs2_extent_rec split_rec;
5122 struct ocfs2_path *left_path = NULL;
5123 struct ocfs2_extent_list *el;
5125 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
5126 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
5128 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5129 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
5130 "that are being written to, but the feature bit "
5131 "is not set in the super block.",
5132 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5138 * XXX: This should be fixed up so that we just re-insert the
5139 * next extent records.
5141 * XXX: This is a hack on the extent tree, maybe it should be
5144 if (et->et_ops == &ocfs2_dinode_et_ops)
5145 ocfs2_extent_map_trunc(inode, 0);
5147 left_path = ocfs2_new_path_from_et(et);
5154 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5159 el = path_leaf_el(left_path);
5161 index = ocfs2_search_extent_list(el, cpos);
5162 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5163 ocfs2_error(inode->i_sb,
5164 "Inode %llu has an extent at cpos %u which can no "
5165 "longer be found.\n",
5166 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5171 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5172 split_rec.e_cpos = cpu_to_le32(cpos);
5173 split_rec.e_leaf_clusters = cpu_to_le16(len);
5174 split_rec.e_blkno = cpu_to_le64(start_blkno);
5175 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
5176 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
5178 ret = __ocfs2_mark_extent_written(inode, et, handle, left_path,
5179 index, &split_rec, meta_ac,
5185 ocfs2_free_path(left_path);
5189 static int ocfs2_split_tree(struct inode *inode, struct ocfs2_extent_tree *et,
5190 handle_t *handle, struct ocfs2_path *path,
5191 int index, u32 new_range,
5192 struct ocfs2_alloc_context *meta_ac)
5194 int ret, depth, credits = handle->h_buffer_credits;
5195 struct buffer_head *last_eb_bh = NULL;
5196 struct ocfs2_extent_block *eb;
5197 struct ocfs2_extent_list *rightmost_el, *el;
5198 struct ocfs2_extent_rec split_rec;
5199 struct ocfs2_extent_rec *rec;
5200 struct ocfs2_insert_type insert;
5203 * Setup the record to split before we grow the tree.
5205 el = path_leaf_el(path);
5206 rec = &el->l_recs[index];
5207 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
5209 depth = path->p_tree_depth;
5211 ret = ocfs2_read_extent_block(et->et_ci,
5212 ocfs2_et_get_last_eb_blk(et),
5219 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5220 rightmost_el = &eb->h_list;
5222 rightmost_el = path_leaf_el(path);
5224 credits += path->p_tree_depth +
5225 ocfs2_extend_meta_needed(et->et_root_el);
5226 ret = ocfs2_extend_trans(handle, credits);
5232 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5233 le16_to_cpu(rightmost_el->l_count)) {
5234 ret = ocfs2_grow_tree(inode, handle, et, &depth, &last_eb_bh,
5242 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5243 insert.ins_appending = APPEND_NONE;
5244 insert.ins_contig = CONTIG_NONE;
5245 insert.ins_split = SPLIT_RIGHT;
5246 insert.ins_tree_depth = depth;
5248 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
5257 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
5258 struct ocfs2_path *path, int index,
5259 struct ocfs2_cached_dealloc_ctxt *dealloc,
5261 struct ocfs2_extent_tree *et)
5264 u32 left_cpos, rec_range, trunc_range;
5265 int wants_rotate = 0, is_rightmost_tree_rec = 0;
5266 struct super_block *sb = inode->i_sb;
5267 struct ocfs2_path *left_path = NULL;
5268 struct ocfs2_extent_list *el = path_leaf_el(path);
5269 struct ocfs2_extent_rec *rec;
5270 struct ocfs2_extent_block *eb;
5272 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5273 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5282 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5283 path->p_tree_depth) {
5285 * Check whether this is the rightmost tree record. If
5286 * we remove all of this record or part of its right
5287 * edge then an update of the record lengths above it
5290 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5291 if (eb->h_next_leaf_blk == 0)
5292 is_rightmost_tree_rec = 1;
5295 rec = &el->l_recs[index];
5296 if (index == 0 && path->p_tree_depth &&
5297 le32_to_cpu(rec->e_cpos) == cpos) {
5299 * Changing the leftmost offset (via partial or whole
5300 * record truncate) of an interior (or rightmost) path
5301 * means we have to update the subtree that is formed
5302 * by this leaf and the one to it's left.
5304 * There are two cases we can skip:
5305 * 1) Path is the leftmost one in our inode tree.
5306 * 2) The leaf is rightmost and will be empty after
5307 * we remove the extent record - the rotate code
5308 * knows how to update the newly formed edge.
5311 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
5318 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5319 left_path = ocfs2_new_path_from_path(path);
5326 ret = ocfs2_find_path(et->et_ci, left_path,
5335 ret = ocfs2_extend_rotate_transaction(handle, 0,
5336 handle->h_buffer_credits,
5343 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5349 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5355 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5356 trunc_range = cpos + len;
5358 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5361 memset(rec, 0, sizeof(*rec));
5362 ocfs2_cleanup_merge(el, index);
5365 next_free = le16_to_cpu(el->l_next_free_rec);
5366 if (is_rightmost_tree_rec && next_free > 1) {
5368 * We skip the edge update if this path will
5369 * be deleted by the rotate code.
5371 rec = &el->l_recs[next_free - 1];
5372 ocfs2_adjust_rightmost_records(inode, handle, path,
5375 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5376 /* Remove leftmost portion of the record. */
5377 le32_add_cpu(&rec->e_cpos, len);
5378 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5379 le16_add_cpu(&rec->e_leaf_clusters, -len);
5380 } else if (rec_range == trunc_range) {
5381 /* Remove rightmost portion of the record */
5382 le16_add_cpu(&rec->e_leaf_clusters, -len);
5383 if (is_rightmost_tree_rec)
5384 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
5386 /* Caller should have trapped this. */
5387 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
5388 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
5389 le32_to_cpu(rec->e_cpos),
5390 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5397 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
5398 ocfs2_complete_edge_insert(inode, handle, left_path, path,
5402 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5404 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5411 ocfs2_free_path(left_path);
5415 int ocfs2_remove_extent(struct inode *inode,
5416 struct ocfs2_extent_tree *et,
5417 u32 cpos, u32 len, handle_t *handle,
5418 struct ocfs2_alloc_context *meta_ac,
5419 struct ocfs2_cached_dealloc_ctxt *dealloc)
5422 u32 rec_range, trunc_range;
5423 struct ocfs2_extent_rec *rec;
5424 struct ocfs2_extent_list *el;
5425 struct ocfs2_path *path = NULL;
5427 ocfs2_extent_map_trunc(inode, 0);
5429 path = ocfs2_new_path_from_et(et);
5436 ret = ocfs2_find_path(et->et_ci, path, cpos);
5442 el = path_leaf_el(path);
5443 index = ocfs2_search_extent_list(el, cpos);
5444 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5445 ocfs2_error(inode->i_sb,
5446 "Inode %llu has an extent at cpos %u which can no "
5447 "longer be found.\n",
5448 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5454 * We have 3 cases of extent removal:
5455 * 1) Range covers the entire extent rec
5456 * 2) Range begins or ends on one edge of the extent rec
5457 * 3) Range is in the middle of the extent rec (no shared edges)
5459 * For case 1 we remove the extent rec and left rotate to
5462 * For case 2 we just shrink the existing extent rec, with a
5463 * tree update if the shrinking edge is also the edge of an
5466 * For case 3 we do a right split to turn the extent rec into
5467 * something case 2 can handle.
5469 rec = &el->l_recs[index];
5470 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5471 trunc_range = cpos + len;
5473 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5475 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
5476 "(cpos %u, len %u)\n",
5477 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
5478 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
5480 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5481 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5488 ret = ocfs2_split_tree(inode, et, handle, path, index,
5489 trunc_range, meta_ac);
5496 * The split could have manipulated the tree enough to
5497 * move the record location, so we have to look for it again.
5499 ocfs2_reinit_path(path, 1);
5501 ret = ocfs2_find_path(et->et_ci, path, cpos);
5507 el = path_leaf_el(path);
5508 index = ocfs2_search_extent_list(el, cpos);
5509 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5510 ocfs2_error(inode->i_sb,
5511 "Inode %llu: split at cpos %u lost record.",
5512 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5519 * Double check our values here. If anything is fishy,
5520 * it's easier to catch it at the top level.
5522 rec = &el->l_recs[index];
5523 rec_range = le32_to_cpu(rec->e_cpos) +
5524 ocfs2_rec_clusters(el, rec);
5525 if (rec_range != trunc_range) {
5526 ocfs2_error(inode->i_sb,
5527 "Inode %llu: error after split at cpos %u"
5528 "trunc len %u, existing record is (%u,%u)",
5529 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5530 cpos, len, le32_to_cpu(rec->e_cpos),
5531 ocfs2_rec_clusters(el, rec));
5536 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5545 ocfs2_free_path(path);
5549 int ocfs2_remove_btree_range(struct inode *inode,
5550 struct ocfs2_extent_tree *et,
5551 u32 cpos, u32 phys_cpos, u32 len,
5552 struct ocfs2_cached_dealloc_ctxt *dealloc)
5555 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5556 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5557 struct inode *tl_inode = osb->osb_tl_inode;
5559 struct ocfs2_alloc_context *meta_ac = NULL;
5561 ret = ocfs2_lock_allocators(inode, et, 0, 1, NULL, &meta_ac);
5567 mutex_lock(&tl_inode->i_mutex);
5569 if (ocfs2_truncate_log_needs_flush(osb)) {
5570 ret = __ocfs2_flush_truncate_log(osb);
5577 handle = ocfs2_start_trans(osb, ocfs2_remove_extent_credits(osb->sb));
5578 if (IS_ERR(handle)) {
5579 ret = PTR_ERR(handle);
5584 ret = ocfs2_et_root_journal_access(handle, et,
5585 OCFS2_JOURNAL_ACCESS_WRITE);
5591 vfs_dq_free_space_nodirty(inode,
5592 ocfs2_clusters_to_bytes(inode->i_sb, len));
5594 ret = ocfs2_remove_extent(inode, et, cpos, len, handle, meta_ac,
5601 ocfs2_et_update_clusters(inode, et, -len);
5603 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
5609 ret = ocfs2_truncate_log_append(osb, handle, phys_blkno, len);
5614 ocfs2_commit_trans(osb, handle);
5616 mutex_unlock(&tl_inode->i_mutex);
5619 ocfs2_free_alloc_context(meta_ac);
5624 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5626 struct buffer_head *tl_bh = osb->osb_tl_bh;
5627 struct ocfs2_dinode *di;
5628 struct ocfs2_truncate_log *tl;
5630 di = (struct ocfs2_dinode *) tl_bh->b_data;
5631 tl = &di->id2.i_dealloc;
5633 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5634 "slot %d, invalid truncate log parameters: used = "
5635 "%u, count = %u\n", osb->slot_num,
5636 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5637 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5640 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5641 unsigned int new_start)
5643 unsigned int tail_index;
5644 unsigned int current_tail;
5646 /* No records, nothing to coalesce */
5647 if (!le16_to_cpu(tl->tl_used))
5650 tail_index = le16_to_cpu(tl->tl_used) - 1;
5651 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5652 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5654 return current_tail == new_start;
5657 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5660 unsigned int num_clusters)
5663 unsigned int start_cluster, tl_count;
5664 struct inode *tl_inode = osb->osb_tl_inode;
5665 struct buffer_head *tl_bh = osb->osb_tl_bh;
5666 struct ocfs2_dinode *di;
5667 struct ocfs2_truncate_log *tl;
5669 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5670 (unsigned long long)start_blk, num_clusters);
5672 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5674 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5676 di = (struct ocfs2_dinode *) tl_bh->b_data;
5678 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5679 * by the underlying call to ocfs2_read_inode_block(), so any
5680 * corruption is a code bug */
5681 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5683 tl = &di->id2.i_dealloc;
5684 tl_count = le16_to_cpu(tl->tl_count);
5685 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5687 "Truncate record count on #%llu invalid "
5688 "wanted %u, actual %u\n",
5689 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5690 ocfs2_truncate_recs_per_inode(osb->sb),
5691 le16_to_cpu(tl->tl_count));
5693 /* Caller should have known to flush before calling us. */
5694 index = le16_to_cpu(tl->tl_used);
5695 if (index >= tl_count) {
5701 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5702 OCFS2_JOURNAL_ACCESS_WRITE);
5708 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5709 "%llu (index = %d)\n", num_clusters, start_cluster,
5710 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
5712 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5714 * Move index back to the record we are coalescing with.
5715 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5719 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5720 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5721 index, le32_to_cpu(tl->tl_recs[index].t_start),
5724 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5725 tl->tl_used = cpu_to_le16(index + 1);
5727 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5729 status = ocfs2_journal_dirty(handle, tl_bh);
5740 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5742 struct inode *data_alloc_inode,
5743 struct buffer_head *data_alloc_bh)
5747 unsigned int num_clusters;
5749 struct ocfs2_truncate_rec rec;
5750 struct ocfs2_dinode *di;
5751 struct ocfs2_truncate_log *tl;
5752 struct inode *tl_inode = osb->osb_tl_inode;
5753 struct buffer_head *tl_bh = osb->osb_tl_bh;
5757 di = (struct ocfs2_dinode *) tl_bh->b_data;
5758 tl = &di->id2.i_dealloc;
5759 i = le16_to_cpu(tl->tl_used) - 1;
5761 /* Caller has given us at least enough credits to
5762 * update the truncate log dinode */
5763 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5764 OCFS2_JOURNAL_ACCESS_WRITE);
5770 tl->tl_used = cpu_to_le16(i);
5772 status = ocfs2_journal_dirty(handle, tl_bh);
5778 /* TODO: Perhaps we can calculate the bulk of the
5779 * credits up front rather than extending like
5781 status = ocfs2_extend_trans(handle,
5782 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5788 rec = tl->tl_recs[i];
5789 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5790 le32_to_cpu(rec.t_start));
5791 num_clusters = le32_to_cpu(rec.t_clusters);
5793 /* if start_blk is not set, we ignore the record as
5796 mlog(0, "free record %d, start = %u, clusters = %u\n",
5797 i, le32_to_cpu(rec.t_start), num_clusters);
5799 status = ocfs2_free_clusters(handle, data_alloc_inode,
5800 data_alloc_bh, start_blk,
5815 /* Expects you to already be holding tl_inode->i_mutex */
5816 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5819 unsigned int num_to_flush;
5821 struct inode *tl_inode = osb->osb_tl_inode;
5822 struct inode *data_alloc_inode = NULL;
5823 struct buffer_head *tl_bh = osb->osb_tl_bh;
5824 struct buffer_head *data_alloc_bh = NULL;
5825 struct ocfs2_dinode *di;
5826 struct ocfs2_truncate_log *tl;
5830 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5832 di = (struct ocfs2_dinode *) tl_bh->b_data;
5834 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5835 * by the underlying call to ocfs2_read_inode_block(), so any
5836 * corruption is a code bug */
5837 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5839 tl = &di->id2.i_dealloc;
5840 num_to_flush = le16_to_cpu(tl->tl_used);
5841 mlog(0, "Flush %u records from truncate log #%llu\n",
5842 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
5843 if (!num_to_flush) {
5848 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5849 GLOBAL_BITMAP_SYSTEM_INODE,
5850 OCFS2_INVALID_SLOT);
5851 if (!data_alloc_inode) {
5853 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5857 mutex_lock(&data_alloc_inode->i_mutex);
5859 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5865 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5866 if (IS_ERR(handle)) {
5867 status = PTR_ERR(handle);
5872 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
5877 ocfs2_commit_trans(osb, handle);
5880 brelse(data_alloc_bh);
5881 ocfs2_inode_unlock(data_alloc_inode, 1);
5884 mutex_unlock(&data_alloc_inode->i_mutex);
5885 iput(data_alloc_inode);
5892 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5895 struct inode *tl_inode = osb->osb_tl_inode;
5897 mutex_lock(&tl_inode->i_mutex);
5898 status = __ocfs2_flush_truncate_log(osb);
5899 mutex_unlock(&tl_inode->i_mutex);
5904 static void ocfs2_truncate_log_worker(struct work_struct *work)
5907 struct ocfs2_super *osb =
5908 container_of(work, struct ocfs2_super,
5909 osb_truncate_log_wq.work);
5913 status = ocfs2_flush_truncate_log(osb);
5917 ocfs2_init_inode_steal_slot(osb);
5922 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5923 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
5926 if (osb->osb_tl_inode) {
5927 /* We want to push off log flushes while truncates are
5930 cancel_delayed_work(&osb->osb_truncate_log_wq);
5932 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
5933 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
5937 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
5939 struct inode **tl_inode,
5940 struct buffer_head **tl_bh)
5943 struct inode *inode = NULL;
5944 struct buffer_head *bh = NULL;
5946 inode = ocfs2_get_system_file_inode(osb,
5947 TRUNCATE_LOG_SYSTEM_INODE,
5951 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
5955 status = ocfs2_read_inode_block(inode, &bh);
5969 /* called during the 1st stage of node recovery. we stamp a clean
5970 * truncate log and pass back a copy for processing later. if the
5971 * truncate log does not require processing, a *tl_copy is set to
5973 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
5975 struct ocfs2_dinode **tl_copy)
5978 struct inode *tl_inode = NULL;
5979 struct buffer_head *tl_bh = NULL;
5980 struct ocfs2_dinode *di;
5981 struct ocfs2_truncate_log *tl;
5985 mlog(0, "recover truncate log from slot %d\n", slot_num);
5987 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
5993 di = (struct ocfs2_dinode *) tl_bh->b_data;
5995 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
5996 * validated by the underlying call to ocfs2_read_inode_block(),
5997 * so any corruption is a code bug */
5998 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6000 tl = &di->id2.i_dealloc;
6001 if (le16_to_cpu(tl->tl_used)) {
6002 mlog(0, "We'll have %u logs to recover\n",
6003 le16_to_cpu(tl->tl_used));
6005 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6012 /* Assuming the write-out below goes well, this copy
6013 * will be passed back to recovery for processing. */
6014 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6016 /* All we need to do to clear the truncate log is set
6020 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6021 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6033 if (status < 0 && (*tl_copy)) {
6042 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6043 struct ocfs2_dinode *tl_copy)
6047 unsigned int clusters, num_recs, start_cluster;
6050 struct inode *tl_inode = osb->osb_tl_inode;
6051 struct ocfs2_truncate_log *tl;
6055 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6056 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6060 tl = &tl_copy->id2.i_dealloc;
6061 num_recs = le16_to_cpu(tl->tl_used);
6062 mlog(0, "cleanup %u records from %llu\n", num_recs,
6063 (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
6065 mutex_lock(&tl_inode->i_mutex);
6066 for(i = 0; i < num_recs; i++) {
6067 if (ocfs2_truncate_log_needs_flush(osb)) {
6068 status = __ocfs2_flush_truncate_log(osb);
6075 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6076 if (IS_ERR(handle)) {
6077 status = PTR_ERR(handle);
6082 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6083 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6084 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6086 status = ocfs2_truncate_log_append(osb, handle,
6087 start_blk, clusters);
6088 ocfs2_commit_trans(osb, handle);
6096 mutex_unlock(&tl_inode->i_mutex);
6102 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6105 struct inode *tl_inode = osb->osb_tl_inode;
6110 cancel_delayed_work(&osb->osb_truncate_log_wq);
6111 flush_workqueue(ocfs2_wq);
6113 status = ocfs2_flush_truncate_log(osb);
6117 brelse(osb->osb_tl_bh);
6118 iput(osb->osb_tl_inode);
6124 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6127 struct inode *tl_inode = NULL;
6128 struct buffer_head *tl_bh = NULL;
6132 status = ocfs2_get_truncate_log_info(osb,
6139 /* ocfs2_truncate_log_shutdown keys on the existence of
6140 * osb->osb_tl_inode so we don't set any of the osb variables
6141 * until we're sure all is well. */
6142 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6143 ocfs2_truncate_log_worker);
6144 osb->osb_tl_bh = tl_bh;
6145 osb->osb_tl_inode = tl_inode;
6152 * Delayed de-allocation of suballocator blocks.
6154 * Some sets of block de-allocations might involve multiple suballocator inodes.
6156 * The locking for this can get extremely complicated, especially when
6157 * the suballocator inodes to delete from aren't known until deep
6158 * within an unrelated codepath.
6160 * ocfs2_extent_block structures are a good example of this - an inode
6161 * btree could have been grown by any number of nodes each allocating
6162 * out of their own suballoc inode.
6164 * These structures allow the delay of block de-allocation until a
6165 * later time, when locking of multiple cluster inodes won't cause
6170 * Describe a single bit freed from a suballocator. For the block
6171 * suballocators, it represents one block. For the global cluster
6172 * allocator, it represents some clusters and free_bit indicates
6175 struct ocfs2_cached_block_free {
6176 struct ocfs2_cached_block_free *free_next;
6178 unsigned int free_bit;
6181 struct ocfs2_per_slot_free_list {
6182 struct ocfs2_per_slot_free_list *f_next_suballocator;
6185 struct ocfs2_cached_block_free *f_first;
6188 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6191 struct ocfs2_cached_block_free *head)
6196 struct inode *inode;
6197 struct buffer_head *di_bh = NULL;
6198 struct ocfs2_cached_block_free *tmp;
6200 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6207 mutex_lock(&inode->i_mutex);
6209 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6215 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6216 if (IS_ERR(handle)) {
6217 ret = PTR_ERR(handle);
6223 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6225 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
6226 head->free_bit, (unsigned long long)head->free_blk);
6228 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6229 head->free_bit, bg_blkno, 1);
6235 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6242 head = head->free_next;
6247 ocfs2_commit_trans(osb, handle);
6250 ocfs2_inode_unlock(inode, 1);
6253 mutex_unlock(&inode->i_mutex);
6257 /* Premature exit may have left some dangling items. */
6259 head = head->free_next;
6266 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6267 u64 blkno, unsigned int bit)
6270 struct ocfs2_cached_block_free *item;
6272 item = kmalloc(sizeof(*item), GFP_NOFS);
6279 mlog(0, "Insert clusters: (bit %u, blk %llu)\n",
6280 bit, (unsigned long long)blkno);
6282 item->free_blk = blkno;
6283 item->free_bit = bit;
6284 item->free_next = ctxt->c_global_allocator;
6286 ctxt->c_global_allocator = item;
6290 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6291 struct ocfs2_cached_block_free *head)
6293 struct ocfs2_cached_block_free *tmp;
6294 struct inode *tl_inode = osb->osb_tl_inode;
6298 mutex_lock(&tl_inode->i_mutex);
6301 if (ocfs2_truncate_log_needs_flush(osb)) {
6302 ret = __ocfs2_flush_truncate_log(osb);
6309 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6310 if (IS_ERR(handle)) {
6311 ret = PTR_ERR(handle);
6316 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6319 ocfs2_commit_trans(osb, handle);
6321 head = head->free_next;
6330 mutex_unlock(&tl_inode->i_mutex);
6333 /* Premature exit may have left some dangling items. */
6335 head = head->free_next;
6342 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6343 struct ocfs2_cached_dealloc_ctxt *ctxt)
6346 struct ocfs2_per_slot_free_list *fl;
6351 while (ctxt->c_first_suballocator) {
6352 fl = ctxt->c_first_suballocator;
6355 mlog(0, "Free items: (type %u, slot %d)\n",
6356 fl->f_inode_type, fl->f_slot);
6357 ret2 = ocfs2_free_cached_blocks(osb,
6367 ctxt->c_first_suballocator = fl->f_next_suballocator;
6371 if (ctxt->c_global_allocator) {
6372 ret2 = ocfs2_free_cached_clusters(osb,
6373 ctxt->c_global_allocator);
6379 ctxt->c_global_allocator = NULL;
6385 static struct ocfs2_per_slot_free_list *
6386 ocfs2_find_per_slot_free_list(int type,
6388 struct ocfs2_cached_dealloc_ctxt *ctxt)
6390 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6393 if (fl->f_inode_type == type && fl->f_slot == slot)
6396 fl = fl->f_next_suballocator;
6399 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6401 fl->f_inode_type = type;
6404 fl->f_next_suballocator = ctxt->c_first_suballocator;
6406 ctxt->c_first_suballocator = fl;
6411 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6412 int type, int slot, u64 blkno,
6416 struct ocfs2_per_slot_free_list *fl;
6417 struct ocfs2_cached_block_free *item;
6419 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6426 item = kmalloc(sizeof(*item), GFP_NOFS);
6433 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
6434 type, slot, bit, (unsigned long long)blkno);
6436 item->free_blk = blkno;
6437 item->free_bit = bit;
6438 item->free_next = fl->f_first;
6447 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6448 struct ocfs2_extent_block *eb)
6450 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6451 le16_to_cpu(eb->h_suballoc_slot),
6452 le64_to_cpu(eb->h_blkno),
6453 le16_to_cpu(eb->h_suballoc_bit));
6456 /* This function will figure out whether the currently last extent
6457 * block will be deleted, and if it will, what the new last extent
6458 * block will be so we can update his h_next_leaf_blk field, as well
6459 * as the dinodes i_last_eb_blk */
6460 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
6461 unsigned int clusters_to_del,
6462 struct ocfs2_path *path,
6463 struct buffer_head **new_last_eb)
6465 int next_free, ret = 0;
6467 struct ocfs2_extent_rec *rec;
6468 struct ocfs2_extent_block *eb;
6469 struct ocfs2_extent_list *el;
6470 struct buffer_head *bh = NULL;
6472 *new_last_eb = NULL;
6474 /* we have no tree, so of course, no last_eb. */
6475 if (!path->p_tree_depth)
6478 /* trunc to zero special case - this makes tree_depth = 0
6479 * regardless of what it is. */
6480 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
6483 el = path_leaf_el(path);
6484 BUG_ON(!el->l_next_free_rec);
6487 * Make sure that this extent list will actually be empty
6488 * after we clear away the data. We can shortcut out if
6489 * there's more than one non-empty extent in the
6490 * list. Otherwise, a check of the remaining extent is
6493 next_free = le16_to_cpu(el->l_next_free_rec);
6495 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6499 /* We may have a valid extent in index 1, check it. */
6501 rec = &el->l_recs[1];
6504 * Fall through - no more nonempty extents, so we want
6505 * to delete this leaf.
6511 rec = &el->l_recs[0];
6516 * Check it we'll only be trimming off the end of this
6519 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
6523 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
6529 ret = ocfs2_find_leaf(INODE_CACHE(inode), path_root_el(path), cpos, &bh);
6535 eb = (struct ocfs2_extent_block *) bh->b_data;
6538 /* ocfs2_find_leaf() gets the eb from ocfs2_read_extent_block().
6539 * Any corruption is a code bug. */
6540 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
6543 get_bh(*new_last_eb);
6544 mlog(0, "returning block %llu, (cpos: %u)\n",
6545 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
6553 * Trim some clusters off the rightmost edge of a tree. Only called
6556 * The caller needs to:
6557 * - start journaling of each path component.
6558 * - compute and fully set up any new last ext block
6560 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
6561 handle_t *handle, struct ocfs2_truncate_context *tc,
6562 u32 clusters_to_del, u64 *delete_start)
6564 int ret, i, index = path->p_tree_depth;
6567 struct buffer_head *bh;
6568 struct ocfs2_extent_list *el;
6569 struct ocfs2_extent_rec *rec;
6573 while (index >= 0) {
6574 bh = path->p_node[index].bh;
6575 el = path->p_node[index].el;
6577 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6578 index, (unsigned long long)bh->b_blocknr);
6580 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
6583 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
6584 ocfs2_error(inode->i_sb,
6585 "Inode %lu has invalid ext. block %llu",
6587 (unsigned long long)bh->b_blocknr);
6593 i = le16_to_cpu(el->l_next_free_rec) - 1;
6594 rec = &el->l_recs[i];
6596 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6597 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
6598 ocfs2_rec_clusters(el, rec),
6599 (unsigned long long)le64_to_cpu(rec->e_blkno),
6600 le16_to_cpu(el->l_next_free_rec));
6602 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
6604 if (le16_to_cpu(el->l_tree_depth) == 0) {
6606 * If the leaf block contains a single empty
6607 * extent and no records, we can just remove
6610 if (i == 0 && ocfs2_is_empty_extent(rec)) {
6612 sizeof(struct ocfs2_extent_rec));
6613 el->l_next_free_rec = cpu_to_le16(0);
6619 * Remove any empty extents by shifting things
6620 * left. That should make life much easier on
6621 * the code below. This condition is rare
6622 * enough that we shouldn't see a performance
6625 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6626 le16_add_cpu(&el->l_next_free_rec, -1);
6629 i < le16_to_cpu(el->l_next_free_rec); i++)
6630 el->l_recs[i] = el->l_recs[i + 1];
6632 memset(&el->l_recs[i], 0,
6633 sizeof(struct ocfs2_extent_rec));
6636 * We've modified our extent list. The
6637 * simplest way to handle this change
6638 * is to being the search from the
6641 goto find_tail_record;
6644 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
6647 * We'll use "new_edge" on our way back up the
6648 * tree to know what our rightmost cpos is.
6650 new_edge = le16_to_cpu(rec->e_leaf_clusters);
6651 new_edge += le32_to_cpu(rec->e_cpos);
6654 * The caller will use this to delete data blocks.
6656 *delete_start = le64_to_cpu(rec->e_blkno)
6657 + ocfs2_clusters_to_blocks(inode->i_sb,
6658 le16_to_cpu(rec->e_leaf_clusters));
6661 * If it's now empty, remove this record.
6663 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
6665 sizeof(struct ocfs2_extent_rec));
6666 le16_add_cpu(&el->l_next_free_rec, -1);
6669 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
6671 sizeof(struct ocfs2_extent_rec));
6672 le16_add_cpu(&el->l_next_free_rec, -1);
6677 /* Can this actually happen? */
6678 if (le16_to_cpu(el->l_next_free_rec) == 0)
6682 * We never actually deleted any clusters
6683 * because our leaf was empty. There's no
6684 * reason to adjust the rightmost edge then.
6689 rec->e_int_clusters = cpu_to_le32(new_edge);
6690 le32_add_cpu(&rec->e_int_clusters,
6691 -le32_to_cpu(rec->e_cpos));
6694 * A deleted child record should have been
6697 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
6701 ret = ocfs2_journal_dirty(handle, bh);
6707 mlog(0, "extent list container %llu, after: record %d: "
6708 "(%u, %u, %llu), next = %u.\n",
6709 (unsigned long long)bh->b_blocknr, i,
6710 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
6711 (unsigned long long)le64_to_cpu(rec->e_blkno),
6712 le16_to_cpu(el->l_next_free_rec));
6715 * We must be careful to only attempt delete of an
6716 * extent block (and not the root inode block).
6718 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
6719 struct ocfs2_extent_block *eb =
6720 (struct ocfs2_extent_block *)bh->b_data;
6723 * Save this for use when processing the
6726 deleted_eb = le64_to_cpu(eb->h_blkno);
6728 mlog(0, "deleting this extent block.\n");
6730 ocfs2_remove_from_cache(INODE_CACHE(inode), bh);
6732 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
6733 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
6734 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
6736 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
6737 /* An error here is not fatal. */
6752 static int ocfs2_do_truncate(struct ocfs2_super *osb,
6753 unsigned int clusters_to_del,
6754 struct inode *inode,
6755 struct buffer_head *fe_bh,
6757 struct ocfs2_truncate_context *tc,
6758 struct ocfs2_path *path)
6761 struct ocfs2_dinode *fe;
6762 struct ocfs2_extent_block *last_eb = NULL;
6763 struct ocfs2_extent_list *el;
6764 struct buffer_head *last_eb_bh = NULL;
6767 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6769 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
6777 * Each component will be touched, so we might as well journal
6778 * here to avoid having to handle errors later.
6780 status = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
6787 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), last_eb_bh,
6788 OCFS2_JOURNAL_ACCESS_WRITE);
6794 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6797 el = &(fe->id2.i_list);
6800 * Lower levels depend on this never happening, but it's best
6801 * to check it up here before changing the tree.
6803 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
6804 ocfs2_error(inode->i_sb,
6805 "Inode %lu has an empty extent record, depth %u\n",
6806 inode->i_ino, le16_to_cpu(el->l_tree_depth));
6811 vfs_dq_free_space_nodirty(inode,
6812 ocfs2_clusters_to_bytes(osb->sb, clusters_to_del));
6813 spin_lock(&OCFS2_I(inode)->ip_lock);
6814 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
6816 spin_unlock(&OCFS2_I(inode)->ip_lock);
6817 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
6818 inode->i_blocks = ocfs2_inode_sector_count(inode);
6820 status = ocfs2_trim_tree(inode, path, handle, tc,
6821 clusters_to_del, &delete_blk);
6827 if (le32_to_cpu(fe->i_clusters) == 0) {
6828 /* trunc to zero is a special case. */
6829 el->l_tree_depth = 0;
6830 fe->i_last_eb_blk = 0;
6832 fe->i_last_eb_blk = last_eb->h_blkno;
6834 status = ocfs2_journal_dirty(handle, fe_bh);
6841 /* If there will be a new last extent block, then by
6842 * definition, there cannot be any leaves to the right of
6844 last_eb->h_next_leaf_blk = 0;
6845 status = ocfs2_journal_dirty(handle, last_eb_bh);
6853 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
6867 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6869 set_buffer_uptodate(bh);
6870 mark_buffer_dirty(bh);
6874 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6875 unsigned int from, unsigned int to,
6876 struct page *page, int zero, u64 *phys)
6878 int ret, partial = 0;
6880 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6885 zero_user_segment(page, from, to);
6888 * Need to set the buffers we zero'd into uptodate
6889 * here if they aren't - ocfs2_map_page_blocks()
6890 * might've skipped some
6892 ret = walk_page_buffers(handle, page_buffers(page),
6897 else if (ocfs2_should_order_data(inode)) {
6898 ret = ocfs2_jbd2_file_inode(handle, inode);
6904 SetPageUptodate(page);
6906 flush_dcache_page(page);
6909 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6910 loff_t end, struct page **pages,
6911 int numpages, u64 phys, handle_t *handle)
6915 unsigned int from, to = PAGE_CACHE_SIZE;
6916 struct super_block *sb = inode->i_sb;
6918 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6923 to = PAGE_CACHE_SIZE;
6924 for(i = 0; i < numpages; i++) {
6927 from = start & (PAGE_CACHE_SIZE - 1);
6928 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6929 to = end & (PAGE_CACHE_SIZE - 1);
6931 BUG_ON(from > PAGE_CACHE_SIZE);
6932 BUG_ON(to > PAGE_CACHE_SIZE);
6934 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6937 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6941 ocfs2_unlock_and_free_pages(pages, numpages);
6944 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6945 struct page **pages, int *num)
6947 int numpages, ret = 0;
6948 struct super_block *sb = inode->i_sb;
6949 struct address_space *mapping = inode->i_mapping;
6950 unsigned long index;
6951 loff_t last_page_bytes;
6953 BUG_ON(start > end);
6955 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6956 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6959 last_page_bytes = PAGE_ALIGN(end);
6960 index = start >> PAGE_CACHE_SHIFT;
6962 pages[numpages] = grab_cache_page(mapping, index);
6963 if (!pages[numpages]) {
6971 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6976 ocfs2_unlock_and_free_pages(pages, numpages);
6986 * Zero the area past i_size but still within an allocated
6987 * cluster. This avoids exposing nonzero data on subsequent file
6990 * We need to call this before i_size is updated on the inode because
6991 * otherwise block_write_full_page() will skip writeout of pages past
6992 * i_size. The new_i_size parameter is passed for this reason.
6994 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6995 u64 range_start, u64 range_end)
6997 int ret = 0, numpages;
6998 struct page **pages = NULL;
7000 unsigned int ext_flags;
7001 struct super_block *sb = inode->i_sb;
7004 * File systems which don't support sparse files zero on every
7007 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
7010 pages = kcalloc(ocfs2_pages_per_cluster(sb),
7011 sizeof(struct page *), GFP_NOFS);
7012 if (pages == NULL) {
7018 if (range_start == range_end)
7021 ret = ocfs2_extent_map_get_blocks(inode,
7022 range_start >> sb->s_blocksize_bits,
7023 &phys, NULL, &ext_flags);
7030 * Tail is a hole, or is marked unwritten. In either case, we
7031 * can count on read and write to return/push zero's.
7033 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
7036 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
7043 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
7044 numpages, phys, handle);
7047 * Initiate writeout of the pages we zero'd here. We don't
7048 * wait on them - the truncate_inode_pages() call later will
7051 ret = do_sync_mapping_range(inode->i_mapping, range_start,
7052 range_end - 1, SYNC_FILE_RANGE_WRITE);
7063 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7064 struct ocfs2_dinode *di)
7066 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7067 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7069 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7070 memset(&di->id2, 0, blocksize -
7071 offsetof(struct ocfs2_dinode, id2) -
7074 memset(&di->id2, 0, blocksize -
7075 offsetof(struct ocfs2_dinode, id2));
7078 void ocfs2_dinode_new_extent_list(struct inode *inode,
7079 struct ocfs2_dinode *di)
7081 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7082 di->id2.i_list.l_tree_depth = 0;
7083 di->id2.i_list.l_next_free_rec = 0;
7084 di->id2.i_list.l_count = cpu_to_le16(
7085 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7088 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7090 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7091 struct ocfs2_inline_data *idata = &di->id2.i_data;
7093 spin_lock(&oi->ip_lock);
7094 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7095 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7096 spin_unlock(&oi->ip_lock);
7099 * We clear the entire i_data structure here so that all
7100 * fields can be properly initialized.
7102 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7104 idata->id_count = cpu_to_le16(
7105 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7108 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7109 struct buffer_head *di_bh)
7111 int ret, i, has_data, num_pages = 0;
7113 u64 uninitialized_var(block);
7114 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7115 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7116 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7117 struct ocfs2_alloc_context *data_ac = NULL;
7118 struct page **pages = NULL;
7119 loff_t end = osb->s_clustersize;
7120 struct ocfs2_extent_tree et;
7123 has_data = i_size_read(inode) ? 1 : 0;
7126 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
7127 sizeof(struct page *), GFP_NOFS);
7128 if (pages == NULL) {
7134 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7141 handle = ocfs2_start_trans(osb,
7142 ocfs2_inline_to_extents_credits(osb->sb));
7143 if (IS_ERR(handle)) {
7144 ret = PTR_ERR(handle);
7149 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7150 OCFS2_JOURNAL_ACCESS_WRITE);
7158 unsigned int page_end;
7161 if (vfs_dq_alloc_space_nodirty(inode,
7162 ocfs2_clusters_to_bytes(osb->sb, 1))) {
7168 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
7176 * Save two copies, one for insert, and one that can
7177 * be changed by ocfs2_map_and_dirty_page() below.
7179 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7182 * Non sparse file systems zero on extend, so no need
7185 if (!ocfs2_sparse_alloc(osb) &&
7186 PAGE_CACHE_SIZE < osb->s_clustersize)
7187 end = PAGE_CACHE_SIZE;
7189 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
7196 * This should populate the 1st page for us and mark
7199 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
7205 page_end = PAGE_CACHE_SIZE;
7206 if (PAGE_CACHE_SIZE > osb->s_clustersize)
7207 page_end = osb->s_clustersize;
7209 for (i = 0; i < num_pages; i++)
7210 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
7211 pages[i], i > 0, &phys);
7214 spin_lock(&oi->ip_lock);
7215 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7216 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7217 spin_unlock(&oi->ip_lock);
7219 ocfs2_dinode_new_extent_list(inode, di);
7221 ocfs2_journal_dirty(handle, di_bh);
7225 * An error at this point should be extremely rare. If
7226 * this proves to be false, we could always re-build
7227 * the in-inode data from our pages.
7229 ocfs2_init_dinode_extent_tree(&et, inode, di_bh);
7230 ret = ocfs2_insert_extent(osb, handle, inode, &et,
7231 0, block, 1, 0, NULL);
7237 inode->i_blocks = ocfs2_inode_sector_count(inode);
7241 if (ret < 0 && did_quota)
7242 vfs_dq_free_space_nodirty(inode,
7243 ocfs2_clusters_to_bytes(osb->sb, 1));
7245 ocfs2_commit_trans(osb, handle);
7249 ocfs2_free_alloc_context(data_ac);
7253 ocfs2_unlock_and_free_pages(pages, num_pages);
7261 * It is expected, that by the time you call this function,
7262 * inode->i_size and fe->i_size have been adjusted.
7264 * WARNING: This will kfree the truncate context
7266 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7267 struct inode *inode,
7268 struct buffer_head *fe_bh,
7269 struct ocfs2_truncate_context *tc)
7271 int status, i, credits, tl_sem = 0;
7272 u32 clusters_to_del, new_highest_cpos, range;
7273 struct ocfs2_extent_list *el;
7274 handle_t *handle = NULL;
7275 struct inode *tl_inode = osb->osb_tl_inode;
7276 struct ocfs2_path *path = NULL;
7277 struct ocfs2_dinode *di = (struct ocfs2_dinode *)fe_bh->b_data;
7281 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7282 i_size_read(inode));
7284 path = ocfs2_new_path(fe_bh, &di->id2.i_list,
7285 ocfs2_journal_access_di);
7292 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7296 * Check that we still have allocation to delete.
7298 if (OCFS2_I(inode)->ip_clusters == 0) {
7304 * Truncate always works against the rightmost tree branch.
7306 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7312 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
7313 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
7316 * By now, el will point to the extent list on the bottom most
7317 * portion of this tree. Only the tail record is considered in
7320 * We handle the following cases, in order:
7321 * - empty extent: delete the remaining branch
7322 * - remove the entire record
7323 * - remove a partial record
7324 * - no record needs to be removed (truncate has completed)
7326 el = path_leaf_el(path);
7327 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7328 ocfs2_error(inode->i_sb,
7329 "Inode %llu has empty extent block at %llu\n",
7330 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7331 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7336 i = le16_to_cpu(el->l_next_free_rec) - 1;
7337 range = le32_to_cpu(el->l_recs[i].e_cpos) +
7338 ocfs2_rec_clusters(el, &el->l_recs[i]);
7339 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
7340 clusters_to_del = 0;
7341 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
7342 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
7343 } else if (range > new_highest_cpos) {
7344 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
7345 le32_to_cpu(el->l_recs[i].e_cpos)) -
7352 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
7353 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
7355 mutex_lock(&tl_inode->i_mutex);
7357 /* ocfs2_truncate_log_needs_flush guarantees us at least one
7358 * record is free for use. If there isn't any, we flush to get
7359 * an empty truncate log. */
7360 if (ocfs2_truncate_log_needs_flush(osb)) {
7361 status = __ocfs2_flush_truncate_log(osb);
7368 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
7369 (struct ocfs2_dinode *)fe_bh->b_data,
7371 handle = ocfs2_start_trans(osb, credits);
7372 if (IS_ERR(handle)) {
7373 status = PTR_ERR(handle);
7379 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
7386 mutex_unlock(&tl_inode->i_mutex);
7389 ocfs2_commit_trans(osb, handle);
7392 ocfs2_reinit_path(path, 1);
7395 * The check above will catch the case where we've truncated
7396 * away all allocation.
7402 ocfs2_schedule_truncate_log_flush(osb, 1);
7405 mutex_unlock(&tl_inode->i_mutex);
7408 ocfs2_commit_trans(osb, handle);
7410 ocfs2_run_deallocs(osb, &tc->tc_dealloc);
7412 ocfs2_free_path(path);
7414 /* This will drop the ext_alloc cluster lock for us */
7415 ocfs2_free_truncate_context(tc);
7422 * Expects the inode to already be locked.
7424 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
7425 struct inode *inode,
7426 struct buffer_head *fe_bh,
7427 struct ocfs2_truncate_context **tc)
7430 unsigned int new_i_clusters;
7431 struct ocfs2_dinode *fe;
7432 struct ocfs2_extent_block *eb;
7433 struct buffer_head *last_eb_bh = NULL;
7439 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
7440 i_size_read(inode));
7441 fe = (struct ocfs2_dinode *) fe_bh->b_data;
7443 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
7444 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
7445 (unsigned long long)le64_to_cpu(fe->i_size));
7447 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
7453 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
7455 if (fe->id2.i_list.l_tree_depth) {
7456 status = ocfs2_read_extent_block(INODE_CACHE(inode),
7457 le64_to_cpu(fe->i_last_eb_blk),
7463 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
7466 (*tc)->tc_last_eb_bh = last_eb_bh;
7472 ocfs2_free_truncate_context(*tc);
7480 * 'start' is inclusive, 'end' is not.
7482 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7483 unsigned int start, unsigned int end, int trunc)
7486 unsigned int numbytes;
7488 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7489 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7490 struct ocfs2_inline_data *idata = &di->id2.i_data;
7492 if (end > i_size_read(inode))
7493 end = i_size_read(inode);
7495 BUG_ON(start >= end);
7497 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7498 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7499 !ocfs2_supports_inline_data(osb)) {
7500 ocfs2_error(inode->i_sb,
7501 "Inline data flags for inode %llu don't agree! "
7502 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7503 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7504 le16_to_cpu(di->i_dyn_features),
7505 OCFS2_I(inode)->ip_dyn_features,
7506 osb->s_feature_incompat);
7511 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7512 if (IS_ERR(handle)) {
7513 ret = PTR_ERR(handle);
7518 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7519 OCFS2_JOURNAL_ACCESS_WRITE);
7525 numbytes = end - start;
7526 memset(idata->id_data + start, 0, numbytes);
7529 * No need to worry about the data page here - it's been
7530 * truncated already and inline data doesn't need it for
7531 * pushing zero's to disk, so we'll let readpage pick it up
7535 i_size_write(inode, start);
7536 di->i_size = cpu_to_le64(start);
7539 inode->i_blocks = ocfs2_inode_sector_count(inode);
7540 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7542 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7543 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7545 ocfs2_journal_dirty(handle, di_bh);
7548 ocfs2_commit_trans(osb, handle);
7554 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
7557 * The caller is responsible for completing deallocation
7558 * before freeing the context.
7560 if (tc->tc_dealloc.c_first_suballocator != NULL)
7562 "Truncate completion has non-empty dealloc context\n");
7564 brelse(tc->tc_last_eb_bh);