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>
34 #include <cluster/masklog.h>
40 #include "blockcheck.h"
42 #include "extent_map.h"
45 #include "localalloc.h"
52 #include "refcounttree.h"
53 #include "ocfs2_trace.h"
55 #include "buffer_head_io.h"
57 enum ocfs2_contig_type {
64 static enum ocfs2_contig_type
65 ocfs2_extent_rec_contig(struct super_block *sb,
66 struct ocfs2_extent_rec *ext,
67 struct ocfs2_extent_rec *insert_rec);
69 * Operations for a specific extent tree type.
71 * To implement an on-disk btree (extent tree) type in ocfs2, add
72 * an ocfs2_extent_tree_operations structure and the matching
73 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
74 * for the allocation portion of the extent tree.
76 struct ocfs2_extent_tree_operations {
78 * last_eb_blk is the block number of the right most leaf extent
79 * block. Most on-disk structures containing an extent tree store
80 * this value for fast access. The ->eo_set_last_eb_blk() and
81 * ->eo_get_last_eb_blk() operations access this value. They are
84 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
86 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
89 * The on-disk structure usually keeps track of how many total
90 * clusters are stored in this extent tree. This function updates
91 * that value. new_clusters is the delta, and must be
92 * added to the total. Required.
94 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
98 * If this extent tree is supported by an extent map, insert
99 * a record into the map.
101 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
102 struct ocfs2_extent_rec *rec);
105 * If this extent tree is supported by an extent map, truncate the
108 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
112 * If ->eo_insert_check() exists, it is called before rec is
113 * inserted into the extent tree. It is optional.
115 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
116 struct ocfs2_extent_rec *rec);
117 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
120 * --------------------------------------------------------------
121 * The remaining are internal to ocfs2_extent_tree and don't have
126 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
129 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
132 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
133 * it exists. If it does not, et->et_max_leaf_clusters is set
134 * to 0 (unlimited). Optional.
136 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
139 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
140 * are contiguous or not. Optional. Don't need to set it if use
141 * ocfs2_extent_rec as the tree leaf.
143 enum ocfs2_contig_type
144 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
145 struct ocfs2_extent_rec *ext,
146 struct ocfs2_extent_rec *insert_rec);
151 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
154 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
155 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
157 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
159 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
160 struct ocfs2_extent_rec *rec);
161 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
163 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
164 struct ocfs2_extent_rec *rec);
165 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
166 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
167 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
168 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
169 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
170 .eo_update_clusters = ocfs2_dinode_update_clusters,
171 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
172 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
173 .eo_insert_check = ocfs2_dinode_insert_check,
174 .eo_sanity_check = ocfs2_dinode_sanity_check,
175 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
178 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
181 struct ocfs2_dinode *di = et->et_object;
183 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
184 di->i_last_eb_blk = cpu_to_le64(blkno);
187 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
189 struct ocfs2_dinode *di = et->et_object;
191 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
192 return le64_to_cpu(di->i_last_eb_blk);
195 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
198 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
199 struct ocfs2_dinode *di = et->et_object;
201 le32_add_cpu(&di->i_clusters, clusters);
202 spin_lock(&oi->ip_lock);
203 oi->ip_clusters = le32_to_cpu(di->i_clusters);
204 spin_unlock(&oi->ip_lock);
207 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
208 struct ocfs2_extent_rec *rec)
210 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
212 ocfs2_extent_map_insert_rec(inode, rec);
215 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
218 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
220 ocfs2_extent_map_trunc(inode, clusters);
223 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
224 struct ocfs2_extent_rec *rec)
226 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
227 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
229 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
230 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
231 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
232 "Device %s, asking for sparse allocation: inode %llu, "
233 "cpos %u, clusters %u\n",
235 (unsigned long long)oi->ip_blkno,
236 rec->e_cpos, oi->ip_clusters);
241 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
243 struct ocfs2_dinode *di = et->et_object;
245 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
246 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
251 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
253 struct ocfs2_dinode *di = et->et_object;
255 et->et_root_el = &di->id2.i_list;
259 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
261 struct ocfs2_xattr_value_buf *vb = et->et_object;
263 et->et_root_el = &vb->vb_xv->xr_list;
266 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
269 struct ocfs2_xattr_value_buf *vb = et->et_object;
271 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
274 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
276 struct ocfs2_xattr_value_buf *vb = et->et_object;
278 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
281 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
284 struct ocfs2_xattr_value_buf *vb = et->et_object;
286 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
289 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
290 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
291 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
292 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
293 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
296 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
298 struct ocfs2_xattr_block *xb = et->et_object;
300 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
303 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
305 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
306 et->et_max_leaf_clusters =
307 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
310 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
313 struct ocfs2_xattr_block *xb = et->et_object;
314 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
316 xt->xt_last_eb_blk = cpu_to_le64(blkno);
319 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
321 struct ocfs2_xattr_block *xb = et->et_object;
322 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
324 return le64_to_cpu(xt->xt_last_eb_blk);
327 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
330 struct ocfs2_xattr_block *xb = et->et_object;
332 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
335 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
336 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
337 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
338 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
339 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
340 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
343 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
346 struct ocfs2_dx_root_block *dx_root = et->et_object;
348 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
351 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
353 struct ocfs2_dx_root_block *dx_root = et->et_object;
355 return le64_to_cpu(dx_root->dr_last_eb_blk);
358 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
361 struct ocfs2_dx_root_block *dx_root = et->et_object;
363 le32_add_cpu(&dx_root->dr_clusters, clusters);
366 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
368 struct ocfs2_dx_root_block *dx_root = et->et_object;
370 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
375 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
377 struct ocfs2_dx_root_block *dx_root = et->et_object;
379 et->et_root_el = &dx_root->dr_list;
382 static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
383 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
384 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
385 .eo_update_clusters = ocfs2_dx_root_update_clusters,
386 .eo_sanity_check = ocfs2_dx_root_sanity_check,
387 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
390 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
392 struct ocfs2_refcount_block *rb = et->et_object;
394 et->et_root_el = &rb->rf_list;
397 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
400 struct ocfs2_refcount_block *rb = et->et_object;
402 rb->rf_last_eb_blk = cpu_to_le64(blkno);
405 static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
407 struct ocfs2_refcount_block *rb = et->et_object;
409 return le64_to_cpu(rb->rf_last_eb_blk);
412 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
415 struct ocfs2_refcount_block *rb = et->et_object;
417 le32_add_cpu(&rb->rf_clusters, clusters);
420 static enum ocfs2_contig_type
421 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
422 struct ocfs2_extent_rec *ext,
423 struct ocfs2_extent_rec *insert_rec)
428 static struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
429 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
430 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
431 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
432 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
433 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
436 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
437 struct ocfs2_caching_info *ci,
438 struct buffer_head *bh,
439 ocfs2_journal_access_func access,
441 struct ocfs2_extent_tree_operations *ops)
446 et->et_root_journal_access = access;
448 obj = (void *)bh->b_data;
451 et->et_ops->eo_fill_root_el(et);
452 if (!et->et_ops->eo_fill_max_leaf_clusters)
453 et->et_max_leaf_clusters = 0;
455 et->et_ops->eo_fill_max_leaf_clusters(et);
458 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
459 struct ocfs2_caching_info *ci,
460 struct buffer_head *bh)
462 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
463 NULL, &ocfs2_dinode_et_ops);
466 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
467 struct ocfs2_caching_info *ci,
468 struct buffer_head *bh)
470 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
471 NULL, &ocfs2_xattr_tree_et_ops);
474 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
475 struct ocfs2_caching_info *ci,
476 struct ocfs2_xattr_value_buf *vb)
478 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
479 &ocfs2_xattr_value_et_ops);
482 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
483 struct ocfs2_caching_info *ci,
484 struct buffer_head *bh)
486 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
487 NULL, &ocfs2_dx_root_et_ops);
490 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
491 struct ocfs2_caching_info *ci,
492 struct buffer_head *bh)
494 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
495 NULL, &ocfs2_refcount_tree_et_ops);
498 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
501 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
504 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
506 return et->et_ops->eo_get_last_eb_blk(et);
509 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
512 et->et_ops->eo_update_clusters(et, clusters);
515 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
516 struct ocfs2_extent_rec *rec)
518 if (et->et_ops->eo_extent_map_insert)
519 et->et_ops->eo_extent_map_insert(et, rec);
522 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
525 if (et->et_ops->eo_extent_map_truncate)
526 et->et_ops->eo_extent_map_truncate(et, clusters);
529 static inline int ocfs2_et_root_journal_access(handle_t *handle,
530 struct ocfs2_extent_tree *et,
533 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
537 static inline enum ocfs2_contig_type
538 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
539 struct ocfs2_extent_rec *rec,
540 struct ocfs2_extent_rec *insert_rec)
542 if (et->et_ops->eo_extent_contig)
543 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
545 return ocfs2_extent_rec_contig(
546 ocfs2_metadata_cache_get_super(et->et_ci),
550 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
551 struct ocfs2_extent_rec *rec)
555 if (et->et_ops->eo_insert_check)
556 ret = et->et_ops->eo_insert_check(et, rec);
560 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
564 if (et->et_ops->eo_sanity_check)
565 ret = et->et_ops->eo_sanity_check(et);
569 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
570 struct ocfs2_extent_block *eb);
571 static void ocfs2_adjust_rightmost_records(handle_t *handle,
572 struct ocfs2_extent_tree *et,
573 struct ocfs2_path *path,
574 struct ocfs2_extent_rec *insert_rec);
576 * Reset the actual path elements so that we can re-use the structure
577 * to build another path. Generally, this involves freeing the buffer
580 void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
582 int i, start = 0, depth = 0;
583 struct ocfs2_path_item *node;
588 for(i = start; i < path_num_items(path); i++) {
589 node = &path->p_node[i];
597 * Tree depth may change during truncate, or insert. If we're
598 * keeping the root extent list, then make sure that our path
599 * structure reflects the proper depth.
602 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
604 path_root_access(path) = NULL;
606 path->p_tree_depth = depth;
609 void ocfs2_free_path(struct ocfs2_path *path)
612 ocfs2_reinit_path(path, 0);
618 * All the elements of src into dest. After this call, src could be freed
619 * without affecting dest.
621 * Both paths should have the same root. Any non-root elements of dest
624 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
628 BUG_ON(path_root_bh(dest) != path_root_bh(src));
629 BUG_ON(path_root_el(dest) != path_root_el(src));
630 BUG_ON(path_root_access(dest) != path_root_access(src));
632 ocfs2_reinit_path(dest, 1);
634 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
635 dest->p_node[i].bh = src->p_node[i].bh;
636 dest->p_node[i].el = src->p_node[i].el;
638 if (dest->p_node[i].bh)
639 get_bh(dest->p_node[i].bh);
644 * Make the *dest path the same as src and re-initialize src path to
647 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
651 BUG_ON(path_root_bh(dest) != path_root_bh(src));
652 BUG_ON(path_root_access(dest) != path_root_access(src));
654 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
655 brelse(dest->p_node[i].bh);
657 dest->p_node[i].bh = src->p_node[i].bh;
658 dest->p_node[i].el = src->p_node[i].el;
660 src->p_node[i].bh = NULL;
661 src->p_node[i].el = NULL;
666 * Insert an extent block at given index.
668 * This will not take an additional reference on eb_bh.
670 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
671 struct buffer_head *eb_bh)
673 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
676 * Right now, no root bh is an extent block, so this helps
677 * catch code errors with dinode trees. The assertion can be
678 * safely removed if we ever need to insert extent block
679 * structures at the root.
683 path->p_node[index].bh = eb_bh;
684 path->p_node[index].el = &eb->h_list;
687 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
688 struct ocfs2_extent_list *root_el,
689 ocfs2_journal_access_func access)
691 struct ocfs2_path *path;
693 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
695 path = kzalloc(sizeof(*path), GFP_NOFS);
697 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
699 path_root_bh(path) = root_bh;
700 path_root_el(path) = root_el;
701 path_root_access(path) = access;
707 struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
709 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
710 path_root_access(path));
713 struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
715 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
716 et->et_root_journal_access);
720 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
721 * otherwise it's the root_access function.
723 * I don't like the way this function's name looks next to
724 * ocfs2_journal_access_path(), but I don't have a better one.
726 int ocfs2_path_bh_journal_access(handle_t *handle,
727 struct ocfs2_caching_info *ci,
728 struct ocfs2_path *path,
731 ocfs2_journal_access_func access = path_root_access(path);
734 access = ocfs2_journal_access;
737 access = ocfs2_journal_access_eb;
739 return access(handle, ci, path->p_node[idx].bh,
740 OCFS2_JOURNAL_ACCESS_WRITE);
744 * Convenience function to journal all components in a path.
746 int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
748 struct ocfs2_path *path)
755 for(i = 0; i < path_num_items(path); i++) {
756 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
768 * Return the index of the extent record which contains cluster #v_cluster.
769 * -1 is returned if it was not found.
771 * Should work fine on interior and exterior nodes.
773 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
777 struct ocfs2_extent_rec *rec;
778 u32 rec_end, rec_start, clusters;
780 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
781 rec = &el->l_recs[i];
783 rec_start = le32_to_cpu(rec->e_cpos);
784 clusters = ocfs2_rec_clusters(el, rec);
786 rec_end = rec_start + clusters;
788 if (v_cluster >= rec_start && v_cluster < rec_end) {
798 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
799 * ocfs2_extent_rec_contig only work properly against leaf nodes!
801 static int ocfs2_block_extent_contig(struct super_block *sb,
802 struct ocfs2_extent_rec *ext,
805 u64 blk_end = le64_to_cpu(ext->e_blkno);
807 blk_end += ocfs2_clusters_to_blocks(sb,
808 le16_to_cpu(ext->e_leaf_clusters));
810 return blkno == blk_end;
813 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
814 struct ocfs2_extent_rec *right)
818 left_range = le32_to_cpu(left->e_cpos) +
819 le16_to_cpu(left->e_leaf_clusters);
821 return (left_range == le32_to_cpu(right->e_cpos));
824 static enum ocfs2_contig_type
825 ocfs2_extent_rec_contig(struct super_block *sb,
826 struct ocfs2_extent_rec *ext,
827 struct ocfs2_extent_rec *insert_rec)
829 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
832 * Refuse to coalesce extent records with different flag
833 * fields - we don't want to mix unwritten extents with user
836 if (ext->e_flags != insert_rec->e_flags)
839 if (ocfs2_extents_adjacent(ext, insert_rec) &&
840 ocfs2_block_extent_contig(sb, ext, blkno))
843 blkno = le64_to_cpu(ext->e_blkno);
844 if (ocfs2_extents_adjacent(insert_rec, ext) &&
845 ocfs2_block_extent_contig(sb, insert_rec, blkno))
852 * NOTE: We can have pretty much any combination of contiguousness and
855 * The usefulness of APPEND_TAIL is more in that it lets us know that
856 * we'll have to update the path to that leaf.
858 enum ocfs2_append_type {
863 enum ocfs2_split_type {
869 struct ocfs2_insert_type {
870 enum ocfs2_split_type ins_split;
871 enum ocfs2_append_type ins_appending;
872 enum ocfs2_contig_type ins_contig;
873 int ins_contig_index;
877 struct ocfs2_merge_ctxt {
878 enum ocfs2_contig_type c_contig_type;
879 int c_has_empty_extent;
880 int c_split_covers_rec;
883 static int ocfs2_validate_extent_block(struct super_block *sb,
884 struct buffer_head *bh)
887 struct ocfs2_extent_block *eb =
888 (struct ocfs2_extent_block *)bh->b_data;
890 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
892 BUG_ON(!buffer_uptodate(bh));
895 * If the ecc fails, we return the error but otherwise
896 * leave the filesystem running. We know any error is
897 * local to this block.
899 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
901 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
902 (unsigned long long)bh->b_blocknr);
907 * Errors after here are fatal.
910 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
912 "Extent block #%llu has bad signature %.*s",
913 (unsigned long long)bh->b_blocknr, 7,
918 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
920 "Extent block #%llu has an invalid h_blkno "
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 "
930 "h_fs_generation of #%u",
931 (unsigned long long)bh->b_blocknr,
932 le32_to_cpu(eb->h_fs_generation));
939 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
940 struct buffer_head **bh)
943 struct buffer_head *tmp = *bh;
945 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
946 ocfs2_validate_extent_block);
948 /* If ocfs2_read_block() got us a new bh, pass it up. */
957 * How many free extents have we got before we need more meta data?
959 int ocfs2_num_free_extents(struct ocfs2_super *osb,
960 struct ocfs2_extent_tree *et)
963 struct ocfs2_extent_list *el = NULL;
964 struct ocfs2_extent_block *eb;
965 struct buffer_head *eb_bh = NULL;
969 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
972 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
978 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
982 BUG_ON(el->l_tree_depth != 0);
984 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
988 trace_ocfs2_num_free_extents(retval);
992 /* expects array to already be allocated
994 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
997 static int ocfs2_create_new_meta_bhs(handle_t *handle,
998 struct ocfs2_extent_tree *et,
1000 struct ocfs2_alloc_context *meta_ac,
1001 struct buffer_head *bhs[])
1003 int count, status, i;
1004 u16 suballoc_bit_start;
1006 u64 suballoc_loc, first_blkno;
1007 struct ocfs2_super *osb =
1008 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1009 struct ocfs2_extent_block *eb;
1012 while (count < wanted) {
1013 status = ocfs2_claim_metadata(handle,
1017 &suballoc_bit_start,
1025 for(i = count; i < (num_got + count); i++) {
1026 bhs[i] = sb_getblk(osb->sb, first_blkno);
1027 if (bhs[i] == NULL) {
1032 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1034 status = ocfs2_journal_access_eb(handle, et->et_ci,
1036 OCFS2_JOURNAL_ACCESS_CREATE);
1042 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1043 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1044 /* Ok, setup the minimal stuff here. */
1045 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1046 eb->h_blkno = cpu_to_le64(first_blkno);
1047 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1048 eb->h_suballoc_slot =
1049 cpu_to_le16(meta_ac->ac_alloc_slot);
1050 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1051 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1052 eb->h_list.l_count =
1053 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1055 suballoc_bit_start++;
1058 /* We'll also be dirtied by the caller, so
1059 * this isn't absolutely necessary. */
1060 ocfs2_journal_dirty(handle, bhs[i]);
1069 for(i = 0; i < wanted; i++) {
1079 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1081 * Returns the sum of the rightmost extent rec logical offset and
1084 * ocfs2_add_branch() uses this to determine what logical cluster
1085 * value should be populated into the leftmost new branch records.
1087 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1088 * value for the new topmost tree record.
1090 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1094 i = le16_to_cpu(el->l_next_free_rec) - 1;
1096 return le32_to_cpu(el->l_recs[i].e_cpos) +
1097 ocfs2_rec_clusters(el, &el->l_recs[i]);
1101 * Change range of the branches in the right most path according to the leaf
1102 * extent block's rightmost record.
1104 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1105 struct ocfs2_extent_tree *et)
1108 struct ocfs2_path *path = NULL;
1109 struct ocfs2_extent_list *el;
1110 struct ocfs2_extent_rec *rec;
1112 path = ocfs2_new_path_from_et(et);
1118 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1124 status = ocfs2_extend_trans(handle, path_num_items(path));
1130 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1136 el = path_leaf_el(path);
1137 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1139 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1142 ocfs2_free_path(path);
1147 * Add an entire tree branch to our inode. eb_bh is the extent block
1148 * to start at, if we don't want to start the branch at the root
1151 * last_eb_bh is required as we have to update it's next_leaf pointer
1152 * for the new last extent block.
1154 * the new branch will be 'empty' in the sense that every block will
1155 * contain a single record with cluster count == 0.
1157 static int ocfs2_add_branch(handle_t *handle,
1158 struct ocfs2_extent_tree *et,
1159 struct buffer_head *eb_bh,
1160 struct buffer_head **last_eb_bh,
1161 struct ocfs2_alloc_context *meta_ac)
1163 int status, new_blocks, i;
1164 u64 next_blkno, new_last_eb_blk;
1165 struct buffer_head *bh;
1166 struct buffer_head **new_eb_bhs = NULL;
1167 struct ocfs2_extent_block *eb;
1168 struct ocfs2_extent_list *eb_el;
1169 struct ocfs2_extent_list *el;
1170 u32 new_cpos, root_end;
1172 BUG_ON(!last_eb_bh || !*last_eb_bh);
1175 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1178 el = et->et_root_el;
1180 /* we never add a branch to a leaf. */
1181 BUG_ON(!el->l_tree_depth);
1183 new_blocks = le16_to_cpu(el->l_tree_depth);
1185 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1186 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1187 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1190 * If there is a gap before the root end and the real end
1191 * of the righmost leaf block, we need to remove the gap
1192 * between new_cpos and root_end first so that the tree
1193 * is consistent after we add a new branch(it will start
1196 if (root_end > new_cpos) {
1197 trace_ocfs2_adjust_rightmost_branch(
1198 (unsigned long long)
1199 ocfs2_metadata_cache_owner(et->et_ci),
1200 root_end, new_cpos);
1202 status = ocfs2_adjust_rightmost_branch(handle, et);
1209 /* allocate the number of new eb blocks we need */
1210 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1218 status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1219 meta_ac, new_eb_bhs);
1225 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1226 * linked with the rest of the tree.
1227 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1229 * when we leave the loop, new_last_eb_blk will point to the
1230 * newest leaf, and next_blkno will point to the topmost extent
1232 next_blkno = new_last_eb_blk = 0;
1233 for(i = 0; i < new_blocks; i++) {
1235 eb = (struct ocfs2_extent_block *) bh->b_data;
1236 /* ocfs2_create_new_meta_bhs() should create it right! */
1237 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1238 eb_el = &eb->h_list;
1240 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1241 OCFS2_JOURNAL_ACCESS_CREATE);
1247 eb->h_next_leaf_blk = 0;
1248 eb_el->l_tree_depth = cpu_to_le16(i);
1249 eb_el->l_next_free_rec = cpu_to_le16(1);
1251 * This actually counts as an empty extent as
1254 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1255 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1257 * eb_el isn't always an interior node, but even leaf
1258 * nodes want a zero'd flags and reserved field so
1259 * this gets the whole 32 bits regardless of use.
1261 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1262 if (!eb_el->l_tree_depth)
1263 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1265 ocfs2_journal_dirty(handle, bh);
1266 next_blkno = le64_to_cpu(eb->h_blkno);
1269 /* This is a bit hairy. We want to update up to three blocks
1270 * here without leaving any of them in an inconsistent state
1271 * in case of error. We don't have to worry about
1272 * journal_dirty erroring as it won't unless we've aborted the
1273 * handle (in which case we would never be here) so reserving
1274 * the write with journal_access is all we need to do. */
1275 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1276 OCFS2_JOURNAL_ACCESS_WRITE);
1281 status = ocfs2_et_root_journal_access(handle, et,
1282 OCFS2_JOURNAL_ACCESS_WRITE);
1288 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1289 OCFS2_JOURNAL_ACCESS_WRITE);
1296 /* Link the new branch into the rest of the tree (el will
1297 * either be on the root_bh, or the extent block passed in. */
1298 i = le16_to_cpu(el->l_next_free_rec);
1299 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1300 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1301 el->l_recs[i].e_int_clusters = 0;
1302 le16_add_cpu(&el->l_next_free_rec, 1);
1304 /* fe needs a new last extent block pointer, as does the
1305 * next_leaf on the previously last-extent-block. */
1306 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1308 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1309 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1311 ocfs2_journal_dirty(handle, *last_eb_bh);
1312 ocfs2_journal_dirty(handle, et->et_root_bh);
1314 ocfs2_journal_dirty(handle, eb_bh);
1317 * Some callers want to track the rightmost leaf so pass it
1320 brelse(*last_eb_bh);
1321 get_bh(new_eb_bhs[0]);
1322 *last_eb_bh = new_eb_bhs[0];
1327 for (i = 0; i < new_blocks; i++)
1328 brelse(new_eb_bhs[i]);
1336 * adds another level to the allocation tree.
1337 * returns back the new extent block so you can add a branch to it
1340 static int ocfs2_shift_tree_depth(handle_t *handle,
1341 struct ocfs2_extent_tree *et,
1342 struct ocfs2_alloc_context *meta_ac,
1343 struct buffer_head **ret_new_eb_bh)
1347 struct buffer_head *new_eb_bh = NULL;
1348 struct ocfs2_extent_block *eb;
1349 struct ocfs2_extent_list *root_el;
1350 struct ocfs2_extent_list *eb_el;
1352 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1359 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1360 /* ocfs2_create_new_meta_bhs() should create it right! */
1361 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1363 eb_el = &eb->h_list;
1364 root_el = et->et_root_el;
1366 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1367 OCFS2_JOURNAL_ACCESS_CREATE);
1373 /* copy the root extent list data into the new extent block */
1374 eb_el->l_tree_depth = root_el->l_tree_depth;
1375 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1376 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1377 eb_el->l_recs[i] = root_el->l_recs[i];
1379 ocfs2_journal_dirty(handle, new_eb_bh);
1381 status = ocfs2_et_root_journal_access(handle, et,
1382 OCFS2_JOURNAL_ACCESS_WRITE);
1388 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1390 /* update root_bh now */
1391 le16_add_cpu(&root_el->l_tree_depth, 1);
1392 root_el->l_recs[0].e_cpos = 0;
1393 root_el->l_recs[0].e_blkno = eb->h_blkno;
1394 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1395 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1396 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1397 root_el->l_next_free_rec = cpu_to_le16(1);
1399 /* If this is our 1st tree depth shift, then last_eb_blk
1400 * becomes the allocated extent block */
1401 if (root_el->l_tree_depth == cpu_to_le16(1))
1402 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1404 ocfs2_journal_dirty(handle, et->et_root_bh);
1406 *ret_new_eb_bh = new_eb_bh;
1416 * Should only be called when there is no space left in any of the
1417 * leaf nodes. What we want to do is find the lowest tree depth
1418 * non-leaf extent block with room for new records. There are three
1419 * valid results of this search:
1421 * 1) a lowest extent block is found, then we pass it back in
1422 * *lowest_eb_bh and return '0'
1424 * 2) the search fails to find anything, but the root_el has room. We
1425 * pass NULL back in *lowest_eb_bh, but still return '0'
1427 * 3) the search fails to find anything AND the root_el is full, in
1428 * which case we return > 0
1430 * return status < 0 indicates an error.
1432 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1433 struct buffer_head **target_bh)
1437 struct ocfs2_extent_block *eb;
1438 struct ocfs2_extent_list *el;
1439 struct buffer_head *bh = NULL;
1440 struct buffer_head *lowest_bh = NULL;
1444 el = et->et_root_el;
1446 while(le16_to_cpu(el->l_tree_depth) > 1) {
1447 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1448 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1449 "Owner %llu has empty "
1450 "extent list (next_free_rec == 0)",
1451 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1455 i = le16_to_cpu(el->l_next_free_rec) - 1;
1456 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1458 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1459 "Owner %llu has extent "
1460 "list where extent # %d has no physical "
1462 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1470 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1476 eb = (struct ocfs2_extent_block *) bh->b_data;
1479 if (le16_to_cpu(el->l_next_free_rec) <
1480 le16_to_cpu(el->l_count)) {
1487 /* If we didn't find one and the fe doesn't have any room,
1488 * then return '1' */
1489 el = et->et_root_el;
1490 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1493 *target_bh = lowest_bh;
1501 * Grow a b-tree so that it has more records.
1503 * We might shift the tree depth in which case existing paths should
1504 * be considered invalid.
1506 * Tree depth after the grow is returned via *final_depth.
1508 * *last_eb_bh will be updated by ocfs2_add_branch().
1510 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1511 int *final_depth, struct buffer_head **last_eb_bh,
1512 struct ocfs2_alloc_context *meta_ac)
1515 struct ocfs2_extent_list *el = et->et_root_el;
1516 int depth = le16_to_cpu(el->l_tree_depth);
1517 struct buffer_head *bh = NULL;
1519 BUG_ON(meta_ac == NULL);
1521 shift = ocfs2_find_branch_target(et, &bh);
1528 /* We traveled all the way to the bottom of the allocation tree
1529 * and didn't find room for any more extents - we need to add
1530 * another tree level */
1533 trace_ocfs2_grow_tree(
1534 (unsigned long long)
1535 ocfs2_metadata_cache_owner(et->et_ci),
1538 /* ocfs2_shift_tree_depth will return us a buffer with
1539 * the new extent block (so we can pass that to
1540 * ocfs2_add_branch). */
1541 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1549 * Special case: we have room now if we shifted from
1550 * tree_depth 0, so no more work needs to be done.
1552 * We won't be calling add_branch, so pass
1553 * back *last_eb_bh as the new leaf. At depth
1554 * zero, it should always be null so there's
1555 * no reason to brelse.
1557 BUG_ON(*last_eb_bh);
1564 /* call ocfs2_add_branch to add the final part of the tree with
1566 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1575 *final_depth = depth;
1581 * This function will discard the rightmost extent record.
1583 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1585 int next_free = le16_to_cpu(el->l_next_free_rec);
1586 int count = le16_to_cpu(el->l_count);
1587 unsigned int num_bytes;
1590 /* This will cause us to go off the end of our extent list. */
1591 BUG_ON(next_free >= count);
1593 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1595 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1598 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1599 struct ocfs2_extent_rec *insert_rec)
1601 int i, insert_index, next_free, has_empty, num_bytes;
1602 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1603 struct ocfs2_extent_rec *rec;
1605 next_free = le16_to_cpu(el->l_next_free_rec);
1606 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1610 /* The tree code before us didn't allow enough room in the leaf. */
1611 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1614 * The easiest way to approach this is to just remove the
1615 * empty extent and temporarily decrement next_free.
1619 * If next_free was 1 (only an empty extent), this
1620 * loop won't execute, which is fine. We still want
1621 * the decrement above to happen.
1623 for(i = 0; i < (next_free - 1); i++)
1624 el->l_recs[i] = el->l_recs[i+1];
1630 * Figure out what the new record index should be.
1632 for(i = 0; i < next_free; i++) {
1633 rec = &el->l_recs[i];
1635 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1640 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1641 has_empty, next_free,
1642 le16_to_cpu(el->l_count));
1644 BUG_ON(insert_index < 0);
1645 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1646 BUG_ON(insert_index > next_free);
1649 * No need to memmove if we're just adding to the tail.
1651 if (insert_index != next_free) {
1652 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1654 num_bytes = next_free - insert_index;
1655 num_bytes *= sizeof(struct ocfs2_extent_rec);
1656 memmove(&el->l_recs[insert_index + 1],
1657 &el->l_recs[insert_index],
1662 * Either we had an empty extent, and need to re-increment or
1663 * there was no empty extent on a non full rightmost leaf node,
1664 * in which case we still need to increment.
1667 el->l_next_free_rec = cpu_to_le16(next_free);
1669 * Make sure none of the math above just messed up our tree.
1671 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1673 el->l_recs[insert_index] = *insert_rec;
1677 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1679 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1681 BUG_ON(num_recs == 0);
1683 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1685 size = num_recs * sizeof(struct ocfs2_extent_rec);
1686 memmove(&el->l_recs[0], &el->l_recs[1], size);
1687 memset(&el->l_recs[num_recs], 0,
1688 sizeof(struct ocfs2_extent_rec));
1689 el->l_next_free_rec = cpu_to_le16(num_recs);
1694 * Create an empty extent record .
1696 * l_next_free_rec may be updated.
1698 * If an empty extent already exists do nothing.
1700 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1702 int next_free = le16_to_cpu(el->l_next_free_rec);
1704 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1709 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1712 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1713 "Asked to create an empty extent in a full list:\n"
1714 "count = %u, tree depth = %u",
1715 le16_to_cpu(el->l_count),
1716 le16_to_cpu(el->l_tree_depth));
1718 ocfs2_shift_records_right(el);
1721 le16_add_cpu(&el->l_next_free_rec, 1);
1722 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1726 * For a rotation which involves two leaf nodes, the "root node" is
1727 * the lowest level tree node which contains a path to both leafs. This
1728 * resulting set of information can be used to form a complete "subtree"
1730 * This function is passed two full paths from the dinode down to a
1731 * pair of adjacent leaves. It's task is to figure out which path
1732 * index contains the subtree root - this can be the root index itself
1733 * in a worst-case rotation.
1735 * The array index of the subtree root is passed back.
1737 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1738 struct ocfs2_path *left,
1739 struct ocfs2_path *right)
1744 * Check that the caller passed in two paths from the same tree.
1746 BUG_ON(path_root_bh(left) != path_root_bh(right));
1752 * The caller didn't pass two adjacent paths.
1754 mlog_bug_on_msg(i > left->p_tree_depth,
1755 "Owner %llu, left depth %u, right depth %u\n"
1756 "left leaf blk %llu, right leaf blk %llu\n",
1757 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1758 left->p_tree_depth, right->p_tree_depth,
1759 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1760 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1761 } while (left->p_node[i].bh->b_blocknr ==
1762 right->p_node[i].bh->b_blocknr);
1767 typedef void (path_insert_t)(void *, struct buffer_head *);
1770 * Traverse a btree path in search of cpos, starting at root_el.
1772 * This code can be called with a cpos larger than the tree, in which
1773 * case it will return the rightmost path.
1775 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1776 struct ocfs2_extent_list *root_el, u32 cpos,
1777 path_insert_t *func, void *data)
1782 struct buffer_head *bh = NULL;
1783 struct ocfs2_extent_block *eb;
1784 struct ocfs2_extent_list *el;
1785 struct ocfs2_extent_rec *rec;
1788 while (el->l_tree_depth) {
1789 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1790 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1791 "Owner %llu has empty extent list at "
1793 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1794 le16_to_cpu(el->l_tree_depth));
1800 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1801 rec = &el->l_recs[i];
1804 * In the case that cpos is off the allocation
1805 * tree, this should just wind up returning the
1808 range = le32_to_cpu(rec->e_cpos) +
1809 ocfs2_rec_clusters(el, rec);
1810 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1814 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1816 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1817 "Owner %llu has bad blkno in extent list "
1818 "at depth %u (index %d)\n",
1819 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1820 le16_to_cpu(el->l_tree_depth), i);
1827 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1833 eb = (struct ocfs2_extent_block *) bh->b_data;
1836 if (le16_to_cpu(el->l_next_free_rec) >
1837 le16_to_cpu(el->l_count)) {
1838 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1839 "Owner %llu has bad count in extent list "
1840 "at block %llu (next free=%u, count=%u)\n",
1841 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1842 (unsigned long long)bh->b_blocknr,
1843 le16_to_cpu(el->l_next_free_rec),
1844 le16_to_cpu(el->l_count));
1855 * Catch any trailing bh that the loop didn't handle.
1863 * Given an initialized path (that is, it has a valid root extent
1864 * list), this function will traverse the btree in search of the path
1865 * which would contain cpos.
1867 * The path traveled is recorded in the path structure.
1869 * Note that this will not do any comparisons on leaf node extent
1870 * records, so it will work fine in the case that we just added a tree
1873 struct find_path_data {
1875 struct ocfs2_path *path;
1877 static void find_path_ins(void *data, struct buffer_head *bh)
1879 struct find_path_data *fp = data;
1882 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1885 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1886 struct ocfs2_path *path, u32 cpos)
1888 struct find_path_data data;
1892 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1893 find_path_ins, &data);
1896 static void find_leaf_ins(void *data, struct buffer_head *bh)
1898 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1899 struct ocfs2_extent_list *el = &eb->h_list;
1900 struct buffer_head **ret = data;
1902 /* We want to retain only the leaf block. */
1903 if (le16_to_cpu(el->l_tree_depth) == 0) {
1909 * Find the leaf block in the tree which would contain cpos. No
1910 * checking of the actual leaf is done.
1912 * Some paths want to call this instead of allocating a path structure
1913 * and calling ocfs2_find_path().
1915 * This function doesn't handle non btree extent lists.
1917 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1918 struct ocfs2_extent_list *root_el, u32 cpos,
1919 struct buffer_head **leaf_bh)
1922 struct buffer_head *bh = NULL;
1924 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1936 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1938 * Basically, we've moved stuff around at the bottom of the tree and
1939 * we need to fix up the extent records above the changes to reflect
1942 * left_rec: the record on the left.
1943 * left_child_el: is the child list pointed to by left_rec
1944 * right_rec: the record to the right of left_rec
1945 * right_child_el: is the child list pointed to by right_rec
1947 * By definition, this only works on interior nodes.
1949 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1950 struct ocfs2_extent_list *left_child_el,
1951 struct ocfs2_extent_rec *right_rec,
1952 struct ocfs2_extent_list *right_child_el)
1954 u32 left_clusters, right_end;
1957 * Interior nodes never have holes. Their cpos is the cpos of
1958 * the leftmost record in their child list. Their cluster
1959 * count covers the full theoretical range of their child list
1960 * - the range between their cpos and the cpos of the record
1961 * immediately to their right.
1963 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1964 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1965 BUG_ON(right_child_el->l_tree_depth);
1966 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1967 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1969 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1970 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1973 * Calculate the rightmost cluster count boundary before
1974 * moving cpos - we will need to adjust clusters after
1975 * updating e_cpos to keep the same highest cluster count.
1977 right_end = le32_to_cpu(right_rec->e_cpos);
1978 right_end += le32_to_cpu(right_rec->e_int_clusters);
1980 right_rec->e_cpos = left_rec->e_cpos;
1981 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1983 right_end -= le32_to_cpu(right_rec->e_cpos);
1984 right_rec->e_int_clusters = cpu_to_le32(right_end);
1988 * Adjust the adjacent root node records involved in a
1989 * rotation. left_el_blkno is passed in as a key so that we can easily
1990 * find it's index in the root list.
1992 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1993 struct ocfs2_extent_list *left_el,
1994 struct ocfs2_extent_list *right_el,
1999 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2000 le16_to_cpu(left_el->l_tree_depth));
2002 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2003 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2008 * The path walking code should have never returned a root and
2009 * two paths which are not adjacent.
2011 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2013 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
2014 &root_el->l_recs[i + 1], right_el);
2018 * We've changed a leaf block (in right_path) and need to reflect that
2019 * change back up the subtree.
2021 * This happens in multiple places:
2022 * - When we've moved an extent record from the left path leaf to the right
2023 * path leaf to make room for an empty extent in the left path leaf.
2024 * - When our insert into the right path leaf is at the leftmost edge
2025 * and requires an update of the path immediately to it's left. This
2026 * can occur at the end of some types of rotation and appending inserts.
2027 * - When we've adjusted the last extent record in the left path leaf and the
2028 * 1st extent record in the right path leaf during cross extent block merge.
2030 static void ocfs2_complete_edge_insert(handle_t *handle,
2031 struct ocfs2_path *left_path,
2032 struct ocfs2_path *right_path,
2036 struct ocfs2_extent_list *el, *left_el, *right_el;
2037 struct ocfs2_extent_rec *left_rec, *right_rec;
2038 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2041 * Update the counts and position values within all the
2042 * interior nodes to reflect the leaf rotation we just did.
2044 * The root node is handled below the loop.
2046 * We begin the loop with right_el and left_el pointing to the
2047 * leaf lists and work our way up.
2049 * NOTE: within this loop, left_el and right_el always refer
2050 * to the *child* lists.
2052 left_el = path_leaf_el(left_path);
2053 right_el = path_leaf_el(right_path);
2054 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2055 trace_ocfs2_complete_edge_insert(i);
2058 * One nice property of knowing that all of these
2059 * nodes are below the root is that we only deal with
2060 * the leftmost right node record and the rightmost
2063 el = left_path->p_node[i].el;
2064 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2065 left_rec = &el->l_recs[idx];
2067 el = right_path->p_node[i].el;
2068 right_rec = &el->l_recs[0];
2070 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2073 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2074 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2077 * Setup our list pointers now so that the current
2078 * parents become children in the next iteration.
2080 left_el = left_path->p_node[i].el;
2081 right_el = right_path->p_node[i].el;
2085 * At the root node, adjust the two adjacent records which
2086 * begin our path to the leaves.
2089 el = left_path->p_node[subtree_index].el;
2090 left_el = left_path->p_node[subtree_index + 1].el;
2091 right_el = right_path->p_node[subtree_index + 1].el;
2093 ocfs2_adjust_root_records(el, left_el, right_el,
2094 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2096 root_bh = left_path->p_node[subtree_index].bh;
2098 ocfs2_journal_dirty(handle, root_bh);
2101 static int ocfs2_rotate_subtree_right(handle_t *handle,
2102 struct ocfs2_extent_tree *et,
2103 struct ocfs2_path *left_path,
2104 struct ocfs2_path *right_path,
2108 struct buffer_head *right_leaf_bh;
2109 struct buffer_head *left_leaf_bh = NULL;
2110 struct buffer_head *root_bh;
2111 struct ocfs2_extent_list *right_el, *left_el;
2112 struct ocfs2_extent_rec move_rec;
2114 left_leaf_bh = path_leaf_bh(left_path);
2115 left_el = path_leaf_el(left_path);
2117 if (left_el->l_next_free_rec != left_el->l_count) {
2118 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2119 "Inode %llu has non-full interior leaf node %llu"
2121 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2122 (unsigned long long)left_leaf_bh->b_blocknr,
2123 le16_to_cpu(left_el->l_next_free_rec));
2128 * This extent block may already have an empty record, so we
2129 * return early if so.
2131 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2134 root_bh = left_path->p_node[subtree_index].bh;
2135 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2137 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2144 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2145 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2152 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2160 right_leaf_bh = path_leaf_bh(right_path);
2161 right_el = path_leaf_el(right_path);
2163 /* This is a code error, not a disk corruption. */
2164 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2165 "because rightmost leaf block %llu is empty\n",
2166 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2167 (unsigned long long)right_leaf_bh->b_blocknr);
2169 ocfs2_create_empty_extent(right_el);
2171 ocfs2_journal_dirty(handle, right_leaf_bh);
2173 /* Do the copy now. */
2174 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2175 move_rec = left_el->l_recs[i];
2176 right_el->l_recs[0] = move_rec;
2179 * Clear out the record we just copied and shift everything
2180 * over, leaving an empty extent in the left leaf.
2182 * We temporarily subtract from next_free_rec so that the
2183 * shift will lose the tail record (which is now defunct).
2185 le16_add_cpu(&left_el->l_next_free_rec, -1);
2186 ocfs2_shift_records_right(left_el);
2187 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2188 le16_add_cpu(&left_el->l_next_free_rec, 1);
2190 ocfs2_journal_dirty(handle, left_leaf_bh);
2192 ocfs2_complete_edge_insert(handle, left_path, right_path,
2200 * Given a full path, determine what cpos value would return us a path
2201 * containing the leaf immediately to the left of the current one.
2203 * Will return zero if the path passed in is already the leftmost path.
2205 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2206 struct ocfs2_path *path, u32 *cpos)
2210 struct ocfs2_extent_list *el;
2212 BUG_ON(path->p_tree_depth == 0);
2216 blkno = path_leaf_bh(path)->b_blocknr;
2218 /* Start at the tree node just above the leaf and work our way up. */
2219 i = path->p_tree_depth - 1;
2221 el = path->p_node[i].el;
2224 * Find the extent record just before the one in our
2227 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2228 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2232 * We've determined that the
2233 * path specified is already
2234 * the leftmost one - return a
2240 * The leftmost record points to our
2241 * leaf - we need to travel up the
2247 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2248 *cpos = *cpos + ocfs2_rec_clusters(el,
2249 &el->l_recs[j - 1]);
2256 * If we got here, we never found a valid node where
2257 * the tree indicated one should be.
2260 "Invalid extent tree at extent block %llu\n",
2261 (unsigned long long)blkno);
2266 blkno = path->p_node[i].bh->b_blocknr;
2275 * Extend the transaction by enough credits to complete the rotation,
2276 * and still leave at least the original number of credits allocated
2277 * to this transaction.
2279 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2281 struct ocfs2_path *path)
2284 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2286 if (handle->h_buffer_credits < credits)
2287 ret = ocfs2_extend_trans(handle,
2288 credits - handle->h_buffer_credits);
2294 * Trap the case where we're inserting into the theoretical range past
2295 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2296 * whose cpos is less than ours into the right leaf.
2298 * It's only necessary to look at the rightmost record of the left
2299 * leaf because the logic that calls us should ensure that the
2300 * theoretical ranges in the path components above the leaves are
2303 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2306 struct ocfs2_extent_list *left_el;
2307 struct ocfs2_extent_rec *rec;
2310 left_el = path_leaf_el(left_path);
2311 next_free = le16_to_cpu(left_el->l_next_free_rec);
2312 rec = &left_el->l_recs[next_free - 1];
2314 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2319 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2321 int next_free = le16_to_cpu(el->l_next_free_rec);
2323 struct ocfs2_extent_rec *rec;
2328 rec = &el->l_recs[0];
2329 if (ocfs2_is_empty_extent(rec)) {
2333 rec = &el->l_recs[1];
2336 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2337 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2343 * Rotate all the records in a btree right one record, starting at insert_cpos.
2345 * The path to the rightmost leaf should be passed in.
2347 * The array is assumed to be large enough to hold an entire path (tree depth).
2349 * Upon successful return from this function:
2351 * - The 'right_path' array will contain a path to the leaf block
2352 * whose range contains e_cpos.
2353 * - That leaf block will have a single empty extent in list index 0.
2354 * - In the case that the rotation requires a post-insert update,
2355 * *ret_left_path will contain a valid path which can be passed to
2356 * ocfs2_insert_path().
2358 static int ocfs2_rotate_tree_right(handle_t *handle,
2359 struct ocfs2_extent_tree *et,
2360 enum ocfs2_split_type split,
2362 struct ocfs2_path *right_path,
2363 struct ocfs2_path **ret_left_path)
2365 int ret, start, orig_credits = handle->h_buffer_credits;
2367 struct ocfs2_path *left_path = NULL;
2368 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2370 *ret_left_path = NULL;
2372 left_path = ocfs2_new_path_from_path(right_path);
2379 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2385 trace_ocfs2_rotate_tree_right(
2386 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2390 * What we want to do here is:
2392 * 1) Start with the rightmost path.
2394 * 2) Determine a path to the leaf block directly to the left
2397 * 3) Determine the 'subtree root' - the lowest level tree node
2398 * which contains a path to both leaves.
2400 * 4) Rotate the subtree.
2402 * 5) Find the next subtree by considering the left path to be
2403 * the new right path.
2405 * The check at the top of this while loop also accepts
2406 * insert_cpos == cpos because cpos is only a _theoretical_
2407 * value to get us the left path - insert_cpos might very well
2408 * be filling that hole.
2410 * Stop at a cpos of '0' because we either started at the
2411 * leftmost branch (i.e., a tree with one branch and a
2412 * rotation inside of it), or we've gone as far as we can in
2413 * rotating subtrees.
2415 while (cpos && insert_cpos <= cpos) {
2416 trace_ocfs2_rotate_tree_right(
2417 (unsigned long long)
2418 ocfs2_metadata_cache_owner(et->et_ci),
2421 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2427 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2428 path_leaf_bh(right_path),
2429 "Owner %llu: error during insert of %u "
2430 "(left path cpos %u) results in two identical "
2431 "paths ending at %llu\n",
2432 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2434 (unsigned long long)
2435 path_leaf_bh(left_path)->b_blocknr);
2437 if (split == SPLIT_NONE &&
2438 ocfs2_rotate_requires_path_adjustment(left_path,
2442 * We've rotated the tree as much as we
2443 * should. The rest is up to
2444 * ocfs2_insert_path() to complete, after the
2445 * record insertion. We indicate this
2446 * situation by returning the left path.
2448 * The reason we don't adjust the records here
2449 * before the record insert is that an error
2450 * later might break the rule where a parent
2451 * record e_cpos will reflect the actual
2452 * e_cpos of the 1st nonempty record of the
2455 *ret_left_path = left_path;
2459 start = ocfs2_find_subtree_root(et, left_path, right_path);
2461 trace_ocfs2_rotate_subtree(start,
2462 (unsigned long long)
2463 right_path->p_node[start].bh->b_blocknr,
2464 right_path->p_tree_depth);
2466 ret = ocfs2_extend_rotate_transaction(handle, start,
2467 orig_credits, right_path);
2473 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2480 if (split != SPLIT_NONE &&
2481 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2484 * A rotate moves the rightmost left leaf
2485 * record over to the leftmost right leaf
2486 * slot. If we're doing an extent split
2487 * instead of a real insert, then we have to
2488 * check that the extent to be split wasn't
2489 * just moved over. If it was, then we can
2490 * exit here, passing left_path back -
2491 * ocfs2_split_extent() is smart enough to
2492 * search both leaves.
2494 *ret_left_path = left_path;
2499 * There is no need to re-read the next right path
2500 * as we know that it'll be our current left
2501 * path. Optimize by copying values instead.
2503 ocfs2_mv_path(right_path, left_path);
2505 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2513 ocfs2_free_path(left_path);
2519 static int ocfs2_update_edge_lengths(handle_t *handle,
2520 struct ocfs2_extent_tree *et,
2521 int subtree_index, struct ocfs2_path *path)
2524 struct ocfs2_extent_rec *rec;
2525 struct ocfs2_extent_list *el;
2526 struct ocfs2_extent_block *eb;
2530 * In normal tree rotation process, we will never touch the
2531 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2532 * doesn't reserve the credits for them either.
2534 * But we do have a special case here which will update the rightmost
2535 * records for all the bh in the path.
2536 * So we have to allocate extra credits and access them.
2538 ret = ocfs2_extend_trans(handle, subtree_index);
2544 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2550 /* Path should always be rightmost. */
2551 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2552 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2555 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2556 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2557 rec = &el->l_recs[idx];
2558 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2560 for (i = 0; i < path->p_tree_depth; i++) {
2561 el = path->p_node[i].el;
2562 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2563 rec = &el->l_recs[idx];
2565 rec->e_int_clusters = cpu_to_le32(range);
2566 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2568 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2574 static void ocfs2_unlink_path(handle_t *handle,
2575 struct ocfs2_extent_tree *et,
2576 struct ocfs2_cached_dealloc_ctxt *dealloc,
2577 struct ocfs2_path *path, int unlink_start)
2580 struct ocfs2_extent_block *eb;
2581 struct ocfs2_extent_list *el;
2582 struct buffer_head *bh;
2584 for(i = unlink_start; i < path_num_items(path); i++) {
2585 bh = path->p_node[i].bh;
2587 eb = (struct ocfs2_extent_block *)bh->b_data;
2589 * Not all nodes might have had their final count
2590 * decremented by the caller - handle this here.
2593 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2595 "Inode %llu, attempted to remove extent block "
2596 "%llu with %u records\n",
2597 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2598 (unsigned long long)le64_to_cpu(eb->h_blkno),
2599 le16_to_cpu(el->l_next_free_rec));
2601 ocfs2_journal_dirty(handle, bh);
2602 ocfs2_remove_from_cache(et->et_ci, bh);
2606 el->l_next_free_rec = 0;
2607 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2609 ocfs2_journal_dirty(handle, bh);
2611 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2615 ocfs2_remove_from_cache(et->et_ci, bh);
2619 static void ocfs2_unlink_subtree(handle_t *handle,
2620 struct ocfs2_extent_tree *et,
2621 struct ocfs2_path *left_path,
2622 struct ocfs2_path *right_path,
2624 struct ocfs2_cached_dealloc_ctxt *dealloc)
2627 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2628 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2629 struct ocfs2_extent_list *el;
2630 struct ocfs2_extent_block *eb;
2632 el = path_leaf_el(left_path);
2634 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2636 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2637 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2640 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2642 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2643 le16_add_cpu(&root_el->l_next_free_rec, -1);
2645 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2646 eb->h_next_leaf_blk = 0;
2648 ocfs2_journal_dirty(handle, root_bh);
2649 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2651 ocfs2_unlink_path(handle, et, dealloc, right_path,
2655 static int ocfs2_rotate_subtree_left(handle_t *handle,
2656 struct ocfs2_extent_tree *et,
2657 struct ocfs2_path *left_path,
2658 struct ocfs2_path *right_path,
2660 struct ocfs2_cached_dealloc_ctxt *dealloc,
2663 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2664 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2665 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2666 struct ocfs2_extent_block *eb;
2670 right_leaf_el = path_leaf_el(right_path);
2671 left_leaf_el = path_leaf_el(left_path);
2672 root_bh = left_path->p_node[subtree_index].bh;
2673 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2675 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2678 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2679 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2681 * It's legal for us to proceed if the right leaf is
2682 * the rightmost one and it has an empty extent. There
2683 * are two cases to handle - whether the leaf will be
2684 * empty after removal or not. If the leaf isn't empty
2685 * then just remove the empty extent up front. The
2686 * next block will handle empty leaves by flagging
2689 * Non rightmost leaves will throw -EAGAIN and the
2690 * caller can manually move the subtree and retry.
2693 if (eb->h_next_leaf_blk != 0ULL)
2696 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2697 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2698 path_leaf_bh(right_path),
2699 OCFS2_JOURNAL_ACCESS_WRITE);
2705 ocfs2_remove_empty_extent(right_leaf_el);
2707 right_has_empty = 1;
2710 if (eb->h_next_leaf_blk == 0ULL &&
2711 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2713 * We have to update i_last_eb_blk during the meta
2716 ret = ocfs2_et_root_journal_access(handle, et,
2717 OCFS2_JOURNAL_ACCESS_WRITE);
2723 del_right_subtree = 1;
2727 * Getting here with an empty extent in the right path implies
2728 * that it's the rightmost path and will be deleted.
2730 BUG_ON(right_has_empty && !del_right_subtree);
2732 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2739 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2740 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2747 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2755 if (!right_has_empty) {
2757 * Only do this if we're moving a real
2758 * record. Otherwise, the action is delayed until
2759 * after removal of the right path in which case we
2760 * can do a simple shift to remove the empty extent.
2762 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2763 memset(&right_leaf_el->l_recs[0], 0,
2764 sizeof(struct ocfs2_extent_rec));
2766 if (eb->h_next_leaf_blk == 0ULL) {
2768 * Move recs over to get rid of empty extent, decrease
2769 * next_free. This is allowed to remove the last
2770 * extent in our leaf (setting l_next_free_rec to
2771 * zero) - the delete code below won't care.
2773 ocfs2_remove_empty_extent(right_leaf_el);
2776 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2777 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2779 if (del_right_subtree) {
2780 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2781 subtree_index, dealloc);
2782 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
2789 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2790 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2793 * Removal of the extent in the left leaf was skipped
2794 * above so we could delete the right path
2797 if (right_has_empty)
2798 ocfs2_remove_empty_extent(left_leaf_el);
2800 ocfs2_journal_dirty(handle, et_root_bh);
2804 ocfs2_complete_edge_insert(handle, left_path, right_path,
2812 * Given a full path, determine what cpos value would return us a path
2813 * containing the leaf immediately to the right of the current one.
2815 * Will return zero if the path passed in is already the rightmost path.
2817 * This looks similar, but is subtly different to
2818 * ocfs2_find_cpos_for_left_leaf().
2820 int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2821 struct ocfs2_path *path, u32 *cpos)
2825 struct ocfs2_extent_list *el;
2829 if (path->p_tree_depth == 0)
2832 blkno = path_leaf_bh(path)->b_blocknr;
2834 /* Start at the tree node just above the leaf and work our way up. */
2835 i = path->p_tree_depth - 1;
2839 el = path->p_node[i].el;
2842 * Find the extent record just after the one in our
2845 next_free = le16_to_cpu(el->l_next_free_rec);
2846 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2847 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2848 if (j == (next_free - 1)) {
2851 * We've determined that the
2852 * path specified is already
2853 * the rightmost one - return a
2859 * The rightmost record points to our
2860 * leaf - we need to travel up the
2866 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2872 * If we got here, we never found a valid node where
2873 * the tree indicated one should be.
2876 "Invalid extent tree at extent block %llu\n",
2877 (unsigned long long)blkno);
2882 blkno = path->p_node[i].bh->b_blocknr;
2890 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2891 struct ocfs2_extent_tree *et,
2892 struct ocfs2_path *path)
2895 struct buffer_head *bh = path_leaf_bh(path);
2896 struct ocfs2_extent_list *el = path_leaf_el(path);
2898 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2901 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2902 path_num_items(path) - 1);
2908 ocfs2_remove_empty_extent(el);
2909 ocfs2_journal_dirty(handle, bh);
2915 static int __ocfs2_rotate_tree_left(handle_t *handle,
2916 struct ocfs2_extent_tree *et,
2918 struct ocfs2_path *path,
2919 struct ocfs2_cached_dealloc_ctxt *dealloc,
2920 struct ocfs2_path **empty_extent_path)
2922 int ret, subtree_root, deleted;
2924 struct ocfs2_path *left_path = NULL;
2925 struct ocfs2_path *right_path = NULL;
2926 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2928 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2930 *empty_extent_path = NULL;
2932 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2938 left_path = ocfs2_new_path_from_path(path);
2945 ocfs2_cp_path(left_path, path);
2947 right_path = ocfs2_new_path_from_path(path);
2954 while (right_cpos) {
2955 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2961 subtree_root = ocfs2_find_subtree_root(et, left_path,
2964 trace_ocfs2_rotate_subtree(subtree_root,
2965 (unsigned long long)
2966 right_path->p_node[subtree_root].bh->b_blocknr,
2967 right_path->p_tree_depth);
2969 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2970 orig_credits, left_path);
2977 * Caller might still want to make changes to the
2978 * tree root, so re-add it to the journal here.
2980 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2987 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2988 right_path, subtree_root,
2990 if (ret == -EAGAIN) {
2992 * The rotation has to temporarily stop due to
2993 * the right subtree having an empty
2994 * extent. Pass it back to the caller for a
2997 *empty_extent_path = right_path;
3007 * The subtree rotate might have removed records on
3008 * the rightmost edge. If so, then rotation is
3014 ocfs2_mv_path(left_path, right_path);
3016 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3025 ocfs2_free_path(right_path);
3026 ocfs2_free_path(left_path);
3031 static int ocfs2_remove_rightmost_path(handle_t *handle,
3032 struct ocfs2_extent_tree *et,
3033 struct ocfs2_path *path,
3034 struct ocfs2_cached_dealloc_ctxt *dealloc)
3036 int ret, subtree_index;
3038 struct ocfs2_path *left_path = NULL;
3039 struct ocfs2_extent_block *eb;
3040 struct ocfs2_extent_list *el;
3043 ret = ocfs2_et_sanity_check(et);
3047 * There's two ways we handle this depending on
3048 * whether path is the only existing one.
3050 ret = ocfs2_extend_rotate_transaction(handle, 0,
3051 handle->h_buffer_credits,
3058 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3064 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3073 * We have a path to the left of this one - it needs
3076 left_path = ocfs2_new_path_from_path(path);
3083 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3089 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3095 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3097 ocfs2_unlink_subtree(handle, et, left_path, path,
3098 subtree_index, dealloc);
3099 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
3106 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3107 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3110 * 'path' is also the leftmost path which
3111 * means it must be the only one. This gets
3112 * handled differently because we want to
3113 * revert the root back to having extents
3116 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3118 el = et->et_root_el;
3119 el->l_tree_depth = 0;
3120 el->l_next_free_rec = 0;
3121 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3123 ocfs2_et_set_last_eb_blk(et, 0);
3126 ocfs2_journal_dirty(handle, path_root_bh(path));
3129 ocfs2_free_path(left_path);
3134 * Left rotation of btree records.
3136 * In many ways, this is (unsurprisingly) the opposite of right
3137 * rotation. We start at some non-rightmost path containing an empty
3138 * extent in the leaf block. The code works its way to the rightmost
3139 * path by rotating records to the left in every subtree.
3141 * This is used by any code which reduces the number of extent records
3142 * in a leaf. After removal, an empty record should be placed in the
3143 * leftmost list position.
3145 * This won't handle a length update of the rightmost path records if
3146 * the rightmost tree leaf record is removed so the caller is
3147 * responsible for detecting and correcting that.
3149 static int ocfs2_rotate_tree_left(handle_t *handle,
3150 struct ocfs2_extent_tree *et,
3151 struct ocfs2_path *path,
3152 struct ocfs2_cached_dealloc_ctxt *dealloc)
3154 int ret, orig_credits = handle->h_buffer_credits;
3155 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3156 struct ocfs2_extent_block *eb;
3157 struct ocfs2_extent_list *el;
3159 el = path_leaf_el(path);
3160 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3163 if (path->p_tree_depth == 0) {
3164 rightmost_no_delete:
3166 * Inline extents. This is trivially handled, so do
3169 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3176 * Handle rightmost branch now. There's several cases:
3177 * 1) simple rotation leaving records in there. That's trivial.
3178 * 2) rotation requiring a branch delete - there's no more
3179 * records left. Two cases of this:
3180 * a) There are branches to the left.
3181 * b) This is also the leftmost (the only) branch.
3183 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3184 * 2a) we need the left branch so that we can update it with the unlink
3185 * 2b) we need to bring the root back to inline extents.
3188 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3190 if (eb->h_next_leaf_blk == 0) {
3192 * This gets a bit tricky if we're going to delete the
3193 * rightmost path. Get the other cases out of the way
3196 if (le16_to_cpu(el->l_next_free_rec) > 1)
3197 goto rightmost_no_delete;
3199 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3201 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3202 "Owner %llu has empty extent block at %llu",
3203 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3204 (unsigned long long)le64_to_cpu(eb->h_blkno));
3209 * XXX: The caller can not trust "path" any more after
3210 * this as it will have been deleted. What do we do?
3212 * In theory the rotate-for-merge code will never get
3213 * here because it'll always ask for a rotate in a
3217 ret = ocfs2_remove_rightmost_path(handle, et, path,
3225 * Now we can loop, remembering the path we get from -EAGAIN
3226 * and restarting from there.
3229 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3230 dealloc, &restart_path);
3231 if (ret && ret != -EAGAIN) {
3236 while (ret == -EAGAIN) {
3237 tmp_path = restart_path;
3238 restart_path = NULL;
3240 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3243 if (ret && ret != -EAGAIN) {
3248 ocfs2_free_path(tmp_path);
3256 ocfs2_free_path(tmp_path);
3257 ocfs2_free_path(restart_path);
3261 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3264 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3267 if (rec->e_leaf_clusters == 0) {
3269 * We consumed all of the merged-from record. An empty
3270 * extent cannot exist anywhere but the 1st array
3271 * position, so move things over if the merged-from
3272 * record doesn't occupy that position.
3274 * This creates a new empty extent so the caller
3275 * should be smart enough to have removed any existing
3279 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3280 size = index * sizeof(struct ocfs2_extent_rec);
3281 memmove(&el->l_recs[1], &el->l_recs[0], size);
3285 * Always memset - the caller doesn't check whether it
3286 * created an empty extent, so there could be junk in
3289 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3293 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3294 struct ocfs2_path *left_path,
3295 struct ocfs2_path **ret_right_path)
3299 struct ocfs2_path *right_path = NULL;
3300 struct ocfs2_extent_list *left_el;
3302 *ret_right_path = NULL;
3304 /* This function shouldn't be called for non-trees. */
3305 BUG_ON(left_path->p_tree_depth == 0);
3307 left_el = path_leaf_el(left_path);
3308 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3310 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3311 left_path, &right_cpos);
3317 /* This function shouldn't be called for the rightmost leaf. */
3318 BUG_ON(right_cpos == 0);
3320 right_path = ocfs2_new_path_from_path(left_path);
3327 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3333 *ret_right_path = right_path;
3336 ocfs2_free_path(right_path);
3341 * Remove split_rec clusters from the record at index and merge them
3342 * onto the beginning of the record "next" to it.
3343 * For index < l_count - 1, the next means the extent rec at index + 1.
3344 * For index == l_count - 1, the "next" means the 1st extent rec of the
3345 * next extent block.
3347 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3349 struct ocfs2_extent_tree *et,
3350 struct ocfs2_extent_rec *split_rec,
3353 int ret, next_free, i;
3354 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3355 struct ocfs2_extent_rec *left_rec;
3356 struct ocfs2_extent_rec *right_rec;
3357 struct ocfs2_extent_list *right_el;
3358 struct ocfs2_path *right_path = NULL;
3359 int subtree_index = 0;
3360 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3361 struct buffer_head *bh = path_leaf_bh(left_path);
3362 struct buffer_head *root_bh = NULL;
3364 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3365 left_rec = &el->l_recs[index];
3367 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3368 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3369 /* we meet with a cross extent block merge. */
3370 ret = ocfs2_get_right_path(et, left_path, &right_path);
3376 right_el = path_leaf_el(right_path);
3377 next_free = le16_to_cpu(right_el->l_next_free_rec);
3378 BUG_ON(next_free <= 0);
3379 right_rec = &right_el->l_recs[0];
3380 if (ocfs2_is_empty_extent(right_rec)) {
3381 BUG_ON(next_free <= 1);
3382 right_rec = &right_el->l_recs[1];
3385 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3386 le16_to_cpu(left_rec->e_leaf_clusters) !=
3387 le32_to_cpu(right_rec->e_cpos));
3389 subtree_index = ocfs2_find_subtree_root(et, left_path,
3392 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3393 handle->h_buffer_credits,
3400 root_bh = left_path->p_node[subtree_index].bh;
3401 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3403 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3410 for (i = subtree_index + 1;
3411 i < path_num_items(right_path); i++) {
3412 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3419 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3428 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3429 right_rec = &el->l_recs[index + 1];
3432 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3433 path_num_items(left_path) - 1);
3439 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3441 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3442 le64_add_cpu(&right_rec->e_blkno,
3443 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3445 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3447 ocfs2_cleanup_merge(el, index);
3449 ocfs2_journal_dirty(handle, bh);
3451 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3452 ocfs2_complete_edge_insert(handle, left_path, right_path,
3456 ocfs2_free_path(right_path);
3460 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3461 struct ocfs2_path *right_path,
3462 struct ocfs2_path **ret_left_path)
3466 struct ocfs2_path *left_path = NULL;
3468 *ret_left_path = NULL;
3470 /* This function shouldn't be called for non-trees. */
3471 BUG_ON(right_path->p_tree_depth == 0);
3473 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3474 right_path, &left_cpos);
3480 /* This function shouldn't be called for the leftmost leaf. */
3481 BUG_ON(left_cpos == 0);
3483 left_path = ocfs2_new_path_from_path(right_path);
3490 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3496 *ret_left_path = left_path;
3499 ocfs2_free_path(left_path);
3504 * Remove split_rec clusters from the record at index and merge them
3505 * onto the tail of the record "before" it.
3506 * For index > 0, the "before" means the extent rec at index - 1.
3508 * For index == 0, the "before" means the last record of the previous
3509 * extent block. And there is also a situation that we may need to
3510 * remove the rightmost leaf extent block in the right_path and change
3511 * the right path to indicate the new rightmost path.
3513 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3515 struct ocfs2_extent_tree *et,
3516 struct ocfs2_extent_rec *split_rec,
3517 struct ocfs2_cached_dealloc_ctxt *dealloc,
3520 int ret, i, subtree_index = 0, has_empty_extent = 0;
3521 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3522 struct ocfs2_extent_rec *left_rec;
3523 struct ocfs2_extent_rec *right_rec;
3524 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3525 struct buffer_head *bh = path_leaf_bh(right_path);
3526 struct buffer_head *root_bh = NULL;
3527 struct ocfs2_path *left_path = NULL;
3528 struct ocfs2_extent_list *left_el;
3532 right_rec = &el->l_recs[index];
3534 /* we meet with a cross extent block merge. */
3535 ret = ocfs2_get_left_path(et, right_path, &left_path);
3541 left_el = path_leaf_el(left_path);
3542 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3543 le16_to_cpu(left_el->l_count));
3545 left_rec = &left_el->l_recs[
3546 le16_to_cpu(left_el->l_next_free_rec) - 1];
3547 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3548 le16_to_cpu(left_rec->e_leaf_clusters) !=
3549 le32_to_cpu(split_rec->e_cpos));
3551 subtree_index = ocfs2_find_subtree_root(et, left_path,
3554 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3555 handle->h_buffer_credits,
3562 root_bh = left_path->p_node[subtree_index].bh;
3563 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3565 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3572 for (i = subtree_index + 1;
3573 i < path_num_items(right_path); i++) {
3574 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3581 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3589 left_rec = &el->l_recs[index - 1];
3590 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3591 has_empty_extent = 1;
3594 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3595 path_num_items(right_path) - 1);
3601 if (has_empty_extent && index == 1) {
3603 * The easy case - we can just plop the record right in.
3605 *left_rec = *split_rec;
3607 has_empty_extent = 0;
3609 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3611 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3612 le64_add_cpu(&right_rec->e_blkno,
3613 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3615 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3617 ocfs2_cleanup_merge(el, index);
3619 ocfs2_journal_dirty(handle, bh);
3621 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3624 * In the situation that the right_rec is empty and the extent
3625 * block is empty also, ocfs2_complete_edge_insert can't handle
3626 * it and we need to delete the right extent block.
3628 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3629 le16_to_cpu(el->l_next_free_rec) == 1) {
3631 ret = ocfs2_remove_rightmost_path(handle, et,
3639 /* Now the rightmost extent block has been deleted.
3640 * So we use the new rightmost path.
3642 ocfs2_mv_path(right_path, left_path);
3645 ocfs2_complete_edge_insert(handle, left_path,
3646 right_path, subtree_index);
3649 ocfs2_free_path(left_path);
3653 static int ocfs2_try_to_merge_extent(handle_t *handle,
3654 struct ocfs2_extent_tree *et,
3655 struct ocfs2_path *path,
3657 struct ocfs2_extent_rec *split_rec,
3658 struct ocfs2_cached_dealloc_ctxt *dealloc,
3659 struct ocfs2_merge_ctxt *ctxt)
3662 struct ocfs2_extent_list *el = path_leaf_el(path);
3663 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3665 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3667 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3669 * The merge code will need to create an empty
3670 * extent to take the place of the newly
3671 * emptied slot. Remove any pre-existing empty
3672 * extents - having more than one in a leaf is
3675 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3681 rec = &el->l_recs[split_index];
3684 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3686 * Left-right contig implies this.
3688 BUG_ON(!ctxt->c_split_covers_rec);
3691 * Since the leftright insert always covers the entire
3692 * extent, this call will delete the insert record
3693 * entirely, resulting in an empty extent record added to
3696 * Since the adding of an empty extent shifts
3697 * everything back to the right, there's no need to
3698 * update split_index here.
3700 * When the split_index is zero, we need to merge it to the
3701 * prevoius extent block. It is more efficient and easier
3702 * if we do merge_right first and merge_left later.
3704 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3712 * We can only get this from logic error above.
3714 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3716 /* The merge left us with an empty extent, remove it. */
3717 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3723 rec = &el->l_recs[split_index];
3726 * Note that we don't pass split_rec here on purpose -
3727 * we've merged it into the rec already.
3729 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3730 dealloc, split_index);
3737 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3739 * Error from this last rotate is not critical, so
3740 * print but don't bubble it up.
3747 * Merge a record to the left or right.
3749 * 'contig_type' is relative to the existing record,
3750 * so for example, if we're "right contig", it's to
3751 * the record on the left (hence the left merge).
3753 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3754 ret = ocfs2_merge_rec_left(path, handle, et,
3762 ret = ocfs2_merge_rec_right(path, handle,
3771 if (ctxt->c_split_covers_rec) {
3773 * The merge may have left an empty extent in
3774 * our leaf. Try to rotate it away.
3776 ret = ocfs2_rotate_tree_left(handle, et, path,
3788 static void ocfs2_subtract_from_rec(struct super_block *sb,
3789 enum ocfs2_split_type split,
3790 struct ocfs2_extent_rec *rec,
3791 struct ocfs2_extent_rec *split_rec)
3795 len_blocks = ocfs2_clusters_to_blocks(sb,
3796 le16_to_cpu(split_rec->e_leaf_clusters));
3798 if (split == SPLIT_LEFT) {
3800 * Region is on the left edge of the existing
3803 le32_add_cpu(&rec->e_cpos,
3804 le16_to_cpu(split_rec->e_leaf_clusters));
3805 le64_add_cpu(&rec->e_blkno, len_blocks);
3806 le16_add_cpu(&rec->e_leaf_clusters,
3807 -le16_to_cpu(split_rec->e_leaf_clusters));
3810 * Region is on the right edge of the existing
3813 le16_add_cpu(&rec->e_leaf_clusters,
3814 -le16_to_cpu(split_rec->e_leaf_clusters));
3819 * Do the final bits of extent record insertion at the target leaf
3820 * list. If this leaf is part of an allocation tree, it is assumed
3821 * that the tree above has been prepared.
3823 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3824 struct ocfs2_extent_rec *insert_rec,
3825 struct ocfs2_extent_list *el,
3826 struct ocfs2_insert_type *insert)
3828 int i = insert->ins_contig_index;
3830 struct ocfs2_extent_rec *rec;
3832 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3834 if (insert->ins_split != SPLIT_NONE) {
3835 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3837 rec = &el->l_recs[i];
3838 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3839 insert->ins_split, rec,
3845 * Contiguous insert - either left or right.
3847 if (insert->ins_contig != CONTIG_NONE) {
3848 rec = &el->l_recs[i];
3849 if (insert->ins_contig == CONTIG_LEFT) {
3850 rec->e_blkno = insert_rec->e_blkno;
3851 rec->e_cpos = insert_rec->e_cpos;
3853 le16_add_cpu(&rec->e_leaf_clusters,
3854 le16_to_cpu(insert_rec->e_leaf_clusters));
3859 * Handle insert into an empty leaf.
3861 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3862 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3863 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3864 el->l_recs[0] = *insert_rec;
3865 el->l_next_free_rec = cpu_to_le16(1);
3872 if (insert->ins_appending == APPEND_TAIL) {
3873 i = le16_to_cpu(el->l_next_free_rec) - 1;
3874 rec = &el->l_recs[i];
3875 range = le32_to_cpu(rec->e_cpos)
3876 + le16_to_cpu(rec->e_leaf_clusters);
3877 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3879 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3880 le16_to_cpu(el->l_count),
3881 "owner %llu, depth %u, count %u, next free %u, "
3882 "rec.cpos %u, rec.clusters %u, "
3883 "insert.cpos %u, insert.clusters %u\n",
3884 ocfs2_metadata_cache_owner(et->et_ci),
3885 le16_to_cpu(el->l_tree_depth),
3886 le16_to_cpu(el->l_count),
3887 le16_to_cpu(el->l_next_free_rec),
3888 le32_to_cpu(el->l_recs[i].e_cpos),
3889 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3890 le32_to_cpu(insert_rec->e_cpos),
3891 le16_to_cpu(insert_rec->e_leaf_clusters));
3893 el->l_recs[i] = *insert_rec;
3894 le16_add_cpu(&el->l_next_free_rec, 1);
3900 * Ok, we have to rotate.
3902 * At this point, it is safe to assume that inserting into an
3903 * empty leaf and appending to a leaf have both been handled
3906 * This leaf needs to have space, either by the empty 1st
3907 * extent record, or by virtue of an l_next_rec < l_count.
3909 ocfs2_rotate_leaf(el, insert_rec);
3912 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3913 struct ocfs2_extent_tree *et,
3914 struct ocfs2_path *path,
3915 struct ocfs2_extent_rec *insert_rec)
3917 int ret, i, next_free;
3918 struct buffer_head *bh;
3919 struct ocfs2_extent_list *el;
3920 struct ocfs2_extent_rec *rec;
3923 * Update everything except the leaf block.
3925 for (i = 0; i < path->p_tree_depth; i++) {
3926 bh = path->p_node[i].bh;
3927 el = path->p_node[i].el;
3929 next_free = le16_to_cpu(el->l_next_free_rec);
3930 if (next_free == 0) {
3931 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3932 "Owner %llu has a bad extent list",
3933 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3938 rec = &el->l_recs[next_free - 1];
3940 rec->e_int_clusters = insert_rec->e_cpos;
3941 le32_add_cpu(&rec->e_int_clusters,
3942 le16_to_cpu(insert_rec->e_leaf_clusters));
3943 le32_add_cpu(&rec->e_int_clusters,
3944 -le32_to_cpu(rec->e_cpos));
3946 ocfs2_journal_dirty(handle, bh);
3950 static int ocfs2_append_rec_to_path(handle_t *handle,
3951 struct ocfs2_extent_tree *et,
3952 struct ocfs2_extent_rec *insert_rec,
3953 struct ocfs2_path *right_path,
3954 struct ocfs2_path **ret_left_path)
3957 struct ocfs2_extent_list *el;
3958 struct ocfs2_path *left_path = NULL;
3960 *ret_left_path = NULL;
3963 * This shouldn't happen for non-trees. The extent rec cluster
3964 * count manipulation below only works for interior nodes.
3966 BUG_ON(right_path->p_tree_depth == 0);
3969 * If our appending insert is at the leftmost edge of a leaf,
3970 * then we might need to update the rightmost records of the
3973 el = path_leaf_el(right_path);
3974 next_free = le16_to_cpu(el->l_next_free_rec);
3975 if (next_free == 0 ||
3976 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3979 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3980 right_path, &left_cpos);
3986 trace_ocfs2_append_rec_to_path(
3987 (unsigned long long)
3988 ocfs2_metadata_cache_owner(et->et_ci),
3989 le32_to_cpu(insert_rec->e_cpos),
3993 * No need to worry if the append is already in the
3997 left_path = ocfs2_new_path_from_path(right_path);
4004 ret = ocfs2_find_path(et->et_ci, left_path,
4012 * ocfs2_insert_path() will pass the left_path to the
4018 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4024 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4026 *ret_left_path = left_path;
4030 ocfs2_free_path(left_path);
4035 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4036 struct ocfs2_path *left_path,
4037 struct ocfs2_path *right_path,
4038 struct ocfs2_extent_rec *split_rec,
4039 enum ocfs2_split_type split)
4042 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4043 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4044 struct ocfs2_extent_rec *rec, *tmprec;
4046 right_el = path_leaf_el(right_path);
4048 left_el = path_leaf_el(left_path);
4051 insert_el = right_el;
4052 index = ocfs2_search_extent_list(el, cpos);
4054 if (index == 0 && left_path) {
4055 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4058 * This typically means that the record
4059 * started in the left path but moved to the
4060 * right as a result of rotation. We either
4061 * move the existing record to the left, or we
4062 * do the later insert there.
4064 * In this case, the left path should always
4065 * exist as the rotate code will have passed
4066 * it back for a post-insert update.
4069 if (split == SPLIT_LEFT) {
4071 * It's a left split. Since we know
4072 * that the rotate code gave us an
4073 * empty extent in the left path, we
4074 * can just do the insert there.
4076 insert_el = left_el;
4079 * Right split - we have to move the
4080 * existing record over to the left
4081 * leaf. The insert will be into the
4082 * newly created empty extent in the
4085 tmprec = &right_el->l_recs[index];
4086 ocfs2_rotate_leaf(left_el, tmprec);
4089 memset(tmprec, 0, sizeof(*tmprec));
4090 index = ocfs2_search_extent_list(left_el, cpos);
4091 BUG_ON(index == -1);
4096 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4098 * Left path is easy - we can just allow the insert to
4102 insert_el = left_el;
4103 index = ocfs2_search_extent_list(el, cpos);
4104 BUG_ON(index == -1);
4107 rec = &el->l_recs[index];
4108 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4109 split, rec, split_rec);
4110 ocfs2_rotate_leaf(insert_el, split_rec);
4114 * This function only does inserts on an allocation b-tree. For tree
4115 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4117 * right_path is the path we want to do the actual insert
4118 * in. left_path should only be passed in if we need to update that
4119 * portion of the tree after an edge insert.
4121 static int ocfs2_insert_path(handle_t *handle,
4122 struct ocfs2_extent_tree *et,
4123 struct ocfs2_path *left_path,
4124 struct ocfs2_path *right_path,
4125 struct ocfs2_extent_rec *insert_rec,
4126 struct ocfs2_insert_type *insert)
4128 int ret, subtree_index;
4129 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4133 * There's a chance that left_path got passed back to
4134 * us without being accounted for in the
4135 * journal. Extend our transaction here to be sure we
4136 * can change those blocks.
4138 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4144 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4152 * Pass both paths to the journal. The majority of inserts
4153 * will be touching all components anyway.
4155 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4161 if (insert->ins_split != SPLIT_NONE) {
4163 * We could call ocfs2_insert_at_leaf() for some types
4164 * of splits, but it's easier to just let one separate
4165 * function sort it all out.
4167 ocfs2_split_record(et, left_path, right_path,
4168 insert_rec, insert->ins_split);
4171 * Split might have modified either leaf and we don't
4172 * have a guarantee that the later edge insert will
4173 * dirty this for us.
4176 ocfs2_journal_dirty(handle,
4177 path_leaf_bh(left_path));
4179 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4182 ocfs2_journal_dirty(handle, leaf_bh);
4186 * The rotate code has indicated that we need to fix
4187 * up portions of the tree after the insert.
4189 * XXX: Should we extend the transaction here?
4191 subtree_index = ocfs2_find_subtree_root(et, left_path,
4193 ocfs2_complete_edge_insert(handle, left_path, right_path,
4202 static int ocfs2_do_insert_extent(handle_t *handle,
4203 struct ocfs2_extent_tree *et,
4204 struct ocfs2_extent_rec *insert_rec,
4205 struct ocfs2_insert_type *type)
4207 int ret, rotate = 0;
4209 struct ocfs2_path *right_path = NULL;
4210 struct ocfs2_path *left_path = NULL;
4211 struct ocfs2_extent_list *el;
4213 el = et->et_root_el;
4215 ret = ocfs2_et_root_journal_access(handle, et,
4216 OCFS2_JOURNAL_ACCESS_WRITE);
4222 if (le16_to_cpu(el->l_tree_depth) == 0) {
4223 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4224 goto out_update_clusters;
4227 right_path = ocfs2_new_path_from_et(et);
4235 * Determine the path to start with. Rotations need the
4236 * rightmost path, everything else can go directly to the
4239 cpos = le32_to_cpu(insert_rec->e_cpos);
4240 if (type->ins_appending == APPEND_NONE &&
4241 type->ins_contig == CONTIG_NONE) {
4246 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4253 * Rotations and appends need special treatment - they modify
4254 * parts of the tree's above them.
4256 * Both might pass back a path immediate to the left of the
4257 * one being inserted to. This will be cause
4258 * ocfs2_insert_path() to modify the rightmost records of
4259 * left_path to account for an edge insert.
4261 * XXX: When modifying this code, keep in mind that an insert
4262 * can wind up skipping both of these two special cases...
4265 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4266 le32_to_cpu(insert_rec->e_cpos),
4267 right_path, &left_path);
4274 * ocfs2_rotate_tree_right() might have extended the
4275 * transaction without re-journaling our tree root.
4277 ret = ocfs2_et_root_journal_access(handle, et,
4278 OCFS2_JOURNAL_ACCESS_WRITE);
4283 } else if (type->ins_appending == APPEND_TAIL
4284 && type->ins_contig != CONTIG_LEFT) {
4285 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4286 right_path, &left_path);
4293 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4300 out_update_clusters:
4301 if (type->ins_split == SPLIT_NONE)
4302 ocfs2_et_update_clusters(et,
4303 le16_to_cpu(insert_rec->e_leaf_clusters));
4305 ocfs2_journal_dirty(handle, et->et_root_bh);
4308 ocfs2_free_path(left_path);
4309 ocfs2_free_path(right_path);
4314 static enum ocfs2_contig_type
4315 ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4316 struct ocfs2_path *path,
4317 struct ocfs2_extent_list *el, int index,
4318 struct ocfs2_extent_rec *split_rec)
4321 enum ocfs2_contig_type ret = CONTIG_NONE;
4322 u32 left_cpos, right_cpos;
4323 struct ocfs2_extent_rec *rec = NULL;
4324 struct ocfs2_extent_list *new_el;
4325 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4326 struct buffer_head *bh;
4327 struct ocfs2_extent_block *eb;
4328 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4331 rec = &el->l_recs[index - 1];
4332 } else if (path->p_tree_depth > 0) {
4333 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4337 if (left_cpos != 0) {
4338 left_path = ocfs2_new_path_from_path(path);
4342 status = ocfs2_find_path(et->et_ci, left_path,
4347 new_el = path_leaf_el(left_path);
4349 if (le16_to_cpu(new_el->l_next_free_rec) !=
4350 le16_to_cpu(new_el->l_count)) {
4351 bh = path_leaf_bh(left_path);
4352 eb = (struct ocfs2_extent_block *)bh->b_data;
4354 "Extent block #%llu has an "
4355 "invalid l_next_free_rec of "
4356 "%d. It should have "
4357 "matched the l_count of %d",
4358 (unsigned long long)le64_to_cpu(eb->h_blkno),
4359 le16_to_cpu(new_el->l_next_free_rec),
4360 le16_to_cpu(new_el->l_count));
4364 rec = &new_el->l_recs[
4365 le16_to_cpu(new_el->l_next_free_rec) - 1];
4370 * We're careful to check for an empty extent record here -
4371 * the merge code will know what to do if it sees one.
4374 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4375 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4378 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4383 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4384 rec = &el->l_recs[index + 1];
4385 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4386 path->p_tree_depth > 0) {
4387 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4391 if (right_cpos == 0)
4394 right_path = ocfs2_new_path_from_path(path);
4398 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4402 new_el = path_leaf_el(right_path);
4403 rec = &new_el->l_recs[0];
4404 if (ocfs2_is_empty_extent(rec)) {
4405 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4406 bh = path_leaf_bh(right_path);
4407 eb = (struct ocfs2_extent_block *)bh->b_data;
4409 "Extent block #%llu has an "
4410 "invalid l_next_free_rec of %d",
4411 (unsigned long long)le64_to_cpu(eb->h_blkno),
4412 le16_to_cpu(new_el->l_next_free_rec));
4416 rec = &new_el->l_recs[1];
4421 enum ocfs2_contig_type contig_type;
4423 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4425 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4426 ret = CONTIG_LEFTRIGHT;
4427 else if (ret == CONTIG_NONE)
4432 ocfs2_free_path(left_path);
4433 ocfs2_free_path(right_path);
4437 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4438 struct ocfs2_insert_type *insert,
4439 struct ocfs2_extent_list *el,
4440 struct ocfs2_extent_rec *insert_rec)
4443 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4445 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4447 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4448 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4450 if (contig_type != CONTIG_NONE) {
4451 insert->ins_contig_index = i;
4455 insert->ins_contig = contig_type;
4457 if (insert->ins_contig != CONTIG_NONE) {
4458 struct ocfs2_extent_rec *rec =
4459 &el->l_recs[insert->ins_contig_index];
4460 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4461 le16_to_cpu(insert_rec->e_leaf_clusters);
4464 * Caller might want us to limit the size of extents, don't
4465 * calculate contiguousness if we might exceed that limit.
4467 if (et->et_max_leaf_clusters &&
4468 (len > et->et_max_leaf_clusters))
4469 insert->ins_contig = CONTIG_NONE;
4474 * This should only be called against the righmost leaf extent list.
4476 * ocfs2_figure_appending_type() will figure out whether we'll have to
4477 * insert at the tail of the rightmost leaf.
4479 * This should also work against the root extent list for tree's with 0
4480 * depth. If we consider the root extent list to be the rightmost leaf node
4481 * then the logic here makes sense.
4483 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4484 struct ocfs2_extent_list *el,
4485 struct ocfs2_extent_rec *insert_rec)
4488 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4489 struct ocfs2_extent_rec *rec;
4491 insert->ins_appending = APPEND_NONE;
4493 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4495 if (!el->l_next_free_rec)
4496 goto set_tail_append;
4498 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4499 /* Were all records empty? */
4500 if (le16_to_cpu(el->l_next_free_rec) == 1)
4501 goto set_tail_append;
4504 i = le16_to_cpu(el->l_next_free_rec) - 1;
4505 rec = &el->l_recs[i];
4508 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4509 goto set_tail_append;
4514 insert->ins_appending = APPEND_TAIL;
4518 * Helper function called at the beginning of an insert.
4520 * This computes a few things that are commonly used in the process of
4521 * inserting into the btree:
4522 * - Whether the new extent is contiguous with an existing one.
4523 * - The current tree depth.
4524 * - Whether the insert is an appending one.
4525 * - The total # of free records in the tree.
4527 * All of the information is stored on the ocfs2_insert_type
4530 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4531 struct buffer_head **last_eb_bh,
4532 struct ocfs2_extent_rec *insert_rec,
4534 struct ocfs2_insert_type *insert)
4537 struct ocfs2_extent_block *eb;
4538 struct ocfs2_extent_list *el;
4539 struct ocfs2_path *path = NULL;
4540 struct buffer_head *bh = NULL;
4542 insert->ins_split = SPLIT_NONE;
4544 el = et->et_root_el;
4545 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4547 if (el->l_tree_depth) {
4549 * If we have tree depth, we read in the
4550 * rightmost extent block ahead of time as
4551 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4552 * may want it later.
4554 ret = ocfs2_read_extent_block(et->et_ci,
4555 ocfs2_et_get_last_eb_blk(et),
4561 eb = (struct ocfs2_extent_block *) bh->b_data;
4566 * Unless we have a contiguous insert, we'll need to know if
4567 * there is room left in our allocation tree for another
4570 * XXX: This test is simplistic, we can search for empty
4571 * extent records too.
4573 *free_records = le16_to_cpu(el->l_count) -
4574 le16_to_cpu(el->l_next_free_rec);
4576 if (!insert->ins_tree_depth) {
4577 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4578 ocfs2_figure_appending_type(insert, el, insert_rec);
4582 path = ocfs2_new_path_from_et(et);
4590 * In the case that we're inserting past what the tree
4591 * currently accounts for, ocfs2_find_path() will return for
4592 * us the rightmost tree path. This is accounted for below in
4593 * the appending code.
4595 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4601 el = path_leaf_el(path);
4604 * Now that we have the path, there's two things we want to determine:
4605 * 1) Contiguousness (also set contig_index if this is so)
4607 * 2) Are we doing an append? We can trivially break this up
4608 * into two types of appends: simple record append, or a
4609 * rotate inside the tail leaf.
4611 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4614 * The insert code isn't quite ready to deal with all cases of
4615 * left contiguousness. Specifically, if it's an insert into
4616 * the 1st record in a leaf, it will require the adjustment of
4617 * cluster count on the last record of the path directly to it's
4618 * left. For now, just catch that case and fool the layers
4619 * above us. This works just fine for tree_depth == 0, which
4620 * is why we allow that above.
4622 if (insert->ins_contig == CONTIG_LEFT &&
4623 insert->ins_contig_index == 0)
4624 insert->ins_contig = CONTIG_NONE;
4627 * Ok, so we can simply compare against last_eb to figure out
4628 * whether the path doesn't exist. This will only happen in
4629 * the case that we're doing a tail append, so maybe we can
4630 * take advantage of that information somehow.
4632 if (ocfs2_et_get_last_eb_blk(et) ==
4633 path_leaf_bh(path)->b_blocknr) {
4635 * Ok, ocfs2_find_path() returned us the rightmost
4636 * tree path. This might be an appending insert. There are
4638 * 1) We're doing a true append at the tail:
4639 * -This might even be off the end of the leaf
4640 * 2) We're "appending" by rotating in the tail
4642 ocfs2_figure_appending_type(insert, el, insert_rec);
4646 ocfs2_free_path(path);
4656 * Insert an extent into a btree.
4658 * The caller needs to update the owning btree's cluster count.
4660 int ocfs2_insert_extent(handle_t *handle,
4661 struct ocfs2_extent_tree *et,
4666 struct ocfs2_alloc_context *meta_ac)
4669 int uninitialized_var(free_records);
4670 struct buffer_head *last_eb_bh = NULL;
4671 struct ocfs2_insert_type insert = {0, };
4672 struct ocfs2_extent_rec rec;
4674 trace_ocfs2_insert_extent_start(
4675 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4676 cpos, new_clusters);
4678 memset(&rec, 0, sizeof(rec));
4679 rec.e_cpos = cpu_to_le32(cpos);
4680 rec.e_blkno = cpu_to_le64(start_blk);
4681 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4682 rec.e_flags = flags;
4683 status = ocfs2_et_insert_check(et, &rec);
4689 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4690 &free_records, &insert);
4696 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4697 insert.ins_contig_index, free_records,
4698 insert.ins_tree_depth);
4700 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4701 status = ocfs2_grow_tree(handle, et,
4702 &insert.ins_tree_depth, &last_eb_bh,
4710 /* Finally, we can add clusters. This might rotate the tree for us. */
4711 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4715 ocfs2_et_extent_map_insert(et, &rec);
4724 * Allcate and add clusters into the extent b-tree.
4725 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4726 * The extent b-tree's root is specified by et, and
4727 * it is not limited to the file storage. Any extent tree can use this
4728 * function if it implements the proper ocfs2_extent_tree.
4730 int ocfs2_add_clusters_in_btree(handle_t *handle,
4731 struct ocfs2_extent_tree *et,
4732 u32 *logical_offset,
4733 u32 clusters_to_add,
4735 struct ocfs2_alloc_context *data_ac,
4736 struct ocfs2_alloc_context *meta_ac,
4737 enum ocfs2_alloc_restarted *reason_ret)
4739 int status = 0, err = 0;
4742 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4743 u32 bit_off, num_bits;
4746 struct ocfs2_super *osb =
4747 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4749 BUG_ON(!clusters_to_add);
4752 flags = OCFS2_EXT_UNWRITTEN;
4754 free_extents = ocfs2_num_free_extents(osb, et);
4755 if (free_extents < 0) {
4756 status = free_extents;
4761 /* there are two cases which could cause us to EAGAIN in the
4762 * we-need-more-metadata case:
4763 * 1) we haven't reserved *any*
4764 * 2) we are so fragmented, we've needed to add metadata too
4766 if (!free_extents && !meta_ac) {
4769 reason = RESTART_META;
4771 } else if ((!free_extents)
4772 && (ocfs2_alloc_context_bits_left(meta_ac)
4773 < ocfs2_extend_meta_needed(et->et_root_el))) {
4776 reason = RESTART_META;
4780 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4781 clusters_to_add, &bit_off, &num_bits);
4783 if (status != -ENOSPC)
4788 BUG_ON(num_bits > clusters_to_add);
4790 /* reserve our write early -- insert_extent may update the tree root */
4791 status = ocfs2_et_root_journal_access(handle, et,
4792 OCFS2_JOURNAL_ACCESS_WRITE);
4799 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4800 trace_ocfs2_add_clusters_in_btree(
4801 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4803 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4804 num_bits, flags, meta_ac);
4811 ocfs2_journal_dirty(handle, et->et_root_bh);
4813 clusters_to_add -= num_bits;
4814 *logical_offset += num_bits;
4816 if (clusters_to_add) {
4817 err = clusters_to_add;
4819 reason = RESTART_TRANS;
4824 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4825 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4828 ocfs2_free_clusters(handle,
4831 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4837 *reason_ret = reason;
4838 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4842 static void ocfs2_make_right_split_rec(struct super_block *sb,
4843 struct ocfs2_extent_rec *split_rec,
4845 struct ocfs2_extent_rec *rec)
4847 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4848 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4850 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4852 split_rec->e_cpos = cpu_to_le32(cpos);
4853 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4855 split_rec->e_blkno = rec->e_blkno;
4856 le64_add_cpu(&split_rec->e_blkno,
4857 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4859 split_rec->e_flags = rec->e_flags;
4862 static int ocfs2_split_and_insert(handle_t *handle,
4863 struct ocfs2_extent_tree *et,
4864 struct ocfs2_path *path,
4865 struct buffer_head **last_eb_bh,
4867 struct ocfs2_extent_rec *orig_split_rec,
4868 struct ocfs2_alloc_context *meta_ac)
4871 unsigned int insert_range, rec_range, do_leftright = 0;
4872 struct ocfs2_extent_rec tmprec;
4873 struct ocfs2_extent_list *rightmost_el;
4874 struct ocfs2_extent_rec rec;
4875 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4876 struct ocfs2_insert_type insert;
4877 struct ocfs2_extent_block *eb;
4881 * Store a copy of the record on the stack - it might move
4882 * around as the tree is manipulated below.
4884 rec = path_leaf_el(path)->l_recs[split_index];
4886 rightmost_el = et->et_root_el;
4888 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4890 BUG_ON(!(*last_eb_bh));
4891 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4892 rightmost_el = &eb->h_list;
4895 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4896 le16_to_cpu(rightmost_el->l_count)) {
4897 ret = ocfs2_grow_tree(handle, et,
4898 &depth, last_eb_bh, meta_ac);
4905 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4906 insert.ins_appending = APPEND_NONE;
4907 insert.ins_contig = CONTIG_NONE;
4908 insert.ins_tree_depth = depth;
4910 insert_range = le32_to_cpu(split_rec.e_cpos) +
4911 le16_to_cpu(split_rec.e_leaf_clusters);
4912 rec_range = le32_to_cpu(rec.e_cpos) +
4913 le16_to_cpu(rec.e_leaf_clusters);
4915 if (split_rec.e_cpos == rec.e_cpos) {
4916 insert.ins_split = SPLIT_LEFT;
4917 } else if (insert_range == rec_range) {
4918 insert.ins_split = SPLIT_RIGHT;
4921 * Left/right split. We fake this as a right split
4922 * first and then make a second pass as a left split.
4924 insert.ins_split = SPLIT_RIGHT;
4926 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4927 &tmprec, insert_range, &rec);
4931 BUG_ON(do_leftright);
4935 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4941 if (do_leftright == 1) {
4943 struct ocfs2_extent_list *el;
4946 split_rec = *orig_split_rec;
4948 ocfs2_reinit_path(path, 1);
4950 cpos = le32_to_cpu(split_rec.e_cpos);
4951 ret = ocfs2_find_path(et->et_ci, path, cpos);
4957 el = path_leaf_el(path);
4958 split_index = ocfs2_search_extent_list(el, cpos);
4959 if (split_index == -1) {
4960 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
4961 "Owner %llu has an extent at cpos %u "
4962 "which can no longer be found.\n",
4963 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4975 static int ocfs2_replace_extent_rec(handle_t *handle,
4976 struct ocfs2_extent_tree *et,
4977 struct ocfs2_path *path,
4978 struct ocfs2_extent_list *el,
4980 struct ocfs2_extent_rec *split_rec)
4984 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
4985 path_num_items(path) - 1);
4991 el->l_recs[split_index] = *split_rec;
4993 ocfs2_journal_dirty(handle, path_leaf_bh(path));
4999 * Split part or all of the extent record at split_index in the leaf
5000 * pointed to by path. Merge with the contiguous extent record if needed.
5002 * Care is taken to handle contiguousness so as to not grow the tree.
5004 * meta_ac is not strictly necessary - we only truly need it if growth
5005 * of the tree is required. All other cases will degrade into a less
5006 * optimal tree layout.
5008 * last_eb_bh should be the rightmost leaf block for any extent
5009 * btree. Since a split may grow the tree or a merge might shrink it,
5010 * the caller cannot trust the contents of that buffer after this call.
5012 * This code is optimized for readability - several passes might be
5013 * made over certain portions of the tree. All of those blocks will
5014 * have been brought into cache (and pinned via the journal), so the
5015 * extra overhead is not expressed in terms of disk reads.
5017 int ocfs2_split_extent(handle_t *handle,
5018 struct ocfs2_extent_tree *et,
5019 struct ocfs2_path *path,
5021 struct ocfs2_extent_rec *split_rec,
5022 struct ocfs2_alloc_context *meta_ac,
5023 struct ocfs2_cached_dealloc_ctxt *dealloc)
5026 struct ocfs2_extent_list *el = path_leaf_el(path);
5027 struct buffer_head *last_eb_bh = NULL;
5028 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5029 struct ocfs2_merge_ctxt ctxt;
5030 struct ocfs2_extent_list *rightmost_el;
5032 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5033 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5034 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5040 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(et, path, el,
5045 * The core merge / split code wants to know how much room is
5046 * left in this allocation tree, so we pass the
5047 * rightmost extent list.
5049 if (path->p_tree_depth) {
5050 struct ocfs2_extent_block *eb;
5052 ret = ocfs2_read_extent_block(et->et_ci,
5053 ocfs2_et_get_last_eb_blk(et),
5060 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5061 rightmost_el = &eb->h_list;
5063 rightmost_el = path_root_el(path);
5065 if (rec->e_cpos == split_rec->e_cpos &&
5066 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5067 ctxt.c_split_covers_rec = 1;
5069 ctxt.c_split_covers_rec = 0;
5071 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5073 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5074 ctxt.c_has_empty_extent,
5075 ctxt.c_split_covers_rec);
5077 if (ctxt.c_contig_type == CONTIG_NONE) {
5078 if (ctxt.c_split_covers_rec)
5079 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5080 split_index, split_rec);
5082 ret = ocfs2_split_and_insert(handle, et, path,
5083 &last_eb_bh, split_index,
5084 split_rec, meta_ac);
5088 ret = ocfs2_try_to_merge_extent(handle, et, path,
5089 split_index, split_rec,
5101 * Change the flags of the already-existing extent at cpos for len clusters.
5103 * new_flags: the flags we want to set.
5104 * clear_flags: the flags we want to clear.
5105 * phys: the new physical offset we want this new extent starts from.
5107 * If the existing extent is larger than the request, initiate a
5108 * split. An attempt will be made at merging with adjacent extents.
5110 * The caller is responsible for passing down meta_ac if we'll need it.
5112 int ocfs2_change_extent_flag(handle_t *handle,
5113 struct ocfs2_extent_tree *et,
5114 u32 cpos, u32 len, u32 phys,
5115 struct ocfs2_alloc_context *meta_ac,
5116 struct ocfs2_cached_dealloc_ctxt *dealloc,
5117 int new_flags, int clear_flags)
5120 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5121 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5122 struct ocfs2_extent_rec split_rec;
5123 struct ocfs2_path *left_path = NULL;
5124 struct ocfs2_extent_list *el;
5125 struct ocfs2_extent_rec *rec;
5127 left_path = ocfs2_new_path_from_et(et);
5134 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5139 el = path_leaf_el(left_path);
5141 index = ocfs2_search_extent_list(el, cpos);
5144 "Owner %llu has an extent at cpos %u which can no "
5145 "longer be found.\n",
5146 (unsigned long long)
5147 ocfs2_metadata_cache_owner(et->et_ci), cpos);
5153 rec = &el->l_recs[index];
5154 if (new_flags && (rec->e_flags & new_flags)) {
5155 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5156 "extent that already had them",
5157 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5162 if (clear_flags && !(rec->e_flags & clear_flags)) {
5163 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5164 "extent that didn't have them",
5165 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5170 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5171 split_rec.e_cpos = cpu_to_le32(cpos);
5172 split_rec.e_leaf_clusters = cpu_to_le16(len);
5173 split_rec.e_blkno = cpu_to_le64(start_blkno);
5174 split_rec.e_flags = rec->e_flags;
5176 split_rec.e_flags |= new_flags;
5178 split_rec.e_flags &= ~clear_flags;
5180 ret = ocfs2_split_extent(handle, et, left_path,
5181 index, &split_rec, meta_ac,
5187 ocfs2_free_path(left_path);
5193 * Mark the already-existing extent at cpos as written for len clusters.
5194 * This removes the unwritten extent flag.
5196 * If the existing extent is larger than the request, initiate a
5197 * split. An attempt will be made at merging with adjacent extents.
5199 * The caller is responsible for passing down meta_ac if we'll need it.
5201 int ocfs2_mark_extent_written(struct inode *inode,
5202 struct ocfs2_extent_tree *et,
5203 handle_t *handle, u32 cpos, u32 len, u32 phys,
5204 struct ocfs2_alloc_context *meta_ac,
5205 struct ocfs2_cached_dealloc_ctxt *dealloc)
5209 trace_ocfs2_mark_extent_written(
5210 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5213 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5214 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
5215 "that are being written to, but the feature bit "
5216 "is not set in the super block.",
5217 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5223 * XXX: This should be fixed up so that we just re-insert the
5224 * next extent records.
5226 ocfs2_et_extent_map_truncate(et, 0);
5228 ret = ocfs2_change_extent_flag(handle, et, cpos,
5229 len, phys, meta_ac, dealloc,
5230 0, OCFS2_EXT_UNWRITTEN);
5238 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5239 struct ocfs2_path *path,
5240 int index, u32 new_range,
5241 struct ocfs2_alloc_context *meta_ac)
5243 int ret, depth, credits;
5244 struct buffer_head *last_eb_bh = NULL;
5245 struct ocfs2_extent_block *eb;
5246 struct ocfs2_extent_list *rightmost_el, *el;
5247 struct ocfs2_extent_rec split_rec;
5248 struct ocfs2_extent_rec *rec;
5249 struct ocfs2_insert_type insert;
5252 * Setup the record to split before we grow the tree.
5254 el = path_leaf_el(path);
5255 rec = &el->l_recs[index];
5256 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5257 &split_rec, new_range, rec);
5259 depth = path->p_tree_depth;
5261 ret = ocfs2_read_extent_block(et->et_ci,
5262 ocfs2_et_get_last_eb_blk(et),
5269 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5270 rightmost_el = &eb->h_list;
5272 rightmost_el = path_leaf_el(path);
5274 credits = path->p_tree_depth +
5275 ocfs2_extend_meta_needed(et->et_root_el);
5276 ret = ocfs2_extend_trans(handle, credits);
5282 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5283 le16_to_cpu(rightmost_el->l_count)) {
5284 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5292 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5293 insert.ins_appending = APPEND_NONE;
5294 insert.ins_contig = CONTIG_NONE;
5295 insert.ins_split = SPLIT_RIGHT;
5296 insert.ins_tree_depth = depth;
5298 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5307 static int ocfs2_truncate_rec(handle_t *handle,
5308 struct ocfs2_extent_tree *et,
5309 struct ocfs2_path *path, int index,
5310 struct ocfs2_cached_dealloc_ctxt *dealloc,
5314 u32 left_cpos, rec_range, trunc_range;
5315 int wants_rotate = 0, is_rightmost_tree_rec = 0;
5316 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5317 struct ocfs2_path *left_path = NULL;
5318 struct ocfs2_extent_list *el = path_leaf_el(path);
5319 struct ocfs2_extent_rec *rec;
5320 struct ocfs2_extent_block *eb;
5322 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5323 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5332 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5333 path->p_tree_depth) {
5335 * Check whether this is the rightmost tree record. If
5336 * we remove all of this record or part of its right
5337 * edge then an update of the record lengths above it
5340 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5341 if (eb->h_next_leaf_blk == 0)
5342 is_rightmost_tree_rec = 1;
5345 rec = &el->l_recs[index];
5346 if (index == 0 && path->p_tree_depth &&
5347 le32_to_cpu(rec->e_cpos) == cpos) {
5349 * Changing the leftmost offset (via partial or whole
5350 * record truncate) of an interior (or rightmost) path
5351 * means we have to update the subtree that is formed
5352 * by this leaf and the one to it's left.
5354 * There are two cases we can skip:
5355 * 1) Path is the leftmost one in our btree.
5356 * 2) The leaf is rightmost and will be empty after
5357 * we remove the extent record - the rotate code
5358 * knows how to update the newly formed edge.
5361 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5367 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5368 left_path = ocfs2_new_path_from_path(path);
5375 ret = ocfs2_find_path(et->et_ci, left_path,
5384 ret = ocfs2_extend_rotate_transaction(handle, 0,
5385 handle->h_buffer_credits,
5392 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5398 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5404 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5405 trunc_range = cpos + len;
5407 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5410 memset(rec, 0, sizeof(*rec));
5411 ocfs2_cleanup_merge(el, index);
5414 next_free = le16_to_cpu(el->l_next_free_rec);
5415 if (is_rightmost_tree_rec && next_free > 1) {
5417 * We skip the edge update if this path will
5418 * be deleted by the rotate code.
5420 rec = &el->l_recs[next_free - 1];
5421 ocfs2_adjust_rightmost_records(handle, et, path,
5424 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5425 /* Remove leftmost portion of the record. */
5426 le32_add_cpu(&rec->e_cpos, len);
5427 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5428 le16_add_cpu(&rec->e_leaf_clusters, -len);
5429 } else if (rec_range == trunc_range) {
5430 /* Remove rightmost portion of the record */
5431 le16_add_cpu(&rec->e_leaf_clusters, -len);
5432 if (is_rightmost_tree_rec)
5433 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5435 /* Caller should have trapped this. */
5436 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5438 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5439 le32_to_cpu(rec->e_cpos),
5440 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5447 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5448 ocfs2_complete_edge_insert(handle, left_path, path,
5452 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5454 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5461 ocfs2_free_path(left_path);
5465 int ocfs2_remove_extent(handle_t *handle,
5466 struct ocfs2_extent_tree *et,
5468 struct ocfs2_alloc_context *meta_ac,
5469 struct ocfs2_cached_dealloc_ctxt *dealloc)
5472 u32 rec_range, trunc_range;
5473 struct ocfs2_extent_rec *rec;
5474 struct ocfs2_extent_list *el;
5475 struct ocfs2_path *path = NULL;
5478 * XXX: Why are we truncating to 0 instead of wherever this
5481 ocfs2_et_extent_map_truncate(et, 0);
5483 path = ocfs2_new_path_from_et(et);
5490 ret = ocfs2_find_path(et->et_ci, path, cpos);
5496 el = path_leaf_el(path);
5497 index = ocfs2_search_extent_list(el, cpos);
5499 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5500 "Owner %llu has an extent at cpos %u which can no "
5501 "longer be found.\n",
5502 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5509 * We have 3 cases of extent removal:
5510 * 1) Range covers the entire extent rec
5511 * 2) Range begins or ends on one edge of the extent rec
5512 * 3) Range is in the middle of the extent rec (no shared edges)
5514 * For case 1 we remove the extent rec and left rotate to
5517 * For case 2 we just shrink the existing extent rec, with a
5518 * tree update if the shrinking edge is also the edge of an
5521 * For case 3 we do a right split to turn the extent rec into
5522 * something case 2 can handle.
5524 rec = &el->l_recs[index];
5525 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5526 trunc_range = cpos + len;
5528 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5530 trace_ocfs2_remove_extent(
5531 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5532 cpos, len, index, le32_to_cpu(rec->e_cpos),
5533 ocfs2_rec_clusters(el, rec));
5535 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5536 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5543 ret = ocfs2_split_tree(handle, et, path, index,
5544 trunc_range, meta_ac);
5551 * The split could have manipulated the tree enough to
5552 * move the record location, so we have to look for it again.
5554 ocfs2_reinit_path(path, 1);
5556 ret = ocfs2_find_path(et->et_ci, path, cpos);
5562 el = path_leaf_el(path);
5563 index = ocfs2_search_extent_list(el, cpos);
5565 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5566 "Owner %llu: split at cpos %u lost record.",
5567 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5574 * Double check our values here. If anything is fishy,
5575 * it's easier to catch it at the top level.
5577 rec = &el->l_recs[index];
5578 rec_range = le32_to_cpu(rec->e_cpos) +
5579 ocfs2_rec_clusters(el, rec);
5580 if (rec_range != trunc_range) {
5581 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5582 "Owner %llu: error after split at cpos %u"
5583 "trunc len %u, existing record is (%u,%u)",
5584 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5585 cpos, len, le32_to_cpu(rec->e_cpos),
5586 ocfs2_rec_clusters(el, rec));
5591 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5600 ocfs2_free_path(path);
5605 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5606 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5607 * number to reserve some extra blocks, and it only handles meta
5610 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5611 * and punching holes.
5613 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5614 struct ocfs2_extent_tree *et,
5615 u32 extents_to_split,
5616 struct ocfs2_alloc_context **ac,
5619 int ret = 0, num_free_extents;
5620 unsigned int max_recs_needed = 2 * extents_to_split;
5621 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5625 num_free_extents = ocfs2_num_free_extents(osb, et);
5626 if (num_free_extents < 0) {
5627 ret = num_free_extents;
5632 if (!num_free_extents ||
5633 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5634 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5637 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5648 ocfs2_free_alloc_context(*ac);
5656 int ocfs2_remove_btree_range(struct inode *inode,
5657 struct ocfs2_extent_tree *et,
5658 u32 cpos, u32 phys_cpos, u32 len, int flags,
5659 struct ocfs2_cached_dealloc_ctxt *dealloc,
5660 u64 refcount_loc, bool refcount_tree_locked)
5662 int ret, credits = 0, extra_blocks = 0;
5663 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5664 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5665 struct inode *tl_inode = osb->osb_tl_inode;
5667 struct ocfs2_alloc_context *meta_ac = NULL;
5668 struct ocfs2_refcount_tree *ref_tree = NULL;
5670 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5671 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
5672 OCFS2_HAS_REFCOUNT_FL));
5674 if (!refcount_tree_locked) {
5675 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5683 ret = ocfs2_prepare_refcount_change_for_del(inode,
5695 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5702 mutex_lock(&tl_inode->i_mutex);
5704 if (ocfs2_truncate_log_needs_flush(osb)) {
5705 ret = __ocfs2_flush_truncate_log(osb);
5712 handle = ocfs2_start_trans(osb,
5713 ocfs2_remove_extent_credits(osb->sb) + credits);
5714 if (IS_ERR(handle)) {
5715 ret = PTR_ERR(handle);
5720 ret = ocfs2_et_root_journal_access(handle, et,
5721 OCFS2_JOURNAL_ACCESS_WRITE);
5727 dquot_free_space_nodirty(inode,
5728 ocfs2_clusters_to_bytes(inode->i_sb, len));
5730 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5736 ocfs2_et_update_clusters(et, -len);
5737 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5739 ocfs2_journal_dirty(handle, et->et_root_bh);
5742 if (flags & OCFS2_EXT_REFCOUNTED)
5743 ret = ocfs2_decrease_refcount(inode, handle,
5744 ocfs2_blocks_to_clusters(osb->sb,
5749 ret = ocfs2_truncate_log_append(osb, handle,
5757 ocfs2_commit_trans(osb, handle);
5759 mutex_unlock(&tl_inode->i_mutex);
5762 ocfs2_free_alloc_context(meta_ac);
5765 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5770 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5772 struct buffer_head *tl_bh = osb->osb_tl_bh;
5773 struct ocfs2_dinode *di;
5774 struct ocfs2_truncate_log *tl;
5776 di = (struct ocfs2_dinode *) tl_bh->b_data;
5777 tl = &di->id2.i_dealloc;
5779 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5780 "slot %d, invalid truncate log parameters: used = "
5781 "%u, count = %u\n", osb->slot_num,
5782 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5783 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5786 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5787 unsigned int new_start)
5789 unsigned int tail_index;
5790 unsigned int current_tail;
5792 /* No records, nothing to coalesce */
5793 if (!le16_to_cpu(tl->tl_used))
5796 tail_index = le16_to_cpu(tl->tl_used) - 1;
5797 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5798 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5800 return current_tail == new_start;
5803 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5806 unsigned int num_clusters)
5809 unsigned int start_cluster, tl_count;
5810 struct inode *tl_inode = osb->osb_tl_inode;
5811 struct buffer_head *tl_bh = osb->osb_tl_bh;
5812 struct ocfs2_dinode *di;
5813 struct ocfs2_truncate_log *tl;
5815 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5817 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5819 di = (struct ocfs2_dinode *) tl_bh->b_data;
5821 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5822 * by the underlying call to ocfs2_read_inode_block(), so any
5823 * corruption is a code bug */
5824 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5826 tl = &di->id2.i_dealloc;
5827 tl_count = le16_to_cpu(tl->tl_count);
5828 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5830 "Truncate record count on #%llu invalid "
5831 "wanted %u, actual %u\n",
5832 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5833 ocfs2_truncate_recs_per_inode(osb->sb),
5834 le16_to_cpu(tl->tl_count));
5836 /* Caller should have known to flush before calling us. */
5837 index = le16_to_cpu(tl->tl_used);
5838 if (index >= tl_count) {
5844 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5845 OCFS2_JOURNAL_ACCESS_WRITE);
5851 trace_ocfs2_truncate_log_append(
5852 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5853 start_cluster, num_clusters);
5854 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5856 * Move index back to the record we are coalescing with.
5857 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5861 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5862 trace_ocfs2_truncate_log_append(
5863 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5864 index, le32_to_cpu(tl->tl_recs[index].t_start),
5867 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5868 tl->tl_used = cpu_to_le16(index + 1);
5870 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5872 ocfs2_journal_dirty(handle, tl_bh);
5874 osb->truncated_clusters += num_clusters;
5879 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5881 struct inode *data_alloc_inode,
5882 struct buffer_head *data_alloc_bh)
5886 unsigned int num_clusters;
5888 struct ocfs2_truncate_rec rec;
5889 struct ocfs2_dinode *di;
5890 struct ocfs2_truncate_log *tl;
5891 struct inode *tl_inode = osb->osb_tl_inode;
5892 struct buffer_head *tl_bh = osb->osb_tl_bh;
5894 di = (struct ocfs2_dinode *) tl_bh->b_data;
5895 tl = &di->id2.i_dealloc;
5896 i = le16_to_cpu(tl->tl_used) - 1;
5898 /* Caller has given us at least enough credits to
5899 * update the truncate log dinode */
5900 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5901 OCFS2_JOURNAL_ACCESS_WRITE);
5907 tl->tl_used = cpu_to_le16(i);
5909 ocfs2_journal_dirty(handle, tl_bh);
5911 /* TODO: Perhaps we can calculate the bulk of the
5912 * credits up front rather than extending like
5914 status = ocfs2_extend_trans(handle,
5915 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5921 rec = tl->tl_recs[i];
5922 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5923 le32_to_cpu(rec.t_start));
5924 num_clusters = le32_to_cpu(rec.t_clusters);
5926 /* if start_blk is not set, we ignore the record as
5929 trace_ocfs2_replay_truncate_records(
5930 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5931 i, le32_to_cpu(rec.t_start), num_clusters);
5933 status = ocfs2_free_clusters(handle, data_alloc_inode,
5934 data_alloc_bh, start_blk,
5944 osb->truncated_clusters = 0;
5950 /* Expects you to already be holding tl_inode->i_mutex */
5951 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5954 unsigned int num_to_flush;
5956 struct inode *tl_inode = osb->osb_tl_inode;
5957 struct inode *data_alloc_inode = NULL;
5958 struct buffer_head *tl_bh = osb->osb_tl_bh;
5959 struct buffer_head *data_alloc_bh = NULL;
5960 struct ocfs2_dinode *di;
5961 struct ocfs2_truncate_log *tl;
5963 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5965 di = (struct ocfs2_dinode *) tl_bh->b_data;
5967 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5968 * by the underlying call to ocfs2_read_inode_block(), so any
5969 * corruption is a code bug */
5970 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5972 tl = &di->id2.i_dealloc;
5973 num_to_flush = le16_to_cpu(tl->tl_used);
5974 trace_ocfs2_flush_truncate_log(
5975 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5977 if (!num_to_flush) {
5982 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5983 GLOBAL_BITMAP_SYSTEM_INODE,
5984 OCFS2_INVALID_SLOT);
5985 if (!data_alloc_inode) {
5987 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5991 mutex_lock(&data_alloc_inode->i_mutex);
5993 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5999 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6000 if (IS_ERR(handle)) {
6001 status = PTR_ERR(handle);
6006 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
6011 ocfs2_commit_trans(osb, handle);
6014 brelse(data_alloc_bh);
6015 ocfs2_inode_unlock(data_alloc_inode, 1);
6018 mutex_unlock(&data_alloc_inode->i_mutex);
6019 iput(data_alloc_inode);
6025 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6028 struct inode *tl_inode = osb->osb_tl_inode;
6030 mutex_lock(&tl_inode->i_mutex);
6031 status = __ocfs2_flush_truncate_log(osb);
6032 mutex_unlock(&tl_inode->i_mutex);
6037 static void ocfs2_truncate_log_worker(struct work_struct *work)
6040 struct ocfs2_super *osb =
6041 container_of(work, struct ocfs2_super,
6042 osb_truncate_log_wq.work);
6044 status = ocfs2_flush_truncate_log(osb);
6048 ocfs2_init_steal_slots(osb);
6051 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6052 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6055 if (osb->osb_tl_inode &&
6056 atomic_read(&osb->osb_tl_disable) == 0) {
6057 /* We want to push off log flushes while truncates are
6060 cancel_delayed_work(&osb->osb_truncate_log_wq);
6062 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
6063 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6067 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6069 struct inode **tl_inode,
6070 struct buffer_head **tl_bh)
6073 struct inode *inode = NULL;
6074 struct buffer_head *bh = NULL;
6076 inode = ocfs2_get_system_file_inode(osb,
6077 TRUNCATE_LOG_SYSTEM_INODE,
6081 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6085 status = ocfs2_read_inode_block(inode, &bh);
6098 /* called during the 1st stage of node recovery. we stamp a clean
6099 * truncate log and pass back a copy for processing later. if the
6100 * truncate log does not require processing, a *tl_copy is set to
6102 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6104 struct ocfs2_dinode **tl_copy)
6107 struct inode *tl_inode = NULL;
6108 struct buffer_head *tl_bh = NULL;
6109 struct ocfs2_dinode *di;
6110 struct ocfs2_truncate_log *tl;
6114 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6116 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6122 di = (struct ocfs2_dinode *) tl_bh->b_data;
6124 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6125 * validated by the underlying call to ocfs2_read_inode_block(),
6126 * so any corruption is a code bug */
6127 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6129 tl = &di->id2.i_dealloc;
6130 if (le16_to_cpu(tl->tl_used)) {
6131 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6133 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6140 /* Assuming the write-out below goes well, this copy
6141 * will be passed back to recovery for processing. */
6142 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6144 /* All we need to do to clear the truncate log is set
6148 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6149 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6161 if (status < 0 && (*tl_copy)) {
6170 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6171 struct ocfs2_dinode *tl_copy)
6175 unsigned int clusters, num_recs, start_cluster;
6178 struct inode *tl_inode = osb->osb_tl_inode;
6179 struct ocfs2_truncate_log *tl;
6181 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6182 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6186 tl = &tl_copy->id2.i_dealloc;
6187 num_recs = le16_to_cpu(tl->tl_used);
6188 trace_ocfs2_complete_truncate_log_recovery(
6189 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6192 mutex_lock(&tl_inode->i_mutex);
6193 for(i = 0; i < num_recs; i++) {
6194 if (ocfs2_truncate_log_needs_flush(osb)) {
6195 status = __ocfs2_flush_truncate_log(osb);
6202 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6203 if (IS_ERR(handle)) {
6204 status = PTR_ERR(handle);
6209 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6210 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6211 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6213 status = ocfs2_truncate_log_append(osb, handle,
6214 start_blk, clusters);
6215 ocfs2_commit_trans(osb, handle);
6223 mutex_unlock(&tl_inode->i_mutex);
6228 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6231 struct inode *tl_inode = osb->osb_tl_inode;
6233 atomic_set(&osb->osb_tl_disable, 1);
6236 cancel_delayed_work(&osb->osb_truncate_log_wq);
6237 flush_workqueue(ocfs2_wq);
6239 status = ocfs2_flush_truncate_log(osb);
6243 brelse(osb->osb_tl_bh);
6244 iput(osb->osb_tl_inode);
6248 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6251 struct inode *tl_inode = NULL;
6252 struct buffer_head *tl_bh = NULL;
6254 status = ocfs2_get_truncate_log_info(osb,
6261 /* ocfs2_truncate_log_shutdown keys on the existence of
6262 * osb->osb_tl_inode so we don't set any of the osb variables
6263 * until we're sure all is well. */
6264 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6265 ocfs2_truncate_log_worker);
6266 atomic_set(&osb->osb_tl_disable, 0);
6267 osb->osb_tl_bh = tl_bh;
6268 osb->osb_tl_inode = tl_inode;
6274 * Delayed de-allocation of suballocator blocks.
6276 * Some sets of block de-allocations might involve multiple suballocator inodes.
6278 * The locking for this can get extremely complicated, especially when
6279 * the suballocator inodes to delete from aren't known until deep
6280 * within an unrelated codepath.
6282 * ocfs2_extent_block structures are a good example of this - an inode
6283 * btree could have been grown by any number of nodes each allocating
6284 * out of their own suballoc inode.
6286 * These structures allow the delay of block de-allocation until a
6287 * later time, when locking of multiple cluster inodes won't cause
6292 * Describe a single bit freed from a suballocator. For the block
6293 * suballocators, it represents one block. For the global cluster
6294 * allocator, it represents some clusters and free_bit indicates
6297 struct ocfs2_cached_block_free {
6298 struct ocfs2_cached_block_free *free_next;
6301 unsigned int free_bit;
6304 struct ocfs2_per_slot_free_list {
6305 struct ocfs2_per_slot_free_list *f_next_suballocator;
6308 struct ocfs2_cached_block_free *f_first;
6311 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6314 struct ocfs2_cached_block_free *head)
6319 struct inode *inode;
6320 struct buffer_head *di_bh = NULL;
6321 struct ocfs2_cached_block_free *tmp;
6323 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6330 mutex_lock(&inode->i_mutex);
6332 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6338 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6339 if (IS_ERR(handle)) {
6340 ret = PTR_ERR(handle);
6347 bg_blkno = head->free_bg;
6349 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6351 trace_ocfs2_free_cached_blocks(
6352 (unsigned long long)head->free_blk, head->free_bit);
6354 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6355 head->free_bit, bg_blkno, 1);
6361 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6368 head = head->free_next;
6373 ocfs2_commit_trans(osb, handle);
6376 ocfs2_inode_unlock(inode, 1);
6379 mutex_unlock(&inode->i_mutex);
6383 /* Premature exit may have left some dangling items. */
6385 head = head->free_next;
6392 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6393 u64 blkno, unsigned int bit)
6396 struct ocfs2_cached_block_free *item;
6398 item = kzalloc(sizeof(*item), GFP_NOFS);
6405 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6407 item->free_blk = blkno;
6408 item->free_bit = bit;
6409 item->free_next = ctxt->c_global_allocator;
6411 ctxt->c_global_allocator = item;
6415 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6416 struct ocfs2_cached_block_free *head)
6418 struct ocfs2_cached_block_free *tmp;
6419 struct inode *tl_inode = osb->osb_tl_inode;
6423 mutex_lock(&tl_inode->i_mutex);
6426 if (ocfs2_truncate_log_needs_flush(osb)) {
6427 ret = __ocfs2_flush_truncate_log(osb);
6434 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6435 if (IS_ERR(handle)) {
6436 ret = PTR_ERR(handle);
6441 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6444 ocfs2_commit_trans(osb, handle);
6446 head = head->free_next;
6455 mutex_unlock(&tl_inode->i_mutex);
6458 /* Premature exit may have left some dangling items. */
6460 head = head->free_next;
6467 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6468 struct ocfs2_cached_dealloc_ctxt *ctxt)
6471 struct ocfs2_per_slot_free_list *fl;
6476 while (ctxt->c_first_suballocator) {
6477 fl = ctxt->c_first_suballocator;
6480 trace_ocfs2_run_deallocs(fl->f_inode_type,
6482 ret2 = ocfs2_free_cached_blocks(osb,
6492 ctxt->c_first_suballocator = fl->f_next_suballocator;
6496 if (ctxt->c_global_allocator) {
6497 ret2 = ocfs2_free_cached_clusters(osb,
6498 ctxt->c_global_allocator);
6504 ctxt->c_global_allocator = NULL;
6510 static struct ocfs2_per_slot_free_list *
6511 ocfs2_find_per_slot_free_list(int type,
6513 struct ocfs2_cached_dealloc_ctxt *ctxt)
6515 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6518 if (fl->f_inode_type == type && fl->f_slot == slot)
6521 fl = fl->f_next_suballocator;
6524 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6526 fl->f_inode_type = type;
6529 fl->f_next_suballocator = ctxt->c_first_suballocator;
6531 ctxt->c_first_suballocator = fl;
6536 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6537 int type, int slot, u64 suballoc,
6538 u64 blkno, unsigned int bit)
6541 struct ocfs2_per_slot_free_list *fl;
6542 struct ocfs2_cached_block_free *item;
6544 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6551 item = kzalloc(sizeof(*item), GFP_NOFS);
6558 trace_ocfs2_cache_block_dealloc(type, slot,
6559 (unsigned long long)suballoc,
6560 (unsigned long long)blkno, bit);
6562 item->free_bg = suballoc;
6563 item->free_blk = blkno;
6564 item->free_bit = bit;
6565 item->free_next = fl->f_first;
6574 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6575 struct ocfs2_extent_block *eb)
6577 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6578 le16_to_cpu(eb->h_suballoc_slot),
6579 le64_to_cpu(eb->h_suballoc_loc),
6580 le64_to_cpu(eb->h_blkno),
6581 le16_to_cpu(eb->h_suballoc_bit));
6584 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6586 set_buffer_uptodate(bh);
6587 mark_buffer_dirty(bh);
6591 void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6592 unsigned int from, unsigned int to,
6593 struct page *page, int zero, u64 *phys)
6595 int ret, partial = 0;
6597 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6602 zero_user_segment(page, from, to);
6605 * Need to set the buffers we zero'd into uptodate
6606 * here if they aren't - ocfs2_map_page_blocks()
6607 * might've skipped some
6609 ret = walk_page_buffers(handle, page_buffers(page),
6614 else if (ocfs2_should_order_data(inode)) {
6615 ret = ocfs2_jbd2_file_inode(handle, inode);
6621 SetPageUptodate(page);
6623 flush_dcache_page(page);
6626 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6627 loff_t end, struct page **pages,
6628 int numpages, u64 phys, handle_t *handle)
6632 unsigned int from, to = PAGE_CACHE_SIZE;
6633 struct super_block *sb = inode->i_sb;
6635 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6640 to = PAGE_CACHE_SIZE;
6641 for(i = 0; i < numpages; i++) {
6644 from = start & (PAGE_CACHE_SIZE - 1);
6645 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6646 to = end & (PAGE_CACHE_SIZE - 1);
6648 BUG_ON(from > PAGE_CACHE_SIZE);
6649 BUG_ON(to > PAGE_CACHE_SIZE);
6651 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6654 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6658 ocfs2_unlock_and_free_pages(pages, numpages);
6661 int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6662 struct page **pages, int *num)
6664 int numpages, ret = 0;
6665 struct address_space *mapping = inode->i_mapping;
6666 unsigned long index;
6667 loff_t last_page_bytes;
6669 BUG_ON(start > end);
6672 last_page_bytes = PAGE_ALIGN(end);
6673 index = start >> PAGE_CACHE_SHIFT;
6675 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6676 if (!pages[numpages]) {
6684 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6689 ocfs2_unlock_and_free_pages(pages, numpages);
6698 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6699 struct page **pages, int *num)
6701 struct super_block *sb = inode->i_sb;
6703 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6704 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6706 return ocfs2_grab_pages(inode, start, end, pages, num);
6710 * Zero the area past i_size but still within an allocated
6711 * cluster. This avoids exposing nonzero data on subsequent file
6714 * We need to call this before i_size is updated on the inode because
6715 * otherwise block_write_full_page() will skip writeout of pages past
6716 * i_size. The new_i_size parameter is passed for this reason.
6718 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6719 u64 range_start, u64 range_end)
6721 int ret = 0, numpages;
6722 struct page **pages = NULL;
6724 unsigned int ext_flags;
6725 struct super_block *sb = inode->i_sb;
6728 * File systems which don't support sparse files zero on every
6731 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6734 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6735 sizeof(struct page *), GFP_NOFS);
6736 if (pages == NULL) {
6742 if (range_start == range_end)
6745 ret = ocfs2_extent_map_get_blocks(inode,
6746 range_start >> sb->s_blocksize_bits,
6747 &phys, NULL, &ext_flags);
6754 * Tail is a hole, or is marked unwritten. In either case, we
6755 * can count on read and write to return/push zero's.
6757 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6760 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6767 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6768 numpages, phys, handle);
6771 * Initiate writeout of the pages we zero'd here. We don't
6772 * wait on them - the truncate_inode_pages() call later will
6775 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6786 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6787 struct ocfs2_dinode *di)
6789 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6790 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6792 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6793 memset(&di->id2, 0, blocksize -
6794 offsetof(struct ocfs2_dinode, id2) -
6797 memset(&di->id2, 0, blocksize -
6798 offsetof(struct ocfs2_dinode, id2));
6801 void ocfs2_dinode_new_extent_list(struct inode *inode,
6802 struct ocfs2_dinode *di)
6804 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6805 di->id2.i_list.l_tree_depth = 0;
6806 di->id2.i_list.l_next_free_rec = 0;
6807 di->id2.i_list.l_count = cpu_to_le16(
6808 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6811 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6813 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6814 struct ocfs2_inline_data *idata = &di->id2.i_data;
6816 spin_lock(&oi->ip_lock);
6817 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6818 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6819 spin_unlock(&oi->ip_lock);
6822 * We clear the entire i_data structure here so that all
6823 * fields can be properly initialized.
6825 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6827 idata->id_count = cpu_to_le16(
6828 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6831 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6832 struct buffer_head *di_bh)
6834 int ret, i, has_data, num_pages = 0;
6838 u64 uninitialized_var(block);
6839 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6840 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6841 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6842 struct ocfs2_alloc_context *data_ac = NULL;
6843 struct page **pages = NULL;
6844 loff_t end = osb->s_clustersize;
6845 struct ocfs2_extent_tree et;
6848 has_data = i_size_read(inode) ? 1 : 0;
6851 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6852 sizeof(struct page *), GFP_NOFS);
6853 if (pages == NULL) {
6859 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6866 handle = ocfs2_start_trans(osb,
6867 ocfs2_inline_to_extents_credits(osb->sb));
6868 if (IS_ERR(handle)) {
6869 ret = PTR_ERR(handle);
6874 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
6875 OCFS2_JOURNAL_ACCESS_WRITE);
6882 unsigned int page_end;
6885 ret = dquot_alloc_space_nodirty(inode,
6886 ocfs2_clusters_to_bytes(osb->sb, 1));
6891 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
6893 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
6901 * Save two copies, one for insert, and one that can
6902 * be changed by ocfs2_map_and_dirty_page() below.
6904 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6907 * Non sparse file systems zero on extend, so no need
6910 if (!ocfs2_sparse_alloc(osb) &&
6911 PAGE_CACHE_SIZE < osb->s_clustersize)
6912 end = PAGE_CACHE_SIZE;
6914 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6922 * This should populate the 1st page for us and mark
6925 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6932 page_end = PAGE_CACHE_SIZE;
6933 if (PAGE_CACHE_SIZE > osb->s_clustersize)
6934 page_end = osb->s_clustersize;
6936 for (i = 0; i < num_pages; i++)
6937 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6938 pages[i], i > 0, &phys);
6941 spin_lock(&oi->ip_lock);
6942 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6943 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6944 spin_unlock(&oi->ip_lock);
6946 ocfs2_update_inode_fsync_trans(handle, inode, 1);
6947 ocfs2_dinode_new_extent_list(inode, di);
6949 ocfs2_journal_dirty(handle, di_bh);
6953 * An error at this point should be extremely rare. If
6954 * this proves to be false, we could always re-build
6955 * the in-inode data from our pages.
6957 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
6958 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
6965 inode->i_blocks = ocfs2_inode_sector_count(inode);
6970 ocfs2_unlock_and_free_pages(pages, num_pages);
6973 if (ret < 0 && did_quota)
6974 dquot_free_space_nodirty(inode,
6975 ocfs2_clusters_to_bytes(osb->sb, 1));
6978 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
6979 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
6982 ocfs2_free_clusters(handle,
6985 ocfs2_clusters_to_blocks(osb->sb, bit_off),
6989 ocfs2_commit_trans(osb, handle);
6993 ocfs2_free_alloc_context(data_ac);
7000 * It is expected, that by the time you call this function,
7001 * inode->i_size and fe->i_size have been adjusted.
7003 * WARNING: This will kfree the truncate context
7005 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7006 struct inode *inode,
7007 struct buffer_head *di_bh)
7009 int status = 0, i, flags = 0;
7010 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7012 struct ocfs2_extent_list *el;
7013 struct ocfs2_extent_rec *rec;
7014 struct ocfs2_path *path = NULL;
7015 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7016 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7017 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7018 struct ocfs2_extent_tree et;
7019 struct ocfs2_cached_dealloc_ctxt dealloc;
7020 struct ocfs2_refcount_tree *ref_tree = NULL;
7022 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7023 ocfs2_init_dealloc_ctxt(&dealloc);
7025 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7026 i_size_read(inode));
7028 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7029 ocfs2_journal_access_di);
7036 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7040 * Check that we still have allocation to delete.
7042 if (OCFS2_I(inode)->ip_clusters == 0) {
7048 * Truncate always works against the rightmost tree branch.
7050 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7056 trace_ocfs2_commit_truncate(
7057 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7059 OCFS2_I(inode)->ip_clusters,
7060 path->p_tree_depth);
7063 * By now, el will point to the extent list on the bottom most
7064 * portion of this tree. Only the tail record is considered in
7067 * We handle the following cases, in order:
7068 * - empty extent: delete the remaining branch
7069 * - remove the entire record
7070 * - remove a partial record
7071 * - no record needs to be removed (truncate has completed)
7073 el = path_leaf_el(path);
7074 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7075 ocfs2_error(inode->i_sb,
7076 "Inode %llu has empty extent block at %llu\n",
7077 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7078 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7083 i = le16_to_cpu(el->l_next_free_rec) - 1;
7084 rec = &el->l_recs[i];
7085 flags = rec->e_flags;
7086 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7088 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7090 * Lower levels depend on this never happening, but it's best
7091 * to check it up here before changing the tree.
7093 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7094 ocfs2_error(inode->i_sb, "Inode %lu has an empty "
7095 "extent record, depth %u\n", inode->i_ino,
7096 le16_to_cpu(root_el->l_tree_depth));
7100 trunc_cpos = le32_to_cpu(rec->e_cpos);
7103 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7105 * Truncate entire record.
7107 trunc_cpos = le32_to_cpu(rec->e_cpos);
7108 trunc_len = ocfs2_rec_clusters(el, rec);
7109 blkno = le64_to_cpu(rec->e_blkno);
7110 } else if (range > new_highest_cpos) {
7112 * Partial truncate. it also should be
7113 * the last truncate we're doing.
7115 trunc_cpos = new_highest_cpos;
7116 trunc_len = range - new_highest_cpos;
7117 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7118 blkno = le64_to_cpu(rec->e_blkno) +
7119 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7122 * Truncate completed, leave happily.
7128 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7130 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7131 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7139 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7140 phys_cpos, trunc_len, flags, &dealloc,
7141 refcount_loc, true);
7147 ocfs2_reinit_path(path, 1);
7150 * The check above will catch the case where we've truncated
7151 * away all allocation.
7157 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7159 ocfs2_schedule_truncate_log_flush(osb, 1);
7161 ocfs2_run_deallocs(osb, &dealloc);
7163 ocfs2_free_path(path);
7169 * 'start' is inclusive, 'end' is not.
7171 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7172 unsigned int start, unsigned int end, int trunc)
7175 unsigned int numbytes;
7177 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7178 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7179 struct ocfs2_inline_data *idata = &di->id2.i_data;
7181 if (end > i_size_read(inode))
7182 end = i_size_read(inode);
7184 BUG_ON(start > end);
7186 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7187 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7188 !ocfs2_supports_inline_data(osb)) {
7189 ocfs2_error(inode->i_sb,
7190 "Inline data flags for inode %llu don't agree! "
7191 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7192 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7193 le16_to_cpu(di->i_dyn_features),
7194 OCFS2_I(inode)->ip_dyn_features,
7195 osb->s_feature_incompat);
7200 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7201 if (IS_ERR(handle)) {
7202 ret = PTR_ERR(handle);
7207 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7208 OCFS2_JOURNAL_ACCESS_WRITE);
7214 numbytes = end - start;
7215 memset(idata->id_data + start, 0, numbytes);
7218 * No need to worry about the data page here - it's been
7219 * truncated already and inline data doesn't need it for
7220 * pushing zero's to disk, so we'll let readpage pick it up
7224 i_size_write(inode, start);
7225 di->i_size = cpu_to_le64(start);
7228 inode->i_blocks = ocfs2_inode_sector_count(inode);
7229 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7231 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7232 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7234 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7235 ocfs2_journal_dirty(handle, di_bh);
7238 ocfs2_commit_trans(osb, handle);
7244 static int ocfs2_trim_extent(struct super_block *sb,
7245 struct ocfs2_group_desc *gd,
7246 u32 start, u32 count)
7248 u64 discard, bcount;
7250 bcount = ocfs2_clusters_to_blocks(sb, count);
7251 discard = le64_to_cpu(gd->bg_blkno) +
7252 ocfs2_clusters_to_blocks(sb, start);
7254 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7256 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7259 static int ocfs2_trim_group(struct super_block *sb,
7260 struct ocfs2_group_desc *gd,
7261 u32 start, u32 max, u32 minbits)
7263 int ret = 0, count = 0, next;
7264 void *bitmap = gd->bg_bitmap;
7266 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7269 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7270 start, max, minbits);
7272 while (start < max) {
7273 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7276 next = ocfs2_find_next_bit(bitmap, max, start);
7278 if ((next - start) >= minbits) {
7279 ret = ocfs2_trim_extent(sb, gd,
7280 start, next - start);
7285 count += next - start;
7289 if (fatal_signal_pending(current)) {
7290 count = -ERESTARTSYS;
7294 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7304 int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7306 struct ocfs2_super *osb = OCFS2_SB(sb);
7307 u64 start, len, trimmed, first_group, last_group, group;
7309 u32 first_bit, last_bit, minlen;
7310 struct buffer_head *main_bm_bh = NULL;
7311 struct inode *main_bm_inode = NULL;
7312 struct buffer_head *gd_bh = NULL;
7313 struct ocfs2_dinode *main_bm;
7314 struct ocfs2_group_desc *gd = NULL;
7316 start = range->start >> osb->s_clustersize_bits;
7317 len = range->len >> osb->s_clustersize_bits;
7318 minlen = range->minlen >> osb->s_clustersize_bits;
7320 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7323 main_bm_inode = ocfs2_get_system_file_inode(osb,
7324 GLOBAL_BITMAP_SYSTEM_INODE,
7325 OCFS2_INVALID_SLOT);
7326 if (!main_bm_inode) {
7332 mutex_lock(&main_bm_inode->i_mutex);
7334 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7339 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7341 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7346 len = range->len >> osb->s_clustersize_bits;
7347 if (start + len > le32_to_cpu(main_bm->i_clusters))
7348 len = le32_to_cpu(main_bm->i_clusters) - start;
7350 trace_ocfs2_trim_fs(start, len, minlen);
7352 /* Determine first and last group to examine based on start and len */
7353 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7354 if (first_group == osb->first_cluster_group_blkno)
7357 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7358 last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7359 last_bit = osb->bitmap_cpg;
7362 for (group = first_group; group <= last_group;) {
7363 if (first_bit + len >= osb->bitmap_cpg)
7364 last_bit = osb->bitmap_cpg;
7366 last_bit = first_bit + len;
7368 ret = ocfs2_read_group_descriptor(main_bm_inode,
7376 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7377 cnt = ocfs2_trim_group(sb, gd, first_bit, last_bit, minlen);
7387 len -= osb->bitmap_cpg - first_bit;
7389 if (group == osb->first_cluster_group_blkno)
7390 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7392 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7394 range->len = trimmed * sb->s_blocksize;
7396 ocfs2_inode_unlock(main_bm_inode, 0);
7399 mutex_unlock(&main_bm_inode->i_mutex);
7400 iput(main_bm_inode);