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_root_journal_access = access;
357 obj = (void *)bh->b_data;
360 et->et_ops->eo_fill_root_el(et);
361 if (!et->et_ops->eo_fill_max_leaf_clusters)
362 et->et_max_leaf_clusters = 0;
364 et->et_ops->eo_fill_max_leaf_clusters(inode, et);
367 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
369 struct buffer_head *bh)
371 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_di,
372 NULL, &ocfs2_dinode_et_ops);
375 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
377 struct buffer_head *bh)
379 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_xb,
380 NULL, &ocfs2_xattr_tree_et_ops);
383 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
385 struct ocfs2_xattr_value_buf *vb)
387 __ocfs2_init_extent_tree(et, inode, vb->vb_bh, vb->vb_access, vb,
388 &ocfs2_xattr_value_et_ops);
391 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
393 struct buffer_head *bh)
395 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_dr,
396 NULL, &ocfs2_dx_root_et_ops);
399 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
402 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
405 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
407 return et->et_ops->eo_get_last_eb_blk(et);
410 static inline void ocfs2_et_update_clusters(struct inode *inode,
411 struct ocfs2_extent_tree *et,
414 et->et_ops->eo_update_clusters(inode, et, clusters);
417 static inline int ocfs2_et_root_journal_access(handle_t *handle,
418 struct ocfs2_caching_info *ci,
419 struct ocfs2_extent_tree *et,
422 return et->et_root_journal_access(handle, 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 inode *inode, struct ocfs2_path *path,
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 inode *inode, u64 eb_blkno,
858 struct buffer_head **bh)
861 struct buffer_head *tmp = *bh;
863 rc = ocfs2_read_block(INODE_CACHE(inode), 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,
879 struct ocfs2_extent_tree *et)
882 struct ocfs2_extent_list *el = NULL;
883 struct ocfs2_extent_block *eb;
884 struct buffer_head *eb_bh = NULL;
890 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
893 retval = ocfs2_read_extent_block(inode, last_eb_blk, &eb_bh);
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(struct ocfs2_super *osb,
921 struct ocfs2_alloc_context *meta_ac,
922 struct buffer_head *bhs[])
924 int count, status, i;
925 u16 suballoc_bit_start;
928 struct ocfs2_extent_block *eb;
933 while (count < wanted) {
934 status = ocfs2_claim_metadata(osb,
946 for(i = count; i < (num_got + count); i++) {
947 bhs[i] = sb_getblk(osb->sb, first_blkno);
948 if (bhs[i] == NULL) {
953 ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode),
956 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), bhs[i],
957 OCFS2_JOURNAL_ACCESS_CREATE);
963 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
964 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
965 /* Ok, setup the minimal stuff here. */
966 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
967 eb->h_blkno = cpu_to_le64(first_blkno);
968 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
969 eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
970 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
972 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
974 suballoc_bit_start++;
977 /* We'll also be dirtied by the caller, so
978 * this isn't absolutely necessary. */
979 status = ocfs2_journal_dirty(handle, bhs[i]);
992 for(i = 0; i < wanted; i++) {
1002 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1004 * Returns the sum of the rightmost extent rec logical offset and
1007 * ocfs2_add_branch() uses this to determine what logical cluster
1008 * value should be populated into the leftmost new branch records.
1010 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1011 * value for the new topmost tree record.
1013 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1017 i = le16_to_cpu(el->l_next_free_rec) - 1;
1019 return le32_to_cpu(el->l_recs[i].e_cpos) +
1020 ocfs2_rec_clusters(el, &el->l_recs[i]);
1024 * Change range of the branches in the right most path according to the leaf
1025 * extent block's rightmost record.
1027 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1028 struct inode *inode,
1029 struct ocfs2_extent_tree *et)
1032 struct ocfs2_path *path = NULL;
1033 struct ocfs2_extent_list *el;
1034 struct ocfs2_extent_rec *rec;
1036 path = ocfs2_new_path_from_et(et);
1042 status = ocfs2_find_path(inode, path, UINT_MAX);
1048 status = ocfs2_extend_trans(handle, path_num_items(path) +
1049 handle->h_buffer_credits);
1055 status = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
1061 el = path_leaf_el(path);
1062 rec = &el->l_recs[le32_to_cpu(el->l_next_free_rec) - 1];
1064 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
1067 ocfs2_free_path(path);
1072 * Add an entire tree branch to our inode. eb_bh is the extent block
1073 * to start at, if we don't want to start the branch at the dinode
1076 * last_eb_bh is required as we have to update it's next_leaf pointer
1077 * for the new last extent block.
1079 * the new branch will be 'empty' in the sense that every block will
1080 * contain a single record with cluster count == 0.
1082 static int ocfs2_add_branch(struct ocfs2_super *osb,
1084 struct inode *inode,
1085 struct ocfs2_extent_tree *et,
1086 struct buffer_head *eb_bh,
1087 struct buffer_head **last_eb_bh,
1088 struct ocfs2_alloc_context *meta_ac)
1090 int status, new_blocks, i;
1091 u64 next_blkno, new_last_eb_blk;
1092 struct buffer_head *bh;
1093 struct buffer_head **new_eb_bhs = NULL;
1094 struct ocfs2_extent_block *eb;
1095 struct ocfs2_extent_list *eb_el;
1096 struct ocfs2_extent_list *el;
1097 u32 new_cpos, root_end;
1101 BUG_ON(!last_eb_bh || !*last_eb_bh);
1104 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1107 el = et->et_root_el;
1109 /* we never add a branch to a leaf. */
1110 BUG_ON(!el->l_tree_depth);
1112 new_blocks = le16_to_cpu(el->l_tree_depth);
1114 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1115 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1116 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1119 * If there is a gap before the root end and the real end
1120 * of the righmost leaf block, we need to remove the gap
1121 * between new_cpos and root_end first so that the tree
1122 * is consistent after we add a new branch(it will start
1125 if (root_end > new_cpos) {
1126 mlog(0, "adjust the cluster end from %u to %u\n",
1127 root_end, new_cpos);
1128 status = ocfs2_adjust_rightmost_branch(handle, inode, et);
1135 /* allocate the number of new eb blocks we need */
1136 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1144 status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
1145 meta_ac, new_eb_bhs);
1151 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1152 * linked with the rest of the tree.
1153 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1155 * when we leave the loop, new_last_eb_blk will point to the
1156 * newest leaf, and next_blkno will point to the topmost extent
1158 next_blkno = new_last_eb_blk = 0;
1159 for(i = 0; i < new_blocks; i++) {
1161 eb = (struct ocfs2_extent_block *) bh->b_data;
1162 /* ocfs2_create_new_meta_bhs() should create it right! */
1163 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1164 eb_el = &eb->h_list;
1166 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), bh,
1167 OCFS2_JOURNAL_ACCESS_CREATE);
1173 eb->h_next_leaf_blk = 0;
1174 eb_el->l_tree_depth = cpu_to_le16(i);
1175 eb_el->l_next_free_rec = cpu_to_le16(1);
1177 * This actually counts as an empty extent as
1180 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1181 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1183 * eb_el isn't always an interior node, but even leaf
1184 * nodes want a zero'd flags and reserved field so
1185 * this gets the whole 32 bits regardless of use.
1187 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1188 if (!eb_el->l_tree_depth)
1189 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1191 status = ocfs2_journal_dirty(handle, bh);
1197 next_blkno = le64_to_cpu(eb->h_blkno);
1200 /* This is a bit hairy. We want to update up to three blocks
1201 * here without leaving any of them in an inconsistent state
1202 * in case of error. We don't have to worry about
1203 * journal_dirty erroring as it won't unless we've aborted the
1204 * handle (in which case we would never be here) so reserving
1205 * the write with journal_access is all we need to do. */
1206 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), *last_eb_bh,
1207 OCFS2_JOURNAL_ACCESS_WRITE);
1212 status = ocfs2_et_root_journal_access(handle, INODE_CACHE(inode), et,
1213 OCFS2_JOURNAL_ACCESS_WRITE);
1219 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), eb_bh,
1220 OCFS2_JOURNAL_ACCESS_WRITE);
1227 /* Link the new branch into the rest of the tree (el will
1228 * either be on the root_bh, or the extent block passed in. */
1229 i = le16_to_cpu(el->l_next_free_rec);
1230 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1231 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1232 el->l_recs[i].e_int_clusters = 0;
1233 le16_add_cpu(&el->l_next_free_rec, 1);
1235 /* fe needs a new last extent block pointer, as does the
1236 * next_leaf on the previously last-extent-block. */
1237 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1239 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1240 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1242 status = ocfs2_journal_dirty(handle, *last_eb_bh);
1245 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1249 status = ocfs2_journal_dirty(handle, eb_bh);
1255 * Some callers want to track the rightmost leaf so pass it
1258 brelse(*last_eb_bh);
1259 get_bh(new_eb_bhs[0]);
1260 *last_eb_bh = new_eb_bhs[0];
1265 for (i = 0; i < new_blocks; i++)
1266 brelse(new_eb_bhs[i]);
1275 * adds another level to the allocation tree.
1276 * returns back the new extent block so you can add a branch to it
1279 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
1281 struct inode *inode,
1282 struct ocfs2_extent_tree *et,
1283 struct ocfs2_alloc_context *meta_ac,
1284 struct buffer_head **ret_new_eb_bh)
1288 struct buffer_head *new_eb_bh = NULL;
1289 struct ocfs2_extent_block *eb;
1290 struct ocfs2_extent_list *root_el;
1291 struct ocfs2_extent_list *eb_el;
1295 status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
1302 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1303 /* ocfs2_create_new_meta_bhs() should create it right! */
1304 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1306 eb_el = &eb->h_list;
1307 root_el = et->et_root_el;
1309 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), new_eb_bh,
1310 OCFS2_JOURNAL_ACCESS_CREATE);
1316 /* copy the root extent list data into the new extent block */
1317 eb_el->l_tree_depth = root_el->l_tree_depth;
1318 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1319 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1320 eb_el->l_recs[i] = root_el->l_recs[i];
1322 status = ocfs2_journal_dirty(handle, new_eb_bh);
1328 status = ocfs2_et_root_journal_access(handle, INODE_CACHE(inode), et,
1329 OCFS2_JOURNAL_ACCESS_WRITE);
1335 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1337 /* update root_bh now */
1338 le16_add_cpu(&root_el->l_tree_depth, 1);
1339 root_el->l_recs[0].e_cpos = 0;
1340 root_el->l_recs[0].e_blkno = eb->h_blkno;
1341 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1342 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1343 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1344 root_el->l_next_free_rec = cpu_to_le16(1);
1346 /* If this is our 1st tree depth shift, then last_eb_blk
1347 * becomes the allocated extent block */
1348 if (root_el->l_tree_depth == cpu_to_le16(1))
1349 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1351 status = ocfs2_journal_dirty(handle, et->et_root_bh);
1357 *ret_new_eb_bh = new_eb_bh;
1368 * Should only be called when there is no space left in any of the
1369 * leaf nodes. What we want to do is find the lowest tree depth
1370 * non-leaf extent block with room for new records. There are three
1371 * valid results of this search:
1373 * 1) a lowest extent block is found, then we pass it back in
1374 * *lowest_eb_bh and return '0'
1376 * 2) the search fails to find anything, but the root_el has room. We
1377 * pass NULL back in *lowest_eb_bh, but still return '0'
1379 * 3) the search fails to find anything AND the root_el is full, in
1380 * which case we return > 0
1382 * return status < 0 indicates an error.
1384 static int ocfs2_find_branch_target(struct ocfs2_super *osb,
1385 struct inode *inode,
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(inode->i_sb, "Dinode %llu has empty "
1405 "extent list (next_free_rec == 0)",
1406 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1410 i = le16_to_cpu(el->l_next_free_rec) - 1;
1411 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1413 ocfs2_error(inode->i_sb, "Dinode %llu has extent "
1414 "list where extent # %d has no physical "
1416 (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
1424 status = ocfs2_read_extent_block(inode, blkno, &bh);
1430 eb = (struct ocfs2_extent_block *) bh->b_data;
1433 if (le16_to_cpu(el->l_next_free_rec) <
1434 le16_to_cpu(el->l_count)) {
1441 /* If we didn't find one and the fe doesn't have any room,
1442 * then return '1' */
1443 el = et->et_root_el;
1444 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1447 *target_bh = lowest_bh;
1456 * Grow a b-tree so that it has more records.
1458 * We might shift the tree depth in which case existing paths should
1459 * be considered invalid.
1461 * Tree depth after the grow is returned via *final_depth.
1463 * *last_eb_bh will be updated by ocfs2_add_branch().
1465 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
1466 struct ocfs2_extent_tree *et, int *final_depth,
1467 struct buffer_head **last_eb_bh,
1468 struct ocfs2_alloc_context *meta_ac)
1471 struct ocfs2_extent_list *el = et->et_root_el;
1472 int depth = le16_to_cpu(el->l_tree_depth);
1473 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1474 struct buffer_head *bh = NULL;
1476 BUG_ON(meta_ac == NULL);
1478 shift = ocfs2_find_branch_target(osb, inode, et, &bh);
1485 /* We traveled all the way to the bottom of the allocation tree
1486 * and didn't find room for any more extents - we need to add
1487 * another tree level */
1490 mlog(0, "need to shift tree depth (current = %d)\n", depth);
1492 /* ocfs2_shift_tree_depth will return us a buffer with
1493 * the new extent block (so we can pass that to
1494 * ocfs2_add_branch). */
1495 ret = ocfs2_shift_tree_depth(osb, handle, inode, et,
1504 * Special case: we have room now if we shifted from
1505 * tree_depth 0, so no more work needs to be done.
1507 * We won't be calling add_branch, so pass
1508 * back *last_eb_bh as the new leaf. At depth
1509 * zero, it should always be null so there's
1510 * no reason to brelse.
1512 BUG_ON(*last_eb_bh);
1519 /* call ocfs2_add_branch to add the final part of the tree with
1521 mlog(0, "add branch. bh = %p\n", bh);
1522 ret = ocfs2_add_branch(osb, handle, inode, et, bh, last_eb_bh,
1531 *final_depth = depth;
1537 * This function will discard the rightmost extent record.
1539 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1541 int next_free = le16_to_cpu(el->l_next_free_rec);
1542 int count = le16_to_cpu(el->l_count);
1543 unsigned int num_bytes;
1546 /* This will cause us to go off the end of our extent list. */
1547 BUG_ON(next_free >= count);
1549 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1551 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1554 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1555 struct ocfs2_extent_rec *insert_rec)
1557 int i, insert_index, next_free, has_empty, num_bytes;
1558 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1559 struct ocfs2_extent_rec *rec;
1561 next_free = le16_to_cpu(el->l_next_free_rec);
1562 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1566 /* The tree code before us didn't allow enough room in the leaf. */
1567 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1570 * The easiest way to approach this is to just remove the
1571 * empty extent and temporarily decrement next_free.
1575 * If next_free was 1 (only an empty extent), this
1576 * loop won't execute, which is fine. We still want
1577 * the decrement above to happen.
1579 for(i = 0; i < (next_free - 1); i++)
1580 el->l_recs[i] = el->l_recs[i+1];
1586 * Figure out what the new record index should be.
1588 for(i = 0; i < next_free; i++) {
1589 rec = &el->l_recs[i];
1591 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1596 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1597 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
1599 BUG_ON(insert_index < 0);
1600 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1601 BUG_ON(insert_index > next_free);
1604 * No need to memmove if we're just adding to the tail.
1606 if (insert_index != next_free) {
1607 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1609 num_bytes = next_free - insert_index;
1610 num_bytes *= sizeof(struct ocfs2_extent_rec);
1611 memmove(&el->l_recs[insert_index + 1],
1612 &el->l_recs[insert_index],
1617 * Either we had an empty extent, and need to re-increment or
1618 * there was no empty extent on a non full rightmost leaf node,
1619 * in which case we still need to increment.
1622 el->l_next_free_rec = cpu_to_le16(next_free);
1624 * Make sure none of the math above just messed up our tree.
1626 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1628 el->l_recs[insert_index] = *insert_rec;
1632 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1634 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1636 BUG_ON(num_recs == 0);
1638 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1640 size = num_recs * sizeof(struct ocfs2_extent_rec);
1641 memmove(&el->l_recs[0], &el->l_recs[1], size);
1642 memset(&el->l_recs[num_recs], 0,
1643 sizeof(struct ocfs2_extent_rec));
1644 el->l_next_free_rec = cpu_to_le16(num_recs);
1649 * Create an empty extent record .
1651 * l_next_free_rec may be updated.
1653 * If an empty extent already exists do nothing.
1655 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1657 int next_free = le16_to_cpu(el->l_next_free_rec);
1659 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1664 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1667 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1668 "Asked to create an empty extent in a full list:\n"
1669 "count = %u, tree depth = %u",
1670 le16_to_cpu(el->l_count),
1671 le16_to_cpu(el->l_tree_depth));
1673 ocfs2_shift_records_right(el);
1676 le16_add_cpu(&el->l_next_free_rec, 1);
1677 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1681 * For a rotation which involves two leaf nodes, the "root node" is
1682 * the lowest level tree node which contains a path to both leafs. This
1683 * resulting set of information can be used to form a complete "subtree"
1685 * This function is passed two full paths from the dinode down to a
1686 * pair of adjacent leaves. It's task is to figure out which path
1687 * index contains the subtree root - this can be the root index itself
1688 * in a worst-case rotation.
1690 * The array index of the subtree root is passed back.
1692 static int ocfs2_find_subtree_root(struct inode *inode,
1693 struct ocfs2_path *left,
1694 struct ocfs2_path *right)
1699 * Check that the caller passed in two paths from the same tree.
1701 BUG_ON(path_root_bh(left) != path_root_bh(right));
1707 * The caller didn't pass two adjacent paths.
1709 mlog_bug_on_msg(i > left->p_tree_depth,
1710 "Inode %lu, left depth %u, right depth %u\n"
1711 "left leaf blk %llu, right leaf blk %llu\n",
1712 inode->i_ino, left->p_tree_depth,
1713 right->p_tree_depth,
1714 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1715 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1716 } while (left->p_node[i].bh->b_blocknr ==
1717 right->p_node[i].bh->b_blocknr);
1722 typedef void (path_insert_t)(void *, struct buffer_head *);
1725 * Traverse a btree path in search of cpos, starting at root_el.
1727 * This code can be called with a cpos larger than the tree, in which
1728 * case it will return the rightmost path.
1730 static int __ocfs2_find_path(struct inode *inode,
1731 struct ocfs2_extent_list *root_el, u32 cpos,
1732 path_insert_t *func, void *data)
1737 struct buffer_head *bh = NULL;
1738 struct ocfs2_extent_block *eb;
1739 struct ocfs2_extent_list *el;
1740 struct ocfs2_extent_rec *rec;
1741 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1744 while (el->l_tree_depth) {
1745 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1746 ocfs2_error(inode->i_sb,
1747 "Inode %llu has empty extent list at "
1749 (unsigned long long)oi->ip_blkno,
1750 le16_to_cpu(el->l_tree_depth));
1756 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1757 rec = &el->l_recs[i];
1760 * In the case that cpos is off the allocation
1761 * tree, this should just wind up returning the
1764 range = le32_to_cpu(rec->e_cpos) +
1765 ocfs2_rec_clusters(el, rec);
1766 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1770 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1772 ocfs2_error(inode->i_sb,
1773 "Inode %llu has bad blkno in extent list "
1774 "at depth %u (index %d)\n",
1775 (unsigned long long)oi->ip_blkno,
1776 le16_to_cpu(el->l_tree_depth), i);
1783 ret = ocfs2_read_extent_block(inode, blkno, &bh);
1789 eb = (struct ocfs2_extent_block *) bh->b_data;
1792 if (le16_to_cpu(el->l_next_free_rec) >
1793 le16_to_cpu(el->l_count)) {
1794 ocfs2_error(inode->i_sb,
1795 "Inode %llu has bad count in extent list "
1796 "at block %llu (next free=%u, count=%u)\n",
1797 (unsigned long long)oi->ip_blkno,
1798 (unsigned long long)bh->b_blocknr,
1799 le16_to_cpu(el->l_next_free_rec),
1800 le16_to_cpu(el->l_count));
1811 * Catch any trailing bh that the loop didn't handle.
1819 * Given an initialized path (that is, it has a valid root extent
1820 * list), this function will traverse the btree in search of the path
1821 * which would contain cpos.
1823 * The path traveled is recorded in the path structure.
1825 * Note that this will not do any comparisons on leaf node extent
1826 * records, so it will work fine in the case that we just added a tree
1829 struct find_path_data {
1831 struct ocfs2_path *path;
1833 static void find_path_ins(void *data, struct buffer_head *bh)
1835 struct find_path_data *fp = data;
1838 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1841 static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
1844 struct find_path_data data;
1848 return __ocfs2_find_path(inode, path_root_el(path), cpos,
1849 find_path_ins, &data);
1852 static void find_leaf_ins(void *data, struct buffer_head *bh)
1854 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1855 struct ocfs2_extent_list *el = &eb->h_list;
1856 struct buffer_head **ret = data;
1858 /* We want to retain only the leaf block. */
1859 if (le16_to_cpu(el->l_tree_depth) == 0) {
1865 * Find the leaf block in the tree which would contain cpos. No
1866 * checking of the actual leaf is done.
1868 * Some paths want to call this instead of allocating a path structure
1869 * and calling ocfs2_find_path().
1871 * This function doesn't handle non btree extent lists.
1873 int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
1874 u32 cpos, struct buffer_head **leaf_bh)
1877 struct buffer_head *bh = NULL;
1879 ret = __ocfs2_find_path(inode, root_el, cpos, find_leaf_ins, &bh);
1891 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1893 * Basically, we've moved stuff around at the bottom of the tree and
1894 * we need to fix up the extent records above the changes to reflect
1897 * left_rec: the record on the left.
1898 * left_child_el: is the child list pointed to by left_rec
1899 * right_rec: the record to the right of left_rec
1900 * right_child_el: is the child list pointed to by right_rec
1902 * By definition, this only works on interior nodes.
1904 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1905 struct ocfs2_extent_list *left_child_el,
1906 struct ocfs2_extent_rec *right_rec,
1907 struct ocfs2_extent_list *right_child_el)
1909 u32 left_clusters, right_end;
1912 * Interior nodes never have holes. Their cpos is the cpos of
1913 * the leftmost record in their child list. Their cluster
1914 * count covers the full theoretical range of their child list
1915 * - the range between their cpos and the cpos of the record
1916 * immediately to their right.
1918 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1919 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1920 BUG_ON(right_child_el->l_tree_depth);
1921 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1922 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1924 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1925 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1928 * Calculate the rightmost cluster count boundary before
1929 * moving cpos - we will need to adjust clusters after
1930 * updating e_cpos to keep the same highest cluster count.
1932 right_end = le32_to_cpu(right_rec->e_cpos);
1933 right_end += le32_to_cpu(right_rec->e_int_clusters);
1935 right_rec->e_cpos = left_rec->e_cpos;
1936 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1938 right_end -= le32_to_cpu(right_rec->e_cpos);
1939 right_rec->e_int_clusters = cpu_to_le32(right_end);
1943 * Adjust the adjacent root node records involved in a
1944 * rotation. left_el_blkno is passed in as a key so that we can easily
1945 * find it's index in the root list.
1947 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1948 struct ocfs2_extent_list *left_el,
1949 struct ocfs2_extent_list *right_el,
1954 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1955 le16_to_cpu(left_el->l_tree_depth));
1957 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1958 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1963 * The path walking code should have never returned a root and
1964 * two paths which are not adjacent.
1966 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1968 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1969 &root_el->l_recs[i + 1], right_el);
1973 * We've changed a leaf block (in right_path) and need to reflect that
1974 * change back up the subtree.
1976 * This happens in multiple places:
1977 * - When we've moved an extent record from the left path leaf to the right
1978 * path leaf to make room for an empty extent in the left path leaf.
1979 * - When our insert into the right path leaf is at the leftmost edge
1980 * and requires an update of the path immediately to it's left. This
1981 * can occur at the end of some types of rotation and appending inserts.
1982 * - When we've adjusted the last extent record in the left path leaf and the
1983 * 1st extent record in the right path leaf during cross extent block merge.
1985 static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1986 struct ocfs2_path *left_path,
1987 struct ocfs2_path *right_path,
1991 struct ocfs2_extent_list *el, *left_el, *right_el;
1992 struct ocfs2_extent_rec *left_rec, *right_rec;
1993 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1996 * Update the counts and position values within all the
1997 * interior nodes to reflect the leaf rotation we just did.
1999 * The root node is handled below the loop.
2001 * We begin the loop with right_el and left_el pointing to the
2002 * leaf lists and work our way up.
2004 * NOTE: within this loop, left_el and right_el always refer
2005 * to the *child* lists.
2007 left_el = path_leaf_el(left_path);
2008 right_el = path_leaf_el(right_path);
2009 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2010 mlog(0, "Adjust records at index %u\n", i);
2013 * One nice property of knowing that all of these
2014 * nodes are below the root is that we only deal with
2015 * the leftmost right node record and the rightmost
2018 el = left_path->p_node[i].el;
2019 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2020 left_rec = &el->l_recs[idx];
2022 el = right_path->p_node[i].el;
2023 right_rec = &el->l_recs[0];
2025 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2028 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2032 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2037 * Setup our list pointers now so that the current
2038 * parents become children in the next iteration.
2040 left_el = left_path->p_node[i].el;
2041 right_el = right_path->p_node[i].el;
2045 * At the root node, adjust the two adjacent records which
2046 * begin our path to the leaves.
2049 el = left_path->p_node[subtree_index].el;
2050 left_el = left_path->p_node[subtree_index + 1].el;
2051 right_el = right_path->p_node[subtree_index + 1].el;
2053 ocfs2_adjust_root_records(el, left_el, right_el,
2054 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2056 root_bh = left_path->p_node[subtree_index].bh;
2058 ret = ocfs2_journal_dirty(handle, root_bh);
2063 static int ocfs2_rotate_subtree_right(struct inode *inode,
2065 struct ocfs2_path *left_path,
2066 struct ocfs2_path *right_path,
2070 struct buffer_head *right_leaf_bh;
2071 struct buffer_head *left_leaf_bh = NULL;
2072 struct buffer_head *root_bh;
2073 struct ocfs2_extent_list *right_el, *left_el;
2074 struct ocfs2_extent_rec move_rec;
2076 left_leaf_bh = path_leaf_bh(left_path);
2077 left_el = path_leaf_el(left_path);
2079 if (left_el->l_next_free_rec != left_el->l_count) {
2080 ocfs2_error(inode->i_sb,
2081 "Inode %llu has non-full interior leaf node %llu"
2083 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2084 (unsigned long long)left_leaf_bh->b_blocknr,
2085 le16_to_cpu(left_el->l_next_free_rec));
2090 * This extent block may already have an empty record, so we
2091 * return early if so.
2093 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2096 root_bh = left_path->p_node[subtree_index].bh;
2097 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2099 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
2106 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2107 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2114 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2122 right_leaf_bh = path_leaf_bh(right_path);
2123 right_el = path_leaf_el(right_path);
2125 /* This is a code error, not a disk corruption. */
2126 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2127 "because rightmost leaf block %llu is empty\n",
2128 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2129 (unsigned long long)right_leaf_bh->b_blocknr);
2131 ocfs2_create_empty_extent(right_el);
2133 ret = ocfs2_journal_dirty(handle, right_leaf_bh);
2139 /* Do the copy now. */
2140 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2141 move_rec = left_el->l_recs[i];
2142 right_el->l_recs[0] = move_rec;
2145 * Clear out the record we just copied and shift everything
2146 * over, leaving an empty extent in the left leaf.
2148 * We temporarily subtract from next_free_rec so that the
2149 * shift will lose the tail record (which is now defunct).
2151 le16_add_cpu(&left_el->l_next_free_rec, -1);
2152 ocfs2_shift_records_right(left_el);
2153 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2154 le16_add_cpu(&left_el->l_next_free_rec, 1);
2156 ret = ocfs2_journal_dirty(handle, left_leaf_bh);
2162 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2170 * Given a full path, determine what cpos value would return us a path
2171 * containing the leaf immediately to the left of the current one.
2173 * Will return zero if the path passed in is already the leftmost path.
2175 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2176 struct ocfs2_path *path, u32 *cpos)
2180 struct ocfs2_extent_list *el;
2182 BUG_ON(path->p_tree_depth == 0);
2186 blkno = path_leaf_bh(path)->b_blocknr;
2188 /* Start at the tree node just above the leaf and work our way up. */
2189 i = path->p_tree_depth - 1;
2191 el = path->p_node[i].el;
2194 * Find the extent record just before the one in our
2197 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2198 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2202 * We've determined that the
2203 * path specified is already
2204 * the leftmost one - return a
2210 * The leftmost record points to our
2211 * leaf - we need to travel up the
2217 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2218 *cpos = *cpos + ocfs2_rec_clusters(el,
2219 &el->l_recs[j - 1]);
2226 * If we got here, we never found a valid node where
2227 * the tree indicated one should be.
2230 "Invalid extent tree at extent block %llu\n",
2231 (unsigned long long)blkno);
2236 blkno = path->p_node[i].bh->b_blocknr;
2245 * Extend the transaction by enough credits to complete the rotation,
2246 * and still leave at least the original number of credits allocated
2247 * to this transaction.
2249 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2251 struct ocfs2_path *path)
2253 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2255 if (handle->h_buffer_credits < credits)
2256 return ocfs2_extend_trans(handle, credits);
2262 * Trap the case where we're inserting into the theoretical range past
2263 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2264 * whose cpos is less than ours into the right leaf.
2266 * It's only necessary to look at the rightmost record of the left
2267 * leaf because the logic that calls us should ensure that the
2268 * theoretical ranges in the path components above the leaves are
2271 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2274 struct ocfs2_extent_list *left_el;
2275 struct ocfs2_extent_rec *rec;
2278 left_el = path_leaf_el(left_path);
2279 next_free = le16_to_cpu(left_el->l_next_free_rec);
2280 rec = &left_el->l_recs[next_free - 1];
2282 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2287 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2289 int next_free = le16_to_cpu(el->l_next_free_rec);
2291 struct ocfs2_extent_rec *rec;
2296 rec = &el->l_recs[0];
2297 if (ocfs2_is_empty_extent(rec)) {
2301 rec = &el->l_recs[1];
2304 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2305 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2311 * Rotate all the records in a btree right one record, starting at insert_cpos.
2313 * The path to the rightmost leaf should be passed in.
2315 * The array is assumed to be large enough to hold an entire path (tree depth).
2317 * Upon succesful return from this function:
2319 * - The 'right_path' array will contain a path to the leaf block
2320 * whose range contains e_cpos.
2321 * - That leaf block will have a single empty extent in list index 0.
2322 * - In the case that the rotation requires a post-insert update,
2323 * *ret_left_path will contain a valid path which can be passed to
2324 * ocfs2_insert_path().
2326 static int ocfs2_rotate_tree_right(struct inode *inode,
2328 enum ocfs2_split_type split,
2330 struct ocfs2_path *right_path,
2331 struct ocfs2_path **ret_left_path)
2333 int ret, start, orig_credits = handle->h_buffer_credits;
2335 struct ocfs2_path *left_path = NULL;
2337 *ret_left_path = NULL;
2339 left_path = ocfs2_new_path_from_path(right_path);
2346 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
2352 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
2355 * What we want to do here is:
2357 * 1) Start with the rightmost path.
2359 * 2) Determine a path to the leaf block directly to the left
2362 * 3) Determine the 'subtree root' - the lowest level tree node
2363 * which contains a path to both leaves.
2365 * 4) Rotate the subtree.
2367 * 5) Find the next subtree by considering the left path to be
2368 * the new right path.
2370 * The check at the top of this while loop also accepts
2371 * insert_cpos == cpos because cpos is only a _theoretical_
2372 * value to get us the left path - insert_cpos might very well
2373 * be filling that hole.
2375 * Stop at a cpos of '0' because we either started at the
2376 * leftmost branch (i.e., a tree with one branch and a
2377 * rotation inside of it), or we've gone as far as we can in
2378 * rotating subtrees.
2380 while (cpos && insert_cpos <= cpos) {
2381 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2384 ret = ocfs2_find_path(inode, left_path, cpos);
2390 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2391 path_leaf_bh(right_path),
2392 "Inode %lu: error during insert of %u "
2393 "(left path cpos %u) results in two identical "
2394 "paths ending at %llu\n",
2395 inode->i_ino, insert_cpos, cpos,
2396 (unsigned long long)
2397 path_leaf_bh(left_path)->b_blocknr);
2399 if (split == SPLIT_NONE &&
2400 ocfs2_rotate_requires_path_adjustment(left_path,
2404 * We've rotated the tree as much as we
2405 * should. The rest is up to
2406 * ocfs2_insert_path() to complete, after the
2407 * record insertion. We indicate this
2408 * situation by returning the left path.
2410 * The reason we don't adjust the records here
2411 * before the record insert is that an error
2412 * later might break the rule where a parent
2413 * record e_cpos will reflect the actual
2414 * e_cpos of the 1st nonempty record of the
2417 *ret_left_path = left_path;
2421 start = ocfs2_find_subtree_root(inode, left_path, right_path);
2423 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2425 (unsigned long long) right_path->p_node[start].bh->b_blocknr,
2426 right_path->p_tree_depth);
2428 ret = ocfs2_extend_rotate_transaction(handle, start,
2429 orig_credits, right_path);
2435 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
2442 if (split != SPLIT_NONE &&
2443 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2446 * A rotate moves the rightmost left leaf
2447 * record over to the leftmost right leaf
2448 * slot. If we're doing an extent split
2449 * instead of a real insert, then we have to
2450 * check that the extent to be split wasn't
2451 * just moved over. If it was, then we can
2452 * exit here, passing left_path back -
2453 * ocfs2_split_extent() is smart enough to
2454 * search both leaves.
2456 *ret_left_path = left_path;
2461 * There is no need to re-read the next right path
2462 * as we know that it'll be our current left
2463 * path. Optimize by copying values instead.
2465 ocfs2_mv_path(right_path, left_path);
2467 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
2476 ocfs2_free_path(left_path);
2482 static int ocfs2_update_edge_lengths(struct inode *inode, handle_t *handle,
2483 int subtree_index, struct ocfs2_path *path)
2486 struct ocfs2_extent_rec *rec;
2487 struct ocfs2_extent_list *el;
2488 struct ocfs2_extent_block *eb;
2492 * In normal tree rotation process, we will never touch the
2493 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2494 * doesn't reserve the credits for them either.
2496 * But we do have a special case here which will update the rightmost
2497 * records for all the bh in the path.
2498 * So we have to allocate extra credits and access them.
2500 ret = ocfs2_extend_trans(handle,
2501 handle->h_buffer_credits + subtree_index);
2507 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
2513 /* Path should always be rightmost. */
2514 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2515 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2518 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2519 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2520 rec = &el->l_recs[idx];
2521 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2523 for (i = 0; i < path->p_tree_depth; i++) {
2524 el = path->p_node[i].el;
2525 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2526 rec = &el->l_recs[idx];
2528 rec->e_int_clusters = cpu_to_le32(range);
2529 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2531 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2537 static void ocfs2_unlink_path(struct inode *inode, handle_t *handle,
2538 struct ocfs2_cached_dealloc_ctxt *dealloc,
2539 struct ocfs2_path *path, int unlink_start)
2542 struct ocfs2_extent_block *eb;
2543 struct ocfs2_extent_list *el;
2544 struct buffer_head *bh;
2546 for(i = unlink_start; i < path_num_items(path); i++) {
2547 bh = path->p_node[i].bh;
2549 eb = (struct ocfs2_extent_block *)bh->b_data;
2551 * Not all nodes might have had their final count
2552 * decremented by the caller - handle this here.
2555 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2557 "Inode %llu, attempted to remove extent block "
2558 "%llu with %u records\n",
2559 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2560 (unsigned long long)le64_to_cpu(eb->h_blkno),
2561 le16_to_cpu(el->l_next_free_rec));
2563 ocfs2_journal_dirty(handle, bh);
2564 ocfs2_remove_from_cache(INODE_CACHE(inode), bh);
2568 el->l_next_free_rec = 0;
2569 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2571 ocfs2_journal_dirty(handle, bh);
2573 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2577 ocfs2_remove_from_cache(INODE_CACHE(inode), bh);
2581 static void ocfs2_unlink_subtree(struct inode *inode, handle_t *handle,
2582 struct ocfs2_path *left_path,
2583 struct ocfs2_path *right_path,
2585 struct ocfs2_cached_dealloc_ctxt *dealloc)
2588 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2589 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2590 struct ocfs2_extent_list *el;
2591 struct ocfs2_extent_block *eb;
2593 el = path_leaf_el(left_path);
2595 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2597 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2598 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2601 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2603 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2604 le16_add_cpu(&root_el->l_next_free_rec, -1);
2606 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2607 eb->h_next_leaf_blk = 0;
2609 ocfs2_journal_dirty(handle, root_bh);
2610 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2612 ocfs2_unlink_path(inode, handle, dealloc, right_path,
2616 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2617 struct ocfs2_path *left_path,
2618 struct ocfs2_path *right_path,
2620 struct ocfs2_cached_dealloc_ctxt *dealloc,
2622 struct ocfs2_extent_tree *et)
2624 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2625 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2626 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2627 struct ocfs2_extent_block *eb;
2631 right_leaf_el = path_leaf_el(right_path);
2632 left_leaf_el = path_leaf_el(left_path);
2633 root_bh = left_path->p_node[subtree_index].bh;
2634 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2636 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2639 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2640 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2642 * It's legal for us to proceed if the right leaf is
2643 * the rightmost one and it has an empty extent. There
2644 * are two cases to handle - whether the leaf will be
2645 * empty after removal or not. If the leaf isn't empty
2646 * then just remove the empty extent up front. The
2647 * next block will handle empty leaves by flagging
2650 * Non rightmost leaves will throw -EAGAIN and the
2651 * caller can manually move the subtree and retry.
2654 if (eb->h_next_leaf_blk != 0ULL)
2657 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2658 ret = ocfs2_journal_access_eb(handle, INODE_CACHE(inode),
2659 path_leaf_bh(right_path),
2660 OCFS2_JOURNAL_ACCESS_WRITE);
2666 ocfs2_remove_empty_extent(right_leaf_el);
2668 right_has_empty = 1;
2671 if (eb->h_next_leaf_blk == 0ULL &&
2672 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2674 * We have to update i_last_eb_blk during the meta
2677 ret = ocfs2_et_root_journal_access(handle, INODE_CACHE(inode), et,
2678 OCFS2_JOURNAL_ACCESS_WRITE);
2684 del_right_subtree = 1;
2688 * Getting here with an empty extent in the right path implies
2689 * that it's the rightmost path and will be deleted.
2691 BUG_ON(right_has_empty && !del_right_subtree);
2693 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
2700 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2701 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2708 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2716 if (!right_has_empty) {
2718 * Only do this if we're moving a real
2719 * record. Otherwise, the action is delayed until
2720 * after removal of the right path in which case we
2721 * can do a simple shift to remove the empty extent.
2723 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2724 memset(&right_leaf_el->l_recs[0], 0,
2725 sizeof(struct ocfs2_extent_rec));
2727 if (eb->h_next_leaf_blk == 0ULL) {
2729 * Move recs over to get rid of empty extent, decrease
2730 * next_free. This is allowed to remove the last
2731 * extent in our leaf (setting l_next_free_rec to
2732 * zero) - the delete code below won't care.
2734 ocfs2_remove_empty_extent(right_leaf_el);
2737 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2740 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2744 if (del_right_subtree) {
2745 ocfs2_unlink_subtree(inode, handle, left_path, right_path,
2746 subtree_index, dealloc);
2747 ret = ocfs2_update_edge_lengths(inode, handle, subtree_index,
2754 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2755 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2758 * Removal of the extent in the left leaf was skipped
2759 * above so we could delete the right path
2762 if (right_has_empty)
2763 ocfs2_remove_empty_extent(left_leaf_el);
2765 ret = ocfs2_journal_dirty(handle, et_root_bh);
2771 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2779 * Given a full path, determine what cpos value would return us a path
2780 * containing the leaf immediately to the right of the current one.
2782 * Will return zero if the path passed in is already the rightmost path.
2784 * This looks similar, but is subtly different to
2785 * ocfs2_find_cpos_for_left_leaf().
2787 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2788 struct ocfs2_path *path, u32 *cpos)
2792 struct ocfs2_extent_list *el;
2796 if (path->p_tree_depth == 0)
2799 blkno = path_leaf_bh(path)->b_blocknr;
2801 /* Start at the tree node just above the leaf and work our way up. */
2802 i = path->p_tree_depth - 1;
2806 el = path->p_node[i].el;
2809 * Find the extent record just after the one in our
2812 next_free = le16_to_cpu(el->l_next_free_rec);
2813 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2814 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2815 if (j == (next_free - 1)) {
2818 * We've determined that the
2819 * path specified is already
2820 * the rightmost one - return a
2826 * The rightmost record points to our
2827 * leaf - we need to travel up the
2833 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2839 * If we got here, we never found a valid node where
2840 * the tree indicated one should be.
2843 "Invalid extent tree at extent block %llu\n",
2844 (unsigned long long)blkno);
2849 blkno = path->p_node[i].bh->b_blocknr;
2857 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2859 struct ocfs2_path *path)
2862 struct buffer_head *bh = path_leaf_bh(path);
2863 struct ocfs2_extent_list *el = path_leaf_el(path);
2865 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2868 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), path,
2869 path_num_items(path) - 1);
2875 ocfs2_remove_empty_extent(el);
2877 ret = ocfs2_journal_dirty(handle, bh);
2885 static int __ocfs2_rotate_tree_left(struct inode *inode,
2886 handle_t *handle, int orig_credits,
2887 struct ocfs2_path *path,
2888 struct ocfs2_cached_dealloc_ctxt *dealloc,
2889 struct ocfs2_path **empty_extent_path,
2890 struct ocfs2_extent_tree *et)
2892 int ret, subtree_root, deleted;
2894 struct ocfs2_path *left_path = NULL;
2895 struct ocfs2_path *right_path = NULL;
2897 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2899 *empty_extent_path = NULL;
2901 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2908 left_path = ocfs2_new_path_from_path(path);
2915 ocfs2_cp_path(left_path, path);
2917 right_path = ocfs2_new_path_from_path(path);
2924 while (right_cpos) {
2925 ret = ocfs2_find_path(inode, right_path, right_cpos);
2931 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2934 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2936 (unsigned long long)
2937 right_path->p_node[subtree_root].bh->b_blocknr,
2938 right_path->p_tree_depth);
2940 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2941 orig_credits, left_path);
2948 * Caller might still want to make changes to the
2949 * tree root, so re-add it to the journal here.
2951 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
2958 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2959 right_path, subtree_root,
2960 dealloc, &deleted, et);
2961 if (ret == -EAGAIN) {
2963 * The rotation has to temporarily stop due to
2964 * the right subtree having an empty
2965 * extent. Pass it back to the caller for a
2968 *empty_extent_path = right_path;
2978 * The subtree rotate might have removed records on
2979 * the rightmost edge. If so, then rotation is
2985 ocfs2_mv_path(left_path, right_path);
2987 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2996 ocfs2_free_path(right_path);
2997 ocfs2_free_path(left_path);
3002 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
3003 struct ocfs2_path *path,
3004 struct ocfs2_cached_dealloc_ctxt *dealloc,
3005 struct ocfs2_extent_tree *et)
3007 int ret, subtree_index;
3009 struct ocfs2_path *left_path = NULL;
3010 struct ocfs2_extent_block *eb;
3011 struct ocfs2_extent_list *el;
3014 ret = ocfs2_et_sanity_check(inode, et);
3018 * There's two ways we handle this depending on
3019 * whether path is the only existing one.
3021 ret = ocfs2_extend_rotate_transaction(handle, 0,
3022 handle->h_buffer_credits,
3029 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
3035 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
3043 * We have a path to the left of this one - it needs
3046 left_path = ocfs2_new_path_from_path(path);
3053 ret = ocfs2_find_path(inode, left_path, cpos);
3059 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, left_path);
3065 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
3067 ocfs2_unlink_subtree(inode, handle, left_path, path,
3068 subtree_index, dealloc);
3069 ret = ocfs2_update_edge_lengths(inode, handle, subtree_index,
3076 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3077 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3080 * 'path' is also the leftmost path which
3081 * means it must be the only one. This gets
3082 * handled differently because we want to
3083 * revert the inode back to having extents
3086 ocfs2_unlink_path(inode, handle, dealloc, path, 1);
3088 el = et->et_root_el;
3089 el->l_tree_depth = 0;
3090 el->l_next_free_rec = 0;
3091 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3093 ocfs2_et_set_last_eb_blk(et, 0);
3096 ocfs2_journal_dirty(handle, path_root_bh(path));
3099 ocfs2_free_path(left_path);
3104 * Left rotation of btree records.
3106 * In many ways, this is (unsurprisingly) the opposite of right
3107 * rotation. We start at some non-rightmost path containing an empty
3108 * extent in the leaf block. The code works its way to the rightmost
3109 * path by rotating records to the left in every subtree.
3111 * This is used by any code which reduces the number of extent records
3112 * in a leaf. After removal, an empty record should be placed in the
3113 * leftmost list position.
3115 * This won't handle a length update of the rightmost path records if
3116 * the rightmost tree leaf record is removed so the caller is
3117 * responsible for detecting and correcting that.
3119 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
3120 struct ocfs2_path *path,
3121 struct ocfs2_cached_dealloc_ctxt *dealloc,
3122 struct ocfs2_extent_tree *et)
3124 int ret, orig_credits = handle->h_buffer_credits;
3125 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3126 struct ocfs2_extent_block *eb;
3127 struct ocfs2_extent_list *el;
3129 el = path_leaf_el(path);
3130 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3133 if (path->p_tree_depth == 0) {
3134 rightmost_no_delete:
3136 * Inline extents. This is trivially handled, so do
3139 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
3147 * Handle rightmost branch now. There's several cases:
3148 * 1) simple rotation leaving records in there. That's trivial.
3149 * 2) rotation requiring a branch delete - there's no more
3150 * records left. Two cases of this:
3151 * a) There are branches to the left.
3152 * b) This is also the leftmost (the only) branch.
3154 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3155 * 2a) we need the left branch so that we can update it with the unlink
3156 * 2b) we need to bring the inode back to inline extents.
3159 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3161 if (eb->h_next_leaf_blk == 0) {
3163 * This gets a bit tricky if we're going to delete the
3164 * rightmost path. Get the other cases out of the way
3167 if (le16_to_cpu(el->l_next_free_rec) > 1)
3168 goto rightmost_no_delete;
3170 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3172 ocfs2_error(inode->i_sb,
3173 "Inode %llu has empty extent block at %llu",
3174 (unsigned long long)OCFS2_I(inode)->ip_blkno,
3175 (unsigned long long)le64_to_cpu(eb->h_blkno));
3180 * XXX: The caller can not trust "path" any more after
3181 * this as it will have been deleted. What do we do?
3183 * In theory the rotate-for-merge code will never get
3184 * here because it'll always ask for a rotate in a
3188 ret = ocfs2_remove_rightmost_path(inode, handle, path,
3196 * Now we can loop, remembering the path we get from -EAGAIN
3197 * and restarting from there.
3200 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
3201 dealloc, &restart_path, et);
3202 if (ret && ret != -EAGAIN) {
3207 while (ret == -EAGAIN) {
3208 tmp_path = restart_path;
3209 restart_path = NULL;
3211 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
3214 if (ret && ret != -EAGAIN) {
3219 ocfs2_free_path(tmp_path);
3227 ocfs2_free_path(tmp_path);
3228 ocfs2_free_path(restart_path);
3232 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3235 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3238 if (rec->e_leaf_clusters == 0) {
3240 * We consumed all of the merged-from record. An empty
3241 * extent cannot exist anywhere but the 1st array
3242 * position, so move things over if the merged-from
3243 * record doesn't occupy that position.
3245 * This creates a new empty extent so the caller
3246 * should be smart enough to have removed any existing
3250 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3251 size = index * sizeof(struct ocfs2_extent_rec);
3252 memmove(&el->l_recs[1], &el->l_recs[0], size);
3256 * Always memset - the caller doesn't check whether it
3257 * created an empty extent, so there could be junk in
3260 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3264 static int ocfs2_get_right_path(struct inode *inode,
3265 struct ocfs2_path *left_path,
3266 struct ocfs2_path **ret_right_path)
3270 struct ocfs2_path *right_path = NULL;
3271 struct ocfs2_extent_list *left_el;
3273 *ret_right_path = NULL;
3275 /* This function shouldn't be called for non-trees. */
3276 BUG_ON(left_path->p_tree_depth == 0);
3278 left_el = path_leaf_el(left_path);
3279 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3281 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
3288 /* This function shouldn't be called for the rightmost leaf. */
3289 BUG_ON(right_cpos == 0);
3291 right_path = ocfs2_new_path_from_path(left_path);
3298 ret = ocfs2_find_path(inode, right_path, right_cpos);
3304 *ret_right_path = right_path;
3307 ocfs2_free_path(right_path);
3312 * Remove split_rec clusters from the record at index and merge them
3313 * onto the beginning of the record "next" to it.
3314 * For index < l_count - 1, the next means the extent rec at index + 1.
3315 * For index == l_count - 1, the "next" means the 1st extent rec of the
3316 * next extent block.
3318 static int ocfs2_merge_rec_right(struct inode *inode,
3319 struct ocfs2_path *left_path,
3321 struct ocfs2_extent_rec *split_rec,
3324 int ret, next_free, i;
3325 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3326 struct ocfs2_extent_rec *left_rec;
3327 struct ocfs2_extent_rec *right_rec;
3328 struct ocfs2_extent_list *right_el;
3329 struct ocfs2_path *right_path = NULL;
3330 int subtree_index = 0;
3331 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3332 struct buffer_head *bh = path_leaf_bh(left_path);
3333 struct buffer_head *root_bh = NULL;
3335 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3336 left_rec = &el->l_recs[index];
3338 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3339 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3340 /* we meet with a cross extent block merge. */
3341 ret = ocfs2_get_right_path(inode, left_path, &right_path);
3347 right_el = path_leaf_el(right_path);
3348 next_free = le16_to_cpu(right_el->l_next_free_rec);
3349 BUG_ON(next_free <= 0);
3350 right_rec = &right_el->l_recs[0];
3351 if (ocfs2_is_empty_extent(right_rec)) {
3352 BUG_ON(next_free <= 1);
3353 right_rec = &right_el->l_recs[1];
3356 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3357 le16_to_cpu(left_rec->e_leaf_clusters) !=
3358 le32_to_cpu(right_rec->e_cpos));
3360 subtree_index = ocfs2_find_subtree_root(inode,
3361 left_path, right_path);
3363 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3364 handle->h_buffer_credits,
3371 root_bh = left_path->p_node[subtree_index].bh;
3372 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3374 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3381 for (i = subtree_index + 1;
3382 i < path_num_items(right_path); i++) {
3383 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3390 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3399 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3400 right_rec = &el->l_recs[index + 1];
3403 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), left_path,
3404 path_num_items(left_path) - 1);
3410 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3412 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3413 le64_add_cpu(&right_rec->e_blkno,
3414 -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3415 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3417 ocfs2_cleanup_merge(el, index);
3419 ret = ocfs2_journal_dirty(handle, bh);
3424 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3428 ocfs2_complete_edge_insert(inode, handle, left_path,
3429 right_path, subtree_index);
3433 ocfs2_free_path(right_path);
3437 static int ocfs2_get_left_path(struct inode *inode,
3438 struct ocfs2_path *right_path,
3439 struct ocfs2_path **ret_left_path)
3443 struct ocfs2_path *left_path = NULL;
3445 *ret_left_path = NULL;
3447 /* This function shouldn't be called for non-trees. */
3448 BUG_ON(right_path->p_tree_depth == 0);
3450 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
3451 right_path, &left_cpos);
3457 /* This function shouldn't be called for the leftmost leaf. */
3458 BUG_ON(left_cpos == 0);
3460 left_path = ocfs2_new_path_from_path(right_path);
3467 ret = ocfs2_find_path(inode, left_path, left_cpos);
3473 *ret_left_path = left_path;
3476 ocfs2_free_path(left_path);
3481 * Remove split_rec clusters from the record at index and merge them
3482 * onto the tail of the record "before" it.
3483 * For index > 0, the "before" means the extent rec at index - 1.
3485 * For index == 0, the "before" means the last record of the previous
3486 * extent block. And there is also a situation that we may need to
3487 * remove the rightmost leaf extent block in the right_path and change
3488 * the right path to indicate the new rightmost path.
3490 static int ocfs2_merge_rec_left(struct inode *inode,
3491 struct ocfs2_path *right_path,
3493 struct ocfs2_extent_rec *split_rec,
3494 struct ocfs2_cached_dealloc_ctxt *dealloc,
3495 struct ocfs2_extent_tree *et,
3498 int ret, i, subtree_index = 0, has_empty_extent = 0;
3499 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3500 struct ocfs2_extent_rec *left_rec;
3501 struct ocfs2_extent_rec *right_rec;
3502 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3503 struct buffer_head *bh = path_leaf_bh(right_path);
3504 struct buffer_head *root_bh = NULL;
3505 struct ocfs2_path *left_path = NULL;
3506 struct ocfs2_extent_list *left_el;
3510 right_rec = &el->l_recs[index];
3512 /* we meet with a cross extent block merge. */
3513 ret = ocfs2_get_left_path(inode, right_path, &left_path);
3519 left_el = path_leaf_el(left_path);
3520 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3521 le16_to_cpu(left_el->l_count));
3523 left_rec = &left_el->l_recs[
3524 le16_to_cpu(left_el->l_next_free_rec) - 1];
3525 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3526 le16_to_cpu(left_rec->e_leaf_clusters) !=
3527 le32_to_cpu(split_rec->e_cpos));
3529 subtree_index = ocfs2_find_subtree_root(inode,
3530 left_path, right_path);
3532 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3533 handle->h_buffer_credits,
3540 root_bh = left_path->p_node[subtree_index].bh;
3541 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3543 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3550 for (i = subtree_index + 1;
3551 i < path_num_items(right_path); i++) {
3552 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3559 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode),
3567 left_rec = &el->l_recs[index - 1];
3568 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3569 has_empty_extent = 1;
3572 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), right_path,
3573 path_num_items(right_path) - 1);
3579 if (has_empty_extent && index == 1) {
3581 * The easy case - we can just plop the record right in.
3583 *left_rec = *split_rec;
3585 has_empty_extent = 0;
3587 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3589 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3590 le64_add_cpu(&right_rec->e_blkno,
3591 ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
3592 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3594 ocfs2_cleanup_merge(el, index);
3596 ret = ocfs2_journal_dirty(handle, bh);
3601 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3606 * In the situation that the right_rec is empty and the extent
3607 * block is empty also, ocfs2_complete_edge_insert can't handle
3608 * it and we need to delete the right extent block.
3610 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3611 le16_to_cpu(el->l_next_free_rec) == 1) {
3613 ret = ocfs2_remove_rightmost_path(inode, handle,
3621 /* Now the rightmost extent block has been deleted.
3622 * So we use the new rightmost path.
3624 ocfs2_mv_path(right_path, left_path);
3627 ocfs2_complete_edge_insert(inode, handle, left_path,
3628 right_path, subtree_index);
3632 ocfs2_free_path(left_path);
3636 static int ocfs2_try_to_merge_extent(struct inode *inode,
3638 struct ocfs2_path *path,
3640 struct ocfs2_extent_rec *split_rec,
3641 struct ocfs2_cached_dealloc_ctxt *dealloc,
3642 struct ocfs2_merge_ctxt *ctxt,
3643 struct ocfs2_extent_tree *et)
3647 struct ocfs2_extent_list *el = path_leaf_el(path);
3648 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3650 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3652 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3654 * The merge code will need to create an empty
3655 * extent to take the place of the newly
3656 * emptied slot. Remove any pre-existing empty
3657 * extents - having more than one in a leaf is
3660 ret = ocfs2_rotate_tree_left(inode, handle, path,
3667 rec = &el->l_recs[split_index];
3670 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3672 * Left-right contig implies this.
3674 BUG_ON(!ctxt->c_split_covers_rec);
3677 * Since the leftright insert always covers the entire
3678 * extent, this call will delete the insert record
3679 * entirely, resulting in an empty extent record added to
3682 * Since the adding of an empty extent shifts
3683 * everything back to the right, there's no need to
3684 * update split_index here.
3686 * When the split_index is zero, we need to merge it to the
3687 * prevoius extent block. It is more efficient and easier
3688 * if we do merge_right first and merge_left later.
3690 ret = ocfs2_merge_rec_right(inode, path,
3699 * We can only get this from logic error above.
3701 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3703 /* The merge left us with an empty extent, remove it. */
3704 ret = ocfs2_rotate_tree_left(inode, handle, path,
3711 rec = &el->l_recs[split_index];
3714 * Note that we don't pass split_rec here on purpose -
3715 * we've merged it into the rec already.
3717 ret = ocfs2_merge_rec_left(inode, path,
3727 ret = ocfs2_rotate_tree_left(inode, handle, path,
3730 * Error from this last rotate is not critical, so
3731 * print but don't bubble it up.
3738 * Merge a record to the left or right.
3740 * 'contig_type' is relative to the existing record,
3741 * so for example, if we're "right contig", it's to
3742 * the record on the left (hence the left merge).
3744 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3745 ret = ocfs2_merge_rec_left(inode,
3755 ret = ocfs2_merge_rec_right(inode,
3765 if (ctxt->c_split_covers_rec) {
3767 * The merge may have left an empty extent in
3768 * our leaf. Try to rotate it away.
3770 ret = ocfs2_rotate_tree_left(inode, handle, path,
3782 static void ocfs2_subtract_from_rec(struct super_block *sb,
3783 enum ocfs2_split_type split,
3784 struct ocfs2_extent_rec *rec,
3785 struct ocfs2_extent_rec *split_rec)
3789 len_blocks = ocfs2_clusters_to_blocks(sb,
3790 le16_to_cpu(split_rec->e_leaf_clusters));
3792 if (split == SPLIT_LEFT) {
3794 * Region is on the left edge of the existing
3797 le32_add_cpu(&rec->e_cpos,
3798 le16_to_cpu(split_rec->e_leaf_clusters));
3799 le64_add_cpu(&rec->e_blkno, len_blocks);
3800 le16_add_cpu(&rec->e_leaf_clusters,
3801 -le16_to_cpu(split_rec->e_leaf_clusters));
3804 * Region is on the right edge of the existing
3807 le16_add_cpu(&rec->e_leaf_clusters,
3808 -le16_to_cpu(split_rec->e_leaf_clusters));
3813 * Do the final bits of extent record insertion at the target leaf
3814 * list. If this leaf is part of an allocation tree, it is assumed
3815 * that the tree above has been prepared.
3817 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
3818 struct ocfs2_extent_list *el,
3819 struct ocfs2_insert_type *insert,
3820 struct inode *inode)
3822 int i = insert->ins_contig_index;
3824 struct ocfs2_extent_rec *rec;
3826 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3828 if (insert->ins_split != SPLIT_NONE) {
3829 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3831 rec = &el->l_recs[i];
3832 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
3838 * Contiguous insert - either left or right.
3840 if (insert->ins_contig != CONTIG_NONE) {
3841 rec = &el->l_recs[i];
3842 if (insert->ins_contig == CONTIG_LEFT) {
3843 rec->e_blkno = insert_rec->e_blkno;
3844 rec->e_cpos = insert_rec->e_cpos;
3846 le16_add_cpu(&rec->e_leaf_clusters,
3847 le16_to_cpu(insert_rec->e_leaf_clusters));
3852 * Handle insert into an empty leaf.
3854 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3855 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3856 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3857 el->l_recs[0] = *insert_rec;
3858 el->l_next_free_rec = cpu_to_le16(1);
3865 if (insert->ins_appending == APPEND_TAIL) {
3866 i = le16_to_cpu(el->l_next_free_rec) - 1;
3867 rec = &el->l_recs[i];
3868 range = le32_to_cpu(rec->e_cpos)
3869 + le16_to_cpu(rec->e_leaf_clusters);
3870 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3872 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3873 le16_to_cpu(el->l_count),
3874 "inode %lu, depth %u, count %u, next free %u, "
3875 "rec.cpos %u, rec.clusters %u, "
3876 "insert.cpos %u, insert.clusters %u\n",
3878 le16_to_cpu(el->l_tree_depth),
3879 le16_to_cpu(el->l_count),
3880 le16_to_cpu(el->l_next_free_rec),
3881 le32_to_cpu(el->l_recs[i].e_cpos),
3882 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3883 le32_to_cpu(insert_rec->e_cpos),
3884 le16_to_cpu(insert_rec->e_leaf_clusters));
3886 el->l_recs[i] = *insert_rec;
3887 le16_add_cpu(&el->l_next_free_rec, 1);
3893 * Ok, we have to rotate.
3895 * At this point, it is safe to assume that inserting into an
3896 * empty leaf and appending to a leaf have both been handled
3899 * This leaf needs to have space, either by the empty 1st
3900 * extent record, or by virtue of an l_next_rec < l_count.
3902 ocfs2_rotate_leaf(el, insert_rec);
3905 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3907 struct ocfs2_path *path,
3908 struct ocfs2_extent_rec *insert_rec)
3910 int ret, i, next_free;
3911 struct buffer_head *bh;
3912 struct ocfs2_extent_list *el;
3913 struct ocfs2_extent_rec *rec;
3916 * Update everything except the leaf block.
3918 for (i = 0; i < path->p_tree_depth; i++) {
3919 bh = path->p_node[i].bh;
3920 el = path->p_node[i].el;
3922 next_free = le16_to_cpu(el->l_next_free_rec);
3923 if (next_free == 0) {
3924 ocfs2_error(inode->i_sb,
3925 "Dinode %llu has a bad extent list",
3926 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3931 rec = &el->l_recs[next_free - 1];
3933 rec->e_int_clusters = insert_rec->e_cpos;
3934 le32_add_cpu(&rec->e_int_clusters,
3935 le16_to_cpu(insert_rec->e_leaf_clusters));
3936 le32_add_cpu(&rec->e_int_clusters,
3937 -le32_to_cpu(rec->e_cpos));
3939 ret = ocfs2_journal_dirty(handle, bh);
3946 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3947 struct ocfs2_extent_rec *insert_rec,
3948 struct ocfs2_path *right_path,
3949 struct ocfs2_path **ret_left_path)
3952 struct ocfs2_extent_list *el;
3953 struct ocfs2_path *left_path = NULL;
3955 *ret_left_path = NULL;
3958 * This shouldn't happen for non-trees. The extent rec cluster
3959 * count manipulation below only works for interior nodes.
3961 BUG_ON(right_path->p_tree_depth == 0);
3964 * If our appending insert is at the leftmost edge of a leaf,
3965 * then we might need to update the rightmost records of the
3968 el = path_leaf_el(right_path);
3969 next_free = le16_to_cpu(el->l_next_free_rec);
3970 if (next_free == 0 ||
3971 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3974 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3981 mlog(0, "Append may need a left path update. cpos: %u, "
3982 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3986 * No need to worry if the append is already in the
3990 left_path = ocfs2_new_path_from_path(right_path);
3997 ret = ocfs2_find_path(inode, left_path, left_cpos);
4004 * ocfs2_insert_path() will pass the left_path to the
4010 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, right_path);
4016 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
4018 *ret_left_path = left_path;
4022 ocfs2_free_path(left_path);
4027 static void ocfs2_split_record(struct inode *inode,
4028 struct ocfs2_path *left_path,
4029 struct ocfs2_path *right_path,
4030 struct ocfs2_extent_rec *split_rec,
4031 enum ocfs2_split_type split)
4034 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4035 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4036 struct ocfs2_extent_rec *rec, *tmprec;
4038 right_el = path_leaf_el(right_path);
4040 left_el = path_leaf_el(left_path);
4043 insert_el = right_el;
4044 index = ocfs2_search_extent_list(el, cpos);
4046 if (index == 0 && left_path) {
4047 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4050 * This typically means that the record
4051 * started in the left path but moved to the
4052 * right as a result of rotation. We either
4053 * move the existing record to the left, or we
4054 * do the later insert there.
4056 * In this case, the left path should always
4057 * exist as the rotate code will have passed
4058 * it back for a post-insert update.
4061 if (split == SPLIT_LEFT) {
4063 * It's a left split. Since we know
4064 * that the rotate code gave us an
4065 * empty extent in the left path, we
4066 * can just do the insert there.
4068 insert_el = left_el;
4071 * Right split - we have to move the
4072 * existing record over to the left
4073 * leaf. The insert will be into the
4074 * newly created empty extent in the
4077 tmprec = &right_el->l_recs[index];
4078 ocfs2_rotate_leaf(left_el, tmprec);
4081 memset(tmprec, 0, sizeof(*tmprec));
4082 index = ocfs2_search_extent_list(left_el, cpos);
4083 BUG_ON(index == -1);
4088 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4090 * Left path is easy - we can just allow the insert to
4094 insert_el = left_el;
4095 index = ocfs2_search_extent_list(el, cpos);
4096 BUG_ON(index == -1);
4099 rec = &el->l_recs[index];
4100 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
4101 ocfs2_rotate_leaf(insert_el, split_rec);
4105 * This function only does inserts on an allocation b-tree. For tree
4106 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4108 * right_path is the path we want to do the actual insert
4109 * in. left_path should only be passed in if we need to update that
4110 * portion of the tree after an edge insert.
4112 static int ocfs2_insert_path(struct inode *inode,
4114 struct ocfs2_path *left_path,
4115 struct ocfs2_path *right_path,
4116 struct ocfs2_extent_rec *insert_rec,
4117 struct ocfs2_insert_type *insert)
4119 int ret, subtree_index;
4120 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4123 int credits = handle->h_buffer_credits;
4126 * There's a chance that left_path got passed back to
4127 * us without being accounted for in the
4128 * journal. Extend our transaction here to be sure we
4129 * can change those blocks.
4131 credits += left_path->p_tree_depth;
4133 ret = ocfs2_extend_trans(handle, credits);
4139 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, left_path);
4147 * Pass both paths to the journal. The majority of inserts
4148 * will be touching all components anyway.
4150 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, right_path);
4156 if (insert->ins_split != SPLIT_NONE) {
4158 * We could call ocfs2_insert_at_leaf() for some types
4159 * of splits, but it's easier to just let one separate
4160 * function sort it all out.
4162 ocfs2_split_record(inode, left_path, right_path,
4163 insert_rec, insert->ins_split);
4166 * Split might have modified either leaf and we don't
4167 * have a guarantee that the later edge insert will
4168 * dirty this for us.
4171 ret = ocfs2_journal_dirty(handle,
4172 path_leaf_bh(left_path));
4176 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
4179 ret = ocfs2_journal_dirty(handle, leaf_bh);
4185 * The rotate code has indicated that we need to fix
4186 * up portions of the tree after the insert.
4188 * XXX: Should we extend the transaction here?
4190 subtree_index = ocfs2_find_subtree_root(inode, left_path,
4192 ocfs2_complete_edge_insert(inode, handle, left_path,
4193 right_path, subtree_index);
4201 static int ocfs2_do_insert_extent(struct inode *inode,
4203 struct ocfs2_extent_tree *et,
4204 struct ocfs2_extent_rec *insert_rec,
4205 struct ocfs2_insert_type *type)
4207 int ret, rotate = 0;
4209 struct ocfs2_path *right_path = NULL;
4210 struct ocfs2_path *left_path = NULL;
4211 struct ocfs2_extent_list *el;
4213 el = et->et_root_el;
4215 ret = ocfs2_et_root_journal_access(handle, INODE_CACHE(inode), et,
4216 OCFS2_JOURNAL_ACCESS_WRITE);
4222 if (le16_to_cpu(el->l_tree_depth) == 0) {
4223 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
4224 goto out_update_clusters;
4227 right_path = ocfs2_new_path_from_et(et);
4235 * Determine the path to start with. Rotations need the
4236 * rightmost path, everything else can go directly to the
4239 cpos = le32_to_cpu(insert_rec->e_cpos);
4240 if (type->ins_appending == APPEND_NONE &&
4241 type->ins_contig == CONTIG_NONE) {
4246 ret = ocfs2_find_path(inode, right_path, cpos);
4253 * Rotations and appends need special treatment - they modify
4254 * parts of the tree's above them.
4256 * Both might pass back a path immediate to the left of the
4257 * one being inserted to. This will be cause
4258 * ocfs2_insert_path() to modify the rightmost records of
4259 * left_path to account for an edge insert.
4261 * XXX: When modifying this code, keep in mind that an insert
4262 * can wind up skipping both of these two special cases...
4265 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
4266 le32_to_cpu(insert_rec->e_cpos),
4267 right_path, &left_path);
4274 * ocfs2_rotate_tree_right() might have extended the
4275 * transaction without re-journaling our tree root.
4277 ret = ocfs2_et_root_journal_access(handle, INODE_CACHE(inode), et,
4278 OCFS2_JOURNAL_ACCESS_WRITE);
4283 } else if (type->ins_appending == APPEND_TAIL
4284 && type->ins_contig != CONTIG_LEFT) {
4285 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
4286 right_path, &left_path);
4293 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
4300 out_update_clusters:
4301 if (type->ins_split == SPLIT_NONE)
4302 ocfs2_et_update_clusters(inode, et,
4303 le16_to_cpu(insert_rec->e_leaf_clusters));
4305 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
4310 ocfs2_free_path(left_path);
4311 ocfs2_free_path(right_path);
4316 static enum ocfs2_contig_type
4317 ocfs2_figure_merge_contig_type(struct inode *inode, struct ocfs2_path *path,
4318 struct ocfs2_extent_list *el, int index,
4319 struct ocfs2_extent_rec *split_rec)
4322 enum ocfs2_contig_type ret = CONTIG_NONE;
4323 u32 left_cpos, right_cpos;
4324 struct ocfs2_extent_rec *rec = NULL;
4325 struct ocfs2_extent_list *new_el;
4326 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4327 struct buffer_head *bh;
4328 struct ocfs2_extent_block *eb;
4331 rec = &el->l_recs[index - 1];
4332 } else if (path->p_tree_depth > 0) {
4333 status = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
4338 if (left_cpos != 0) {
4339 left_path = ocfs2_new_path_from_path(path);
4343 status = ocfs2_find_path(inode, left_path, left_cpos);
4347 new_el = path_leaf_el(left_path);
4349 if (le16_to_cpu(new_el->l_next_free_rec) !=
4350 le16_to_cpu(new_el->l_count)) {
4351 bh = path_leaf_bh(left_path);
4352 eb = (struct ocfs2_extent_block *)bh->b_data;
4353 ocfs2_error(inode->i_sb,
4354 "Extent block #%llu has an "
4355 "invalid l_next_free_rec of "
4356 "%d. It should have "
4357 "matched the l_count of %d",
4358 (unsigned long long)le64_to_cpu(eb->h_blkno),
4359 le16_to_cpu(new_el->l_next_free_rec),
4360 le16_to_cpu(new_el->l_count));
4364 rec = &new_el->l_recs[
4365 le16_to_cpu(new_el->l_next_free_rec) - 1];
4370 * We're careful to check for an empty extent record here -
4371 * the merge code will know what to do if it sees one.
4374 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4375 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4378 ret = ocfs2_extent_contig(inode, rec, split_rec);
4383 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4384 rec = &el->l_recs[index + 1];
4385 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4386 path->p_tree_depth > 0) {
4387 status = ocfs2_find_cpos_for_right_leaf(inode->i_sb,
4392 if (right_cpos == 0)
4395 right_path = ocfs2_new_path_from_path(path);
4399 status = ocfs2_find_path(inode, right_path, right_cpos);
4403 new_el = path_leaf_el(right_path);
4404 rec = &new_el->l_recs[0];
4405 if (ocfs2_is_empty_extent(rec)) {
4406 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4407 bh = path_leaf_bh(right_path);
4408 eb = (struct ocfs2_extent_block *)bh->b_data;
4409 ocfs2_error(inode->i_sb,
4410 "Extent block #%llu has an "
4411 "invalid l_next_free_rec of %d",
4412 (unsigned long long)le64_to_cpu(eb->h_blkno),
4413 le16_to_cpu(new_el->l_next_free_rec));
4417 rec = &new_el->l_recs[1];
4422 enum ocfs2_contig_type contig_type;
4424 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
4426 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4427 ret = CONTIG_LEFTRIGHT;
4428 else if (ret == CONTIG_NONE)
4434 ocfs2_free_path(left_path);
4436 ocfs2_free_path(right_path);
4441 static void ocfs2_figure_contig_type(struct inode *inode,
4442 struct ocfs2_insert_type *insert,
4443 struct ocfs2_extent_list *el,
4444 struct ocfs2_extent_rec *insert_rec,
4445 struct ocfs2_extent_tree *et)
4448 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4450 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4452 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4453 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
4455 if (contig_type != CONTIG_NONE) {
4456 insert->ins_contig_index = i;
4460 insert->ins_contig = contig_type;
4462 if (insert->ins_contig != CONTIG_NONE) {
4463 struct ocfs2_extent_rec *rec =
4464 &el->l_recs[insert->ins_contig_index];
4465 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4466 le16_to_cpu(insert_rec->e_leaf_clusters);
4469 * Caller might want us to limit the size of extents, don't
4470 * calculate contiguousness if we might exceed that limit.
4472 if (et->et_max_leaf_clusters &&
4473 (len > et->et_max_leaf_clusters))
4474 insert->ins_contig = CONTIG_NONE;
4479 * This should only be called against the righmost leaf extent list.
4481 * ocfs2_figure_appending_type() will figure out whether we'll have to
4482 * insert at the tail of the rightmost leaf.
4484 * This should also work against the root extent list for tree's with 0
4485 * depth. If we consider the root extent list to be the rightmost leaf node
4486 * then the logic here makes sense.
4488 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4489 struct ocfs2_extent_list *el,
4490 struct ocfs2_extent_rec *insert_rec)
4493 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4494 struct ocfs2_extent_rec *rec;
4496 insert->ins_appending = APPEND_NONE;
4498 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4500 if (!el->l_next_free_rec)
4501 goto set_tail_append;
4503 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4504 /* Were all records empty? */
4505 if (le16_to_cpu(el->l_next_free_rec) == 1)
4506 goto set_tail_append;
4509 i = le16_to_cpu(el->l_next_free_rec) - 1;
4510 rec = &el->l_recs[i];
4513 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4514 goto set_tail_append;
4519 insert->ins_appending = APPEND_TAIL;
4523 * Helper function called at the begining of an insert.
4525 * This computes a few things that are commonly used in the process of
4526 * inserting into the btree:
4527 * - Whether the new extent is contiguous with an existing one.
4528 * - The current tree depth.
4529 * - Whether the insert is an appending one.
4530 * - The total # of free records in the tree.
4532 * All of the information is stored on the ocfs2_insert_type
4535 static int ocfs2_figure_insert_type(struct inode *inode,
4536 struct ocfs2_extent_tree *et,
4537 struct buffer_head **last_eb_bh,
4538 struct ocfs2_extent_rec *insert_rec,
4540 struct ocfs2_insert_type *insert)
4543 struct ocfs2_extent_block *eb;
4544 struct ocfs2_extent_list *el;
4545 struct ocfs2_path *path = NULL;
4546 struct buffer_head *bh = NULL;
4548 insert->ins_split = SPLIT_NONE;
4550 el = et->et_root_el;
4551 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4553 if (el->l_tree_depth) {
4555 * If we have tree depth, we read in the
4556 * rightmost extent block ahead of time as
4557 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4558 * may want it later.
4560 ret = ocfs2_read_extent_block(inode,
4561 ocfs2_et_get_last_eb_blk(et),
4567 eb = (struct ocfs2_extent_block *) bh->b_data;
4572 * Unless we have a contiguous insert, we'll need to know if
4573 * there is room left in our allocation tree for another
4576 * XXX: This test is simplistic, we can search for empty
4577 * extent records too.
4579 *free_records = le16_to_cpu(el->l_count) -
4580 le16_to_cpu(el->l_next_free_rec);
4582 if (!insert->ins_tree_depth) {
4583 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4584 ocfs2_figure_appending_type(insert, el, insert_rec);
4588 path = ocfs2_new_path_from_et(et);
4596 * In the case that we're inserting past what the tree
4597 * currently accounts for, ocfs2_find_path() will return for
4598 * us the rightmost tree path. This is accounted for below in
4599 * the appending code.
4601 ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
4607 el = path_leaf_el(path);
4610 * Now that we have the path, there's two things we want to determine:
4611 * 1) Contiguousness (also set contig_index if this is so)
4613 * 2) Are we doing an append? We can trivially break this up
4614 * into two types of appends: simple record append, or a
4615 * rotate inside the tail leaf.
4617 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et);
4620 * The insert code isn't quite ready to deal with all cases of
4621 * left contiguousness. Specifically, if it's an insert into
4622 * the 1st record in a leaf, it will require the adjustment of
4623 * cluster count on the last record of the path directly to it's
4624 * left. For now, just catch that case and fool the layers
4625 * above us. This works just fine for tree_depth == 0, which
4626 * is why we allow that above.
4628 if (insert->ins_contig == CONTIG_LEFT &&
4629 insert->ins_contig_index == 0)
4630 insert->ins_contig = CONTIG_NONE;
4633 * Ok, so we can simply compare against last_eb to figure out
4634 * whether the path doesn't exist. This will only happen in
4635 * the case that we're doing a tail append, so maybe we can
4636 * take advantage of that information somehow.
4638 if (ocfs2_et_get_last_eb_blk(et) ==
4639 path_leaf_bh(path)->b_blocknr) {
4641 * Ok, ocfs2_find_path() returned us the rightmost
4642 * tree path. This might be an appending insert. There are
4644 * 1) We're doing a true append at the tail:
4645 * -This might even be off the end of the leaf
4646 * 2) We're "appending" by rotating in the tail
4648 ocfs2_figure_appending_type(insert, el, insert_rec);
4652 ocfs2_free_path(path);
4662 * Insert an extent into an inode btree.
4664 * The caller needs to update fe->i_clusters
4666 int ocfs2_insert_extent(struct ocfs2_super *osb,
4668 struct inode *inode,
4669 struct ocfs2_extent_tree *et,
4674 struct ocfs2_alloc_context *meta_ac)
4677 int uninitialized_var(free_records);
4678 struct buffer_head *last_eb_bh = NULL;
4679 struct ocfs2_insert_type insert = {0, };
4680 struct ocfs2_extent_rec rec;
4682 mlog(0, "add %u clusters at position %u to inode %llu\n",
4683 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4685 memset(&rec, 0, sizeof(rec));
4686 rec.e_cpos = cpu_to_le32(cpos);
4687 rec.e_blkno = cpu_to_le64(start_blk);
4688 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4689 rec.e_flags = flags;
4690 status = ocfs2_et_insert_check(inode, et, &rec);
4696 status = ocfs2_figure_insert_type(inode, et, &last_eb_bh, &rec,
4697 &free_records, &insert);
4703 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4704 "Insert.contig_index: %d, Insert.free_records: %d, "
4705 "Insert.tree_depth: %d\n",
4706 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
4707 free_records, insert.ins_tree_depth);
4709 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4710 status = ocfs2_grow_tree(inode, handle, et,
4711 &insert.ins_tree_depth, &last_eb_bh,
4719 /* Finally, we can add clusters. This might rotate the tree for us. */
4720 status = ocfs2_do_insert_extent(inode, handle, et, &rec, &insert);
4723 else if (et->et_ops == &ocfs2_dinode_et_ops)
4724 ocfs2_extent_map_insert_rec(inode, &rec);
4734 * Allcate and add clusters into the extent b-tree.
4735 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4736 * The extent b-tree's root is specified by et, and
4737 * it is not limited to the file storage. Any extent tree can use this
4738 * function if it implements the proper ocfs2_extent_tree.
4740 int ocfs2_add_clusters_in_btree(struct ocfs2_super *osb,
4741 struct inode *inode,
4742 u32 *logical_offset,
4743 u32 clusters_to_add,
4745 struct ocfs2_extent_tree *et,
4747 struct ocfs2_alloc_context *data_ac,
4748 struct ocfs2_alloc_context *meta_ac,
4749 enum ocfs2_alloc_restarted *reason_ret)
4753 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4754 u32 bit_off, num_bits;
4758 BUG_ON(!clusters_to_add);
4761 flags = OCFS2_EXT_UNWRITTEN;
4763 free_extents = ocfs2_num_free_extents(osb, inode, et);
4764 if (free_extents < 0) {
4765 status = free_extents;
4770 /* there are two cases which could cause us to EAGAIN in the
4771 * we-need-more-metadata case:
4772 * 1) we haven't reserved *any*
4773 * 2) we are so fragmented, we've needed to add metadata too
4775 if (!free_extents && !meta_ac) {
4776 mlog(0, "we haven't reserved any metadata!\n");
4778 reason = RESTART_META;
4780 } else if ((!free_extents)
4781 && (ocfs2_alloc_context_bits_left(meta_ac)
4782 < ocfs2_extend_meta_needed(et->et_root_el))) {
4783 mlog(0, "filesystem is really fragmented...\n");
4785 reason = RESTART_META;
4789 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
4790 clusters_to_add, &bit_off, &num_bits);
4792 if (status != -ENOSPC)
4797 BUG_ON(num_bits > clusters_to_add);
4799 /* reserve our write early -- insert_extent may update the tree root */
4800 status = ocfs2_et_root_journal_access(handle, INODE_CACHE(inode), et,
4801 OCFS2_JOURNAL_ACCESS_WRITE);
4807 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4808 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
4809 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
4810 status = ocfs2_insert_extent(osb, handle, inode, et,
4811 *logical_offset, block,
4812 num_bits, flags, meta_ac);
4818 status = ocfs2_journal_dirty(handle, et->et_root_bh);
4824 clusters_to_add -= num_bits;
4825 *logical_offset += num_bits;
4827 if (clusters_to_add) {
4828 mlog(0, "need to alloc once more, wanted = %u\n",
4831 reason = RESTART_TRANS;
4837 *reason_ret = reason;
4841 static void ocfs2_make_right_split_rec(struct super_block *sb,
4842 struct ocfs2_extent_rec *split_rec,
4844 struct ocfs2_extent_rec *rec)
4846 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4847 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4849 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4851 split_rec->e_cpos = cpu_to_le32(cpos);
4852 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4854 split_rec->e_blkno = rec->e_blkno;
4855 le64_add_cpu(&split_rec->e_blkno,
4856 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4858 split_rec->e_flags = rec->e_flags;
4861 static int ocfs2_split_and_insert(struct inode *inode,
4863 struct ocfs2_path *path,
4864 struct ocfs2_extent_tree *et,
4865 struct buffer_head **last_eb_bh,
4867 struct ocfs2_extent_rec *orig_split_rec,
4868 struct ocfs2_alloc_context *meta_ac)
4871 unsigned int insert_range, rec_range, do_leftright = 0;
4872 struct ocfs2_extent_rec tmprec;
4873 struct ocfs2_extent_list *rightmost_el;
4874 struct ocfs2_extent_rec rec;
4875 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4876 struct ocfs2_insert_type insert;
4877 struct ocfs2_extent_block *eb;
4881 * Store a copy of the record on the stack - it might move
4882 * around as the tree is manipulated below.
4884 rec = path_leaf_el(path)->l_recs[split_index];
4886 rightmost_el = et->et_root_el;
4888 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4890 BUG_ON(!(*last_eb_bh));
4891 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4892 rightmost_el = &eb->h_list;
4895 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4896 le16_to_cpu(rightmost_el->l_count)) {
4897 ret = ocfs2_grow_tree(inode, handle, et,
4898 &depth, last_eb_bh, meta_ac);
4905 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4906 insert.ins_appending = APPEND_NONE;
4907 insert.ins_contig = CONTIG_NONE;
4908 insert.ins_tree_depth = depth;
4910 insert_range = le32_to_cpu(split_rec.e_cpos) +
4911 le16_to_cpu(split_rec.e_leaf_clusters);
4912 rec_range = le32_to_cpu(rec.e_cpos) +
4913 le16_to_cpu(rec.e_leaf_clusters);
4915 if (split_rec.e_cpos == rec.e_cpos) {
4916 insert.ins_split = SPLIT_LEFT;
4917 } else if (insert_range == rec_range) {
4918 insert.ins_split = SPLIT_RIGHT;
4921 * Left/right split. We fake this as a right split
4922 * first and then make a second pass as a left split.
4924 insert.ins_split = SPLIT_RIGHT;
4926 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
4931 BUG_ON(do_leftright);
4935 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
4941 if (do_leftright == 1) {
4943 struct ocfs2_extent_list *el;
4946 split_rec = *orig_split_rec;
4948 ocfs2_reinit_path(path, 1);
4950 cpos = le32_to_cpu(split_rec.e_cpos);
4951 ret = ocfs2_find_path(inode, path, cpos);
4957 el = path_leaf_el(path);
4958 split_index = ocfs2_search_extent_list(el, cpos);
4966 static int ocfs2_replace_extent_rec(struct inode *inode,
4968 struct ocfs2_path *path,
4969 struct ocfs2_extent_list *el,
4971 struct ocfs2_extent_rec *split_rec)
4975 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), path,
4976 path_num_items(path) - 1);
4982 el->l_recs[split_index] = *split_rec;
4984 ocfs2_journal_dirty(handle, path_leaf_bh(path));
4990 * Mark part or all of the extent record at split_index in the leaf
4991 * pointed to by path as written. This removes the unwritten
4994 * Care is taken to handle contiguousness so as to not grow the tree.
4996 * meta_ac is not strictly necessary - we only truly need it if growth
4997 * of the tree is required. All other cases will degrade into a less
4998 * optimal tree layout.
5000 * last_eb_bh should be the rightmost leaf block for any extent
5001 * btree. Since a split may grow the tree or a merge might shrink it,
5002 * the caller cannot trust the contents of that buffer after this call.
5004 * This code is optimized for readability - several passes might be
5005 * made over certain portions of the tree. All of those blocks will
5006 * have been brought into cache (and pinned via the journal), so the
5007 * extra overhead is not expressed in terms of disk reads.
5009 static int __ocfs2_mark_extent_written(struct inode *inode,
5010 struct ocfs2_extent_tree *et,
5012 struct ocfs2_path *path,
5014 struct ocfs2_extent_rec *split_rec,
5015 struct ocfs2_alloc_context *meta_ac,
5016 struct ocfs2_cached_dealloc_ctxt *dealloc)
5019 struct ocfs2_extent_list *el = path_leaf_el(path);
5020 struct buffer_head *last_eb_bh = NULL;
5021 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5022 struct ocfs2_merge_ctxt ctxt;
5023 struct ocfs2_extent_list *rightmost_el;
5025 if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) {
5031 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5032 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5033 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5039 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, path, el,
5044 * The core merge / split code wants to know how much room is
5045 * left in this inodes allocation tree, so we pass the
5046 * rightmost extent list.
5048 if (path->p_tree_depth) {
5049 struct ocfs2_extent_block *eb;
5051 ret = ocfs2_read_extent_block(inode,
5052 ocfs2_et_get_last_eb_blk(et),
5059 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5060 rightmost_el = &eb->h_list;
5062 rightmost_el = path_root_el(path);
5064 if (rec->e_cpos == split_rec->e_cpos &&
5065 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5066 ctxt.c_split_covers_rec = 1;
5068 ctxt.c_split_covers_rec = 0;
5070 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5072 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
5073 split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent,
5074 ctxt.c_split_covers_rec);
5076 if (ctxt.c_contig_type == CONTIG_NONE) {
5077 if (ctxt.c_split_covers_rec)
5078 ret = ocfs2_replace_extent_rec(inode, handle,
5080 split_index, split_rec);
5082 ret = ocfs2_split_and_insert(inode, handle, path, et,
5083 &last_eb_bh, split_index,
5084 split_rec, meta_ac);
5088 ret = ocfs2_try_to_merge_extent(inode, handle, path,
5089 split_index, split_rec,
5090 dealloc, &ctxt, et);
5101 * Mark the already-existing extent at cpos as written for len clusters.
5103 * If the existing extent is larger than the request, initiate a
5104 * split. An attempt will be made at merging with adjacent extents.
5106 * The caller is responsible for passing down meta_ac if we'll need it.
5108 int ocfs2_mark_extent_written(struct inode *inode,
5109 struct ocfs2_extent_tree *et,
5110 handle_t *handle, u32 cpos, u32 len, u32 phys,
5111 struct ocfs2_alloc_context *meta_ac,
5112 struct ocfs2_cached_dealloc_ctxt *dealloc)
5115 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
5116 struct ocfs2_extent_rec split_rec;
5117 struct ocfs2_path *left_path = NULL;
5118 struct ocfs2_extent_list *el;
5120 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
5121 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
5123 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5124 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
5125 "that are being written to, but the feature bit "
5126 "is not set in the super block.",
5127 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5133 * XXX: This should be fixed up so that we just re-insert the
5134 * next extent records.
5136 * XXX: This is a hack on the extent tree, maybe it should be
5139 if (et->et_ops == &ocfs2_dinode_et_ops)
5140 ocfs2_extent_map_trunc(inode, 0);
5142 left_path = ocfs2_new_path_from_et(et);
5149 ret = ocfs2_find_path(inode, left_path, cpos);
5154 el = path_leaf_el(left_path);
5156 index = ocfs2_search_extent_list(el, cpos);
5157 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5158 ocfs2_error(inode->i_sb,
5159 "Inode %llu has an extent at cpos %u which can no "
5160 "longer be found.\n",
5161 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5166 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5167 split_rec.e_cpos = cpu_to_le32(cpos);
5168 split_rec.e_leaf_clusters = cpu_to_le16(len);
5169 split_rec.e_blkno = cpu_to_le64(start_blkno);
5170 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
5171 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
5173 ret = __ocfs2_mark_extent_written(inode, et, handle, left_path,
5174 index, &split_rec, meta_ac,
5180 ocfs2_free_path(left_path);
5184 static int ocfs2_split_tree(struct inode *inode, struct ocfs2_extent_tree *et,
5185 handle_t *handle, struct ocfs2_path *path,
5186 int index, u32 new_range,
5187 struct ocfs2_alloc_context *meta_ac)
5189 int ret, depth, credits = handle->h_buffer_credits;
5190 struct buffer_head *last_eb_bh = NULL;
5191 struct ocfs2_extent_block *eb;
5192 struct ocfs2_extent_list *rightmost_el, *el;
5193 struct ocfs2_extent_rec split_rec;
5194 struct ocfs2_extent_rec *rec;
5195 struct ocfs2_insert_type insert;
5198 * Setup the record to split before we grow the tree.
5200 el = path_leaf_el(path);
5201 rec = &el->l_recs[index];
5202 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
5204 depth = path->p_tree_depth;
5206 ret = ocfs2_read_extent_block(inode,
5207 ocfs2_et_get_last_eb_blk(et),
5214 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5215 rightmost_el = &eb->h_list;
5217 rightmost_el = path_leaf_el(path);
5219 credits += path->p_tree_depth +
5220 ocfs2_extend_meta_needed(et->et_root_el);
5221 ret = ocfs2_extend_trans(handle, credits);
5227 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5228 le16_to_cpu(rightmost_el->l_count)) {
5229 ret = ocfs2_grow_tree(inode, handle, et, &depth, &last_eb_bh,
5237 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5238 insert.ins_appending = APPEND_NONE;
5239 insert.ins_contig = CONTIG_NONE;
5240 insert.ins_split = SPLIT_RIGHT;
5241 insert.ins_tree_depth = depth;
5243 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert);
5252 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
5253 struct ocfs2_path *path, int index,
5254 struct ocfs2_cached_dealloc_ctxt *dealloc,
5256 struct ocfs2_extent_tree *et)
5259 u32 left_cpos, rec_range, trunc_range;
5260 int wants_rotate = 0, is_rightmost_tree_rec = 0;
5261 struct super_block *sb = inode->i_sb;
5262 struct ocfs2_path *left_path = NULL;
5263 struct ocfs2_extent_list *el = path_leaf_el(path);
5264 struct ocfs2_extent_rec *rec;
5265 struct ocfs2_extent_block *eb;
5267 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5268 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5277 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5278 path->p_tree_depth) {
5280 * Check whether this is the rightmost tree record. If
5281 * we remove all of this record or part of its right
5282 * edge then an update of the record lengths above it
5285 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5286 if (eb->h_next_leaf_blk == 0)
5287 is_rightmost_tree_rec = 1;
5290 rec = &el->l_recs[index];
5291 if (index == 0 && path->p_tree_depth &&
5292 le32_to_cpu(rec->e_cpos) == cpos) {
5294 * Changing the leftmost offset (via partial or whole
5295 * record truncate) of an interior (or rightmost) path
5296 * means we have to update the subtree that is formed
5297 * by this leaf and the one to it's left.
5299 * There are two cases we can skip:
5300 * 1) Path is the leftmost one in our inode tree.
5301 * 2) The leaf is rightmost and will be empty after
5302 * we remove the extent record - the rotate code
5303 * knows how to update the newly formed edge.
5306 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
5313 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5314 left_path = ocfs2_new_path_from_path(path);
5321 ret = ocfs2_find_path(inode, left_path, left_cpos);
5329 ret = ocfs2_extend_rotate_transaction(handle, 0,
5330 handle->h_buffer_credits,
5337 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
5343 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, left_path);
5349 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5350 trunc_range = cpos + len;
5352 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5355 memset(rec, 0, sizeof(*rec));
5356 ocfs2_cleanup_merge(el, index);
5359 next_free = le16_to_cpu(el->l_next_free_rec);
5360 if (is_rightmost_tree_rec && next_free > 1) {
5362 * We skip the edge update if this path will
5363 * be deleted by the rotate code.
5365 rec = &el->l_recs[next_free - 1];
5366 ocfs2_adjust_rightmost_records(inode, handle, path,
5369 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5370 /* Remove leftmost portion of the record. */
5371 le32_add_cpu(&rec->e_cpos, len);
5372 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5373 le16_add_cpu(&rec->e_leaf_clusters, -len);
5374 } else if (rec_range == trunc_range) {
5375 /* Remove rightmost portion of the record */
5376 le16_add_cpu(&rec->e_leaf_clusters, -len);
5377 if (is_rightmost_tree_rec)
5378 ocfs2_adjust_rightmost_records(inode, handle, path, rec);
5380 /* Caller should have trapped this. */
5381 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
5382 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
5383 le32_to_cpu(rec->e_cpos),
5384 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5391 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
5392 ocfs2_complete_edge_insert(inode, handle, left_path, path,
5396 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5398 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc, et);
5405 ocfs2_free_path(left_path);
5409 int ocfs2_remove_extent(struct inode *inode,
5410 struct ocfs2_extent_tree *et,
5411 u32 cpos, u32 len, handle_t *handle,
5412 struct ocfs2_alloc_context *meta_ac,
5413 struct ocfs2_cached_dealloc_ctxt *dealloc)
5416 u32 rec_range, trunc_range;
5417 struct ocfs2_extent_rec *rec;
5418 struct ocfs2_extent_list *el;
5419 struct ocfs2_path *path = NULL;
5421 ocfs2_extent_map_trunc(inode, 0);
5423 path = ocfs2_new_path_from_et(et);
5430 ret = ocfs2_find_path(inode, path, cpos);
5436 el = path_leaf_el(path);
5437 index = ocfs2_search_extent_list(el, cpos);
5438 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5439 ocfs2_error(inode->i_sb,
5440 "Inode %llu has an extent at cpos %u which can no "
5441 "longer be found.\n",
5442 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
5448 * We have 3 cases of extent removal:
5449 * 1) Range covers the entire extent rec
5450 * 2) Range begins or ends on one edge of the extent rec
5451 * 3) Range is in the middle of the extent rec (no shared edges)
5453 * For case 1 we remove the extent rec and left rotate to
5456 * For case 2 we just shrink the existing extent rec, with a
5457 * tree update if the shrinking edge is also the edge of an
5460 * For case 3 we do a right split to turn the extent rec into
5461 * something case 2 can handle.
5463 rec = &el->l_recs[index];
5464 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5465 trunc_range = cpos + len;
5467 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5469 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
5470 "(cpos %u, len %u)\n",
5471 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
5472 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
5474 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5475 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5482 ret = ocfs2_split_tree(inode, et, handle, path, index,
5483 trunc_range, meta_ac);
5490 * The split could have manipulated the tree enough to
5491 * move the record location, so we have to look for it again.
5493 ocfs2_reinit_path(path, 1);
5495 ret = ocfs2_find_path(inode, path, cpos);
5501 el = path_leaf_el(path);
5502 index = ocfs2_search_extent_list(el, cpos);
5503 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5504 ocfs2_error(inode->i_sb,
5505 "Inode %llu: split at cpos %u lost record.",
5506 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5513 * Double check our values here. If anything is fishy,
5514 * it's easier to catch it at the top level.
5516 rec = &el->l_recs[index];
5517 rec_range = le32_to_cpu(rec->e_cpos) +
5518 ocfs2_rec_clusters(el, rec);
5519 if (rec_range != trunc_range) {
5520 ocfs2_error(inode->i_sb,
5521 "Inode %llu: error after split at cpos %u"
5522 "trunc len %u, existing record is (%u,%u)",
5523 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5524 cpos, len, le32_to_cpu(rec->e_cpos),
5525 ocfs2_rec_clusters(el, rec));
5530 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
5539 ocfs2_free_path(path);
5543 int ocfs2_remove_btree_range(struct inode *inode,
5544 struct ocfs2_extent_tree *et,
5545 u32 cpos, u32 phys_cpos, u32 len,
5546 struct ocfs2_cached_dealloc_ctxt *dealloc)
5549 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5550 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5551 struct inode *tl_inode = osb->osb_tl_inode;
5553 struct ocfs2_alloc_context *meta_ac = NULL;
5555 ret = ocfs2_lock_allocators(inode, et, 0, 1, NULL, &meta_ac);
5561 mutex_lock(&tl_inode->i_mutex);
5563 if (ocfs2_truncate_log_needs_flush(osb)) {
5564 ret = __ocfs2_flush_truncate_log(osb);
5571 handle = ocfs2_start_trans(osb, ocfs2_remove_extent_credits(osb->sb));
5572 if (IS_ERR(handle)) {
5573 ret = PTR_ERR(handle);
5578 ret = ocfs2_et_root_journal_access(handle, INODE_CACHE(inode), et,
5579 OCFS2_JOURNAL_ACCESS_WRITE);
5585 vfs_dq_free_space_nodirty(inode,
5586 ocfs2_clusters_to_bytes(inode->i_sb, len));
5588 ret = ocfs2_remove_extent(inode, et, cpos, len, handle, meta_ac,
5595 ocfs2_et_update_clusters(inode, et, -len);
5597 ret = ocfs2_journal_dirty(handle, et->et_root_bh);
5603 ret = ocfs2_truncate_log_append(osb, handle, phys_blkno, len);
5608 ocfs2_commit_trans(osb, handle);
5610 mutex_unlock(&tl_inode->i_mutex);
5613 ocfs2_free_alloc_context(meta_ac);
5618 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5620 struct buffer_head *tl_bh = osb->osb_tl_bh;
5621 struct ocfs2_dinode *di;
5622 struct ocfs2_truncate_log *tl;
5624 di = (struct ocfs2_dinode *) tl_bh->b_data;
5625 tl = &di->id2.i_dealloc;
5627 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5628 "slot %d, invalid truncate log parameters: used = "
5629 "%u, count = %u\n", osb->slot_num,
5630 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5631 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5634 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5635 unsigned int new_start)
5637 unsigned int tail_index;
5638 unsigned int current_tail;
5640 /* No records, nothing to coalesce */
5641 if (!le16_to_cpu(tl->tl_used))
5644 tail_index = le16_to_cpu(tl->tl_used) - 1;
5645 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5646 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5648 return current_tail == new_start;
5651 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5654 unsigned int num_clusters)
5657 unsigned int start_cluster, tl_count;
5658 struct inode *tl_inode = osb->osb_tl_inode;
5659 struct buffer_head *tl_bh = osb->osb_tl_bh;
5660 struct ocfs2_dinode *di;
5661 struct ocfs2_truncate_log *tl;
5663 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5664 (unsigned long long)start_blk, num_clusters);
5666 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5668 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5670 di = (struct ocfs2_dinode *) tl_bh->b_data;
5672 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5673 * by the underlying call to ocfs2_read_inode_block(), so any
5674 * corruption is a code bug */
5675 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5677 tl = &di->id2.i_dealloc;
5678 tl_count = le16_to_cpu(tl->tl_count);
5679 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5681 "Truncate record count on #%llu invalid "
5682 "wanted %u, actual %u\n",
5683 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5684 ocfs2_truncate_recs_per_inode(osb->sb),
5685 le16_to_cpu(tl->tl_count));
5687 /* Caller should have known to flush before calling us. */
5688 index = le16_to_cpu(tl->tl_used);
5689 if (index >= tl_count) {
5695 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5696 OCFS2_JOURNAL_ACCESS_WRITE);
5702 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5703 "%llu (index = %d)\n", num_clusters, start_cluster,
5704 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
5706 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5708 * Move index back to the record we are coalescing with.
5709 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5713 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5714 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5715 index, le32_to_cpu(tl->tl_recs[index].t_start),
5718 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5719 tl->tl_used = cpu_to_le16(index + 1);
5721 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5723 status = ocfs2_journal_dirty(handle, tl_bh);
5734 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5736 struct inode *data_alloc_inode,
5737 struct buffer_head *data_alloc_bh)
5741 unsigned int num_clusters;
5743 struct ocfs2_truncate_rec rec;
5744 struct ocfs2_dinode *di;
5745 struct ocfs2_truncate_log *tl;
5746 struct inode *tl_inode = osb->osb_tl_inode;
5747 struct buffer_head *tl_bh = osb->osb_tl_bh;
5751 di = (struct ocfs2_dinode *) tl_bh->b_data;
5752 tl = &di->id2.i_dealloc;
5753 i = le16_to_cpu(tl->tl_used) - 1;
5755 /* Caller has given us at least enough credits to
5756 * update the truncate log dinode */
5757 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5758 OCFS2_JOURNAL_ACCESS_WRITE);
5764 tl->tl_used = cpu_to_le16(i);
5766 status = ocfs2_journal_dirty(handle, tl_bh);
5772 /* TODO: Perhaps we can calculate the bulk of the
5773 * credits up front rather than extending like
5775 status = ocfs2_extend_trans(handle,
5776 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5782 rec = tl->tl_recs[i];
5783 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5784 le32_to_cpu(rec.t_start));
5785 num_clusters = le32_to_cpu(rec.t_clusters);
5787 /* if start_blk is not set, we ignore the record as
5790 mlog(0, "free record %d, start = %u, clusters = %u\n",
5791 i, le32_to_cpu(rec.t_start), num_clusters);
5793 status = ocfs2_free_clusters(handle, data_alloc_inode,
5794 data_alloc_bh, start_blk,
5809 /* Expects you to already be holding tl_inode->i_mutex */
5810 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5813 unsigned int num_to_flush;
5815 struct inode *tl_inode = osb->osb_tl_inode;
5816 struct inode *data_alloc_inode = NULL;
5817 struct buffer_head *tl_bh = osb->osb_tl_bh;
5818 struct buffer_head *data_alloc_bh = NULL;
5819 struct ocfs2_dinode *di;
5820 struct ocfs2_truncate_log *tl;
5824 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5826 di = (struct ocfs2_dinode *) tl_bh->b_data;
5828 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5829 * by the underlying call to ocfs2_read_inode_block(), so any
5830 * corruption is a code bug */
5831 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5833 tl = &di->id2.i_dealloc;
5834 num_to_flush = le16_to_cpu(tl->tl_used);
5835 mlog(0, "Flush %u records from truncate log #%llu\n",
5836 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
5837 if (!num_to_flush) {
5842 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5843 GLOBAL_BITMAP_SYSTEM_INODE,
5844 OCFS2_INVALID_SLOT);
5845 if (!data_alloc_inode) {
5847 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5851 mutex_lock(&data_alloc_inode->i_mutex);
5853 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5859 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5860 if (IS_ERR(handle)) {
5861 status = PTR_ERR(handle);
5866 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
5871 ocfs2_commit_trans(osb, handle);
5874 brelse(data_alloc_bh);
5875 ocfs2_inode_unlock(data_alloc_inode, 1);
5878 mutex_unlock(&data_alloc_inode->i_mutex);
5879 iput(data_alloc_inode);
5886 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5889 struct inode *tl_inode = osb->osb_tl_inode;
5891 mutex_lock(&tl_inode->i_mutex);
5892 status = __ocfs2_flush_truncate_log(osb);
5893 mutex_unlock(&tl_inode->i_mutex);
5898 static void ocfs2_truncate_log_worker(struct work_struct *work)
5901 struct ocfs2_super *osb =
5902 container_of(work, struct ocfs2_super,
5903 osb_truncate_log_wq.work);
5907 status = ocfs2_flush_truncate_log(osb);
5911 ocfs2_init_inode_steal_slot(osb);
5916 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5917 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
5920 if (osb->osb_tl_inode) {
5921 /* We want to push off log flushes while truncates are
5924 cancel_delayed_work(&osb->osb_truncate_log_wq);
5926 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
5927 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
5931 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
5933 struct inode **tl_inode,
5934 struct buffer_head **tl_bh)
5937 struct inode *inode = NULL;
5938 struct buffer_head *bh = NULL;
5940 inode = ocfs2_get_system_file_inode(osb,
5941 TRUNCATE_LOG_SYSTEM_INODE,
5945 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
5949 status = ocfs2_read_inode_block(inode, &bh);
5963 /* called during the 1st stage of node recovery. we stamp a clean
5964 * truncate log and pass back a copy for processing later. if the
5965 * truncate log does not require processing, a *tl_copy is set to
5967 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
5969 struct ocfs2_dinode **tl_copy)
5972 struct inode *tl_inode = NULL;
5973 struct buffer_head *tl_bh = NULL;
5974 struct ocfs2_dinode *di;
5975 struct ocfs2_truncate_log *tl;
5979 mlog(0, "recover truncate log from slot %d\n", slot_num);
5981 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
5987 di = (struct ocfs2_dinode *) tl_bh->b_data;
5989 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
5990 * validated by the underlying call to ocfs2_read_inode_block(),
5991 * so any corruption is a code bug */
5992 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5994 tl = &di->id2.i_dealloc;
5995 if (le16_to_cpu(tl->tl_used)) {
5996 mlog(0, "We'll have %u logs to recover\n",
5997 le16_to_cpu(tl->tl_used));
5999 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6006 /* Assuming the write-out below goes well, this copy
6007 * will be passed back to recovery for processing. */
6008 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6010 /* All we need to do to clear the truncate log is set
6014 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6015 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6027 if (status < 0 && (*tl_copy)) {
6036 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6037 struct ocfs2_dinode *tl_copy)
6041 unsigned int clusters, num_recs, start_cluster;
6044 struct inode *tl_inode = osb->osb_tl_inode;
6045 struct ocfs2_truncate_log *tl;
6049 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6050 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6054 tl = &tl_copy->id2.i_dealloc;
6055 num_recs = le16_to_cpu(tl->tl_used);
6056 mlog(0, "cleanup %u records from %llu\n", num_recs,
6057 (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
6059 mutex_lock(&tl_inode->i_mutex);
6060 for(i = 0; i < num_recs; i++) {
6061 if (ocfs2_truncate_log_needs_flush(osb)) {
6062 status = __ocfs2_flush_truncate_log(osb);
6069 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6070 if (IS_ERR(handle)) {
6071 status = PTR_ERR(handle);
6076 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6077 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6078 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6080 status = ocfs2_truncate_log_append(osb, handle,
6081 start_blk, clusters);
6082 ocfs2_commit_trans(osb, handle);
6090 mutex_unlock(&tl_inode->i_mutex);
6096 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6099 struct inode *tl_inode = osb->osb_tl_inode;
6104 cancel_delayed_work(&osb->osb_truncate_log_wq);
6105 flush_workqueue(ocfs2_wq);
6107 status = ocfs2_flush_truncate_log(osb);
6111 brelse(osb->osb_tl_bh);
6112 iput(osb->osb_tl_inode);
6118 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6121 struct inode *tl_inode = NULL;
6122 struct buffer_head *tl_bh = NULL;
6126 status = ocfs2_get_truncate_log_info(osb,
6133 /* ocfs2_truncate_log_shutdown keys on the existence of
6134 * osb->osb_tl_inode so we don't set any of the osb variables
6135 * until we're sure all is well. */
6136 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6137 ocfs2_truncate_log_worker);
6138 osb->osb_tl_bh = tl_bh;
6139 osb->osb_tl_inode = tl_inode;
6146 * Delayed de-allocation of suballocator blocks.
6148 * Some sets of block de-allocations might involve multiple suballocator inodes.
6150 * The locking for this can get extremely complicated, especially when
6151 * the suballocator inodes to delete from aren't known until deep
6152 * within an unrelated codepath.
6154 * ocfs2_extent_block structures are a good example of this - an inode
6155 * btree could have been grown by any number of nodes each allocating
6156 * out of their own suballoc inode.
6158 * These structures allow the delay of block de-allocation until a
6159 * later time, when locking of multiple cluster inodes won't cause
6164 * Describe a single bit freed from a suballocator. For the block
6165 * suballocators, it represents one block. For the global cluster
6166 * allocator, it represents some clusters and free_bit indicates
6169 struct ocfs2_cached_block_free {
6170 struct ocfs2_cached_block_free *free_next;
6172 unsigned int free_bit;
6175 struct ocfs2_per_slot_free_list {
6176 struct ocfs2_per_slot_free_list *f_next_suballocator;
6179 struct ocfs2_cached_block_free *f_first;
6182 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6185 struct ocfs2_cached_block_free *head)
6190 struct inode *inode;
6191 struct buffer_head *di_bh = NULL;
6192 struct ocfs2_cached_block_free *tmp;
6194 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6201 mutex_lock(&inode->i_mutex);
6203 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6209 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6210 if (IS_ERR(handle)) {
6211 ret = PTR_ERR(handle);
6217 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6219 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
6220 head->free_bit, (unsigned long long)head->free_blk);
6222 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6223 head->free_bit, bg_blkno, 1);
6229 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6236 head = head->free_next;
6241 ocfs2_commit_trans(osb, handle);
6244 ocfs2_inode_unlock(inode, 1);
6247 mutex_unlock(&inode->i_mutex);
6251 /* Premature exit may have left some dangling items. */
6253 head = head->free_next;
6260 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6261 u64 blkno, unsigned int bit)
6264 struct ocfs2_cached_block_free *item;
6266 item = kmalloc(sizeof(*item), GFP_NOFS);
6273 mlog(0, "Insert clusters: (bit %u, blk %llu)\n",
6274 bit, (unsigned long long)blkno);
6276 item->free_blk = blkno;
6277 item->free_bit = bit;
6278 item->free_next = ctxt->c_global_allocator;
6280 ctxt->c_global_allocator = item;
6284 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6285 struct ocfs2_cached_block_free *head)
6287 struct ocfs2_cached_block_free *tmp;
6288 struct inode *tl_inode = osb->osb_tl_inode;
6292 mutex_lock(&tl_inode->i_mutex);
6295 if (ocfs2_truncate_log_needs_flush(osb)) {
6296 ret = __ocfs2_flush_truncate_log(osb);
6303 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6304 if (IS_ERR(handle)) {
6305 ret = PTR_ERR(handle);
6310 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6313 ocfs2_commit_trans(osb, handle);
6315 head = head->free_next;
6324 mutex_unlock(&tl_inode->i_mutex);
6327 /* Premature exit may have left some dangling items. */
6329 head = head->free_next;
6336 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6337 struct ocfs2_cached_dealloc_ctxt *ctxt)
6340 struct ocfs2_per_slot_free_list *fl;
6345 while (ctxt->c_first_suballocator) {
6346 fl = ctxt->c_first_suballocator;
6349 mlog(0, "Free items: (type %u, slot %d)\n",
6350 fl->f_inode_type, fl->f_slot);
6351 ret2 = ocfs2_free_cached_blocks(osb,
6361 ctxt->c_first_suballocator = fl->f_next_suballocator;
6365 if (ctxt->c_global_allocator) {
6366 ret2 = ocfs2_free_cached_clusters(osb,
6367 ctxt->c_global_allocator);
6373 ctxt->c_global_allocator = NULL;
6379 static struct ocfs2_per_slot_free_list *
6380 ocfs2_find_per_slot_free_list(int type,
6382 struct ocfs2_cached_dealloc_ctxt *ctxt)
6384 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6387 if (fl->f_inode_type == type && fl->f_slot == slot)
6390 fl = fl->f_next_suballocator;
6393 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6395 fl->f_inode_type = type;
6398 fl->f_next_suballocator = ctxt->c_first_suballocator;
6400 ctxt->c_first_suballocator = fl;
6405 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6406 int type, int slot, u64 blkno,
6410 struct ocfs2_per_slot_free_list *fl;
6411 struct ocfs2_cached_block_free *item;
6413 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6420 item = kmalloc(sizeof(*item), GFP_NOFS);
6427 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
6428 type, slot, bit, (unsigned long long)blkno);
6430 item->free_blk = blkno;
6431 item->free_bit = bit;
6432 item->free_next = fl->f_first;
6441 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6442 struct ocfs2_extent_block *eb)
6444 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6445 le16_to_cpu(eb->h_suballoc_slot),
6446 le64_to_cpu(eb->h_blkno),
6447 le16_to_cpu(eb->h_suballoc_bit));
6450 /* This function will figure out whether the currently last extent
6451 * block will be deleted, and if it will, what the new last extent
6452 * block will be so we can update his h_next_leaf_blk field, as well
6453 * as the dinodes i_last_eb_blk */
6454 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
6455 unsigned int clusters_to_del,
6456 struct ocfs2_path *path,
6457 struct buffer_head **new_last_eb)
6459 int next_free, ret = 0;
6461 struct ocfs2_extent_rec *rec;
6462 struct ocfs2_extent_block *eb;
6463 struct ocfs2_extent_list *el;
6464 struct buffer_head *bh = NULL;
6466 *new_last_eb = NULL;
6468 /* we have no tree, so of course, no last_eb. */
6469 if (!path->p_tree_depth)
6472 /* trunc to zero special case - this makes tree_depth = 0
6473 * regardless of what it is. */
6474 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
6477 el = path_leaf_el(path);
6478 BUG_ON(!el->l_next_free_rec);
6481 * Make sure that this extent list will actually be empty
6482 * after we clear away the data. We can shortcut out if
6483 * there's more than one non-empty extent in the
6484 * list. Otherwise, a check of the remaining extent is
6487 next_free = le16_to_cpu(el->l_next_free_rec);
6489 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6493 /* We may have a valid extent in index 1, check it. */
6495 rec = &el->l_recs[1];
6498 * Fall through - no more nonempty extents, so we want
6499 * to delete this leaf.
6505 rec = &el->l_recs[0];
6510 * Check it we'll only be trimming off the end of this
6513 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
6517 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
6523 ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
6529 eb = (struct ocfs2_extent_block *) bh->b_data;
6532 /* ocfs2_find_leaf() gets the eb from ocfs2_read_extent_block().
6533 * Any corruption is a code bug. */
6534 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
6537 get_bh(*new_last_eb);
6538 mlog(0, "returning block %llu, (cpos: %u)\n",
6539 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
6547 * Trim some clusters off the rightmost edge of a tree. Only called
6550 * The caller needs to:
6551 * - start journaling of each path component.
6552 * - compute and fully set up any new last ext block
6554 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
6555 handle_t *handle, struct ocfs2_truncate_context *tc,
6556 u32 clusters_to_del, u64 *delete_start)
6558 int ret, i, index = path->p_tree_depth;
6561 struct buffer_head *bh;
6562 struct ocfs2_extent_list *el;
6563 struct ocfs2_extent_rec *rec;
6567 while (index >= 0) {
6568 bh = path->p_node[index].bh;
6569 el = path->p_node[index].el;
6571 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6572 index, (unsigned long long)bh->b_blocknr);
6574 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
6577 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
6578 ocfs2_error(inode->i_sb,
6579 "Inode %lu has invalid ext. block %llu",
6581 (unsigned long long)bh->b_blocknr);
6587 i = le16_to_cpu(el->l_next_free_rec) - 1;
6588 rec = &el->l_recs[i];
6590 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6591 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
6592 ocfs2_rec_clusters(el, rec),
6593 (unsigned long long)le64_to_cpu(rec->e_blkno),
6594 le16_to_cpu(el->l_next_free_rec));
6596 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
6598 if (le16_to_cpu(el->l_tree_depth) == 0) {
6600 * If the leaf block contains a single empty
6601 * extent and no records, we can just remove
6604 if (i == 0 && ocfs2_is_empty_extent(rec)) {
6606 sizeof(struct ocfs2_extent_rec));
6607 el->l_next_free_rec = cpu_to_le16(0);
6613 * Remove any empty extents by shifting things
6614 * left. That should make life much easier on
6615 * the code below. This condition is rare
6616 * enough that we shouldn't see a performance
6619 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
6620 le16_add_cpu(&el->l_next_free_rec, -1);
6623 i < le16_to_cpu(el->l_next_free_rec); i++)
6624 el->l_recs[i] = el->l_recs[i + 1];
6626 memset(&el->l_recs[i], 0,
6627 sizeof(struct ocfs2_extent_rec));
6630 * We've modified our extent list. The
6631 * simplest way to handle this change
6632 * is to being the search from the
6635 goto find_tail_record;
6638 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
6641 * We'll use "new_edge" on our way back up the
6642 * tree to know what our rightmost cpos is.
6644 new_edge = le16_to_cpu(rec->e_leaf_clusters);
6645 new_edge += le32_to_cpu(rec->e_cpos);
6648 * The caller will use this to delete data blocks.
6650 *delete_start = le64_to_cpu(rec->e_blkno)
6651 + ocfs2_clusters_to_blocks(inode->i_sb,
6652 le16_to_cpu(rec->e_leaf_clusters));
6655 * If it's now empty, remove this record.
6657 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
6659 sizeof(struct ocfs2_extent_rec));
6660 le16_add_cpu(&el->l_next_free_rec, -1);
6663 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
6665 sizeof(struct ocfs2_extent_rec));
6666 le16_add_cpu(&el->l_next_free_rec, -1);
6671 /* Can this actually happen? */
6672 if (le16_to_cpu(el->l_next_free_rec) == 0)
6676 * We never actually deleted any clusters
6677 * because our leaf was empty. There's no
6678 * reason to adjust the rightmost edge then.
6683 rec->e_int_clusters = cpu_to_le32(new_edge);
6684 le32_add_cpu(&rec->e_int_clusters,
6685 -le32_to_cpu(rec->e_cpos));
6688 * A deleted child record should have been
6691 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
6695 ret = ocfs2_journal_dirty(handle, bh);
6701 mlog(0, "extent list container %llu, after: record %d: "
6702 "(%u, %u, %llu), next = %u.\n",
6703 (unsigned long long)bh->b_blocknr, i,
6704 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
6705 (unsigned long long)le64_to_cpu(rec->e_blkno),
6706 le16_to_cpu(el->l_next_free_rec));
6709 * We must be careful to only attempt delete of an
6710 * extent block (and not the root inode block).
6712 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
6713 struct ocfs2_extent_block *eb =
6714 (struct ocfs2_extent_block *)bh->b_data;
6717 * Save this for use when processing the
6720 deleted_eb = le64_to_cpu(eb->h_blkno);
6722 mlog(0, "deleting this extent block.\n");
6724 ocfs2_remove_from_cache(INODE_CACHE(inode), bh);
6726 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
6727 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
6728 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
6730 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
6731 /* An error here is not fatal. */
6746 static int ocfs2_do_truncate(struct ocfs2_super *osb,
6747 unsigned int clusters_to_del,
6748 struct inode *inode,
6749 struct buffer_head *fe_bh,
6751 struct ocfs2_truncate_context *tc,
6752 struct ocfs2_path *path)
6755 struct ocfs2_dinode *fe;
6756 struct ocfs2_extent_block *last_eb = NULL;
6757 struct ocfs2_extent_list *el;
6758 struct buffer_head *last_eb_bh = NULL;
6761 fe = (struct ocfs2_dinode *) fe_bh->b_data;
6763 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
6771 * Each component will be touched, so we might as well journal
6772 * here to avoid having to handle errors later.
6774 status = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path);
6781 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), last_eb_bh,
6782 OCFS2_JOURNAL_ACCESS_WRITE);
6788 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6791 el = &(fe->id2.i_list);
6794 * Lower levels depend on this never happening, but it's best
6795 * to check it up here before changing the tree.
6797 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
6798 ocfs2_error(inode->i_sb,
6799 "Inode %lu has an empty extent record, depth %u\n",
6800 inode->i_ino, le16_to_cpu(el->l_tree_depth));
6805 vfs_dq_free_space_nodirty(inode,
6806 ocfs2_clusters_to_bytes(osb->sb, clusters_to_del));
6807 spin_lock(&OCFS2_I(inode)->ip_lock);
6808 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
6810 spin_unlock(&OCFS2_I(inode)->ip_lock);
6811 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
6812 inode->i_blocks = ocfs2_inode_sector_count(inode);
6814 status = ocfs2_trim_tree(inode, path, handle, tc,
6815 clusters_to_del, &delete_blk);
6821 if (le32_to_cpu(fe->i_clusters) == 0) {
6822 /* trunc to zero is a special case. */
6823 el->l_tree_depth = 0;
6824 fe->i_last_eb_blk = 0;
6826 fe->i_last_eb_blk = last_eb->h_blkno;
6828 status = ocfs2_journal_dirty(handle, fe_bh);
6835 /* If there will be a new last extent block, then by
6836 * definition, there cannot be any leaves to the right of
6838 last_eb->h_next_leaf_blk = 0;
6839 status = ocfs2_journal_dirty(handle, last_eb_bh);
6847 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
6861 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6863 set_buffer_uptodate(bh);
6864 mark_buffer_dirty(bh);
6868 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6869 unsigned int from, unsigned int to,
6870 struct page *page, int zero, u64 *phys)
6872 int ret, partial = 0;
6874 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6879 zero_user_segment(page, from, to);
6882 * Need to set the buffers we zero'd into uptodate
6883 * here if they aren't - ocfs2_map_page_blocks()
6884 * might've skipped some
6886 ret = walk_page_buffers(handle, page_buffers(page),
6891 else if (ocfs2_should_order_data(inode)) {
6892 ret = ocfs2_jbd2_file_inode(handle, inode);
6898 SetPageUptodate(page);
6900 flush_dcache_page(page);
6903 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6904 loff_t end, struct page **pages,
6905 int numpages, u64 phys, handle_t *handle)
6909 unsigned int from, to = PAGE_CACHE_SIZE;
6910 struct super_block *sb = inode->i_sb;
6912 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6917 to = PAGE_CACHE_SIZE;
6918 for(i = 0; i < numpages; i++) {
6921 from = start & (PAGE_CACHE_SIZE - 1);
6922 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6923 to = end & (PAGE_CACHE_SIZE - 1);
6925 BUG_ON(from > PAGE_CACHE_SIZE);
6926 BUG_ON(to > PAGE_CACHE_SIZE);
6928 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6931 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6935 ocfs2_unlock_and_free_pages(pages, numpages);
6938 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6939 struct page **pages, int *num)
6941 int numpages, ret = 0;
6942 struct super_block *sb = inode->i_sb;
6943 struct address_space *mapping = inode->i_mapping;
6944 unsigned long index;
6945 loff_t last_page_bytes;
6947 BUG_ON(start > end);
6949 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6950 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6953 last_page_bytes = PAGE_ALIGN(end);
6954 index = start >> PAGE_CACHE_SHIFT;
6956 pages[numpages] = grab_cache_page(mapping, index);
6957 if (!pages[numpages]) {
6965 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6970 ocfs2_unlock_and_free_pages(pages, numpages);
6980 * Zero the area past i_size but still within an allocated
6981 * cluster. This avoids exposing nonzero data on subsequent file
6984 * We need to call this before i_size is updated on the inode because
6985 * otherwise block_write_full_page() will skip writeout of pages past
6986 * i_size. The new_i_size parameter is passed for this reason.
6988 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6989 u64 range_start, u64 range_end)
6991 int ret = 0, numpages;
6992 struct page **pages = NULL;
6994 unsigned int ext_flags;
6995 struct super_block *sb = inode->i_sb;
6998 * File systems which don't support sparse files zero on every
7001 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
7004 pages = kcalloc(ocfs2_pages_per_cluster(sb),
7005 sizeof(struct page *), GFP_NOFS);
7006 if (pages == NULL) {
7012 if (range_start == range_end)
7015 ret = ocfs2_extent_map_get_blocks(inode,
7016 range_start >> sb->s_blocksize_bits,
7017 &phys, NULL, &ext_flags);
7024 * Tail is a hole, or is marked unwritten. In either case, we
7025 * can count on read and write to return/push zero's.
7027 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
7030 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
7037 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
7038 numpages, phys, handle);
7041 * Initiate writeout of the pages we zero'd here. We don't
7042 * wait on them - the truncate_inode_pages() call later will
7045 ret = do_sync_mapping_range(inode->i_mapping, range_start,
7046 range_end - 1, SYNC_FILE_RANGE_WRITE);
7057 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7058 struct ocfs2_dinode *di)
7060 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7061 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7063 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7064 memset(&di->id2, 0, blocksize -
7065 offsetof(struct ocfs2_dinode, id2) -
7068 memset(&di->id2, 0, blocksize -
7069 offsetof(struct ocfs2_dinode, id2));
7072 void ocfs2_dinode_new_extent_list(struct inode *inode,
7073 struct ocfs2_dinode *di)
7075 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7076 di->id2.i_list.l_tree_depth = 0;
7077 di->id2.i_list.l_next_free_rec = 0;
7078 di->id2.i_list.l_count = cpu_to_le16(
7079 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7082 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7084 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7085 struct ocfs2_inline_data *idata = &di->id2.i_data;
7087 spin_lock(&oi->ip_lock);
7088 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7089 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7090 spin_unlock(&oi->ip_lock);
7093 * We clear the entire i_data structure here so that all
7094 * fields can be properly initialized.
7096 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7098 idata->id_count = cpu_to_le16(
7099 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7102 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7103 struct buffer_head *di_bh)
7105 int ret, i, has_data, num_pages = 0;
7107 u64 uninitialized_var(block);
7108 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7109 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7110 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7111 struct ocfs2_alloc_context *data_ac = NULL;
7112 struct page **pages = NULL;
7113 loff_t end = osb->s_clustersize;
7114 struct ocfs2_extent_tree et;
7117 has_data = i_size_read(inode) ? 1 : 0;
7120 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
7121 sizeof(struct page *), GFP_NOFS);
7122 if (pages == NULL) {
7128 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7135 handle = ocfs2_start_trans(osb,
7136 ocfs2_inline_to_extents_credits(osb->sb));
7137 if (IS_ERR(handle)) {
7138 ret = PTR_ERR(handle);
7143 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7144 OCFS2_JOURNAL_ACCESS_WRITE);
7152 unsigned int page_end;
7155 if (vfs_dq_alloc_space_nodirty(inode,
7156 ocfs2_clusters_to_bytes(osb->sb, 1))) {
7162 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
7170 * Save two copies, one for insert, and one that can
7171 * be changed by ocfs2_map_and_dirty_page() below.
7173 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7176 * Non sparse file systems zero on extend, so no need
7179 if (!ocfs2_sparse_alloc(osb) &&
7180 PAGE_CACHE_SIZE < osb->s_clustersize)
7181 end = PAGE_CACHE_SIZE;
7183 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
7190 * This should populate the 1st page for us and mark
7193 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
7199 page_end = PAGE_CACHE_SIZE;
7200 if (PAGE_CACHE_SIZE > osb->s_clustersize)
7201 page_end = osb->s_clustersize;
7203 for (i = 0; i < num_pages; i++)
7204 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
7205 pages[i], i > 0, &phys);
7208 spin_lock(&oi->ip_lock);
7209 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7210 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7211 spin_unlock(&oi->ip_lock);
7213 ocfs2_dinode_new_extent_list(inode, di);
7215 ocfs2_journal_dirty(handle, di_bh);
7219 * An error at this point should be extremely rare. If
7220 * this proves to be false, we could always re-build
7221 * the in-inode data from our pages.
7223 ocfs2_init_dinode_extent_tree(&et, inode, di_bh);
7224 ret = ocfs2_insert_extent(osb, handle, inode, &et,
7225 0, block, 1, 0, NULL);
7231 inode->i_blocks = ocfs2_inode_sector_count(inode);
7235 if (ret < 0 && did_quota)
7236 vfs_dq_free_space_nodirty(inode,
7237 ocfs2_clusters_to_bytes(osb->sb, 1));
7239 ocfs2_commit_trans(osb, handle);
7243 ocfs2_free_alloc_context(data_ac);
7247 ocfs2_unlock_and_free_pages(pages, num_pages);
7255 * It is expected, that by the time you call this function,
7256 * inode->i_size and fe->i_size have been adjusted.
7258 * WARNING: This will kfree the truncate context
7260 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7261 struct inode *inode,
7262 struct buffer_head *fe_bh,
7263 struct ocfs2_truncate_context *tc)
7265 int status, i, credits, tl_sem = 0;
7266 u32 clusters_to_del, new_highest_cpos, range;
7267 struct ocfs2_extent_list *el;
7268 handle_t *handle = NULL;
7269 struct inode *tl_inode = osb->osb_tl_inode;
7270 struct ocfs2_path *path = NULL;
7271 struct ocfs2_dinode *di = (struct ocfs2_dinode *)fe_bh->b_data;
7275 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7276 i_size_read(inode));
7278 path = ocfs2_new_path(fe_bh, &di->id2.i_list,
7279 ocfs2_journal_access_di);
7286 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7290 * Check that we still have allocation to delete.
7292 if (OCFS2_I(inode)->ip_clusters == 0) {
7298 * Truncate always works against the rightmost tree branch.
7300 status = ocfs2_find_path(inode, path, UINT_MAX);
7306 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
7307 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
7310 * By now, el will point to the extent list on the bottom most
7311 * portion of this tree. Only the tail record is considered in
7314 * We handle the following cases, in order:
7315 * - empty extent: delete the remaining branch
7316 * - remove the entire record
7317 * - remove a partial record
7318 * - no record needs to be removed (truncate has completed)
7320 el = path_leaf_el(path);
7321 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7322 ocfs2_error(inode->i_sb,
7323 "Inode %llu has empty extent block at %llu\n",
7324 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7325 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7330 i = le16_to_cpu(el->l_next_free_rec) - 1;
7331 range = le32_to_cpu(el->l_recs[i].e_cpos) +
7332 ocfs2_rec_clusters(el, &el->l_recs[i]);
7333 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
7334 clusters_to_del = 0;
7335 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
7336 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
7337 } else if (range > new_highest_cpos) {
7338 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
7339 le32_to_cpu(el->l_recs[i].e_cpos)) -
7346 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
7347 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
7349 mutex_lock(&tl_inode->i_mutex);
7351 /* ocfs2_truncate_log_needs_flush guarantees us at least one
7352 * record is free for use. If there isn't any, we flush to get
7353 * an empty truncate log. */
7354 if (ocfs2_truncate_log_needs_flush(osb)) {
7355 status = __ocfs2_flush_truncate_log(osb);
7362 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
7363 (struct ocfs2_dinode *)fe_bh->b_data,
7365 handle = ocfs2_start_trans(osb, credits);
7366 if (IS_ERR(handle)) {
7367 status = PTR_ERR(handle);
7373 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
7380 mutex_unlock(&tl_inode->i_mutex);
7383 ocfs2_commit_trans(osb, handle);
7386 ocfs2_reinit_path(path, 1);
7389 * The check above will catch the case where we've truncated
7390 * away all allocation.
7396 ocfs2_schedule_truncate_log_flush(osb, 1);
7399 mutex_unlock(&tl_inode->i_mutex);
7402 ocfs2_commit_trans(osb, handle);
7404 ocfs2_run_deallocs(osb, &tc->tc_dealloc);
7406 ocfs2_free_path(path);
7408 /* This will drop the ext_alloc cluster lock for us */
7409 ocfs2_free_truncate_context(tc);
7416 * Expects the inode to already be locked.
7418 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
7419 struct inode *inode,
7420 struct buffer_head *fe_bh,
7421 struct ocfs2_truncate_context **tc)
7424 unsigned int new_i_clusters;
7425 struct ocfs2_dinode *fe;
7426 struct ocfs2_extent_block *eb;
7427 struct buffer_head *last_eb_bh = NULL;
7433 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
7434 i_size_read(inode));
7435 fe = (struct ocfs2_dinode *) fe_bh->b_data;
7437 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
7438 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
7439 (unsigned long long)le64_to_cpu(fe->i_size));
7441 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
7447 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
7449 if (fe->id2.i_list.l_tree_depth) {
7450 status = ocfs2_read_extent_block(inode,
7451 le64_to_cpu(fe->i_last_eb_blk),
7457 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
7460 (*tc)->tc_last_eb_bh = last_eb_bh;
7466 ocfs2_free_truncate_context(*tc);
7474 * 'start' is inclusive, 'end' is not.
7476 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7477 unsigned int start, unsigned int end, int trunc)
7480 unsigned int numbytes;
7482 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7483 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7484 struct ocfs2_inline_data *idata = &di->id2.i_data;
7486 if (end > i_size_read(inode))
7487 end = i_size_read(inode);
7489 BUG_ON(start >= end);
7491 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7492 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7493 !ocfs2_supports_inline_data(osb)) {
7494 ocfs2_error(inode->i_sb,
7495 "Inline data flags for inode %llu don't agree! "
7496 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7497 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7498 le16_to_cpu(di->i_dyn_features),
7499 OCFS2_I(inode)->ip_dyn_features,
7500 osb->s_feature_incompat);
7505 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7506 if (IS_ERR(handle)) {
7507 ret = PTR_ERR(handle);
7512 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7513 OCFS2_JOURNAL_ACCESS_WRITE);
7519 numbytes = end - start;
7520 memset(idata->id_data + start, 0, numbytes);
7523 * No need to worry about the data page here - it's been
7524 * truncated already and inline data doesn't need it for
7525 * pushing zero's to disk, so we'll let readpage pick it up
7529 i_size_write(inode, start);
7530 di->i_size = cpu_to_le64(start);
7533 inode->i_blocks = ocfs2_inode_sector_count(inode);
7534 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7536 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7537 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7539 ocfs2_journal_dirty(handle, di_bh);
7542 ocfs2_commit_trans(osb, handle);
7548 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
7551 * The caller is responsible for completing deallocation
7552 * before freeing the context.
7554 if (tc->tc_dealloc.c_first_suballocator != NULL)
7556 "Truncate completion has non-empty dealloc context\n");
7558 brelse(tc->tc_last_eb_bh);