1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Extent allocs and frees
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
31 #include <linux/quotaops.h>
33 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
34 #include <cluster/masklog.h>
40 #include "blockcheck.h"
42 #include "extent_map.h"
45 #include "localalloc.h"
52 #include "refcounttree.h"
54 #include "buffer_head_io.h"
56 enum ocfs2_contig_type
{
63 static enum ocfs2_contig_type
64 ocfs2_extent_rec_contig(struct super_block
*sb
,
65 struct ocfs2_extent_rec
*ext
,
66 struct ocfs2_extent_rec
*insert_rec
);
68 * Operations for a specific extent tree type.
70 * To implement an on-disk btree (extent tree) type in ocfs2, add
71 * an ocfs2_extent_tree_operations structure and the matching
72 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
73 * for the allocation portion of the extent tree.
75 struct ocfs2_extent_tree_operations
{
77 * last_eb_blk is the block number of the right most leaf extent
78 * block. Most on-disk structures containing an extent tree store
79 * this value for fast access. The ->eo_set_last_eb_blk() and
80 * ->eo_get_last_eb_blk() operations access this value. They are
83 void (*eo_set_last_eb_blk
)(struct ocfs2_extent_tree
*et
,
85 u64 (*eo_get_last_eb_blk
)(struct ocfs2_extent_tree
*et
);
88 * The on-disk structure usually keeps track of how many total
89 * clusters are stored in this extent tree. This function updates
90 * that value. new_clusters is the delta, and must be
91 * added to the total. Required.
93 void (*eo_update_clusters
)(struct ocfs2_extent_tree
*et
,
97 * If this extent tree is supported by an extent map, insert
98 * a record into the map.
100 void (*eo_extent_map_insert
)(struct ocfs2_extent_tree
*et
,
101 struct ocfs2_extent_rec
*rec
);
104 * If this extent tree is supported by an extent map, truncate the
107 void (*eo_extent_map_truncate
)(struct ocfs2_extent_tree
*et
,
111 * If ->eo_insert_check() exists, it is called before rec is
112 * inserted into the extent tree. It is optional.
114 int (*eo_insert_check
)(struct ocfs2_extent_tree
*et
,
115 struct ocfs2_extent_rec
*rec
);
116 int (*eo_sanity_check
)(struct ocfs2_extent_tree
*et
);
119 * --------------------------------------------------------------
120 * The remaining are internal to ocfs2_extent_tree and don't have
125 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
128 void (*eo_fill_root_el
)(struct ocfs2_extent_tree
*et
);
131 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
132 * it exists. If it does not, et->et_max_leaf_clusters is set
133 * to 0 (unlimited). Optional.
135 void (*eo_fill_max_leaf_clusters
)(struct ocfs2_extent_tree
*et
);
138 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
139 * are contiguous or not. Optional. Don't need to set it if use
140 * ocfs2_extent_rec as the tree leaf.
142 enum ocfs2_contig_type
143 (*eo_extent_contig
)(struct ocfs2_extent_tree
*et
,
144 struct ocfs2_extent_rec
*ext
,
145 struct ocfs2_extent_rec
*insert_rec
);
150 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
153 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
);
154 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
156 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree
*et
,
158 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree
*et
,
159 struct ocfs2_extent_rec
*rec
);
160 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree
*et
,
162 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree
*et
,
163 struct ocfs2_extent_rec
*rec
);
164 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree
*et
);
165 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
);
166 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops
= {
167 .eo_set_last_eb_blk
= ocfs2_dinode_set_last_eb_blk
,
168 .eo_get_last_eb_blk
= ocfs2_dinode_get_last_eb_blk
,
169 .eo_update_clusters
= ocfs2_dinode_update_clusters
,
170 .eo_extent_map_insert
= ocfs2_dinode_extent_map_insert
,
171 .eo_extent_map_truncate
= ocfs2_dinode_extent_map_truncate
,
172 .eo_insert_check
= ocfs2_dinode_insert_check
,
173 .eo_sanity_check
= ocfs2_dinode_sanity_check
,
174 .eo_fill_root_el
= ocfs2_dinode_fill_root_el
,
177 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
180 struct ocfs2_dinode
*di
= et
->et_object
;
182 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
183 di
->i_last_eb_blk
= cpu_to_le64(blkno
);
186 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
188 struct ocfs2_dinode
*di
= et
->et_object
;
190 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
191 return le64_to_cpu(di
->i_last_eb_blk
);
194 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree
*et
,
197 struct ocfs2_inode_info
*oi
= cache_info_to_inode(et
->et_ci
);
198 struct ocfs2_dinode
*di
= et
->et_object
;
200 le32_add_cpu(&di
->i_clusters
, clusters
);
201 spin_lock(&oi
->ip_lock
);
202 oi
->ip_clusters
= le32_to_cpu(di
->i_clusters
);
203 spin_unlock(&oi
->ip_lock
);
206 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree
*et
,
207 struct ocfs2_extent_rec
*rec
)
209 struct inode
*inode
= &cache_info_to_inode(et
->et_ci
)->vfs_inode
;
211 ocfs2_extent_map_insert_rec(inode
, rec
);
214 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree
*et
,
217 struct inode
*inode
= &cache_info_to_inode(et
->et_ci
)->vfs_inode
;
219 ocfs2_extent_map_trunc(inode
, clusters
);
222 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree
*et
,
223 struct ocfs2_extent_rec
*rec
)
225 struct ocfs2_inode_info
*oi
= cache_info_to_inode(et
->et_ci
);
226 struct ocfs2_super
*osb
= OCFS2_SB(oi
->vfs_inode
.i_sb
);
228 BUG_ON(oi
->ip_dyn_features
& OCFS2_INLINE_DATA_FL
);
229 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb
) &&
230 (oi
->ip_clusters
!= le32_to_cpu(rec
->e_cpos
)),
231 "Device %s, asking for sparse allocation: inode %llu, "
232 "cpos %u, clusters %u\n",
234 (unsigned long long)oi
->ip_blkno
,
235 rec
->e_cpos
, oi
->ip_clusters
);
240 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree
*et
)
242 struct ocfs2_dinode
*di
= et
->et_object
;
244 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
245 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
250 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
)
252 struct ocfs2_dinode
*di
= et
->et_object
;
254 et
->et_root_el
= &di
->id2
.i_list
;
258 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree
*et
)
260 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
262 et
->et_root_el
= &vb
->vb_xv
->xr_list
;
265 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
268 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
270 vb
->vb_xv
->xr_last_eb_blk
= cpu_to_le64(blkno
);
273 static u64
ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
275 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
277 return le64_to_cpu(vb
->vb_xv
->xr_last_eb_blk
);
280 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree
*et
,
283 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
285 le32_add_cpu(&vb
->vb_xv
->xr_clusters
, clusters
);
288 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops
= {
289 .eo_set_last_eb_blk
= ocfs2_xattr_value_set_last_eb_blk
,
290 .eo_get_last_eb_blk
= ocfs2_xattr_value_get_last_eb_blk
,
291 .eo_update_clusters
= ocfs2_xattr_value_update_clusters
,
292 .eo_fill_root_el
= ocfs2_xattr_value_fill_root_el
,
295 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree
*et
)
297 struct ocfs2_xattr_block
*xb
= et
->et_object
;
299 et
->et_root_el
= &xb
->xb_attrs
.xb_root
.xt_list
;
302 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree
*et
)
304 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
305 et
->et_max_leaf_clusters
=
306 ocfs2_clusters_for_bytes(sb
, OCFS2_MAX_XATTR_TREE_LEAF_SIZE
);
309 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
312 struct ocfs2_xattr_block
*xb
= et
->et_object
;
313 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
315 xt
->xt_last_eb_blk
= cpu_to_le64(blkno
);
318 static u64
ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
320 struct ocfs2_xattr_block
*xb
= et
->et_object
;
321 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
323 return le64_to_cpu(xt
->xt_last_eb_blk
);
326 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree
*et
,
329 struct ocfs2_xattr_block
*xb
= et
->et_object
;
331 le32_add_cpu(&xb
->xb_attrs
.xb_root
.xt_clusters
, clusters
);
334 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops
= {
335 .eo_set_last_eb_blk
= ocfs2_xattr_tree_set_last_eb_blk
,
336 .eo_get_last_eb_blk
= ocfs2_xattr_tree_get_last_eb_blk
,
337 .eo_update_clusters
= ocfs2_xattr_tree_update_clusters
,
338 .eo_fill_root_el
= ocfs2_xattr_tree_fill_root_el
,
339 .eo_fill_max_leaf_clusters
= ocfs2_xattr_tree_fill_max_leaf_clusters
,
342 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
345 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
347 dx_root
->dr_last_eb_blk
= cpu_to_le64(blkno
);
350 static u64
ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
352 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
354 return le64_to_cpu(dx_root
->dr_last_eb_blk
);
357 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree
*et
,
360 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
362 le32_add_cpu(&dx_root
->dr_clusters
, clusters
);
365 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree
*et
)
367 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
369 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root
));
374 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree
*et
)
376 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
378 et
->et_root_el
= &dx_root
->dr_list
;
381 static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops
= {
382 .eo_set_last_eb_blk
= ocfs2_dx_root_set_last_eb_blk
,
383 .eo_get_last_eb_blk
= ocfs2_dx_root_get_last_eb_blk
,
384 .eo_update_clusters
= ocfs2_dx_root_update_clusters
,
385 .eo_sanity_check
= ocfs2_dx_root_sanity_check
,
386 .eo_fill_root_el
= ocfs2_dx_root_fill_root_el
,
389 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree
*et
)
391 struct ocfs2_refcount_block
*rb
= et
->et_object
;
393 et
->et_root_el
= &rb
->rf_list
;
396 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
399 struct ocfs2_refcount_block
*rb
= et
->et_object
;
401 rb
->rf_last_eb_blk
= cpu_to_le64(blkno
);
404 static u64
ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
406 struct ocfs2_refcount_block
*rb
= et
->et_object
;
408 return le64_to_cpu(rb
->rf_last_eb_blk
);
411 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree
*et
,
414 struct ocfs2_refcount_block
*rb
= et
->et_object
;
416 le32_add_cpu(&rb
->rf_clusters
, clusters
);
419 static enum ocfs2_contig_type
420 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree
*et
,
421 struct ocfs2_extent_rec
*ext
,
422 struct ocfs2_extent_rec
*insert_rec
)
427 static struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops
= {
428 .eo_set_last_eb_blk
= ocfs2_refcount_tree_set_last_eb_blk
,
429 .eo_get_last_eb_blk
= ocfs2_refcount_tree_get_last_eb_blk
,
430 .eo_update_clusters
= ocfs2_refcount_tree_update_clusters
,
431 .eo_fill_root_el
= ocfs2_refcount_tree_fill_root_el
,
432 .eo_extent_contig
= ocfs2_refcount_tree_extent_contig
,
435 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree
*et
,
436 struct ocfs2_caching_info
*ci
,
437 struct buffer_head
*bh
,
438 ocfs2_journal_access_func access
,
440 struct ocfs2_extent_tree_operations
*ops
)
445 et
->et_root_journal_access
= access
;
447 obj
= (void *)bh
->b_data
;
450 et
->et_ops
->eo_fill_root_el(et
);
451 if (!et
->et_ops
->eo_fill_max_leaf_clusters
)
452 et
->et_max_leaf_clusters
= 0;
454 et
->et_ops
->eo_fill_max_leaf_clusters(et
);
457 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree
*et
,
458 struct ocfs2_caching_info
*ci
,
459 struct buffer_head
*bh
)
461 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_di
,
462 NULL
, &ocfs2_dinode_et_ops
);
465 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree
*et
,
466 struct ocfs2_caching_info
*ci
,
467 struct buffer_head
*bh
)
469 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_xb
,
470 NULL
, &ocfs2_xattr_tree_et_ops
);
473 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree
*et
,
474 struct ocfs2_caching_info
*ci
,
475 struct ocfs2_xattr_value_buf
*vb
)
477 __ocfs2_init_extent_tree(et
, ci
, vb
->vb_bh
, vb
->vb_access
, vb
,
478 &ocfs2_xattr_value_et_ops
);
481 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree
*et
,
482 struct ocfs2_caching_info
*ci
,
483 struct buffer_head
*bh
)
485 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_dr
,
486 NULL
, &ocfs2_dx_root_et_ops
);
489 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree
*et
,
490 struct ocfs2_caching_info
*ci
,
491 struct buffer_head
*bh
)
493 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_rb
,
494 NULL
, &ocfs2_refcount_tree_et_ops
);
497 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
500 et
->et_ops
->eo_set_last_eb_blk(et
, new_last_eb_blk
);
503 static inline u64
ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
505 return et
->et_ops
->eo_get_last_eb_blk(et
);
508 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree
*et
,
511 et
->et_ops
->eo_update_clusters(et
, clusters
);
514 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree
*et
,
515 struct ocfs2_extent_rec
*rec
)
517 if (et
->et_ops
->eo_extent_map_insert
)
518 et
->et_ops
->eo_extent_map_insert(et
, rec
);
521 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree
*et
,
524 if (et
->et_ops
->eo_extent_map_truncate
)
525 et
->et_ops
->eo_extent_map_truncate(et
, clusters
);
528 static inline int ocfs2_et_root_journal_access(handle_t
*handle
,
529 struct ocfs2_extent_tree
*et
,
532 return et
->et_root_journal_access(handle
, et
->et_ci
, et
->et_root_bh
,
536 static inline enum ocfs2_contig_type
537 ocfs2_et_extent_contig(struct ocfs2_extent_tree
*et
,
538 struct ocfs2_extent_rec
*rec
,
539 struct ocfs2_extent_rec
*insert_rec
)
541 if (et
->et_ops
->eo_extent_contig
)
542 return et
->et_ops
->eo_extent_contig(et
, rec
, insert_rec
);
544 return ocfs2_extent_rec_contig(
545 ocfs2_metadata_cache_get_super(et
->et_ci
),
549 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree
*et
,
550 struct ocfs2_extent_rec
*rec
)
554 if (et
->et_ops
->eo_insert_check
)
555 ret
= et
->et_ops
->eo_insert_check(et
, rec
);
559 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree
*et
)
563 if (et
->et_ops
->eo_sanity_check
)
564 ret
= et
->et_ops
->eo_sanity_check(et
);
568 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
);
569 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
570 struct ocfs2_extent_block
*eb
);
571 static void ocfs2_adjust_rightmost_records(handle_t
*handle
,
572 struct ocfs2_extent_tree
*et
,
573 struct ocfs2_path
*path
,
574 struct ocfs2_extent_rec
*insert_rec
);
576 * Reset the actual path elements so that we can re-use the structure
577 * to build another path. Generally, this involves freeing the buffer
580 void ocfs2_reinit_path(struct ocfs2_path
*path
, int keep_root
)
582 int i
, start
= 0, depth
= 0;
583 struct ocfs2_path_item
*node
;
588 for(i
= start
; i
< path_num_items(path
); i
++) {
589 node
= &path
->p_node
[i
];
597 * Tree depth may change during truncate, or insert. If we're
598 * keeping the root extent list, then make sure that our path
599 * structure reflects the proper depth.
602 depth
= le16_to_cpu(path_root_el(path
)->l_tree_depth
);
604 path_root_access(path
) = NULL
;
606 path
->p_tree_depth
= depth
;
609 void ocfs2_free_path(struct ocfs2_path
*path
)
612 ocfs2_reinit_path(path
, 0);
618 * All the elements of src into dest. After this call, src could be freed
619 * without affecting dest.
621 * Both paths should have the same root. Any non-root elements of dest
624 static void ocfs2_cp_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
628 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
629 BUG_ON(path_root_el(dest
) != path_root_el(src
));
630 BUG_ON(path_root_access(dest
) != path_root_access(src
));
632 ocfs2_reinit_path(dest
, 1);
634 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
635 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
636 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
638 if (dest
->p_node
[i
].bh
)
639 get_bh(dest
->p_node
[i
].bh
);
644 * Make the *dest path the same as src and re-initialize src path to
647 static void ocfs2_mv_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
651 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
652 BUG_ON(path_root_access(dest
) != path_root_access(src
));
654 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
655 brelse(dest
->p_node
[i
].bh
);
657 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
658 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
660 src
->p_node
[i
].bh
= NULL
;
661 src
->p_node
[i
].el
= NULL
;
666 * Insert an extent block at given index.
668 * This will not take an additional reference on eb_bh.
670 static inline void ocfs2_path_insert_eb(struct ocfs2_path
*path
, int index
,
671 struct buffer_head
*eb_bh
)
673 struct ocfs2_extent_block
*eb
= (struct ocfs2_extent_block
*)eb_bh
->b_data
;
676 * Right now, no root bh is an extent block, so this helps
677 * catch code errors with dinode trees. The assertion can be
678 * safely removed if we ever need to insert extent block
679 * structures at the root.
683 path
->p_node
[index
].bh
= eb_bh
;
684 path
->p_node
[index
].el
= &eb
->h_list
;
687 static struct ocfs2_path
*ocfs2_new_path(struct buffer_head
*root_bh
,
688 struct ocfs2_extent_list
*root_el
,
689 ocfs2_journal_access_func access
)
691 struct ocfs2_path
*path
;
693 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) >= OCFS2_MAX_PATH_DEPTH
);
695 path
= kzalloc(sizeof(*path
), GFP_NOFS
);
697 path
->p_tree_depth
= le16_to_cpu(root_el
->l_tree_depth
);
699 path_root_bh(path
) = root_bh
;
700 path_root_el(path
) = root_el
;
701 path_root_access(path
) = access
;
707 struct ocfs2_path
*ocfs2_new_path_from_path(struct ocfs2_path
*path
)
709 return ocfs2_new_path(path_root_bh(path
), path_root_el(path
),
710 path_root_access(path
));
713 struct ocfs2_path
*ocfs2_new_path_from_et(struct ocfs2_extent_tree
*et
)
715 return ocfs2_new_path(et
->et_root_bh
, et
->et_root_el
,
716 et
->et_root_journal_access
);
720 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
721 * otherwise it's the root_access function.
723 * I don't like the way this function's name looks next to
724 * ocfs2_journal_access_path(), but I don't have a better one.
726 int ocfs2_path_bh_journal_access(handle_t
*handle
,
727 struct ocfs2_caching_info
*ci
,
728 struct ocfs2_path
*path
,
731 ocfs2_journal_access_func access
= path_root_access(path
);
734 access
= ocfs2_journal_access
;
737 access
= ocfs2_journal_access_eb
;
739 return access(handle
, ci
, path
->p_node
[idx
].bh
,
740 OCFS2_JOURNAL_ACCESS_WRITE
);
744 * Convenience function to journal all components in a path.
746 int ocfs2_journal_access_path(struct ocfs2_caching_info
*ci
,
748 struct ocfs2_path
*path
)
755 for(i
= 0; i
< path_num_items(path
); i
++) {
756 ret
= ocfs2_path_bh_journal_access(handle
, ci
, path
, i
);
768 * Return the index of the extent record which contains cluster #v_cluster.
769 * -1 is returned if it was not found.
771 * Should work fine on interior and exterior nodes.
773 int ocfs2_search_extent_list(struct ocfs2_extent_list
*el
, u32 v_cluster
)
777 struct ocfs2_extent_rec
*rec
;
778 u32 rec_end
, rec_start
, clusters
;
780 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
781 rec
= &el
->l_recs
[i
];
783 rec_start
= le32_to_cpu(rec
->e_cpos
);
784 clusters
= ocfs2_rec_clusters(el
, rec
);
786 rec_end
= rec_start
+ clusters
;
788 if (v_cluster
>= rec_start
&& v_cluster
< rec_end
) {
798 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
799 * ocfs2_extent_rec_contig only work properly against leaf nodes!
801 static int ocfs2_block_extent_contig(struct super_block
*sb
,
802 struct ocfs2_extent_rec
*ext
,
805 u64 blk_end
= le64_to_cpu(ext
->e_blkno
);
807 blk_end
+= ocfs2_clusters_to_blocks(sb
,
808 le16_to_cpu(ext
->e_leaf_clusters
));
810 return blkno
== blk_end
;
813 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec
*left
,
814 struct ocfs2_extent_rec
*right
)
818 left_range
= le32_to_cpu(left
->e_cpos
) +
819 le16_to_cpu(left
->e_leaf_clusters
);
821 return (left_range
== le32_to_cpu(right
->e_cpos
));
824 static enum ocfs2_contig_type
825 ocfs2_extent_rec_contig(struct super_block
*sb
,
826 struct ocfs2_extent_rec
*ext
,
827 struct ocfs2_extent_rec
*insert_rec
)
829 u64 blkno
= le64_to_cpu(insert_rec
->e_blkno
);
832 * Refuse to coalesce extent records with different flag
833 * fields - we don't want to mix unwritten extents with user
836 if (ext
->e_flags
!= insert_rec
->e_flags
)
839 if (ocfs2_extents_adjacent(ext
, insert_rec
) &&
840 ocfs2_block_extent_contig(sb
, ext
, blkno
))
843 blkno
= le64_to_cpu(ext
->e_blkno
);
844 if (ocfs2_extents_adjacent(insert_rec
, ext
) &&
845 ocfs2_block_extent_contig(sb
, insert_rec
, blkno
))
852 * NOTE: We can have pretty much any combination of contiguousness and
855 * The usefulness of APPEND_TAIL is more in that it lets us know that
856 * we'll have to update the path to that leaf.
858 enum ocfs2_append_type
{
863 enum ocfs2_split_type
{
869 struct ocfs2_insert_type
{
870 enum ocfs2_split_type ins_split
;
871 enum ocfs2_append_type ins_appending
;
872 enum ocfs2_contig_type ins_contig
;
873 int ins_contig_index
;
877 struct ocfs2_merge_ctxt
{
878 enum ocfs2_contig_type c_contig_type
;
879 int c_has_empty_extent
;
880 int c_split_covers_rec
;
883 static int ocfs2_validate_extent_block(struct super_block
*sb
,
884 struct buffer_head
*bh
)
887 struct ocfs2_extent_block
*eb
=
888 (struct ocfs2_extent_block
*)bh
->b_data
;
890 mlog(0, "Validating extent block %llu\n",
891 (unsigned long long)bh
->b_blocknr
);
893 BUG_ON(!buffer_uptodate(bh
));
896 * If the ecc fails, we return the error but otherwise
897 * leave the filesystem running. We know any error is
898 * local to this block.
900 rc
= ocfs2_validate_meta_ecc(sb
, bh
->b_data
, &eb
->h_check
);
902 mlog(ML_ERROR
, "Checksum failed for extent block %llu\n",
903 (unsigned long long)bh
->b_blocknr
);
908 * Errors after here are fatal.
911 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
913 "Extent block #%llu has bad signature %.*s",
914 (unsigned long long)bh
->b_blocknr
, 7,
919 if (le64_to_cpu(eb
->h_blkno
) != bh
->b_blocknr
) {
921 "Extent block #%llu has an invalid h_blkno "
923 (unsigned long long)bh
->b_blocknr
,
924 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
928 if (le32_to_cpu(eb
->h_fs_generation
) != OCFS2_SB(sb
)->fs_generation
) {
930 "Extent block #%llu has an invalid "
931 "h_fs_generation of #%u",
932 (unsigned long long)bh
->b_blocknr
,
933 le32_to_cpu(eb
->h_fs_generation
));
940 int ocfs2_read_extent_block(struct ocfs2_caching_info
*ci
, u64 eb_blkno
,
941 struct buffer_head
**bh
)
944 struct buffer_head
*tmp
= *bh
;
946 rc
= ocfs2_read_block(ci
, eb_blkno
, &tmp
,
947 ocfs2_validate_extent_block
);
949 /* If ocfs2_read_block() got us a new bh, pass it up. */
958 * How many free extents have we got before we need more meta data?
960 int ocfs2_num_free_extents(struct ocfs2_super
*osb
,
961 struct ocfs2_extent_tree
*et
)
964 struct ocfs2_extent_list
*el
= NULL
;
965 struct ocfs2_extent_block
*eb
;
966 struct buffer_head
*eb_bh
= NULL
;
972 last_eb_blk
= ocfs2_et_get_last_eb_blk(et
);
975 retval
= ocfs2_read_extent_block(et
->et_ci
, last_eb_blk
,
981 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
985 BUG_ON(el
->l_tree_depth
!= 0);
987 retval
= le16_to_cpu(el
->l_count
) - le16_to_cpu(el
->l_next_free_rec
);
995 /* expects array to already be allocated
997 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
1000 static int ocfs2_create_new_meta_bhs(handle_t
*handle
,
1001 struct ocfs2_extent_tree
*et
,
1003 struct ocfs2_alloc_context
*meta_ac
,
1004 struct buffer_head
*bhs
[])
1006 int count
, status
, i
;
1007 u16 suballoc_bit_start
;
1010 struct ocfs2_super
*osb
=
1011 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
1012 struct ocfs2_extent_block
*eb
;
1017 while (count
< wanted
) {
1018 status
= ocfs2_claim_metadata(osb
,
1022 &suballoc_bit_start
,
1030 for(i
= count
; i
< (num_got
+ count
); i
++) {
1031 bhs
[i
] = sb_getblk(osb
->sb
, first_blkno
);
1032 if (bhs
[i
] == NULL
) {
1037 ocfs2_set_new_buffer_uptodate(et
->et_ci
, bhs
[i
]);
1039 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
1041 OCFS2_JOURNAL_ACCESS_CREATE
);
1047 memset(bhs
[i
]->b_data
, 0, osb
->sb
->s_blocksize
);
1048 eb
= (struct ocfs2_extent_block
*) bhs
[i
]->b_data
;
1049 /* Ok, setup the minimal stuff here. */
1050 strcpy(eb
->h_signature
, OCFS2_EXTENT_BLOCK_SIGNATURE
);
1051 eb
->h_blkno
= cpu_to_le64(first_blkno
);
1052 eb
->h_fs_generation
= cpu_to_le32(osb
->fs_generation
);
1053 eb
->h_suballoc_slot
=
1054 cpu_to_le16(meta_ac
->ac_alloc_slot
);
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 status
= ocfs2_journal_dirty(handle
, bhs
[i
]);
1077 for(i
= 0; i
< wanted
; i
++) {
1087 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1089 * Returns the sum of the rightmost extent rec logical offset and
1092 * ocfs2_add_branch() uses this to determine what logical cluster
1093 * value should be populated into the leftmost new branch records.
1095 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1096 * value for the new topmost tree record.
1098 static inline u32
ocfs2_sum_rightmost_rec(struct ocfs2_extent_list
*el
)
1102 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1104 return le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
1105 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
1109 * Change range of the branches in the right most path according to the leaf
1110 * extent block's rightmost record.
1112 static int ocfs2_adjust_rightmost_branch(handle_t
*handle
,
1113 struct ocfs2_extent_tree
*et
)
1116 struct ocfs2_path
*path
= NULL
;
1117 struct ocfs2_extent_list
*el
;
1118 struct ocfs2_extent_rec
*rec
;
1120 path
= ocfs2_new_path_from_et(et
);
1126 status
= ocfs2_find_path(et
->et_ci
, path
, UINT_MAX
);
1132 status
= ocfs2_extend_trans(handle
, path_num_items(path
) +
1133 handle
->h_buffer_credits
);
1139 status
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
1145 el
= path_leaf_el(path
);
1146 rec
= &el
->l_recs
[le32_to_cpu(el
->l_next_free_rec
) - 1];
1148 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
1151 ocfs2_free_path(path
);
1156 * Add an entire tree branch to our inode. eb_bh is the extent block
1157 * to start at, if we don't want to start the branch at the root
1160 * last_eb_bh is required as we have to update it's next_leaf pointer
1161 * for the new last extent block.
1163 * the new branch will be 'empty' in the sense that every block will
1164 * contain a single record with cluster count == 0.
1166 static int ocfs2_add_branch(handle_t
*handle
,
1167 struct ocfs2_extent_tree
*et
,
1168 struct buffer_head
*eb_bh
,
1169 struct buffer_head
**last_eb_bh
,
1170 struct ocfs2_alloc_context
*meta_ac
)
1172 int status
, new_blocks
, i
;
1173 u64 next_blkno
, new_last_eb_blk
;
1174 struct buffer_head
*bh
;
1175 struct buffer_head
**new_eb_bhs
= NULL
;
1176 struct ocfs2_extent_block
*eb
;
1177 struct ocfs2_extent_list
*eb_el
;
1178 struct ocfs2_extent_list
*el
;
1179 u32 new_cpos
, root_end
;
1183 BUG_ON(!last_eb_bh
|| !*last_eb_bh
);
1186 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
1189 el
= et
->et_root_el
;
1191 /* we never add a branch to a leaf. */
1192 BUG_ON(!el
->l_tree_depth
);
1194 new_blocks
= le16_to_cpu(el
->l_tree_depth
);
1196 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
1197 new_cpos
= ocfs2_sum_rightmost_rec(&eb
->h_list
);
1198 root_end
= ocfs2_sum_rightmost_rec(et
->et_root_el
);
1201 * If there is a gap before the root end and the real end
1202 * of the righmost leaf block, we need to remove the gap
1203 * between new_cpos and root_end first so that the tree
1204 * is consistent after we add a new branch(it will start
1207 if (root_end
> new_cpos
) {
1208 mlog(0, "adjust the cluster end from %u to %u\n",
1209 root_end
, new_cpos
);
1210 status
= ocfs2_adjust_rightmost_branch(handle
, et
);
1217 /* allocate the number of new eb blocks we need */
1218 new_eb_bhs
= kcalloc(new_blocks
, sizeof(struct buffer_head
*),
1226 status
= ocfs2_create_new_meta_bhs(handle
, et
, new_blocks
,
1227 meta_ac
, new_eb_bhs
);
1233 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1234 * linked with the rest of the tree.
1235 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1237 * when we leave the loop, new_last_eb_blk will point to the
1238 * newest leaf, and next_blkno will point to the topmost extent
1240 next_blkno
= new_last_eb_blk
= 0;
1241 for(i
= 0; i
< new_blocks
; i
++) {
1243 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1244 /* ocfs2_create_new_meta_bhs() should create it right! */
1245 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1246 eb_el
= &eb
->h_list
;
1248 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, bh
,
1249 OCFS2_JOURNAL_ACCESS_CREATE
);
1255 eb
->h_next_leaf_blk
= 0;
1256 eb_el
->l_tree_depth
= cpu_to_le16(i
);
1257 eb_el
->l_next_free_rec
= cpu_to_le16(1);
1259 * This actually counts as an empty extent as
1262 eb_el
->l_recs
[0].e_cpos
= cpu_to_le32(new_cpos
);
1263 eb_el
->l_recs
[0].e_blkno
= cpu_to_le64(next_blkno
);
1265 * eb_el isn't always an interior node, but even leaf
1266 * nodes want a zero'd flags and reserved field so
1267 * this gets the whole 32 bits regardless of use.
1269 eb_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(0);
1270 if (!eb_el
->l_tree_depth
)
1271 new_last_eb_blk
= le64_to_cpu(eb
->h_blkno
);
1273 status
= ocfs2_journal_dirty(handle
, bh
);
1279 next_blkno
= le64_to_cpu(eb
->h_blkno
);
1282 /* This is a bit hairy. We want to update up to three blocks
1283 * here without leaving any of them in an inconsistent state
1284 * in case of error. We don't have to worry about
1285 * journal_dirty erroring as it won't unless we've aborted the
1286 * handle (in which case we would never be here) so reserving
1287 * the write with journal_access is all we need to do. */
1288 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, *last_eb_bh
,
1289 OCFS2_JOURNAL_ACCESS_WRITE
);
1294 status
= ocfs2_et_root_journal_access(handle
, et
,
1295 OCFS2_JOURNAL_ACCESS_WRITE
);
1301 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, eb_bh
,
1302 OCFS2_JOURNAL_ACCESS_WRITE
);
1309 /* Link the new branch into the rest of the tree (el will
1310 * either be on the root_bh, or the extent block passed in. */
1311 i
= le16_to_cpu(el
->l_next_free_rec
);
1312 el
->l_recs
[i
].e_blkno
= cpu_to_le64(next_blkno
);
1313 el
->l_recs
[i
].e_cpos
= cpu_to_le32(new_cpos
);
1314 el
->l_recs
[i
].e_int_clusters
= 0;
1315 le16_add_cpu(&el
->l_next_free_rec
, 1);
1317 /* fe needs a new last extent block pointer, as does the
1318 * next_leaf on the previously last-extent-block. */
1319 ocfs2_et_set_last_eb_blk(et
, new_last_eb_blk
);
1321 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
1322 eb
->h_next_leaf_blk
= cpu_to_le64(new_last_eb_blk
);
1324 status
= ocfs2_journal_dirty(handle
, *last_eb_bh
);
1327 status
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1331 status
= ocfs2_journal_dirty(handle
, eb_bh
);
1337 * Some callers want to track the rightmost leaf so pass it
1340 brelse(*last_eb_bh
);
1341 get_bh(new_eb_bhs
[0]);
1342 *last_eb_bh
= new_eb_bhs
[0];
1347 for (i
= 0; i
< new_blocks
; i
++)
1348 brelse(new_eb_bhs
[i
]);
1357 * adds another level to the allocation tree.
1358 * returns back the new extent block so you can add a branch to it
1361 static int ocfs2_shift_tree_depth(handle_t
*handle
,
1362 struct ocfs2_extent_tree
*et
,
1363 struct ocfs2_alloc_context
*meta_ac
,
1364 struct buffer_head
**ret_new_eb_bh
)
1368 struct buffer_head
*new_eb_bh
= NULL
;
1369 struct ocfs2_extent_block
*eb
;
1370 struct ocfs2_extent_list
*root_el
;
1371 struct ocfs2_extent_list
*eb_el
;
1375 status
= ocfs2_create_new_meta_bhs(handle
, et
, 1, meta_ac
,
1382 eb
= (struct ocfs2_extent_block
*) new_eb_bh
->b_data
;
1383 /* ocfs2_create_new_meta_bhs() should create it right! */
1384 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1386 eb_el
= &eb
->h_list
;
1387 root_el
= et
->et_root_el
;
1389 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, new_eb_bh
,
1390 OCFS2_JOURNAL_ACCESS_CREATE
);
1396 /* copy the root extent list data into the new extent block */
1397 eb_el
->l_tree_depth
= root_el
->l_tree_depth
;
1398 eb_el
->l_next_free_rec
= root_el
->l_next_free_rec
;
1399 for (i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1400 eb_el
->l_recs
[i
] = root_el
->l_recs
[i
];
1402 status
= ocfs2_journal_dirty(handle
, new_eb_bh
);
1408 status
= ocfs2_et_root_journal_access(handle
, et
,
1409 OCFS2_JOURNAL_ACCESS_WRITE
);
1415 new_clusters
= ocfs2_sum_rightmost_rec(eb_el
);
1417 /* update root_bh now */
1418 le16_add_cpu(&root_el
->l_tree_depth
, 1);
1419 root_el
->l_recs
[0].e_cpos
= 0;
1420 root_el
->l_recs
[0].e_blkno
= eb
->h_blkno
;
1421 root_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(new_clusters
);
1422 for (i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1423 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
1424 root_el
->l_next_free_rec
= cpu_to_le16(1);
1426 /* If this is our 1st tree depth shift, then last_eb_blk
1427 * becomes the allocated extent block */
1428 if (root_el
->l_tree_depth
== cpu_to_le16(1))
1429 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
1431 status
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1437 *ret_new_eb_bh
= new_eb_bh
;
1448 * Should only be called when there is no space left in any of the
1449 * leaf nodes. What we want to do is find the lowest tree depth
1450 * non-leaf extent block with room for new records. There are three
1451 * valid results of this search:
1453 * 1) a lowest extent block is found, then we pass it back in
1454 * *lowest_eb_bh and return '0'
1456 * 2) the search fails to find anything, but the root_el has room. We
1457 * pass NULL back in *lowest_eb_bh, but still return '0'
1459 * 3) the search fails to find anything AND the root_el is full, in
1460 * which case we return > 0
1462 * return status < 0 indicates an error.
1464 static int ocfs2_find_branch_target(struct ocfs2_extent_tree
*et
,
1465 struct buffer_head
**target_bh
)
1469 struct ocfs2_extent_block
*eb
;
1470 struct ocfs2_extent_list
*el
;
1471 struct buffer_head
*bh
= NULL
;
1472 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 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1483 "Owner %llu has empty "
1484 "extent list (next_free_rec == 0)",
1485 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
1489 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1490 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1492 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1493 "Owner %llu has extent "
1494 "list where extent # %d has no physical "
1496 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
), i
);
1504 status
= ocfs2_read_extent_block(et
->et_ci
, blkno
, &bh
);
1510 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1513 if (le16_to_cpu(el
->l_next_free_rec
) <
1514 le16_to_cpu(el
->l_count
)) {
1521 /* If we didn't find one and the fe doesn't have any room,
1522 * then return '1' */
1523 el
= et
->et_root_el
;
1524 if (!lowest_bh
&& (el
->l_next_free_rec
== el
->l_count
))
1527 *target_bh
= lowest_bh
;
1536 * Grow a b-tree so that it has more records.
1538 * We might shift the tree depth in which case existing paths should
1539 * be considered invalid.
1541 * Tree depth after the grow is returned via *final_depth.
1543 * *last_eb_bh will be updated by ocfs2_add_branch().
1545 static int ocfs2_grow_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
1546 int *final_depth
, struct buffer_head
**last_eb_bh
,
1547 struct ocfs2_alloc_context
*meta_ac
)
1550 struct ocfs2_extent_list
*el
= et
->et_root_el
;
1551 int depth
= le16_to_cpu(el
->l_tree_depth
);
1552 struct buffer_head
*bh
= NULL
;
1554 BUG_ON(meta_ac
== NULL
);
1556 shift
= ocfs2_find_branch_target(et
, &bh
);
1563 /* We traveled all the way to the bottom of the allocation tree
1564 * and didn't find room for any more extents - we need to add
1565 * another tree level */
1568 mlog(0, "need to shift tree depth (current = %d)\n", depth
);
1570 /* ocfs2_shift_tree_depth will return us a buffer with
1571 * the new extent block (so we can pass that to
1572 * ocfs2_add_branch). */
1573 ret
= ocfs2_shift_tree_depth(handle
, et
, meta_ac
, &bh
);
1581 * Special case: we have room now if we shifted from
1582 * tree_depth 0, so no more work needs to be done.
1584 * We won't be calling add_branch, so pass
1585 * back *last_eb_bh as the new leaf. At depth
1586 * zero, it should always be null so there's
1587 * no reason to brelse.
1589 BUG_ON(*last_eb_bh
);
1596 /* call ocfs2_add_branch to add the final part of the tree with
1598 mlog(0, "add branch. bh = %p\n", bh
);
1599 ret
= ocfs2_add_branch(handle
, et
, bh
, last_eb_bh
,
1608 *final_depth
= depth
;
1614 * This function will discard the rightmost extent record.
1616 static void ocfs2_shift_records_right(struct ocfs2_extent_list
*el
)
1618 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1619 int count
= le16_to_cpu(el
->l_count
);
1620 unsigned int num_bytes
;
1623 /* This will cause us to go off the end of our extent list. */
1624 BUG_ON(next_free
>= count
);
1626 num_bytes
= sizeof(struct ocfs2_extent_rec
) * next_free
;
1628 memmove(&el
->l_recs
[1], &el
->l_recs
[0], num_bytes
);
1631 static void ocfs2_rotate_leaf(struct ocfs2_extent_list
*el
,
1632 struct ocfs2_extent_rec
*insert_rec
)
1634 int i
, insert_index
, next_free
, has_empty
, num_bytes
;
1635 u32 insert_cpos
= le32_to_cpu(insert_rec
->e_cpos
);
1636 struct ocfs2_extent_rec
*rec
;
1638 next_free
= le16_to_cpu(el
->l_next_free_rec
);
1639 has_empty
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
1643 /* The tree code before us didn't allow enough room in the leaf. */
1644 BUG_ON(el
->l_next_free_rec
== el
->l_count
&& !has_empty
);
1647 * The easiest way to approach this is to just remove the
1648 * empty extent and temporarily decrement next_free.
1652 * If next_free was 1 (only an empty extent), this
1653 * loop won't execute, which is fine. We still want
1654 * the decrement above to happen.
1656 for(i
= 0; i
< (next_free
- 1); i
++)
1657 el
->l_recs
[i
] = el
->l_recs
[i
+1];
1663 * Figure out what the new record index should be.
1665 for(i
= 0; i
< next_free
; i
++) {
1666 rec
= &el
->l_recs
[i
];
1668 if (insert_cpos
< le32_to_cpu(rec
->e_cpos
))
1673 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1674 insert_cpos
, insert_index
, has_empty
, next_free
, le16_to_cpu(el
->l_count
));
1676 BUG_ON(insert_index
< 0);
1677 BUG_ON(insert_index
>= le16_to_cpu(el
->l_count
));
1678 BUG_ON(insert_index
> next_free
);
1681 * No need to memmove if we're just adding to the tail.
1683 if (insert_index
!= next_free
) {
1684 BUG_ON(next_free
>= le16_to_cpu(el
->l_count
));
1686 num_bytes
= next_free
- insert_index
;
1687 num_bytes
*= sizeof(struct ocfs2_extent_rec
);
1688 memmove(&el
->l_recs
[insert_index
+ 1],
1689 &el
->l_recs
[insert_index
],
1694 * Either we had an empty extent, and need to re-increment or
1695 * there was no empty extent on a non full rightmost leaf node,
1696 * in which case we still need to increment.
1699 el
->l_next_free_rec
= cpu_to_le16(next_free
);
1701 * Make sure none of the math above just messed up our tree.
1703 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) > le16_to_cpu(el
->l_count
));
1705 el
->l_recs
[insert_index
] = *insert_rec
;
1709 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list
*el
)
1711 int size
, num_recs
= le16_to_cpu(el
->l_next_free_rec
);
1713 BUG_ON(num_recs
== 0);
1715 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
1717 size
= num_recs
* sizeof(struct ocfs2_extent_rec
);
1718 memmove(&el
->l_recs
[0], &el
->l_recs
[1], size
);
1719 memset(&el
->l_recs
[num_recs
], 0,
1720 sizeof(struct ocfs2_extent_rec
));
1721 el
->l_next_free_rec
= cpu_to_le16(num_recs
);
1726 * Create an empty extent record .
1728 * l_next_free_rec may be updated.
1730 * If an empty extent already exists do nothing.
1732 static void ocfs2_create_empty_extent(struct ocfs2_extent_list
*el
)
1734 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1736 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
1741 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
1744 mlog_bug_on_msg(el
->l_count
== el
->l_next_free_rec
,
1745 "Asked to create an empty extent in a full list:\n"
1746 "count = %u, tree depth = %u",
1747 le16_to_cpu(el
->l_count
),
1748 le16_to_cpu(el
->l_tree_depth
));
1750 ocfs2_shift_records_right(el
);
1753 le16_add_cpu(&el
->l_next_free_rec
, 1);
1754 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1758 * For a rotation which involves two leaf nodes, the "root node" is
1759 * the lowest level tree node which contains a path to both leafs. This
1760 * resulting set of information can be used to form a complete "subtree"
1762 * This function is passed two full paths from the dinode down to a
1763 * pair of adjacent leaves. It's task is to figure out which path
1764 * index contains the subtree root - this can be the root index itself
1765 * in a worst-case rotation.
1767 * The array index of the subtree root is passed back.
1769 int ocfs2_find_subtree_root(struct ocfs2_extent_tree
*et
,
1770 struct ocfs2_path
*left
,
1771 struct ocfs2_path
*right
)
1776 * Check that the caller passed in two paths from the same tree.
1778 BUG_ON(path_root_bh(left
) != path_root_bh(right
));
1784 * The caller didn't pass two adjacent paths.
1786 mlog_bug_on_msg(i
> left
->p_tree_depth
,
1787 "Owner %llu, left depth %u, right depth %u\n"
1788 "left leaf blk %llu, right leaf blk %llu\n",
1789 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
1790 left
->p_tree_depth
, right
->p_tree_depth
,
1791 (unsigned long long)path_leaf_bh(left
)->b_blocknr
,
1792 (unsigned long long)path_leaf_bh(right
)->b_blocknr
);
1793 } while (left
->p_node
[i
].bh
->b_blocknr
==
1794 right
->p_node
[i
].bh
->b_blocknr
);
1799 typedef void (path_insert_t
)(void *, struct buffer_head
*);
1802 * Traverse a btree path in search of cpos, starting at root_el.
1804 * This code can be called with a cpos larger than the tree, in which
1805 * case it will return the rightmost path.
1807 static int __ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1808 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1809 path_insert_t
*func
, void *data
)
1814 struct buffer_head
*bh
= NULL
;
1815 struct ocfs2_extent_block
*eb
;
1816 struct ocfs2_extent_list
*el
;
1817 struct ocfs2_extent_rec
*rec
;
1820 while (el
->l_tree_depth
) {
1821 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1822 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1823 "Owner %llu has empty extent list at "
1825 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1826 le16_to_cpu(el
->l_tree_depth
));
1832 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
) - 1; i
++) {
1833 rec
= &el
->l_recs
[i
];
1836 * In the case that cpos is off the allocation
1837 * tree, this should just wind up returning the
1840 range
= le32_to_cpu(rec
->e_cpos
) +
1841 ocfs2_rec_clusters(el
, rec
);
1842 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1846 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1848 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1849 "Owner %llu has bad blkno in extent list "
1850 "at depth %u (index %d)\n",
1851 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1852 le16_to_cpu(el
->l_tree_depth
), i
);
1859 ret
= ocfs2_read_extent_block(ci
, blkno
, &bh
);
1865 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1868 if (le16_to_cpu(el
->l_next_free_rec
) >
1869 le16_to_cpu(el
->l_count
)) {
1870 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1871 "Owner %llu has bad count in extent list "
1872 "at block %llu (next free=%u, count=%u)\n",
1873 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1874 (unsigned long long)bh
->b_blocknr
,
1875 le16_to_cpu(el
->l_next_free_rec
),
1876 le16_to_cpu(el
->l_count
));
1887 * Catch any trailing bh that the loop didn't handle.
1895 * Given an initialized path (that is, it has a valid root extent
1896 * list), this function will traverse the btree in search of the path
1897 * which would contain cpos.
1899 * The path traveled is recorded in the path structure.
1901 * Note that this will not do any comparisons on leaf node extent
1902 * records, so it will work fine in the case that we just added a tree
1905 struct find_path_data
{
1907 struct ocfs2_path
*path
;
1909 static void find_path_ins(void *data
, struct buffer_head
*bh
)
1911 struct find_path_data
*fp
= data
;
1914 ocfs2_path_insert_eb(fp
->path
, fp
->index
, bh
);
1917 int ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1918 struct ocfs2_path
*path
, u32 cpos
)
1920 struct find_path_data data
;
1924 return __ocfs2_find_path(ci
, path_root_el(path
), cpos
,
1925 find_path_ins
, &data
);
1928 static void find_leaf_ins(void *data
, struct buffer_head
*bh
)
1930 struct ocfs2_extent_block
*eb
=(struct ocfs2_extent_block
*)bh
->b_data
;
1931 struct ocfs2_extent_list
*el
= &eb
->h_list
;
1932 struct buffer_head
**ret
= data
;
1934 /* We want to retain only the leaf block. */
1935 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
1941 * Find the leaf block in the tree which would contain cpos. No
1942 * checking of the actual leaf is done.
1944 * Some paths want to call this instead of allocating a path structure
1945 * and calling ocfs2_find_path().
1947 * This function doesn't handle non btree extent lists.
1949 int ocfs2_find_leaf(struct ocfs2_caching_info
*ci
,
1950 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1951 struct buffer_head
**leaf_bh
)
1954 struct buffer_head
*bh
= NULL
;
1956 ret
= __ocfs2_find_path(ci
, root_el
, cpos
, find_leaf_ins
, &bh
);
1968 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1970 * Basically, we've moved stuff around at the bottom of the tree and
1971 * we need to fix up the extent records above the changes to reflect
1974 * left_rec: the record on the left.
1975 * left_child_el: is the child list pointed to by left_rec
1976 * right_rec: the record to the right of left_rec
1977 * right_child_el: is the child list pointed to by right_rec
1979 * By definition, this only works on interior nodes.
1981 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec
*left_rec
,
1982 struct ocfs2_extent_list
*left_child_el
,
1983 struct ocfs2_extent_rec
*right_rec
,
1984 struct ocfs2_extent_list
*right_child_el
)
1986 u32 left_clusters
, right_end
;
1989 * Interior nodes never have holes. Their cpos is the cpos of
1990 * the leftmost record in their child list. Their cluster
1991 * count covers the full theoretical range of their child list
1992 * - the range between their cpos and the cpos of the record
1993 * immediately to their right.
1995 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[0].e_cpos
);
1996 if (!ocfs2_rec_clusters(right_child_el
, &right_child_el
->l_recs
[0])) {
1997 BUG_ON(right_child_el
->l_tree_depth
);
1998 BUG_ON(le16_to_cpu(right_child_el
->l_next_free_rec
) <= 1);
1999 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[1].e_cpos
);
2001 left_clusters
-= le32_to_cpu(left_rec
->e_cpos
);
2002 left_rec
->e_int_clusters
= cpu_to_le32(left_clusters
);
2005 * Calculate the rightmost cluster count boundary before
2006 * moving cpos - we will need to adjust clusters after
2007 * updating e_cpos to keep the same highest cluster count.
2009 right_end
= le32_to_cpu(right_rec
->e_cpos
);
2010 right_end
+= le32_to_cpu(right_rec
->e_int_clusters
);
2012 right_rec
->e_cpos
= left_rec
->e_cpos
;
2013 le32_add_cpu(&right_rec
->e_cpos
, left_clusters
);
2015 right_end
-= le32_to_cpu(right_rec
->e_cpos
);
2016 right_rec
->e_int_clusters
= cpu_to_le32(right_end
);
2020 * Adjust the adjacent root node records involved in a
2021 * rotation. left_el_blkno is passed in as a key so that we can easily
2022 * find it's index in the root list.
2024 static void ocfs2_adjust_root_records(struct ocfs2_extent_list
*root_el
,
2025 struct ocfs2_extent_list
*left_el
,
2026 struct ocfs2_extent_list
*right_el
,
2031 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) <=
2032 le16_to_cpu(left_el
->l_tree_depth
));
2034 for(i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
) - 1; i
++) {
2035 if (le64_to_cpu(root_el
->l_recs
[i
].e_blkno
) == left_el_blkno
)
2040 * The path walking code should have never returned a root and
2041 * two paths which are not adjacent.
2043 BUG_ON(i
>= (le16_to_cpu(root_el
->l_next_free_rec
) - 1));
2045 ocfs2_adjust_adjacent_records(&root_el
->l_recs
[i
], left_el
,
2046 &root_el
->l_recs
[i
+ 1], right_el
);
2050 * We've changed a leaf block (in right_path) and need to reflect that
2051 * change back up the subtree.
2053 * This happens in multiple places:
2054 * - When we've moved an extent record from the left path leaf to the right
2055 * path leaf to make room for an empty extent in the left path leaf.
2056 * - When our insert into the right path leaf is at the leftmost edge
2057 * and requires an update of the path immediately to it's left. This
2058 * can occur at the end of some types of rotation and appending inserts.
2059 * - When we've adjusted the last extent record in the left path leaf and the
2060 * 1st extent record in the right path leaf during cross extent block merge.
2062 static void ocfs2_complete_edge_insert(handle_t
*handle
,
2063 struct ocfs2_path
*left_path
,
2064 struct ocfs2_path
*right_path
,
2068 struct ocfs2_extent_list
*el
, *left_el
, *right_el
;
2069 struct ocfs2_extent_rec
*left_rec
, *right_rec
;
2070 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2073 * Update the counts and position values within all the
2074 * interior nodes to reflect the leaf rotation we just did.
2076 * The root node is handled below the loop.
2078 * We begin the loop with right_el and left_el pointing to the
2079 * leaf lists and work our way up.
2081 * NOTE: within this loop, left_el and right_el always refer
2082 * to the *child* lists.
2084 left_el
= path_leaf_el(left_path
);
2085 right_el
= path_leaf_el(right_path
);
2086 for(i
= left_path
->p_tree_depth
- 1; i
> subtree_index
; i
--) {
2087 mlog(0, "Adjust records at index %u\n", i
);
2090 * One nice property of knowing that all of these
2091 * nodes are below the root is that we only deal with
2092 * the leftmost right node record and the rightmost
2095 el
= left_path
->p_node
[i
].el
;
2096 idx
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2097 left_rec
= &el
->l_recs
[idx
];
2099 el
= right_path
->p_node
[i
].el
;
2100 right_rec
= &el
->l_recs
[0];
2102 ocfs2_adjust_adjacent_records(left_rec
, left_el
, right_rec
,
2105 ret
= ocfs2_journal_dirty(handle
, left_path
->p_node
[i
].bh
);
2109 ret
= ocfs2_journal_dirty(handle
, right_path
->p_node
[i
].bh
);
2114 * Setup our list pointers now so that the current
2115 * parents become children in the next iteration.
2117 left_el
= left_path
->p_node
[i
].el
;
2118 right_el
= right_path
->p_node
[i
].el
;
2122 * At the root node, adjust the two adjacent records which
2123 * begin our path to the leaves.
2126 el
= left_path
->p_node
[subtree_index
].el
;
2127 left_el
= left_path
->p_node
[subtree_index
+ 1].el
;
2128 right_el
= right_path
->p_node
[subtree_index
+ 1].el
;
2130 ocfs2_adjust_root_records(el
, left_el
, right_el
,
2131 left_path
->p_node
[subtree_index
+ 1].bh
->b_blocknr
);
2133 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2135 ret
= ocfs2_journal_dirty(handle
, root_bh
);
2140 static int ocfs2_rotate_subtree_right(handle_t
*handle
,
2141 struct ocfs2_extent_tree
*et
,
2142 struct ocfs2_path
*left_path
,
2143 struct ocfs2_path
*right_path
,
2147 struct buffer_head
*right_leaf_bh
;
2148 struct buffer_head
*left_leaf_bh
= NULL
;
2149 struct buffer_head
*root_bh
;
2150 struct ocfs2_extent_list
*right_el
, *left_el
;
2151 struct ocfs2_extent_rec move_rec
;
2153 left_leaf_bh
= path_leaf_bh(left_path
);
2154 left_el
= path_leaf_el(left_path
);
2156 if (left_el
->l_next_free_rec
!= left_el
->l_count
) {
2157 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
2158 "Inode %llu has non-full interior leaf node %llu"
2160 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2161 (unsigned long long)left_leaf_bh
->b_blocknr
,
2162 le16_to_cpu(left_el
->l_next_free_rec
));
2167 * This extent block may already have an empty record, so we
2168 * return early if so.
2170 if (ocfs2_is_empty_extent(&left_el
->l_recs
[0]))
2173 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2174 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2176 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2183 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2184 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2191 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2199 right_leaf_bh
= path_leaf_bh(right_path
);
2200 right_el
= path_leaf_el(right_path
);
2202 /* This is a code error, not a disk corruption. */
2203 mlog_bug_on_msg(!right_el
->l_next_free_rec
, "Inode %llu: Rotate fails "
2204 "because rightmost leaf block %llu is empty\n",
2205 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2206 (unsigned long long)right_leaf_bh
->b_blocknr
);
2208 ocfs2_create_empty_extent(right_el
);
2210 ret
= ocfs2_journal_dirty(handle
, right_leaf_bh
);
2216 /* Do the copy now. */
2217 i
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2218 move_rec
= left_el
->l_recs
[i
];
2219 right_el
->l_recs
[0] = move_rec
;
2222 * Clear out the record we just copied and shift everything
2223 * over, leaving an empty extent in the left leaf.
2225 * We temporarily subtract from next_free_rec so that the
2226 * shift will lose the tail record (which is now defunct).
2228 le16_add_cpu(&left_el
->l_next_free_rec
, -1);
2229 ocfs2_shift_records_right(left_el
);
2230 memset(&left_el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2231 le16_add_cpu(&left_el
->l_next_free_rec
, 1);
2233 ret
= ocfs2_journal_dirty(handle
, left_leaf_bh
);
2239 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2247 * Given a full path, determine what cpos value would return us a path
2248 * containing the leaf immediately to the left of the current one.
2250 * Will return zero if the path passed in is already the leftmost path.
2252 static int ocfs2_find_cpos_for_left_leaf(struct super_block
*sb
,
2253 struct ocfs2_path
*path
, u32
*cpos
)
2257 struct ocfs2_extent_list
*el
;
2259 BUG_ON(path
->p_tree_depth
== 0);
2263 blkno
= path_leaf_bh(path
)->b_blocknr
;
2265 /* Start at the tree node just above the leaf and work our way up. */
2266 i
= path
->p_tree_depth
- 1;
2268 el
= path
->p_node
[i
].el
;
2271 * Find the extent record just before the one in our
2274 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2275 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2279 * We've determined that the
2280 * path specified is already
2281 * the leftmost one - return a
2287 * The leftmost record points to our
2288 * leaf - we need to travel up the
2294 *cpos
= le32_to_cpu(el
->l_recs
[j
- 1].e_cpos
);
2295 *cpos
= *cpos
+ ocfs2_rec_clusters(el
,
2296 &el
->l_recs
[j
- 1]);
2303 * If we got here, we never found a valid node where
2304 * the tree indicated one should be.
2307 "Invalid extent tree at extent block %llu\n",
2308 (unsigned long long)blkno
);
2313 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2322 * Extend the transaction by enough credits to complete the rotation,
2323 * and still leave at least the original number of credits allocated
2324 * to this transaction.
2326 static int ocfs2_extend_rotate_transaction(handle_t
*handle
, int subtree_depth
,
2328 struct ocfs2_path
*path
)
2331 int credits
= (path
->p_tree_depth
- subtree_depth
) * 2 + 1 + op_credits
;
2333 if (handle
->h_buffer_credits
< credits
) {
2334 ret
= ocfs2_extend_trans(handle
,
2335 credits
- handle
->h_buffer_credits
);
2339 if (unlikely(handle
->h_buffer_credits
< credits
))
2340 return ocfs2_extend_trans(handle
, credits
);
2347 * Trap the case where we're inserting into the theoretical range past
2348 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2349 * whose cpos is less than ours into the right leaf.
2351 * It's only necessary to look at the rightmost record of the left
2352 * leaf because the logic that calls us should ensure that the
2353 * theoretical ranges in the path components above the leaves are
2356 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path
*left_path
,
2359 struct ocfs2_extent_list
*left_el
;
2360 struct ocfs2_extent_rec
*rec
;
2363 left_el
= path_leaf_el(left_path
);
2364 next_free
= le16_to_cpu(left_el
->l_next_free_rec
);
2365 rec
= &left_el
->l_recs
[next_free
- 1];
2367 if (insert_cpos
> le32_to_cpu(rec
->e_cpos
))
2372 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list
*el
, u32 cpos
)
2374 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
2376 struct ocfs2_extent_rec
*rec
;
2381 rec
= &el
->l_recs
[0];
2382 if (ocfs2_is_empty_extent(rec
)) {
2386 rec
= &el
->l_recs
[1];
2389 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2390 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
2396 * Rotate all the records in a btree right one record, starting at insert_cpos.
2398 * The path to the rightmost leaf should be passed in.
2400 * The array is assumed to be large enough to hold an entire path (tree depth).
2402 * Upon successful return from this function:
2404 * - The 'right_path' array will contain a path to the leaf block
2405 * whose range contains e_cpos.
2406 * - That leaf block will have a single empty extent in list index 0.
2407 * - In the case that the rotation requires a post-insert update,
2408 * *ret_left_path will contain a valid path which can be passed to
2409 * ocfs2_insert_path().
2411 static int ocfs2_rotate_tree_right(handle_t
*handle
,
2412 struct ocfs2_extent_tree
*et
,
2413 enum ocfs2_split_type split
,
2415 struct ocfs2_path
*right_path
,
2416 struct ocfs2_path
**ret_left_path
)
2418 int ret
, start
, orig_credits
= handle
->h_buffer_credits
;
2420 struct ocfs2_path
*left_path
= NULL
;
2421 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2423 *ret_left_path
= NULL
;
2425 left_path
= ocfs2_new_path_from_path(right_path
);
2432 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2438 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos
, cpos
);
2441 * What we want to do here is:
2443 * 1) Start with the rightmost path.
2445 * 2) Determine a path to the leaf block directly to the left
2448 * 3) Determine the 'subtree root' - the lowest level tree node
2449 * which contains a path to both leaves.
2451 * 4) Rotate the subtree.
2453 * 5) Find the next subtree by considering the left path to be
2454 * the new right path.
2456 * The check at the top of this while loop also accepts
2457 * insert_cpos == cpos because cpos is only a _theoretical_
2458 * value to get us the left path - insert_cpos might very well
2459 * be filling that hole.
2461 * Stop at a cpos of '0' because we either started at the
2462 * leftmost branch (i.e., a tree with one branch and a
2463 * rotation inside of it), or we've gone as far as we can in
2464 * rotating subtrees.
2466 while (cpos
&& insert_cpos
<= cpos
) {
2467 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2470 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
2476 mlog_bug_on_msg(path_leaf_bh(left_path
) ==
2477 path_leaf_bh(right_path
),
2478 "Owner %llu: error during insert of %u "
2479 "(left path cpos %u) results in two identical "
2480 "paths ending at %llu\n",
2481 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2483 (unsigned long long)
2484 path_leaf_bh(left_path
)->b_blocknr
);
2486 if (split
== SPLIT_NONE
&&
2487 ocfs2_rotate_requires_path_adjustment(left_path
,
2491 * We've rotated the tree as much as we
2492 * should. The rest is up to
2493 * ocfs2_insert_path() to complete, after the
2494 * record insertion. We indicate this
2495 * situation by returning the left path.
2497 * The reason we don't adjust the records here
2498 * before the record insert is that an error
2499 * later might break the rule where a parent
2500 * record e_cpos will reflect the actual
2501 * e_cpos of the 1st nonempty record of the
2504 *ret_left_path
= left_path
;
2508 start
= ocfs2_find_subtree_root(et
, left_path
, right_path
);
2510 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2512 (unsigned long long) right_path
->p_node
[start
].bh
->b_blocknr
,
2513 right_path
->p_tree_depth
);
2515 ret
= ocfs2_extend_rotate_transaction(handle
, start
,
2516 orig_credits
, right_path
);
2522 ret
= ocfs2_rotate_subtree_right(handle
, et
, left_path
,
2529 if (split
!= SPLIT_NONE
&&
2530 ocfs2_leftmost_rec_contains(path_leaf_el(right_path
),
2533 * A rotate moves the rightmost left leaf
2534 * record over to the leftmost right leaf
2535 * slot. If we're doing an extent split
2536 * instead of a real insert, then we have to
2537 * check that the extent to be split wasn't
2538 * just moved over. If it was, then we can
2539 * exit here, passing left_path back -
2540 * ocfs2_split_extent() is smart enough to
2541 * search both leaves.
2543 *ret_left_path
= left_path
;
2548 * There is no need to re-read the next right path
2549 * as we know that it'll be our current left
2550 * path. Optimize by copying values instead.
2552 ocfs2_mv_path(right_path
, left_path
);
2554 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2562 ocfs2_free_path(left_path
);
2568 static int ocfs2_update_edge_lengths(handle_t
*handle
,
2569 struct ocfs2_extent_tree
*et
,
2570 int subtree_index
, struct ocfs2_path
*path
)
2573 struct ocfs2_extent_rec
*rec
;
2574 struct ocfs2_extent_list
*el
;
2575 struct ocfs2_extent_block
*eb
;
2579 * In normal tree rotation process, we will never touch the
2580 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2581 * doesn't reserve the credits for them either.
2583 * But we do have a special case here which will update the rightmost
2584 * records for all the bh in the path.
2585 * So we have to allocate extra credits and access them.
2587 ret
= ocfs2_extend_trans(handle
,
2588 handle
->h_buffer_credits
+ subtree_index
);
2594 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
2600 /* Path should always be rightmost. */
2601 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2602 BUG_ON(eb
->h_next_leaf_blk
!= 0ULL);
2605 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
2606 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2607 rec
= &el
->l_recs
[idx
];
2608 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2610 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
2611 el
= path
->p_node
[i
].el
;
2612 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2613 rec
= &el
->l_recs
[idx
];
2615 rec
->e_int_clusters
= cpu_to_le32(range
);
2616 le32_add_cpu(&rec
->e_int_clusters
, -le32_to_cpu(rec
->e_cpos
));
2618 ocfs2_journal_dirty(handle
, path
->p_node
[i
].bh
);
2624 static void ocfs2_unlink_path(handle_t
*handle
,
2625 struct ocfs2_extent_tree
*et
,
2626 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2627 struct ocfs2_path
*path
, int unlink_start
)
2630 struct ocfs2_extent_block
*eb
;
2631 struct ocfs2_extent_list
*el
;
2632 struct buffer_head
*bh
;
2634 for(i
= unlink_start
; i
< path_num_items(path
); i
++) {
2635 bh
= path
->p_node
[i
].bh
;
2637 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
2639 * Not all nodes might have had their final count
2640 * decremented by the caller - handle this here.
2643 if (le16_to_cpu(el
->l_next_free_rec
) > 1) {
2645 "Inode %llu, attempted to remove extent block "
2646 "%llu with %u records\n",
2647 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2648 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
2649 le16_to_cpu(el
->l_next_free_rec
));
2651 ocfs2_journal_dirty(handle
, bh
);
2652 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2656 el
->l_next_free_rec
= 0;
2657 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2659 ocfs2_journal_dirty(handle
, bh
);
2661 ret
= ocfs2_cache_extent_block_free(dealloc
, eb
);
2665 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2669 static void ocfs2_unlink_subtree(handle_t
*handle
,
2670 struct ocfs2_extent_tree
*et
,
2671 struct ocfs2_path
*left_path
,
2672 struct ocfs2_path
*right_path
,
2674 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2677 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2678 struct ocfs2_extent_list
*root_el
= left_path
->p_node
[subtree_index
].el
;
2679 struct ocfs2_extent_list
*el
;
2680 struct ocfs2_extent_block
*eb
;
2682 el
= path_leaf_el(left_path
);
2684 eb
= (struct ocfs2_extent_block
*)right_path
->p_node
[subtree_index
+ 1].bh
->b_data
;
2686 for(i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
2687 if (root_el
->l_recs
[i
].e_blkno
== eb
->h_blkno
)
2690 BUG_ON(i
>= le16_to_cpu(root_el
->l_next_free_rec
));
2692 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
2693 le16_add_cpu(&root_el
->l_next_free_rec
, -1);
2695 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2696 eb
->h_next_leaf_blk
= 0;
2698 ocfs2_journal_dirty(handle
, root_bh
);
2699 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2701 ocfs2_unlink_path(handle
, et
, dealloc
, right_path
,
2705 static int ocfs2_rotate_subtree_left(handle_t
*handle
,
2706 struct ocfs2_extent_tree
*et
,
2707 struct ocfs2_path
*left_path
,
2708 struct ocfs2_path
*right_path
,
2710 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2713 int ret
, i
, del_right_subtree
= 0, right_has_empty
= 0;
2714 struct buffer_head
*root_bh
, *et_root_bh
= path_root_bh(right_path
);
2715 struct ocfs2_extent_list
*right_leaf_el
, *left_leaf_el
;
2716 struct ocfs2_extent_block
*eb
;
2720 right_leaf_el
= path_leaf_el(right_path
);
2721 left_leaf_el
= path_leaf_el(left_path
);
2722 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2723 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2725 if (!ocfs2_is_empty_extent(&left_leaf_el
->l_recs
[0]))
2728 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(right_path
)->b_data
;
2729 if (ocfs2_is_empty_extent(&right_leaf_el
->l_recs
[0])) {
2731 * It's legal for us to proceed if the right leaf is
2732 * the rightmost one and it has an empty extent. There
2733 * are two cases to handle - whether the leaf will be
2734 * empty after removal or not. If the leaf isn't empty
2735 * then just remove the empty extent up front. The
2736 * next block will handle empty leaves by flagging
2739 * Non rightmost leaves will throw -EAGAIN and the
2740 * caller can manually move the subtree and retry.
2743 if (eb
->h_next_leaf_blk
!= 0ULL)
2746 if (le16_to_cpu(right_leaf_el
->l_next_free_rec
) > 1) {
2747 ret
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
2748 path_leaf_bh(right_path
),
2749 OCFS2_JOURNAL_ACCESS_WRITE
);
2755 ocfs2_remove_empty_extent(right_leaf_el
);
2757 right_has_empty
= 1;
2760 if (eb
->h_next_leaf_blk
== 0ULL &&
2761 le16_to_cpu(right_leaf_el
->l_next_free_rec
) == 1) {
2763 * We have to update i_last_eb_blk during the meta
2766 ret
= ocfs2_et_root_journal_access(handle
, et
,
2767 OCFS2_JOURNAL_ACCESS_WRITE
);
2773 del_right_subtree
= 1;
2777 * Getting here with an empty extent in the right path implies
2778 * that it's the rightmost path and will be deleted.
2780 BUG_ON(right_has_empty
&& !del_right_subtree
);
2782 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2789 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2790 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2797 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2805 if (!right_has_empty
) {
2807 * Only do this if we're moving a real
2808 * record. Otherwise, the action is delayed until
2809 * after removal of the right path in which case we
2810 * can do a simple shift to remove the empty extent.
2812 ocfs2_rotate_leaf(left_leaf_el
, &right_leaf_el
->l_recs
[0]);
2813 memset(&right_leaf_el
->l_recs
[0], 0,
2814 sizeof(struct ocfs2_extent_rec
));
2816 if (eb
->h_next_leaf_blk
== 0ULL) {
2818 * Move recs over to get rid of empty extent, decrease
2819 * next_free. This is allowed to remove the last
2820 * extent in our leaf (setting l_next_free_rec to
2821 * zero) - the delete code below won't care.
2823 ocfs2_remove_empty_extent(right_leaf_el
);
2826 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2829 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2833 if (del_right_subtree
) {
2834 ocfs2_unlink_subtree(handle
, et
, left_path
, right_path
,
2835 subtree_index
, dealloc
);
2836 ret
= ocfs2_update_edge_lengths(handle
, et
, subtree_index
,
2843 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2844 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
2847 * Removal of the extent in the left leaf was skipped
2848 * above so we could delete the right path
2851 if (right_has_empty
)
2852 ocfs2_remove_empty_extent(left_leaf_el
);
2854 ret
= ocfs2_journal_dirty(handle
, et_root_bh
);
2860 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2868 * Given a full path, determine what cpos value would return us a path
2869 * containing the leaf immediately to the right of the current one.
2871 * Will return zero if the path passed in is already the rightmost path.
2873 * This looks similar, but is subtly different to
2874 * ocfs2_find_cpos_for_left_leaf().
2876 int ocfs2_find_cpos_for_right_leaf(struct super_block
*sb
,
2877 struct ocfs2_path
*path
, u32
*cpos
)
2881 struct ocfs2_extent_list
*el
;
2885 if (path
->p_tree_depth
== 0)
2888 blkno
= path_leaf_bh(path
)->b_blocknr
;
2890 /* Start at the tree node just above the leaf and work our way up. */
2891 i
= path
->p_tree_depth
- 1;
2895 el
= path
->p_node
[i
].el
;
2898 * Find the extent record just after the one in our
2901 next_free
= le16_to_cpu(el
->l_next_free_rec
);
2902 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2903 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2904 if (j
== (next_free
- 1)) {
2907 * We've determined that the
2908 * path specified is already
2909 * the rightmost one - return a
2915 * The rightmost record points to our
2916 * leaf - we need to travel up the
2922 *cpos
= le32_to_cpu(el
->l_recs
[j
+ 1].e_cpos
);
2928 * If we got here, we never found a valid node where
2929 * the tree indicated one should be.
2932 "Invalid extent tree at extent block %llu\n",
2933 (unsigned long long)blkno
);
2938 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2946 static int ocfs2_rotate_rightmost_leaf_left(handle_t
*handle
,
2947 struct ocfs2_extent_tree
*et
,
2948 struct ocfs2_path
*path
)
2951 struct buffer_head
*bh
= path_leaf_bh(path
);
2952 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
2954 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2957 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
2958 path_num_items(path
) - 1);
2964 ocfs2_remove_empty_extent(el
);
2966 ret
= ocfs2_journal_dirty(handle
, bh
);
2974 static int __ocfs2_rotate_tree_left(handle_t
*handle
,
2975 struct ocfs2_extent_tree
*et
,
2977 struct ocfs2_path
*path
,
2978 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2979 struct ocfs2_path
**empty_extent_path
)
2981 int ret
, subtree_root
, deleted
;
2983 struct ocfs2_path
*left_path
= NULL
;
2984 struct ocfs2_path
*right_path
= NULL
;
2985 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2987 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path
)->l_recs
[0])));
2989 *empty_extent_path
= NULL
;
2991 ret
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
2997 left_path
= ocfs2_new_path_from_path(path
);
3004 ocfs2_cp_path(left_path
, path
);
3006 right_path
= ocfs2_new_path_from_path(path
);
3013 while (right_cpos
) {
3014 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
3020 subtree_root
= ocfs2_find_subtree_root(et
, left_path
,
3023 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
3025 (unsigned long long)
3026 right_path
->p_node
[subtree_root
].bh
->b_blocknr
,
3027 right_path
->p_tree_depth
);
3029 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_root
,
3030 orig_credits
, left_path
);
3037 * Caller might still want to make changes to the
3038 * tree root, so re-add it to the journal here.
3040 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3047 ret
= ocfs2_rotate_subtree_left(handle
, et
, left_path
,
3048 right_path
, subtree_root
,
3050 if (ret
== -EAGAIN
) {
3052 * The rotation has to temporarily stop due to
3053 * the right subtree having an empty
3054 * extent. Pass it back to the caller for a
3057 *empty_extent_path
= right_path
;
3067 * The subtree rotate might have removed records on
3068 * the rightmost edge. If so, then rotation is
3074 ocfs2_mv_path(left_path
, right_path
);
3076 ret
= ocfs2_find_cpos_for_right_leaf(sb
, left_path
,
3085 ocfs2_free_path(right_path
);
3086 ocfs2_free_path(left_path
);
3091 static int ocfs2_remove_rightmost_path(handle_t
*handle
,
3092 struct ocfs2_extent_tree
*et
,
3093 struct ocfs2_path
*path
,
3094 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3096 int ret
, subtree_index
;
3098 struct ocfs2_path
*left_path
= NULL
;
3099 struct ocfs2_extent_block
*eb
;
3100 struct ocfs2_extent_list
*el
;
3103 ret
= ocfs2_et_sanity_check(et
);
3107 * There's two ways we handle this depending on
3108 * whether path is the only existing one.
3110 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3111 handle
->h_buffer_credits
,
3118 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
3124 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3133 * We have a path to the left of this one - it needs
3136 left_path
= ocfs2_new_path_from_path(path
);
3143 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
3149 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
3155 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
3157 ocfs2_unlink_subtree(handle
, et
, left_path
, path
,
3158 subtree_index
, dealloc
);
3159 ret
= ocfs2_update_edge_lengths(handle
, et
, subtree_index
,
3166 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
3167 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
3170 * 'path' is also the leftmost path which
3171 * means it must be the only one. This gets
3172 * handled differently because we want to
3173 * revert the root back to having extents
3176 ocfs2_unlink_path(handle
, et
, dealloc
, path
, 1);
3178 el
= et
->et_root_el
;
3179 el
->l_tree_depth
= 0;
3180 el
->l_next_free_rec
= 0;
3181 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3183 ocfs2_et_set_last_eb_blk(et
, 0);
3186 ocfs2_journal_dirty(handle
, path_root_bh(path
));
3189 ocfs2_free_path(left_path
);
3194 * Left rotation of btree records.
3196 * In many ways, this is (unsurprisingly) the opposite of right
3197 * rotation. We start at some non-rightmost path containing an empty
3198 * extent in the leaf block. The code works its way to the rightmost
3199 * path by rotating records to the left in every subtree.
3201 * This is used by any code which reduces the number of extent records
3202 * in a leaf. After removal, an empty record should be placed in the
3203 * leftmost list position.
3205 * This won't handle a length update of the rightmost path records if
3206 * the rightmost tree leaf record is removed so the caller is
3207 * responsible for detecting and correcting that.
3209 static int ocfs2_rotate_tree_left(handle_t
*handle
,
3210 struct ocfs2_extent_tree
*et
,
3211 struct ocfs2_path
*path
,
3212 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3214 int ret
, orig_credits
= handle
->h_buffer_credits
;
3215 struct ocfs2_path
*tmp_path
= NULL
, *restart_path
= NULL
;
3216 struct ocfs2_extent_block
*eb
;
3217 struct ocfs2_extent_list
*el
;
3219 el
= path_leaf_el(path
);
3220 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
3223 if (path
->p_tree_depth
== 0) {
3224 rightmost_no_delete
:
3226 * Inline extents. This is trivially handled, so do
3229 ret
= ocfs2_rotate_rightmost_leaf_left(handle
, et
, path
);
3236 * Handle rightmost branch now. There's several cases:
3237 * 1) simple rotation leaving records in there. That's trivial.
3238 * 2) rotation requiring a branch delete - there's no more
3239 * records left. Two cases of this:
3240 * a) There are branches to the left.
3241 * b) This is also the leftmost (the only) branch.
3243 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3244 * 2a) we need the left branch so that we can update it with the unlink
3245 * 2b) we need to bring the root back to inline extents.
3248 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
3250 if (eb
->h_next_leaf_blk
== 0) {
3252 * This gets a bit tricky if we're going to delete the
3253 * rightmost path. Get the other cases out of the way
3256 if (le16_to_cpu(el
->l_next_free_rec
) > 1)
3257 goto rightmost_no_delete
;
3259 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
3261 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
3262 "Owner %llu has empty extent block at %llu",
3263 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
3264 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
3269 * XXX: The caller can not trust "path" any more after
3270 * this as it will have been deleted. What do we do?
3272 * In theory the rotate-for-merge code will never get
3273 * here because it'll always ask for a rotate in a
3277 ret
= ocfs2_remove_rightmost_path(handle
, et
, path
,
3285 * Now we can loop, remembering the path we get from -EAGAIN
3286 * and restarting from there.
3289 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
, path
,
3290 dealloc
, &restart_path
);
3291 if (ret
&& ret
!= -EAGAIN
) {
3296 while (ret
== -EAGAIN
) {
3297 tmp_path
= restart_path
;
3298 restart_path
= NULL
;
3300 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
,
3303 if (ret
&& ret
!= -EAGAIN
) {
3308 ocfs2_free_path(tmp_path
);
3316 ocfs2_free_path(tmp_path
);
3317 ocfs2_free_path(restart_path
);
3321 static void ocfs2_cleanup_merge(struct ocfs2_extent_list
*el
,
3324 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[index
];
3327 if (rec
->e_leaf_clusters
== 0) {
3329 * We consumed all of the merged-from record. An empty
3330 * extent cannot exist anywhere but the 1st array
3331 * position, so move things over if the merged-from
3332 * record doesn't occupy that position.
3334 * This creates a new empty extent so the caller
3335 * should be smart enough to have removed any existing
3339 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3340 size
= index
* sizeof(struct ocfs2_extent_rec
);
3341 memmove(&el
->l_recs
[1], &el
->l_recs
[0], size
);
3345 * Always memset - the caller doesn't check whether it
3346 * created an empty extent, so there could be junk in
3349 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3353 static int ocfs2_get_right_path(struct ocfs2_extent_tree
*et
,
3354 struct ocfs2_path
*left_path
,
3355 struct ocfs2_path
**ret_right_path
)
3359 struct ocfs2_path
*right_path
= NULL
;
3360 struct ocfs2_extent_list
*left_el
;
3362 *ret_right_path
= NULL
;
3364 /* This function shouldn't be called for non-trees. */
3365 BUG_ON(left_path
->p_tree_depth
== 0);
3367 left_el
= path_leaf_el(left_path
);
3368 BUG_ON(left_el
->l_next_free_rec
!= left_el
->l_count
);
3370 ret
= ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3371 left_path
, &right_cpos
);
3377 /* This function shouldn't be called for the rightmost leaf. */
3378 BUG_ON(right_cpos
== 0);
3380 right_path
= ocfs2_new_path_from_path(left_path
);
3387 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
3393 *ret_right_path
= right_path
;
3396 ocfs2_free_path(right_path
);
3401 * Remove split_rec clusters from the record at index and merge them
3402 * onto the beginning of the record "next" to it.
3403 * For index < l_count - 1, the next means the extent rec at index + 1.
3404 * For index == l_count - 1, the "next" means the 1st extent rec of the
3405 * next extent block.
3407 static int ocfs2_merge_rec_right(struct ocfs2_path
*left_path
,
3409 struct ocfs2_extent_tree
*et
,
3410 struct ocfs2_extent_rec
*split_rec
,
3413 int ret
, next_free
, i
;
3414 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3415 struct ocfs2_extent_rec
*left_rec
;
3416 struct ocfs2_extent_rec
*right_rec
;
3417 struct ocfs2_extent_list
*right_el
;
3418 struct ocfs2_path
*right_path
= NULL
;
3419 int subtree_index
= 0;
3420 struct ocfs2_extent_list
*el
= path_leaf_el(left_path
);
3421 struct buffer_head
*bh
= path_leaf_bh(left_path
);
3422 struct buffer_head
*root_bh
= NULL
;
3424 BUG_ON(index
>= le16_to_cpu(el
->l_next_free_rec
));
3425 left_rec
= &el
->l_recs
[index
];
3427 if (index
== le16_to_cpu(el
->l_next_free_rec
) - 1 &&
3428 le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
)) {
3429 /* we meet with a cross extent block merge. */
3430 ret
= ocfs2_get_right_path(et
, left_path
, &right_path
);
3436 right_el
= path_leaf_el(right_path
);
3437 next_free
= le16_to_cpu(right_el
->l_next_free_rec
);
3438 BUG_ON(next_free
<= 0);
3439 right_rec
= &right_el
->l_recs
[0];
3440 if (ocfs2_is_empty_extent(right_rec
)) {
3441 BUG_ON(next_free
<= 1);
3442 right_rec
= &right_el
->l_recs
[1];
3445 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3446 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3447 le32_to_cpu(right_rec
->e_cpos
));
3449 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3452 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3453 handle
->h_buffer_credits
,
3460 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3461 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3463 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3470 for (i
= subtree_index
+ 1;
3471 i
< path_num_items(right_path
); i
++) {
3472 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3479 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3488 BUG_ON(index
== le16_to_cpu(el
->l_next_free_rec
) - 1);
3489 right_rec
= &el
->l_recs
[index
+ 1];
3492 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, left_path
,
3493 path_num_items(left_path
) - 1);
3499 le16_add_cpu(&left_rec
->e_leaf_clusters
, -split_clusters
);
3501 le32_add_cpu(&right_rec
->e_cpos
, -split_clusters
);
3502 le64_add_cpu(&right_rec
->e_blkno
,
3503 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3505 le16_add_cpu(&right_rec
->e_leaf_clusters
, split_clusters
);
3507 ocfs2_cleanup_merge(el
, index
);
3509 ret
= ocfs2_journal_dirty(handle
, bh
);
3514 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
3518 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
3523 ocfs2_free_path(right_path
);
3527 static int ocfs2_get_left_path(struct ocfs2_extent_tree
*et
,
3528 struct ocfs2_path
*right_path
,
3529 struct ocfs2_path
**ret_left_path
)
3533 struct ocfs2_path
*left_path
= NULL
;
3535 *ret_left_path
= NULL
;
3537 /* This function shouldn't be called for non-trees. */
3538 BUG_ON(right_path
->p_tree_depth
== 0);
3540 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3541 right_path
, &left_cpos
);
3547 /* This function shouldn't be called for the leftmost leaf. */
3548 BUG_ON(left_cpos
== 0);
3550 left_path
= ocfs2_new_path_from_path(right_path
);
3557 ret
= ocfs2_find_path(et
->et_ci
, left_path
, left_cpos
);
3563 *ret_left_path
= left_path
;
3566 ocfs2_free_path(left_path
);
3571 * Remove split_rec clusters from the record at index and merge them
3572 * onto the tail of the record "before" it.
3573 * For index > 0, the "before" means the extent rec at index - 1.
3575 * For index == 0, the "before" means the last record of the previous
3576 * extent block. And there is also a situation that we may need to
3577 * remove the rightmost leaf extent block in the right_path and change
3578 * the right path to indicate the new rightmost path.
3580 static int ocfs2_merge_rec_left(struct ocfs2_path
*right_path
,
3582 struct ocfs2_extent_tree
*et
,
3583 struct ocfs2_extent_rec
*split_rec
,
3584 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3587 int ret
, i
, subtree_index
= 0, has_empty_extent
= 0;
3588 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3589 struct ocfs2_extent_rec
*left_rec
;
3590 struct ocfs2_extent_rec
*right_rec
;
3591 struct ocfs2_extent_list
*el
= path_leaf_el(right_path
);
3592 struct buffer_head
*bh
= path_leaf_bh(right_path
);
3593 struct buffer_head
*root_bh
= NULL
;
3594 struct ocfs2_path
*left_path
= NULL
;
3595 struct ocfs2_extent_list
*left_el
;
3599 right_rec
= &el
->l_recs
[index
];
3601 /* we meet with a cross extent block merge. */
3602 ret
= ocfs2_get_left_path(et
, right_path
, &left_path
);
3608 left_el
= path_leaf_el(left_path
);
3609 BUG_ON(le16_to_cpu(left_el
->l_next_free_rec
) !=
3610 le16_to_cpu(left_el
->l_count
));
3612 left_rec
= &left_el
->l_recs
[
3613 le16_to_cpu(left_el
->l_next_free_rec
) - 1];
3614 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3615 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3616 le32_to_cpu(split_rec
->e_cpos
));
3618 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3621 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3622 handle
->h_buffer_credits
,
3629 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3630 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3632 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3639 for (i
= subtree_index
+ 1;
3640 i
< path_num_items(right_path
); i
++) {
3641 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3648 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3656 left_rec
= &el
->l_recs
[index
- 1];
3657 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
3658 has_empty_extent
= 1;
3661 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3662 path_num_items(right_path
) - 1);
3668 if (has_empty_extent
&& index
== 1) {
3670 * The easy case - we can just plop the record right in.
3672 *left_rec
= *split_rec
;
3674 has_empty_extent
= 0;
3676 le16_add_cpu(&left_rec
->e_leaf_clusters
, split_clusters
);
3678 le32_add_cpu(&right_rec
->e_cpos
, split_clusters
);
3679 le64_add_cpu(&right_rec
->e_blkno
,
3680 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3682 le16_add_cpu(&right_rec
->e_leaf_clusters
, -split_clusters
);
3684 ocfs2_cleanup_merge(el
, index
);
3686 ret
= ocfs2_journal_dirty(handle
, bh
);
3691 ret
= ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
3696 * In the situation that the right_rec is empty and the extent
3697 * block is empty also, ocfs2_complete_edge_insert can't handle
3698 * it and we need to delete the right extent block.
3700 if (le16_to_cpu(right_rec
->e_leaf_clusters
) == 0 &&
3701 le16_to_cpu(el
->l_next_free_rec
) == 1) {
3703 ret
= ocfs2_remove_rightmost_path(handle
, et
,
3711 /* Now the rightmost extent block has been deleted.
3712 * So we use the new rightmost path.
3714 ocfs2_mv_path(right_path
, left_path
);
3717 ocfs2_complete_edge_insert(handle
, left_path
,
3718 right_path
, subtree_index
);
3722 ocfs2_free_path(left_path
);
3726 static int ocfs2_try_to_merge_extent(handle_t
*handle
,
3727 struct ocfs2_extent_tree
*et
,
3728 struct ocfs2_path
*path
,
3730 struct ocfs2_extent_rec
*split_rec
,
3731 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3732 struct ocfs2_merge_ctxt
*ctxt
)
3735 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
3736 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
3738 BUG_ON(ctxt
->c_contig_type
== CONTIG_NONE
);
3740 if (ctxt
->c_split_covers_rec
&& ctxt
->c_has_empty_extent
) {
3742 * The merge code will need to create an empty
3743 * extent to take the place of the newly
3744 * emptied slot. Remove any pre-existing empty
3745 * extents - having more than one in a leaf is
3748 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3754 rec
= &el
->l_recs
[split_index
];
3757 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
) {
3759 * Left-right contig implies this.
3761 BUG_ON(!ctxt
->c_split_covers_rec
);
3764 * Since the leftright insert always covers the entire
3765 * extent, this call will delete the insert record
3766 * entirely, resulting in an empty extent record added to
3769 * Since the adding of an empty extent shifts
3770 * everything back to the right, there's no need to
3771 * update split_index here.
3773 * When the split_index is zero, we need to merge it to the
3774 * prevoius extent block. It is more efficient and easier
3775 * if we do merge_right first and merge_left later.
3777 ret
= ocfs2_merge_rec_right(path
, handle
, et
, split_rec
,
3785 * We can only get this from logic error above.
3787 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
3789 /* The merge left us with an empty extent, remove it. */
3790 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3796 rec
= &el
->l_recs
[split_index
];
3799 * Note that we don't pass split_rec here on purpose -
3800 * we've merged it into the rec already.
3802 ret
= ocfs2_merge_rec_left(path
, handle
, et
, rec
,
3803 dealloc
, split_index
);
3810 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3812 * Error from this last rotate is not critical, so
3813 * print but don't bubble it up.
3820 * Merge a record to the left or right.
3822 * 'contig_type' is relative to the existing record,
3823 * so for example, if we're "right contig", it's to
3824 * the record on the left (hence the left merge).
3826 if (ctxt
->c_contig_type
== CONTIG_RIGHT
) {
3827 ret
= ocfs2_merge_rec_left(path
, handle
, et
,
3835 ret
= ocfs2_merge_rec_right(path
, handle
,
3844 if (ctxt
->c_split_covers_rec
) {
3846 * The merge may have left an empty extent in
3847 * our leaf. Try to rotate it away.
3849 ret
= ocfs2_rotate_tree_left(handle
, et
, path
,
3861 static void ocfs2_subtract_from_rec(struct super_block
*sb
,
3862 enum ocfs2_split_type split
,
3863 struct ocfs2_extent_rec
*rec
,
3864 struct ocfs2_extent_rec
*split_rec
)
3868 len_blocks
= ocfs2_clusters_to_blocks(sb
,
3869 le16_to_cpu(split_rec
->e_leaf_clusters
));
3871 if (split
== SPLIT_LEFT
) {
3873 * Region is on the left edge of the existing
3876 le32_add_cpu(&rec
->e_cpos
,
3877 le16_to_cpu(split_rec
->e_leaf_clusters
));
3878 le64_add_cpu(&rec
->e_blkno
, len_blocks
);
3879 le16_add_cpu(&rec
->e_leaf_clusters
,
3880 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3883 * Region is on the right edge of the existing
3886 le16_add_cpu(&rec
->e_leaf_clusters
,
3887 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3892 * Do the final bits of extent record insertion at the target leaf
3893 * list. If this leaf is part of an allocation tree, it is assumed
3894 * that the tree above has been prepared.
3896 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree
*et
,
3897 struct ocfs2_extent_rec
*insert_rec
,
3898 struct ocfs2_extent_list
*el
,
3899 struct ocfs2_insert_type
*insert
)
3901 int i
= insert
->ins_contig_index
;
3903 struct ocfs2_extent_rec
*rec
;
3905 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3907 if (insert
->ins_split
!= SPLIT_NONE
) {
3908 i
= ocfs2_search_extent_list(el
, le32_to_cpu(insert_rec
->e_cpos
));
3910 rec
= &el
->l_recs
[i
];
3911 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
3912 insert
->ins_split
, rec
,
3918 * Contiguous insert - either left or right.
3920 if (insert
->ins_contig
!= CONTIG_NONE
) {
3921 rec
= &el
->l_recs
[i
];
3922 if (insert
->ins_contig
== CONTIG_LEFT
) {
3923 rec
->e_blkno
= insert_rec
->e_blkno
;
3924 rec
->e_cpos
= insert_rec
->e_cpos
;
3926 le16_add_cpu(&rec
->e_leaf_clusters
,
3927 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3932 * Handle insert into an empty leaf.
3934 if (le16_to_cpu(el
->l_next_free_rec
) == 0 ||
3935 ((le16_to_cpu(el
->l_next_free_rec
) == 1) &&
3936 ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3937 el
->l_recs
[0] = *insert_rec
;
3938 el
->l_next_free_rec
= cpu_to_le16(1);
3945 if (insert
->ins_appending
== APPEND_TAIL
) {
3946 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3947 rec
= &el
->l_recs
[i
];
3948 range
= le32_to_cpu(rec
->e_cpos
)
3949 + le16_to_cpu(rec
->e_leaf_clusters
);
3950 BUG_ON(le32_to_cpu(insert_rec
->e_cpos
) < range
);
3952 mlog_bug_on_msg(le16_to_cpu(el
->l_next_free_rec
) >=
3953 le16_to_cpu(el
->l_count
),
3954 "owner %llu, depth %u, count %u, next free %u, "
3955 "rec.cpos %u, rec.clusters %u, "
3956 "insert.cpos %u, insert.clusters %u\n",
3957 ocfs2_metadata_cache_owner(et
->et_ci
),
3958 le16_to_cpu(el
->l_tree_depth
),
3959 le16_to_cpu(el
->l_count
),
3960 le16_to_cpu(el
->l_next_free_rec
),
3961 le32_to_cpu(el
->l_recs
[i
].e_cpos
),
3962 le16_to_cpu(el
->l_recs
[i
].e_leaf_clusters
),
3963 le32_to_cpu(insert_rec
->e_cpos
),
3964 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3966 el
->l_recs
[i
] = *insert_rec
;
3967 le16_add_cpu(&el
->l_next_free_rec
, 1);
3973 * Ok, we have to rotate.
3975 * At this point, it is safe to assume that inserting into an
3976 * empty leaf and appending to a leaf have both been handled
3979 * This leaf needs to have space, either by the empty 1st
3980 * extent record, or by virtue of an l_next_rec < l_count.
3982 ocfs2_rotate_leaf(el
, insert_rec
);
3985 static void ocfs2_adjust_rightmost_records(handle_t
*handle
,
3986 struct ocfs2_extent_tree
*et
,
3987 struct ocfs2_path
*path
,
3988 struct ocfs2_extent_rec
*insert_rec
)
3990 int ret
, i
, next_free
;
3991 struct buffer_head
*bh
;
3992 struct ocfs2_extent_list
*el
;
3993 struct ocfs2_extent_rec
*rec
;
3996 * Update everything except the leaf block.
3998 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
3999 bh
= path
->p_node
[i
].bh
;
4000 el
= path
->p_node
[i
].el
;
4002 next_free
= le16_to_cpu(el
->l_next_free_rec
);
4003 if (next_free
== 0) {
4004 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
4005 "Owner %llu has a bad extent list",
4006 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
4011 rec
= &el
->l_recs
[next_free
- 1];
4013 rec
->e_int_clusters
= insert_rec
->e_cpos
;
4014 le32_add_cpu(&rec
->e_int_clusters
,
4015 le16_to_cpu(insert_rec
->e_leaf_clusters
));
4016 le32_add_cpu(&rec
->e_int_clusters
,
4017 -le32_to_cpu(rec
->e_cpos
));
4019 ret
= ocfs2_journal_dirty(handle
, bh
);
4026 static int ocfs2_append_rec_to_path(handle_t
*handle
,
4027 struct ocfs2_extent_tree
*et
,
4028 struct ocfs2_extent_rec
*insert_rec
,
4029 struct ocfs2_path
*right_path
,
4030 struct ocfs2_path
**ret_left_path
)
4033 struct ocfs2_extent_list
*el
;
4034 struct ocfs2_path
*left_path
= NULL
;
4036 *ret_left_path
= NULL
;
4039 * This shouldn't happen for non-trees. The extent rec cluster
4040 * count manipulation below only works for interior nodes.
4042 BUG_ON(right_path
->p_tree_depth
== 0);
4045 * If our appending insert is at the leftmost edge of a leaf,
4046 * then we might need to update the rightmost records of the
4049 el
= path_leaf_el(right_path
);
4050 next_free
= le16_to_cpu(el
->l_next_free_rec
);
4051 if (next_free
== 0 ||
4052 (next_free
== 1 && ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
4055 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
4056 right_path
, &left_cpos
);
4062 mlog(0, "Append may need a left path update. cpos: %u, "
4063 "left_cpos: %u\n", le32_to_cpu(insert_rec
->e_cpos
),
4067 * No need to worry if the append is already in the
4071 left_path
= ocfs2_new_path_from_path(right_path
);
4078 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
4086 * ocfs2_insert_path() will pass the left_path to the
4092 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4098 ocfs2_adjust_rightmost_records(handle
, et
, right_path
, insert_rec
);
4100 *ret_left_path
= left_path
;
4104 ocfs2_free_path(left_path
);
4109 static void ocfs2_split_record(struct ocfs2_extent_tree
*et
,
4110 struct ocfs2_path
*left_path
,
4111 struct ocfs2_path
*right_path
,
4112 struct ocfs2_extent_rec
*split_rec
,
4113 enum ocfs2_split_type split
)
4116 u32 cpos
= le32_to_cpu(split_rec
->e_cpos
);
4117 struct ocfs2_extent_list
*left_el
= NULL
, *right_el
, *insert_el
, *el
;
4118 struct ocfs2_extent_rec
*rec
, *tmprec
;
4120 right_el
= path_leaf_el(right_path
);
4122 left_el
= path_leaf_el(left_path
);
4125 insert_el
= right_el
;
4126 index
= ocfs2_search_extent_list(el
, cpos
);
4128 if (index
== 0 && left_path
) {
4129 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
4132 * This typically means that the record
4133 * started in the left path but moved to the
4134 * right as a result of rotation. We either
4135 * move the existing record to the left, or we
4136 * do the later insert there.
4138 * In this case, the left path should always
4139 * exist as the rotate code will have passed
4140 * it back for a post-insert update.
4143 if (split
== SPLIT_LEFT
) {
4145 * It's a left split. Since we know
4146 * that the rotate code gave us an
4147 * empty extent in the left path, we
4148 * can just do the insert there.
4150 insert_el
= left_el
;
4153 * Right split - we have to move the
4154 * existing record over to the left
4155 * leaf. The insert will be into the
4156 * newly created empty extent in the
4159 tmprec
= &right_el
->l_recs
[index
];
4160 ocfs2_rotate_leaf(left_el
, tmprec
);
4163 memset(tmprec
, 0, sizeof(*tmprec
));
4164 index
= ocfs2_search_extent_list(left_el
, cpos
);
4165 BUG_ON(index
== -1);
4170 BUG_ON(!ocfs2_is_empty_extent(&left_el
->l_recs
[0]));
4172 * Left path is easy - we can just allow the insert to
4176 insert_el
= left_el
;
4177 index
= ocfs2_search_extent_list(el
, cpos
);
4178 BUG_ON(index
== -1);
4181 rec
= &el
->l_recs
[index
];
4182 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
4183 split
, rec
, split_rec
);
4184 ocfs2_rotate_leaf(insert_el
, split_rec
);
4188 * This function only does inserts on an allocation b-tree. For tree
4189 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4191 * right_path is the path we want to do the actual insert
4192 * in. left_path should only be passed in if we need to update that
4193 * portion of the tree after an edge insert.
4195 static int ocfs2_insert_path(handle_t
*handle
,
4196 struct ocfs2_extent_tree
*et
,
4197 struct ocfs2_path
*left_path
,
4198 struct ocfs2_path
*right_path
,
4199 struct ocfs2_extent_rec
*insert_rec
,
4200 struct ocfs2_insert_type
*insert
)
4202 int ret
, subtree_index
;
4203 struct buffer_head
*leaf_bh
= path_leaf_bh(right_path
);
4206 int credits
= handle
->h_buffer_credits
;
4209 * There's a chance that left_path got passed back to
4210 * us without being accounted for in the
4211 * journal. Extend our transaction here to be sure we
4212 * can change those blocks.
4214 credits
+= left_path
->p_tree_depth
;
4216 ret
= ocfs2_extend_trans(handle
, credits
);
4222 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
4230 * Pass both paths to the journal. The majority of inserts
4231 * will be touching all components anyway.
4233 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4239 if (insert
->ins_split
!= SPLIT_NONE
) {
4241 * We could call ocfs2_insert_at_leaf() for some types
4242 * of splits, but it's easier to just let one separate
4243 * function sort it all out.
4245 ocfs2_split_record(et
, left_path
, right_path
,
4246 insert_rec
, insert
->ins_split
);
4249 * Split might have modified either leaf and we don't
4250 * have a guarantee that the later edge insert will
4251 * dirty this for us.
4254 ret
= ocfs2_journal_dirty(handle
,
4255 path_leaf_bh(left_path
));
4259 ocfs2_insert_at_leaf(et
, insert_rec
, path_leaf_el(right_path
),
4262 ret
= ocfs2_journal_dirty(handle
, leaf_bh
);
4268 * The rotate code has indicated that we need to fix
4269 * up portions of the tree after the insert.
4271 * XXX: Should we extend the transaction here?
4273 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
4275 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
4284 static int ocfs2_do_insert_extent(handle_t
*handle
,
4285 struct ocfs2_extent_tree
*et
,
4286 struct ocfs2_extent_rec
*insert_rec
,
4287 struct ocfs2_insert_type
*type
)
4289 int ret
, rotate
= 0;
4291 struct ocfs2_path
*right_path
= NULL
;
4292 struct ocfs2_path
*left_path
= NULL
;
4293 struct ocfs2_extent_list
*el
;
4295 el
= et
->et_root_el
;
4297 ret
= ocfs2_et_root_journal_access(handle
, et
,
4298 OCFS2_JOURNAL_ACCESS_WRITE
);
4304 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
4305 ocfs2_insert_at_leaf(et
, insert_rec
, el
, type
);
4306 goto out_update_clusters
;
4309 right_path
= ocfs2_new_path_from_et(et
);
4317 * Determine the path to start with. Rotations need the
4318 * rightmost path, everything else can go directly to the
4321 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4322 if (type
->ins_appending
== APPEND_NONE
&&
4323 type
->ins_contig
== CONTIG_NONE
) {
4328 ret
= ocfs2_find_path(et
->et_ci
, right_path
, cpos
);
4335 * Rotations and appends need special treatment - they modify
4336 * parts of the tree's above them.
4338 * Both might pass back a path immediate to the left of the
4339 * one being inserted to. This will be cause
4340 * ocfs2_insert_path() to modify the rightmost records of
4341 * left_path to account for an edge insert.
4343 * XXX: When modifying this code, keep in mind that an insert
4344 * can wind up skipping both of these two special cases...
4347 ret
= ocfs2_rotate_tree_right(handle
, et
, type
->ins_split
,
4348 le32_to_cpu(insert_rec
->e_cpos
),
4349 right_path
, &left_path
);
4356 * ocfs2_rotate_tree_right() might have extended the
4357 * transaction without re-journaling our tree root.
4359 ret
= ocfs2_et_root_journal_access(handle
, et
,
4360 OCFS2_JOURNAL_ACCESS_WRITE
);
4365 } else if (type
->ins_appending
== APPEND_TAIL
4366 && type
->ins_contig
!= CONTIG_LEFT
) {
4367 ret
= ocfs2_append_rec_to_path(handle
, et
, insert_rec
,
4368 right_path
, &left_path
);
4375 ret
= ocfs2_insert_path(handle
, et
, left_path
, right_path
,
4382 out_update_clusters
:
4383 if (type
->ins_split
== SPLIT_NONE
)
4384 ocfs2_et_update_clusters(et
,
4385 le16_to_cpu(insert_rec
->e_leaf_clusters
));
4387 ret
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4392 ocfs2_free_path(left_path
);
4393 ocfs2_free_path(right_path
);
4398 static enum ocfs2_contig_type
4399 ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree
*et
,
4400 struct ocfs2_path
*path
,
4401 struct ocfs2_extent_list
*el
, int index
,
4402 struct ocfs2_extent_rec
*split_rec
)
4405 enum ocfs2_contig_type ret
= CONTIG_NONE
;
4406 u32 left_cpos
, right_cpos
;
4407 struct ocfs2_extent_rec
*rec
= NULL
;
4408 struct ocfs2_extent_list
*new_el
;
4409 struct ocfs2_path
*left_path
= NULL
, *right_path
= NULL
;
4410 struct buffer_head
*bh
;
4411 struct ocfs2_extent_block
*eb
;
4412 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
4415 rec
= &el
->l_recs
[index
- 1];
4416 } else if (path
->p_tree_depth
> 0) {
4417 status
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
4421 if (left_cpos
!= 0) {
4422 left_path
= ocfs2_new_path_from_path(path
);
4426 status
= ocfs2_find_path(et
->et_ci
, left_path
,
4431 new_el
= path_leaf_el(left_path
);
4433 if (le16_to_cpu(new_el
->l_next_free_rec
) !=
4434 le16_to_cpu(new_el
->l_count
)) {
4435 bh
= path_leaf_bh(left_path
);
4436 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4438 "Extent block #%llu has an "
4439 "invalid l_next_free_rec of "
4440 "%d. It should have "
4441 "matched the l_count of %d",
4442 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4443 le16_to_cpu(new_el
->l_next_free_rec
),
4444 le16_to_cpu(new_el
->l_count
));
4448 rec
= &new_el
->l_recs
[
4449 le16_to_cpu(new_el
->l_next_free_rec
) - 1];
4454 * We're careful to check for an empty extent record here -
4455 * the merge code will know what to do if it sees one.
4458 if (index
== 1 && ocfs2_is_empty_extent(rec
)) {
4459 if (split_rec
->e_cpos
== el
->l_recs
[index
].e_cpos
)
4462 ret
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4467 if (index
< (le16_to_cpu(el
->l_next_free_rec
) - 1))
4468 rec
= &el
->l_recs
[index
+ 1];
4469 else if (le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
) &&
4470 path
->p_tree_depth
> 0) {
4471 status
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
4475 if (right_cpos
== 0)
4478 right_path
= ocfs2_new_path_from_path(path
);
4482 status
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
4486 new_el
= path_leaf_el(right_path
);
4487 rec
= &new_el
->l_recs
[0];
4488 if (ocfs2_is_empty_extent(rec
)) {
4489 if (le16_to_cpu(new_el
->l_next_free_rec
) <= 1) {
4490 bh
= path_leaf_bh(right_path
);
4491 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4493 "Extent block #%llu has an "
4494 "invalid l_next_free_rec of %d",
4495 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4496 le16_to_cpu(new_el
->l_next_free_rec
));
4500 rec
= &new_el
->l_recs
[1];
4505 enum ocfs2_contig_type contig_type
;
4507 contig_type
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4509 if (contig_type
== CONTIG_LEFT
&& ret
== CONTIG_RIGHT
)
4510 ret
= CONTIG_LEFTRIGHT
;
4511 else if (ret
== CONTIG_NONE
)
4517 ocfs2_free_path(left_path
);
4519 ocfs2_free_path(right_path
);
4524 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree
*et
,
4525 struct ocfs2_insert_type
*insert
,
4526 struct ocfs2_extent_list
*el
,
4527 struct ocfs2_extent_rec
*insert_rec
)
4530 enum ocfs2_contig_type contig_type
= CONTIG_NONE
;
4532 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4534 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
4535 contig_type
= ocfs2_et_extent_contig(et
, &el
->l_recs
[i
],
4537 if (contig_type
!= CONTIG_NONE
) {
4538 insert
->ins_contig_index
= i
;
4542 insert
->ins_contig
= contig_type
;
4544 if (insert
->ins_contig
!= CONTIG_NONE
) {
4545 struct ocfs2_extent_rec
*rec
=
4546 &el
->l_recs
[insert
->ins_contig_index
];
4547 unsigned int len
= le16_to_cpu(rec
->e_leaf_clusters
) +
4548 le16_to_cpu(insert_rec
->e_leaf_clusters
);
4551 * Caller might want us to limit the size of extents, don't
4552 * calculate contiguousness if we might exceed that limit.
4554 if (et
->et_max_leaf_clusters
&&
4555 (len
> et
->et_max_leaf_clusters
))
4556 insert
->ins_contig
= CONTIG_NONE
;
4561 * This should only be called against the righmost leaf extent list.
4563 * ocfs2_figure_appending_type() will figure out whether we'll have to
4564 * insert at the tail of the rightmost leaf.
4566 * This should also work against the root extent list for tree's with 0
4567 * depth. If we consider the root extent list to be the rightmost leaf node
4568 * then the logic here makes sense.
4570 static void ocfs2_figure_appending_type(struct ocfs2_insert_type
*insert
,
4571 struct ocfs2_extent_list
*el
,
4572 struct ocfs2_extent_rec
*insert_rec
)
4575 u32 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4576 struct ocfs2_extent_rec
*rec
;
4578 insert
->ins_appending
= APPEND_NONE
;
4580 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4582 if (!el
->l_next_free_rec
)
4583 goto set_tail_append
;
4585 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
4586 /* Were all records empty? */
4587 if (le16_to_cpu(el
->l_next_free_rec
) == 1)
4588 goto set_tail_append
;
4591 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
4592 rec
= &el
->l_recs
[i
];
4595 (le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)))
4596 goto set_tail_append
;
4601 insert
->ins_appending
= APPEND_TAIL
;
4605 * Helper function called at the begining of an insert.
4607 * This computes a few things that are commonly used in the process of
4608 * inserting into the btree:
4609 * - Whether the new extent is contiguous with an existing one.
4610 * - The current tree depth.
4611 * - Whether the insert is an appending one.
4612 * - The total # of free records in the tree.
4614 * All of the information is stored on the ocfs2_insert_type
4617 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree
*et
,
4618 struct buffer_head
**last_eb_bh
,
4619 struct ocfs2_extent_rec
*insert_rec
,
4621 struct ocfs2_insert_type
*insert
)
4624 struct ocfs2_extent_block
*eb
;
4625 struct ocfs2_extent_list
*el
;
4626 struct ocfs2_path
*path
= NULL
;
4627 struct buffer_head
*bh
= NULL
;
4629 insert
->ins_split
= SPLIT_NONE
;
4631 el
= et
->et_root_el
;
4632 insert
->ins_tree_depth
= le16_to_cpu(el
->l_tree_depth
);
4634 if (el
->l_tree_depth
) {
4636 * If we have tree depth, we read in the
4637 * rightmost extent block ahead of time as
4638 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4639 * may want it later.
4641 ret
= ocfs2_read_extent_block(et
->et_ci
,
4642 ocfs2_et_get_last_eb_blk(et
),
4648 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
4653 * Unless we have a contiguous insert, we'll need to know if
4654 * there is room left in our allocation tree for another
4657 * XXX: This test is simplistic, we can search for empty
4658 * extent records too.
4660 *free_records
= le16_to_cpu(el
->l_count
) -
4661 le16_to_cpu(el
->l_next_free_rec
);
4663 if (!insert
->ins_tree_depth
) {
4664 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4665 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4669 path
= ocfs2_new_path_from_et(et
);
4677 * In the case that we're inserting past what the tree
4678 * currently accounts for, ocfs2_find_path() will return for
4679 * us the rightmost tree path. This is accounted for below in
4680 * the appending code.
4682 ret
= ocfs2_find_path(et
->et_ci
, path
, le32_to_cpu(insert_rec
->e_cpos
));
4688 el
= path_leaf_el(path
);
4691 * Now that we have the path, there's two things we want to determine:
4692 * 1) Contiguousness (also set contig_index if this is so)
4694 * 2) Are we doing an append? We can trivially break this up
4695 * into two types of appends: simple record append, or a
4696 * rotate inside the tail leaf.
4698 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4701 * The insert code isn't quite ready to deal with all cases of
4702 * left contiguousness. Specifically, if it's an insert into
4703 * the 1st record in a leaf, it will require the adjustment of
4704 * cluster count on the last record of the path directly to it's
4705 * left. For now, just catch that case and fool the layers
4706 * above us. This works just fine for tree_depth == 0, which
4707 * is why we allow that above.
4709 if (insert
->ins_contig
== CONTIG_LEFT
&&
4710 insert
->ins_contig_index
== 0)
4711 insert
->ins_contig
= CONTIG_NONE
;
4714 * Ok, so we can simply compare against last_eb to figure out
4715 * whether the path doesn't exist. This will only happen in
4716 * the case that we're doing a tail append, so maybe we can
4717 * take advantage of that information somehow.
4719 if (ocfs2_et_get_last_eb_blk(et
) ==
4720 path_leaf_bh(path
)->b_blocknr
) {
4722 * Ok, ocfs2_find_path() returned us the rightmost
4723 * tree path. This might be an appending insert. There are
4725 * 1) We're doing a true append at the tail:
4726 * -This might even be off the end of the leaf
4727 * 2) We're "appending" by rotating in the tail
4729 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4733 ocfs2_free_path(path
);
4743 * Insert an extent into a btree.
4745 * The caller needs to update the owning btree's cluster count.
4747 int ocfs2_insert_extent(handle_t
*handle
,
4748 struct ocfs2_extent_tree
*et
,
4753 struct ocfs2_alloc_context
*meta_ac
)
4756 int uninitialized_var(free_records
);
4757 struct buffer_head
*last_eb_bh
= NULL
;
4758 struct ocfs2_insert_type insert
= {0, };
4759 struct ocfs2_extent_rec rec
;
4761 mlog(0, "add %u clusters at position %u to owner %llu\n",
4763 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
4765 memset(&rec
, 0, sizeof(rec
));
4766 rec
.e_cpos
= cpu_to_le32(cpos
);
4767 rec
.e_blkno
= cpu_to_le64(start_blk
);
4768 rec
.e_leaf_clusters
= cpu_to_le16(new_clusters
);
4769 rec
.e_flags
= flags
;
4770 status
= ocfs2_et_insert_check(et
, &rec
);
4776 status
= ocfs2_figure_insert_type(et
, &last_eb_bh
, &rec
,
4777 &free_records
, &insert
);
4783 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4784 "Insert.contig_index: %d, Insert.free_records: %d, "
4785 "Insert.tree_depth: %d\n",
4786 insert
.ins_appending
, insert
.ins_contig
, insert
.ins_contig_index
,
4787 free_records
, insert
.ins_tree_depth
);
4789 if (insert
.ins_contig
== CONTIG_NONE
&& free_records
== 0) {
4790 status
= ocfs2_grow_tree(handle
, et
,
4791 &insert
.ins_tree_depth
, &last_eb_bh
,
4799 /* Finally, we can add clusters. This might rotate the tree for us. */
4800 status
= ocfs2_do_insert_extent(handle
, et
, &rec
, &insert
);
4804 ocfs2_et_extent_map_insert(et
, &rec
);
4814 * Allcate and add clusters into the extent b-tree.
4815 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4816 * The extent b-tree's root is specified by et, and
4817 * it is not limited to the file storage. Any extent tree can use this
4818 * function if it implements the proper ocfs2_extent_tree.
4820 int ocfs2_add_clusters_in_btree(handle_t
*handle
,
4821 struct ocfs2_extent_tree
*et
,
4822 u32
*logical_offset
,
4823 u32 clusters_to_add
,
4825 struct ocfs2_alloc_context
*data_ac
,
4826 struct ocfs2_alloc_context
*meta_ac
,
4827 enum ocfs2_alloc_restarted
*reason_ret
)
4831 enum ocfs2_alloc_restarted reason
= RESTART_NONE
;
4832 u32 bit_off
, num_bits
;
4835 struct ocfs2_super
*osb
=
4836 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
4838 BUG_ON(!clusters_to_add
);
4841 flags
= OCFS2_EXT_UNWRITTEN
;
4843 free_extents
= ocfs2_num_free_extents(osb
, et
);
4844 if (free_extents
< 0) {
4845 status
= free_extents
;
4850 /* there are two cases which could cause us to EAGAIN in the
4851 * we-need-more-metadata case:
4852 * 1) we haven't reserved *any*
4853 * 2) we are so fragmented, we've needed to add metadata too
4855 if (!free_extents
&& !meta_ac
) {
4856 mlog(0, "we haven't reserved any metadata!\n");
4858 reason
= RESTART_META
;
4860 } else if ((!free_extents
)
4861 && (ocfs2_alloc_context_bits_left(meta_ac
)
4862 < ocfs2_extend_meta_needed(et
->et_root_el
))) {
4863 mlog(0, "filesystem is really fragmented...\n");
4865 reason
= RESTART_META
;
4869 status
= __ocfs2_claim_clusters(osb
, handle
, data_ac
, 1,
4870 clusters_to_add
, &bit_off
, &num_bits
);
4872 if (status
!= -ENOSPC
)
4877 BUG_ON(num_bits
> clusters_to_add
);
4879 /* reserve our write early -- insert_extent may update the tree root */
4880 status
= ocfs2_et_root_journal_access(handle
, et
,
4881 OCFS2_JOURNAL_ACCESS_WRITE
);
4887 block
= ocfs2_clusters_to_blocks(osb
->sb
, bit_off
);
4888 mlog(0, "Allocating %u clusters at block %u for owner %llu\n",
4890 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
4891 status
= ocfs2_insert_extent(handle
, et
, *logical_offset
, block
,
4892 num_bits
, flags
, meta_ac
);
4898 status
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4904 clusters_to_add
-= num_bits
;
4905 *logical_offset
+= num_bits
;
4907 if (clusters_to_add
) {
4908 mlog(0, "need to alloc once more, wanted = %u\n",
4911 reason
= RESTART_TRANS
;
4917 *reason_ret
= reason
;
4921 static void ocfs2_make_right_split_rec(struct super_block
*sb
,
4922 struct ocfs2_extent_rec
*split_rec
,
4924 struct ocfs2_extent_rec
*rec
)
4926 u32 rec_cpos
= le32_to_cpu(rec
->e_cpos
);
4927 u32 rec_range
= rec_cpos
+ le16_to_cpu(rec
->e_leaf_clusters
);
4929 memset(split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4931 split_rec
->e_cpos
= cpu_to_le32(cpos
);
4932 split_rec
->e_leaf_clusters
= cpu_to_le16(rec_range
- cpos
);
4934 split_rec
->e_blkno
= rec
->e_blkno
;
4935 le64_add_cpu(&split_rec
->e_blkno
,
4936 ocfs2_clusters_to_blocks(sb
, cpos
- rec_cpos
));
4938 split_rec
->e_flags
= rec
->e_flags
;
4941 static int ocfs2_split_and_insert(handle_t
*handle
,
4942 struct ocfs2_extent_tree
*et
,
4943 struct ocfs2_path
*path
,
4944 struct buffer_head
**last_eb_bh
,
4946 struct ocfs2_extent_rec
*orig_split_rec
,
4947 struct ocfs2_alloc_context
*meta_ac
)
4950 unsigned int insert_range
, rec_range
, do_leftright
= 0;
4951 struct ocfs2_extent_rec tmprec
;
4952 struct ocfs2_extent_list
*rightmost_el
;
4953 struct ocfs2_extent_rec rec
;
4954 struct ocfs2_extent_rec split_rec
= *orig_split_rec
;
4955 struct ocfs2_insert_type insert
;
4956 struct ocfs2_extent_block
*eb
;
4960 * Store a copy of the record on the stack - it might move
4961 * around as the tree is manipulated below.
4963 rec
= path_leaf_el(path
)->l_recs
[split_index
];
4965 rightmost_el
= et
->et_root_el
;
4967 depth
= le16_to_cpu(rightmost_el
->l_tree_depth
);
4969 BUG_ON(!(*last_eb_bh
));
4970 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
4971 rightmost_el
= &eb
->h_list
;
4974 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4975 le16_to_cpu(rightmost_el
->l_count
)) {
4976 ret
= ocfs2_grow_tree(handle
, et
,
4977 &depth
, last_eb_bh
, meta_ac
);
4984 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4985 insert
.ins_appending
= APPEND_NONE
;
4986 insert
.ins_contig
= CONTIG_NONE
;
4987 insert
.ins_tree_depth
= depth
;
4989 insert_range
= le32_to_cpu(split_rec
.e_cpos
) +
4990 le16_to_cpu(split_rec
.e_leaf_clusters
);
4991 rec_range
= le32_to_cpu(rec
.e_cpos
) +
4992 le16_to_cpu(rec
.e_leaf_clusters
);
4994 if (split_rec
.e_cpos
== rec
.e_cpos
) {
4995 insert
.ins_split
= SPLIT_LEFT
;
4996 } else if (insert_range
== rec_range
) {
4997 insert
.ins_split
= SPLIT_RIGHT
;
5000 * Left/right split. We fake this as a right split
5001 * first and then make a second pass as a left split.
5003 insert
.ins_split
= SPLIT_RIGHT
;
5005 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
5006 &tmprec
, insert_range
, &rec
);
5010 BUG_ON(do_leftright
);
5014 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
5020 if (do_leftright
== 1) {
5022 struct ocfs2_extent_list
*el
;
5025 split_rec
= *orig_split_rec
;
5027 ocfs2_reinit_path(path
, 1);
5029 cpos
= le32_to_cpu(split_rec
.e_cpos
);
5030 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5036 el
= path_leaf_el(path
);
5037 split_index
= ocfs2_search_extent_list(el
, cpos
);
5045 static int ocfs2_replace_extent_rec(handle_t
*handle
,
5046 struct ocfs2_extent_tree
*et
,
5047 struct ocfs2_path
*path
,
5048 struct ocfs2_extent_list
*el
,
5050 struct ocfs2_extent_rec
*split_rec
)
5054 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
5055 path_num_items(path
) - 1);
5061 el
->l_recs
[split_index
] = *split_rec
;
5063 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5069 * Split part or all of the extent record at split_index in the leaf
5070 * pointed to by path. Merge with the contiguous extent record if needed.
5072 * Care is taken to handle contiguousness so as to not grow the tree.
5074 * meta_ac is not strictly necessary - we only truly need it if growth
5075 * of the tree is required. All other cases will degrade into a less
5076 * optimal tree layout.
5078 * last_eb_bh should be the rightmost leaf block for any extent
5079 * btree. Since a split may grow the tree or a merge might shrink it,
5080 * the caller cannot trust the contents of that buffer after this call.
5082 * This code is optimized for readability - several passes might be
5083 * made over certain portions of the tree. All of those blocks will
5084 * have been brought into cache (and pinned via the journal), so the
5085 * extra overhead is not expressed in terms of disk reads.
5087 int ocfs2_split_extent(handle_t
*handle
,
5088 struct ocfs2_extent_tree
*et
,
5089 struct ocfs2_path
*path
,
5091 struct ocfs2_extent_rec
*split_rec
,
5092 struct ocfs2_alloc_context
*meta_ac
,
5093 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5096 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5097 struct buffer_head
*last_eb_bh
= NULL
;
5098 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
5099 struct ocfs2_merge_ctxt ctxt
;
5100 struct ocfs2_extent_list
*rightmost_el
;
5102 if (le32_to_cpu(rec
->e_cpos
) > le32_to_cpu(split_rec
->e_cpos
) ||
5103 ((le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)) <
5104 (le32_to_cpu(split_rec
->e_cpos
) + le16_to_cpu(split_rec
->e_leaf_clusters
)))) {
5110 ctxt
.c_contig_type
= ocfs2_figure_merge_contig_type(et
, path
, el
,
5115 * The core merge / split code wants to know how much room is
5116 * left in this allocation tree, so we pass the
5117 * rightmost extent list.
5119 if (path
->p_tree_depth
) {
5120 struct ocfs2_extent_block
*eb
;
5122 ret
= ocfs2_read_extent_block(et
->et_ci
,
5123 ocfs2_et_get_last_eb_blk(et
),
5130 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5131 rightmost_el
= &eb
->h_list
;
5133 rightmost_el
= path_root_el(path
);
5135 if (rec
->e_cpos
== split_rec
->e_cpos
&&
5136 rec
->e_leaf_clusters
== split_rec
->e_leaf_clusters
)
5137 ctxt
.c_split_covers_rec
= 1;
5139 ctxt
.c_split_covers_rec
= 0;
5141 ctxt
.c_has_empty_extent
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
5143 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
5144 split_index
, ctxt
.c_contig_type
, ctxt
.c_has_empty_extent
,
5145 ctxt
.c_split_covers_rec
);
5147 if (ctxt
.c_contig_type
== CONTIG_NONE
) {
5148 if (ctxt
.c_split_covers_rec
)
5149 ret
= ocfs2_replace_extent_rec(handle
, et
, path
, el
,
5150 split_index
, split_rec
);
5152 ret
= ocfs2_split_and_insert(handle
, et
, path
,
5153 &last_eb_bh
, split_index
,
5154 split_rec
, meta_ac
);
5158 ret
= ocfs2_try_to_merge_extent(handle
, et
, path
,
5159 split_index
, split_rec
,
5171 * Change the flags of the already-existing extent at cpos for len clusters.
5173 * new_flags: the flags we want to set.
5174 * clear_flags: the flags we want to clear.
5175 * phys: the new physical offset we want this new extent starts from.
5177 * If the existing extent is larger than the request, initiate a
5178 * split. An attempt will be made at merging with adjacent extents.
5180 * The caller is responsible for passing down meta_ac if we'll need it.
5182 int ocfs2_change_extent_flag(handle_t
*handle
,
5183 struct ocfs2_extent_tree
*et
,
5184 u32 cpos
, u32 len
, u32 phys
,
5185 struct ocfs2_alloc_context
*meta_ac
,
5186 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5187 int new_flags
, int clear_flags
)
5190 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5191 u64 start_blkno
= ocfs2_clusters_to_blocks(sb
, phys
);
5192 struct ocfs2_extent_rec split_rec
;
5193 struct ocfs2_path
*left_path
= NULL
;
5194 struct ocfs2_extent_list
*el
;
5195 struct ocfs2_extent_rec
*rec
;
5197 left_path
= ocfs2_new_path_from_et(et
);
5204 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
5209 el
= path_leaf_el(left_path
);
5211 index
= ocfs2_search_extent_list(el
, cpos
);
5212 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5214 "Owner %llu has an extent at cpos %u which can no "
5215 "longer be found.\n",
5216 (unsigned long long)
5217 ocfs2_metadata_cache_owner(et
->et_ci
), cpos
);
5223 rec
= &el
->l_recs
[index
];
5224 if (new_flags
&& (rec
->e_flags
& new_flags
)) {
5225 mlog(ML_ERROR
, "Owner %llu tried to set %d flags on an "
5226 "extent that already had them",
5227 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5232 if (clear_flags
&& !(rec
->e_flags
& clear_flags
)) {
5233 mlog(ML_ERROR
, "Owner %llu tried to clear %d flags on an "
5234 "extent that didn't have them",
5235 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5240 memset(&split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
5241 split_rec
.e_cpos
= cpu_to_le32(cpos
);
5242 split_rec
.e_leaf_clusters
= cpu_to_le16(len
);
5243 split_rec
.e_blkno
= cpu_to_le64(start_blkno
);
5244 split_rec
.e_flags
= rec
->e_flags
;
5246 split_rec
.e_flags
|= new_flags
;
5248 split_rec
.e_flags
&= ~clear_flags
;
5250 ret
= ocfs2_split_extent(handle
, et
, left_path
,
5251 index
, &split_rec
, meta_ac
,
5257 ocfs2_free_path(left_path
);
5263 * Mark the already-existing extent at cpos as written for len clusters.
5264 * This removes the unwritten extent flag.
5266 * If the existing extent is larger than the request, initiate a
5267 * split. An attempt will be made at merging with adjacent extents.
5269 * The caller is responsible for passing down meta_ac if we'll need it.
5271 int ocfs2_mark_extent_written(struct inode
*inode
,
5272 struct ocfs2_extent_tree
*et
,
5273 handle_t
*handle
, u32 cpos
, u32 len
, u32 phys
,
5274 struct ocfs2_alloc_context
*meta_ac
,
5275 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5279 mlog(0, "Inode %lu cpos %u, len %u, phys clusters %u\n",
5280 inode
->i_ino
, cpos
, len
, phys
);
5282 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode
->i_sb
))) {
5283 ocfs2_error(inode
->i_sb
, "Inode %llu has unwritten extents "
5284 "that are being written to, but the feature bit "
5285 "is not set in the super block.",
5286 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
5292 * XXX: This should be fixed up so that we just re-insert the
5293 * next extent records.
5295 ocfs2_et_extent_map_truncate(et
, 0);
5297 ret
= ocfs2_change_extent_flag(handle
, et
, cpos
,
5298 len
, phys
, meta_ac
, dealloc
,
5299 0, OCFS2_EXT_UNWRITTEN
);
5307 static int ocfs2_split_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
5308 struct ocfs2_path
*path
,
5309 int index
, u32 new_range
,
5310 struct ocfs2_alloc_context
*meta_ac
)
5312 int ret
, depth
, credits
= handle
->h_buffer_credits
;
5313 struct buffer_head
*last_eb_bh
= NULL
;
5314 struct ocfs2_extent_block
*eb
;
5315 struct ocfs2_extent_list
*rightmost_el
, *el
;
5316 struct ocfs2_extent_rec split_rec
;
5317 struct ocfs2_extent_rec
*rec
;
5318 struct ocfs2_insert_type insert
;
5321 * Setup the record to split before we grow the tree.
5323 el
= path_leaf_el(path
);
5324 rec
= &el
->l_recs
[index
];
5325 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
5326 &split_rec
, new_range
, rec
);
5328 depth
= path
->p_tree_depth
;
5330 ret
= ocfs2_read_extent_block(et
->et_ci
,
5331 ocfs2_et_get_last_eb_blk(et
),
5338 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5339 rightmost_el
= &eb
->h_list
;
5341 rightmost_el
= path_leaf_el(path
);
5343 credits
+= path
->p_tree_depth
+
5344 ocfs2_extend_meta_needed(et
->et_root_el
);
5345 ret
= ocfs2_extend_trans(handle
, credits
);
5351 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
5352 le16_to_cpu(rightmost_el
->l_count
)) {
5353 ret
= ocfs2_grow_tree(handle
, et
, &depth
, &last_eb_bh
,
5361 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
5362 insert
.ins_appending
= APPEND_NONE
;
5363 insert
.ins_contig
= CONTIG_NONE
;
5364 insert
.ins_split
= SPLIT_RIGHT
;
5365 insert
.ins_tree_depth
= depth
;
5367 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
5376 static int ocfs2_truncate_rec(handle_t
*handle
,
5377 struct ocfs2_extent_tree
*et
,
5378 struct ocfs2_path
*path
, int index
,
5379 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5383 u32 left_cpos
, rec_range
, trunc_range
;
5384 int wants_rotate
= 0, is_rightmost_tree_rec
= 0;
5385 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5386 struct ocfs2_path
*left_path
= NULL
;
5387 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5388 struct ocfs2_extent_rec
*rec
;
5389 struct ocfs2_extent_block
*eb
;
5391 if (ocfs2_is_empty_extent(&el
->l_recs
[0]) && index
> 0) {
5392 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5401 if (index
== (le16_to_cpu(el
->l_next_free_rec
) - 1) &&
5402 path
->p_tree_depth
) {
5404 * Check whether this is the rightmost tree record. If
5405 * we remove all of this record or part of its right
5406 * edge then an update of the record lengths above it
5409 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
5410 if (eb
->h_next_leaf_blk
== 0)
5411 is_rightmost_tree_rec
= 1;
5414 rec
= &el
->l_recs
[index
];
5415 if (index
== 0 && path
->p_tree_depth
&&
5416 le32_to_cpu(rec
->e_cpos
) == cpos
) {
5418 * Changing the leftmost offset (via partial or whole
5419 * record truncate) of an interior (or rightmost) path
5420 * means we have to update the subtree that is formed
5421 * by this leaf and the one to it's left.
5423 * There are two cases we can skip:
5424 * 1) Path is the leftmost one in our btree.
5425 * 2) The leaf is rightmost and will be empty after
5426 * we remove the extent record - the rotate code
5427 * knows how to update the newly formed edge.
5430 ret
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
5436 if (left_cpos
&& le16_to_cpu(el
->l_next_free_rec
) > 1) {
5437 left_path
= ocfs2_new_path_from_path(path
);
5444 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
5453 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
5454 handle
->h_buffer_credits
,
5461 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
5467 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
5473 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5474 trunc_range
= cpos
+ len
;
5476 if (le32_to_cpu(rec
->e_cpos
) == cpos
&& rec_range
== trunc_range
) {
5479 memset(rec
, 0, sizeof(*rec
));
5480 ocfs2_cleanup_merge(el
, index
);
5483 next_free
= le16_to_cpu(el
->l_next_free_rec
);
5484 if (is_rightmost_tree_rec
&& next_free
> 1) {
5486 * We skip the edge update if this path will
5487 * be deleted by the rotate code.
5489 rec
= &el
->l_recs
[next_free
- 1];
5490 ocfs2_adjust_rightmost_records(handle
, et
, path
,
5493 } else if (le32_to_cpu(rec
->e_cpos
) == cpos
) {
5494 /* Remove leftmost portion of the record. */
5495 le32_add_cpu(&rec
->e_cpos
, len
);
5496 le64_add_cpu(&rec
->e_blkno
, ocfs2_clusters_to_blocks(sb
, len
));
5497 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5498 } else if (rec_range
== trunc_range
) {
5499 /* Remove rightmost portion of the record */
5500 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5501 if (is_rightmost_tree_rec
)
5502 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
5504 /* Caller should have trapped this. */
5505 mlog(ML_ERROR
, "Owner %llu: Invalid record truncate: (%u, %u) "
5507 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5508 le32_to_cpu(rec
->e_cpos
),
5509 le16_to_cpu(rec
->e_leaf_clusters
), cpos
, len
);
5516 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
5517 ocfs2_complete_edge_insert(handle
, left_path
, path
,
5521 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5523 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5530 ocfs2_free_path(left_path
);
5534 int ocfs2_remove_extent(handle_t
*handle
,
5535 struct ocfs2_extent_tree
*et
,
5537 struct ocfs2_alloc_context
*meta_ac
,
5538 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5541 u32 rec_range
, trunc_range
;
5542 struct ocfs2_extent_rec
*rec
;
5543 struct ocfs2_extent_list
*el
;
5544 struct ocfs2_path
*path
= NULL
;
5547 * XXX: Why are we truncating to 0 instead of wherever this
5550 ocfs2_et_extent_map_truncate(et
, 0);
5552 path
= ocfs2_new_path_from_et(et
);
5559 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5565 el
= path_leaf_el(path
);
5566 index
= ocfs2_search_extent_list(el
, cpos
);
5567 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5568 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5569 "Owner %llu has an extent at cpos %u which can no "
5570 "longer be found.\n",
5571 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5578 * We have 3 cases of extent removal:
5579 * 1) Range covers the entire extent rec
5580 * 2) Range begins or ends on one edge of the extent rec
5581 * 3) Range is in the middle of the extent rec (no shared edges)
5583 * For case 1 we remove the extent rec and left rotate to
5586 * For case 2 we just shrink the existing extent rec, with a
5587 * tree update if the shrinking edge is also the edge of an
5590 * For case 3 we do a right split to turn the extent rec into
5591 * something case 2 can handle.
5593 rec
= &el
->l_recs
[index
];
5594 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5595 trunc_range
= cpos
+ len
;
5597 BUG_ON(cpos
< le32_to_cpu(rec
->e_cpos
) || trunc_range
> rec_range
);
5599 mlog(0, "Owner %llu, remove (cpos %u, len %u). Existing index %d "
5600 "(cpos %u, len %u)\n",
5601 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5603 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
));
5605 if (le32_to_cpu(rec
->e_cpos
) == cpos
|| rec_range
== trunc_range
) {
5606 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5613 ret
= ocfs2_split_tree(handle
, et
, path
, index
,
5614 trunc_range
, meta_ac
);
5621 * The split could have manipulated the tree enough to
5622 * move the record location, so we have to look for it again.
5624 ocfs2_reinit_path(path
, 1);
5626 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5632 el
= path_leaf_el(path
);
5633 index
= ocfs2_search_extent_list(el
, cpos
);
5634 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5635 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5636 "Owner %llu: split at cpos %u lost record.",
5637 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5644 * Double check our values here. If anything is fishy,
5645 * it's easier to catch it at the top level.
5647 rec
= &el
->l_recs
[index
];
5648 rec_range
= le32_to_cpu(rec
->e_cpos
) +
5649 ocfs2_rec_clusters(el
, rec
);
5650 if (rec_range
!= trunc_range
) {
5651 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5652 "Owner %llu: error after split at cpos %u"
5653 "trunc len %u, existing record is (%u,%u)",
5654 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5655 cpos
, len
, le32_to_cpu(rec
->e_cpos
),
5656 ocfs2_rec_clusters(el
, rec
));
5661 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5670 ocfs2_free_path(path
);
5674 int ocfs2_remove_btree_range(struct inode
*inode
,
5675 struct ocfs2_extent_tree
*et
,
5676 u32 cpos
, u32 phys_cpos
, u32 len
,
5677 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5680 u64 phys_blkno
= ocfs2_clusters_to_blocks(inode
->i_sb
, phys_cpos
);
5681 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
5682 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5684 struct ocfs2_alloc_context
*meta_ac
= NULL
;
5686 ret
= ocfs2_lock_allocators(inode
, et
, 0, 1, NULL
, &meta_ac
);
5692 mutex_lock(&tl_inode
->i_mutex
);
5694 if (ocfs2_truncate_log_needs_flush(osb
)) {
5695 ret
= __ocfs2_flush_truncate_log(osb
);
5702 handle
= ocfs2_start_trans(osb
, ocfs2_remove_extent_credits(osb
->sb
));
5703 if (IS_ERR(handle
)) {
5704 ret
= PTR_ERR(handle
);
5709 ret
= ocfs2_et_root_journal_access(handle
, et
,
5710 OCFS2_JOURNAL_ACCESS_WRITE
);
5716 dquot_free_space_nodirty(inode
,
5717 ocfs2_clusters_to_bytes(inode
->i_sb
, len
));
5719 ret
= ocfs2_remove_extent(handle
, et
, cpos
, len
, meta_ac
, dealloc
);
5725 ocfs2_et_update_clusters(et
, -len
);
5727 ret
= ocfs2_journal_dirty(handle
, et
->et_root_bh
);
5733 ret
= ocfs2_truncate_log_append(osb
, handle
, phys_blkno
, len
);
5738 ocfs2_commit_trans(osb
, handle
);
5740 mutex_unlock(&tl_inode
->i_mutex
);
5743 ocfs2_free_alloc_context(meta_ac
);
5748 int ocfs2_truncate_log_needs_flush(struct ocfs2_super
*osb
)
5750 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5751 struct ocfs2_dinode
*di
;
5752 struct ocfs2_truncate_log
*tl
;
5754 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5755 tl
= &di
->id2
.i_dealloc
;
5757 mlog_bug_on_msg(le16_to_cpu(tl
->tl_used
) > le16_to_cpu(tl
->tl_count
),
5758 "slot %d, invalid truncate log parameters: used = "
5759 "%u, count = %u\n", osb
->slot_num
,
5760 le16_to_cpu(tl
->tl_used
), le16_to_cpu(tl
->tl_count
));
5761 return le16_to_cpu(tl
->tl_used
) == le16_to_cpu(tl
->tl_count
);
5764 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log
*tl
,
5765 unsigned int new_start
)
5767 unsigned int tail_index
;
5768 unsigned int current_tail
;
5770 /* No records, nothing to coalesce */
5771 if (!le16_to_cpu(tl
->tl_used
))
5774 tail_index
= le16_to_cpu(tl
->tl_used
) - 1;
5775 current_tail
= le32_to_cpu(tl
->tl_recs
[tail_index
].t_start
);
5776 current_tail
+= le32_to_cpu(tl
->tl_recs
[tail_index
].t_clusters
);
5778 return current_tail
== new_start
;
5781 int ocfs2_truncate_log_append(struct ocfs2_super
*osb
,
5784 unsigned int num_clusters
)
5787 unsigned int start_cluster
, tl_count
;
5788 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5789 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5790 struct ocfs2_dinode
*di
;
5791 struct ocfs2_truncate_log
*tl
;
5793 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5794 (unsigned long long)start_blk
, num_clusters
);
5796 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
5798 start_cluster
= ocfs2_blocks_to_clusters(osb
->sb
, start_blk
);
5800 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5802 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5803 * by the underlying call to ocfs2_read_inode_block(), so any
5804 * corruption is a code bug */
5805 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
5807 tl
= &di
->id2
.i_dealloc
;
5808 tl_count
= le16_to_cpu(tl
->tl_count
);
5809 mlog_bug_on_msg(tl_count
> ocfs2_truncate_recs_per_inode(osb
->sb
) ||
5811 "Truncate record count on #%llu invalid "
5812 "wanted %u, actual %u\n",
5813 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5814 ocfs2_truncate_recs_per_inode(osb
->sb
),
5815 le16_to_cpu(tl
->tl_count
));
5817 /* Caller should have known to flush before calling us. */
5818 index
= le16_to_cpu(tl
->tl_used
);
5819 if (index
>= tl_count
) {
5825 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5826 OCFS2_JOURNAL_ACCESS_WRITE
);
5832 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5833 "%llu (index = %d)\n", num_clusters
, start_cluster
,
5834 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
, index
);
5836 if (ocfs2_truncate_log_can_coalesce(tl
, start_cluster
)) {
5838 * Move index back to the record we are coalescing with.
5839 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5843 num_clusters
+= le32_to_cpu(tl
->tl_recs
[index
].t_clusters
);
5844 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5845 index
, le32_to_cpu(tl
->tl_recs
[index
].t_start
),
5848 tl
->tl_recs
[index
].t_start
= cpu_to_le32(start_cluster
);
5849 tl
->tl_used
= cpu_to_le16(index
+ 1);
5851 tl
->tl_recs
[index
].t_clusters
= cpu_to_le32(num_clusters
);
5853 status
= ocfs2_journal_dirty(handle
, tl_bh
);
5864 static int ocfs2_replay_truncate_records(struct ocfs2_super
*osb
,
5866 struct inode
*data_alloc_inode
,
5867 struct buffer_head
*data_alloc_bh
)
5871 unsigned int num_clusters
;
5873 struct ocfs2_truncate_rec rec
;
5874 struct ocfs2_dinode
*di
;
5875 struct ocfs2_truncate_log
*tl
;
5876 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5877 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5881 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5882 tl
= &di
->id2
.i_dealloc
;
5883 i
= le16_to_cpu(tl
->tl_used
) - 1;
5885 /* Caller has given us at least enough credits to
5886 * update the truncate log dinode */
5887 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5888 OCFS2_JOURNAL_ACCESS_WRITE
);
5894 tl
->tl_used
= cpu_to_le16(i
);
5896 status
= ocfs2_journal_dirty(handle
, tl_bh
);
5902 /* TODO: Perhaps we can calculate the bulk of the
5903 * credits up front rather than extending like
5905 status
= ocfs2_extend_trans(handle
,
5906 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC
);
5912 rec
= tl
->tl_recs
[i
];
5913 start_blk
= ocfs2_clusters_to_blocks(data_alloc_inode
->i_sb
,
5914 le32_to_cpu(rec
.t_start
));
5915 num_clusters
= le32_to_cpu(rec
.t_clusters
);
5917 /* if start_blk is not set, we ignore the record as
5920 mlog(0, "free record %d, start = %u, clusters = %u\n",
5921 i
, le32_to_cpu(rec
.t_start
), num_clusters
);
5923 status
= ocfs2_free_clusters(handle
, data_alloc_inode
,
5924 data_alloc_bh
, start_blk
,
5939 /* Expects you to already be holding tl_inode->i_mutex */
5940 int __ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5943 unsigned int num_to_flush
;
5945 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5946 struct inode
*data_alloc_inode
= NULL
;
5947 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5948 struct buffer_head
*data_alloc_bh
= NULL
;
5949 struct ocfs2_dinode
*di
;
5950 struct ocfs2_truncate_log
*tl
;
5954 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
5956 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5958 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5959 * by the underlying call to ocfs2_read_inode_block(), so any
5960 * corruption is a code bug */
5961 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
5963 tl
= &di
->id2
.i_dealloc
;
5964 num_to_flush
= le16_to_cpu(tl
->tl_used
);
5965 mlog(0, "Flush %u records from truncate log #%llu\n",
5966 num_to_flush
, (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
);
5967 if (!num_to_flush
) {
5972 data_alloc_inode
= ocfs2_get_system_file_inode(osb
,
5973 GLOBAL_BITMAP_SYSTEM_INODE
,
5974 OCFS2_INVALID_SLOT
);
5975 if (!data_alloc_inode
) {
5977 mlog(ML_ERROR
, "Could not get bitmap inode!\n");
5981 mutex_lock(&data_alloc_inode
->i_mutex
);
5983 status
= ocfs2_inode_lock(data_alloc_inode
, &data_alloc_bh
, 1);
5989 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
5990 if (IS_ERR(handle
)) {
5991 status
= PTR_ERR(handle
);
5996 status
= ocfs2_replay_truncate_records(osb
, handle
, data_alloc_inode
,
6001 ocfs2_commit_trans(osb
, handle
);
6004 brelse(data_alloc_bh
);
6005 ocfs2_inode_unlock(data_alloc_inode
, 1);
6008 mutex_unlock(&data_alloc_inode
->i_mutex
);
6009 iput(data_alloc_inode
);
6016 int ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
6019 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6021 mutex_lock(&tl_inode
->i_mutex
);
6022 status
= __ocfs2_flush_truncate_log(osb
);
6023 mutex_unlock(&tl_inode
->i_mutex
);
6028 static void ocfs2_truncate_log_worker(struct work_struct
*work
)
6031 struct ocfs2_super
*osb
=
6032 container_of(work
, struct ocfs2_super
,
6033 osb_truncate_log_wq
.work
);
6037 status
= ocfs2_flush_truncate_log(osb
);
6041 ocfs2_init_steal_slots(osb
);
6046 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6047 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super
*osb
,
6050 if (osb
->osb_tl_inode
) {
6051 /* We want to push off log flushes while truncates are
6054 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
6056 queue_delayed_work(ocfs2_wq
, &osb
->osb_truncate_log_wq
,
6057 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL
);
6061 static int ocfs2_get_truncate_log_info(struct ocfs2_super
*osb
,
6063 struct inode
**tl_inode
,
6064 struct buffer_head
**tl_bh
)
6067 struct inode
*inode
= NULL
;
6068 struct buffer_head
*bh
= NULL
;
6070 inode
= ocfs2_get_system_file_inode(osb
,
6071 TRUNCATE_LOG_SYSTEM_INODE
,
6075 mlog(ML_ERROR
, "Could not get load truncate log inode!\n");
6079 status
= ocfs2_read_inode_block(inode
, &bh
);
6093 /* called during the 1st stage of node recovery. we stamp a clean
6094 * truncate log and pass back a copy for processing later. if the
6095 * truncate log does not require processing, a *tl_copy is set to
6097 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super
*osb
,
6099 struct ocfs2_dinode
**tl_copy
)
6102 struct inode
*tl_inode
= NULL
;
6103 struct buffer_head
*tl_bh
= NULL
;
6104 struct ocfs2_dinode
*di
;
6105 struct ocfs2_truncate_log
*tl
;
6109 mlog(0, "recover truncate log from slot %d\n", slot_num
);
6111 status
= ocfs2_get_truncate_log_info(osb
, slot_num
, &tl_inode
, &tl_bh
);
6117 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
6119 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6120 * validated by the underlying call to ocfs2_read_inode_block(),
6121 * so any corruption is a code bug */
6122 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
6124 tl
= &di
->id2
.i_dealloc
;
6125 if (le16_to_cpu(tl
->tl_used
)) {
6126 mlog(0, "We'll have %u logs to recover\n",
6127 le16_to_cpu(tl
->tl_used
));
6129 *tl_copy
= kmalloc(tl_bh
->b_size
, GFP_KERNEL
);
6136 /* Assuming the write-out below goes well, this copy
6137 * will be passed back to recovery for processing. */
6138 memcpy(*tl_copy
, tl_bh
->b_data
, tl_bh
->b_size
);
6140 /* All we need to do to clear the truncate log is set
6144 ocfs2_compute_meta_ecc(osb
->sb
, tl_bh
->b_data
, &di
->i_check
);
6145 status
= ocfs2_write_block(osb
, tl_bh
, INODE_CACHE(tl_inode
));
6157 if (status
< 0 && (*tl_copy
)) {
6166 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super
*osb
,
6167 struct ocfs2_dinode
*tl_copy
)
6171 unsigned int clusters
, num_recs
, start_cluster
;
6174 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6175 struct ocfs2_truncate_log
*tl
;
6179 if (OCFS2_I(tl_inode
)->ip_blkno
== le64_to_cpu(tl_copy
->i_blkno
)) {
6180 mlog(ML_ERROR
, "Asked to recover my own truncate log!\n");
6184 tl
= &tl_copy
->id2
.i_dealloc
;
6185 num_recs
= le16_to_cpu(tl
->tl_used
);
6186 mlog(0, "cleanup %u records from %llu\n", num_recs
,
6187 (unsigned long long)le64_to_cpu(tl_copy
->i_blkno
));
6189 mutex_lock(&tl_inode
->i_mutex
);
6190 for(i
= 0; i
< num_recs
; i
++) {
6191 if (ocfs2_truncate_log_needs_flush(osb
)) {
6192 status
= __ocfs2_flush_truncate_log(osb
);
6199 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
6200 if (IS_ERR(handle
)) {
6201 status
= PTR_ERR(handle
);
6206 clusters
= le32_to_cpu(tl
->tl_recs
[i
].t_clusters
);
6207 start_cluster
= le32_to_cpu(tl
->tl_recs
[i
].t_start
);
6208 start_blk
= ocfs2_clusters_to_blocks(osb
->sb
, start_cluster
);
6210 status
= ocfs2_truncate_log_append(osb
, handle
,
6211 start_blk
, clusters
);
6212 ocfs2_commit_trans(osb
, handle
);
6220 mutex_unlock(&tl_inode
->i_mutex
);
6226 void ocfs2_truncate_log_shutdown(struct ocfs2_super
*osb
)
6229 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6234 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
6235 flush_workqueue(ocfs2_wq
);
6237 status
= ocfs2_flush_truncate_log(osb
);
6241 brelse(osb
->osb_tl_bh
);
6242 iput(osb
->osb_tl_inode
);
6248 int ocfs2_truncate_log_init(struct ocfs2_super
*osb
)
6251 struct inode
*tl_inode
= NULL
;
6252 struct buffer_head
*tl_bh
= NULL
;
6256 status
= ocfs2_get_truncate_log_info(osb
,
6263 /* ocfs2_truncate_log_shutdown keys on the existence of
6264 * osb->osb_tl_inode so we don't set any of the osb variables
6265 * until we're sure all is well. */
6266 INIT_DELAYED_WORK(&osb
->osb_truncate_log_wq
,
6267 ocfs2_truncate_log_worker
);
6268 osb
->osb_tl_bh
= tl_bh
;
6269 osb
->osb_tl_inode
= tl_inode
;
6276 * Delayed de-allocation of suballocator blocks.
6278 * Some sets of block de-allocations might involve multiple suballocator inodes.
6280 * The locking for this can get extremely complicated, especially when
6281 * the suballocator inodes to delete from aren't known until deep
6282 * within an unrelated codepath.
6284 * ocfs2_extent_block structures are a good example of this - an inode
6285 * btree could have been grown by any number of nodes each allocating
6286 * out of their own suballoc inode.
6288 * These structures allow the delay of block de-allocation until a
6289 * later time, when locking of multiple cluster inodes won't cause
6294 * Describe a single bit freed from a suballocator. For the block
6295 * suballocators, it represents one block. For the global cluster
6296 * allocator, it represents some clusters and free_bit indicates
6299 struct ocfs2_cached_block_free
{
6300 struct ocfs2_cached_block_free
*free_next
;
6302 unsigned int free_bit
;
6305 struct ocfs2_per_slot_free_list
{
6306 struct ocfs2_per_slot_free_list
*f_next_suballocator
;
6309 struct ocfs2_cached_block_free
*f_first
;
6312 static int ocfs2_free_cached_blocks(struct ocfs2_super
*osb
,
6315 struct ocfs2_cached_block_free
*head
)
6320 struct inode
*inode
;
6321 struct buffer_head
*di_bh
= NULL
;
6322 struct ocfs2_cached_block_free
*tmp
;
6324 inode
= ocfs2_get_system_file_inode(osb
, sysfile_type
, slot
);
6331 mutex_lock(&inode
->i_mutex
);
6333 ret
= ocfs2_inode_lock(inode
, &di_bh
, 1);
6339 handle
= ocfs2_start_trans(osb
, OCFS2_SUBALLOC_FREE
);
6340 if (IS_ERR(handle
)) {
6341 ret
= PTR_ERR(handle
);
6347 bg_blkno
= ocfs2_which_suballoc_group(head
->free_blk
,
6349 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
6350 head
->free_bit
, (unsigned long long)head
->free_blk
);
6352 ret
= ocfs2_free_suballoc_bits(handle
, inode
, di_bh
,
6353 head
->free_bit
, bg_blkno
, 1);
6359 ret
= ocfs2_extend_trans(handle
, OCFS2_SUBALLOC_FREE
);
6366 head
= head
->free_next
;
6371 ocfs2_commit_trans(osb
, handle
);
6374 ocfs2_inode_unlock(inode
, 1);
6377 mutex_unlock(&inode
->i_mutex
);
6381 /* Premature exit may have left some dangling items. */
6383 head
= head
->free_next
;
6390 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6391 u64 blkno
, unsigned int bit
)
6394 struct ocfs2_cached_block_free
*item
;
6396 item
= kmalloc(sizeof(*item
), GFP_NOFS
);
6403 mlog(0, "Insert clusters: (bit %u, blk %llu)\n",
6404 bit
, (unsigned long long)blkno
);
6406 item
->free_blk
= blkno
;
6407 item
->free_bit
= bit
;
6408 item
->free_next
= ctxt
->c_global_allocator
;
6410 ctxt
->c_global_allocator
= item
;
6414 static int ocfs2_free_cached_clusters(struct ocfs2_super
*osb
,
6415 struct ocfs2_cached_block_free
*head
)
6417 struct ocfs2_cached_block_free
*tmp
;
6418 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6422 mutex_lock(&tl_inode
->i_mutex
);
6425 if (ocfs2_truncate_log_needs_flush(osb
)) {
6426 ret
= __ocfs2_flush_truncate_log(osb
);
6433 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
6434 if (IS_ERR(handle
)) {
6435 ret
= PTR_ERR(handle
);
6440 ret
= ocfs2_truncate_log_append(osb
, handle
, head
->free_blk
,
6443 ocfs2_commit_trans(osb
, handle
);
6445 head
= head
->free_next
;
6454 mutex_unlock(&tl_inode
->i_mutex
);
6457 /* Premature exit may have left some dangling items. */
6459 head
= head
->free_next
;
6466 int ocfs2_run_deallocs(struct ocfs2_super
*osb
,
6467 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
6470 struct ocfs2_per_slot_free_list
*fl
;
6475 while (ctxt
->c_first_suballocator
) {
6476 fl
= ctxt
->c_first_suballocator
;
6479 mlog(0, "Free items: (type %u, slot %d)\n",
6480 fl
->f_inode_type
, fl
->f_slot
);
6481 ret2
= ocfs2_free_cached_blocks(osb
,
6491 ctxt
->c_first_suballocator
= fl
->f_next_suballocator
;
6495 if (ctxt
->c_global_allocator
) {
6496 ret2
= ocfs2_free_cached_clusters(osb
,
6497 ctxt
->c_global_allocator
);
6503 ctxt
->c_global_allocator
= NULL
;
6509 static struct ocfs2_per_slot_free_list
*
6510 ocfs2_find_per_slot_free_list(int type
,
6512 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
6514 struct ocfs2_per_slot_free_list
*fl
= ctxt
->c_first_suballocator
;
6517 if (fl
->f_inode_type
== type
&& fl
->f_slot
== slot
)
6520 fl
= fl
->f_next_suballocator
;
6523 fl
= kmalloc(sizeof(*fl
), GFP_NOFS
);
6525 fl
->f_inode_type
= type
;
6528 fl
->f_next_suballocator
= ctxt
->c_first_suballocator
;
6530 ctxt
->c_first_suballocator
= fl
;
6535 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6536 int type
, int slot
, u64 blkno
,
6540 struct ocfs2_per_slot_free_list
*fl
;
6541 struct ocfs2_cached_block_free
*item
;
6543 fl
= ocfs2_find_per_slot_free_list(type
, slot
, ctxt
);
6550 item
= kmalloc(sizeof(*item
), GFP_NOFS
);
6557 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
6558 type
, slot
, bit
, (unsigned long long)blkno
);
6560 item
->free_blk
= blkno
;
6561 item
->free_bit
= bit
;
6562 item
->free_next
= fl
->f_first
;
6571 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6572 struct ocfs2_extent_block
*eb
)
6574 return ocfs2_cache_block_dealloc(ctxt
, EXTENT_ALLOC_SYSTEM_INODE
,
6575 le16_to_cpu(eb
->h_suballoc_slot
),
6576 le64_to_cpu(eb
->h_blkno
),
6577 le16_to_cpu(eb
->h_suballoc_bit
));
6580 /* This function will figure out whether the currently last extent
6581 * block will be deleted, and if it will, what the new last extent
6582 * block will be so we can update his h_next_leaf_blk field, as well
6583 * as the dinodes i_last_eb_blk */
6584 static int ocfs2_find_new_last_ext_blk(struct inode
*inode
,
6585 unsigned int clusters_to_del
,
6586 struct ocfs2_path
*path
,
6587 struct buffer_head
**new_last_eb
)
6589 int next_free
, ret
= 0;
6591 struct ocfs2_extent_rec
*rec
;
6592 struct ocfs2_extent_block
*eb
;
6593 struct ocfs2_extent_list
*el
;
6594 struct buffer_head
*bh
= NULL
;
6596 *new_last_eb
= NULL
;
6598 /* we have no tree, so of course, no last_eb. */
6599 if (!path
->p_tree_depth
)
6602 /* trunc to zero special case - this makes tree_depth = 0
6603 * regardless of what it is. */
6604 if (OCFS2_I(inode
)->ip_clusters
== clusters_to_del
)
6607 el
= path_leaf_el(path
);
6608 BUG_ON(!el
->l_next_free_rec
);
6611 * Make sure that this extent list will actually be empty
6612 * after we clear away the data. We can shortcut out if
6613 * there's more than one non-empty extent in the
6614 * list. Otherwise, a check of the remaining extent is
6617 next_free
= le16_to_cpu(el
->l_next_free_rec
);
6619 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
6623 /* We may have a valid extent in index 1, check it. */
6625 rec
= &el
->l_recs
[1];
6628 * Fall through - no more nonempty extents, so we want
6629 * to delete this leaf.
6635 rec
= &el
->l_recs
[0];
6640 * Check it we'll only be trimming off the end of this
6643 if (le16_to_cpu(rec
->e_leaf_clusters
) > clusters_to_del
)
6647 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
6653 ret
= ocfs2_find_leaf(INODE_CACHE(inode
), path_root_el(path
), cpos
, &bh
);
6659 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
6662 /* ocfs2_find_leaf() gets the eb from ocfs2_read_extent_block().
6663 * Any corruption is a code bug. */
6664 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
6667 get_bh(*new_last_eb
);
6668 mlog(0, "returning block %llu, (cpos: %u)\n",
6669 (unsigned long long)le64_to_cpu(eb
->h_blkno
), cpos
);
6677 * Trim some clusters off the rightmost edge of a tree. Only called
6680 * The caller needs to:
6681 * - start journaling of each path component.
6682 * - compute and fully set up any new last ext block
6684 static int ocfs2_trim_tree(struct inode
*inode
, struct ocfs2_path
*path
,
6685 handle_t
*handle
, struct ocfs2_truncate_context
*tc
,
6686 u32 clusters_to_del
, u64
*delete_start
, u8
*flags
)
6688 int ret
, i
, index
= path
->p_tree_depth
;
6691 struct buffer_head
*bh
;
6692 struct ocfs2_extent_list
*el
;
6693 struct ocfs2_extent_rec
*rec
;
6698 while (index
>= 0) {
6699 bh
= path
->p_node
[index
].bh
;
6700 el
= path
->p_node
[index
].el
;
6702 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6703 index
, (unsigned long long)bh
->b_blocknr
);
6705 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
6708 (path
->p_tree_depth
- le16_to_cpu(el
->l_tree_depth
))) {
6709 ocfs2_error(inode
->i_sb
,
6710 "Inode %lu has invalid ext. block %llu",
6712 (unsigned long long)bh
->b_blocknr
);
6718 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
6719 rec
= &el
->l_recs
[i
];
6721 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6722 "next = %u\n", i
, le32_to_cpu(rec
->e_cpos
),
6723 ocfs2_rec_clusters(el
, rec
),
6724 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
6725 le16_to_cpu(el
->l_next_free_rec
));
6727 BUG_ON(ocfs2_rec_clusters(el
, rec
) < clusters_to_del
);
6729 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
6731 * If the leaf block contains a single empty
6732 * extent and no records, we can just remove
6735 if (i
== 0 && ocfs2_is_empty_extent(rec
)) {
6737 sizeof(struct ocfs2_extent_rec
));
6738 el
->l_next_free_rec
= cpu_to_le16(0);
6744 * Remove any empty extents by shifting things
6745 * left. That should make life much easier on
6746 * the code below. This condition is rare
6747 * enough that we shouldn't see a performance
6750 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
6751 le16_add_cpu(&el
->l_next_free_rec
, -1);
6754 i
< le16_to_cpu(el
->l_next_free_rec
); i
++)
6755 el
->l_recs
[i
] = el
->l_recs
[i
+ 1];
6757 memset(&el
->l_recs
[i
], 0,
6758 sizeof(struct ocfs2_extent_rec
));
6761 * We've modified our extent list. The
6762 * simplest way to handle this change
6763 * is to being the search from the
6766 goto find_tail_record
;
6769 le16_add_cpu(&rec
->e_leaf_clusters
, -clusters_to_del
);
6772 * We'll use "new_edge" on our way back up the
6773 * tree to know what our rightmost cpos is.
6775 new_edge
= le16_to_cpu(rec
->e_leaf_clusters
);
6776 new_edge
+= le32_to_cpu(rec
->e_cpos
);
6779 * The caller will use this to delete data blocks.
6781 *delete_start
= le64_to_cpu(rec
->e_blkno
)
6782 + ocfs2_clusters_to_blocks(inode
->i_sb
,
6783 le16_to_cpu(rec
->e_leaf_clusters
));
6784 *flags
= rec
->e_flags
;
6787 * If it's now empty, remove this record.
6789 if (le16_to_cpu(rec
->e_leaf_clusters
) == 0) {
6791 sizeof(struct ocfs2_extent_rec
));
6792 le16_add_cpu(&el
->l_next_free_rec
, -1);
6795 if (le64_to_cpu(rec
->e_blkno
) == deleted_eb
) {
6797 sizeof(struct ocfs2_extent_rec
));
6798 le16_add_cpu(&el
->l_next_free_rec
, -1);
6803 /* Can this actually happen? */
6804 if (le16_to_cpu(el
->l_next_free_rec
) == 0)
6808 * We never actually deleted any clusters
6809 * because our leaf was empty. There's no
6810 * reason to adjust the rightmost edge then.
6815 rec
->e_int_clusters
= cpu_to_le32(new_edge
);
6816 le32_add_cpu(&rec
->e_int_clusters
,
6817 -le32_to_cpu(rec
->e_cpos
));
6820 * A deleted child record should have been
6823 BUG_ON(le32_to_cpu(rec
->e_int_clusters
) == 0);
6827 ret
= ocfs2_journal_dirty(handle
, bh
);
6833 mlog(0, "extent list container %llu, after: record %d: "
6834 "(%u, %u, %llu), next = %u.\n",
6835 (unsigned long long)bh
->b_blocknr
, i
,
6836 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
),
6837 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
6838 le16_to_cpu(el
->l_next_free_rec
));
6841 * We must be careful to only attempt delete of an
6842 * extent block (and not the root inode block).
6844 if (index
> 0 && le16_to_cpu(el
->l_next_free_rec
) == 0) {
6845 struct ocfs2_extent_block
*eb
=
6846 (struct ocfs2_extent_block
*)bh
->b_data
;
6849 * Save this for use when processing the
6852 deleted_eb
= le64_to_cpu(eb
->h_blkno
);
6854 mlog(0, "deleting this extent block.\n");
6856 ocfs2_remove_from_cache(INODE_CACHE(inode
), bh
);
6858 BUG_ON(ocfs2_rec_clusters(el
, &el
->l_recs
[0]));
6859 BUG_ON(le32_to_cpu(el
->l_recs
[0].e_cpos
));
6860 BUG_ON(le64_to_cpu(el
->l_recs
[0].e_blkno
));
6862 ret
= ocfs2_cache_extent_block_free(&tc
->tc_dealloc
, eb
);
6863 /* An error here is not fatal. */
6878 static int ocfs2_do_truncate(struct ocfs2_super
*osb
,
6879 unsigned int clusters_to_del
,
6880 struct inode
*inode
,
6881 struct buffer_head
*fe_bh
,
6883 struct ocfs2_truncate_context
*tc
,
6884 struct ocfs2_path
*path
,
6885 struct ocfs2_alloc_context
*meta_ac
)
6888 struct ocfs2_dinode
*fe
;
6889 struct ocfs2_extent_block
*last_eb
= NULL
;
6890 struct ocfs2_extent_list
*el
;
6891 struct buffer_head
*last_eb_bh
= NULL
;
6895 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
6897 status
= ocfs2_find_new_last_ext_blk(inode
, clusters_to_del
,
6905 * Each component will be touched, so we might as well journal
6906 * here to avoid having to handle errors later.
6908 status
= ocfs2_journal_access_path(INODE_CACHE(inode
), handle
, path
);
6915 status
= ocfs2_journal_access_eb(handle
, INODE_CACHE(inode
), last_eb_bh
,
6916 OCFS2_JOURNAL_ACCESS_WRITE
);
6922 last_eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
6925 el
= &(fe
->id2
.i_list
);
6928 * Lower levels depend on this never happening, but it's best
6929 * to check it up here before changing the tree.
6931 if (el
->l_tree_depth
&& el
->l_recs
[0].e_int_clusters
== 0) {
6932 ocfs2_error(inode
->i_sb
,
6933 "Inode %lu has an empty extent record, depth %u\n",
6934 inode
->i_ino
, le16_to_cpu(el
->l_tree_depth
));
6939 dquot_free_space_nodirty(inode
,
6940 ocfs2_clusters_to_bytes(osb
->sb
, clusters_to_del
));
6941 spin_lock(&OCFS2_I(inode
)->ip_lock
);
6942 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
) -
6944 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
6945 le32_add_cpu(&fe
->i_clusters
, -clusters_to_del
);
6946 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
6948 status
= ocfs2_trim_tree(inode
, path
, handle
, tc
,
6949 clusters_to_del
, &delete_blk
, &rec_flags
);
6955 if (le32_to_cpu(fe
->i_clusters
) == 0) {
6956 /* trunc to zero is a special case. */
6957 el
->l_tree_depth
= 0;
6958 fe
->i_last_eb_blk
= 0;
6960 fe
->i_last_eb_blk
= last_eb
->h_blkno
;
6962 status
= ocfs2_journal_dirty(handle
, fe_bh
);
6969 /* If there will be a new last extent block, then by
6970 * definition, there cannot be any leaves to the right of
6972 last_eb
->h_next_leaf_blk
= 0;
6973 status
= ocfs2_journal_dirty(handle
, last_eb_bh
);
6981 if (rec_flags
& OCFS2_EXT_REFCOUNTED
)
6982 status
= ocfs2_decrease_refcount(inode
, handle
,
6983 ocfs2_blocks_to_clusters(osb
->sb
,
6985 clusters_to_del
, meta_ac
,
6986 &tc
->tc_dealloc
, 1);
6988 status
= ocfs2_truncate_log_append(osb
, handle
,
7003 static int ocfs2_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
7005 set_buffer_uptodate(bh
);
7006 mark_buffer_dirty(bh
);
7010 void ocfs2_map_and_dirty_page(struct inode
*inode
, handle_t
*handle
,
7011 unsigned int from
, unsigned int to
,
7012 struct page
*page
, int zero
, u64
*phys
)
7014 int ret
, partial
= 0;
7016 ret
= ocfs2_map_page_blocks(page
, phys
, inode
, from
, to
, 0);
7021 zero_user_segment(page
, from
, to
);
7024 * Need to set the buffers we zero'd into uptodate
7025 * here if they aren't - ocfs2_map_page_blocks()
7026 * might've skipped some
7028 ret
= walk_page_buffers(handle
, page_buffers(page
),
7033 else if (ocfs2_should_order_data(inode
)) {
7034 ret
= ocfs2_jbd2_file_inode(handle
, inode
);
7040 SetPageUptodate(page
);
7042 flush_dcache_page(page
);
7045 static void ocfs2_zero_cluster_pages(struct inode
*inode
, loff_t start
,
7046 loff_t end
, struct page
**pages
,
7047 int numpages
, u64 phys
, handle_t
*handle
)
7051 unsigned int from
, to
= PAGE_CACHE_SIZE
;
7052 struct super_block
*sb
= inode
->i_sb
;
7054 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb
)));
7059 to
= PAGE_CACHE_SIZE
;
7060 for(i
= 0; i
< numpages
; i
++) {
7063 from
= start
& (PAGE_CACHE_SIZE
- 1);
7064 if ((end
>> PAGE_CACHE_SHIFT
) == page
->index
)
7065 to
= end
& (PAGE_CACHE_SIZE
- 1);
7067 BUG_ON(from
> PAGE_CACHE_SIZE
);
7068 BUG_ON(to
> PAGE_CACHE_SIZE
);
7070 ocfs2_map_and_dirty_page(inode
, handle
, from
, to
, page
, 1,
7073 start
= (page
->index
+ 1) << PAGE_CACHE_SHIFT
;
7077 ocfs2_unlock_and_free_pages(pages
, numpages
);
7080 int ocfs2_grab_pages(struct inode
*inode
, loff_t start
, loff_t end
,
7081 struct page
**pages
, int *num
)
7083 int numpages
, ret
= 0;
7084 struct address_space
*mapping
= inode
->i_mapping
;
7085 unsigned long index
;
7086 loff_t last_page_bytes
;
7088 BUG_ON(start
> end
);
7091 last_page_bytes
= PAGE_ALIGN(end
);
7092 index
= start
>> PAGE_CACHE_SHIFT
;
7094 pages
[numpages
] = grab_cache_page(mapping
, index
);
7095 if (!pages
[numpages
]) {
7103 } while (index
< (last_page_bytes
>> PAGE_CACHE_SHIFT
));
7108 ocfs2_unlock_and_free_pages(pages
, numpages
);
7117 static int ocfs2_grab_eof_pages(struct inode
*inode
, loff_t start
, loff_t end
,
7118 struct page
**pages
, int *num
)
7120 struct super_block
*sb
= inode
->i_sb
;
7122 BUG_ON(start
>> OCFS2_SB(sb
)->s_clustersize_bits
!=
7123 (end
- 1) >> OCFS2_SB(sb
)->s_clustersize_bits
);
7125 return ocfs2_grab_pages(inode
, start
, end
, pages
, num
);
7129 * Zero the area past i_size but still within an allocated
7130 * cluster. This avoids exposing nonzero data on subsequent file
7133 * We need to call this before i_size is updated on the inode because
7134 * otherwise block_write_full_page() will skip writeout of pages past
7135 * i_size. The new_i_size parameter is passed for this reason.
7137 int ocfs2_zero_range_for_truncate(struct inode
*inode
, handle_t
*handle
,
7138 u64 range_start
, u64 range_end
)
7140 int ret
= 0, numpages
;
7141 struct page
**pages
= NULL
;
7143 unsigned int ext_flags
;
7144 struct super_block
*sb
= inode
->i_sb
;
7147 * File systems which don't support sparse files zero on every
7150 if (!ocfs2_sparse_alloc(OCFS2_SB(sb
)))
7153 pages
= kcalloc(ocfs2_pages_per_cluster(sb
),
7154 sizeof(struct page
*), GFP_NOFS
);
7155 if (pages
== NULL
) {
7161 if (range_start
== range_end
)
7164 ret
= ocfs2_extent_map_get_blocks(inode
,
7165 range_start
>> sb
->s_blocksize_bits
,
7166 &phys
, NULL
, &ext_flags
);
7173 * Tail is a hole, or is marked unwritten. In either case, we
7174 * can count on read and write to return/push zero's.
7176 if (phys
== 0 || ext_flags
& OCFS2_EXT_UNWRITTEN
)
7179 ret
= ocfs2_grab_eof_pages(inode
, range_start
, range_end
, pages
,
7186 ocfs2_zero_cluster_pages(inode
, range_start
, range_end
, pages
,
7187 numpages
, phys
, handle
);
7190 * Initiate writeout of the pages we zero'd here. We don't
7191 * wait on them - the truncate_inode_pages() call later will
7194 ret
= filemap_fdatawrite_range(inode
->i_mapping
, range_start
,
7206 static void ocfs2_zero_dinode_id2_with_xattr(struct inode
*inode
,
7207 struct ocfs2_dinode
*di
)
7209 unsigned int blocksize
= 1 << inode
->i_sb
->s_blocksize_bits
;
7210 unsigned int xattrsize
= le16_to_cpu(di
->i_xattr_inline_size
);
7212 if (le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_XATTR_FL
)
7213 memset(&di
->id2
, 0, blocksize
-
7214 offsetof(struct ocfs2_dinode
, id2
) -
7217 memset(&di
->id2
, 0, blocksize
-
7218 offsetof(struct ocfs2_dinode
, id2
));
7221 void ocfs2_dinode_new_extent_list(struct inode
*inode
,
7222 struct ocfs2_dinode
*di
)
7224 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7225 di
->id2
.i_list
.l_tree_depth
= 0;
7226 di
->id2
.i_list
.l_next_free_rec
= 0;
7227 di
->id2
.i_list
.l_count
= cpu_to_le16(
7228 ocfs2_extent_recs_per_inode_with_xattr(inode
->i_sb
, di
));
7231 void ocfs2_set_inode_data_inline(struct inode
*inode
, struct ocfs2_dinode
*di
)
7233 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7234 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7236 spin_lock(&oi
->ip_lock
);
7237 oi
->ip_dyn_features
|= OCFS2_INLINE_DATA_FL
;
7238 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7239 spin_unlock(&oi
->ip_lock
);
7242 * We clear the entire i_data structure here so that all
7243 * fields can be properly initialized.
7245 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7247 idata
->id_count
= cpu_to_le16(
7248 ocfs2_max_inline_data_with_xattr(inode
->i_sb
, di
));
7251 int ocfs2_convert_inline_data_to_extents(struct inode
*inode
,
7252 struct buffer_head
*di_bh
)
7254 int ret
, i
, has_data
, num_pages
= 0;
7256 u64
uninitialized_var(block
);
7257 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7258 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7259 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7260 struct ocfs2_alloc_context
*data_ac
= NULL
;
7261 struct page
**pages
= NULL
;
7262 loff_t end
= osb
->s_clustersize
;
7263 struct ocfs2_extent_tree et
;
7266 has_data
= i_size_read(inode
) ? 1 : 0;
7269 pages
= kcalloc(ocfs2_pages_per_cluster(osb
->sb
),
7270 sizeof(struct page
*), GFP_NOFS
);
7271 if (pages
== NULL
) {
7277 ret
= ocfs2_reserve_clusters(osb
, 1, &data_ac
);
7284 handle
= ocfs2_start_trans(osb
,
7285 ocfs2_inline_to_extents_credits(osb
->sb
));
7286 if (IS_ERR(handle
)) {
7287 ret
= PTR_ERR(handle
);
7292 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7293 OCFS2_JOURNAL_ACCESS_WRITE
);
7301 unsigned int page_end
;
7304 ret
= dquot_alloc_space_nodirty(inode
,
7305 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7310 ret
= ocfs2_claim_clusters(osb
, handle
, data_ac
, 1, &bit_off
,
7318 * Save two copies, one for insert, and one that can
7319 * be changed by ocfs2_map_and_dirty_page() below.
7321 block
= phys
= ocfs2_clusters_to_blocks(inode
->i_sb
, bit_off
);
7324 * Non sparse file systems zero on extend, so no need
7327 if (!ocfs2_sparse_alloc(osb
) &&
7328 PAGE_CACHE_SIZE
< osb
->s_clustersize
)
7329 end
= PAGE_CACHE_SIZE
;
7331 ret
= ocfs2_grab_eof_pages(inode
, 0, end
, pages
, &num_pages
);
7338 * This should populate the 1st page for us and mark
7341 ret
= ocfs2_read_inline_data(inode
, pages
[0], di_bh
);
7347 page_end
= PAGE_CACHE_SIZE
;
7348 if (PAGE_CACHE_SIZE
> osb
->s_clustersize
)
7349 page_end
= osb
->s_clustersize
;
7351 for (i
= 0; i
< num_pages
; i
++)
7352 ocfs2_map_and_dirty_page(inode
, handle
, 0, page_end
,
7353 pages
[i
], i
> 0, &phys
);
7356 spin_lock(&oi
->ip_lock
);
7357 oi
->ip_dyn_features
&= ~OCFS2_INLINE_DATA_FL
;
7358 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7359 spin_unlock(&oi
->ip_lock
);
7361 ocfs2_dinode_new_extent_list(inode
, di
);
7363 ocfs2_journal_dirty(handle
, di_bh
);
7367 * An error at this point should be extremely rare. If
7368 * this proves to be false, we could always re-build
7369 * the in-inode data from our pages.
7371 ocfs2_init_dinode_extent_tree(&et
, INODE_CACHE(inode
), di_bh
);
7372 ret
= ocfs2_insert_extent(handle
, &et
, 0, block
, 1, 0, NULL
);
7378 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7382 if (ret
< 0 && did_quota
)
7383 dquot_free_space_nodirty(inode
,
7384 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7386 ocfs2_commit_trans(osb
, handle
);
7390 ocfs2_free_alloc_context(data_ac
);
7394 ocfs2_unlock_and_free_pages(pages
, num_pages
);
7402 * It is expected, that by the time you call this function,
7403 * inode->i_size and fe->i_size have been adjusted.
7405 * WARNING: This will kfree the truncate context
7407 int ocfs2_commit_truncate(struct ocfs2_super
*osb
,
7408 struct inode
*inode
,
7409 struct buffer_head
*fe_bh
,
7410 struct ocfs2_truncate_context
*tc
)
7412 int status
, i
, credits
, tl_sem
= 0;
7413 u32 clusters_to_del
, new_highest_cpos
, range
;
7415 struct ocfs2_extent_list
*el
;
7416 handle_t
*handle
= NULL
;
7417 struct inode
*tl_inode
= osb
->osb_tl_inode
;
7418 struct ocfs2_path
*path
= NULL
;
7419 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)fe_bh
->b_data
;
7420 struct ocfs2_alloc_context
*meta_ac
= NULL
;
7421 struct ocfs2_refcount_tree
*ref_tree
= NULL
;
7425 new_highest_cpos
= ocfs2_clusters_for_bytes(osb
->sb
,
7426 i_size_read(inode
));
7428 path
= ocfs2_new_path(fe_bh
, &di
->id2
.i_list
,
7429 ocfs2_journal_access_di
);
7436 ocfs2_extent_map_trunc(inode
, new_highest_cpos
);
7440 * Check that we still have allocation to delete.
7442 if (OCFS2_I(inode
)->ip_clusters
== 0) {
7450 * Truncate always works against the rightmost tree branch.
7452 status
= ocfs2_find_path(INODE_CACHE(inode
), path
, UINT_MAX
);
7458 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
7459 OCFS2_I(inode
)->ip_clusters
, path
->p_tree_depth
);
7462 * By now, el will point to the extent list on the bottom most
7463 * portion of this tree. Only the tail record is considered in
7466 * We handle the following cases, in order:
7467 * - empty extent: delete the remaining branch
7468 * - remove the entire record
7469 * - remove a partial record
7470 * - no record needs to be removed (truncate has completed)
7472 el
= path_leaf_el(path
);
7473 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
7474 ocfs2_error(inode
->i_sb
,
7475 "Inode %llu has empty extent block at %llu\n",
7476 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7477 (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
7482 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
7483 range
= le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
7484 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
7485 if (i
== 0 && ocfs2_is_empty_extent(&el
->l_recs
[i
])) {
7486 clusters_to_del
= 0;
7487 } else if (le32_to_cpu(el
->l_recs
[i
].e_cpos
) >= new_highest_cpos
) {
7488 clusters_to_del
= ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
7489 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
7490 } else if (range
> new_highest_cpos
) {
7491 clusters_to_del
= (ocfs2_rec_clusters(el
, &el
->l_recs
[i
]) +
7492 le32_to_cpu(el
->l_recs
[i
].e_cpos
)) -
7494 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
) +
7495 ocfs2_clusters_to_blocks(inode
->i_sb
,
7496 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]) -
7503 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
7504 clusters_to_del
, (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
7506 if (el
->l_recs
[i
].e_flags
& OCFS2_EXT_REFCOUNTED
&& clusters_to_del
) {
7507 BUG_ON(!(OCFS2_I(inode
)->ip_dyn_features
&
7508 OCFS2_HAS_REFCOUNT_FL
));
7510 status
= ocfs2_lock_refcount_tree(osb
,
7511 le64_to_cpu(di
->i_refcount_loc
),
7512 1, &ref_tree
, NULL
);
7518 status
= ocfs2_prepare_refcount_change_for_del(inode
, fe_bh
,
7529 mutex_lock(&tl_inode
->i_mutex
);
7531 /* ocfs2_truncate_log_needs_flush guarantees us at least one
7532 * record is free for use. If there isn't any, we flush to get
7533 * an empty truncate log. */
7534 if (ocfs2_truncate_log_needs_flush(osb
)) {
7535 status
= __ocfs2_flush_truncate_log(osb
);
7542 credits
+= ocfs2_calc_tree_trunc_credits(osb
->sb
, clusters_to_del
,
7543 (struct ocfs2_dinode
*)fe_bh
->b_data
,
7545 handle
= ocfs2_start_trans(osb
, credits
);
7546 if (IS_ERR(handle
)) {
7547 status
= PTR_ERR(handle
);
7553 status
= ocfs2_do_truncate(osb
, clusters_to_del
, inode
, fe_bh
, handle
,
7560 mutex_unlock(&tl_inode
->i_mutex
);
7563 ocfs2_commit_trans(osb
, handle
);
7566 ocfs2_reinit_path(path
, 1);
7569 ocfs2_free_alloc_context(meta_ac
);
7574 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
7579 * The check above will catch the case where we've truncated
7580 * away all allocation.
7586 ocfs2_schedule_truncate_log_flush(osb
, 1);
7589 mutex_unlock(&tl_inode
->i_mutex
);
7592 ocfs2_commit_trans(osb
, handle
);
7595 ocfs2_free_alloc_context(meta_ac
);
7598 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
7600 ocfs2_run_deallocs(osb
, &tc
->tc_dealloc
);
7602 ocfs2_free_path(path
);
7604 /* This will drop the ext_alloc cluster lock for us */
7605 ocfs2_free_truncate_context(tc
);
7612 * Expects the inode to already be locked.
7614 int ocfs2_prepare_truncate(struct ocfs2_super
*osb
,
7615 struct inode
*inode
,
7616 struct buffer_head
*fe_bh
,
7617 struct ocfs2_truncate_context
**tc
)
7620 unsigned int new_i_clusters
;
7621 struct ocfs2_dinode
*fe
;
7622 struct ocfs2_extent_block
*eb
;
7623 struct buffer_head
*last_eb_bh
= NULL
;
7629 new_i_clusters
= ocfs2_clusters_for_bytes(osb
->sb
,
7630 i_size_read(inode
));
7631 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
7633 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
7634 "%llu\n", le32_to_cpu(fe
->i_clusters
), new_i_clusters
,
7635 (unsigned long long)le64_to_cpu(fe
->i_size
));
7637 *tc
= kzalloc(sizeof(struct ocfs2_truncate_context
), GFP_KERNEL
);
7643 ocfs2_init_dealloc_ctxt(&(*tc
)->tc_dealloc
);
7645 if (fe
->id2
.i_list
.l_tree_depth
) {
7646 status
= ocfs2_read_extent_block(INODE_CACHE(inode
),
7647 le64_to_cpu(fe
->i_last_eb_blk
),
7653 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
7656 (*tc
)->tc_last_eb_bh
= last_eb_bh
;
7662 ocfs2_free_truncate_context(*tc
);
7670 * 'start' is inclusive, 'end' is not.
7672 int ocfs2_truncate_inline(struct inode
*inode
, struct buffer_head
*di_bh
,
7673 unsigned int start
, unsigned int end
, int trunc
)
7676 unsigned int numbytes
;
7678 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7679 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7680 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7682 if (end
> i_size_read(inode
))
7683 end
= i_size_read(inode
);
7685 BUG_ON(start
>= end
);
7687 if (!(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
) ||
7688 !(le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_DATA_FL
) ||
7689 !ocfs2_supports_inline_data(osb
)) {
7690 ocfs2_error(inode
->i_sb
,
7691 "Inline data flags for inode %llu don't agree! "
7692 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7693 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7694 le16_to_cpu(di
->i_dyn_features
),
7695 OCFS2_I(inode
)->ip_dyn_features
,
7696 osb
->s_feature_incompat
);
7701 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
7702 if (IS_ERR(handle
)) {
7703 ret
= PTR_ERR(handle
);
7708 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7709 OCFS2_JOURNAL_ACCESS_WRITE
);
7715 numbytes
= end
- start
;
7716 memset(idata
->id_data
+ start
, 0, numbytes
);
7719 * No need to worry about the data page here - it's been
7720 * truncated already and inline data doesn't need it for
7721 * pushing zero's to disk, so we'll let readpage pick it up
7725 i_size_write(inode
, start
);
7726 di
->i_size
= cpu_to_le64(start
);
7729 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7730 inode
->i_ctime
= inode
->i_mtime
= CURRENT_TIME
;
7732 di
->i_ctime
= di
->i_mtime
= cpu_to_le64(inode
->i_ctime
.tv_sec
);
7733 di
->i_ctime_nsec
= di
->i_mtime_nsec
= cpu_to_le32(inode
->i_ctime
.tv_nsec
);
7735 ocfs2_journal_dirty(handle
, di_bh
);
7738 ocfs2_commit_trans(osb
, handle
);
7744 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
)
7747 * The caller is responsible for completing deallocation
7748 * before freeing the context.
7750 if (tc
->tc_dealloc
.c_first_suballocator
!= NULL
)
7752 "Truncate completion has non-empty dealloc context\n");
7754 brelse(tc
->tc_last_eb_bh
);