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>
32 #include <linux/blkdev.h>
33 #include <linux/sched/signal.h>
35 #include <cluster/masklog.h>
41 #include "blockcheck.h"
43 #include "extent_map.h"
46 #include "localalloc.h"
53 #include "refcounttree.h"
54 #include "ocfs2_trace.h"
56 #include "buffer_head_io.h"
58 enum ocfs2_contig_type {
65 static enum ocfs2_contig_type
66 ocfs2_extent_rec_contig(struct super_block *sb,
67 struct ocfs2_extent_rec *ext,
68 struct ocfs2_extent_rec *insert_rec);
70 * Operations for a specific extent tree type.
72 * To implement an on-disk btree (extent tree) type in ocfs2, add
73 * an ocfs2_extent_tree_operations structure and the matching
74 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
75 * for the allocation portion of the extent tree.
77 struct ocfs2_extent_tree_operations {
79 * last_eb_blk is the block number of the right most leaf extent
80 * block. Most on-disk structures containing an extent tree store
81 * this value for fast access. The ->eo_set_last_eb_blk() and
82 * ->eo_get_last_eb_blk() operations access this value. They are
85 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
87 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
90 * The on-disk structure usually keeps track of how many total
91 * clusters are stored in this extent tree. This function updates
92 * that value. new_clusters is the delta, and must be
93 * added to the total. Required.
95 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
99 * If this extent tree is supported by an extent map, insert
100 * a record into the map.
102 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
103 struct ocfs2_extent_rec *rec);
106 * If this extent tree is supported by an extent map, truncate the
109 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
113 * If ->eo_insert_check() exists, it is called before rec is
114 * inserted into the extent tree. It is optional.
116 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
117 struct ocfs2_extent_rec *rec);
118 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
121 * --------------------------------------------------------------
122 * The remaining are internal to ocfs2_extent_tree and don't have
127 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
130 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
133 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
134 * it exists. If it does not, et->et_max_leaf_clusters is set
135 * to 0 (unlimited). Optional.
137 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
140 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
141 * are contiguous or not. Optional. Don't need to set it if use
142 * ocfs2_extent_rec as the tree leaf.
144 enum ocfs2_contig_type
145 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
146 struct ocfs2_extent_rec *ext,
147 struct ocfs2_extent_rec *insert_rec);
152 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
155 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
156 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
158 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
160 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
161 struct ocfs2_extent_rec *rec);
162 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
164 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
165 struct ocfs2_extent_rec *rec);
166 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
167 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
168 static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
169 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
170 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
171 .eo_update_clusters = ocfs2_dinode_update_clusters,
172 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
173 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
174 .eo_insert_check = ocfs2_dinode_insert_check,
175 .eo_sanity_check = ocfs2_dinode_sanity_check,
176 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
179 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
182 struct ocfs2_dinode *di = et->et_object;
184 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
185 di->i_last_eb_blk = cpu_to_le64(blkno);
188 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
190 struct ocfs2_dinode *di = et->et_object;
192 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
193 return le64_to_cpu(di->i_last_eb_blk);
196 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
199 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
200 struct ocfs2_dinode *di = et->et_object;
202 le32_add_cpu(&di->i_clusters, clusters);
203 spin_lock(&oi->ip_lock);
204 oi->ip_clusters = le32_to_cpu(di->i_clusters);
205 spin_unlock(&oi->ip_lock);
208 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
209 struct ocfs2_extent_rec *rec)
211 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
213 ocfs2_extent_map_insert_rec(inode, rec);
216 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
219 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
221 ocfs2_extent_map_trunc(inode, clusters);
224 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
225 struct ocfs2_extent_rec *rec)
227 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
228 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
230 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
231 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
232 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
233 "Device %s, asking for sparse allocation: inode %llu, "
234 "cpos %u, clusters %u\n",
236 (unsigned long long)oi->ip_blkno,
237 rec->e_cpos, oi->ip_clusters);
242 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
244 struct ocfs2_dinode *di = et->et_object;
246 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
247 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
252 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
254 struct ocfs2_dinode *di = et->et_object;
256 et->et_root_el = &di->id2.i_list;
260 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
262 struct ocfs2_xattr_value_buf *vb = et->et_object;
264 et->et_root_el = &vb->vb_xv->xr_list;
267 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
270 struct ocfs2_xattr_value_buf *vb = et->et_object;
272 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
275 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
277 struct ocfs2_xattr_value_buf *vb = et->et_object;
279 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
282 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
285 struct ocfs2_xattr_value_buf *vb = et->et_object;
287 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
290 static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
291 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
292 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
293 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
294 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
297 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
299 struct ocfs2_xattr_block *xb = et->et_object;
301 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
304 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
306 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
307 et->et_max_leaf_clusters =
308 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
311 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
314 struct ocfs2_xattr_block *xb = et->et_object;
315 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
317 xt->xt_last_eb_blk = cpu_to_le64(blkno);
320 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
322 struct ocfs2_xattr_block *xb = et->et_object;
323 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
325 return le64_to_cpu(xt->xt_last_eb_blk);
328 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
331 struct ocfs2_xattr_block *xb = et->et_object;
333 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
336 static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
337 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
338 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
339 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
340 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
341 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
344 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
347 struct ocfs2_dx_root_block *dx_root = et->et_object;
349 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
352 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
354 struct ocfs2_dx_root_block *dx_root = et->et_object;
356 return le64_to_cpu(dx_root->dr_last_eb_blk);
359 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
362 struct ocfs2_dx_root_block *dx_root = et->et_object;
364 le32_add_cpu(&dx_root->dr_clusters, clusters);
367 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
369 struct ocfs2_dx_root_block *dx_root = et->et_object;
371 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
376 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
378 struct ocfs2_dx_root_block *dx_root = et->et_object;
380 et->et_root_el = &dx_root->dr_list;
383 static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
384 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
385 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
386 .eo_update_clusters = ocfs2_dx_root_update_clusters,
387 .eo_sanity_check = ocfs2_dx_root_sanity_check,
388 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
391 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
393 struct ocfs2_refcount_block *rb = et->et_object;
395 et->et_root_el = &rb->rf_list;
398 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
401 struct ocfs2_refcount_block *rb = et->et_object;
403 rb->rf_last_eb_blk = cpu_to_le64(blkno);
406 static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
408 struct ocfs2_refcount_block *rb = et->et_object;
410 return le64_to_cpu(rb->rf_last_eb_blk);
413 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
416 struct ocfs2_refcount_block *rb = et->et_object;
418 le32_add_cpu(&rb->rf_clusters, clusters);
421 static enum ocfs2_contig_type
422 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
423 struct ocfs2_extent_rec *ext,
424 struct ocfs2_extent_rec *insert_rec)
429 static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
430 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
431 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
432 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
433 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
434 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
437 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
438 struct ocfs2_caching_info *ci,
439 struct buffer_head *bh,
440 ocfs2_journal_access_func access,
442 const struct ocfs2_extent_tree_operations *ops)
447 et->et_root_journal_access = access;
449 obj = (void *)bh->b_data;
452 et->et_ops->eo_fill_root_el(et);
453 if (!et->et_ops->eo_fill_max_leaf_clusters)
454 et->et_max_leaf_clusters = 0;
456 et->et_ops->eo_fill_max_leaf_clusters(et);
459 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
460 struct ocfs2_caching_info *ci,
461 struct buffer_head *bh)
463 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
464 NULL, &ocfs2_dinode_et_ops);
467 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
468 struct ocfs2_caching_info *ci,
469 struct buffer_head *bh)
471 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
472 NULL, &ocfs2_xattr_tree_et_ops);
475 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
476 struct ocfs2_caching_info *ci,
477 struct ocfs2_xattr_value_buf *vb)
479 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
480 &ocfs2_xattr_value_et_ops);
483 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
484 struct ocfs2_caching_info *ci,
485 struct buffer_head *bh)
487 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
488 NULL, &ocfs2_dx_root_et_ops);
491 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
492 struct ocfs2_caching_info *ci,
493 struct buffer_head *bh)
495 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
496 NULL, &ocfs2_refcount_tree_et_ops);
499 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
502 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
505 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
507 return et->et_ops->eo_get_last_eb_blk(et);
510 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
513 et->et_ops->eo_update_clusters(et, clusters);
516 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
517 struct ocfs2_extent_rec *rec)
519 if (et->et_ops->eo_extent_map_insert)
520 et->et_ops->eo_extent_map_insert(et, rec);
523 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
526 if (et->et_ops->eo_extent_map_truncate)
527 et->et_ops->eo_extent_map_truncate(et, clusters);
530 static inline int ocfs2_et_root_journal_access(handle_t *handle,
531 struct ocfs2_extent_tree *et,
534 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
538 static inline enum ocfs2_contig_type
539 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
540 struct ocfs2_extent_rec *rec,
541 struct ocfs2_extent_rec *insert_rec)
543 if (et->et_ops->eo_extent_contig)
544 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
546 return ocfs2_extent_rec_contig(
547 ocfs2_metadata_cache_get_super(et->et_ci),
551 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
552 struct ocfs2_extent_rec *rec)
556 if (et->et_ops->eo_insert_check)
557 ret = et->et_ops->eo_insert_check(et, rec);
561 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
565 if (et->et_ops->eo_sanity_check)
566 ret = et->et_ops->eo_sanity_check(et);
570 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
571 struct ocfs2_extent_block *eb);
572 static void ocfs2_adjust_rightmost_records(handle_t *handle,
573 struct ocfs2_extent_tree *et,
574 struct ocfs2_path *path,
575 struct ocfs2_extent_rec *insert_rec);
577 * Reset the actual path elements so that we can re-use the structure
578 * to build another path. Generally, this involves freeing the buffer
581 void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
583 int i, start = 0, depth = 0;
584 struct ocfs2_path_item *node;
589 for(i = start; i < path_num_items(path); i++) {
590 node = &path->p_node[i];
598 * Tree depth may change during truncate, or insert. If we're
599 * keeping the root extent list, then make sure that our path
600 * structure reflects the proper depth.
603 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
605 path_root_access(path) = NULL;
607 path->p_tree_depth = depth;
610 void ocfs2_free_path(struct ocfs2_path *path)
613 ocfs2_reinit_path(path, 0);
619 * All the elements of src into dest. After this call, src could be freed
620 * without affecting dest.
622 * Both paths should have the same root. Any non-root elements of dest
625 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
629 BUG_ON(path_root_bh(dest) != path_root_bh(src));
630 BUG_ON(path_root_el(dest) != path_root_el(src));
631 BUG_ON(path_root_access(dest) != path_root_access(src));
633 ocfs2_reinit_path(dest, 1);
635 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
636 dest->p_node[i].bh = src->p_node[i].bh;
637 dest->p_node[i].el = src->p_node[i].el;
639 if (dest->p_node[i].bh)
640 get_bh(dest->p_node[i].bh);
645 * Make the *dest path the same as src and re-initialize src path to
648 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
652 BUG_ON(path_root_bh(dest) != path_root_bh(src));
653 BUG_ON(path_root_access(dest) != path_root_access(src));
655 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
656 brelse(dest->p_node[i].bh);
658 dest->p_node[i].bh = src->p_node[i].bh;
659 dest->p_node[i].el = src->p_node[i].el;
661 src->p_node[i].bh = NULL;
662 src->p_node[i].el = NULL;
667 * Insert an extent block at given index.
669 * This will not take an additional reference on eb_bh.
671 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
672 struct buffer_head *eb_bh)
674 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
677 * Right now, no root bh is an extent block, so this helps
678 * catch code errors with dinode trees. The assertion can be
679 * safely removed if we ever need to insert extent block
680 * structures at the root.
684 path->p_node[index].bh = eb_bh;
685 path->p_node[index].el = &eb->h_list;
688 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
689 struct ocfs2_extent_list *root_el,
690 ocfs2_journal_access_func access)
692 struct ocfs2_path *path;
694 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
696 path = kzalloc(sizeof(*path), GFP_NOFS);
698 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
700 path_root_bh(path) = root_bh;
701 path_root_el(path) = root_el;
702 path_root_access(path) = access;
708 struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
710 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
711 path_root_access(path));
714 struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
716 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
717 et->et_root_journal_access);
721 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
722 * otherwise it's the root_access function.
724 * I don't like the way this function's name looks next to
725 * ocfs2_journal_access_path(), but I don't have a better one.
727 int ocfs2_path_bh_journal_access(handle_t *handle,
728 struct ocfs2_caching_info *ci,
729 struct ocfs2_path *path,
732 ocfs2_journal_access_func access = path_root_access(path);
735 access = ocfs2_journal_access;
738 access = ocfs2_journal_access_eb;
740 return access(handle, ci, path->p_node[idx].bh,
741 OCFS2_JOURNAL_ACCESS_WRITE);
745 * Convenience function to journal all components in a path.
747 int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
749 struct ocfs2_path *path)
756 for(i = 0; i < path_num_items(path); i++) {
757 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
769 * Return the index of the extent record which contains cluster #v_cluster.
770 * -1 is returned if it was not found.
772 * Should work fine on interior and exterior nodes.
774 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
778 struct ocfs2_extent_rec *rec;
779 u32 rec_end, rec_start, clusters;
781 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
782 rec = &el->l_recs[i];
784 rec_start = le32_to_cpu(rec->e_cpos);
785 clusters = ocfs2_rec_clusters(el, rec);
787 rec_end = rec_start + clusters;
789 if (v_cluster >= rec_start && v_cluster < rec_end) {
799 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
800 * ocfs2_extent_rec_contig only work properly against leaf nodes!
802 static int ocfs2_block_extent_contig(struct super_block *sb,
803 struct ocfs2_extent_rec *ext,
806 u64 blk_end = le64_to_cpu(ext->e_blkno);
808 blk_end += ocfs2_clusters_to_blocks(sb,
809 le16_to_cpu(ext->e_leaf_clusters));
811 return blkno == blk_end;
814 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
815 struct ocfs2_extent_rec *right)
819 left_range = le32_to_cpu(left->e_cpos) +
820 le16_to_cpu(left->e_leaf_clusters);
822 return (left_range == le32_to_cpu(right->e_cpos));
825 static enum ocfs2_contig_type
826 ocfs2_extent_rec_contig(struct super_block *sb,
827 struct ocfs2_extent_rec *ext,
828 struct ocfs2_extent_rec *insert_rec)
830 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
833 * Refuse to coalesce extent records with different flag
834 * fields - we don't want to mix unwritten extents with user
837 if (ext->e_flags != insert_rec->e_flags)
840 if (ocfs2_extents_adjacent(ext, insert_rec) &&
841 ocfs2_block_extent_contig(sb, ext, blkno))
844 blkno = le64_to_cpu(ext->e_blkno);
845 if (ocfs2_extents_adjacent(insert_rec, ext) &&
846 ocfs2_block_extent_contig(sb, insert_rec, blkno))
853 * NOTE: We can have pretty much any combination of contiguousness and
856 * The usefulness of APPEND_TAIL is more in that it lets us know that
857 * we'll have to update the path to that leaf.
859 enum ocfs2_append_type {
864 enum ocfs2_split_type {
870 struct ocfs2_insert_type {
871 enum ocfs2_split_type ins_split;
872 enum ocfs2_append_type ins_appending;
873 enum ocfs2_contig_type ins_contig;
874 int ins_contig_index;
878 struct ocfs2_merge_ctxt {
879 enum ocfs2_contig_type c_contig_type;
880 int c_has_empty_extent;
881 int c_split_covers_rec;
884 static int ocfs2_validate_extent_block(struct super_block *sb,
885 struct buffer_head *bh)
888 struct ocfs2_extent_block *eb =
889 (struct ocfs2_extent_block *)bh->b_data;
891 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
893 BUG_ON(!buffer_uptodate(bh));
896 * If the ecc fails, we return the error but otherwise
897 * leave the filesystem running. We know any error is
898 * local to this block.
900 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
902 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
903 (unsigned long long)bh->b_blocknr);
908 * Errors after here are fatal.
911 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
913 "Extent block #%llu has bad signature %.*s\n",
914 (unsigned long long)bh->b_blocknr, 7,
919 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
921 "Extent block #%llu has an invalid h_blkno of %llu\n",
922 (unsigned long long)bh->b_blocknr,
923 (unsigned long long)le64_to_cpu(eb->h_blkno));
927 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
929 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
930 (unsigned long long)bh->b_blocknr,
931 le32_to_cpu(eb->h_fs_generation));
938 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
939 struct buffer_head **bh)
942 struct buffer_head *tmp = *bh;
944 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
945 ocfs2_validate_extent_block);
947 /* If ocfs2_read_block() got us a new bh, pass it up. */
956 * How many free extents have we got before we need more meta data?
958 int ocfs2_num_free_extents(struct ocfs2_super *osb,
959 struct ocfs2_extent_tree *et)
962 struct ocfs2_extent_list *el = NULL;
963 struct ocfs2_extent_block *eb;
964 struct buffer_head *eb_bh = NULL;
968 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
971 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
977 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
981 BUG_ON(el->l_tree_depth != 0);
983 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
987 trace_ocfs2_num_free_extents(retval);
991 /* expects array to already be allocated
993 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
996 static int ocfs2_create_new_meta_bhs(handle_t *handle,
997 struct ocfs2_extent_tree *et,
999 struct ocfs2_alloc_context *meta_ac,
1000 struct buffer_head *bhs[])
1002 int count, status, i;
1003 u16 suballoc_bit_start;
1005 u64 suballoc_loc, first_blkno;
1006 struct ocfs2_super *osb =
1007 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1008 struct ocfs2_extent_block *eb;
1011 while (count < wanted) {
1012 status = ocfs2_claim_metadata(handle,
1016 &suballoc_bit_start,
1024 for(i = count; i < (num_got + count); i++) {
1025 bhs[i] = sb_getblk(osb->sb, first_blkno);
1026 if (bhs[i] == NULL) {
1031 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1033 status = ocfs2_journal_access_eb(handle, et->et_ci,
1035 OCFS2_JOURNAL_ACCESS_CREATE);
1041 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1042 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1043 /* Ok, setup the minimal stuff here. */
1044 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1045 eb->h_blkno = cpu_to_le64(first_blkno);
1046 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1047 eb->h_suballoc_slot =
1048 cpu_to_le16(meta_ac->ac_alloc_slot);
1049 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1050 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1051 eb->h_list.l_count =
1052 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1054 suballoc_bit_start++;
1057 /* We'll also be dirtied by the caller, so
1058 * this isn't absolutely necessary. */
1059 ocfs2_journal_dirty(handle, bhs[i]);
1068 for(i = 0; i < wanted; i++) {
1078 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1080 * Returns the sum of the rightmost extent rec logical offset and
1083 * ocfs2_add_branch() uses this to determine what logical cluster
1084 * value should be populated into the leftmost new branch records.
1086 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1087 * value for the new topmost tree record.
1089 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1093 i = le16_to_cpu(el->l_next_free_rec) - 1;
1095 return le32_to_cpu(el->l_recs[i].e_cpos) +
1096 ocfs2_rec_clusters(el, &el->l_recs[i]);
1100 * Change range of the branches in the right most path according to the leaf
1101 * extent block's rightmost record.
1103 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1104 struct ocfs2_extent_tree *et)
1107 struct ocfs2_path *path = NULL;
1108 struct ocfs2_extent_list *el;
1109 struct ocfs2_extent_rec *rec;
1111 path = ocfs2_new_path_from_et(et);
1117 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1123 status = ocfs2_extend_trans(handle, path_num_items(path));
1129 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1135 el = path_leaf_el(path);
1136 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1138 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1141 ocfs2_free_path(path);
1146 * Add an entire tree branch to our inode. eb_bh is the extent block
1147 * to start at, if we don't want to start the branch at the root
1150 * last_eb_bh is required as we have to update it's next_leaf pointer
1151 * for the new last extent block.
1153 * the new branch will be 'empty' in the sense that every block will
1154 * contain a single record with cluster count == 0.
1156 static int ocfs2_add_branch(handle_t *handle,
1157 struct ocfs2_extent_tree *et,
1158 struct buffer_head *eb_bh,
1159 struct buffer_head **last_eb_bh,
1160 struct ocfs2_alloc_context *meta_ac)
1162 int status, new_blocks, i;
1163 u64 next_blkno, new_last_eb_blk;
1164 struct buffer_head *bh;
1165 struct buffer_head **new_eb_bhs = NULL;
1166 struct ocfs2_extent_block *eb;
1167 struct ocfs2_extent_list *eb_el;
1168 struct ocfs2_extent_list *el;
1169 u32 new_cpos, root_end;
1171 BUG_ON(!last_eb_bh || !*last_eb_bh);
1174 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1177 el = et->et_root_el;
1179 /* we never add a branch to a leaf. */
1180 BUG_ON(!el->l_tree_depth);
1182 new_blocks = le16_to_cpu(el->l_tree_depth);
1184 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1185 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1186 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1189 * If there is a gap before the root end and the real end
1190 * of the righmost leaf block, we need to remove the gap
1191 * between new_cpos and root_end first so that the tree
1192 * is consistent after we add a new branch(it will start
1195 if (root_end > new_cpos) {
1196 trace_ocfs2_adjust_rightmost_branch(
1197 (unsigned long long)
1198 ocfs2_metadata_cache_owner(et->et_ci),
1199 root_end, new_cpos);
1201 status = ocfs2_adjust_rightmost_branch(handle, et);
1208 /* allocate the number of new eb blocks we need */
1209 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1217 status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1218 meta_ac, new_eb_bhs);
1224 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1225 * linked with the rest of the tree.
1226 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1228 * when we leave the loop, new_last_eb_blk will point to the
1229 * newest leaf, and next_blkno will point to the topmost extent
1231 next_blkno = new_last_eb_blk = 0;
1232 for(i = 0; i < new_blocks; i++) {
1234 eb = (struct ocfs2_extent_block *) bh->b_data;
1235 /* ocfs2_create_new_meta_bhs() should create it right! */
1236 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1237 eb_el = &eb->h_list;
1239 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1240 OCFS2_JOURNAL_ACCESS_CREATE);
1246 eb->h_next_leaf_blk = 0;
1247 eb_el->l_tree_depth = cpu_to_le16(i);
1248 eb_el->l_next_free_rec = cpu_to_le16(1);
1250 * This actually counts as an empty extent as
1253 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1254 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1256 * eb_el isn't always an interior node, but even leaf
1257 * nodes want a zero'd flags and reserved field so
1258 * this gets the whole 32 bits regardless of use.
1260 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1261 if (!eb_el->l_tree_depth)
1262 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1264 ocfs2_journal_dirty(handle, bh);
1265 next_blkno = le64_to_cpu(eb->h_blkno);
1268 /* This is a bit hairy. We want to update up to three blocks
1269 * here without leaving any of them in an inconsistent state
1270 * in case of error. We don't have to worry about
1271 * journal_dirty erroring as it won't unless we've aborted the
1272 * handle (in which case we would never be here) so reserving
1273 * the write with journal_access is all we need to do. */
1274 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1275 OCFS2_JOURNAL_ACCESS_WRITE);
1280 status = ocfs2_et_root_journal_access(handle, et,
1281 OCFS2_JOURNAL_ACCESS_WRITE);
1287 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1288 OCFS2_JOURNAL_ACCESS_WRITE);
1295 /* Link the new branch into the rest of the tree (el will
1296 * either be on the root_bh, or the extent block passed in. */
1297 i = le16_to_cpu(el->l_next_free_rec);
1298 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1299 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1300 el->l_recs[i].e_int_clusters = 0;
1301 le16_add_cpu(&el->l_next_free_rec, 1);
1303 /* fe needs a new last extent block pointer, as does the
1304 * next_leaf on the previously last-extent-block. */
1305 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1307 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1308 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1310 ocfs2_journal_dirty(handle, *last_eb_bh);
1311 ocfs2_journal_dirty(handle, et->et_root_bh);
1313 ocfs2_journal_dirty(handle, eb_bh);
1316 * Some callers want to track the rightmost leaf so pass it
1319 brelse(*last_eb_bh);
1320 get_bh(new_eb_bhs[0]);
1321 *last_eb_bh = new_eb_bhs[0];
1326 for (i = 0; i < new_blocks; i++)
1327 brelse(new_eb_bhs[i]);
1335 * adds another level to the allocation tree.
1336 * returns back the new extent block so you can add a branch to it
1339 static int ocfs2_shift_tree_depth(handle_t *handle,
1340 struct ocfs2_extent_tree *et,
1341 struct ocfs2_alloc_context *meta_ac,
1342 struct buffer_head **ret_new_eb_bh)
1346 struct buffer_head *new_eb_bh = NULL;
1347 struct ocfs2_extent_block *eb;
1348 struct ocfs2_extent_list *root_el;
1349 struct ocfs2_extent_list *eb_el;
1351 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1358 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1359 /* ocfs2_create_new_meta_bhs() should create it right! */
1360 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1362 eb_el = &eb->h_list;
1363 root_el = et->et_root_el;
1365 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1366 OCFS2_JOURNAL_ACCESS_CREATE);
1372 /* copy the root extent list data into the new extent block */
1373 eb_el->l_tree_depth = root_el->l_tree_depth;
1374 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1375 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1376 eb_el->l_recs[i] = root_el->l_recs[i];
1378 ocfs2_journal_dirty(handle, new_eb_bh);
1380 status = ocfs2_et_root_journal_access(handle, et,
1381 OCFS2_JOURNAL_ACCESS_WRITE);
1387 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1389 /* update root_bh now */
1390 le16_add_cpu(&root_el->l_tree_depth, 1);
1391 root_el->l_recs[0].e_cpos = 0;
1392 root_el->l_recs[0].e_blkno = eb->h_blkno;
1393 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1394 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1395 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1396 root_el->l_next_free_rec = cpu_to_le16(1);
1398 /* If this is our 1st tree depth shift, then last_eb_blk
1399 * becomes the allocated extent block */
1400 if (root_el->l_tree_depth == cpu_to_le16(1))
1401 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1403 ocfs2_journal_dirty(handle, et->et_root_bh);
1405 *ret_new_eb_bh = new_eb_bh;
1415 * Should only be called when there is no space left in any of the
1416 * leaf nodes. What we want to do is find the lowest tree depth
1417 * non-leaf extent block with room for new records. There are three
1418 * valid results of this search:
1420 * 1) a lowest extent block is found, then we pass it back in
1421 * *lowest_eb_bh and return '0'
1423 * 2) the search fails to find anything, but the root_el has room. We
1424 * pass NULL back in *lowest_eb_bh, but still return '0'
1426 * 3) the search fails to find anything AND the root_el is full, in
1427 * which case we return > 0
1429 * return status < 0 indicates an error.
1431 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1432 struct buffer_head **target_bh)
1436 struct ocfs2_extent_block *eb;
1437 struct ocfs2_extent_list *el;
1438 struct buffer_head *bh = NULL;
1439 struct buffer_head *lowest_bh = NULL;
1443 el = et->et_root_el;
1445 while(le16_to_cpu(el->l_tree_depth) > 1) {
1446 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1447 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1448 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1449 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1453 i = le16_to_cpu(el->l_next_free_rec) - 1;
1454 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1456 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1457 "Owner %llu has extent list where extent # %d has no physical block start\n",
1458 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1466 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1472 eb = (struct ocfs2_extent_block *) bh->b_data;
1475 if (le16_to_cpu(el->l_next_free_rec) <
1476 le16_to_cpu(el->l_count)) {
1483 /* If we didn't find one and the fe doesn't have any room,
1484 * then return '1' */
1485 el = et->et_root_el;
1486 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1489 *target_bh = lowest_bh;
1497 * Grow a b-tree so that it has more records.
1499 * We might shift the tree depth in which case existing paths should
1500 * be considered invalid.
1502 * Tree depth after the grow is returned via *final_depth.
1504 * *last_eb_bh will be updated by ocfs2_add_branch().
1506 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1507 int *final_depth, struct buffer_head **last_eb_bh,
1508 struct ocfs2_alloc_context *meta_ac)
1511 struct ocfs2_extent_list *el = et->et_root_el;
1512 int depth = le16_to_cpu(el->l_tree_depth);
1513 struct buffer_head *bh = NULL;
1515 BUG_ON(meta_ac == NULL);
1517 shift = ocfs2_find_branch_target(et, &bh);
1524 /* We traveled all the way to the bottom of the allocation tree
1525 * and didn't find room for any more extents - we need to add
1526 * another tree level */
1529 trace_ocfs2_grow_tree(
1530 (unsigned long long)
1531 ocfs2_metadata_cache_owner(et->et_ci),
1534 /* ocfs2_shift_tree_depth will return us a buffer with
1535 * the new extent block (so we can pass that to
1536 * ocfs2_add_branch). */
1537 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1545 * Special case: we have room now if we shifted from
1546 * tree_depth 0, so no more work needs to be done.
1548 * We won't be calling add_branch, so pass
1549 * back *last_eb_bh as the new leaf. At depth
1550 * zero, it should always be null so there's
1551 * no reason to brelse.
1553 BUG_ON(*last_eb_bh);
1560 /* call ocfs2_add_branch to add the final part of the tree with
1562 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1571 *final_depth = depth;
1577 * This function will discard the rightmost extent record.
1579 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1581 int next_free = le16_to_cpu(el->l_next_free_rec);
1582 int count = le16_to_cpu(el->l_count);
1583 unsigned int num_bytes;
1586 /* This will cause us to go off the end of our extent list. */
1587 BUG_ON(next_free >= count);
1589 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1591 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1594 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1595 struct ocfs2_extent_rec *insert_rec)
1597 int i, insert_index, next_free, has_empty, num_bytes;
1598 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1599 struct ocfs2_extent_rec *rec;
1601 next_free = le16_to_cpu(el->l_next_free_rec);
1602 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1606 /* The tree code before us didn't allow enough room in the leaf. */
1607 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1610 * The easiest way to approach this is to just remove the
1611 * empty extent and temporarily decrement next_free.
1615 * If next_free was 1 (only an empty extent), this
1616 * loop won't execute, which is fine. We still want
1617 * the decrement above to happen.
1619 for(i = 0; i < (next_free - 1); i++)
1620 el->l_recs[i] = el->l_recs[i+1];
1626 * Figure out what the new record index should be.
1628 for(i = 0; i < next_free; i++) {
1629 rec = &el->l_recs[i];
1631 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1636 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1637 has_empty, next_free,
1638 le16_to_cpu(el->l_count));
1640 BUG_ON(insert_index < 0);
1641 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1642 BUG_ON(insert_index > next_free);
1645 * No need to memmove if we're just adding to the tail.
1647 if (insert_index != next_free) {
1648 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1650 num_bytes = next_free - insert_index;
1651 num_bytes *= sizeof(struct ocfs2_extent_rec);
1652 memmove(&el->l_recs[insert_index + 1],
1653 &el->l_recs[insert_index],
1658 * Either we had an empty extent, and need to re-increment or
1659 * there was no empty extent on a non full rightmost leaf node,
1660 * in which case we still need to increment.
1663 el->l_next_free_rec = cpu_to_le16(next_free);
1665 * Make sure none of the math above just messed up our tree.
1667 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1669 el->l_recs[insert_index] = *insert_rec;
1673 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1675 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1677 BUG_ON(num_recs == 0);
1679 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1681 size = num_recs * sizeof(struct ocfs2_extent_rec);
1682 memmove(&el->l_recs[0], &el->l_recs[1], size);
1683 memset(&el->l_recs[num_recs], 0,
1684 sizeof(struct ocfs2_extent_rec));
1685 el->l_next_free_rec = cpu_to_le16(num_recs);
1690 * Create an empty extent record .
1692 * l_next_free_rec may be updated.
1694 * If an empty extent already exists do nothing.
1696 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1698 int next_free = le16_to_cpu(el->l_next_free_rec);
1700 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1705 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1708 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1709 "Asked to create an empty extent in a full list:\n"
1710 "count = %u, tree depth = %u",
1711 le16_to_cpu(el->l_count),
1712 le16_to_cpu(el->l_tree_depth));
1714 ocfs2_shift_records_right(el);
1717 le16_add_cpu(&el->l_next_free_rec, 1);
1718 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1722 * For a rotation which involves two leaf nodes, the "root node" is
1723 * the lowest level tree node which contains a path to both leafs. This
1724 * resulting set of information can be used to form a complete "subtree"
1726 * This function is passed two full paths from the dinode down to a
1727 * pair of adjacent leaves. It's task is to figure out which path
1728 * index contains the subtree root - this can be the root index itself
1729 * in a worst-case rotation.
1731 * The array index of the subtree root is passed back.
1733 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1734 struct ocfs2_path *left,
1735 struct ocfs2_path *right)
1740 * Check that the caller passed in two paths from the same tree.
1742 BUG_ON(path_root_bh(left) != path_root_bh(right));
1748 * The caller didn't pass two adjacent paths.
1750 mlog_bug_on_msg(i > left->p_tree_depth,
1751 "Owner %llu, left depth %u, right depth %u\n"
1752 "left leaf blk %llu, right leaf blk %llu\n",
1753 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1754 left->p_tree_depth, right->p_tree_depth,
1755 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1756 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1757 } while (left->p_node[i].bh->b_blocknr ==
1758 right->p_node[i].bh->b_blocknr);
1763 typedef void (path_insert_t)(void *, struct buffer_head *);
1766 * Traverse a btree path in search of cpos, starting at root_el.
1768 * This code can be called with a cpos larger than the tree, in which
1769 * case it will return the rightmost path.
1771 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1772 struct ocfs2_extent_list *root_el, u32 cpos,
1773 path_insert_t *func, void *data)
1778 struct buffer_head *bh = NULL;
1779 struct ocfs2_extent_block *eb;
1780 struct ocfs2_extent_list *el;
1781 struct ocfs2_extent_rec *rec;
1784 while (el->l_tree_depth) {
1785 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1786 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1787 "Owner %llu has empty extent list at depth %u\n",
1788 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1789 le16_to_cpu(el->l_tree_depth));
1795 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1796 rec = &el->l_recs[i];
1799 * In the case that cpos is off the allocation
1800 * tree, this should just wind up returning the
1803 range = le32_to_cpu(rec->e_cpos) +
1804 ocfs2_rec_clusters(el, rec);
1805 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1809 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1811 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1812 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1813 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1814 le16_to_cpu(el->l_tree_depth), i);
1821 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1827 eb = (struct ocfs2_extent_block *) bh->b_data;
1830 if (le16_to_cpu(el->l_next_free_rec) >
1831 le16_to_cpu(el->l_count)) {
1832 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1833 "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1834 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1835 (unsigned long long)bh->b_blocknr,
1836 le16_to_cpu(el->l_next_free_rec),
1837 le16_to_cpu(el->l_count));
1848 * Catch any trailing bh that the loop didn't handle.
1856 * Given an initialized path (that is, it has a valid root extent
1857 * list), this function will traverse the btree in search of the path
1858 * which would contain cpos.
1860 * The path traveled is recorded in the path structure.
1862 * Note that this will not do any comparisons on leaf node extent
1863 * records, so it will work fine in the case that we just added a tree
1866 struct find_path_data {
1868 struct ocfs2_path *path;
1870 static void find_path_ins(void *data, struct buffer_head *bh)
1872 struct find_path_data *fp = data;
1875 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1878 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1879 struct ocfs2_path *path, u32 cpos)
1881 struct find_path_data data;
1885 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1886 find_path_ins, &data);
1889 static void find_leaf_ins(void *data, struct buffer_head *bh)
1891 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1892 struct ocfs2_extent_list *el = &eb->h_list;
1893 struct buffer_head **ret = data;
1895 /* We want to retain only the leaf block. */
1896 if (le16_to_cpu(el->l_tree_depth) == 0) {
1902 * Find the leaf block in the tree which would contain cpos. No
1903 * checking of the actual leaf is done.
1905 * Some paths want to call this instead of allocating a path structure
1906 * and calling ocfs2_find_path().
1908 * This function doesn't handle non btree extent lists.
1910 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1911 struct ocfs2_extent_list *root_el, u32 cpos,
1912 struct buffer_head **leaf_bh)
1915 struct buffer_head *bh = NULL;
1917 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1929 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1931 * Basically, we've moved stuff around at the bottom of the tree and
1932 * we need to fix up the extent records above the changes to reflect
1935 * left_rec: the record on the left.
1936 * left_child_el: is the child list pointed to by left_rec
1937 * right_rec: the record to the right of left_rec
1938 * right_child_el: is the child list pointed to by right_rec
1940 * By definition, this only works on interior nodes.
1942 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1943 struct ocfs2_extent_list *left_child_el,
1944 struct ocfs2_extent_rec *right_rec,
1945 struct ocfs2_extent_list *right_child_el)
1947 u32 left_clusters, right_end;
1950 * Interior nodes never have holes. Their cpos is the cpos of
1951 * the leftmost record in their child list. Their cluster
1952 * count covers the full theoretical range of their child list
1953 * - the range between their cpos and the cpos of the record
1954 * immediately to their right.
1956 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1957 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1958 BUG_ON(right_child_el->l_tree_depth);
1959 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1960 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1962 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1963 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1966 * Calculate the rightmost cluster count boundary before
1967 * moving cpos - we will need to adjust clusters after
1968 * updating e_cpos to keep the same highest cluster count.
1970 right_end = le32_to_cpu(right_rec->e_cpos);
1971 right_end += le32_to_cpu(right_rec->e_int_clusters);
1973 right_rec->e_cpos = left_rec->e_cpos;
1974 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1976 right_end -= le32_to_cpu(right_rec->e_cpos);
1977 right_rec->e_int_clusters = cpu_to_le32(right_end);
1981 * Adjust the adjacent root node records involved in a
1982 * rotation. left_el_blkno is passed in as a key so that we can easily
1983 * find it's index in the root list.
1985 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1986 struct ocfs2_extent_list *left_el,
1987 struct ocfs2_extent_list *right_el,
1992 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1993 le16_to_cpu(left_el->l_tree_depth));
1995 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1996 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2001 * The path walking code should have never returned a root and
2002 * two paths which are not adjacent.
2004 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2006 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
2007 &root_el->l_recs[i + 1], right_el);
2011 * We've changed a leaf block (in right_path) and need to reflect that
2012 * change back up the subtree.
2014 * This happens in multiple places:
2015 * - When we've moved an extent record from the left path leaf to the right
2016 * path leaf to make room for an empty extent in the left path leaf.
2017 * - When our insert into the right path leaf is at the leftmost edge
2018 * and requires an update of the path immediately to it's left. This
2019 * can occur at the end of some types of rotation and appending inserts.
2020 * - When we've adjusted the last extent record in the left path leaf and the
2021 * 1st extent record in the right path leaf during cross extent block merge.
2023 static void ocfs2_complete_edge_insert(handle_t *handle,
2024 struct ocfs2_path *left_path,
2025 struct ocfs2_path *right_path,
2029 struct ocfs2_extent_list *el, *left_el, *right_el;
2030 struct ocfs2_extent_rec *left_rec, *right_rec;
2031 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2034 * Update the counts and position values within all the
2035 * interior nodes to reflect the leaf rotation we just did.
2037 * The root node is handled below the loop.
2039 * We begin the loop with right_el and left_el pointing to the
2040 * leaf lists and work our way up.
2042 * NOTE: within this loop, left_el and right_el always refer
2043 * to the *child* lists.
2045 left_el = path_leaf_el(left_path);
2046 right_el = path_leaf_el(right_path);
2047 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2048 trace_ocfs2_complete_edge_insert(i);
2051 * One nice property of knowing that all of these
2052 * nodes are below the root is that we only deal with
2053 * the leftmost right node record and the rightmost
2056 el = left_path->p_node[i].el;
2057 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2058 left_rec = &el->l_recs[idx];
2060 el = right_path->p_node[i].el;
2061 right_rec = &el->l_recs[0];
2063 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2066 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2067 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2070 * Setup our list pointers now so that the current
2071 * parents become children in the next iteration.
2073 left_el = left_path->p_node[i].el;
2074 right_el = right_path->p_node[i].el;
2078 * At the root node, adjust the two adjacent records which
2079 * begin our path to the leaves.
2082 el = left_path->p_node[subtree_index].el;
2083 left_el = left_path->p_node[subtree_index + 1].el;
2084 right_el = right_path->p_node[subtree_index + 1].el;
2086 ocfs2_adjust_root_records(el, left_el, right_el,
2087 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2089 root_bh = left_path->p_node[subtree_index].bh;
2091 ocfs2_journal_dirty(handle, root_bh);
2094 static int ocfs2_rotate_subtree_right(handle_t *handle,
2095 struct ocfs2_extent_tree *et,
2096 struct ocfs2_path *left_path,
2097 struct ocfs2_path *right_path,
2101 struct buffer_head *right_leaf_bh;
2102 struct buffer_head *left_leaf_bh = NULL;
2103 struct buffer_head *root_bh;
2104 struct ocfs2_extent_list *right_el, *left_el;
2105 struct ocfs2_extent_rec move_rec;
2107 left_leaf_bh = path_leaf_bh(left_path);
2108 left_el = path_leaf_el(left_path);
2110 if (left_el->l_next_free_rec != left_el->l_count) {
2111 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2112 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2113 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2114 (unsigned long long)left_leaf_bh->b_blocknr,
2115 le16_to_cpu(left_el->l_next_free_rec));
2120 * This extent block may already have an empty record, so we
2121 * return early if so.
2123 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2126 root_bh = left_path->p_node[subtree_index].bh;
2127 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2129 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2136 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2137 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2144 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2152 right_leaf_bh = path_leaf_bh(right_path);
2153 right_el = path_leaf_el(right_path);
2155 /* This is a code error, not a disk corruption. */
2156 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2157 "because rightmost leaf block %llu is empty\n",
2158 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2159 (unsigned long long)right_leaf_bh->b_blocknr);
2161 ocfs2_create_empty_extent(right_el);
2163 ocfs2_journal_dirty(handle, right_leaf_bh);
2165 /* Do the copy now. */
2166 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2167 move_rec = left_el->l_recs[i];
2168 right_el->l_recs[0] = move_rec;
2171 * Clear out the record we just copied and shift everything
2172 * over, leaving an empty extent in the left leaf.
2174 * We temporarily subtract from next_free_rec so that the
2175 * shift will lose the tail record (which is now defunct).
2177 le16_add_cpu(&left_el->l_next_free_rec, -1);
2178 ocfs2_shift_records_right(left_el);
2179 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2180 le16_add_cpu(&left_el->l_next_free_rec, 1);
2182 ocfs2_journal_dirty(handle, left_leaf_bh);
2184 ocfs2_complete_edge_insert(handle, left_path, right_path,
2192 * Given a full path, determine what cpos value would return us a path
2193 * containing the leaf immediately to the left of the current one.
2195 * Will return zero if the path passed in is already the leftmost path.
2197 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2198 struct ocfs2_path *path, u32 *cpos)
2202 struct ocfs2_extent_list *el;
2204 BUG_ON(path->p_tree_depth == 0);
2208 blkno = path_leaf_bh(path)->b_blocknr;
2210 /* Start at the tree node just above the leaf and work our way up. */
2211 i = path->p_tree_depth - 1;
2213 el = path->p_node[i].el;
2216 * Find the extent record just before the one in our
2219 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2220 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2224 * We've determined that the
2225 * path specified is already
2226 * the leftmost one - return a
2232 * The leftmost record points to our
2233 * leaf - we need to travel up the
2239 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2240 *cpos = *cpos + ocfs2_rec_clusters(el,
2241 &el->l_recs[j - 1]);
2248 * If we got here, we never found a valid node where
2249 * the tree indicated one should be.
2251 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2252 (unsigned long long)blkno);
2257 blkno = path->p_node[i].bh->b_blocknr;
2266 * Extend the transaction by enough credits to complete the rotation,
2267 * and still leave at least the original number of credits allocated
2268 * to this transaction.
2270 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2272 struct ocfs2_path *path)
2275 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2277 if (handle->h_buffer_credits < credits)
2278 ret = ocfs2_extend_trans(handle,
2279 credits - handle->h_buffer_credits);
2285 * Trap the case where we're inserting into the theoretical range past
2286 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2287 * whose cpos is less than ours into the right leaf.
2289 * It's only necessary to look at the rightmost record of the left
2290 * leaf because the logic that calls us should ensure that the
2291 * theoretical ranges in the path components above the leaves are
2294 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2297 struct ocfs2_extent_list *left_el;
2298 struct ocfs2_extent_rec *rec;
2301 left_el = path_leaf_el(left_path);
2302 next_free = le16_to_cpu(left_el->l_next_free_rec);
2303 rec = &left_el->l_recs[next_free - 1];
2305 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2310 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2312 int next_free = le16_to_cpu(el->l_next_free_rec);
2314 struct ocfs2_extent_rec *rec;
2319 rec = &el->l_recs[0];
2320 if (ocfs2_is_empty_extent(rec)) {
2324 rec = &el->l_recs[1];
2327 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2328 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2334 * Rotate all the records in a btree right one record, starting at insert_cpos.
2336 * The path to the rightmost leaf should be passed in.
2338 * The array is assumed to be large enough to hold an entire path (tree depth).
2340 * Upon successful return from this function:
2342 * - The 'right_path' array will contain a path to the leaf block
2343 * whose range contains e_cpos.
2344 * - That leaf block will have a single empty extent in list index 0.
2345 * - In the case that the rotation requires a post-insert update,
2346 * *ret_left_path will contain a valid path which can be passed to
2347 * ocfs2_insert_path().
2349 static int ocfs2_rotate_tree_right(handle_t *handle,
2350 struct ocfs2_extent_tree *et,
2351 enum ocfs2_split_type split,
2353 struct ocfs2_path *right_path,
2354 struct ocfs2_path **ret_left_path)
2356 int ret, start, orig_credits = handle->h_buffer_credits;
2358 struct ocfs2_path *left_path = NULL;
2359 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2361 *ret_left_path = NULL;
2363 left_path = ocfs2_new_path_from_path(right_path);
2370 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2376 trace_ocfs2_rotate_tree_right(
2377 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2381 * What we want to do here is:
2383 * 1) Start with the rightmost path.
2385 * 2) Determine a path to the leaf block directly to the left
2388 * 3) Determine the 'subtree root' - the lowest level tree node
2389 * which contains a path to both leaves.
2391 * 4) Rotate the subtree.
2393 * 5) Find the next subtree by considering the left path to be
2394 * the new right path.
2396 * The check at the top of this while loop also accepts
2397 * insert_cpos == cpos because cpos is only a _theoretical_
2398 * value to get us the left path - insert_cpos might very well
2399 * be filling that hole.
2401 * Stop at a cpos of '0' because we either started at the
2402 * leftmost branch (i.e., a tree with one branch and a
2403 * rotation inside of it), or we've gone as far as we can in
2404 * rotating subtrees.
2406 while (cpos && insert_cpos <= cpos) {
2407 trace_ocfs2_rotate_tree_right(
2408 (unsigned long long)
2409 ocfs2_metadata_cache_owner(et->et_ci),
2412 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2418 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2419 path_leaf_bh(right_path),
2420 "Owner %llu: error during insert of %u "
2421 "(left path cpos %u) results in two identical "
2422 "paths ending at %llu\n",
2423 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2425 (unsigned long long)
2426 path_leaf_bh(left_path)->b_blocknr);
2428 if (split == SPLIT_NONE &&
2429 ocfs2_rotate_requires_path_adjustment(left_path,
2433 * We've rotated the tree as much as we
2434 * should. The rest is up to
2435 * ocfs2_insert_path() to complete, after the
2436 * record insertion. We indicate this
2437 * situation by returning the left path.
2439 * The reason we don't adjust the records here
2440 * before the record insert is that an error
2441 * later might break the rule where a parent
2442 * record e_cpos will reflect the actual
2443 * e_cpos of the 1st nonempty record of the
2446 *ret_left_path = left_path;
2450 start = ocfs2_find_subtree_root(et, left_path, right_path);
2452 trace_ocfs2_rotate_subtree(start,
2453 (unsigned long long)
2454 right_path->p_node[start].bh->b_blocknr,
2455 right_path->p_tree_depth);
2457 ret = ocfs2_extend_rotate_transaction(handle, start,
2458 orig_credits, right_path);
2464 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2471 if (split != SPLIT_NONE &&
2472 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2475 * A rotate moves the rightmost left leaf
2476 * record over to the leftmost right leaf
2477 * slot. If we're doing an extent split
2478 * instead of a real insert, then we have to
2479 * check that the extent to be split wasn't
2480 * just moved over. If it was, then we can
2481 * exit here, passing left_path back -
2482 * ocfs2_split_extent() is smart enough to
2483 * search both leaves.
2485 *ret_left_path = left_path;
2490 * There is no need to re-read the next right path
2491 * as we know that it'll be our current left
2492 * path. Optimize by copying values instead.
2494 ocfs2_mv_path(right_path, left_path);
2496 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2504 ocfs2_free_path(left_path);
2510 static int ocfs2_update_edge_lengths(handle_t *handle,
2511 struct ocfs2_extent_tree *et,
2512 int subtree_index, struct ocfs2_path *path)
2515 struct ocfs2_extent_rec *rec;
2516 struct ocfs2_extent_list *el;
2517 struct ocfs2_extent_block *eb;
2520 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2526 /* Path should always be rightmost. */
2527 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2528 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2531 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2532 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2533 rec = &el->l_recs[idx];
2534 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2536 for (i = 0; i < path->p_tree_depth; i++) {
2537 el = path->p_node[i].el;
2538 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2539 rec = &el->l_recs[idx];
2541 rec->e_int_clusters = cpu_to_le32(range);
2542 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2544 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2550 static void ocfs2_unlink_path(handle_t *handle,
2551 struct ocfs2_extent_tree *et,
2552 struct ocfs2_cached_dealloc_ctxt *dealloc,
2553 struct ocfs2_path *path, int unlink_start)
2556 struct ocfs2_extent_block *eb;
2557 struct ocfs2_extent_list *el;
2558 struct buffer_head *bh;
2560 for(i = unlink_start; i < path_num_items(path); i++) {
2561 bh = path->p_node[i].bh;
2563 eb = (struct ocfs2_extent_block *)bh->b_data;
2565 * Not all nodes might have had their final count
2566 * decremented by the caller - handle this here.
2569 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2571 "Inode %llu, attempted to remove extent block "
2572 "%llu with %u records\n",
2573 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2574 (unsigned long long)le64_to_cpu(eb->h_blkno),
2575 le16_to_cpu(el->l_next_free_rec));
2577 ocfs2_journal_dirty(handle, bh);
2578 ocfs2_remove_from_cache(et->et_ci, bh);
2582 el->l_next_free_rec = 0;
2583 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2585 ocfs2_journal_dirty(handle, bh);
2587 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2591 ocfs2_remove_from_cache(et->et_ci, bh);
2595 static void ocfs2_unlink_subtree(handle_t *handle,
2596 struct ocfs2_extent_tree *et,
2597 struct ocfs2_path *left_path,
2598 struct ocfs2_path *right_path,
2600 struct ocfs2_cached_dealloc_ctxt *dealloc)
2603 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2604 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2605 struct ocfs2_extent_list *el;
2606 struct ocfs2_extent_block *eb;
2608 el = path_leaf_el(left_path);
2610 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2612 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2613 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2616 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2618 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2619 le16_add_cpu(&root_el->l_next_free_rec, -1);
2621 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2622 eb->h_next_leaf_blk = 0;
2624 ocfs2_journal_dirty(handle, root_bh);
2625 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2627 ocfs2_unlink_path(handle, et, dealloc, right_path,
2631 static int ocfs2_rotate_subtree_left(handle_t *handle,
2632 struct ocfs2_extent_tree *et,
2633 struct ocfs2_path *left_path,
2634 struct ocfs2_path *right_path,
2636 struct ocfs2_cached_dealloc_ctxt *dealloc,
2639 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2640 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2641 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2642 struct ocfs2_extent_block *eb;
2646 right_leaf_el = path_leaf_el(right_path);
2647 left_leaf_el = path_leaf_el(left_path);
2648 root_bh = left_path->p_node[subtree_index].bh;
2649 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2651 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2654 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2655 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2657 * It's legal for us to proceed if the right leaf is
2658 * the rightmost one and it has an empty extent. There
2659 * are two cases to handle - whether the leaf will be
2660 * empty after removal or not. If the leaf isn't empty
2661 * then just remove the empty extent up front. The
2662 * next block will handle empty leaves by flagging
2665 * Non rightmost leaves will throw -EAGAIN and the
2666 * caller can manually move the subtree and retry.
2669 if (eb->h_next_leaf_blk != 0ULL)
2672 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2673 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2674 path_leaf_bh(right_path),
2675 OCFS2_JOURNAL_ACCESS_WRITE);
2681 ocfs2_remove_empty_extent(right_leaf_el);
2683 right_has_empty = 1;
2686 if (eb->h_next_leaf_blk == 0ULL &&
2687 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2689 * We have to update i_last_eb_blk during the meta
2692 ret = ocfs2_et_root_journal_access(handle, et,
2693 OCFS2_JOURNAL_ACCESS_WRITE);
2699 del_right_subtree = 1;
2703 * Getting here with an empty extent in the right path implies
2704 * that it's the rightmost path and will be deleted.
2706 BUG_ON(right_has_empty && !del_right_subtree);
2708 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2715 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2716 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2723 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2731 if (!right_has_empty) {
2733 * Only do this if we're moving a real
2734 * record. Otherwise, the action is delayed until
2735 * after removal of the right path in which case we
2736 * can do a simple shift to remove the empty extent.
2738 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2739 memset(&right_leaf_el->l_recs[0], 0,
2740 sizeof(struct ocfs2_extent_rec));
2742 if (eb->h_next_leaf_blk == 0ULL) {
2744 * Move recs over to get rid of empty extent, decrease
2745 * next_free. This is allowed to remove the last
2746 * extent in our leaf (setting l_next_free_rec to
2747 * zero) - the delete code below won't care.
2749 ocfs2_remove_empty_extent(right_leaf_el);
2752 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2753 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2755 if (del_right_subtree) {
2756 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2757 subtree_index, dealloc);
2758 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
2765 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2766 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2769 * Removal of the extent in the left leaf was skipped
2770 * above so we could delete the right path
2773 if (right_has_empty)
2774 ocfs2_remove_empty_extent(left_leaf_el);
2776 ocfs2_journal_dirty(handle, et_root_bh);
2780 ocfs2_complete_edge_insert(handle, left_path, right_path,
2788 * Given a full path, determine what cpos value would return us a path
2789 * containing the leaf immediately to the right of the current one.
2791 * Will return zero if the path passed in is already the rightmost path.
2793 * This looks similar, but is subtly different to
2794 * ocfs2_find_cpos_for_left_leaf().
2796 int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2797 struct ocfs2_path *path, u32 *cpos)
2801 struct ocfs2_extent_list *el;
2805 if (path->p_tree_depth == 0)
2808 blkno = path_leaf_bh(path)->b_blocknr;
2810 /* Start at the tree node just above the leaf and work our way up. */
2811 i = path->p_tree_depth - 1;
2815 el = path->p_node[i].el;
2818 * Find the extent record just after the one in our
2821 next_free = le16_to_cpu(el->l_next_free_rec);
2822 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2823 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2824 if (j == (next_free - 1)) {
2827 * We've determined that the
2828 * path specified is already
2829 * the rightmost one - return a
2835 * The rightmost record points to our
2836 * leaf - we need to travel up the
2842 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2848 * If we got here, we never found a valid node where
2849 * the tree indicated one should be.
2851 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2852 (unsigned long long)blkno);
2857 blkno = path->p_node[i].bh->b_blocknr;
2865 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2866 struct ocfs2_extent_tree *et,
2867 struct ocfs2_path *path)
2870 struct buffer_head *bh = path_leaf_bh(path);
2871 struct ocfs2_extent_list *el = path_leaf_el(path);
2873 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2876 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2877 path_num_items(path) - 1);
2883 ocfs2_remove_empty_extent(el);
2884 ocfs2_journal_dirty(handle, bh);
2890 static int __ocfs2_rotate_tree_left(handle_t *handle,
2891 struct ocfs2_extent_tree *et,
2893 struct ocfs2_path *path,
2894 struct ocfs2_cached_dealloc_ctxt *dealloc,
2895 struct ocfs2_path **empty_extent_path)
2897 int ret, subtree_root, deleted;
2899 struct ocfs2_path *left_path = NULL;
2900 struct ocfs2_path *right_path = NULL;
2901 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2903 if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2906 *empty_extent_path = NULL;
2908 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2914 left_path = ocfs2_new_path_from_path(path);
2921 ocfs2_cp_path(left_path, path);
2923 right_path = ocfs2_new_path_from_path(path);
2930 while (right_cpos) {
2931 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2937 subtree_root = ocfs2_find_subtree_root(et, left_path,
2940 trace_ocfs2_rotate_subtree(subtree_root,
2941 (unsigned long long)
2942 right_path->p_node[subtree_root].bh->b_blocknr,
2943 right_path->p_tree_depth);
2945 ret = ocfs2_extend_rotate_transaction(handle, 0,
2946 orig_credits, left_path);
2953 * Caller might still want to make changes to the
2954 * tree root, so re-add it to the journal here.
2956 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2963 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2964 right_path, subtree_root,
2966 if (ret == -EAGAIN) {
2968 * The rotation has to temporarily stop due to
2969 * the right subtree having an empty
2970 * extent. Pass it back to the caller for a
2973 *empty_extent_path = right_path;
2983 * The subtree rotate might have removed records on
2984 * the rightmost edge. If so, then rotation is
2990 ocfs2_mv_path(left_path, right_path);
2992 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3001 ocfs2_free_path(right_path);
3002 ocfs2_free_path(left_path);
3007 static int ocfs2_remove_rightmost_path(handle_t *handle,
3008 struct ocfs2_extent_tree *et,
3009 struct ocfs2_path *path,
3010 struct ocfs2_cached_dealloc_ctxt *dealloc)
3012 int ret, subtree_index;
3014 struct ocfs2_path *left_path = NULL;
3015 struct ocfs2_extent_block *eb;
3016 struct ocfs2_extent_list *el;
3018 ret = ocfs2_et_sanity_check(et);
3022 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3028 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3037 * We have a path to the left of this one - it needs
3040 left_path = ocfs2_new_path_from_path(path);
3047 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3053 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3059 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3061 ocfs2_unlink_subtree(handle, et, left_path, path,
3062 subtree_index, dealloc);
3063 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
3070 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3071 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3074 * 'path' is also the leftmost path which
3075 * means it must be the only one. This gets
3076 * handled differently because we want to
3077 * revert the root back to having extents
3080 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3082 el = et->et_root_el;
3083 el->l_tree_depth = 0;
3084 el->l_next_free_rec = 0;
3085 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3087 ocfs2_et_set_last_eb_blk(et, 0);
3090 ocfs2_journal_dirty(handle, path_root_bh(path));
3093 ocfs2_free_path(left_path);
3097 static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3098 struct ocfs2_extent_tree *et,
3099 struct ocfs2_path *path,
3100 struct ocfs2_cached_dealloc_ctxt *dealloc)
3104 int credits = path->p_tree_depth * 2 + 1;
3106 handle = ocfs2_start_trans(osb, credits);
3107 if (IS_ERR(handle)) {
3108 ret = PTR_ERR(handle);
3113 ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3117 ocfs2_commit_trans(osb, handle);
3122 * Left rotation of btree records.
3124 * In many ways, this is (unsurprisingly) the opposite of right
3125 * rotation. We start at some non-rightmost path containing an empty
3126 * extent in the leaf block. The code works its way to the rightmost
3127 * path by rotating records to the left in every subtree.
3129 * This is used by any code which reduces the number of extent records
3130 * in a leaf. After removal, an empty record should be placed in the
3131 * leftmost list position.
3133 * This won't handle a length update of the rightmost path records if
3134 * the rightmost tree leaf record is removed so the caller is
3135 * responsible for detecting and correcting that.
3137 static int ocfs2_rotate_tree_left(handle_t *handle,
3138 struct ocfs2_extent_tree *et,
3139 struct ocfs2_path *path,
3140 struct ocfs2_cached_dealloc_ctxt *dealloc)
3142 int ret, orig_credits = handle->h_buffer_credits;
3143 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3144 struct ocfs2_extent_block *eb;
3145 struct ocfs2_extent_list *el;
3147 el = path_leaf_el(path);
3148 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3151 if (path->p_tree_depth == 0) {
3152 rightmost_no_delete:
3154 * Inline extents. This is trivially handled, so do
3157 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3164 * Handle rightmost branch now. There's several cases:
3165 * 1) simple rotation leaving records in there. That's trivial.
3166 * 2) rotation requiring a branch delete - there's no more
3167 * records left. Two cases of this:
3168 * a) There are branches to the left.
3169 * b) This is also the leftmost (the only) branch.
3171 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3172 * 2a) we need the left branch so that we can update it with the unlink
3173 * 2b) we need to bring the root back to inline extents.
3176 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3178 if (eb->h_next_leaf_blk == 0) {
3180 * This gets a bit tricky if we're going to delete the
3181 * rightmost path. Get the other cases out of the way
3184 if (le16_to_cpu(el->l_next_free_rec) > 1)
3185 goto rightmost_no_delete;
3187 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3189 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3190 "Owner %llu has empty extent block at %llu\n",
3191 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3192 (unsigned long long)le64_to_cpu(eb->h_blkno));
3197 * XXX: The caller can not trust "path" any more after
3198 * this as it will have been deleted. What do we do?
3200 * In theory the rotate-for-merge code will never get
3201 * here because it'll always ask for a rotate in a
3205 ret = ocfs2_remove_rightmost_path(handle, et, path,
3213 * Now we can loop, remembering the path we get from -EAGAIN
3214 * and restarting from there.
3217 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3218 dealloc, &restart_path);
3219 if (ret && ret != -EAGAIN) {
3224 while (ret == -EAGAIN) {
3225 tmp_path = restart_path;
3226 restart_path = NULL;
3228 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3231 if (ret && ret != -EAGAIN) {
3236 ocfs2_free_path(tmp_path);
3244 ocfs2_free_path(tmp_path);
3245 ocfs2_free_path(restart_path);
3249 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3252 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3255 if (rec->e_leaf_clusters == 0) {
3257 * We consumed all of the merged-from record. An empty
3258 * extent cannot exist anywhere but the 1st array
3259 * position, so move things over if the merged-from
3260 * record doesn't occupy that position.
3262 * This creates a new empty extent so the caller
3263 * should be smart enough to have removed any existing
3267 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3268 size = index * sizeof(struct ocfs2_extent_rec);
3269 memmove(&el->l_recs[1], &el->l_recs[0], size);
3273 * Always memset - the caller doesn't check whether it
3274 * created an empty extent, so there could be junk in
3277 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3281 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3282 struct ocfs2_path *left_path,
3283 struct ocfs2_path **ret_right_path)
3287 struct ocfs2_path *right_path = NULL;
3288 struct ocfs2_extent_list *left_el;
3290 *ret_right_path = NULL;
3292 /* This function shouldn't be called for non-trees. */
3293 BUG_ON(left_path->p_tree_depth == 0);
3295 left_el = path_leaf_el(left_path);
3296 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3298 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3299 left_path, &right_cpos);
3305 /* This function shouldn't be called for the rightmost leaf. */
3306 BUG_ON(right_cpos == 0);
3308 right_path = ocfs2_new_path_from_path(left_path);
3315 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3321 *ret_right_path = right_path;
3324 ocfs2_free_path(right_path);
3329 * Remove split_rec clusters from the record at index and merge them
3330 * onto the beginning of the record "next" to it.
3331 * For index < l_count - 1, the next means the extent rec at index + 1.
3332 * For index == l_count - 1, the "next" means the 1st extent rec of the
3333 * next extent block.
3335 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3337 struct ocfs2_extent_tree *et,
3338 struct ocfs2_extent_rec *split_rec,
3341 int ret, next_free, i;
3342 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3343 struct ocfs2_extent_rec *left_rec;
3344 struct ocfs2_extent_rec *right_rec;
3345 struct ocfs2_extent_list *right_el;
3346 struct ocfs2_path *right_path = NULL;
3347 int subtree_index = 0;
3348 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3349 struct buffer_head *bh = path_leaf_bh(left_path);
3350 struct buffer_head *root_bh = NULL;
3352 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3353 left_rec = &el->l_recs[index];
3355 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3356 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3357 /* we meet with a cross extent block merge. */
3358 ret = ocfs2_get_right_path(et, left_path, &right_path);
3364 right_el = path_leaf_el(right_path);
3365 next_free = le16_to_cpu(right_el->l_next_free_rec);
3366 BUG_ON(next_free <= 0);
3367 right_rec = &right_el->l_recs[0];
3368 if (ocfs2_is_empty_extent(right_rec)) {
3369 BUG_ON(next_free <= 1);
3370 right_rec = &right_el->l_recs[1];
3373 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3374 le16_to_cpu(left_rec->e_leaf_clusters) !=
3375 le32_to_cpu(right_rec->e_cpos));
3377 subtree_index = ocfs2_find_subtree_root(et, left_path,
3380 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3381 handle->h_buffer_credits,
3388 root_bh = left_path->p_node[subtree_index].bh;
3389 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3391 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3398 for (i = subtree_index + 1;
3399 i < path_num_items(right_path); i++) {
3400 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3407 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3416 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3417 right_rec = &el->l_recs[index + 1];
3420 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3421 path_num_items(left_path) - 1);
3427 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3429 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3430 le64_add_cpu(&right_rec->e_blkno,
3431 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3433 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3435 ocfs2_cleanup_merge(el, index);
3437 ocfs2_journal_dirty(handle, bh);
3439 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3440 ocfs2_complete_edge_insert(handle, left_path, right_path,
3444 ocfs2_free_path(right_path);
3448 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3449 struct ocfs2_path *right_path,
3450 struct ocfs2_path **ret_left_path)
3454 struct ocfs2_path *left_path = NULL;
3456 *ret_left_path = NULL;
3458 /* This function shouldn't be called for non-trees. */
3459 BUG_ON(right_path->p_tree_depth == 0);
3461 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3462 right_path, &left_cpos);
3468 /* This function shouldn't be called for the leftmost leaf. */
3469 BUG_ON(left_cpos == 0);
3471 left_path = ocfs2_new_path_from_path(right_path);
3478 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3484 *ret_left_path = left_path;
3487 ocfs2_free_path(left_path);
3492 * Remove split_rec clusters from the record at index and merge them
3493 * onto the tail of the record "before" it.
3494 * For index > 0, the "before" means the extent rec at index - 1.
3496 * For index == 0, the "before" means the last record of the previous
3497 * extent block. And there is also a situation that we may need to
3498 * remove the rightmost leaf extent block in the right_path and change
3499 * the right path to indicate the new rightmost path.
3501 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3503 struct ocfs2_extent_tree *et,
3504 struct ocfs2_extent_rec *split_rec,
3505 struct ocfs2_cached_dealloc_ctxt *dealloc,
3508 int ret, i, subtree_index = 0, has_empty_extent = 0;
3509 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3510 struct ocfs2_extent_rec *left_rec;
3511 struct ocfs2_extent_rec *right_rec;
3512 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3513 struct buffer_head *bh = path_leaf_bh(right_path);
3514 struct buffer_head *root_bh = NULL;
3515 struct ocfs2_path *left_path = NULL;
3516 struct ocfs2_extent_list *left_el;
3520 right_rec = &el->l_recs[index];
3522 /* we meet with a cross extent block merge. */
3523 ret = ocfs2_get_left_path(et, right_path, &left_path);
3529 left_el = path_leaf_el(left_path);
3530 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3531 le16_to_cpu(left_el->l_count));
3533 left_rec = &left_el->l_recs[
3534 le16_to_cpu(left_el->l_next_free_rec) - 1];
3535 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3536 le16_to_cpu(left_rec->e_leaf_clusters) !=
3537 le32_to_cpu(split_rec->e_cpos));
3539 subtree_index = ocfs2_find_subtree_root(et, left_path,
3542 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3543 handle->h_buffer_credits,
3550 root_bh = left_path->p_node[subtree_index].bh;
3551 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3553 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3560 for (i = subtree_index + 1;
3561 i < path_num_items(right_path); i++) {
3562 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3569 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3577 left_rec = &el->l_recs[index - 1];
3578 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3579 has_empty_extent = 1;
3582 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3583 path_num_items(right_path) - 1);
3589 if (has_empty_extent && index == 1) {
3591 * The easy case - we can just plop the record right in.
3593 *left_rec = *split_rec;
3595 has_empty_extent = 0;
3597 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3599 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3600 le64_add_cpu(&right_rec->e_blkno,
3601 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3603 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3605 ocfs2_cleanup_merge(el, index);
3607 ocfs2_journal_dirty(handle, bh);
3609 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3612 * In the situation that the right_rec is empty and the extent
3613 * block is empty also, ocfs2_complete_edge_insert can't handle
3614 * it and we need to delete the right extent block.
3616 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3617 le16_to_cpu(el->l_next_free_rec) == 1) {
3618 /* extend credit for ocfs2_remove_rightmost_path */
3619 ret = ocfs2_extend_rotate_transaction(handle, 0,
3620 handle->h_buffer_credits,
3627 ret = ocfs2_remove_rightmost_path(handle, et,
3635 /* Now the rightmost extent block has been deleted.
3636 * So we use the new rightmost path.
3638 ocfs2_mv_path(right_path, left_path);
3641 ocfs2_complete_edge_insert(handle, left_path,
3642 right_path, subtree_index);
3645 ocfs2_free_path(left_path);
3649 static int ocfs2_try_to_merge_extent(handle_t *handle,
3650 struct ocfs2_extent_tree *et,
3651 struct ocfs2_path *path,
3653 struct ocfs2_extent_rec *split_rec,
3654 struct ocfs2_cached_dealloc_ctxt *dealloc,
3655 struct ocfs2_merge_ctxt *ctxt)
3658 struct ocfs2_extent_list *el = path_leaf_el(path);
3659 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3661 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3663 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3664 /* extend credit for ocfs2_remove_rightmost_path */
3665 ret = ocfs2_extend_rotate_transaction(handle, 0,
3666 handle->h_buffer_credits,
3673 * The merge code will need to create an empty
3674 * extent to take the place of the newly
3675 * emptied slot. Remove any pre-existing empty
3676 * extents - having more than one in a leaf is
3679 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3685 rec = &el->l_recs[split_index];
3688 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3690 * Left-right contig implies this.
3692 BUG_ON(!ctxt->c_split_covers_rec);
3695 * Since the leftright insert always covers the entire
3696 * extent, this call will delete the insert record
3697 * entirely, resulting in an empty extent record added to
3700 * Since the adding of an empty extent shifts
3701 * everything back to the right, there's no need to
3702 * update split_index here.
3704 * When the split_index is zero, we need to merge it to the
3705 * prevoius extent block. It is more efficient and easier
3706 * if we do merge_right first and merge_left later.
3708 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3716 * We can only get this from logic error above.
3718 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3720 /* extend credit for ocfs2_remove_rightmost_path */
3721 ret = ocfs2_extend_rotate_transaction(handle, 0,
3722 handle->h_buffer_credits,
3729 /* The merge left us with an empty extent, remove it. */
3730 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3736 rec = &el->l_recs[split_index];
3739 * Note that we don't pass split_rec here on purpose -
3740 * we've merged it into the rec already.
3742 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3743 dealloc, split_index);
3750 /* extend credit for ocfs2_remove_rightmost_path */
3751 ret = ocfs2_extend_rotate_transaction(handle, 0,
3752 handle->h_buffer_credits,
3759 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3761 * Error from this last rotate is not critical, so
3762 * print but don't bubble it up.
3769 * Merge a record to the left or right.
3771 * 'contig_type' is relative to the existing record,
3772 * so for example, if we're "right contig", it's to
3773 * the record on the left (hence the left merge).
3775 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3776 ret = ocfs2_merge_rec_left(path, handle, et,
3784 ret = ocfs2_merge_rec_right(path, handle,
3793 if (ctxt->c_split_covers_rec) {
3794 /* extend credit for ocfs2_remove_rightmost_path */
3795 ret = ocfs2_extend_rotate_transaction(handle, 0,
3796 handle->h_buffer_credits,
3805 * The merge may have left an empty extent in
3806 * our leaf. Try to rotate it away.
3808 ret = ocfs2_rotate_tree_left(handle, et, path,
3820 static void ocfs2_subtract_from_rec(struct super_block *sb,
3821 enum ocfs2_split_type split,
3822 struct ocfs2_extent_rec *rec,
3823 struct ocfs2_extent_rec *split_rec)
3827 len_blocks = ocfs2_clusters_to_blocks(sb,
3828 le16_to_cpu(split_rec->e_leaf_clusters));
3830 if (split == SPLIT_LEFT) {
3832 * Region is on the left edge of the existing
3835 le32_add_cpu(&rec->e_cpos,
3836 le16_to_cpu(split_rec->e_leaf_clusters));
3837 le64_add_cpu(&rec->e_blkno, len_blocks);
3838 le16_add_cpu(&rec->e_leaf_clusters,
3839 -le16_to_cpu(split_rec->e_leaf_clusters));
3842 * Region is on the right edge of the existing
3845 le16_add_cpu(&rec->e_leaf_clusters,
3846 -le16_to_cpu(split_rec->e_leaf_clusters));
3851 * Do the final bits of extent record insertion at the target leaf
3852 * list. If this leaf is part of an allocation tree, it is assumed
3853 * that the tree above has been prepared.
3855 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3856 struct ocfs2_extent_rec *insert_rec,
3857 struct ocfs2_extent_list *el,
3858 struct ocfs2_insert_type *insert)
3860 int i = insert->ins_contig_index;
3862 struct ocfs2_extent_rec *rec;
3864 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3866 if (insert->ins_split != SPLIT_NONE) {
3867 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3869 rec = &el->l_recs[i];
3870 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3871 insert->ins_split, rec,
3877 * Contiguous insert - either left or right.
3879 if (insert->ins_contig != CONTIG_NONE) {
3880 rec = &el->l_recs[i];
3881 if (insert->ins_contig == CONTIG_LEFT) {
3882 rec->e_blkno = insert_rec->e_blkno;
3883 rec->e_cpos = insert_rec->e_cpos;
3885 le16_add_cpu(&rec->e_leaf_clusters,
3886 le16_to_cpu(insert_rec->e_leaf_clusters));
3891 * Handle insert into an empty leaf.
3893 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3894 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3895 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3896 el->l_recs[0] = *insert_rec;
3897 el->l_next_free_rec = cpu_to_le16(1);
3904 if (insert->ins_appending == APPEND_TAIL) {
3905 i = le16_to_cpu(el->l_next_free_rec) - 1;
3906 rec = &el->l_recs[i];
3907 range = le32_to_cpu(rec->e_cpos)
3908 + le16_to_cpu(rec->e_leaf_clusters);
3909 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3911 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3912 le16_to_cpu(el->l_count),
3913 "owner %llu, depth %u, count %u, next free %u, "
3914 "rec.cpos %u, rec.clusters %u, "
3915 "insert.cpos %u, insert.clusters %u\n",
3916 ocfs2_metadata_cache_owner(et->et_ci),
3917 le16_to_cpu(el->l_tree_depth),
3918 le16_to_cpu(el->l_count),
3919 le16_to_cpu(el->l_next_free_rec),
3920 le32_to_cpu(el->l_recs[i].e_cpos),
3921 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3922 le32_to_cpu(insert_rec->e_cpos),
3923 le16_to_cpu(insert_rec->e_leaf_clusters));
3925 el->l_recs[i] = *insert_rec;
3926 le16_add_cpu(&el->l_next_free_rec, 1);
3932 * Ok, we have to rotate.
3934 * At this point, it is safe to assume that inserting into an
3935 * empty leaf and appending to a leaf have both been handled
3938 * This leaf needs to have space, either by the empty 1st
3939 * extent record, or by virtue of an l_next_rec < l_count.
3941 ocfs2_rotate_leaf(el, insert_rec);
3944 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3945 struct ocfs2_extent_tree *et,
3946 struct ocfs2_path *path,
3947 struct ocfs2_extent_rec *insert_rec)
3949 int ret, i, next_free;
3950 struct buffer_head *bh;
3951 struct ocfs2_extent_list *el;
3952 struct ocfs2_extent_rec *rec;
3955 * Update everything except the leaf block.
3957 for (i = 0; i < path->p_tree_depth; i++) {
3958 bh = path->p_node[i].bh;
3959 el = path->p_node[i].el;
3961 next_free = le16_to_cpu(el->l_next_free_rec);
3962 if (next_free == 0) {
3963 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3964 "Owner %llu has a bad extent list\n",
3965 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3970 rec = &el->l_recs[next_free - 1];
3972 rec->e_int_clusters = insert_rec->e_cpos;
3973 le32_add_cpu(&rec->e_int_clusters,
3974 le16_to_cpu(insert_rec->e_leaf_clusters));
3975 le32_add_cpu(&rec->e_int_clusters,
3976 -le32_to_cpu(rec->e_cpos));
3978 ocfs2_journal_dirty(handle, bh);
3982 static int ocfs2_append_rec_to_path(handle_t *handle,
3983 struct ocfs2_extent_tree *et,
3984 struct ocfs2_extent_rec *insert_rec,
3985 struct ocfs2_path *right_path,
3986 struct ocfs2_path **ret_left_path)
3989 struct ocfs2_extent_list *el;
3990 struct ocfs2_path *left_path = NULL;
3992 *ret_left_path = NULL;
3995 * This shouldn't happen for non-trees. The extent rec cluster
3996 * count manipulation below only works for interior nodes.
3998 BUG_ON(right_path->p_tree_depth == 0);
4001 * If our appending insert is at the leftmost edge of a leaf,
4002 * then we might need to update the rightmost records of the
4005 el = path_leaf_el(right_path);
4006 next_free = le16_to_cpu(el->l_next_free_rec);
4007 if (next_free == 0 ||
4008 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4011 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4012 right_path, &left_cpos);
4018 trace_ocfs2_append_rec_to_path(
4019 (unsigned long long)
4020 ocfs2_metadata_cache_owner(et->et_ci),
4021 le32_to_cpu(insert_rec->e_cpos),
4025 * No need to worry if the append is already in the
4029 left_path = ocfs2_new_path_from_path(right_path);
4036 ret = ocfs2_find_path(et->et_ci, left_path,
4044 * ocfs2_insert_path() will pass the left_path to the
4050 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4056 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4058 *ret_left_path = left_path;
4062 ocfs2_free_path(left_path);
4067 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4068 struct ocfs2_path *left_path,
4069 struct ocfs2_path *right_path,
4070 struct ocfs2_extent_rec *split_rec,
4071 enum ocfs2_split_type split)
4074 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4075 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4076 struct ocfs2_extent_rec *rec, *tmprec;
4078 right_el = path_leaf_el(right_path);
4080 left_el = path_leaf_el(left_path);
4083 insert_el = right_el;
4084 index = ocfs2_search_extent_list(el, cpos);
4086 if (index == 0 && left_path) {
4087 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4090 * This typically means that the record
4091 * started in the left path but moved to the
4092 * right as a result of rotation. We either
4093 * move the existing record to the left, or we
4094 * do the later insert there.
4096 * In this case, the left path should always
4097 * exist as the rotate code will have passed
4098 * it back for a post-insert update.
4101 if (split == SPLIT_LEFT) {
4103 * It's a left split. Since we know
4104 * that the rotate code gave us an
4105 * empty extent in the left path, we
4106 * can just do the insert there.
4108 insert_el = left_el;
4111 * Right split - we have to move the
4112 * existing record over to the left
4113 * leaf. The insert will be into the
4114 * newly created empty extent in the
4117 tmprec = &right_el->l_recs[index];
4118 ocfs2_rotate_leaf(left_el, tmprec);
4121 memset(tmprec, 0, sizeof(*tmprec));
4122 index = ocfs2_search_extent_list(left_el, cpos);
4123 BUG_ON(index == -1);
4128 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4130 * Left path is easy - we can just allow the insert to
4134 insert_el = left_el;
4135 index = ocfs2_search_extent_list(el, cpos);
4136 BUG_ON(index == -1);
4139 rec = &el->l_recs[index];
4140 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4141 split, rec, split_rec);
4142 ocfs2_rotate_leaf(insert_el, split_rec);
4146 * This function only does inserts on an allocation b-tree. For tree
4147 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4149 * right_path is the path we want to do the actual insert
4150 * in. left_path should only be passed in if we need to update that
4151 * portion of the tree after an edge insert.
4153 static int ocfs2_insert_path(handle_t *handle,
4154 struct ocfs2_extent_tree *et,
4155 struct ocfs2_path *left_path,
4156 struct ocfs2_path *right_path,
4157 struct ocfs2_extent_rec *insert_rec,
4158 struct ocfs2_insert_type *insert)
4160 int ret, subtree_index;
4161 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4165 * There's a chance that left_path got passed back to
4166 * us without being accounted for in the
4167 * journal. Extend our transaction here to be sure we
4168 * can change those blocks.
4170 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4176 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4184 * Pass both paths to the journal. The majority of inserts
4185 * will be touching all components anyway.
4187 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4193 if (insert->ins_split != SPLIT_NONE) {
4195 * We could call ocfs2_insert_at_leaf() for some types
4196 * of splits, but it's easier to just let one separate
4197 * function sort it all out.
4199 ocfs2_split_record(et, left_path, right_path,
4200 insert_rec, insert->ins_split);
4203 * Split might have modified either leaf and we don't
4204 * have a guarantee that the later edge insert will
4205 * dirty this for us.
4208 ocfs2_journal_dirty(handle,
4209 path_leaf_bh(left_path));
4211 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4214 ocfs2_journal_dirty(handle, leaf_bh);
4218 * The rotate code has indicated that we need to fix
4219 * up portions of the tree after the insert.
4221 * XXX: Should we extend the transaction here?
4223 subtree_index = ocfs2_find_subtree_root(et, left_path,
4225 ocfs2_complete_edge_insert(handle, left_path, right_path,
4234 static int ocfs2_do_insert_extent(handle_t *handle,
4235 struct ocfs2_extent_tree *et,
4236 struct ocfs2_extent_rec *insert_rec,
4237 struct ocfs2_insert_type *type)
4239 int ret, rotate = 0;
4241 struct ocfs2_path *right_path = NULL;
4242 struct ocfs2_path *left_path = NULL;
4243 struct ocfs2_extent_list *el;
4245 el = et->et_root_el;
4247 ret = ocfs2_et_root_journal_access(handle, et,
4248 OCFS2_JOURNAL_ACCESS_WRITE);
4254 if (le16_to_cpu(el->l_tree_depth) == 0) {
4255 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4256 goto out_update_clusters;
4259 right_path = ocfs2_new_path_from_et(et);
4267 * Determine the path to start with. Rotations need the
4268 * rightmost path, everything else can go directly to the
4271 cpos = le32_to_cpu(insert_rec->e_cpos);
4272 if (type->ins_appending == APPEND_NONE &&
4273 type->ins_contig == CONTIG_NONE) {
4278 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4285 * Rotations and appends need special treatment - they modify
4286 * parts of the tree's above them.
4288 * Both might pass back a path immediate to the left of the
4289 * one being inserted to. This will be cause
4290 * ocfs2_insert_path() to modify the rightmost records of
4291 * left_path to account for an edge insert.
4293 * XXX: When modifying this code, keep in mind that an insert
4294 * can wind up skipping both of these two special cases...
4297 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4298 le32_to_cpu(insert_rec->e_cpos),
4299 right_path, &left_path);
4306 * ocfs2_rotate_tree_right() might have extended the
4307 * transaction without re-journaling our tree root.
4309 ret = ocfs2_et_root_journal_access(handle, et,
4310 OCFS2_JOURNAL_ACCESS_WRITE);
4315 } else if (type->ins_appending == APPEND_TAIL
4316 && type->ins_contig != CONTIG_LEFT) {
4317 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4318 right_path, &left_path);
4325 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4332 out_update_clusters:
4333 if (type->ins_split == SPLIT_NONE)
4334 ocfs2_et_update_clusters(et,
4335 le16_to_cpu(insert_rec->e_leaf_clusters));
4337 ocfs2_journal_dirty(handle, et->et_root_bh);
4340 ocfs2_free_path(left_path);
4341 ocfs2_free_path(right_path);
4346 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4347 struct ocfs2_path *path,
4348 struct ocfs2_extent_list *el, int index,
4349 struct ocfs2_extent_rec *split_rec,
4350 struct ocfs2_merge_ctxt *ctxt)
4353 enum ocfs2_contig_type ret = CONTIG_NONE;
4354 u32 left_cpos, right_cpos;
4355 struct ocfs2_extent_rec *rec = NULL;
4356 struct ocfs2_extent_list *new_el;
4357 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4358 struct buffer_head *bh;
4359 struct ocfs2_extent_block *eb;
4360 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4363 rec = &el->l_recs[index - 1];
4364 } else if (path->p_tree_depth > 0) {
4365 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4369 if (left_cpos != 0) {
4370 left_path = ocfs2_new_path_from_path(path);
4377 status = ocfs2_find_path(et->et_ci, left_path,
4380 goto free_left_path;
4382 new_el = path_leaf_el(left_path);
4384 if (le16_to_cpu(new_el->l_next_free_rec) !=
4385 le16_to_cpu(new_el->l_count)) {
4386 bh = path_leaf_bh(left_path);
4387 eb = (struct ocfs2_extent_block *)bh->b_data;
4389 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4390 (unsigned long long)le64_to_cpu(eb->h_blkno),
4391 le16_to_cpu(new_el->l_next_free_rec),
4392 le16_to_cpu(new_el->l_count));
4394 goto free_left_path;
4396 rec = &new_el->l_recs[
4397 le16_to_cpu(new_el->l_next_free_rec) - 1];
4402 * We're careful to check for an empty extent record here -
4403 * the merge code will know what to do if it sees one.
4406 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4407 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4410 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4415 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4416 rec = &el->l_recs[index + 1];
4417 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4418 path->p_tree_depth > 0) {
4419 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4421 goto free_left_path;
4423 if (right_cpos == 0)
4424 goto free_left_path;
4426 right_path = ocfs2_new_path_from_path(path);
4430 goto free_left_path;
4433 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4435 goto free_right_path;
4437 new_el = path_leaf_el(right_path);
4438 rec = &new_el->l_recs[0];
4439 if (ocfs2_is_empty_extent(rec)) {
4440 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4441 bh = path_leaf_bh(right_path);
4442 eb = (struct ocfs2_extent_block *)bh->b_data;
4444 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4445 (unsigned long long)le64_to_cpu(eb->h_blkno),
4446 le16_to_cpu(new_el->l_next_free_rec));
4448 goto free_right_path;
4450 rec = &new_el->l_recs[1];
4455 enum ocfs2_contig_type contig_type;
4457 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4459 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4460 ret = CONTIG_LEFTRIGHT;
4461 else if (ret == CONTIG_NONE)
4466 ocfs2_free_path(right_path);
4468 ocfs2_free_path(left_path);
4471 ctxt->c_contig_type = ret;
4476 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4477 struct ocfs2_insert_type *insert,
4478 struct ocfs2_extent_list *el,
4479 struct ocfs2_extent_rec *insert_rec)
4482 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4484 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4486 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4487 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4489 if (contig_type != CONTIG_NONE) {
4490 insert->ins_contig_index = i;
4494 insert->ins_contig = contig_type;
4496 if (insert->ins_contig != CONTIG_NONE) {
4497 struct ocfs2_extent_rec *rec =
4498 &el->l_recs[insert->ins_contig_index];
4499 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4500 le16_to_cpu(insert_rec->e_leaf_clusters);
4503 * Caller might want us to limit the size of extents, don't
4504 * calculate contiguousness if we might exceed that limit.
4506 if (et->et_max_leaf_clusters &&
4507 (len > et->et_max_leaf_clusters))
4508 insert->ins_contig = CONTIG_NONE;
4513 * This should only be called against the righmost leaf extent list.
4515 * ocfs2_figure_appending_type() will figure out whether we'll have to
4516 * insert at the tail of the rightmost leaf.
4518 * This should also work against the root extent list for tree's with 0
4519 * depth. If we consider the root extent list to be the rightmost leaf node
4520 * then the logic here makes sense.
4522 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4523 struct ocfs2_extent_list *el,
4524 struct ocfs2_extent_rec *insert_rec)
4527 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4528 struct ocfs2_extent_rec *rec;
4530 insert->ins_appending = APPEND_NONE;
4532 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4534 if (!el->l_next_free_rec)
4535 goto set_tail_append;
4537 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4538 /* Were all records empty? */
4539 if (le16_to_cpu(el->l_next_free_rec) == 1)
4540 goto set_tail_append;
4543 i = le16_to_cpu(el->l_next_free_rec) - 1;
4544 rec = &el->l_recs[i];
4547 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4548 goto set_tail_append;
4553 insert->ins_appending = APPEND_TAIL;
4557 * Helper function called at the beginning of an insert.
4559 * This computes a few things that are commonly used in the process of
4560 * inserting into the btree:
4561 * - Whether the new extent is contiguous with an existing one.
4562 * - The current tree depth.
4563 * - Whether the insert is an appending one.
4564 * - The total # of free records in the tree.
4566 * All of the information is stored on the ocfs2_insert_type
4569 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4570 struct buffer_head **last_eb_bh,
4571 struct ocfs2_extent_rec *insert_rec,
4573 struct ocfs2_insert_type *insert)
4576 struct ocfs2_extent_block *eb;
4577 struct ocfs2_extent_list *el;
4578 struct ocfs2_path *path = NULL;
4579 struct buffer_head *bh = NULL;
4581 insert->ins_split = SPLIT_NONE;
4583 el = et->et_root_el;
4584 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4586 if (el->l_tree_depth) {
4588 * If we have tree depth, we read in the
4589 * rightmost extent block ahead of time as
4590 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4591 * may want it later.
4593 ret = ocfs2_read_extent_block(et->et_ci,
4594 ocfs2_et_get_last_eb_blk(et),
4600 eb = (struct ocfs2_extent_block *) bh->b_data;
4605 * Unless we have a contiguous insert, we'll need to know if
4606 * there is room left in our allocation tree for another
4609 * XXX: This test is simplistic, we can search for empty
4610 * extent records too.
4612 *free_records = le16_to_cpu(el->l_count) -
4613 le16_to_cpu(el->l_next_free_rec);
4615 if (!insert->ins_tree_depth) {
4616 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4617 ocfs2_figure_appending_type(insert, el, insert_rec);
4621 path = ocfs2_new_path_from_et(et);
4629 * In the case that we're inserting past what the tree
4630 * currently accounts for, ocfs2_find_path() will return for
4631 * us the rightmost tree path. This is accounted for below in
4632 * the appending code.
4634 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4640 el = path_leaf_el(path);
4643 * Now that we have the path, there's two things we want to determine:
4644 * 1) Contiguousness (also set contig_index if this is so)
4646 * 2) Are we doing an append? We can trivially break this up
4647 * into two types of appends: simple record append, or a
4648 * rotate inside the tail leaf.
4650 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4653 * The insert code isn't quite ready to deal with all cases of
4654 * left contiguousness. Specifically, if it's an insert into
4655 * the 1st record in a leaf, it will require the adjustment of
4656 * cluster count on the last record of the path directly to it's
4657 * left. For now, just catch that case and fool the layers
4658 * above us. This works just fine for tree_depth == 0, which
4659 * is why we allow that above.
4661 if (insert->ins_contig == CONTIG_LEFT &&
4662 insert->ins_contig_index == 0)
4663 insert->ins_contig = CONTIG_NONE;
4666 * Ok, so we can simply compare against last_eb to figure out
4667 * whether the path doesn't exist. This will only happen in
4668 * the case that we're doing a tail append, so maybe we can
4669 * take advantage of that information somehow.
4671 if (ocfs2_et_get_last_eb_blk(et) ==
4672 path_leaf_bh(path)->b_blocknr) {
4674 * Ok, ocfs2_find_path() returned us the rightmost
4675 * tree path. This might be an appending insert. There are
4677 * 1) We're doing a true append at the tail:
4678 * -This might even be off the end of the leaf
4679 * 2) We're "appending" by rotating in the tail
4681 ocfs2_figure_appending_type(insert, el, insert_rec);
4685 ocfs2_free_path(path);
4695 * Insert an extent into a btree.
4697 * The caller needs to update the owning btree's cluster count.
4699 int ocfs2_insert_extent(handle_t *handle,
4700 struct ocfs2_extent_tree *et,
4705 struct ocfs2_alloc_context *meta_ac)
4708 int uninitialized_var(free_records);
4709 struct buffer_head *last_eb_bh = NULL;
4710 struct ocfs2_insert_type insert = {0, };
4711 struct ocfs2_extent_rec rec;
4713 trace_ocfs2_insert_extent_start(
4714 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4715 cpos, new_clusters);
4717 memset(&rec, 0, sizeof(rec));
4718 rec.e_cpos = cpu_to_le32(cpos);
4719 rec.e_blkno = cpu_to_le64(start_blk);
4720 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4721 rec.e_flags = flags;
4722 status = ocfs2_et_insert_check(et, &rec);
4728 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4729 &free_records, &insert);
4735 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4736 insert.ins_contig_index, free_records,
4737 insert.ins_tree_depth);
4739 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4740 status = ocfs2_grow_tree(handle, et,
4741 &insert.ins_tree_depth, &last_eb_bh,
4749 /* Finally, we can add clusters. This might rotate the tree for us. */
4750 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4754 ocfs2_et_extent_map_insert(et, &rec);
4763 * Allcate and add clusters into the extent b-tree.
4764 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4765 * The extent b-tree's root is specified by et, and
4766 * it is not limited to the file storage. Any extent tree can use this
4767 * function if it implements the proper ocfs2_extent_tree.
4769 int ocfs2_add_clusters_in_btree(handle_t *handle,
4770 struct ocfs2_extent_tree *et,
4771 u32 *logical_offset,
4772 u32 clusters_to_add,
4774 struct ocfs2_alloc_context *data_ac,
4775 struct ocfs2_alloc_context *meta_ac,
4776 enum ocfs2_alloc_restarted *reason_ret)
4778 int status = 0, err = 0;
4781 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4782 u32 bit_off, num_bits;
4785 struct ocfs2_super *osb =
4786 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4788 BUG_ON(!clusters_to_add);
4791 flags = OCFS2_EXT_UNWRITTEN;
4793 free_extents = ocfs2_num_free_extents(osb, et);
4794 if (free_extents < 0) {
4795 status = free_extents;
4800 /* there are two cases which could cause us to EAGAIN in the
4801 * we-need-more-metadata case:
4802 * 1) we haven't reserved *any*
4803 * 2) we are so fragmented, we've needed to add metadata too
4805 if (!free_extents && !meta_ac) {
4808 reason = RESTART_META;
4810 } else if ((!free_extents)
4811 && (ocfs2_alloc_context_bits_left(meta_ac)
4812 < ocfs2_extend_meta_needed(et->et_root_el))) {
4815 reason = RESTART_META;
4819 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4820 clusters_to_add, &bit_off, &num_bits);
4822 if (status != -ENOSPC)
4827 BUG_ON(num_bits > clusters_to_add);
4829 /* reserve our write early -- insert_extent may update the tree root */
4830 status = ocfs2_et_root_journal_access(handle, et,
4831 OCFS2_JOURNAL_ACCESS_WRITE);
4838 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4839 trace_ocfs2_add_clusters_in_btree(
4840 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4842 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4843 num_bits, flags, meta_ac);
4850 ocfs2_journal_dirty(handle, et->et_root_bh);
4852 clusters_to_add -= num_bits;
4853 *logical_offset += num_bits;
4855 if (clusters_to_add) {
4856 err = clusters_to_add;
4858 reason = RESTART_TRANS;
4863 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4864 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4867 ocfs2_free_clusters(handle,
4870 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4876 *reason_ret = reason;
4877 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4881 static void ocfs2_make_right_split_rec(struct super_block *sb,
4882 struct ocfs2_extent_rec *split_rec,
4884 struct ocfs2_extent_rec *rec)
4886 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4887 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4889 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4891 split_rec->e_cpos = cpu_to_le32(cpos);
4892 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4894 split_rec->e_blkno = rec->e_blkno;
4895 le64_add_cpu(&split_rec->e_blkno,
4896 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4898 split_rec->e_flags = rec->e_flags;
4901 static int ocfs2_split_and_insert(handle_t *handle,
4902 struct ocfs2_extent_tree *et,
4903 struct ocfs2_path *path,
4904 struct buffer_head **last_eb_bh,
4906 struct ocfs2_extent_rec *orig_split_rec,
4907 struct ocfs2_alloc_context *meta_ac)
4910 unsigned int insert_range, rec_range, do_leftright = 0;
4911 struct ocfs2_extent_rec tmprec;
4912 struct ocfs2_extent_list *rightmost_el;
4913 struct ocfs2_extent_rec rec;
4914 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4915 struct ocfs2_insert_type insert;
4916 struct ocfs2_extent_block *eb;
4920 * Store a copy of the record on the stack - it might move
4921 * around as the tree is manipulated below.
4923 rec = path_leaf_el(path)->l_recs[split_index];
4925 rightmost_el = et->et_root_el;
4927 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4929 BUG_ON(!(*last_eb_bh));
4930 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4931 rightmost_el = &eb->h_list;
4934 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4935 le16_to_cpu(rightmost_el->l_count)) {
4936 ret = ocfs2_grow_tree(handle, et,
4937 &depth, last_eb_bh, meta_ac);
4944 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4945 insert.ins_appending = APPEND_NONE;
4946 insert.ins_contig = CONTIG_NONE;
4947 insert.ins_tree_depth = depth;
4949 insert_range = le32_to_cpu(split_rec.e_cpos) +
4950 le16_to_cpu(split_rec.e_leaf_clusters);
4951 rec_range = le32_to_cpu(rec.e_cpos) +
4952 le16_to_cpu(rec.e_leaf_clusters);
4954 if (split_rec.e_cpos == rec.e_cpos) {
4955 insert.ins_split = SPLIT_LEFT;
4956 } else if (insert_range == rec_range) {
4957 insert.ins_split = SPLIT_RIGHT;
4960 * Left/right split. We fake this as a right split
4961 * first and then make a second pass as a left split.
4963 insert.ins_split = SPLIT_RIGHT;
4965 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4966 &tmprec, insert_range, &rec);
4970 BUG_ON(do_leftright);
4974 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4980 if (do_leftright == 1) {
4982 struct ocfs2_extent_list *el;
4985 split_rec = *orig_split_rec;
4987 ocfs2_reinit_path(path, 1);
4989 cpos = le32_to_cpu(split_rec.e_cpos);
4990 ret = ocfs2_find_path(et->et_ci, path, cpos);
4996 el = path_leaf_el(path);
4997 split_index = ocfs2_search_extent_list(el, cpos);
4998 if (split_index == -1) {
4999 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5000 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5001 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5013 static int ocfs2_replace_extent_rec(handle_t *handle,
5014 struct ocfs2_extent_tree *et,
5015 struct ocfs2_path *path,
5016 struct ocfs2_extent_list *el,
5018 struct ocfs2_extent_rec *split_rec)
5022 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5023 path_num_items(path) - 1);
5029 el->l_recs[split_index] = *split_rec;
5031 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5037 * Split part or all of the extent record at split_index in the leaf
5038 * pointed to by path. Merge with the contiguous extent record if needed.
5040 * Care is taken to handle contiguousness so as to not grow the tree.
5042 * meta_ac is not strictly necessary - we only truly need it if growth
5043 * of the tree is required. All other cases will degrade into a less
5044 * optimal tree layout.
5046 * last_eb_bh should be the rightmost leaf block for any extent
5047 * btree. Since a split may grow the tree or a merge might shrink it,
5048 * the caller cannot trust the contents of that buffer after this call.
5050 * This code is optimized for readability - several passes might be
5051 * made over certain portions of the tree. All of those blocks will
5052 * have been brought into cache (and pinned via the journal), so the
5053 * extra overhead is not expressed in terms of disk reads.
5055 int ocfs2_split_extent(handle_t *handle,
5056 struct ocfs2_extent_tree *et,
5057 struct ocfs2_path *path,
5059 struct ocfs2_extent_rec *split_rec,
5060 struct ocfs2_alloc_context *meta_ac,
5061 struct ocfs2_cached_dealloc_ctxt *dealloc)
5064 struct ocfs2_extent_list *el = path_leaf_el(path);
5065 struct buffer_head *last_eb_bh = NULL;
5066 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5067 struct ocfs2_merge_ctxt ctxt;
5068 struct ocfs2_extent_list *rightmost_el;
5070 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5071 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5072 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5078 ret = ocfs2_figure_merge_contig_type(et, path, el,
5088 * The core merge / split code wants to know how much room is
5089 * left in this allocation tree, so we pass the
5090 * rightmost extent list.
5092 if (path->p_tree_depth) {
5093 struct ocfs2_extent_block *eb;
5095 ret = ocfs2_read_extent_block(et->et_ci,
5096 ocfs2_et_get_last_eb_blk(et),
5103 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5104 rightmost_el = &eb->h_list;
5106 rightmost_el = path_root_el(path);
5108 if (rec->e_cpos == split_rec->e_cpos &&
5109 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5110 ctxt.c_split_covers_rec = 1;
5112 ctxt.c_split_covers_rec = 0;
5114 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5116 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5117 ctxt.c_has_empty_extent,
5118 ctxt.c_split_covers_rec);
5120 if (ctxt.c_contig_type == CONTIG_NONE) {
5121 if (ctxt.c_split_covers_rec)
5122 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5123 split_index, split_rec);
5125 ret = ocfs2_split_and_insert(handle, et, path,
5126 &last_eb_bh, split_index,
5127 split_rec, meta_ac);
5131 ret = ocfs2_try_to_merge_extent(handle, et, path,
5132 split_index, split_rec,
5144 * Change the flags of the already-existing extent at cpos for len clusters.
5146 * new_flags: the flags we want to set.
5147 * clear_flags: the flags we want to clear.
5148 * phys: the new physical offset we want this new extent starts from.
5150 * If the existing extent is larger than the request, initiate a
5151 * split. An attempt will be made at merging with adjacent extents.
5153 * The caller is responsible for passing down meta_ac if we'll need it.
5155 int ocfs2_change_extent_flag(handle_t *handle,
5156 struct ocfs2_extent_tree *et,
5157 u32 cpos, u32 len, u32 phys,
5158 struct ocfs2_alloc_context *meta_ac,
5159 struct ocfs2_cached_dealloc_ctxt *dealloc,
5160 int new_flags, int clear_flags)
5163 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5164 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5165 struct ocfs2_extent_rec split_rec;
5166 struct ocfs2_path *left_path = NULL;
5167 struct ocfs2_extent_list *el;
5168 struct ocfs2_extent_rec *rec;
5170 left_path = ocfs2_new_path_from_et(et);
5177 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5182 el = path_leaf_el(left_path);
5184 index = ocfs2_search_extent_list(el, cpos);
5187 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5188 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5195 rec = &el->l_recs[index];
5196 if (new_flags && (rec->e_flags & new_flags)) {
5197 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5198 "extent that already had them\n",
5199 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5204 if (clear_flags && !(rec->e_flags & clear_flags)) {
5205 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5206 "extent that didn't have them\n",
5207 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5212 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5213 split_rec.e_cpos = cpu_to_le32(cpos);
5214 split_rec.e_leaf_clusters = cpu_to_le16(len);
5215 split_rec.e_blkno = cpu_to_le64(start_blkno);
5216 split_rec.e_flags = rec->e_flags;
5218 split_rec.e_flags |= new_flags;
5220 split_rec.e_flags &= ~clear_flags;
5222 ret = ocfs2_split_extent(handle, et, left_path,
5223 index, &split_rec, meta_ac,
5229 ocfs2_free_path(left_path);
5235 * Mark the already-existing extent at cpos as written for len clusters.
5236 * This removes the unwritten extent flag.
5238 * If the existing extent is larger than the request, initiate a
5239 * split. An attempt will be made at merging with adjacent extents.
5241 * The caller is responsible for passing down meta_ac if we'll need it.
5243 int ocfs2_mark_extent_written(struct inode *inode,
5244 struct ocfs2_extent_tree *et,
5245 handle_t *handle, u32 cpos, u32 len, u32 phys,
5246 struct ocfs2_alloc_context *meta_ac,
5247 struct ocfs2_cached_dealloc_ctxt *dealloc)
5251 trace_ocfs2_mark_extent_written(
5252 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5255 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5256 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents that are being written to, but the feature bit is not set in the super block\n",
5257 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5263 * XXX: This should be fixed up so that we just re-insert the
5264 * next extent records.
5266 ocfs2_et_extent_map_truncate(et, 0);
5268 ret = ocfs2_change_extent_flag(handle, et, cpos,
5269 len, phys, meta_ac, dealloc,
5270 0, OCFS2_EXT_UNWRITTEN);
5278 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5279 struct ocfs2_path *path,
5280 int index, u32 new_range,
5281 struct ocfs2_alloc_context *meta_ac)
5283 int ret, depth, credits;
5284 struct buffer_head *last_eb_bh = NULL;
5285 struct ocfs2_extent_block *eb;
5286 struct ocfs2_extent_list *rightmost_el, *el;
5287 struct ocfs2_extent_rec split_rec;
5288 struct ocfs2_extent_rec *rec;
5289 struct ocfs2_insert_type insert;
5292 * Setup the record to split before we grow the tree.
5294 el = path_leaf_el(path);
5295 rec = &el->l_recs[index];
5296 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5297 &split_rec, new_range, rec);
5299 depth = path->p_tree_depth;
5301 ret = ocfs2_read_extent_block(et->et_ci,
5302 ocfs2_et_get_last_eb_blk(et),
5309 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5310 rightmost_el = &eb->h_list;
5312 rightmost_el = path_leaf_el(path);
5314 credits = path->p_tree_depth +
5315 ocfs2_extend_meta_needed(et->et_root_el);
5316 ret = ocfs2_extend_trans(handle, credits);
5322 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5323 le16_to_cpu(rightmost_el->l_count)) {
5324 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5332 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5333 insert.ins_appending = APPEND_NONE;
5334 insert.ins_contig = CONTIG_NONE;
5335 insert.ins_split = SPLIT_RIGHT;
5336 insert.ins_tree_depth = depth;
5338 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5347 static int ocfs2_truncate_rec(handle_t *handle,
5348 struct ocfs2_extent_tree *et,
5349 struct ocfs2_path *path, int index,
5350 struct ocfs2_cached_dealloc_ctxt *dealloc,
5354 u32 left_cpos, rec_range, trunc_range;
5355 int is_rightmost_tree_rec = 0;
5356 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5357 struct ocfs2_path *left_path = NULL;
5358 struct ocfs2_extent_list *el = path_leaf_el(path);
5359 struct ocfs2_extent_rec *rec;
5360 struct ocfs2_extent_block *eb;
5362 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5363 /* extend credit for ocfs2_remove_rightmost_path */
5364 ret = ocfs2_extend_rotate_transaction(handle, 0,
5365 handle->h_buffer_credits,
5372 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5381 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5382 path->p_tree_depth) {
5384 * Check whether this is the rightmost tree record. If
5385 * we remove all of this record or part of its right
5386 * edge then an update of the record lengths above it
5389 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5390 if (eb->h_next_leaf_blk == 0)
5391 is_rightmost_tree_rec = 1;
5394 rec = &el->l_recs[index];
5395 if (index == 0 && path->p_tree_depth &&
5396 le32_to_cpu(rec->e_cpos) == cpos) {
5398 * Changing the leftmost offset (via partial or whole
5399 * record truncate) of an interior (or rightmost) path
5400 * means we have to update the subtree that is formed
5401 * by this leaf and the one to it's left.
5403 * There are two cases we can skip:
5404 * 1) Path is the leftmost one in our btree.
5405 * 2) The leaf is rightmost and will be empty after
5406 * we remove the extent record - the rotate code
5407 * knows how to update the newly formed edge.
5410 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5416 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5417 left_path = ocfs2_new_path_from_path(path);
5424 ret = ocfs2_find_path(et->et_ci, left_path,
5433 ret = ocfs2_extend_rotate_transaction(handle, 0,
5434 handle->h_buffer_credits,
5441 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5447 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5453 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5454 trunc_range = cpos + len;
5456 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5459 memset(rec, 0, sizeof(*rec));
5460 ocfs2_cleanup_merge(el, index);
5462 next_free = le16_to_cpu(el->l_next_free_rec);
5463 if (is_rightmost_tree_rec && next_free > 1) {
5465 * We skip the edge update if this path will
5466 * be deleted by the rotate code.
5468 rec = &el->l_recs[next_free - 1];
5469 ocfs2_adjust_rightmost_records(handle, et, path,
5472 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5473 /* Remove leftmost portion of the record. */
5474 le32_add_cpu(&rec->e_cpos, len);
5475 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5476 le16_add_cpu(&rec->e_leaf_clusters, -len);
5477 } else if (rec_range == trunc_range) {
5478 /* Remove rightmost portion of the record */
5479 le16_add_cpu(&rec->e_leaf_clusters, -len);
5480 if (is_rightmost_tree_rec)
5481 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5483 /* Caller should have trapped this. */
5484 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5486 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5487 le32_to_cpu(rec->e_cpos),
5488 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5495 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5496 ocfs2_complete_edge_insert(handle, left_path, path,
5500 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5502 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5509 ocfs2_free_path(left_path);
5513 int ocfs2_remove_extent(handle_t *handle,
5514 struct ocfs2_extent_tree *et,
5516 struct ocfs2_alloc_context *meta_ac,
5517 struct ocfs2_cached_dealloc_ctxt *dealloc)
5520 u32 rec_range, trunc_range;
5521 struct ocfs2_extent_rec *rec;
5522 struct ocfs2_extent_list *el;
5523 struct ocfs2_path *path = NULL;
5526 * XXX: Why are we truncating to 0 instead of wherever this
5529 ocfs2_et_extent_map_truncate(et, 0);
5531 path = ocfs2_new_path_from_et(et);
5538 ret = ocfs2_find_path(et->et_ci, path, cpos);
5544 el = path_leaf_el(path);
5545 index = ocfs2_search_extent_list(el, cpos);
5547 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5548 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5549 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5556 * We have 3 cases of extent removal:
5557 * 1) Range covers the entire extent rec
5558 * 2) Range begins or ends on one edge of the extent rec
5559 * 3) Range is in the middle of the extent rec (no shared edges)
5561 * For case 1 we remove the extent rec and left rotate to
5564 * For case 2 we just shrink the existing extent rec, with a
5565 * tree update if the shrinking edge is also the edge of an
5568 * For case 3 we do a right split to turn the extent rec into
5569 * something case 2 can handle.
5571 rec = &el->l_recs[index];
5572 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5573 trunc_range = cpos + len;
5575 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5577 trace_ocfs2_remove_extent(
5578 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5579 cpos, len, index, le32_to_cpu(rec->e_cpos),
5580 ocfs2_rec_clusters(el, rec));
5582 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5583 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5590 ret = ocfs2_split_tree(handle, et, path, index,
5591 trunc_range, meta_ac);
5598 * The split could have manipulated the tree enough to
5599 * move the record location, so we have to look for it again.
5601 ocfs2_reinit_path(path, 1);
5603 ret = ocfs2_find_path(et->et_ci, path, cpos);
5609 el = path_leaf_el(path);
5610 index = ocfs2_search_extent_list(el, cpos);
5612 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5613 "Owner %llu: split at cpos %u lost record\n",
5614 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5621 * Double check our values here. If anything is fishy,
5622 * it's easier to catch it at the top level.
5624 rec = &el->l_recs[index];
5625 rec_range = le32_to_cpu(rec->e_cpos) +
5626 ocfs2_rec_clusters(el, rec);
5627 if (rec_range != trunc_range) {
5628 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5629 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5630 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5631 cpos, len, le32_to_cpu(rec->e_cpos),
5632 ocfs2_rec_clusters(el, rec));
5637 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5646 ocfs2_free_path(path);
5651 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5652 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5653 * number to reserve some extra blocks, and it only handles meta
5656 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5657 * and punching holes.
5659 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5660 struct ocfs2_extent_tree *et,
5661 u32 extents_to_split,
5662 struct ocfs2_alloc_context **ac,
5665 int ret = 0, num_free_extents;
5666 unsigned int max_recs_needed = 2 * extents_to_split;
5667 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5671 num_free_extents = ocfs2_num_free_extents(osb, et);
5672 if (num_free_extents < 0) {
5673 ret = num_free_extents;
5678 if (!num_free_extents ||
5679 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5680 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5683 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5694 ocfs2_free_alloc_context(*ac);
5702 int ocfs2_remove_btree_range(struct inode *inode,
5703 struct ocfs2_extent_tree *et,
5704 u32 cpos, u32 phys_cpos, u32 len, int flags,
5705 struct ocfs2_cached_dealloc_ctxt *dealloc,
5706 u64 refcount_loc, bool refcount_tree_locked)
5708 int ret, credits = 0, extra_blocks = 0;
5709 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5710 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5711 struct inode *tl_inode = osb->osb_tl_inode;
5713 struct ocfs2_alloc_context *meta_ac = NULL;
5714 struct ocfs2_refcount_tree *ref_tree = NULL;
5716 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5717 BUG_ON(!ocfs2_is_refcount_inode(inode));
5719 if (!refcount_tree_locked) {
5720 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5728 ret = ocfs2_prepare_refcount_change_for_del(inode,
5740 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5747 inode_lock(tl_inode);
5749 if (ocfs2_truncate_log_needs_flush(osb)) {
5750 ret = __ocfs2_flush_truncate_log(osb);
5757 handle = ocfs2_start_trans(osb,
5758 ocfs2_remove_extent_credits(osb->sb) + credits);
5759 if (IS_ERR(handle)) {
5760 ret = PTR_ERR(handle);
5765 ret = ocfs2_et_root_journal_access(handle, et,
5766 OCFS2_JOURNAL_ACCESS_WRITE);
5772 dquot_free_space_nodirty(inode,
5773 ocfs2_clusters_to_bytes(inode->i_sb, len));
5775 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5781 ocfs2_et_update_clusters(et, -len);
5782 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5784 ocfs2_journal_dirty(handle, et->et_root_bh);
5787 if (flags & OCFS2_EXT_REFCOUNTED)
5788 ret = ocfs2_decrease_refcount(inode, handle,
5789 ocfs2_blocks_to_clusters(osb->sb,
5794 ret = ocfs2_truncate_log_append(osb, handle,
5802 ocfs2_commit_trans(osb, handle);
5804 inode_unlock(tl_inode);
5807 ocfs2_free_alloc_context(meta_ac);
5810 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5815 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5817 struct buffer_head *tl_bh = osb->osb_tl_bh;
5818 struct ocfs2_dinode *di;
5819 struct ocfs2_truncate_log *tl;
5821 di = (struct ocfs2_dinode *) tl_bh->b_data;
5822 tl = &di->id2.i_dealloc;
5824 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5825 "slot %d, invalid truncate log parameters: used = "
5826 "%u, count = %u\n", osb->slot_num,
5827 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5828 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5831 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5832 unsigned int new_start)
5834 unsigned int tail_index;
5835 unsigned int current_tail;
5837 /* No records, nothing to coalesce */
5838 if (!le16_to_cpu(tl->tl_used))
5841 tail_index = le16_to_cpu(tl->tl_used) - 1;
5842 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5843 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5845 return current_tail == new_start;
5848 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5851 unsigned int num_clusters)
5854 unsigned int start_cluster, tl_count;
5855 struct inode *tl_inode = osb->osb_tl_inode;
5856 struct buffer_head *tl_bh = osb->osb_tl_bh;
5857 struct ocfs2_dinode *di;
5858 struct ocfs2_truncate_log *tl;
5860 BUG_ON(inode_trylock(tl_inode));
5862 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5864 di = (struct ocfs2_dinode *) tl_bh->b_data;
5866 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5867 * by the underlying call to ocfs2_read_inode_block(), so any
5868 * corruption is a code bug */
5869 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5871 tl = &di->id2.i_dealloc;
5872 tl_count = le16_to_cpu(tl->tl_count);
5873 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5875 "Truncate record count on #%llu invalid "
5876 "wanted %u, actual %u\n",
5877 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5878 ocfs2_truncate_recs_per_inode(osb->sb),
5879 le16_to_cpu(tl->tl_count));
5881 /* Caller should have known to flush before calling us. */
5882 index = le16_to_cpu(tl->tl_used);
5883 if (index >= tl_count) {
5889 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5890 OCFS2_JOURNAL_ACCESS_WRITE);
5896 trace_ocfs2_truncate_log_append(
5897 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5898 start_cluster, num_clusters);
5899 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5901 * Move index back to the record we are coalescing with.
5902 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5906 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5907 trace_ocfs2_truncate_log_append(
5908 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5909 index, le32_to_cpu(tl->tl_recs[index].t_start),
5912 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5913 tl->tl_used = cpu_to_le16(index + 1);
5915 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5917 ocfs2_journal_dirty(handle, tl_bh);
5919 osb->truncated_clusters += num_clusters;
5924 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5925 struct inode *data_alloc_inode,
5926 struct buffer_head *data_alloc_bh)
5930 unsigned int num_clusters;
5932 struct ocfs2_truncate_rec rec;
5933 struct ocfs2_dinode *di;
5934 struct ocfs2_truncate_log *tl;
5935 struct inode *tl_inode = osb->osb_tl_inode;
5936 struct buffer_head *tl_bh = osb->osb_tl_bh;
5939 di = (struct ocfs2_dinode *) tl_bh->b_data;
5940 tl = &di->id2.i_dealloc;
5941 i = le16_to_cpu(tl->tl_used) - 1;
5943 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5944 if (IS_ERR(handle)) {
5945 status = PTR_ERR(handle);
5950 /* Caller has given us at least enough credits to
5951 * update the truncate log dinode */
5952 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5953 OCFS2_JOURNAL_ACCESS_WRITE);
5959 tl->tl_used = cpu_to_le16(i);
5961 ocfs2_journal_dirty(handle, tl_bh);
5963 rec = tl->tl_recs[i];
5964 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5965 le32_to_cpu(rec.t_start));
5966 num_clusters = le32_to_cpu(rec.t_clusters);
5968 /* if start_blk is not set, we ignore the record as
5971 trace_ocfs2_replay_truncate_records(
5972 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5973 i, le32_to_cpu(rec.t_start), num_clusters);
5975 status = ocfs2_free_clusters(handle, data_alloc_inode,
5976 data_alloc_bh, start_blk,
5984 ocfs2_commit_trans(osb, handle);
5988 osb->truncated_clusters = 0;
5994 /* Expects you to already be holding tl_inode->i_mutex */
5995 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5998 unsigned int num_to_flush;
5999 struct inode *tl_inode = osb->osb_tl_inode;
6000 struct inode *data_alloc_inode = NULL;
6001 struct buffer_head *tl_bh = osb->osb_tl_bh;
6002 struct buffer_head *data_alloc_bh = NULL;
6003 struct ocfs2_dinode *di;
6004 struct ocfs2_truncate_log *tl;
6006 BUG_ON(inode_trylock(tl_inode));
6008 di = (struct ocfs2_dinode *) tl_bh->b_data;
6010 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
6011 * by the underlying call to ocfs2_read_inode_block(), so any
6012 * corruption is a code bug */
6013 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6015 tl = &di->id2.i_dealloc;
6016 num_to_flush = le16_to_cpu(tl->tl_used);
6017 trace_ocfs2_flush_truncate_log(
6018 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6020 if (!num_to_flush) {
6025 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6026 GLOBAL_BITMAP_SYSTEM_INODE,
6027 OCFS2_INVALID_SLOT);
6028 if (!data_alloc_inode) {
6030 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6034 inode_lock(data_alloc_inode);
6036 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6042 status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6047 brelse(data_alloc_bh);
6048 ocfs2_inode_unlock(data_alloc_inode, 1);
6051 inode_unlock(data_alloc_inode);
6052 iput(data_alloc_inode);
6058 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6061 struct inode *tl_inode = osb->osb_tl_inode;
6063 inode_lock(tl_inode);
6064 status = __ocfs2_flush_truncate_log(osb);
6065 inode_unlock(tl_inode);
6070 static void ocfs2_truncate_log_worker(struct work_struct *work)
6073 struct ocfs2_super *osb =
6074 container_of(work, struct ocfs2_super,
6075 osb_truncate_log_wq.work);
6077 status = ocfs2_flush_truncate_log(osb);
6081 ocfs2_init_steal_slots(osb);
6084 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6085 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6088 if (osb->osb_tl_inode &&
6089 atomic_read(&osb->osb_tl_disable) == 0) {
6090 /* We want to push off log flushes while truncates are
6093 cancel_delayed_work(&osb->osb_truncate_log_wq);
6095 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6096 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6101 * Try to flush truncate logs if we can free enough clusters from it.
6102 * As for return value, "< 0" means error, "0" no space and "1" means
6103 * we have freed enough spaces and let the caller try to allocate again.
6105 int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6106 unsigned int needed)
6110 unsigned int truncated_clusters;
6112 inode_lock(osb->osb_tl_inode);
6113 truncated_clusters = osb->truncated_clusters;
6114 inode_unlock(osb->osb_tl_inode);
6117 * Check whether we can succeed in allocating if we free
6120 if (truncated_clusters < needed)
6123 ret = ocfs2_flush_truncate_log(osb);
6129 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6130 jbd2_log_wait_commit(osb->journal->j_journal, target);
6137 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6139 struct inode **tl_inode,
6140 struct buffer_head **tl_bh)
6143 struct inode *inode = NULL;
6144 struct buffer_head *bh = NULL;
6146 inode = ocfs2_get_system_file_inode(osb,
6147 TRUNCATE_LOG_SYSTEM_INODE,
6151 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6155 status = ocfs2_read_inode_block(inode, &bh);
6168 /* called during the 1st stage of node recovery. we stamp a clean
6169 * truncate log and pass back a copy for processing later. if the
6170 * truncate log does not require processing, a *tl_copy is set to
6172 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6174 struct ocfs2_dinode **tl_copy)
6177 struct inode *tl_inode = NULL;
6178 struct buffer_head *tl_bh = NULL;
6179 struct ocfs2_dinode *di;
6180 struct ocfs2_truncate_log *tl;
6184 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6186 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6192 di = (struct ocfs2_dinode *) tl_bh->b_data;
6194 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6195 * validated by the underlying call to ocfs2_read_inode_block(),
6196 * so any corruption is a code bug */
6197 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6199 tl = &di->id2.i_dealloc;
6200 if (le16_to_cpu(tl->tl_used)) {
6201 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6203 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6210 /* Assuming the write-out below goes well, this copy
6211 * will be passed back to recovery for processing. */
6212 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6214 /* All we need to do to clear the truncate log is set
6218 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6219 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6239 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6240 struct ocfs2_dinode *tl_copy)
6244 unsigned int clusters, num_recs, start_cluster;
6247 struct inode *tl_inode = osb->osb_tl_inode;
6248 struct ocfs2_truncate_log *tl;
6250 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6251 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6255 tl = &tl_copy->id2.i_dealloc;
6256 num_recs = le16_to_cpu(tl->tl_used);
6257 trace_ocfs2_complete_truncate_log_recovery(
6258 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6261 inode_lock(tl_inode);
6262 for(i = 0; i < num_recs; i++) {
6263 if (ocfs2_truncate_log_needs_flush(osb)) {
6264 status = __ocfs2_flush_truncate_log(osb);
6271 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6272 if (IS_ERR(handle)) {
6273 status = PTR_ERR(handle);
6278 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6279 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6280 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6282 status = ocfs2_truncate_log_append(osb, handle,
6283 start_blk, clusters);
6284 ocfs2_commit_trans(osb, handle);
6292 inode_unlock(tl_inode);
6297 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6300 struct inode *tl_inode = osb->osb_tl_inode;
6302 atomic_set(&osb->osb_tl_disable, 1);
6305 cancel_delayed_work(&osb->osb_truncate_log_wq);
6306 flush_workqueue(osb->ocfs2_wq);
6308 status = ocfs2_flush_truncate_log(osb);
6312 brelse(osb->osb_tl_bh);
6313 iput(osb->osb_tl_inode);
6317 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6320 struct inode *tl_inode = NULL;
6321 struct buffer_head *tl_bh = NULL;
6323 status = ocfs2_get_truncate_log_info(osb,
6330 /* ocfs2_truncate_log_shutdown keys on the existence of
6331 * osb->osb_tl_inode so we don't set any of the osb variables
6332 * until we're sure all is well. */
6333 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6334 ocfs2_truncate_log_worker);
6335 atomic_set(&osb->osb_tl_disable, 0);
6336 osb->osb_tl_bh = tl_bh;
6337 osb->osb_tl_inode = tl_inode;
6343 * Delayed de-allocation of suballocator blocks.
6345 * Some sets of block de-allocations might involve multiple suballocator inodes.
6347 * The locking for this can get extremely complicated, especially when
6348 * the suballocator inodes to delete from aren't known until deep
6349 * within an unrelated codepath.
6351 * ocfs2_extent_block structures are a good example of this - an inode
6352 * btree could have been grown by any number of nodes each allocating
6353 * out of their own suballoc inode.
6355 * These structures allow the delay of block de-allocation until a
6356 * later time, when locking of multiple cluster inodes won't cause
6361 * Describe a single bit freed from a suballocator. For the block
6362 * suballocators, it represents one block. For the global cluster
6363 * allocator, it represents some clusters and free_bit indicates
6366 struct ocfs2_cached_block_free {
6367 struct ocfs2_cached_block_free *free_next;
6370 unsigned int free_bit;
6373 struct ocfs2_per_slot_free_list {
6374 struct ocfs2_per_slot_free_list *f_next_suballocator;
6377 struct ocfs2_cached_block_free *f_first;
6380 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6383 struct ocfs2_cached_block_free *head)
6388 struct inode *inode;
6389 struct buffer_head *di_bh = NULL;
6390 struct ocfs2_cached_block_free *tmp;
6392 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6401 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6409 bg_blkno = head->free_bg;
6411 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6413 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6414 if (IS_ERR(handle)) {
6415 ret = PTR_ERR(handle);
6420 trace_ocfs2_free_cached_blocks(
6421 (unsigned long long)head->free_blk, head->free_bit);
6423 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6424 head->free_bit, bg_blkno, 1);
6428 ocfs2_commit_trans(osb, handle);
6431 head = head->free_next;
6436 ocfs2_inode_unlock(inode, 1);
6439 inode_unlock(inode);
6443 /* Premature exit may have left some dangling items. */
6445 head = head->free_next;
6452 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6453 u64 blkno, unsigned int bit)
6456 struct ocfs2_cached_block_free *item;
6458 item = kzalloc(sizeof(*item), GFP_NOFS);
6465 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6467 item->free_blk = blkno;
6468 item->free_bit = bit;
6469 item->free_next = ctxt->c_global_allocator;
6471 ctxt->c_global_allocator = item;
6475 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6476 struct ocfs2_cached_block_free *head)
6478 struct ocfs2_cached_block_free *tmp;
6479 struct inode *tl_inode = osb->osb_tl_inode;
6483 inode_lock(tl_inode);
6486 if (ocfs2_truncate_log_needs_flush(osb)) {
6487 ret = __ocfs2_flush_truncate_log(osb);
6494 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6495 if (IS_ERR(handle)) {
6496 ret = PTR_ERR(handle);
6501 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6504 ocfs2_commit_trans(osb, handle);
6506 head = head->free_next;
6515 inode_unlock(tl_inode);
6518 /* Premature exit may have left some dangling items. */
6520 head = head->free_next;
6527 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6528 struct ocfs2_cached_dealloc_ctxt *ctxt)
6531 struct ocfs2_per_slot_free_list *fl;
6536 while (ctxt->c_first_suballocator) {
6537 fl = ctxt->c_first_suballocator;
6540 trace_ocfs2_run_deallocs(fl->f_inode_type,
6542 ret2 = ocfs2_free_cached_blocks(osb,
6552 ctxt->c_first_suballocator = fl->f_next_suballocator;
6556 if (ctxt->c_global_allocator) {
6557 ret2 = ocfs2_free_cached_clusters(osb,
6558 ctxt->c_global_allocator);
6564 ctxt->c_global_allocator = NULL;
6570 static struct ocfs2_per_slot_free_list *
6571 ocfs2_find_per_slot_free_list(int type,
6573 struct ocfs2_cached_dealloc_ctxt *ctxt)
6575 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6578 if (fl->f_inode_type == type && fl->f_slot == slot)
6581 fl = fl->f_next_suballocator;
6584 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6586 fl->f_inode_type = type;
6589 fl->f_next_suballocator = ctxt->c_first_suballocator;
6591 ctxt->c_first_suballocator = fl;
6596 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6597 int type, int slot, u64 suballoc,
6598 u64 blkno, unsigned int bit)
6601 struct ocfs2_per_slot_free_list *fl;
6602 struct ocfs2_cached_block_free *item;
6604 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6611 item = kzalloc(sizeof(*item), GFP_NOFS);
6618 trace_ocfs2_cache_block_dealloc(type, slot,
6619 (unsigned long long)suballoc,
6620 (unsigned long long)blkno, bit);
6622 item->free_bg = suballoc;
6623 item->free_blk = blkno;
6624 item->free_bit = bit;
6625 item->free_next = fl->f_first;
6634 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6635 struct ocfs2_extent_block *eb)
6637 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6638 le16_to_cpu(eb->h_suballoc_slot),
6639 le64_to_cpu(eb->h_suballoc_loc),
6640 le64_to_cpu(eb->h_blkno),
6641 le16_to_cpu(eb->h_suballoc_bit));
6644 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6646 set_buffer_uptodate(bh);
6647 mark_buffer_dirty(bh);
6651 void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6652 unsigned int from, unsigned int to,
6653 struct page *page, int zero, u64 *phys)
6655 int ret, partial = 0;
6657 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6662 zero_user_segment(page, from, to);
6665 * Need to set the buffers we zero'd into uptodate
6666 * here if they aren't - ocfs2_map_page_blocks()
6667 * might've skipped some
6669 ret = walk_page_buffers(handle, page_buffers(page),
6674 else if (ocfs2_should_order_data(inode)) {
6675 ret = ocfs2_jbd2_file_inode(handle, inode);
6681 SetPageUptodate(page);
6683 flush_dcache_page(page);
6686 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6687 loff_t end, struct page **pages,
6688 int numpages, u64 phys, handle_t *handle)
6692 unsigned int from, to = PAGE_SIZE;
6693 struct super_block *sb = inode->i_sb;
6695 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6701 for(i = 0; i < numpages; i++) {
6704 from = start & (PAGE_SIZE - 1);
6705 if ((end >> PAGE_SHIFT) == page->index)
6706 to = end & (PAGE_SIZE - 1);
6708 BUG_ON(from > PAGE_SIZE);
6709 BUG_ON(to > PAGE_SIZE);
6711 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6714 start = (page->index + 1) << PAGE_SHIFT;
6718 ocfs2_unlock_and_free_pages(pages, numpages);
6721 int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6722 struct page **pages, int *num)
6724 int numpages, ret = 0;
6725 struct address_space *mapping = inode->i_mapping;
6726 unsigned long index;
6727 loff_t last_page_bytes;
6729 BUG_ON(start > end);
6732 last_page_bytes = PAGE_ALIGN(end);
6733 index = start >> PAGE_SHIFT;
6735 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6736 if (!pages[numpages]) {
6744 } while (index < (last_page_bytes >> PAGE_SHIFT));
6749 ocfs2_unlock_and_free_pages(pages, numpages);
6758 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6759 struct page **pages, int *num)
6761 struct super_block *sb = inode->i_sb;
6763 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6764 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6766 return ocfs2_grab_pages(inode, start, end, pages, num);
6770 * Zero the area past i_size but still within an allocated
6771 * cluster. This avoids exposing nonzero data on subsequent file
6774 * We need to call this before i_size is updated on the inode because
6775 * otherwise block_write_full_page() will skip writeout of pages past
6776 * i_size. The new_i_size parameter is passed for this reason.
6778 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6779 u64 range_start, u64 range_end)
6781 int ret = 0, numpages;
6782 struct page **pages = NULL;
6784 unsigned int ext_flags;
6785 struct super_block *sb = inode->i_sb;
6788 * File systems which don't support sparse files zero on every
6791 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6794 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6795 sizeof(struct page *), GFP_NOFS);
6796 if (pages == NULL) {
6802 if (range_start == range_end)
6805 ret = ocfs2_extent_map_get_blocks(inode,
6806 range_start >> sb->s_blocksize_bits,
6807 &phys, NULL, &ext_flags);
6814 * Tail is a hole, or is marked unwritten. In either case, we
6815 * can count on read and write to return/push zero's.
6817 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6820 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6827 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6828 numpages, phys, handle);
6831 * Initiate writeout of the pages we zero'd here. We don't
6832 * wait on them - the truncate_inode_pages() call later will
6835 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6846 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6847 struct ocfs2_dinode *di)
6849 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6850 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6852 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6853 memset(&di->id2, 0, blocksize -
6854 offsetof(struct ocfs2_dinode, id2) -
6857 memset(&di->id2, 0, blocksize -
6858 offsetof(struct ocfs2_dinode, id2));
6861 void ocfs2_dinode_new_extent_list(struct inode *inode,
6862 struct ocfs2_dinode *di)
6864 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6865 di->id2.i_list.l_tree_depth = 0;
6866 di->id2.i_list.l_next_free_rec = 0;
6867 di->id2.i_list.l_count = cpu_to_le16(
6868 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6871 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6873 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6874 struct ocfs2_inline_data *idata = &di->id2.i_data;
6876 spin_lock(&oi->ip_lock);
6877 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6878 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6879 spin_unlock(&oi->ip_lock);
6882 * We clear the entire i_data structure here so that all
6883 * fields can be properly initialized.
6885 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6887 idata->id_count = cpu_to_le16(
6888 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6891 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6892 struct buffer_head *di_bh)
6894 int ret, i, has_data, num_pages = 0;
6898 u64 uninitialized_var(block);
6899 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6900 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6901 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6902 struct ocfs2_alloc_context *data_ac = NULL;
6903 struct page **pages = NULL;
6904 loff_t end = osb->s_clustersize;
6905 struct ocfs2_extent_tree et;
6908 has_data = i_size_read(inode) ? 1 : 0;
6911 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6912 sizeof(struct page *), GFP_NOFS);
6913 if (pages == NULL) {
6919 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6926 handle = ocfs2_start_trans(osb,
6927 ocfs2_inline_to_extents_credits(osb->sb));
6928 if (IS_ERR(handle)) {
6929 ret = PTR_ERR(handle);
6934 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
6935 OCFS2_JOURNAL_ACCESS_WRITE);
6942 unsigned int page_end;
6945 ret = dquot_alloc_space_nodirty(inode,
6946 ocfs2_clusters_to_bytes(osb->sb, 1));
6951 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
6953 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
6961 * Save two copies, one for insert, and one that can
6962 * be changed by ocfs2_map_and_dirty_page() below.
6964 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6967 * Non sparse file systems zero on extend, so no need
6970 if (!ocfs2_sparse_alloc(osb) &&
6971 PAGE_SIZE < osb->s_clustersize)
6974 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6982 * This should populate the 1st page for us and mark
6985 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6992 page_end = PAGE_SIZE;
6993 if (PAGE_SIZE > osb->s_clustersize)
6994 page_end = osb->s_clustersize;
6996 for (i = 0; i < num_pages; i++)
6997 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6998 pages[i], i > 0, &phys);
7001 spin_lock(&oi->ip_lock);
7002 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7003 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7004 spin_unlock(&oi->ip_lock);
7006 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7007 ocfs2_dinode_new_extent_list(inode, di);
7009 ocfs2_journal_dirty(handle, di_bh);
7013 * An error at this point should be extremely rare. If
7014 * this proves to be false, we could always re-build
7015 * the in-inode data from our pages.
7017 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7018 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7025 inode->i_blocks = ocfs2_inode_sector_count(inode);
7030 ocfs2_unlock_and_free_pages(pages, num_pages);
7033 if (ret < 0 && did_quota)
7034 dquot_free_space_nodirty(inode,
7035 ocfs2_clusters_to_bytes(osb->sb, 1));
7038 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7039 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7042 ocfs2_free_clusters(handle,
7045 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7049 ocfs2_commit_trans(osb, handle);
7053 ocfs2_free_alloc_context(data_ac);
7060 * It is expected, that by the time you call this function,
7061 * inode->i_size and fe->i_size have been adjusted.
7063 * WARNING: This will kfree the truncate context
7065 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7066 struct inode *inode,
7067 struct buffer_head *di_bh)
7069 int status = 0, i, flags = 0;
7070 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7072 struct ocfs2_extent_list *el;
7073 struct ocfs2_extent_rec *rec;
7074 struct ocfs2_path *path = NULL;
7075 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7076 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7077 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7078 struct ocfs2_extent_tree et;
7079 struct ocfs2_cached_dealloc_ctxt dealloc;
7080 struct ocfs2_refcount_tree *ref_tree = NULL;
7082 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7083 ocfs2_init_dealloc_ctxt(&dealloc);
7085 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7086 i_size_read(inode));
7088 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7089 ocfs2_journal_access_di);
7096 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7100 * Check that we still have allocation to delete.
7102 if (OCFS2_I(inode)->ip_clusters == 0) {
7108 * Truncate always works against the rightmost tree branch.
7110 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7116 trace_ocfs2_commit_truncate(
7117 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7119 OCFS2_I(inode)->ip_clusters,
7120 path->p_tree_depth);
7123 * By now, el will point to the extent list on the bottom most
7124 * portion of this tree. Only the tail record is considered in
7127 * We handle the following cases, in order:
7128 * - empty extent: delete the remaining branch
7129 * - remove the entire record
7130 * - remove a partial record
7131 * - no record needs to be removed (truncate has completed)
7133 el = path_leaf_el(path);
7134 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7135 ocfs2_error(inode->i_sb,
7136 "Inode %llu has empty extent block at %llu\n",
7137 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7138 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7143 i = le16_to_cpu(el->l_next_free_rec) - 1;
7144 rec = &el->l_recs[i];
7145 flags = rec->e_flags;
7146 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7148 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7150 * Lower levels depend on this never happening, but it's best
7151 * to check it up here before changing the tree.
7153 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7154 mlog(ML_ERROR, "Inode %lu has an empty "
7155 "extent record, depth %u\n", inode->i_ino,
7156 le16_to_cpu(root_el->l_tree_depth));
7157 status = ocfs2_remove_rightmost_empty_extent(osb,
7158 &et, path, &dealloc);
7164 ocfs2_reinit_path(path, 1);
7167 trunc_cpos = le32_to_cpu(rec->e_cpos);
7171 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7173 * Truncate entire record.
7175 trunc_cpos = le32_to_cpu(rec->e_cpos);
7176 trunc_len = ocfs2_rec_clusters(el, rec);
7177 blkno = le64_to_cpu(rec->e_blkno);
7178 } else if (range > new_highest_cpos) {
7180 * Partial truncate. it also should be
7181 * the last truncate we're doing.
7183 trunc_cpos = new_highest_cpos;
7184 trunc_len = range - new_highest_cpos;
7185 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7186 blkno = le64_to_cpu(rec->e_blkno) +
7187 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7190 * Truncate completed, leave happily.
7196 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7198 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7199 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7207 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7208 phys_cpos, trunc_len, flags, &dealloc,
7209 refcount_loc, true);
7215 ocfs2_reinit_path(path, 1);
7218 * The check above will catch the case where we've truncated
7219 * away all allocation.
7225 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7227 ocfs2_schedule_truncate_log_flush(osb, 1);
7229 ocfs2_run_deallocs(osb, &dealloc);
7231 ocfs2_free_path(path);
7237 * 'start' is inclusive, 'end' is not.
7239 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7240 unsigned int start, unsigned int end, int trunc)
7243 unsigned int numbytes;
7245 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7246 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7247 struct ocfs2_inline_data *idata = &di->id2.i_data;
7249 if (end > i_size_read(inode))
7250 end = i_size_read(inode);
7252 BUG_ON(start > end);
7254 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7255 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7256 !ocfs2_supports_inline_data(osb)) {
7257 ocfs2_error(inode->i_sb,
7258 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7259 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7260 le16_to_cpu(di->i_dyn_features),
7261 OCFS2_I(inode)->ip_dyn_features,
7262 osb->s_feature_incompat);
7267 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7268 if (IS_ERR(handle)) {
7269 ret = PTR_ERR(handle);
7274 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7275 OCFS2_JOURNAL_ACCESS_WRITE);
7281 numbytes = end - start;
7282 memset(idata->id_data + start, 0, numbytes);
7285 * No need to worry about the data page here - it's been
7286 * truncated already and inline data doesn't need it for
7287 * pushing zero's to disk, so we'll let readpage pick it up
7291 i_size_write(inode, start);
7292 di->i_size = cpu_to_le64(start);
7295 inode->i_blocks = ocfs2_inode_sector_count(inode);
7296 inode->i_ctime = inode->i_mtime = current_time(inode);
7298 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7299 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7301 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7302 ocfs2_journal_dirty(handle, di_bh);
7305 ocfs2_commit_trans(osb, handle);
7311 static int ocfs2_trim_extent(struct super_block *sb,
7312 struct ocfs2_group_desc *gd,
7313 u32 start, u32 count)
7315 u64 discard, bcount;
7317 bcount = ocfs2_clusters_to_blocks(sb, count);
7318 discard = le64_to_cpu(gd->bg_blkno) +
7319 ocfs2_clusters_to_blocks(sb, start);
7321 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7323 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7326 static int ocfs2_trim_group(struct super_block *sb,
7327 struct ocfs2_group_desc *gd,
7328 u32 start, u32 max, u32 minbits)
7330 int ret = 0, count = 0, next;
7331 void *bitmap = gd->bg_bitmap;
7333 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7336 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7337 start, max, minbits);
7339 while (start < max) {
7340 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7343 next = ocfs2_find_next_bit(bitmap, max, start);
7345 if ((next - start) >= minbits) {
7346 ret = ocfs2_trim_extent(sb, gd,
7347 start, next - start);
7352 count += next - start;
7356 if (fatal_signal_pending(current)) {
7357 count = -ERESTARTSYS;
7361 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7371 int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7373 struct ocfs2_super *osb = OCFS2_SB(sb);
7374 u64 start, len, trimmed, first_group, last_group, group;
7376 u32 first_bit, last_bit, minlen;
7377 struct buffer_head *main_bm_bh = NULL;
7378 struct inode *main_bm_inode = NULL;
7379 struct buffer_head *gd_bh = NULL;
7380 struct ocfs2_dinode *main_bm;
7381 struct ocfs2_group_desc *gd = NULL;
7383 start = range->start >> osb->s_clustersize_bits;
7384 len = range->len >> osb->s_clustersize_bits;
7385 minlen = range->minlen >> osb->s_clustersize_bits;
7387 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7390 main_bm_inode = ocfs2_get_system_file_inode(osb,
7391 GLOBAL_BITMAP_SYSTEM_INODE,
7392 OCFS2_INVALID_SLOT);
7393 if (!main_bm_inode) {
7399 inode_lock(main_bm_inode);
7401 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7406 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7408 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7413 len = range->len >> osb->s_clustersize_bits;
7414 if (start + len > le32_to_cpu(main_bm->i_clusters))
7415 len = le32_to_cpu(main_bm->i_clusters) - start;
7417 trace_ocfs2_trim_fs(start, len, minlen);
7419 /* Determine first and last group to examine based on start and len */
7420 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7421 if (first_group == osb->first_cluster_group_blkno)
7424 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7425 last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7426 last_bit = osb->bitmap_cpg;
7429 for (group = first_group; group <= last_group;) {
7430 if (first_bit + len >= osb->bitmap_cpg)
7431 last_bit = osb->bitmap_cpg;
7433 last_bit = first_bit + len;
7435 ret = ocfs2_read_group_descriptor(main_bm_inode,
7443 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7444 cnt = ocfs2_trim_group(sb, gd, first_bit, last_bit, minlen);
7454 len -= osb->bitmap_cpg - first_bit;
7456 if (group == osb->first_cluster_group_blkno)
7457 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7459 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7461 range->len = trimmed * sb->s_blocksize;
7463 ocfs2_inode_unlock(main_bm_inode, 0);
7466 inode_unlock(main_bm_inode);
7467 iput(main_bm_inode);