1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Extent allocs and frees
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
31 #include <linux/quotaops.h>
32 #include <linux/blkdev.h>
33 #include <linux/sched/signal.h>
35 #include <cluster/masklog.h>
41 #include "blockcheck.h"
43 #include "extent_map.h"
46 #include "localalloc.h"
53 #include "refcounttree.h"
54 #include "ocfs2_trace.h"
56 #include "buffer_head_io.h"
58 enum ocfs2_contig_type {
65 static enum ocfs2_contig_type
66 ocfs2_extent_rec_contig(struct super_block *sb,
67 struct ocfs2_extent_rec *ext,
68 struct ocfs2_extent_rec *insert_rec);
70 * Operations for a specific extent tree type.
72 * To implement an on-disk btree (extent tree) type in ocfs2, add
73 * an ocfs2_extent_tree_operations structure and the matching
74 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
75 * for the allocation portion of the extent tree.
77 struct ocfs2_extent_tree_operations {
79 * last_eb_blk is the block number of the right most leaf extent
80 * block. Most on-disk structures containing an extent tree store
81 * this value for fast access. The ->eo_set_last_eb_blk() and
82 * ->eo_get_last_eb_blk() operations access this value. They are
85 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
87 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
90 * The on-disk structure usually keeps track of how many total
91 * clusters are stored in this extent tree. This function updates
92 * that value. new_clusters is the delta, and must be
93 * added to the total. Required.
95 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
99 * If this extent tree is supported by an extent map, insert
100 * a record into the map.
102 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
103 struct ocfs2_extent_rec *rec);
106 * If this extent tree is supported by an extent map, truncate the
109 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
113 * If ->eo_insert_check() exists, it is called before rec is
114 * inserted into the extent tree. It is optional.
116 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
117 struct ocfs2_extent_rec *rec);
118 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
121 * --------------------------------------------------------------
122 * The remaining are internal to ocfs2_extent_tree and don't have
127 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
130 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
133 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
134 * it exists. If it does not, et->et_max_leaf_clusters is set
135 * to 0 (unlimited). Optional.
137 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
140 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
141 * are contiguous or not. Optional. Don't need to set it if use
142 * ocfs2_extent_rec as the tree leaf.
144 enum ocfs2_contig_type
145 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
146 struct ocfs2_extent_rec *ext,
147 struct ocfs2_extent_rec *insert_rec);
152 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
155 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
156 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
158 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
160 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
161 struct ocfs2_extent_rec *rec);
162 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
164 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
165 struct ocfs2_extent_rec *rec);
166 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
167 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
169 static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
170 struct ocfs2_extent_tree *et,
171 struct buffer_head **new_eb_bh,
172 int blk_wanted, int *blk_given);
173 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et);
175 static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
176 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
177 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
178 .eo_update_clusters = ocfs2_dinode_update_clusters,
179 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
180 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
181 .eo_insert_check = ocfs2_dinode_insert_check,
182 .eo_sanity_check = ocfs2_dinode_sanity_check,
183 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
186 static void ocfs2_dinode_set_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 di->i_last_eb_blk = cpu_to_le64(blkno);
195 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
197 struct ocfs2_dinode *di = et->et_object;
199 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
200 return le64_to_cpu(di->i_last_eb_blk);
203 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
206 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
207 struct ocfs2_dinode *di = et->et_object;
209 le32_add_cpu(&di->i_clusters, clusters);
210 spin_lock(&oi->ip_lock);
211 oi->ip_clusters = le32_to_cpu(di->i_clusters);
212 spin_unlock(&oi->ip_lock);
215 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
216 struct ocfs2_extent_rec *rec)
218 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
220 ocfs2_extent_map_insert_rec(inode, rec);
223 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
226 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
228 ocfs2_extent_map_trunc(inode, clusters);
231 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
232 struct ocfs2_extent_rec *rec)
234 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
235 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
237 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
238 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
239 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
240 "Device %s, asking for sparse allocation: inode %llu, "
241 "cpos %u, clusters %u\n",
243 (unsigned long long)oi->ip_blkno,
244 rec->e_cpos, oi->ip_clusters);
249 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
251 struct ocfs2_dinode *di = et->et_object;
253 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
254 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
259 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
261 struct ocfs2_dinode *di = et->et_object;
263 et->et_root_el = &di->id2.i_list;
267 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
269 struct ocfs2_xattr_value_buf *vb = et->et_object;
271 et->et_root_el = &vb->vb_xv->xr_list;
274 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
277 struct ocfs2_xattr_value_buf *vb = et->et_object;
279 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
282 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
284 struct ocfs2_xattr_value_buf *vb = et->et_object;
286 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
289 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
292 struct ocfs2_xattr_value_buf *vb = et->et_object;
294 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
297 static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
298 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
299 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
300 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
301 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
304 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
306 struct ocfs2_xattr_block *xb = et->et_object;
308 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
311 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
313 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
314 et->et_max_leaf_clusters =
315 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
318 static void ocfs2_xattr_tree_set_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 xt->xt_last_eb_blk = cpu_to_le64(blkno);
327 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
329 struct ocfs2_xattr_block *xb = et->et_object;
330 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
332 return le64_to_cpu(xt->xt_last_eb_blk);
335 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
338 struct ocfs2_xattr_block *xb = et->et_object;
340 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
343 static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
344 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
345 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
346 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
347 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
348 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
351 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
354 struct ocfs2_dx_root_block *dx_root = et->et_object;
356 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
359 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
361 struct ocfs2_dx_root_block *dx_root = et->et_object;
363 return le64_to_cpu(dx_root->dr_last_eb_blk);
366 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
369 struct ocfs2_dx_root_block *dx_root = et->et_object;
371 le32_add_cpu(&dx_root->dr_clusters, clusters);
374 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
376 struct ocfs2_dx_root_block *dx_root = et->et_object;
378 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
383 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
385 struct ocfs2_dx_root_block *dx_root = et->et_object;
387 et->et_root_el = &dx_root->dr_list;
390 static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
391 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
392 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
393 .eo_update_clusters = ocfs2_dx_root_update_clusters,
394 .eo_sanity_check = ocfs2_dx_root_sanity_check,
395 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
398 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
400 struct ocfs2_refcount_block *rb = et->et_object;
402 et->et_root_el = &rb->rf_list;
405 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
408 struct ocfs2_refcount_block *rb = et->et_object;
410 rb->rf_last_eb_blk = cpu_to_le64(blkno);
413 static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
415 struct ocfs2_refcount_block *rb = et->et_object;
417 return le64_to_cpu(rb->rf_last_eb_blk);
420 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
423 struct ocfs2_refcount_block *rb = et->et_object;
425 le32_add_cpu(&rb->rf_clusters, clusters);
428 static enum ocfs2_contig_type
429 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
430 struct ocfs2_extent_rec *ext,
431 struct ocfs2_extent_rec *insert_rec)
436 static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
437 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
438 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
439 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
440 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
441 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
444 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
445 struct ocfs2_caching_info *ci,
446 struct buffer_head *bh,
447 ocfs2_journal_access_func access,
449 const struct ocfs2_extent_tree_operations *ops)
454 et->et_root_journal_access = access;
456 obj = (void *)bh->b_data;
458 et->et_dealloc = NULL;
460 et->et_ops->eo_fill_root_el(et);
461 if (!et->et_ops->eo_fill_max_leaf_clusters)
462 et->et_max_leaf_clusters = 0;
464 et->et_ops->eo_fill_max_leaf_clusters(et);
467 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
468 struct ocfs2_caching_info *ci,
469 struct buffer_head *bh)
471 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
472 NULL, &ocfs2_dinode_et_ops);
475 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
476 struct ocfs2_caching_info *ci,
477 struct buffer_head *bh)
479 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
480 NULL, &ocfs2_xattr_tree_et_ops);
483 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
484 struct ocfs2_caching_info *ci,
485 struct ocfs2_xattr_value_buf *vb)
487 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
488 &ocfs2_xattr_value_et_ops);
491 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
492 struct ocfs2_caching_info *ci,
493 struct buffer_head *bh)
495 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
496 NULL, &ocfs2_dx_root_et_ops);
499 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
500 struct ocfs2_caching_info *ci,
501 struct buffer_head *bh)
503 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
504 NULL, &ocfs2_refcount_tree_et_ops);
507 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
510 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
513 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
515 return et->et_ops->eo_get_last_eb_blk(et);
518 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
521 et->et_ops->eo_update_clusters(et, clusters);
524 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
525 struct ocfs2_extent_rec *rec)
527 if (et->et_ops->eo_extent_map_insert)
528 et->et_ops->eo_extent_map_insert(et, rec);
531 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
534 if (et->et_ops->eo_extent_map_truncate)
535 et->et_ops->eo_extent_map_truncate(et, clusters);
538 static inline int ocfs2_et_root_journal_access(handle_t *handle,
539 struct ocfs2_extent_tree *et,
542 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
546 static inline enum ocfs2_contig_type
547 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
548 struct ocfs2_extent_rec *rec,
549 struct ocfs2_extent_rec *insert_rec)
551 if (et->et_ops->eo_extent_contig)
552 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
554 return ocfs2_extent_rec_contig(
555 ocfs2_metadata_cache_get_super(et->et_ci),
559 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
560 struct ocfs2_extent_rec *rec)
564 if (et->et_ops->eo_insert_check)
565 ret = et->et_ops->eo_insert_check(et, rec);
569 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
573 if (et->et_ops->eo_sanity_check)
574 ret = et->et_ops->eo_sanity_check(et);
578 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
579 struct ocfs2_extent_block *eb);
580 static void ocfs2_adjust_rightmost_records(handle_t *handle,
581 struct ocfs2_extent_tree *et,
582 struct ocfs2_path *path,
583 struct ocfs2_extent_rec *insert_rec);
585 * Reset the actual path elements so that we can re-use the structure
586 * to build another path. Generally, this involves freeing the buffer
589 void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
591 int i, start = 0, depth = 0;
592 struct ocfs2_path_item *node;
597 for(i = start; i < path_num_items(path); i++) {
598 node = &path->p_node[i];
606 * Tree depth may change during truncate, or insert. If we're
607 * keeping the root extent list, then make sure that our path
608 * structure reflects the proper depth.
611 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
613 path_root_access(path) = NULL;
615 path->p_tree_depth = depth;
618 void ocfs2_free_path(struct ocfs2_path *path)
621 ocfs2_reinit_path(path, 0);
627 * All the elements of src into dest. After this call, src could be freed
628 * without affecting dest.
630 * Both paths should have the same root. Any non-root elements of dest
633 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
637 BUG_ON(path_root_bh(dest) != path_root_bh(src));
638 BUG_ON(path_root_el(dest) != path_root_el(src));
639 BUG_ON(path_root_access(dest) != path_root_access(src));
641 ocfs2_reinit_path(dest, 1);
643 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
644 dest->p_node[i].bh = src->p_node[i].bh;
645 dest->p_node[i].el = src->p_node[i].el;
647 if (dest->p_node[i].bh)
648 get_bh(dest->p_node[i].bh);
653 * Make the *dest path the same as src and re-initialize src path to
656 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
660 BUG_ON(path_root_bh(dest) != path_root_bh(src));
661 BUG_ON(path_root_access(dest) != path_root_access(src));
663 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
664 brelse(dest->p_node[i].bh);
666 dest->p_node[i].bh = src->p_node[i].bh;
667 dest->p_node[i].el = src->p_node[i].el;
669 src->p_node[i].bh = NULL;
670 src->p_node[i].el = NULL;
675 * Insert an extent block at given index.
677 * This will not take an additional reference on eb_bh.
679 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
680 struct buffer_head *eb_bh)
682 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
685 * Right now, no root bh is an extent block, so this helps
686 * catch code errors with dinode trees. The assertion can be
687 * safely removed if we ever need to insert extent block
688 * structures at the root.
692 path->p_node[index].bh = eb_bh;
693 path->p_node[index].el = &eb->h_list;
696 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
697 struct ocfs2_extent_list *root_el,
698 ocfs2_journal_access_func access)
700 struct ocfs2_path *path;
702 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
704 path = kzalloc(sizeof(*path), GFP_NOFS);
706 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
708 path_root_bh(path) = root_bh;
709 path_root_el(path) = root_el;
710 path_root_access(path) = access;
716 struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
718 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
719 path_root_access(path));
722 struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
724 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
725 et->et_root_journal_access);
729 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
730 * otherwise it's the root_access function.
732 * I don't like the way this function's name looks next to
733 * ocfs2_journal_access_path(), but I don't have a better one.
735 int ocfs2_path_bh_journal_access(handle_t *handle,
736 struct ocfs2_caching_info *ci,
737 struct ocfs2_path *path,
740 ocfs2_journal_access_func access = path_root_access(path);
743 access = ocfs2_journal_access;
746 access = ocfs2_journal_access_eb;
748 return access(handle, ci, path->p_node[idx].bh,
749 OCFS2_JOURNAL_ACCESS_WRITE);
753 * Convenience function to journal all components in a path.
755 int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
757 struct ocfs2_path *path)
764 for(i = 0; i < path_num_items(path); i++) {
765 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
777 * Return the index of the extent record which contains cluster #v_cluster.
778 * -1 is returned if it was not found.
780 * Should work fine on interior and exterior nodes.
782 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
786 struct ocfs2_extent_rec *rec;
787 u32 rec_end, rec_start, clusters;
789 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
790 rec = &el->l_recs[i];
792 rec_start = le32_to_cpu(rec->e_cpos);
793 clusters = ocfs2_rec_clusters(el, rec);
795 rec_end = rec_start + clusters;
797 if (v_cluster >= rec_start && v_cluster < rec_end) {
807 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
808 * ocfs2_extent_rec_contig only work properly against leaf nodes!
810 static int ocfs2_block_extent_contig(struct super_block *sb,
811 struct ocfs2_extent_rec *ext,
814 u64 blk_end = le64_to_cpu(ext->e_blkno);
816 blk_end += ocfs2_clusters_to_blocks(sb,
817 le16_to_cpu(ext->e_leaf_clusters));
819 return blkno == blk_end;
822 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
823 struct ocfs2_extent_rec *right)
827 left_range = le32_to_cpu(left->e_cpos) +
828 le16_to_cpu(left->e_leaf_clusters);
830 return (left_range == le32_to_cpu(right->e_cpos));
833 static enum ocfs2_contig_type
834 ocfs2_extent_rec_contig(struct super_block *sb,
835 struct ocfs2_extent_rec *ext,
836 struct ocfs2_extent_rec *insert_rec)
838 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
841 * Refuse to coalesce extent records with different flag
842 * fields - we don't want to mix unwritten extents with user
845 if (ext->e_flags != insert_rec->e_flags)
848 if (ocfs2_extents_adjacent(ext, insert_rec) &&
849 ocfs2_block_extent_contig(sb, ext, blkno))
852 blkno = le64_to_cpu(ext->e_blkno);
853 if (ocfs2_extents_adjacent(insert_rec, ext) &&
854 ocfs2_block_extent_contig(sb, insert_rec, blkno))
861 * NOTE: We can have pretty much any combination of contiguousness and
864 * The usefulness of APPEND_TAIL is more in that it lets us know that
865 * we'll have to update the path to that leaf.
867 enum ocfs2_append_type {
872 enum ocfs2_split_type {
878 struct ocfs2_insert_type {
879 enum ocfs2_split_type ins_split;
880 enum ocfs2_append_type ins_appending;
881 enum ocfs2_contig_type ins_contig;
882 int ins_contig_index;
886 struct ocfs2_merge_ctxt {
887 enum ocfs2_contig_type c_contig_type;
888 int c_has_empty_extent;
889 int c_split_covers_rec;
892 static int ocfs2_validate_extent_block(struct super_block *sb,
893 struct buffer_head *bh)
896 struct ocfs2_extent_block *eb =
897 (struct ocfs2_extent_block *)bh->b_data;
899 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
901 BUG_ON(!buffer_uptodate(bh));
904 * If the ecc fails, we return the error but otherwise
905 * leave the filesystem running. We know any error is
906 * local to this block.
908 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
910 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
911 (unsigned long long)bh->b_blocknr);
916 * Errors after here are fatal.
919 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
921 "Extent block #%llu has bad signature %.*s\n",
922 (unsigned long long)bh->b_blocknr, 7,
927 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
929 "Extent block #%llu has an invalid h_blkno of %llu\n",
930 (unsigned long long)bh->b_blocknr,
931 (unsigned long long)le64_to_cpu(eb->h_blkno));
935 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation)
937 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
938 (unsigned long long)bh->b_blocknr,
939 le32_to_cpu(eb->h_fs_generation));
944 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
945 struct buffer_head **bh)
948 struct buffer_head *tmp = *bh;
950 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
951 ocfs2_validate_extent_block);
953 /* If ocfs2_read_block() got us a new bh, pass it up. */
962 * How many free extents have we got before we need more meta data?
964 int ocfs2_num_free_extents(struct ocfs2_extent_tree *et)
967 struct ocfs2_extent_list *el = NULL;
968 struct ocfs2_extent_block *eb;
969 struct buffer_head *eb_bh = NULL;
973 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
976 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
982 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
986 BUG_ON(el->l_tree_depth != 0);
988 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
992 trace_ocfs2_num_free_extents(retval);
996 /* expects array to already be allocated
998 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
1001 static int ocfs2_create_new_meta_bhs(handle_t *handle,
1002 struct ocfs2_extent_tree *et,
1004 struct ocfs2_alloc_context *meta_ac,
1005 struct buffer_head *bhs[])
1007 int count, status, i;
1008 u16 suballoc_bit_start;
1010 u64 suballoc_loc, first_blkno;
1011 struct ocfs2_super *osb =
1012 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1013 struct ocfs2_extent_block *eb;
1016 while (count < wanted) {
1017 status = ocfs2_claim_metadata(handle,
1021 &suballoc_bit_start,
1029 for(i = count; i < (num_got + count); i++) {
1030 bhs[i] = sb_getblk(osb->sb, first_blkno);
1031 if (bhs[i] == NULL) {
1036 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1038 status = ocfs2_journal_access_eb(handle, et->et_ci,
1040 OCFS2_JOURNAL_ACCESS_CREATE);
1046 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1047 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1048 /* Ok, setup the minimal stuff here. */
1049 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1050 eb->h_blkno = cpu_to_le64(first_blkno);
1051 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1052 eb->h_suballoc_slot =
1053 cpu_to_le16(meta_ac->ac_alloc_slot);
1054 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1055 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1056 eb->h_list.l_count =
1057 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1059 suballoc_bit_start++;
1062 /* We'll also be dirtied by the caller, so
1063 * this isn't absolutely necessary. */
1064 ocfs2_journal_dirty(handle, bhs[i]);
1073 for(i = 0; i < wanted; i++) {
1083 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1085 * Returns the sum of the rightmost extent rec logical offset and
1088 * ocfs2_add_branch() uses this to determine what logical cluster
1089 * value should be populated into the leftmost new branch records.
1091 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1092 * value for the new topmost tree record.
1094 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1098 i = le16_to_cpu(el->l_next_free_rec) - 1;
1100 return le32_to_cpu(el->l_recs[i].e_cpos) +
1101 ocfs2_rec_clusters(el, &el->l_recs[i]);
1105 * Change range of the branches in the right most path according to the leaf
1106 * extent block's rightmost record.
1108 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1109 struct ocfs2_extent_tree *et)
1112 struct ocfs2_path *path = NULL;
1113 struct ocfs2_extent_list *el;
1114 struct ocfs2_extent_rec *rec;
1116 path = ocfs2_new_path_from_et(et);
1122 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1128 status = ocfs2_extend_trans(handle, path_num_items(path));
1134 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1140 el = path_leaf_el(path);
1141 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1143 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1146 ocfs2_free_path(path);
1151 * Add an entire tree branch to our inode. eb_bh is the extent block
1152 * to start at, if we don't want to start the branch at the root
1155 * last_eb_bh is required as we have to update it's next_leaf pointer
1156 * for the new last extent block.
1158 * the new branch will be 'empty' in the sense that every block will
1159 * contain a single record with cluster count == 0.
1161 static int ocfs2_add_branch(handle_t *handle,
1162 struct ocfs2_extent_tree *et,
1163 struct buffer_head *eb_bh,
1164 struct buffer_head **last_eb_bh,
1165 struct ocfs2_alloc_context *meta_ac)
1167 int status, new_blocks, i, block_given = 0;
1168 u64 next_blkno, new_last_eb_blk;
1169 struct buffer_head *bh;
1170 struct buffer_head **new_eb_bhs = NULL;
1171 struct ocfs2_extent_block *eb;
1172 struct ocfs2_extent_list *eb_el;
1173 struct ocfs2_extent_list *el;
1174 u32 new_cpos, root_end;
1176 BUG_ON(!last_eb_bh || !*last_eb_bh);
1179 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1182 el = et->et_root_el;
1184 /* we never add a branch to a leaf. */
1185 BUG_ON(!el->l_tree_depth);
1187 new_blocks = le16_to_cpu(el->l_tree_depth);
1189 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1190 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1191 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1194 * If there is a gap before the root end and the real end
1195 * of the righmost leaf block, we need to remove the gap
1196 * between new_cpos and root_end first so that the tree
1197 * is consistent after we add a new branch(it will start
1200 if (root_end > new_cpos) {
1201 trace_ocfs2_adjust_rightmost_branch(
1202 (unsigned long long)
1203 ocfs2_metadata_cache_owner(et->et_ci),
1204 root_end, new_cpos);
1206 status = ocfs2_adjust_rightmost_branch(handle, et);
1213 /* allocate the number of new eb blocks we need */
1214 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1222 /* Firstyly, try to reuse dealloc since we have already estimated how
1223 * many extent blocks we may use.
1225 if (!ocfs2_is_dealloc_empty(et)) {
1226 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1227 new_eb_bhs, new_blocks,
1235 BUG_ON(block_given > new_blocks);
1237 if (block_given < new_blocks) {
1239 status = ocfs2_create_new_meta_bhs(handle, et,
1240 new_blocks - block_given,
1242 &new_eb_bhs[block_given]);
1249 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1250 * linked with the rest of the tree.
1251 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1253 * when we leave the loop, new_last_eb_blk will point to the
1254 * newest leaf, and next_blkno will point to the topmost extent
1256 next_blkno = new_last_eb_blk = 0;
1257 for(i = 0; i < new_blocks; i++) {
1259 eb = (struct ocfs2_extent_block *) bh->b_data;
1260 /* ocfs2_create_new_meta_bhs() should create it right! */
1261 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1262 eb_el = &eb->h_list;
1264 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1265 OCFS2_JOURNAL_ACCESS_CREATE);
1271 eb->h_next_leaf_blk = 0;
1272 eb_el->l_tree_depth = cpu_to_le16(i);
1273 eb_el->l_next_free_rec = cpu_to_le16(1);
1275 * This actually counts as an empty extent as
1278 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1279 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1281 * eb_el isn't always an interior node, but even leaf
1282 * nodes want a zero'd flags and reserved field so
1283 * this gets the whole 32 bits regardless of use.
1285 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1286 if (!eb_el->l_tree_depth)
1287 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1289 ocfs2_journal_dirty(handle, bh);
1290 next_blkno = le64_to_cpu(eb->h_blkno);
1293 /* This is a bit hairy. We want to update up to three blocks
1294 * here without leaving any of them in an inconsistent state
1295 * in case of error. We don't have to worry about
1296 * journal_dirty erroring as it won't unless we've aborted the
1297 * handle (in which case we would never be here) so reserving
1298 * the write with journal_access is all we need to do. */
1299 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1300 OCFS2_JOURNAL_ACCESS_WRITE);
1305 status = ocfs2_et_root_journal_access(handle, et,
1306 OCFS2_JOURNAL_ACCESS_WRITE);
1312 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1313 OCFS2_JOURNAL_ACCESS_WRITE);
1320 /* Link the new branch into the rest of the tree (el will
1321 * either be on the root_bh, or the extent block passed in. */
1322 i = le16_to_cpu(el->l_next_free_rec);
1323 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1324 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1325 el->l_recs[i].e_int_clusters = 0;
1326 le16_add_cpu(&el->l_next_free_rec, 1);
1328 /* fe needs a new last extent block pointer, as does the
1329 * next_leaf on the previously last-extent-block. */
1330 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1332 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1333 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1335 ocfs2_journal_dirty(handle, *last_eb_bh);
1336 ocfs2_journal_dirty(handle, et->et_root_bh);
1338 ocfs2_journal_dirty(handle, eb_bh);
1341 * Some callers want to track the rightmost leaf so pass it
1344 brelse(*last_eb_bh);
1345 get_bh(new_eb_bhs[0]);
1346 *last_eb_bh = new_eb_bhs[0];
1351 for (i = 0; i < new_blocks; i++)
1352 brelse(new_eb_bhs[i]);
1360 * adds another level to the allocation tree.
1361 * returns back the new extent block so you can add a branch to it
1364 static int ocfs2_shift_tree_depth(handle_t *handle,
1365 struct ocfs2_extent_tree *et,
1366 struct ocfs2_alloc_context *meta_ac,
1367 struct buffer_head **ret_new_eb_bh)
1369 int status, i, block_given = 0;
1371 struct buffer_head *new_eb_bh = NULL;
1372 struct ocfs2_extent_block *eb;
1373 struct ocfs2_extent_list *root_el;
1374 struct ocfs2_extent_list *eb_el;
1376 if (!ocfs2_is_dealloc_empty(et)) {
1377 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1380 } else if (meta_ac) {
1381 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1393 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1394 /* ocfs2_create_new_meta_bhs() should create it right! */
1395 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1397 eb_el = &eb->h_list;
1398 root_el = et->et_root_el;
1400 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1401 OCFS2_JOURNAL_ACCESS_CREATE);
1407 /* copy the root extent list data into the new extent block */
1408 eb_el->l_tree_depth = root_el->l_tree_depth;
1409 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1410 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1411 eb_el->l_recs[i] = root_el->l_recs[i];
1413 ocfs2_journal_dirty(handle, new_eb_bh);
1415 status = ocfs2_et_root_journal_access(handle, et,
1416 OCFS2_JOURNAL_ACCESS_WRITE);
1422 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1424 /* update root_bh now */
1425 le16_add_cpu(&root_el->l_tree_depth, 1);
1426 root_el->l_recs[0].e_cpos = 0;
1427 root_el->l_recs[0].e_blkno = eb->h_blkno;
1428 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1429 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1430 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1431 root_el->l_next_free_rec = cpu_to_le16(1);
1433 /* If this is our 1st tree depth shift, then last_eb_blk
1434 * becomes the allocated extent block */
1435 if (root_el->l_tree_depth == cpu_to_le16(1))
1436 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1438 ocfs2_journal_dirty(handle, et->et_root_bh);
1440 *ret_new_eb_bh = new_eb_bh;
1450 * Should only be called when there is no space left in any of the
1451 * leaf nodes. What we want to do is find the lowest tree depth
1452 * non-leaf extent block with room for new records. There are three
1453 * valid results of this search:
1455 * 1) a lowest extent block is found, then we pass it back in
1456 * *lowest_eb_bh and return '0'
1458 * 2) the search fails to find anything, but the root_el has room. We
1459 * pass NULL back in *lowest_eb_bh, but still return '0'
1461 * 3) the search fails to find anything AND the root_el is full, in
1462 * which case we return > 0
1464 * return status < 0 indicates an error.
1466 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1467 struct buffer_head **target_bh)
1471 struct ocfs2_extent_block *eb;
1472 struct ocfs2_extent_list *el;
1473 struct buffer_head *bh = NULL;
1474 struct buffer_head *lowest_bh = NULL;
1478 el = et->et_root_el;
1480 while(le16_to_cpu(el->l_tree_depth) > 1) {
1481 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1482 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1483 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1484 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1487 i = le16_to_cpu(el->l_next_free_rec) - 1;
1488 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1490 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1491 "Owner %llu has extent list where extent # %d has no physical block start\n",
1492 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1499 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1505 eb = (struct ocfs2_extent_block *) bh->b_data;
1508 if (le16_to_cpu(el->l_next_free_rec) <
1509 le16_to_cpu(el->l_count)) {
1516 /* If we didn't find one and the fe doesn't have any room,
1517 * then return '1' */
1518 el = et->et_root_el;
1519 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1522 *target_bh = lowest_bh;
1530 * Grow a b-tree so that it has more records.
1532 * We might shift the tree depth in which case existing paths should
1533 * be considered invalid.
1535 * Tree depth after the grow is returned via *final_depth.
1537 * *last_eb_bh will be updated by ocfs2_add_branch().
1539 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1540 int *final_depth, struct buffer_head **last_eb_bh,
1541 struct ocfs2_alloc_context *meta_ac)
1544 struct ocfs2_extent_list *el = et->et_root_el;
1545 int depth = le16_to_cpu(el->l_tree_depth);
1546 struct buffer_head *bh = NULL;
1548 BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));
1550 shift = ocfs2_find_branch_target(et, &bh);
1557 /* We traveled all the way to the bottom of the allocation tree
1558 * and didn't find room for any more extents - we need to add
1559 * another tree level */
1562 trace_ocfs2_grow_tree(
1563 (unsigned long long)
1564 ocfs2_metadata_cache_owner(et->et_ci),
1567 /* ocfs2_shift_tree_depth will return us a buffer with
1568 * the new extent block (so we can pass that to
1569 * ocfs2_add_branch). */
1570 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1578 * Special case: we have room now if we shifted from
1579 * tree_depth 0, so no more work needs to be done.
1581 * We won't be calling add_branch, so pass
1582 * back *last_eb_bh as the new leaf. At depth
1583 * zero, it should always be null so there's
1584 * no reason to brelse.
1586 BUG_ON(*last_eb_bh);
1593 /* call ocfs2_add_branch to add the final part of the tree with
1595 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1602 *final_depth = depth;
1608 * This function will discard the rightmost extent record.
1610 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1612 int next_free = le16_to_cpu(el->l_next_free_rec);
1613 int count = le16_to_cpu(el->l_count);
1614 unsigned int num_bytes;
1617 /* This will cause us to go off the end of our extent list. */
1618 BUG_ON(next_free >= count);
1620 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1622 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1625 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1626 struct ocfs2_extent_rec *insert_rec)
1628 int i, insert_index, next_free, has_empty, num_bytes;
1629 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1630 struct ocfs2_extent_rec *rec;
1632 next_free = le16_to_cpu(el->l_next_free_rec);
1633 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1637 /* The tree code before us didn't allow enough room in the leaf. */
1638 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1641 * The easiest way to approach this is to just remove the
1642 * empty extent and temporarily decrement next_free.
1646 * If next_free was 1 (only an empty extent), this
1647 * loop won't execute, which is fine. We still want
1648 * the decrement above to happen.
1650 for(i = 0; i < (next_free - 1); i++)
1651 el->l_recs[i] = el->l_recs[i+1];
1657 * Figure out what the new record index should be.
1659 for(i = 0; i < next_free; i++) {
1660 rec = &el->l_recs[i];
1662 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1667 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1668 has_empty, next_free,
1669 le16_to_cpu(el->l_count));
1671 BUG_ON(insert_index < 0);
1672 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1673 BUG_ON(insert_index > next_free);
1676 * No need to memmove if we're just adding to the tail.
1678 if (insert_index != next_free) {
1679 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1681 num_bytes = next_free - insert_index;
1682 num_bytes *= sizeof(struct ocfs2_extent_rec);
1683 memmove(&el->l_recs[insert_index + 1],
1684 &el->l_recs[insert_index],
1689 * Either we had an empty extent, and need to re-increment or
1690 * there was no empty extent on a non full rightmost leaf node,
1691 * in which case we still need to increment.
1694 el->l_next_free_rec = cpu_to_le16(next_free);
1696 * Make sure none of the math above just messed up our tree.
1698 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1700 el->l_recs[insert_index] = *insert_rec;
1704 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1706 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1708 BUG_ON(num_recs == 0);
1710 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1712 size = num_recs * sizeof(struct ocfs2_extent_rec);
1713 memmove(&el->l_recs[0], &el->l_recs[1], size);
1714 memset(&el->l_recs[num_recs], 0,
1715 sizeof(struct ocfs2_extent_rec));
1716 el->l_next_free_rec = cpu_to_le16(num_recs);
1721 * Create an empty extent record .
1723 * l_next_free_rec may be updated.
1725 * If an empty extent already exists do nothing.
1727 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1729 int next_free = le16_to_cpu(el->l_next_free_rec);
1731 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1736 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1739 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1740 "Asked to create an empty extent in a full list:\n"
1741 "count = %u, tree depth = %u",
1742 le16_to_cpu(el->l_count),
1743 le16_to_cpu(el->l_tree_depth));
1745 ocfs2_shift_records_right(el);
1748 le16_add_cpu(&el->l_next_free_rec, 1);
1749 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1753 * For a rotation which involves two leaf nodes, the "root node" is
1754 * the lowest level tree node which contains a path to both leafs. This
1755 * resulting set of information can be used to form a complete "subtree"
1757 * This function is passed two full paths from the dinode down to a
1758 * pair of adjacent leaves. It's task is to figure out which path
1759 * index contains the subtree root - this can be the root index itself
1760 * in a worst-case rotation.
1762 * The array index of the subtree root is passed back.
1764 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1765 struct ocfs2_path *left,
1766 struct ocfs2_path *right)
1771 * Check that the caller passed in two paths from the same tree.
1773 BUG_ON(path_root_bh(left) != path_root_bh(right));
1779 * The caller didn't pass two adjacent paths.
1781 mlog_bug_on_msg(i > left->p_tree_depth,
1782 "Owner %llu, left depth %u, right depth %u\n"
1783 "left leaf blk %llu, right leaf blk %llu\n",
1784 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1785 left->p_tree_depth, right->p_tree_depth,
1786 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1787 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1788 } while (left->p_node[i].bh->b_blocknr ==
1789 right->p_node[i].bh->b_blocknr);
1794 typedef void (path_insert_t)(void *, struct buffer_head *);
1797 * Traverse a btree path in search of cpos, starting at root_el.
1799 * This code can be called with a cpos larger than the tree, in which
1800 * case it will return the rightmost path.
1802 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1803 struct ocfs2_extent_list *root_el, u32 cpos,
1804 path_insert_t *func, void *data)
1809 struct buffer_head *bh = NULL;
1810 struct ocfs2_extent_block *eb;
1811 struct ocfs2_extent_list *el;
1812 struct ocfs2_extent_rec *rec;
1815 while (el->l_tree_depth) {
1816 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1817 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1818 "Owner %llu has empty extent list at depth %u\n",
1819 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1820 le16_to_cpu(el->l_tree_depth));
1826 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1827 rec = &el->l_recs[i];
1830 * In the case that cpos is off the allocation
1831 * tree, this should just wind up returning the
1834 range = le32_to_cpu(rec->e_cpos) +
1835 ocfs2_rec_clusters(el, rec);
1836 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1840 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1842 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1843 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1844 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1845 le16_to_cpu(el->l_tree_depth), i);
1852 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1858 eb = (struct ocfs2_extent_block *) bh->b_data;
1861 if (le16_to_cpu(el->l_next_free_rec) >
1862 le16_to_cpu(el->l_count)) {
1863 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1864 "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1865 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1866 (unsigned long long)bh->b_blocknr,
1867 le16_to_cpu(el->l_next_free_rec),
1868 le16_to_cpu(el->l_count));
1879 * Catch any trailing bh that the loop didn't handle.
1887 * Given an initialized path (that is, it has a valid root extent
1888 * list), this function will traverse the btree in search of the path
1889 * which would contain cpos.
1891 * The path traveled is recorded in the path structure.
1893 * Note that this will not do any comparisons on leaf node extent
1894 * records, so it will work fine in the case that we just added a tree
1897 struct find_path_data {
1899 struct ocfs2_path *path;
1901 static void find_path_ins(void *data, struct buffer_head *bh)
1903 struct find_path_data *fp = data;
1906 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1909 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1910 struct ocfs2_path *path, u32 cpos)
1912 struct find_path_data data;
1916 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1917 find_path_ins, &data);
1920 static void find_leaf_ins(void *data, struct buffer_head *bh)
1922 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1923 struct ocfs2_extent_list *el = &eb->h_list;
1924 struct buffer_head **ret = data;
1926 /* We want to retain only the leaf block. */
1927 if (le16_to_cpu(el->l_tree_depth) == 0) {
1933 * Find the leaf block in the tree which would contain cpos. No
1934 * checking of the actual leaf is done.
1936 * Some paths want to call this instead of allocating a path structure
1937 * and calling ocfs2_find_path().
1939 * This function doesn't handle non btree extent lists.
1941 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1942 struct ocfs2_extent_list *root_el, u32 cpos,
1943 struct buffer_head **leaf_bh)
1946 struct buffer_head *bh = NULL;
1948 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1960 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1962 * Basically, we've moved stuff around at the bottom of the tree and
1963 * we need to fix up the extent records above the changes to reflect
1966 * left_rec: the record on the left.
1967 * right_rec: the record to the right of left_rec
1968 * right_child_el: is the child list pointed to by right_rec
1970 * By definition, this only works on interior nodes.
1972 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1973 struct ocfs2_extent_rec *right_rec,
1974 struct ocfs2_extent_list *right_child_el)
1976 u32 left_clusters, right_end;
1979 * Interior nodes never have holes. Their cpos is the cpos of
1980 * the leftmost record in their child list. Their cluster
1981 * count covers the full theoretical range of their child list
1982 * - the range between their cpos and the cpos of the record
1983 * immediately to their right.
1985 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1986 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1987 BUG_ON(right_child_el->l_tree_depth);
1988 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1989 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1991 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1992 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1995 * Calculate the rightmost cluster count boundary before
1996 * moving cpos - we will need to adjust clusters after
1997 * updating e_cpos to keep the same highest cluster count.
1999 right_end = le32_to_cpu(right_rec->e_cpos);
2000 right_end += le32_to_cpu(right_rec->e_int_clusters);
2002 right_rec->e_cpos = left_rec->e_cpos;
2003 le32_add_cpu(&right_rec->e_cpos, left_clusters);
2005 right_end -= le32_to_cpu(right_rec->e_cpos);
2006 right_rec->e_int_clusters = cpu_to_le32(right_end);
2010 * Adjust the adjacent root node records involved in a
2011 * rotation. left_el_blkno is passed in as a key so that we can easily
2012 * find it's index in the root list.
2014 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
2015 struct ocfs2_extent_list *left_el,
2016 struct ocfs2_extent_list *right_el,
2021 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2022 le16_to_cpu(left_el->l_tree_depth));
2024 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2025 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2030 * The path walking code should have never returned a root and
2031 * two paths which are not adjacent.
2033 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2035 ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
2036 &root_el->l_recs[i + 1], right_el);
2040 * We've changed a leaf block (in right_path) and need to reflect that
2041 * change back up the subtree.
2043 * This happens in multiple places:
2044 * - When we've moved an extent record from the left path leaf to the right
2045 * path leaf to make room for an empty extent in the left path leaf.
2046 * - When our insert into the right path leaf is at the leftmost edge
2047 * and requires an update of the path immediately to it's left. This
2048 * can occur at the end of some types of rotation and appending inserts.
2049 * - When we've adjusted the last extent record in the left path leaf and the
2050 * 1st extent record in the right path leaf during cross extent block merge.
2052 static void ocfs2_complete_edge_insert(handle_t *handle,
2053 struct ocfs2_path *left_path,
2054 struct ocfs2_path *right_path,
2058 struct ocfs2_extent_list *el, *left_el, *right_el;
2059 struct ocfs2_extent_rec *left_rec, *right_rec;
2060 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2063 * Update the counts and position values within all the
2064 * interior nodes to reflect the leaf rotation we just did.
2066 * The root node is handled below the loop.
2068 * We begin the loop with right_el and left_el pointing to the
2069 * leaf lists and work our way up.
2071 * NOTE: within this loop, left_el and right_el always refer
2072 * to the *child* lists.
2074 left_el = path_leaf_el(left_path);
2075 right_el = path_leaf_el(right_path);
2076 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2077 trace_ocfs2_complete_edge_insert(i);
2080 * One nice property of knowing that all of these
2081 * nodes are below the root is that we only deal with
2082 * the leftmost right node record and the rightmost
2085 el = left_path->p_node[i].el;
2086 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2087 left_rec = &el->l_recs[idx];
2089 el = right_path->p_node[i].el;
2090 right_rec = &el->l_recs[0];
2092 ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);
2094 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2095 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2098 * Setup our list pointers now so that the current
2099 * parents become children in the next iteration.
2101 left_el = left_path->p_node[i].el;
2102 right_el = right_path->p_node[i].el;
2106 * At the root node, adjust the two adjacent records which
2107 * begin our path to the leaves.
2110 el = left_path->p_node[subtree_index].el;
2111 left_el = left_path->p_node[subtree_index + 1].el;
2112 right_el = right_path->p_node[subtree_index + 1].el;
2114 ocfs2_adjust_root_records(el, left_el, right_el,
2115 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2117 root_bh = left_path->p_node[subtree_index].bh;
2119 ocfs2_journal_dirty(handle, root_bh);
2122 static int ocfs2_rotate_subtree_right(handle_t *handle,
2123 struct ocfs2_extent_tree *et,
2124 struct ocfs2_path *left_path,
2125 struct ocfs2_path *right_path,
2129 struct buffer_head *right_leaf_bh;
2130 struct buffer_head *left_leaf_bh = NULL;
2131 struct buffer_head *root_bh;
2132 struct ocfs2_extent_list *right_el, *left_el;
2133 struct ocfs2_extent_rec move_rec;
2135 left_leaf_bh = path_leaf_bh(left_path);
2136 left_el = path_leaf_el(left_path);
2138 if (left_el->l_next_free_rec != left_el->l_count) {
2139 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2140 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2141 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2142 (unsigned long long)left_leaf_bh->b_blocknr,
2143 le16_to_cpu(left_el->l_next_free_rec));
2148 * This extent block may already have an empty record, so we
2149 * return early if so.
2151 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2154 root_bh = left_path->p_node[subtree_index].bh;
2155 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2157 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2164 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2165 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2172 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2180 right_leaf_bh = path_leaf_bh(right_path);
2181 right_el = path_leaf_el(right_path);
2183 /* This is a code error, not a disk corruption. */
2184 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2185 "because rightmost leaf block %llu is empty\n",
2186 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2187 (unsigned long long)right_leaf_bh->b_blocknr);
2189 ocfs2_create_empty_extent(right_el);
2191 ocfs2_journal_dirty(handle, right_leaf_bh);
2193 /* Do the copy now. */
2194 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2195 move_rec = left_el->l_recs[i];
2196 right_el->l_recs[0] = move_rec;
2199 * Clear out the record we just copied and shift everything
2200 * over, leaving an empty extent in the left leaf.
2202 * We temporarily subtract from next_free_rec so that the
2203 * shift will lose the tail record (which is now defunct).
2205 le16_add_cpu(&left_el->l_next_free_rec, -1);
2206 ocfs2_shift_records_right(left_el);
2207 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2208 le16_add_cpu(&left_el->l_next_free_rec, 1);
2210 ocfs2_journal_dirty(handle, left_leaf_bh);
2212 ocfs2_complete_edge_insert(handle, left_path, right_path,
2220 * Given a full path, determine what cpos value would return us a path
2221 * containing the leaf immediately to the left of the current one.
2223 * Will return zero if the path passed in is already the leftmost path.
2225 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2226 struct ocfs2_path *path, u32 *cpos)
2230 struct ocfs2_extent_list *el;
2232 BUG_ON(path->p_tree_depth == 0);
2236 blkno = path_leaf_bh(path)->b_blocknr;
2238 /* Start at the tree node just above the leaf and work our way up. */
2239 i = path->p_tree_depth - 1;
2241 el = path->p_node[i].el;
2244 * Find the extent record just before the one in our
2247 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2248 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2252 * We've determined that the
2253 * path specified is already
2254 * the leftmost one - return a
2260 * The leftmost record points to our
2261 * leaf - we need to travel up the
2267 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2268 *cpos = *cpos + ocfs2_rec_clusters(el,
2269 &el->l_recs[j - 1]);
2276 * If we got here, we never found a valid node where
2277 * the tree indicated one should be.
2279 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2280 (unsigned long long)blkno);
2285 blkno = path->p_node[i].bh->b_blocknr;
2294 * Extend the transaction by enough credits to complete the rotation,
2295 * and still leave at least the original number of credits allocated
2296 * to this transaction.
2298 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2300 struct ocfs2_path *path)
2303 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2305 if (handle->h_buffer_credits < credits)
2306 ret = ocfs2_extend_trans(handle,
2307 credits - handle->h_buffer_credits);
2313 * Trap the case where we're inserting into the theoretical range past
2314 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2315 * whose cpos is less than ours into the right leaf.
2317 * It's only necessary to look at the rightmost record of the left
2318 * leaf because the logic that calls us should ensure that the
2319 * theoretical ranges in the path components above the leaves are
2322 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2325 struct ocfs2_extent_list *left_el;
2326 struct ocfs2_extent_rec *rec;
2329 left_el = path_leaf_el(left_path);
2330 next_free = le16_to_cpu(left_el->l_next_free_rec);
2331 rec = &left_el->l_recs[next_free - 1];
2333 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2338 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2340 int next_free = le16_to_cpu(el->l_next_free_rec);
2342 struct ocfs2_extent_rec *rec;
2347 rec = &el->l_recs[0];
2348 if (ocfs2_is_empty_extent(rec)) {
2352 rec = &el->l_recs[1];
2355 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2356 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2362 * Rotate all the records in a btree right one record, starting at insert_cpos.
2364 * The path to the rightmost leaf should be passed in.
2366 * The array is assumed to be large enough to hold an entire path (tree depth).
2368 * Upon successful return from this function:
2370 * - The 'right_path' array will contain a path to the leaf block
2371 * whose range contains e_cpos.
2372 * - That leaf block will have a single empty extent in list index 0.
2373 * - In the case that the rotation requires a post-insert update,
2374 * *ret_left_path will contain a valid path which can be passed to
2375 * ocfs2_insert_path().
2377 static int ocfs2_rotate_tree_right(handle_t *handle,
2378 struct ocfs2_extent_tree *et,
2379 enum ocfs2_split_type split,
2381 struct ocfs2_path *right_path,
2382 struct ocfs2_path **ret_left_path)
2384 int ret, start, orig_credits = handle->h_buffer_credits;
2386 struct ocfs2_path *left_path = NULL;
2387 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2389 *ret_left_path = NULL;
2391 left_path = ocfs2_new_path_from_path(right_path);
2398 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2404 trace_ocfs2_rotate_tree_right(
2405 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2409 * What we want to do here is:
2411 * 1) Start with the rightmost path.
2413 * 2) Determine a path to the leaf block directly to the left
2416 * 3) Determine the 'subtree root' - the lowest level tree node
2417 * which contains a path to both leaves.
2419 * 4) Rotate the subtree.
2421 * 5) Find the next subtree by considering the left path to be
2422 * the new right path.
2424 * The check at the top of this while loop also accepts
2425 * insert_cpos == cpos because cpos is only a _theoretical_
2426 * value to get us the left path - insert_cpos might very well
2427 * be filling that hole.
2429 * Stop at a cpos of '0' because we either started at the
2430 * leftmost branch (i.e., a tree with one branch and a
2431 * rotation inside of it), or we've gone as far as we can in
2432 * rotating subtrees.
2434 while (cpos && insert_cpos <= cpos) {
2435 trace_ocfs2_rotate_tree_right(
2436 (unsigned long long)
2437 ocfs2_metadata_cache_owner(et->et_ci),
2440 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2446 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2447 path_leaf_bh(right_path),
2448 "Owner %llu: error during insert of %u "
2449 "(left path cpos %u) results in two identical "
2450 "paths ending at %llu\n",
2451 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2453 (unsigned long long)
2454 path_leaf_bh(left_path)->b_blocknr);
2456 if (split == SPLIT_NONE &&
2457 ocfs2_rotate_requires_path_adjustment(left_path,
2461 * We've rotated the tree as much as we
2462 * should. The rest is up to
2463 * ocfs2_insert_path() to complete, after the
2464 * record insertion. We indicate this
2465 * situation by returning the left path.
2467 * The reason we don't adjust the records here
2468 * before the record insert is that an error
2469 * later might break the rule where a parent
2470 * record e_cpos will reflect the actual
2471 * e_cpos of the 1st nonempty record of the
2474 *ret_left_path = left_path;
2478 start = ocfs2_find_subtree_root(et, left_path, right_path);
2480 trace_ocfs2_rotate_subtree(start,
2481 (unsigned long long)
2482 right_path->p_node[start].bh->b_blocknr,
2483 right_path->p_tree_depth);
2485 ret = ocfs2_extend_rotate_transaction(handle, start,
2486 orig_credits, right_path);
2492 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2499 if (split != SPLIT_NONE &&
2500 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2503 * A rotate moves the rightmost left leaf
2504 * record over to the leftmost right leaf
2505 * slot. If we're doing an extent split
2506 * instead of a real insert, then we have to
2507 * check that the extent to be split wasn't
2508 * just moved over. If it was, then we can
2509 * exit here, passing left_path back -
2510 * ocfs2_split_extent() is smart enough to
2511 * search both leaves.
2513 *ret_left_path = left_path;
2518 * There is no need to re-read the next right path
2519 * as we know that it'll be our current left
2520 * path. Optimize by copying values instead.
2522 ocfs2_mv_path(right_path, left_path);
2524 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2532 ocfs2_free_path(left_path);
2538 static int ocfs2_update_edge_lengths(handle_t *handle,
2539 struct ocfs2_extent_tree *et,
2540 struct ocfs2_path *path)
2543 struct ocfs2_extent_rec *rec;
2544 struct ocfs2_extent_list *el;
2545 struct ocfs2_extent_block *eb;
2548 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2554 /* Path should always be rightmost. */
2555 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2556 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2559 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2560 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2561 rec = &el->l_recs[idx];
2562 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2564 for (i = 0; i < path->p_tree_depth; i++) {
2565 el = path->p_node[i].el;
2566 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2567 rec = &el->l_recs[idx];
2569 rec->e_int_clusters = cpu_to_le32(range);
2570 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2572 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2578 static void ocfs2_unlink_path(handle_t *handle,
2579 struct ocfs2_extent_tree *et,
2580 struct ocfs2_cached_dealloc_ctxt *dealloc,
2581 struct ocfs2_path *path, int unlink_start)
2584 struct ocfs2_extent_block *eb;
2585 struct ocfs2_extent_list *el;
2586 struct buffer_head *bh;
2588 for(i = unlink_start; i < path_num_items(path); i++) {
2589 bh = path->p_node[i].bh;
2591 eb = (struct ocfs2_extent_block *)bh->b_data;
2593 * Not all nodes might have had their final count
2594 * decremented by the caller - handle this here.
2597 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2599 "Inode %llu, attempted to remove extent block "
2600 "%llu with %u records\n",
2601 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2602 (unsigned long long)le64_to_cpu(eb->h_blkno),
2603 le16_to_cpu(el->l_next_free_rec));
2605 ocfs2_journal_dirty(handle, bh);
2606 ocfs2_remove_from_cache(et->et_ci, bh);
2610 el->l_next_free_rec = 0;
2611 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2613 ocfs2_journal_dirty(handle, bh);
2615 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2619 ocfs2_remove_from_cache(et->et_ci, bh);
2623 static void ocfs2_unlink_subtree(handle_t *handle,
2624 struct ocfs2_extent_tree *et,
2625 struct ocfs2_path *left_path,
2626 struct ocfs2_path *right_path,
2628 struct ocfs2_cached_dealloc_ctxt *dealloc)
2631 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2632 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2633 struct ocfs2_extent_block *eb;
2635 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2637 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2638 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2641 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2643 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2644 le16_add_cpu(&root_el->l_next_free_rec, -1);
2646 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2647 eb->h_next_leaf_blk = 0;
2649 ocfs2_journal_dirty(handle, root_bh);
2650 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2652 ocfs2_unlink_path(handle, et, dealloc, right_path,
2656 static int ocfs2_rotate_subtree_left(handle_t *handle,
2657 struct ocfs2_extent_tree *et,
2658 struct ocfs2_path *left_path,
2659 struct ocfs2_path *right_path,
2661 struct ocfs2_cached_dealloc_ctxt *dealloc,
2664 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2665 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2666 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2667 struct ocfs2_extent_block *eb;
2671 right_leaf_el = path_leaf_el(right_path);
2672 left_leaf_el = path_leaf_el(left_path);
2673 root_bh = left_path->p_node[subtree_index].bh;
2674 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2676 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2679 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2680 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2682 * It's legal for us to proceed if the right leaf is
2683 * the rightmost one and it has an empty extent. There
2684 * are two cases to handle - whether the leaf will be
2685 * empty after removal or not. If the leaf isn't empty
2686 * then just remove the empty extent up front. The
2687 * next block will handle empty leaves by flagging
2690 * Non rightmost leaves will throw -EAGAIN and the
2691 * caller can manually move the subtree and retry.
2694 if (eb->h_next_leaf_blk != 0ULL)
2697 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2698 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2699 path_leaf_bh(right_path),
2700 OCFS2_JOURNAL_ACCESS_WRITE);
2706 ocfs2_remove_empty_extent(right_leaf_el);
2708 right_has_empty = 1;
2711 if (eb->h_next_leaf_blk == 0ULL &&
2712 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2714 * We have to update i_last_eb_blk during the meta
2717 ret = ocfs2_et_root_journal_access(handle, et,
2718 OCFS2_JOURNAL_ACCESS_WRITE);
2724 del_right_subtree = 1;
2728 * Getting here with an empty extent in the right path implies
2729 * that it's the rightmost path and will be deleted.
2731 BUG_ON(right_has_empty && !del_right_subtree);
2733 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2740 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2741 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2748 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2756 if (!right_has_empty) {
2758 * Only do this if we're moving a real
2759 * record. Otherwise, the action is delayed until
2760 * after removal of the right path in which case we
2761 * can do a simple shift to remove the empty extent.
2763 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2764 memset(&right_leaf_el->l_recs[0], 0,
2765 sizeof(struct ocfs2_extent_rec));
2767 if (eb->h_next_leaf_blk == 0ULL) {
2769 * Move recs over to get rid of empty extent, decrease
2770 * next_free. This is allowed to remove the last
2771 * extent in our leaf (setting l_next_free_rec to
2772 * zero) - the delete code below won't care.
2774 ocfs2_remove_empty_extent(right_leaf_el);
2777 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2778 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2780 if (del_right_subtree) {
2781 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2782 subtree_index, dealloc);
2783 ret = ocfs2_update_edge_lengths(handle, et, left_path);
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.
2875 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2876 (unsigned long long)blkno);
2881 blkno = path->p_node[i].bh->b_blocknr;
2889 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2890 struct ocfs2_extent_tree *et,
2891 struct ocfs2_path *path)
2894 struct buffer_head *bh = path_leaf_bh(path);
2895 struct ocfs2_extent_list *el = path_leaf_el(path);
2897 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2900 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2901 path_num_items(path) - 1);
2907 ocfs2_remove_empty_extent(el);
2908 ocfs2_journal_dirty(handle, bh);
2914 static int __ocfs2_rotate_tree_left(handle_t *handle,
2915 struct ocfs2_extent_tree *et,
2917 struct ocfs2_path *path,
2918 struct ocfs2_cached_dealloc_ctxt *dealloc,
2919 struct ocfs2_path **empty_extent_path)
2921 int ret, subtree_root, deleted;
2923 struct ocfs2_path *left_path = NULL;
2924 struct ocfs2_path *right_path = NULL;
2925 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2927 if (!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, 0,
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;
3042 ret = ocfs2_et_sanity_check(et);
3046 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3052 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3061 * We have a path to the left of this one - it needs
3064 left_path = ocfs2_new_path_from_path(path);
3071 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3077 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3083 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3085 ocfs2_unlink_subtree(handle, et, left_path, path,
3086 subtree_index, dealloc);
3087 ret = ocfs2_update_edge_lengths(handle, et, left_path);
3093 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3094 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3097 * 'path' is also the leftmost path which
3098 * means it must be the only one. This gets
3099 * handled differently because we want to
3100 * revert the root back to having extents
3103 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3105 el = et->et_root_el;
3106 el->l_tree_depth = 0;
3107 el->l_next_free_rec = 0;
3108 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3110 ocfs2_et_set_last_eb_blk(et, 0);
3113 ocfs2_journal_dirty(handle, path_root_bh(path));
3116 ocfs2_free_path(left_path);
3120 static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3121 struct ocfs2_extent_tree *et,
3122 struct ocfs2_path *path,
3123 struct ocfs2_cached_dealloc_ctxt *dealloc)
3127 int credits = path->p_tree_depth * 2 + 1;
3129 handle = ocfs2_start_trans(osb, credits);
3130 if (IS_ERR(handle)) {
3131 ret = PTR_ERR(handle);
3136 ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3140 ocfs2_commit_trans(osb, handle);
3145 * Left rotation of btree records.
3147 * In many ways, this is (unsurprisingly) the opposite of right
3148 * rotation. We start at some non-rightmost path containing an empty
3149 * extent in the leaf block. The code works its way to the rightmost
3150 * path by rotating records to the left in every subtree.
3152 * This is used by any code which reduces the number of extent records
3153 * in a leaf. After removal, an empty record should be placed in the
3154 * leftmost list position.
3156 * This won't handle a length update of the rightmost path records if
3157 * the rightmost tree leaf record is removed so the caller is
3158 * responsible for detecting and correcting that.
3160 static int ocfs2_rotate_tree_left(handle_t *handle,
3161 struct ocfs2_extent_tree *et,
3162 struct ocfs2_path *path,
3163 struct ocfs2_cached_dealloc_ctxt *dealloc)
3165 int ret, orig_credits = handle->h_buffer_credits;
3166 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3167 struct ocfs2_extent_block *eb;
3168 struct ocfs2_extent_list *el;
3170 el = path_leaf_el(path);
3171 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3174 if (path->p_tree_depth == 0) {
3175 rightmost_no_delete:
3177 * Inline extents. This is trivially handled, so do
3180 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3187 * Handle rightmost branch now. There's several cases:
3188 * 1) simple rotation leaving records in there. That's trivial.
3189 * 2) rotation requiring a branch delete - there's no more
3190 * records left. Two cases of this:
3191 * a) There are branches to the left.
3192 * b) This is also the leftmost (the only) branch.
3194 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3195 * 2a) we need the left branch so that we can update it with the unlink
3196 * 2b) we need to bring the root back to inline extents.
3199 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3201 if (eb->h_next_leaf_blk == 0) {
3203 * This gets a bit tricky if we're going to delete the
3204 * rightmost path. Get the other cases out of the way
3207 if (le16_to_cpu(el->l_next_free_rec) > 1)
3208 goto rightmost_no_delete;
3210 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3211 ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3212 "Owner %llu has empty extent block at %llu\n",
3213 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3214 (unsigned long long)le64_to_cpu(eb->h_blkno));
3219 * XXX: The caller can not trust "path" any more after
3220 * this as it will have been deleted. What do we do?
3222 * In theory the rotate-for-merge code will never get
3223 * here because it'll always ask for a rotate in a
3227 ret = ocfs2_remove_rightmost_path(handle, et, path,
3235 * Now we can loop, remembering the path we get from -EAGAIN
3236 * and restarting from there.
3239 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3240 dealloc, &restart_path);
3241 if (ret && ret != -EAGAIN) {
3246 while (ret == -EAGAIN) {
3247 tmp_path = restart_path;
3248 restart_path = NULL;
3250 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3253 if (ret && ret != -EAGAIN) {
3258 ocfs2_free_path(tmp_path);
3266 ocfs2_free_path(tmp_path);
3267 ocfs2_free_path(restart_path);
3271 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3274 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3277 if (rec->e_leaf_clusters == 0) {
3279 * We consumed all of the merged-from record. An empty
3280 * extent cannot exist anywhere but the 1st array
3281 * position, so move things over if the merged-from
3282 * record doesn't occupy that position.
3284 * This creates a new empty extent so the caller
3285 * should be smart enough to have removed any existing
3289 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3290 size = index * sizeof(struct ocfs2_extent_rec);
3291 memmove(&el->l_recs[1], &el->l_recs[0], size);
3295 * Always memset - the caller doesn't check whether it
3296 * created an empty extent, so there could be junk in
3299 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3303 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3304 struct ocfs2_path *left_path,
3305 struct ocfs2_path **ret_right_path)
3309 struct ocfs2_path *right_path = NULL;
3310 struct ocfs2_extent_list *left_el;
3312 *ret_right_path = NULL;
3314 /* This function shouldn't be called for non-trees. */
3315 BUG_ON(left_path->p_tree_depth == 0);
3317 left_el = path_leaf_el(left_path);
3318 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3320 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3321 left_path, &right_cpos);
3327 /* This function shouldn't be called for the rightmost leaf. */
3328 BUG_ON(right_cpos == 0);
3330 right_path = ocfs2_new_path_from_path(left_path);
3337 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3343 *ret_right_path = right_path;
3346 ocfs2_free_path(right_path);
3351 * Remove split_rec clusters from the record at index and merge them
3352 * onto the beginning of the record "next" to it.
3353 * For index < l_count - 1, the next means the extent rec at index + 1.
3354 * For index == l_count - 1, the "next" means the 1st extent rec of the
3355 * next extent block.
3357 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3359 struct ocfs2_extent_tree *et,
3360 struct ocfs2_extent_rec *split_rec,
3363 int ret, next_free, i;
3364 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3365 struct ocfs2_extent_rec *left_rec;
3366 struct ocfs2_extent_rec *right_rec;
3367 struct ocfs2_extent_list *right_el;
3368 struct ocfs2_path *right_path = NULL;
3369 int subtree_index = 0;
3370 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3371 struct buffer_head *bh = path_leaf_bh(left_path);
3372 struct buffer_head *root_bh = NULL;
3374 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3375 left_rec = &el->l_recs[index];
3377 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3378 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3379 /* we meet with a cross extent block merge. */
3380 ret = ocfs2_get_right_path(et, left_path, &right_path);
3386 right_el = path_leaf_el(right_path);
3387 next_free = le16_to_cpu(right_el->l_next_free_rec);
3388 BUG_ON(next_free <= 0);
3389 right_rec = &right_el->l_recs[0];
3390 if (ocfs2_is_empty_extent(right_rec)) {
3391 BUG_ON(next_free <= 1);
3392 right_rec = &right_el->l_recs[1];
3395 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3396 le16_to_cpu(left_rec->e_leaf_clusters) !=
3397 le32_to_cpu(right_rec->e_cpos));
3399 subtree_index = ocfs2_find_subtree_root(et, left_path,
3402 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3403 handle->h_buffer_credits,
3410 root_bh = left_path->p_node[subtree_index].bh;
3411 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3413 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3420 for (i = subtree_index + 1;
3421 i < path_num_items(right_path); i++) {
3422 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3429 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3438 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3439 right_rec = &el->l_recs[index + 1];
3442 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3443 path_num_items(left_path) - 1);
3449 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3451 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3452 le64_add_cpu(&right_rec->e_blkno,
3453 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3455 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3457 ocfs2_cleanup_merge(el, index);
3459 ocfs2_journal_dirty(handle, bh);
3461 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3462 ocfs2_complete_edge_insert(handle, left_path, right_path,
3466 ocfs2_free_path(right_path);
3470 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3471 struct ocfs2_path *right_path,
3472 struct ocfs2_path **ret_left_path)
3476 struct ocfs2_path *left_path = NULL;
3478 *ret_left_path = NULL;
3480 /* This function shouldn't be called for non-trees. */
3481 BUG_ON(right_path->p_tree_depth == 0);
3483 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3484 right_path, &left_cpos);
3490 /* This function shouldn't be called for the leftmost leaf. */
3491 BUG_ON(left_cpos == 0);
3493 left_path = ocfs2_new_path_from_path(right_path);
3500 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3506 *ret_left_path = left_path;
3509 ocfs2_free_path(left_path);
3514 * Remove split_rec clusters from the record at index and merge them
3515 * onto the tail of the record "before" it.
3516 * For index > 0, the "before" means the extent rec at index - 1.
3518 * For index == 0, the "before" means the last record of the previous
3519 * extent block. And there is also a situation that we may need to
3520 * remove the rightmost leaf extent block in the right_path and change
3521 * the right path to indicate the new rightmost path.
3523 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3525 struct ocfs2_extent_tree *et,
3526 struct ocfs2_extent_rec *split_rec,
3527 struct ocfs2_cached_dealloc_ctxt *dealloc,
3530 int ret, i, subtree_index = 0, has_empty_extent = 0;
3531 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3532 struct ocfs2_extent_rec *left_rec;
3533 struct ocfs2_extent_rec *right_rec;
3534 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3535 struct buffer_head *bh = path_leaf_bh(right_path);
3536 struct buffer_head *root_bh = NULL;
3537 struct ocfs2_path *left_path = NULL;
3538 struct ocfs2_extent_list *left_el;
3542 right_rec = &el->l_recs[index];
3544 /* we meet with a cross extent block merge. */
3545 ret = ocfs2_get_left_path(et, right_path, &left_path);
3551 left_el = path_leaf_el(left_path);
3552 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3553 le16_to_cpu(left_el->l_count));
3555 left_rec = &left_el->l_recs[
3556 le16_to_cpu(left_el->l_next_free_rec) - 1];
3557 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3558 le16_to_cpu(left_rec->e_leaf_clusters) !=
3559 le32_to_cpu(split_rec->e_cpos));
3561 subtree_index = ocfs2_find_subtree_root(et, left_path,
3564 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3565 handle->h_buffer_credits,
3572 root_bh = left_path->p_node[subtree_index].bh;
3573 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3575 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3582 for (i = subtree_index + 1;
3583 i < path_num_items(right_path); i++) {
3584 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3591 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3599 left_rec = &el->l_recs[index - 1];
3600 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3601 has_empty_extent = 1;
3604 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3605 path_num_items(right_path) - 1);
3611 if (has_empty_extent && index == 1) {
3613 * The easy case - we can just plop the record right in.
3615 *left_rec = *split_rec;
3617 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3619 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3620 le64_add_cpu(&right_rec->e_blkno,
3621 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3623 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3625 ocfs2_cleanup_merge(el, index);
3627 ocfs2_journal_dirty(handle, bh);
3629 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3632 * In the situation that the right_rec is empty and the extent
3633 * block is empty also, ocfs2_complete_edge_insert can't handle
3634 * it and we need to delete the right extent block.
3636 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3637 le16_to_cpu(el->l_next_free_rec) == 1) {
3638 /* extend credit for ocfs2_remove_rightmost_path */
3639 ret = ocfs2_extend_rotate_transaction(handle, 0,
3640 handle->h_buffer_credits,
3647 ret = ocfs2_remove_rightmost_path(handle, et,
3655 /* Now the rightmost extent block has been deleted.
3656 * So we use the new rightmost path.
3658 ocfs2_mv_path(right_path, left_path);
3661 ocfs2_complete_edge_insert(handle, left_path,
3662 right_path, subtree_index);
3665 ocfs2_free_path(left_path);
3669 static int ocfs2_try_to_merge_extent(handle_t *handle,
3670 struct ocfs2_extent_tree *et,
3671 struct ocfs2_path *path,
3673 struct ocfs2_extent_rec *split_rec,
3674 struct ocfs2_cached_dealloc_ctxt *dealloc,
3675 struct ocfs2_merge_ctxt *ctxt)
3678 struct ocfs2_extent_list *el = path_leaf_el(path);
3679 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3681 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3683 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3684 /* extend credit for ocfs2_remove_rightmost_path */
3685 ret = ocfs2_extend_rotate_transaction(handle, 0,
3686 handle->h_buffer_credits,
3693 * The merge code will need to create an empty
3694 * extent to take the place of the newly
3695 * emptied slot. Remove any pre-existing empty
3696 * extents - having more than one in a leaf is
3699 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3705 rec = &el->l_recs[split_index];
3708 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3710 * Left-right contig implies this.
3712 BUG_ON(!ctxt->c_split_covers_rec);
3715 * Since the leftright insert always covers the entire
3716 * extent, this call will delete the insert record
3717 * entirely, resulting in an empty extent record added to
3720 * Since the adding of an empty extent shifts
3721 * everything back to the right, there's no need to
3722 * update split_index here.
3724 * When the split_index is zero, we need to merge it to the
3725 * prevoius extent block. It is more efficient and easier
3726 * if we do merge_right first and merge_left later.
3728 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3736 * We can only get this from logic error above.
3738 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3740 /* extend credit for ocfs2_remove_rightmost_path */
3741 ret = ocfs2_extend_rotate_transaction(handle, 0,
3742 handle->h_buffer_credits,
3749 /* The merge left us with an empty extent, remove it. */
3750 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3756 rec = &el->l_recs[split_index];
3759 * Note that we don't pass split_rec here on purpose -
3760 * we've merged it into the rec already.
3762 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3763 dealloc, split_index);
3770 /* extend credit for ocfs2_remove_rightmost_path */
3771 ret = ocfs2_extend_rotate_transaction(handle, 0,
3772 handle->h_buffer_credits,
3779 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3781 * Error from this last rotate is not critical, so
3782 * print but don't bubble it up.
3789 * Merge a record to the left or right.
3791 * 'contig_type' is relative to the existing record,
3792 * so for example, if we're "right contig", it's to
3793 * the record on the left (hence the left merge).
3795 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3796 ret = ocfs2_merge_rec_left(path, handle, et,
3804 ret = ocfs2_merge_rec_right(path, handle,
3813 if (ctxt->c_split_covers_rec) {
3814 /* extend credit for ocfs2_remove_rightmost_path */
3815 ret = ocfs2_extend_rotate_transaction(handle, 0,
3816 handle->h_buffer_credits,
3825 * The merge may have left an empty extent in
3826 * our leaf. Try to rotate it away.
3828 ret = ocfs2_rotate_tree_left(handle, et, path,
3840 static void ocfs2_subtract_from_rec(struct super_block *sb,
3841 enum ocfs2_split_type split,
3842 struct ocfs2_extent_rec *rec,
3843 struct ocfs2_extent_rec *split_rec)
3847 len_blocks = ocfs2_clusters_to_blocks(sb,
3848 le16_to_cpu(split_rec->e_leaf_clusters));
3850 if (split == SPLIT_LEFT) {
3852 * Region is on the left edge of the existing
3855 le32_add_cpu(&rec->e_cpos,
3856 le16_to_cpu(split_rec->e_leaf_clusters));
3857 le64_add_cpu(&rec->e_blkno, len_blocks);
3858 le16_add_cpu(&rec->e_leaf_clusters,
3859 -le16_to_cpu(split_rec->e_leaf_clusters));
3862 * Region is on the right edge of the existing
3865 le16_add_cpu(&rec->e_leaf_clusters,
3866 -le16_to_cpu(split_rec->e_leaf_clusters));
3871 * Do the final bits of extent record insertion at the target leaf
3872 * list. If this leaf is part of an allocation tree, it is assumed
3873 * that the tree above has been prepared.
3875 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3876 struct ocfs2_extent_rec *insert_rec,
3877 struct ocfs2_extent_list *el,
3878 struct ocfs2_insert_type *insert)
3880 int i = insert->ins_contig_index;
3882 struct ocfs2_extent_rec *rec;
3884 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3886 if (insert->ins_split != SPLIT_NONE) {
3887 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3889 rec = &el->l_recs[i];
3890 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3891 insert->ins_split, rec,
3897 * Contiguous insert - either left or right.
3899 if (insert->ins_contig != CONTIG_NONE) {
3900 rec = &el->l_recs[i];
3901 if (insert->ins_contig == CONTIG_LEFT) {
3902 rec->e_blkno = insert_rec->e_blkno;
3903 rec->e_cpos = insert_rec->e_cpos;
3905 le16_add_cpu(&rec->e_leaf_clusters,
3906 le16_to_cpu(insert_rec->e_leaf_clusters));
3911 * Handle insert into an empty leaf.
3913 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3914 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3915 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3916 el->l_recs[0] = *insert_rec;
3917 el->l_next_free_rec = cpu_to_le16(1);
3924 if (insert->ins_appending == APPEND_TAIL) {
3925 i = le16_to_cpu(el->l_next_free_rec) - 1;
3926 rec = &el->l_recs[i];
3927 range = le32_to_cpu(rec->e_cpos)
3928 + le16_to_cpu(rec->e_leaf_clusters);
3929 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3931 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3932 le16_to_cpu(el->l_count),
3933 "owner %llu, depth %u, count %u, next free %u, "
3934 "rec.cpos %u, rec.clusters %u, "
3935 "insert.cpos %u, insert.clusters %u\n",
3936 ocfs2_metadata_cache_owner(et->et_ci),
3937 le16_to_cpu(el->l_tree_depth),
3938 le16_to_cpu(el->l_count),
3939 le16_to_cpu(el->l_next_free_rec),
3940 le32_to_cpu(el->l_recs[i].e_cpos),
3941 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3942 le32_to_cpu(insert_rec->e_cpos),
3943 le16_to_cpu(insert_rec->e_leaf_clusters));
3945 el->l_recs[i] = *insert_rec;
3946 le16_add_cpu(&el->l_next_free_rec, 1);
3952 * Ok, we have to rotate.
3954 * At this point, it is safe to assume that inserting into an
3955 * empty leaf and appending to a leaf have both been handled
3958 * This leaf needs to have space, either by the empty 1st
3959 * extent record, or by virtue of an l_next_rec < l_count.
3961 ocfs2_rotate_leaf(el, insert_rec);
3964 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3965 struct ocfs2_extent_tree *et,
3966 struct ocfs2_path *path,
3967 struct ocfs2_extent_rec *insert_rec)
3970 struct buffer_head *bh;
3971 struct ocfs2_extent_list *el;
3972 struct ocfs2_extent_rec *rec;
3975 * Update everything except the leaf block.
3977 for (i = 0; i < path->p_tree_depth; i++) {
3978 bh = path->p_node[i].bh;
3979 el = path->p_node[i].el;
3981 next_free = le16_to_cpu(el->l_next_free_rec);
3982 if (next_free == 0) {
3983 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3984 "Owner %llu has a bad extent list\n",
3985 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3989 rec = &el->l_recs[next_free - 1];
3991 rec->e_int_clusters = insert_rec->e_cpos;
3992 le32_add_cpu(&rec->e_int_clusters,
3993 le16_to_cpu(insert_rec->e_leaf_clusters));
3994 le32_add_cpu(&rec->e_int_clusters,
3995 -le32_to_cpu(rec->e_cpos));
3997 ocfs2_journal_dirty(handle, bh);
4001 static int ocfs2_append_rec_to_path(handle_t *handle,
4002 struct ocfs2_extent_tree *et,
4003 struct ocfs2_extent_rec *insert_rec,
4004 struct ocfs2_path *right_path,
4005 struct ocfs2_path **ret_left_path)
4008 struct ocfs2_extent_list *el;
4009 struct ocfs2_path *left_path = NULL;
4011 *ret_left_path = NULL;
4014 * This shouldn't happen for non-trees. The extent rec cluster
4015 * count manipulation below only works for interior nodes.
4017 BUG_ON(right_path->p_tree_depth == 0);
4020 * If our appending insert is at the leftmost edge of a leaf,
4021 * then we might need to update the rightmost records of the
4024 el = path_leaf_el(right_path);
4025 next_free = le16_to_cpu(el->l_next_free_rec);
4026 if (next_free == 0 ||
4027 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4030 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4031 right_path, &left_cpos);
4037 trace_ocfs2_append_rec_to_path(
4038 (unsigned long long)
4039 ocfs2_metadata_cache_owner(et->et_ci),
4040 le32_to_cpu(insert_rec->e_cpos),
4044 * No need to worry if the append is already in the
4048 left_path = ocfs2_new_path_from_path(right_path);
4055 ret = ocfs2_find_path(et->et_ci, left_path,
4063 * ocfs2_insert_path() will pass the left_path to the
4069 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4075 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4077 *ret_left_path = left_path;
4081 ocfs2_free_path(left_path);
4086 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4087 struct ocfs2_path *left_path,
4088 struct ocfs2_path *right_path,
4089 struct ocfs2_extent_rec *split_rec,
4090 enum ocfs2_split_type split)
4093 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4094 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4095 struct ocfs2_extent_rec *rec, *tmprec;
4097 right_el = path_leaf_el(right_path);
4099 left_el = path_leaf_el(left_path);
4102 insert_el = right_el;
4103 index = ocfs2_search_extent_list(el, cpos);
4105 if (index == 0 && left_path) {
4106 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4109 * This typically means that the record
4110 * started in the left path but moved to the
4111 * right as a result of rotation. We either
4112 * move the existing record to the left, or we
4113 * do the later insert there.
4115 * In this case, the left path should always
4116 * exist as the rotate code will have passed
4117 * it back for a post-insert update.
4120 if (split == SPLIT_LEFT) {
4122 * It's a left split. Since we know
4123 * that the rotate code gave us an
4124 * empty extent in the left path, we
4125 * can just do the insert there.
4127 insert_el = left_el;
4130 * Right split - we have to move the
4131 * existing record over to the left
4132 * leaf. The insert will be into the
4133 * newly created empty extent in the
4136 tmprec = &right_el->l_recs[index];
4137 ocfs2_rotate_leaf(left_el, tmprec);
4140 memset(tmprec, 0, sizeof(*tmprec));
4141 index = ocfs2_search_extent_list(left_el, cpos);
4142 BUG_ON(index == -1);
4147 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4149 * Left path is easy - we can just allow the insert to
4153 insert_el = left_el;
4154 index = ocfs2_search_extent_list(el, cpos);
4155 BUG_ON(index == -1);
4158 rec = &el->l_recs[index];
4159 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4160 split, rec, split_rec);
4161 ocfs2_rotate_leaf(insert_el, split_rec);
4165 * This function only does inserts on an allocation b-tree. For tree
4166 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4168 * right_path is the path we want to do the actual insert
4169 * in. left_path should only be passed in if we need to update that
4170 * portion of the tree after an edge insert.
4172 static int ocfs2_insert_path(handle_t *handle,
4173 struct ocfs2_extent_tree *et,
4174 struct ocfs2_path *left_path,
4175 struct ocfs2_path *right_path,
4176 struct ocfs2_extent_rec *insert_rec,
4177 struct ocfs2_insert_type *insert)
4179 int ret, subtree_index;
4180 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4184 * There's a chance that left_path got passed back to
4185 * us without being accounted for in the
4186 * journal. Extend our transaction here to be sure we
4187 * can change those blocks.
4189 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4195 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4203 * Pass both paths to the journal. The majority of inserts
4204 * will be touching all components anyway.
4206 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4212 if (insert->ins_split != SPLIT_NONE) {
4214 * We could call ocfs2_insert_at_leaf() for some types
4215 * of splits, but it's easier to just let one separate
4216 * function sort it all out.
4218 ocfs2_split_record(et, left_path, right_path,
4219 insert_rec, insert->ins_split);
4222 * Split might have modified either leaf and we don't
4223 * have a guarantee that the later edge insert will
4224 * dirty this for us.
4227 ocfs2_journal_dirty(handle,
4228 path_leaf_bh(left_path));
4230 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4233 ocfs2_journal_dirty(handle, leaf_bh);
4237 * The rotate code has indicated that we need to fix
4238 * up portions of the tree after the insert.
4240 * XXX: Should we extend the transaction here?
4242 subtree_index = ocfs2_find_subtree_root(et, left_path,
4244 ocfs2_complete_edge_insert(handle, left_path, right_path,
4253 static int ocfs2_do_insert_extent(handle_t *handle,
4254 struct ocfs2_extent_tree *et,
4255 struct ocfs2_extent_rec *insert_rec,
4256 struct ocfs2_insert_type *type)
4258 int ret, rotate = 0;
4260 struct ocfs2_path *right_path = NULL;
4261 struct ocfs2_path *left_path = NULL;
4262 struct ocfs2_extent_list *el;
4264 el = et->et_root_el;
4266 ret = ocfs2_et_root_journal_access(handle, et,
4267 OCFS2_JOURNAL_ACCESS_WRITE);
4273 if (le16_to_cpu(el->l_tree_depth) == 0) {
4274 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4275 goto out_update_clusters;
4278 right_path = ocfs2_new_path_from_et(et);
4286 * Determine the path to start with. Rotations need the
4287 * rightmost path, everything else can go directly to the
4290 cpos = le32_to_cpu(insert_rec->e_cpos);
4291 if (type->ins_appending == APPEND_NONE &&
4292 type->ins_contig == CONTIG_NONE) {
4297 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4304 * Rotations and appends need special treatment - they modify
4305 * parts of the tree's above them.
4307 * Both might pass back a path immediate to the left of the
4308 * one being inserted to. This will be cause
4309 * ocfs2_insert_path() to modify the rightmost records of
4310 * left_path to account for an edge insert.
4312 * XXX: When modifying this code, keep in mind that an insert
4313 * can wind up skipping both of these two special cases...
4316 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4317 le32_to_cpu(insert_rec->e_cpos),
4318 right_path, &left_path);
4325 * ocfs2_rotate_tree_right() might have extended the
4326 * transaction without re-journaling our tree root.
4328 ret = ocfs2_et_root_journal_access(handle, et,
4329 OCFS2_JOURNAL_ACCESS_WRITE);
4334 } else if (type->ins_appending == APPEND_TAIL
4335 && type->ins_contig != CONTIG_LEFT) {
4336 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4337 right_path, &left_path);
4344 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4351 out_update_clusters:
4352 if (type->ins_split == SPLIT_NONE)
4353 ocfs2_et_update_clusters(et,
4354 le16_to_cpu(insert_rec->e_leaf_clusters));
4356 ocfs2_journal_dirty(handle, et->et_root_bh);
4359 ocfs2_free_path(left_path);
4360 ocfs2_free_path(right_path);
4365 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4366 struct ocfs2_path *path,
4367 struct ocfs2_extent_list *el, int index,
4368 struct ocfs2_extent_rec *split_rec,
4369 struct ocfs2_merge_ctxt *ctxt)
4372 enum ocfs2_contig_type ret = CONTIG_NONE;
4373 u32 left_cpos, right_cpos;
4374 struct ocfs2_extent_rec *rec = NULL;
4375 struct ocfs2_extent_list *new_el;
4376 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4377 struct buffer_head *bh;
4378 struct ocfs2_extent_block *eb;
4379 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4382 rec = &el->l_recs[index - 1];
4383 } else if (path->p_tree_depth > 0) {
4384 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4388 if (left_cpos != 0) {
4389 left_path = ocfs2_new_path_from_path(path);
4396 status = ocfs2_find_path(et->et_ci, left_path,
4399 goto free_left_path;
4401 new_el = path_leaf_el(left_path);
4403 if (le16_to_cpu(new_el->l_next_free_rec) !=
4404 le16_to_cpu(new_el->l_count)) {
4405 bh = path_leaf_bh(left_path);
4406 eb = (struct ocfs2_extent_block *)bh->b_data;
4407 status = ocfs2_error(sb,
4408 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4409 (unsigned long long)le64_to_cpu(eb->h_blkno),
4410 le16_to_cpu(new_el->l_next_free_rec),
4411 le16_to_cpu(new_el->l_count));
4412 goto free_left_path;
4414 rec = &new_el->l_recs[
4415 le16_to_cpu(new_el->l_next_free_rec) - 1];
4420 * We're careful to check for an empty extent record here -
4421 * the merge code will know what to do if it sees one.
4424 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4425 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4428 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4433 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4434 rec = &el->l_recs[index + 1];
4435 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4436 path->p_tree_depth > 0) {
4437 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4439 goto free_left_path;
4441 if (right_cpos == 0)
4442 goto free_left_path;
4444 right_path = ocfs2_new_path_from_path(path);
4448 goto free_left_path;
4451 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4453 goto free_right_path;
4455 new_el = path_leaf_el(right_path);
4456 rec = &new_el->l_recs[0];
4457 if (ocfs2_is_empty_extent(rec)) {
4458 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4459 bh = path_leaf_bh(right_path);
4460 eb = (struct ocfs2_extent_block *)bh->b_data;
4461 status = ocfs2_error(sb,
4462 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4463 (unsigned long long)le64_to_cpu(eb->h_blkno),
4464 le16_to_cpu(new_el->l_next_free_rec));
4465 goto free_right_path;
4467 rec = &new_el->l_recs[1];
4472 enum ocfs2_contig_type contig_type;
4474 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4476 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4477 ret = CONTIG_LEFTRIGHT;
4478 else if (ret == CONTIG_NONE)
4483 ocfs2_free_path(right_path);
4485 ocfs2_free_path(left_path);
4488 ctxt->c_contig_type = ret;
4493 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4494 struct ocfs2_insert_type *insert,
4495 struct ocfs2_extent_list *el,
4496 struct ocfs2_extent_rec *insert_rec)
4499 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4501 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4503 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4504 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4506 if (contig_type != CONTIG_NONE) {
4507 insert->ins_contig_index = i;
4511 insert->ins_contig = contig_type;
4513 if (insert->ins_contig != CONTIG_NONE) {
4514 struct ocfs2_extent_rec *rec =
4515 &el->l_recs[insert->ins_contig_index];
4516 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4517 le16_to_cpu(insert_rec->e_leaf_clusters);
4520 * Caller might want us to limit the size of extents, don't
4521 * calculate contiguousness if we might exceed that limit.
4523 if (et->et_max_leaf_clusters &&
4524 (len > et->et_max_leaf_clusters))
4525 insert->ins_contig = CONTIG_NONE;
4530 * This should only be called against the righmost leaf extent list.
4532 * ocfs2_figure_appending_type() will figure out whether we'll have to
4533 * insert at the tail of the rightmost leaf.
4535 * This should also work against the root extent list for tree's with 0
4536 * depth. If we consider the root extent list to be the rightmost leaf node
4537 * then the logic here makes sense.
4539 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4540 struct ocfs2_extent_list *el,
4541 struct ocfs2_extent_rec *insert_rec)
4544 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4545 struct ocfs2_extent_rec *rec;
4547 insert->ins_appending = APPEND_NONE;
4549 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4551 if (!el->l_next_free_rec)
4552 goto set_tail_append;
4554 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4555 /* Were all records empty? */
4556 if (le16_to_cpu(el->l_next_free_rec) == 1)
4557 goto set_tail_append;
4560 i = le16_to_cpu(el->l_next_free_rec) - 1;
4561 rec = &el->l_recs[i];
4564 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4565 goto set_tail_append;
4570 insert->ins_appending = APPEND_TAIL;
4574 * Helper function called at the beginning of an insert.
4576 * This computes a few things that are commonly used in the process of
4577 * inserting into the btree:
4578 * - Whether the new extent is contiguous with an existing one.
4579 * - The current tree depth.
4580 * - Whether the insert is an appending one.
4581 * - The total # of free records in the tree.
4583 * All of the information is stored on the ocfs2_insert_type
4586 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4587 struct buffer_head **last_eb_bh,
4588 struct ocfs2_extent_rec *insert_rec,
4590 struct ocfs2_insert_type *insert)
4593 struct ocfs2_extent_block *eb;
4594 struct ocfs2_extent_list *el;
4595 struct ocfs2_path *path = NULL;
4596 struct buffer_head *bh = NULL;
4598 insert->ins_split = SPLIT_NONE;
4600 el = et->et_root_el;
4601 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4603 if (el->l_tree_depth) {
4605 * If we have tree depth, we read in the
4606 * rightmost extent block ahead of time as
4607 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4608 * may want it later.
4610 ret = ocfs2_read_extent_block(et->et_ci,
4611 ocfs2_et_get_last_eb_blk(et),
4617 eb = (struct ocfs2_extent_block *) bh->b_data;
4622 * Unless we have a contiguous insert, we'll need to know if
4623 * there is room left in our allocation tree for another
4626 * XXX: This test is simplistic, we can search for empty
4627 * extent records too.
4629 *free_records = le16_to_cpu(el->l_count) -
4630 le16_to_cpu(el->l_next_free_rec);
4632 if (!insert->ins_tree_depth) {
4633 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4634 ocfs2_figure_appending_type(insert, el, insert_rec);
4638 path = ocfs2_new_path_from_et(et);
4646 * In the case that we're inserting past what the tree
4647 * currently accounts for, ocfs2_find_path() will return for
4648 * us the rightmost tree path. This is accounted for below in
4649 * the appending code.
4651 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4657 el = path_leaf_el(path);
4660 * Now that we have the path, there's two things we want to determine:
4661 * 1) Contiguousness (also set contig_index if this is so)
4663 * 2) Are we doing an append? We can trivially break this up
4664 * into two types of appends: simple record append, or a
4665 * rotate inside the tail leaf.
4667 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4670 * The insert code isn't quite ready to deal with all cases of
4671 * left contiguousness. Specifically, if it's an insert into
4672 * the 1st record in a leaf, it will require the adjustment of
4673 * cluster count on the last record of the path directly to it's
4674 * left. For now, just catch that case and fool the layers
4675 * above us. This works just fine for tree_depth == 0, which
4676 * is why we allow that above.
4678 if (insert->ins_contig == CONTIG_LEFT &&
4679 insert->ins_contig_index == 0)
4680 insert->ins_contig = CONTIG_NONE;
4683 * Ok, so we can simply compare against last_eb to figure out
4684 * whether the path doesn't exist. This will only happen in
4685 * the case that we're doing a tail append, so maybe we can
4686 * take advantage of that information somehow.
4688 if (ocfs2_et_get_last_eb_blk(et) ==
4689 path_leaf_bh(path)->b_blocknr) {
4691 * Ok, ocfs2_find_path() returned us the rightmost
4692 * tree path. This might be an appending insert. There are
4694 * 1) We're doing a true append at the tail:
4695 * -This might even be off the end of the leaf
4696 * 2) We're "appending" by rotating in the tail
4698 ocfs2_figure_appending_type(insert, el, insert_rec);
4702 ocfs2_free_path(path);
4712 * Insert an extent into a btree.
4714 * The caller needs to update the owning btree's cluster count.
4716 int ocfs2_insert_extent(handle_t *handle,
4717 struct ocfs2_extent_tree *et,
4722 struct ocfs2_alloc_context *meta_ac)
4726 struct buffer_head *last_eb_bh = NULL;
4727 struct ocfs2_insert_type insert = {0, };
4728 struct ocfs2_extent_rec rec;
4730 trace_ocfs2_insert_extent_start(
4731 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4732 cpos, new_clusters);
4734 memset(&rec, 0, sizeof(rec));
4735 rec.e_cpos = cpu_to_le32(cpos);
4736 rec.e_blkno = cpu_to_le64(start_blk);
4737 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4738 rec.e_flags = flags;
4739 status = ocfs2_et_insert_check(et, &rec);
4745 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4746 &free_records, &insert);
4752 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4753 insert.ins_contig_index, free_records,
4754 insert.ins_tree_depth);
4756 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4757 status = ocfs2_grow_tree(handle, et,
4758 &insert.ins_tree_depth, &last_eb_bh,
4766 /* Finally, we can add clusters. This might rotate the tree for us. */
4767 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4771 ocfs2_et_extent_map_insert(et, &rec);
4780 * Allcate and add clusters into the extent b-tree.
4781 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4782 * The extent b-tree's root is specified by et, and
4783 * it is not limited to the file storage. Any extent tree can use this
4784 * function if it implements the proper ocfs2_extent_tree.
4786 int ocfs2_add_clusters_in_btree(handle_t *handle,
4787 struct ocfs2_extent_tree *et,
4788 u32 *logical_offset,
4789 u32 clusters_to_add,
4791 struct ocfs2_alloc_context *data_ac,
4792 struct ocfs2_alloc_context *meta_ac,
4793 enum ocfs2_alloc_restarted *reason_ret)
4795 int status = 0, err = 0;
4798 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4799 u32 bit_off, num_bits;
4802 struct ocfs2_super *osb =
4803 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4805 BUG_ON(!clusters_to_add);
4808 flags = OCFS2_EXT_UNWRITTEN;
4810 free_extents = ocfs2_num_free_extents(et);
4811 if (free_extents < 0) {
4812 status = free_extents;
4817 /* there are two cases which could cause us to EAGAIN in the
4818 * we-need-more-metadata case:
4819 * 1) we haven't reserved *any*
4820 * 2) we are so fragmented, we've needed to add metadata too
4822 if (!free_extents && !meta_ac) {
4825 reason = RESTART_META;
4827 } else if ((!free_extents)
4828 && (ocfs2_alloc_context_bits_left(meta_ac)
4829 < ocfs2_extend_meta_needed(et->et_root_el))) {
4832 reason = RESTART_META;
4836 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4837 clusters_to_add, &bit_off, &num_bits);
4839 if (status != -ENOSPC)
4844 BUG_ON(num_bits > clusters_to_add);
4846 /* reserve our write early -- insert_extent may update the tree root */
4847 status = ocfs2_et_root_journal_access(handle, et,
4848 OCFS2_JOURNAL_ACCESS_WRITE);
4855 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4856 trace_ocfs2_add_clusters_in_btree(
4857 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4859 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4860 num_bits, flags, meta_ac);
4867 ocfs2_journal_dirty(handle, et->et_root_bh);
4869 clusters_to_add -= num_bits;
4870 *logical_offset += num_bits;
4872 if (clusters_to_add) {
4873 err = clusters_to_add;
4875 reason = RESTART_TRANS;
4880 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4881 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4884 ocfs2_free_clusters(handle,
4887 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4893 *reason_ret = reason;
4894 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4898 static void ocfs2_make_right_split_rec(struct super_block *sb,
4899 struct ocfs2_extent_rec *split_rec,
4901 struct ocfs2_extent_rec *rec)
4903 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4904 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4906 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4908 split_rec->e_cpos = cpu_to_le32(cpos);
4909 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4911 split_rec->e_blkno = rec->e_blkno;
4912 le64_add_cpu(&split_rec->e_blkno,
4913 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4915 split_rec->e_flags = rec->e_flags;
4918 static int ocfs2_split_and_insert(handle_t *handle,
4919 struct ocfs2_extent_tree *et,
4920 struct ocfs2_path *path,
4921 struct buffer_head **last_eb_bh,
4923 struct ocfs2_extent_rec *orig_split_rec,
4924 struct ocfs2_alloc_context *meta_ac)
4927 unsigned int insert_range, rec_range, do_leftright = 0;
4928 struct ocfs2_extent_rec tmprec;
4929 struct ocfs2_extent_list *rightmost_el;
4930 struct ocfs2_extent_rec rec;
4931 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4932 struct ocfs2_insert_type insert;
4933 struct ocfs2_extent_block *eb;
4937 * Store a copy of the record on the stack - it might move
4938 * around as the tree is manipulated below.
4940 rec = path_leaf_el(path)->l_recs[split_index];
4942 rightmost_el = et->et_root_el;
4944 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4946 BUG_ON(!(*last_eb_bh));
4947 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4948 rightmost_el = &eb->h_list;
4951 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4952 le16_to_cpu(rightmost_el->l_count)) {
4953 ret = ocfs2_grow_tree(handle, et,
4954 &depth, last_eb_bh, meta_ac);
4961 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4962 insert.ins_appending = APPEND_NONE;
4963 insert.ins_contig = CONTIG_NONE;
4964 insert.ins_tree_depth = depth;
4966 insert_range = le32_to_cpu(split_rec.e_cpos) +
4967 le16_to_cpu(split_rec.e_leaf_clusters);
4968 rec_range = le32_to_cpu(rec.e_cpos) +
4969 le16_to_cpu(rec.e_leaf_clusters);
4971 if (split_rec.e_cpos == rec.e_cpos) {
4972 insert.ins_split = SPLIT_LEFT;
4973 } else if (insert_range == rec_range) {
4974 insert.ins_split = SPLIT_RIGHT;
4977 * Left/right split. We fake this as a right split
4978 * first and then make a second pass as a left split.
4980 insert.ins_split = SPLIT_RIGHT;
4982 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4983 &tmprec, insert_range, &rec);
4987 BUG_ON(do_leftright);
4991 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4997 if (do_leftright == 1) {
4999 struct ocfs2_extent_list *el;
5002 split_rec = *orig_split_rec;
5004 ocfs2_reinit_path(path, 1);
5006 cpos = le32_to_cpu(split_rec.e_cpos);
5007 ret = ocfs2_find_path(et->et_ci, path, cpos);
5013 el = path_leaf_el(path);
5014 split_index = ocfs2_search_extent_list(el, cpos);
5015 if (split_index == -1) {
5016 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5017 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5018 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5030 static int ocfs2_replace_extent_rec(handle_t *handle,
5031 struct ocfs2_extent_tree *et,
5032 struct ocfs2_path *path,
5033 struct ocfs2_extent_list *el,
5035 struct ocfs2_extent_rec *split_rec)
5039 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5040 path_num_items(path) - 1);
5046 el->l_recs[split_index] = *split_rec;
5048 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5054 * Split part or all of the extent record at split_index in the leaf
5055 * pointed to by path. Merge with the contiguous extent record if needed.
5057 * Care is taken to handle contiguousness so as to not grow the tree.
5059 * meta_ac is not strictly necessary - we only truly need it if growth
5060 * of the tree is required. All other cases will degrade into a less
5061 * optimal tree layout.
5063 * last_eb_bh should be the rightmost leaf block for any extent
5064 * btree. Since a split may grow the tree or a merge might shrink it,
5065 * the caller cannot trust the contents of that buffer after this call.
5067 * This code is optimized for readability - several passes might be
5068 * made over certain portions of the tree. All of those blocks will
5069 * have been brought into cache (and pinned via the journal), so the
5070 * extra overhead is not expressed in terms of disk reads.
5072 int ocfs2_split_extent(handle_t *handle,
5073 struct ocfs2_extent_tree *et,
5074 struct ocfs2_path *path,
5076 struct ocfs2_extent_rec *split_rec,
5077 struct ocfs2_alloc_context *meta_ac,
5078 struct ocfs2_cached_dealloc_ctxt *dealloc)
5081 struct ocfs2_extent_list *el = path_leaf_el(path);
5082 struct buffer_head *last_eb_bh = NULL;
5083 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5084 struct ocfs2_merge_ctxt ctxt;
5086 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5087 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5088 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5094 ret = ocfs2_figure_merge_contig_type(et, path, el,
5104 * The core merge / split code wants to know how much room is
5105 * left in this allocation tree, so we pass the
5106 * rightmost extent list.
5108 if (path->p_tree_depth) {
5109 struct ocfs2_extent_block *eb;
5111 ret = ocfs2_read_extent_block(et->et_ci,
5112 ocfs2_et_get_last_eb_blk(et),
5119 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5122 if (rec->e_cpos == split_rec->e_cpos &&
5123 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5124 ctxt.c_split_covers_rec = 1;
5126 ctxt.c_split_covers_rec = 0;
5128 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5130 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5131 ctxt.c_has_empty_extent,
5132 ctxt.c_split_covers_rec);
5134 if (ctxt.c_contig_type == CONTIG_NONE) {
5135 if (ctxt.c_split_covers_rec)
5136 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5137 split_index, split_rec);
5139 ret = ocfs2_split_and_insert(handle, et, path,
5140 &last_eb_bh, split_index,
5141 split_rec, meta_ac);
5145 ret = ocfs2_try_to_merge_extent(handle, et, path,
5146 split_index, split_rec,
5158 * Change the flags of the already-existing extent at cpos for len clusters.
5160 * new_flags: the flags we want to set.
5161 * clear_flags: the flags we want to clear.
5162 * phys: the new physical offset we want this new extent starts from.
5164 * If the existing extent is larger than the request, initiate a
5165 * split. An attempt will be made at merging with adjacent extents.
5167 * The caller is responsible for passing down meta_ac if we'll need it.
5169 int ocfs2_change_extent_flag(handle_t *handle,
5170 struct ocfs2_extent_tree *et,
5171 u32 cpos, u32 len, u32 phys,
5172 struct ocfs2_alloc_context *meta_ac,
5173 struct ocfs2_cached_dealloc_ctxt *dealloc,
5174 int new_flags, int clear_flags)
5177 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5178 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5179 struct ocfs2_extent_rec split_rec;
5180 struct ocfs2_path *left_path = NULL;
5181 struct ocfs2_extent_list *el;
5182 struct ocfs2_extent_rec *rec;
5184 left_path = ocfs2_new_path_from_et(et);
5191 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5196 el = path_leaf_el(left_path);
5198 index = ocfs2_search_extent_list(el, cpos);
5201 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5202 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5209 rec = &el->l_recs[index];
5210 if (new_flags && (rec->e_flags & new_flags)) {
5211 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5212 "extent that already had them\n",
5213 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5218 if (clear_flags && !(rec->e_flags & clear_flags)) {
5219 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5220 "extent that didn't have them\n",
5221 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5226 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5227 split_rec.e_cpos = cpu_to_le32(cpos);
5228 split_rec.e_leaf_clusters = cpu_to_le16(len);
5229 split_rec.e_blkno = cpu_to_le64(start_blkno);
5230 split_rec.e_flags = rec->e_flags;
5232 split_rec.e_flags |= new_flags;
5234 split_rec.e_flags &= ~clear_flags;
5236 ret = ocfs2_split_extent(handle, et, left_path,
5237 index, &split_rec, meta_ac,
5243 ocfs2_free_path(left_path);
5249 * Mark the already-existing extent at cpos as written for len clusters.
5250 * This removes the unwritten extent flag.
5252 * If the existing extent is larger than the request, initiate a
5253 * split. An attempt will be made at merging with adjacent extents.
5255 * The caller is responsible for passing down meta_ac if we'll need it.
5257 int ocfs2_mark_extent_written(struct inode *inode,
5258 struct ocfs2_extent_tree *et,
5259 handle_t *handle, u32 cpos, u32 len, u32 phys,
5260 struct ocfs2_alloc_context *meta_ac,
5261 struct ocfs2_cached_dealloc_ctxt *dealloc)
5265 trace_ocfs2_mark_extent_written(
5266 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5269 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5270 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents that are being written to, but the feature bit is not set in the super block\n",
5271 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5277 * XXX: This should be fixed up so that we just re-insert the
5278 * next extent records.
5280 ocfs2_et_extent_map_truncate(et, 0);
5282 ret = ocfs2_change_extent_flag(handle, et, cpos,
5283 len, phys, meta_ac, dealloc,
5284 0, OCFS2_EXT_UNWRITTEN);
5292 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5293 struct ocfs2_path *path,
5294 int index, u32 new_range,
5295 struct ocfs2_alloc_context *meta_ac)
5297 int ret, depth, credits;
5298 struct buffer_head *last_eb_bh = NULL;
5299 struct ocfs2_extent_block *eb;
5300 struct ocfs2_extent_list *rightmost_el, *el;
5301 struct ocfs2_extent_rec split_rec;
5302 struct ocfs2_extent_rec *rec;
5303 struct ocfs2_insert_type insert;
5306 * Setup the record to split before we grow the tree.
5308 el = path_leaf_el(path);
5309 rec = &el->l_recs[index];
5310 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5311 &split_rec, new_range, rec);
5313 depth = path->p_tree_depth;
5315 ret = ocfs2_read_extent_block(et->et_ci,
5316 ocfs2_et_get_last_eb_blk(et),
5323 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5324 rightmost_el = &eb->h_list;
5326 rightmost_el = path_leaf_el(path);
5328 credits = path->p_tree_depth +
5329 ocfs2_extend_meta_needed(et->et_root_el);
5330 ret = ocfs2_extend_trans(handle, credits);
5336 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5337 le16_to_cpu(rightmost_el->l_count)) {
5338 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5346 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5347 insert.ins_appending = APPEND_NONE;
5348 insert.ins_contig = CONTIG_NONE;
5349 insert.ins_split = SPLIT_RIGHT;
5350 insert.ins_tree_depth = depth;
5352 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5361 static int ocfs2_truncate_rec(handle_t *handle,
5362 struct ocfs2_extent_tree *et,
5363 struct ocfs2_path *path, int index,
5364 struct ocfs2_cached_dealloc_ctxt *dealloc,
5368 u32 left_cpos, rec_range, trunc_range;
5369 int is_rightmost_tree_rec = 0;
5370 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5371 struct ocfs2_path *left_path = NULL;
5372 struct ocfs2_extent_list *el = path_leaf_el(path);
5373 struct ocfs2_extent_rec *rec;
5374 struct ocfs2_extent_block *eb;
5376 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5377 /* extend credit for ocfs2_remove_rightmost_path */
5378 ret = ocfs2_extend_rotate_transaction(handle, 0,
5379 handle->h_buffer_credits,
5386 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5395 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5396 path->p_tree_depth) {
5398 * Check whether this is the rightmost tree record. If
5399 * we remove all of this record or part of its right
5400 * edge then an update of the record lengths above it
5403 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5404 if (eb->h_next_leaf_blk == 0)
5405 is_rightmost_tree_rec = 1;
5408 rec = &el->l_recs[index];
5409 if (index == 0 && path->p_tree_depth &&
5410 le32_to_cpu(rec->e_cpos) == cpos) {
5412 * Changing the leftmost offset (via partial or whole
5413 * record truncate) of an interior (or rightmost) path
5414 * means we have to update the subtree that is formed
5415 * by this leaf and the one to it's left.
5417 * There are two cases we can skip:
5418 * 1) Path is the leftmost one in our btree.
5419 * 2) The leaf is rightmost and will be empty after
5420 * we remove the extent record - the rotate code
5421 * knows how to update the newly formed edge.
5424 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5430 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5431 left_path = ocfs2_new_path_from_path(path);
5438 ret = ocfs2_find_path(et->et_ci, left_path,
5447 ret = ocfs2_extend_rotate_transaction(handle, 0,
5448 handle->h_buffer_credits,
5455 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5461 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5467 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5468 trunc_range = cpos + len;
5470 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5473 memset(rec, 0, sizeof(*rec));
5474 ocfs2_cleanup_merge(el, index);
5476 next_free = le16_to_cpu(el->l_next_free_rec);
5477 if (is_rightmost_tree_rec && next_free > 1) {
5479 * We skip the edge update if this path will
5480 * be deleted by the rotate code.
5482 rec = &el->l_recs[next_free - 1];
5483 ocfs2_adjust_rightmost_records(handle, et, path,
5486 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5487 /* Remove leftmost portion of the record. */
5488 le32_add_cpu(&rec->e_cpos, len);
5489 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5490 le16_add_cpu(&rec->e_leaf_clusters, -len);
5491 } else if (rec_range == trunc_range) {
5492 /* Remove rightmost portion of the record */
5493 le16_add_cpu(&rec->e_leaf_clusters, -len);
5494 if (is_rightmost_tree_rec)
5495 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5497 /* Caller should have trapped this. */
5498 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5500 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5501 le32_to_cpu(rec->e_cpos),
5502 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5509 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5510 ocfs2_complete_edge_insert(handle, left_path, path,
5514 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5516 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5521 ocfs2_free_path(left_path);
5525 int ocfs2_remove_extent(handle_t *handle,
5526 struct ocfs2_extent_tree *et,
5528 struct ocfs2_alloc_context *meta_ac,
5529 struct ocfs2_cached_dealloc_ctxt *dealloc)
5532 u32 rec_range, trunc_range;
5533 struct ocfs2_extent_rec *rec;
5534 struct ocfs2_extent_list *el;
5535 struct ocfs2_path *path = NULL;
5538 * XXX: Why are we truncating to 0 instead of wherever this
5541 ocfs2_et_extent_map_truncate(et, 0);
5543 path = ocfs2_new_path_from_et(et);
5550 ret = ocfs2_find_path(et->et_ci, path, cpos);
5556 el = path_leaf_el(path);
5557 index = ocfs2_search_extent_list(el, cpos);
5559 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5560 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5561 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5568 * We have 3 cases of extent removal:
5569 * 1) Range covers the entire extent rec
5570 * 2) Range begins or ends on one edge of the extent rec
5571 * 3) Range is in the middle of the extent rec (no shared edges)
5573 * For case 1 we remove the extent rec and left rotate to
5576 * For case 2 we just shrink the existing extent rec, with a
5577 * tree update if the shrinking edge is also the edge of an
5580 * For case 3 we do a right split to turn the extent rec into
5581 * something case 2 can handle.
5583 rec = &el->l_recs[index];
5584 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5585 trunc_range = cpos + len;
5587 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5589 trace_ocfs2_remove_extent(
5590 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5591 cpos, len, index, le32_to_cpu(rec->e_cpos),
5592 ocfs2_rec_clusters(el, rec));
5594 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5595 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5602 ret = ocfs2_split_tree(handle, et, path, index,
5603 trunc_range, meta_ac);
5610 * The split could have manipulated the tree enough to
5611 * move the record location, so we have to look for it again.
5613 ocfs2_reinit_path(path, 1);
5615 ret = ocfs2_find_path(et->et_ci, path, cpos);
5621 el = path_leaf_el(path);
5622 index = ocfs2_search_extent_list(el, cpos);
5624 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5625 "Owner %llu: split at cpos %u lost record\n",
5626 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5633 * Double check our values here. If anything is fishy,
5634 * it's easier to catch it at the top level.
5636 rec = &el->l_recs[index];
5637 rec_range = le32_to_cpu(rec->e_cpos) +
5638 ocfs2_rec_clusters(el, rec);
5639 if (rec_range != trunc_range) {
5640 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5641 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5642 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5643 cpos, len, le32_to_cpu(rec->e_cpos),
5644 ocfs2_rec_clusters(el, rec));
5649 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5656 ocfs2_free_path(path);
5661 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5662 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5663 * number to reserve some extra blocks, and it only handles meta
5666 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5667 * and punching holes.
5669 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5670 struct ocfs2_extent_tree *et,
5671 u32 extents_to_split,
5672 struct ocfs2_alloc_context **ac,
5675 int ret = 0, num_free_extents;
5676 unsigned int max_recs_needed = 2 * extents_to_split;
5677 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5681 num_free_extents = ocfs2_num_free_extents(et);
5682 if (num_free_extents < 0) {
5683 ret = num_free_extents;
5688 if (!num_free_extents ||
5689 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5690 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5693 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5703 ocfs2_free_alloc_context(*ac);
5711 int ocfs2_remove_btree_range(struct inode *inode,
5712 struct ocfs2_extent_tree *et,
5713 u32 cpos, u32 phys_cpos, u32 len, int flags,
5714 struct ocfs2_cached_dealloc_ctxt *dealloc,
5715 u64 refcount_loc, bool refcount_tree_locked)
5717 int ret, credits = 0, extra_blocks = 0;
5718 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5719 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5720 struct inode *tl_inode = osb->osb_tl_inode;
5722 struct ocfs2_alloc_context *meta_ac = NULL;
5723 struct ocfs2_refcount_tree *ref_tree = NULL;
5725 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5726 BUG_ON(!ocfs2_is_refcount_inode(inode));
5728 if (!refcount_tree_locked) {
5729 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5737 ret = ocfs2_prepare_refcount_change_for_del(inode,
5749 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5756 inode_lock(tl_inode);
5758 if (ocfs2_truncate_log_needs_flush(osb)) {
5759 ret = __ocfs2_flush_truncate_log(osb);
5766 handle = ocfs2_start_trans(osb,
5767 ocfs2_remove_extent_credits(osb->sb) + credits);
5768 if (IS_ERR(handle)) {
5769 ret = PTR_ERR(handle);
5774 ret = ocfs2_et_root_journal_access(handle, et,
5775 OCFS2_JOURNAL_ACCESS_WRITE);
5781 dquot_free_space_nodirty(inode,
5782 ocfs2_clusters_to_bytes(inode->i_sb, len));
5784 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5790 ocfs2_et_update_clusters(et, -len);
5791 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5793 ocfs2_journal_dirty(handle, et->et_root_bh);
5796 if (flags & OCFS2_EXT_REFCOUNTED)
5797 ret = ocfs2_decrease_refcount(inode, handle,
5798 ocfs2_blocks_to_clusters(osb->sb,
5803 ret = ocfs2_truncate_log_append(osb, handle,
5811 ocfs2_commit_trans(osb, handle);
5813 inode_unlock(tl_inode);
5816 ocfs2_free_alloc_context(meta_ac);
5819 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5824 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5826 struct buffer_head *tl_bh = osb->osb_tl_bh;
5827 struct ocfs2_dinode *di;
5828 struct ocfs2_truncate_log *tl;
5830 di = (struct ocfs2_dinode *) tl_bh->b_data;
5831 tl = &di->id2.i_dealloc;
5833 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5834 "slot %d, invalid truncate log parameters: used = "
5835 "%u, count = %u\n", osb->slot_num,
5836 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5837 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5840 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5841 unsigned int new_start)
5843 unsigned int tail_index;
5844 unsigned int current_tail;
5846 /* No records, nothing to coalesce */
5847 if (!le16_to_cpu(tl->tl_used))
5850 tail_index = le16_to_cpu(tl->tl_used) - 1;
5851 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5852 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5854 return current_tail == new_start;
5857 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5860 unsigned int num_clusters)
5863 unsigned int start_cluster, tl_count;
5864 struct inode *tl_inode = osb->osb_tl_inode;
5865 struct buffer_head *tl_bh = osb->osb_tl_bh;
5866 struct ocfs2_dinode *di;
5867 struct ocfs2_truncate_log *tl;
5869 BUG_ON(inode_trylock(tl_inode));
5871 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5873 di = (struct ocfs2_dinode *) tl_bh->b_data;
5875 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5876 * by the underlying call to ocfs2_read_inode_block(), so any
5877 * corruption is a code bug */
5878 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5880 tl = &di->id2.i_dealloc;
5881 tl_count = le16_to_cpu(tl->tl_count);
5882 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5884 "Truncate record count on #%llu invalid "
5885 "wanted %u, actual %u\n",
5886 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5887 ocfs2_truncate_recs_per_inode(osb->sb),
5888 le16_to_cpu(tl->tl_count));
5890 /* Caller should have known to flush before calling us. */
5891 index = le16_to_cpu(tl->tl_used);
5892 if (index >= tl_count) {
5898 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5899 OCFS2_JOURNAL_ACCESS_WRITE);
5905 trace_ocfs2_truncate_log_append(
5906 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5907 start_cluster, num_clusters);
5908 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5910 * Move index back to the record we are coalescing with.
5911 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5915 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5916 trace_ocfs2_truncate_log_append(
5917 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5918 index, le32_to_cpu(tl->tl_recs[index].t_start),
5921 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5922 tl->tl_used = cpu_to_le16(index + 1);
5924 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5926 ocfs2_journal_dirty(handle, tl_bh);
5928 osb->truncated_clusters += num_clusters;
5933 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5934 struct inode *data_alloc_inode,
5935 struct buffer_head *data_alloc_bh)
5939 unsigned int num_clusters;
5941 struct ocfs2_truncate_rec rec;
5942 struct ocfs2_dinode *di;
5943 struct ocfs2_truncate_log *tl;
5944 struct inode *tl_inode = osb->osb_tl_inode;
5945 struct buffer_head *tl_bh = osb->osb_tl_bh;
5948 di = (struct ocfs2_dinode *) tl_bh->b_data;
5949 tl = &di->id2.i_dealloc;
5950 i = le16_to_cpu(tl->tl_used) - 1;
5952 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5953 if (IS_ERR(handle)) {
5954 status = PTR_ERR(handle);
5959 /* Caller has given us at least enough credits to
5960 * update the truncate log dinode */
5961 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5962 OCFS2_JOURNAL_ACCESS_WRITE);
5968 tl->tl_used = cpu_to_le16(i);
5970 ocfs2_journal_dirty(handle, tl_bh);
5972 rec = tl->tl_recs[i];
5973 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5974 le32_to_cpu(rec.t_start));
5975 num_clusters = le32_to_cpu(rec.t_clusters);
5977 /* if start_blk is not set, we ignore the record as
5980 trace_ocfs2_replay_truncate_records(
5981 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5982 i, le32_to_cpu(rec.t_start), num_clusters);
5984 status = ocfs2_free_clusters(handle, data_alloc_inode,
5985 data_alloc_bh, start_blk,
5993 ocfs2_commit_trans(osb, handle);
5997 osb->truncated_clusters = 0;
6003 /* Expects you to already be holding tl_inode->i_mutex */
6004 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6007 unsigned int num_to_flush;
6008 struct inode *tl_inode = osb->osb_tl_inode;
6009 struct inode *data_alloc_inode = NULL;
6010 struct buffer_head *tl_bh = osb->osb_tl_bh;
6011 struct buffer_head *data_alloc_bh = NULL;
6012 struct ocfs2_dinode *di;
6013 struct ocfs2_truncate_log *tl;
6015 BUG_ON(inode_trylock(tl_inode));
6017 di = (struct ocfs2_dinode *) tl_bh->b_data;
6019 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
6020 * by the underlying call to ocfs2_read_inode_block(), so any
6021 * corruption is a code bug */
6022 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6024 tl = &di->id2.i_dealloc;
6025 num_to_flush = le16_to_cpu(tl->tl_used);
6026 trace_ocfs2_flush_truncate_log(
6027 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6029 if (!num_to_flush) {
6034 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6035 GLOBAL_BITMAP_SYSTEM_INODE,
6036 OCFS2_INVALID_SLOT);
6037 if (!data_alloc_inode) {
6039 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6043 inode_lock(data_alloc_inode);
6045 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6051 status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6056 brelse(data_alloc_bh);
6057 ocfs2_inode_unlock(data_alloc_inode, 1);
6060 inode_unlock(data_alloc_inode);
6061 iput(data_alloc_inode);
6067 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6070 struct inode *tl_inode = osb->osb_tl_inode;
6072 inode_lock(tl_inode);
6073 status = __ocfs2_flush_truncate_log(osb);
6074 inode_unlock(tl_inode);
6079 static void ocfs2_truncate_log_worker(struct work_struct *work)
6082 struct ocfs2_super *osb =
6083 container_of(work, struct ocfs2_super,
6084 osb_truncate_log_wq.work);
6086 status = ocfs2_flush_truncate_log(osb);
6090 ocfs2_init_steal_slots(osb);
6093 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6094 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6097 if (osb->osb_tl_inode &&
6098 atomic_read(&osb->osb_tl_disable) == 0) {
6099 /* We want to push off log flushes while truncates are
6102 cancel_delayed_work(&osb->osb_truncate_log_wq);
6104 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6105 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6110 * Try to flush truncate logs if we can free enough clusters from it.
6111 * As for return value, "< 0" means error, "0" no space and "1" means
6112 * we have freed enough spaces and let the caller try to allocate again.
6114 int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6115 unsigned int needed)
6119 unsigned int truncated_clusters;
6121 inode_lock(osb->osb_tl_inode);
6122 truncated_clusters = osb->truncated_clusters;
6123 inode_unlock(osb->osb_tl_inode);
6126 * Check whether we can succeed in allocating if we free
6129 if (truncated_clusters < needed)
6132 ret = ocfs2_flush_truncate_log(osb);
6138 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6139 jbd2_log_wait_commit(osb->journal->j_journal, target);
6146 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6148 struct inode **tl_inode,
6149 struct buffer_head **tl_bh)
6152 struct inode *inode = NULL;
6153 struct buffer_head *bh = NULL;
6155 inode = ocfs2_get_system_file_inode(osb,
6156 TRUNCATE_LOG_SYSTEM_INODE,
6160 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6164 status = ocfs2_read_inode_block(inode, &bh);
6177 /* called during the 1st stage of node recovery. we stamp a clean
6178 * truncate log and pass back a copy for processing later. if the
6179 * truncate log does not require processing, a *tl_copy is set to
6181 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6183 struct ocfs2_dinode **tl_copy)
6186 struct inode *tl_inode = NULL;
6187 struct buffer_head *tl_bh = NULL;
6188 struct ocfs2_dinode *di;
6189 struct ocfs2_truncate_log *tl;
6193 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6195 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6201 di = (struct ocfs2_dinode *) tl_bh->b_data;
6203 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6204 * validated by the underlying call to ocfs2_read_inode_block(),
6205 * so any corruption is a code bug */
6206 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6208 tl = &di->id2.i_dealloc;
6209 if (le16_to_cpu(tl->tl_used)) {
6210 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6212 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6219 /* Assuming the write-out below goes well, this copy
6220 * will be passed back to recovery for processing. */
6221 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6223 /* All we need to do to clear the truncate log is set
6227 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6228 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6248 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6249 struct ocfs2_dinode *tl_copy)
6253 unsigned int clusters, num_recs, start_cluster;
6256 struct inode *tl_inode = osb->osb_tl_inode;
6257 struct ocfs2_truncate_log *tl;
6259 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6260 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6264 tl = &tl_copy->id2.i_dealloc;
6265 num_recs = le16_to_cpu(tl->tl_used);
6266 trace_ocfs2_complete_truncate_log_recovery(
6267 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6270 inode_lock(tl_inode);
6271 for(i = 0; i < num_recs; i++) {
6272 if (ocfs2_truncate_log_needs_flush(osb)) {
6273 status = __ocfs2_flush_truncate_log(osb);
6280 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6281 if (IS_ERR(handle)) {
6282 status = PTR_ERR(handle);
6287 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6288 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6289 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6291 status = ocfs2_truncate_log_append(osb, handle,
6292 start_blk, clusters);
6293 ocfs2_commit_trans(osb, handle);
6301 inode_unlock(tl_inode);
6306 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6309 struct inode *tl_inode = osb->osb_tl_inode;
6311 atomic_set(&osb->osb_tl_disable, 1);
6314 cancel_delayed_work(&osb->osb_truncate_log_wq);
6315 flush_workqueue(osb->ocfs2_wq);
6317 status = ocfs2_flush_truncate_log(osb);
6321 brelse(osb->osb_tl_bh);
6322 iput(osb->osb_tl_inode);
6326 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6329 struct inode *tl_inode = NULL;
6330 struct buffer_head *tl_bh = NULL;
6332 status = ocfs2_get_truncate_log_info(osb,
6339 /* ocfs2_truncate_log_shutdown keys on the existence of
6340 * osb->osb_tl_inode so we don't set any of the osb variables
6341 * until we're sure all is well. */
6342 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6343 ocfs2_truncate_log_worker);
6344 atomic_set(&osb->osb_tl_disable, 0);
6345 osb->osb_tl_bh = tl_bh;
6346 osb->osb_tl_inode = tl_inode;
6352 * Delayed de-allocation of suballocator blocks.
6354 * Some sets of block de-allocations might involve multiple suballocator inodes.
6356 * The locking for this can get extremely complicated, especially when
6357 * the suballocator inodes to delete from aren't known until deep
6358 * within an unrelated codepath.
6360 * ocfs2_extent_block structures are a good example of this - an inode
6361 * btree could have been grown by any number of nodes each allocating
6362 * out of their own suballoc inode.
6364 * These structures allow the delay of block de-allocation until a
6365 * later time, when locking of multiple cluster inodes won't cause
6370 * Describe a single bit freed from a suballocator. For the block
6371 * suballocators, it represents one block. For the global cluster
6372 * allocator, it represents some clusters and free_bit indicates
6375 struct ocfs2_cached_block_free {
6376 struct ocfs2_cached_block_free *free_next;
6379 unsigned int free_bit;
6382 struct ocfs2_per_slot_free_list {
6383 struct ocfs2_per_slot_free_list *f_next_suballocator;
6386 struct ocfs2_cached_block_free *f_first;
6389 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6392 struct ocfs2_cached_block_free *head)
6397 struct inode *inode;
6398 struct buffer_head *di_bh = NULL;
6399 struct ocfs2_cached_block_free *tmp;
6401 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6410 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6418 bg_blkno = head->free_bg;
6420 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6422 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6423 if (IS_ERR(handle)) {
6424 ret = PTR_ERR(handle);
6429 trace_ocfs2_free_cached_blocks(
6430 (unsigned long long)head->free_blk, head->free_bit);
6432 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6433 head->free_bit, bg_blkno, 1);
6437 ocfs2_commit_trans(osb, handle);
6440 head = head->free_next;
6445 ocfs2_inode_unlock(inode, 1);
6448 inode_unlock(inode);
6452 /* Premature exit may have left some dangling items. */
6454 head = head->free_next;
6461 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6462 u64 blkno, unsigned int bit)
6465 struct ocfs2_cached_block_free *item;
6467 item = kzalloc(sizeof(*item), GFP_NOFS);
6474 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6476 item->free_blk = blkno;
6477 item->free_bit = bit;
6478 item->free_next = ctxt->c_global_allocator;
6480 ctxt->c_global_allocator = item;
6484 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6485 struct ocfs2_cached_block_free *head)
6487 struct ocfs2_cached_block_free *tmp;
6488 struct inode *tl_inode = osb->osb_tl_inode;
6492 inode_lock(tl_inode);
6495 if (ocfs2_truncate_log_needs_flush(osb)) {
6496 ret = __ocfs2_flush_truncate_log(osb);
6503 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6504 if (IS_ERR(handle)) {
6505 ret = PTR_ERR(handle);
6510 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6513 ocfs2_commit_trans(osb, handle);
6515 head = head->free_next;
6524 inode_unlock(tl_inode);
6527 /* Premature exit may have left some dangling items. */
6529 head = head->free_next;
6536 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6537 struct ocfs2_cached_dealloc_ctxt *ctxt)
6540 struct ocfs2_per_slot_free_list *fl;
6545 while (ctxt->c_first_suballocator) {
6546 fl = ctxt->c_first_suballocator;
6549 trace_ocfs2_run_deallocs(fl->f_inode_type,
6551 ret2 = ocfs2_free_cached_blocks(osb,
6561 ctxt->c_first_suballocator = fl->f_next_suballocator;
6565 if (ctxt->c_global_allocator) {
6566 ret2 = ocfs2_free_cached_clusters(osb,
6567 ctxt->c_global_allocator);
6573 ctxt->c_global_allocator = NULL;
6579 static struct ocfs2_per_slot_free_list *
6580 ocfs2_find_per_slot_free_list(int type,
6582 struct ocfs2_cached_dealloc_ctxt *ctxt)
6584 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6587 if (fl->f_inode_type == type && fl->f_slot == slot)
6590 fl = fl->f_next_suballocator;
6593 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6595 fl->f_inode_type = type;
6598 fl->f_next_suballocator = ctxt->c_first_suballocator;
6600 ctxt->c_first_suballocator = fl;
6605 static struct ocfs2_per_slot_free_list *
6606 ocfs2_find_preferred_free_list(int type,
6609 struct ocfs2_cached_dealloc_ctxt *ctxt)
6611 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6614 if (fl->f_inode_type == type && fl->f_slot == preferred_slot) {
6615 *real_slot = fl->f_slot;
6619 fl = fl->f_next_suballocator;
6622 /* If we can't find any free list matching preferred slot, just use
6625 fl = ctxt->c_first_suballocator;
6626 *real_slot = fl->f_slot;
6631 /* Return Value 1 indicates empty */
6632 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et)
6634 struct ocfs2_per_slot_free_list *fl = NULL;
6636 if (!et->et_dealloc)
6639 fl = et->et_dealloc->c_first_suballocator;
6649 /* If extent was deleted from tree due to extent rotation and merging, and
6650 * no metadata is reserved ahead of time. Try to reuse some extents
6651 * just deleted. This is only used to reuse extent blocks.
6652 * It is supposed to find enough extent blocks in dealloc if our estimation
6653 * on metadata is accurate.
6655 static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
6656 struct ocfs2_extent_tree *et,
6657 struct buffer_head **new_eb_bh,
6658 int blk_wanted, int *blk_given)
6660 int i, status = 0, real_slot;
6661 struct ocfs2_cached_dealloc_ctxt *dealloc;
6662 struct ocfs2_per_slot_free_list *fl;
6663 struct ocfs2_cached_block_free *bf;
6664 struct ocfs2_extent_block *eb;
6665 struct ocfs2_super *osb =
6666 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
6670 /* If extent tree doesn't have a dealloc, this is not faulty. Just
6671 * tell upper caller dealloc can't provide any block and it should
6672 * ask for alloc to claim more space.
6674 dealloc = et->et_dealloc;
6678 for (i = 0; i < blk_wanted; i++) {
6679 /* Prefer to use local slot */
6680 fl = ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE,
6681 osb->slot_num, &real_slot,
6683 /* If no more block can be reused, we should claim more
6684 * from alloc. Just return here normally.
6692 fl->f_first = bf->free_next;
6694 new_eb_bh[i] = sb_getblk(osb->sb, bf->free_blk);
6695 if (new_eb_bh[i] == NULL) {
6701 mlog(0, "Reusing block(%llu) from "
6702 "dealloc(local slot:%d, real slot:%d)\n",
6703 bf->free_blk, osb->slot_num, real_slot);
6705 ocfs2_set_new_buffer_uptodate(et->et_ci, new_eb_bh[i]);
6707 status = ocfs2_journal_access_eb(handle, et->et_ci,
6709 OCFS2_JOURNAL_ACCESS_CREATE);
6715 memset(new_eb_bh[i]->b_data, 0, osb->sb->s_blocksize);
6716 eb = (struct ocfs2_extent_block *) new_eb_bh[i]->b_data;
6718 /* We can't guarantee that buffer head is still cached, so
6719 * polutlate the extent block again.
6721 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
6722 eb->h_blkno = cpu_to_le64(bf->free_blk);
6723 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
6724 eb->h_suballoc_slot = cpu_to_le16(real_slot);
6725 eb->h_suballoc_loc = cpu_to_le64(bf->free_bg);
6726 eb->h_suballoc_bit = cpu_to_le16(bf->free_bit);
6727 eb->h_list.l_count =
6728 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
6730 /* We'll also be dirtied by the caller, so
6731 * this isn't absolutely necessary.
6733 ocfs2_journal_dirty(handle, new_eb_bh[i]);
6736 dealloc->c_first_suballocator = fl->f_next_suballocator;
6745 if (unlikely(status < 0)) {
6746 for (i = 0; i < blk_wanted; i++)
6747 brelse(new_eb_bh[i]);
6753 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6754 int type, int slot, u64 suballoc,
6755 u64 blkno, unsigned int bit)
6758 struct ocfs2_per_slot_free_list *fl;
6759 struct ocfs2_cached_block_free *item;
6761 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6768 item = kzalloc(sizeof(*item), GFP_NOFS);
6775 trace_ocfs2_cache_block_dealloc(type, slot,
6776 (unsigned long long)suballoc,
6777 (unsigned long long)blkno, bit);
6779 item->free_bg = suballoc;
6780 item->free_blk = blkno;
6781 item->free_bit = bit;
6782 item->free_next = fl->f_first;
6791 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6792 struct ocfs2_extent_block *eb)
6794 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6795 le16_to_cpu(eb->h_suballoc_slot),
6796 le64_to_cpu(eb->h_suballoc_loc),
6797 le64_to_cpu(eb->h_blkno),
6798 le16_to_cpu(eb->h_suballoc_bit));
6801 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6803 set_buffer_uptodate(bh);
6804 mark_buffer_dirty(bh);
6808 void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6809 unsigned int from, unsigned int to,
6810 struct page *page, int zero, u64 *phys)
6812 int ret, partial = 0;
6814 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6819 zero_user_segment(page, from, to);
6822 * Need to set the buffers we zero'd into uptodate
6823 * here if they aren't - ocfs2_map_page_blocks()
6824 * might've skipped some
6826 ret = walk_page_buffers(handle, page_buffers(page),
6831 else if (ocfs2_should_order_data(inode)) {
6832 ret = ocfs2_jbd2_file_inode(handle, inode);
6838 SetPageUptodate(page);
6840 flush_dcache_page(page);
6843 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6844 loff_t end, struct page **pages,
6845 int numpages, u64 phys, handle_t *handle)
6849 unsigned int from, to = PAGE_SIZE;
6850 struct super_block *sb = inode->i_sb;
6852 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6858 for(i = 0; i < numpages; i++) {
6861 from = start & (PAGE_SIZE - 1);
6862 if ((end >> PAGE_SHIFT) == page->index)
6863 to = end & (PAGE_SIZE - 1);
6865 BUG_ON(from > PAGE_SIZE);
6866 BUG_ON(to > PAGE_SIZE);
6868 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6871 start = (page->index + 1) << PAGE_SHIFT;
6875 ocfs2_unlock_and_free_pages(pages, numpages);
6878 int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6879 struct page **pages, int *num)
6881 int numpages, ret = 0;
6882 struct address_space *mapping = inode->i_mapping;
6883 unsigned long index;
6884 loff_t last_page_bytes;
6886 BUG_ON(start > end);
6889 last_page_bytes = PAGE_ALIGN(end);
6890 index = start >> PAGE_SHIFT;
6892 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6893 if (!pages[numpages]) {
6901 } while (index < (last_page_bytes >> PAGE_SHIFT));
6906 ocfs2_unlock_and_free_pages(pages, numpages);
6915 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6916 struct page **pages, int *num)
6918 struct super_block *sb = inode->i_sb;
6920 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6921 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6923 return ocfs2_grab_pages(inode, start, end, pages, num);
6927 * Zero the area past i_size but still within an allocated
6928 * cluster. This avoids exposing nonzero data on subsequent file
6931 * We need to call this before i_size is updated on the inode because
6932 * otherwise block_write_full_page() will skip writeout of pages past
6933 * i_size. The new_i_size parameter is passed for this reason.
6935 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6936 u64 range_start, u64 range_end)
6938 int ret = 0, numpages;
6939 struct page **pages = NULL;
6941 unsigned int ext_flags;
6942 struct super_block *sb = inode->i_sb;
6945 * File systems which don't support sparse files zero on every
6948 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6951 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6952 sizeof(struct page *), GFP_NOFS);
6953 if (pages == NULL) {
6959 if (range_start == range_end)
6962 ret = ocfs2_extent_map_get_blocks(inode,
6963 range_start >> sb->s_blocksize_bits,
6964 &phys, NULL, &ext_flags);
6971 * Tail is a hole, or is marked unwritten. In either case, we
6972 * can count on read and write to return/push zero's.
6974 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6977 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6984 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6985 numpages, phys, handle);
6988 * Initiate writeout of the pages we zero'd here. We don't
6989 * wait on them - the truncate_inode_pages() call later will
6992 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
7003 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7004 struct ocfs2_dinode *di)
7006 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7007 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7009 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7010 memset(&di->id2, 0, blocksize -
7011 offsetof(struct ocfs2_dinode, id2) -
7014 memset(&di->id2, 0, blocksize -
7015 offsetof(struct ocfs2_dinode, id2));
7018 void ocfs2_dinode_new_extent_list(struct inode *inode,
7019 struct ocfs2_dinode *di)
7021 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7022 di->id2.i_list.l_tree_depth = 0;
7023 di->id2.i_list.l_next_free_rec = 0;
7024 di->id2.i_list.l_count = cpu_to_le16(
7025 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7028 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7030 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7031 struct ocfs2_inline_data *idata = &di->id2.i_data;
7033 spin_lock(&oi->ip_lock);
7034 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7035 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7036 spin_unlock(&oi->ip_lock);
7039 * We clear the entire i_data structure here so that all
7040 * fields can be properly initialized.
7042 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7044 idata->id_count = cpu_to_le16(
7045 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7048 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7049 struct buffer_head *di_bh)
7051 int ret, has_data, num_pages = 0;
7056 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7057 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7058 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7059 struct ocfs2_alloc_context *data_ac = NULL;
7060 struct page *page = NULL;
7061 struct ocfs2_extent_tree et;
7064 has_data = i_size_read(inode) ? 1 : 0;
7067 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7074 handle = ocfs2_start_trans(osb,
7075 ocfs2_inline_to_extents_credits(osb->sb));
7076 if (IS_ERR(handle)) {
7077 ret = PTR_ERR(handle);
7082 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7083 OCFS2_JOURNAL_ACCESS_WRITE);
7090 unsigned int page_end = min_t(unsigned, PAGE_SIZE,
7091 osb->s_clustersize);
7094 ret = dquot_alloc_space_nodirty(inode,
7095 ocfs2_clusters_to_bytes(osb->sb, 1));
7100 data_ac->ac_resv = &oi->ip_la_data_resv;
7102 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
7110 * Save two copies, one for insert, and one that can
7111 * be changed by ocfs2_map_and_dirty_page() below.
7113 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7115 ret = ocfs2_grab_eof_pages(inode, 0, page_end, &page,
7124 * This should populate the 1st page for us and mark
7127 ret = ocfs2_read_inline_data(inode, page, di_bh);
7134 ocfs2_map_and_dirty_page(inode, handle, 0, page_end, page, 0,
7138 spin_lock(&oi->ip_lock);
7139 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7140 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7141 spin_unlock(&oi->ip_lock);
7143 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7144 ocfs2_dinode_new_extent_list(inode, di);
7146 ocfs2_journal_dirty(handle, di_bh);
7150 * An error at this point should be extremely rare. If
7151 * this proves to be false, we could always re-build
7152 * the in-inode data from our pages.
7154 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7155 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7162 inode->i_blocks = ocfs2_inode_sector_count(inode);
7167 ocfs2_unlock_and_free_pages(&page, num_pages);
7170 if (ret < 0 && did_quota)
7171 dquot_free_space_nodirty(inode,
7172 ocfs2_clusters_to_bytes(osb->sb, 1));
7175 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7176 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7179 ocfs2_free_clusters(handle,
7182 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7186 ocfs2_commit_trans(osb, handle);
7190 ocfs2_free_alloc_context(data_ac);
7195 * It is expected, that by the time you call this function,
7196 * inode->i_size and fe->i_size have been adjusted.
7198 * WARNING: This will kfree the truncate context
7200 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7201 struct inode *inode,
7202 struct buffer_head *di_bh)
7204 int status = 0, i, flags = 0;
7205 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7207 struct ocfs2_extent_list *el;
7208 struct ocfs2_extent_rec *rec;
7209 struct ocfs2_path *path = NULL;
7210 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7211 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7212 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7213 struct ocfs2_extent_tree et;
7214 struct ocfs2_cached_dealloc_ctxt dealloc;
7215 struct ocfs2_refcount_tree *ref_tree = NULL;
7217 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7218 ocfs2_init_dealloc_ctxt(&dealloc);
7220 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7221 i_size_read(inode));
7223 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7224 ocfs2_journal_access_di);
7231 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7235 * Check that we still have allocation to delete.
7237 if (OCFS2_I(inode)->ip_clusters == 0) {
7243 * Truncate always works against the rightmost tree branch.
7245 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7251 trace_ocfs2_commit_truncate(
7252 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7254 OCFS2_I(inode)->ip_clusters,
7255 path->p_tree_depth);
7258 * By now, el will point to the extent list on the bottom most
7259 * portion of this tree. Only the tail record is considered in
7262 * We handle the following cases, in order:
7263 * - empty extent: delete the remaining branch
7264 * - remove the entire record
7265 * - remove a partial record
7266 * - no record needs to be removed (truncate has completed)
7268 el = path_leaf_el(path);
7269 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7270 ocfs2_error(inode->i_sb,
7271 "Inode %llu has empty extent block at %llu\n",
7272 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7273 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7278 i = le16_to_cpu(el->l_next_free_rec) - 1;
7279 rec = &el->l_recs[i];
7280 flags = rec->e_flags;
7281 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7283 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7285 * Lower levels depend on this never happening, but it's best
7286 * to check it up here before changing the tree.
7288 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7289 mlog(ML_ERROR, "Inode %lu has an empty "
7290 "extent record, depth %u\n", inode->i_ino,
7291 le16_to_cpu(root_el->l_tree_depth));
7292 status = ocfs2_remove_rightmost_empty_extent(osb,
7293 &et, path, &dealloc);
7299 ocfs2_reinit_path(path, 1);
7302 trunc_cpos = le32_to_cpu(rec->e_cpos);
7306 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7308 * Truncate entire record.
7310 trunc_cpos = le32_to_cpu(rec->e_cpos);
7311 trunc_len = ocfs2_rec_clusters(el, rec);
7312 blkno = le64_to_cpu(rec->e_blkno);
7313 } else if (range > new_highest_cpos) {
7315 * Partial truncate. it also should be
7316 * the last truncate we're doing.
7318 trunc_cpos = new_highest_cpos;
7319 trunc_len = range - new_highest_cpos;
7320 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7321 blkno = le64_to_cpu(rec->e_blkno) +
7322 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7325 * Truncate completed, leave happily.
7331 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7333 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7334 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7342 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7343 phys_cpos, trunc_len, flags, &dealloc,
7344 refcount_loc, true);
7350 ocfs2_reinit_path(path, 1);
7353 * The check above will catch the case where we've truncated
7354 * away all allocation.
7360 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7362 ocfs2_schedule_truncate_log_flush(osb, 1);
7364 ocfs2_run_deallocs(osb, &dealloc);
7366 ocfs2_free_path(path);
7372 * 'start' is inclusive, 'end' is not.
7374 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7375 unsigned int start, unsigned int end, int trunc)
7378 unsigned int numbytes;
7380 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7381 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7382 struct ocfs2_inline_data *idata = &di->id2.i_data;
7384 /* No need to punch hole beyond i_size. */
7385 if (start >= i_size_read(inode))
7388 if (end > i_size_read(inode))
7389 end = i_size_read(inode);
7391 BUG_ON(start > end);
7393 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7394 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7395 !ocfs2_supports_inline_data(osb)) {
7396 ocfs2_error(inode->i_sb,
7397 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7398 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7399 le16_to_cpu(di->i_dyn_features),
7400 OCFS2_I(inode)->ip_dyn_features,
7401 osb->s_feature_incompat);
7406 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7407 if (IS_ERR(handle)) {
7408 ret = PTR_ERR(handle);
7413 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7414 OCFS2_JOURNAL_ACCESS_WRITE);
7420 numbytes = end - start;
7421 memset(idata->id_data + start, 0, numbytes);
7424 * No need to worry about the data page here - it's been
7425 * truncated already and inline data doesn't need it for
7426 * pushing zero's to disk, so we'll let readpage pick it up
7430 i_size_write(inode, start);
7431 di->i_size = cpu_to_le64(start);
7434 inode->i_blocks = ocfs2_inode_sector_count(inode);
7435 inode->i_ctime = inode->i_mtime = current_time(inode);
7437 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7438 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7440 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7441 ocfs2_journal_dirty(handle, di_bh);
7444 ocfs2_commit_trans(osb, handle);
7450 static int ocfs2_trim_extent(struct super_block *sb,
7451 struct ocfs2_group_desc *gd,
7452 u64 group, u32 start, u32 count)
7454 u64 discard, bcount;
7455 struct ocfs2_super *osb = OCFS2_SB(sb);
7457 bcount = ocfs2_clusters_to_blocks(sb, count);
7458 discard = ocfs2_clusters_to_blocks(sb, start);
7461 * For the first cluster group, the gd->bg_blkno is not at the start
7462 * of the group, but at an offset from the start. If we add it while
7463 * calculating discard for first group, we will wrongly start fstrim a
7464 * few blocks after the desried start block and the range can cross
7465 * over into the next cluster group. So, add it only if this is not
7466 * the first cluster group.
7468 if (group != osb->first_cluster_group_blkno)
7469 discard += le64_to_cpu(gd->bg_blkno);
7471 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7473 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7476 static int ocfs2_trim_group(struct super_block *sb,
7477 struct ocfs2_group_desc *gd, u64 group,
7478 u32 start, u32 max, u32 minbits)
7480 int ret = 0, count = 0, next;
7481 void *bitmap = gd->bg_bitmap;
7483 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7486 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7487 start, max, minbits);
7489 while (start < max) {
7490 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7493 next = ocfs2_find_next_bit(bitmap, max, start);
7495 if ((next - start) >= minbits) {
7496 ret = ocfs2_trim_extent(sb, gd, group,
7497 start, next - start);
7502 count += next - start;
7506 if (fatal_signal_pending(current)) {
7507 count = -ERESTARTSYS;
7511 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7521 int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7523 struct ocfs2_super *osb = OCFS2_SB(sb);
7524 u64 start, len, trimmed, first_group, last_group, group;
7526 u32 first_bit, last_bit, minlen;
7527 struct buffer_head *main_bm_bh = NULL;
7528 struct inode *main_bm_inode = NULL;
7529 struct buffer_head *gd_bh = NULL;
7530 struct ocfs2_dinode *main_bm;
7531 struct ocfs2_group_desc *gd = NULL;
7532 struct ocfs2_trim_fs_info info, *pinfo = NULL;
7534 start = range->start >> osb->s_clustersize_bits;
7535 len = range->len >> osb->s_clustersize_bits;
7536 minlen = range->minlen >> osb->s_clustersize_bits;
7538 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7541 main_bm_inode = ocfs2_get_system_file_inode(osb,
7542 GLOBAL_BITMAP_SYSTEM_INODE,
7543 OCFS2_INVALID_SLOT);
7544 if (!main_bm_inode) {
7550 inode_lock(main_bm_inode);
7552 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7557 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7559 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7564 len = range->len >> osb->s_clustersize_bits;
7565 if (start + len > le32_to_cpu(main_bm->i_clusters))
7566 len = le32_to_cpu(main_bm->i_clusters) - start;
7568 trace_ocfs2_trim_fs(start, len, minlen);
7570 ocfs2_trim_fs_lock_res_init(osb);
7571 ret = ocfs2_trim_fs_lock(osb, NULL, 1);
7573 if (ret != -EAGAIN) {
7575 ocfs2_trim_fs_lock_res_uninit(osb);
7579 mlog(ML_NOTICE, "Wait for trim on device (%s) to "
7580 "finish, which is running from another node.\n",
7582 ret = ocfs2_trim_fs_lock(osb, &info, 0);
7585 ocfs2_trim_fs_lock_res_uninit(osb);
7589 if (info.tf_valid && info.tf_success &&
7590 info.tf_start == start && info.tf_len == len &&
7591 info.tf_minlen == minlen) {
7592 /* Avoid sending duplicated trim to a shared device */
7593 mlog(ML_NOTICE, "The same trim on device (%s) was "
7594 "just done from node (%u), return.\n",
7595 osb->dev_str, info.tf_nodenum);
7596 range->len = info.tf_trimlen;
7597 goto out_trimunlock;
7601 info.tf_nodenum = osb->node_num;
7602 info.tf_start = start;
7604 info.tf_minlen = minlen;
7606 /* Determine first and last group to examine based on start and len */
7607 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7608 if (first_group == osb->first_cluster_group_blkno)
7611 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7612 last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7613 last_bit = osb->bitmap_cpg;
7616 for (group = first_group; group <= last_group;) {
7617 if (first_bit + len >= osb->bitmap_cpg)
7618 last_bit = osb->bitmap_cpg;
7620 last_bit = first_bit + len;
7622 ret = ocfs2_read_group_descriptor(main_bm_inode,
7630 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7631 cnt = ocfs2_trim_group(sb, gd, group,
7632 first_bit, last_bit, minlen);
7642 len -= osb->bitmap_cpg - first_bit;
7644 if (group == osb->first_cluster_group_blkno)
7645 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7647 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7649 range->len = trimmed * sb->s_blocksize;
7651 info.tf_trimlen = range->len;
7652 info.tf_success = (ret ? 0 : 1);
7655 ocfs2_trim_fs_unlock(osb, pinfo);
7656 ocfs2_trim_fs_lock_res_uninit(osb);
7658 ocfs2_inode_unlock(main_bm_inode, 0);
7661 inode_unlock(main_bm_inode);
7662 iput(main_bm_inode);