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 ocfs2_journal_dirty(handle
, bhs
[i
]);
1073 for(i
= 0; i
< wanted
; i
++) {
1083 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1085 * Returns the sum of the rightmost extent rec logical offset and
1088 * ocfs2_add_branch() uses this to determine what logical cluster
1089 * value should be populated into the leftmost new branch records.
1091 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1092 * value for the new topmost tree record.
1094 static inline u32
ocfs2_sum_rightmost_rec(struct ocfs2_extent_list
*el
)
1098 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1100 return le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
1101 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
1105 * Change range of the branches in the right most path according to the leaf
1106 * extent block's rightmost record.
1108 static int ocfs2_adjust_rightmost_branch(handle_t
*handle
,
1109 struct ocfs2_extent_tree
*et
)
1112 struct ocfs2_path
*path
= NULL
;
1113 struct ocfs2_extent_list
*el
;
1114 struct ocfs2_extent_rec
*rec
;
1116 path
= ocfs2_new_path_from_et(et
);
1122 status
= ocfs2_find_path(et
->et_ci
, path
, UINT_MAX
);
1128 status
= ocfs2_extend_trans(handle
, path_num_items(path
) +
1129 handle
->h_buffer_credits
);
1135 status
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
1141 el
= path_leaf_el(path
);
1142 rec
= &el
->l_recs
[le32_to_cpu(el
->l_next_free_rec
) - 1];
1144 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
1147 ocfs2_free_path(path
);
1152 * Add an entire tree branch to our inode. eb_bh is the extent block
1153 * to start at, if we don't want to start the branch at the root
1156 * last_eb_bh is required as we have to update it's next_leaf pointer
1157 * for the new last extent block.
1159 * the new branch will be 'empty' in the sense that every block will
1160 * contain a single record with cluster count == 0.
1162 static int ocfs2_add_branch(handle_t
*handle
,
1163 struct ocfs2_extent_tree
*et
,
1164 struct buffer_head
*eb_bh
,
1165 struct buffer_head
**last_eb_bh
,
1166 struct ocfs2_alloc_context
*meta_ac
)
1168 int status
, new_blocks
, i
;
1169 u64 next_blkno
, new_last_eb_blk
;
1170 struct buffer_head
*bh
;
1171 struct buffer_head
**new_eb_bhs
= NULL
;
1172 struct ocfs2_extent_block
*eb
;
1173 struct ocfs2_extent_list
*eb_el
;
1174 struct ocfs2_extent_list
*el
;
1175 u32 new_cpos
, root_end
;
1179 BUG_ON(!last_eb_bh
|| !*last_eb_bh
);
1182 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
1185 el
= et
->et_root_el
;
1187 /* we never add a branch to a leaf. */
1188 BUG_ON(!el
->l_tree_depth
);
1190 new_blocks
= le16_to_cpu(el
->l_tree_depth
);
1192 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
1193 new_cpos
= ocfs2_sum_rightmost_rec(&eb
->h_list
);
1194 root_end
= ocfs2_sum_rightmost_rec(et
->et_root_el
);
1197 * If there is a gap before the root end and the real end
1198 * of the righmost leaf block, we need to remove the gap
1199 * between new_cpos and root_end first so that the tree
1200 * is consistent after we add a new branch(it will start
1203 if (root_end
> new_cpos
) {
1204 mlog(0, "adjust the cluster end from %u to %u\n",
1205 root_end
, new_cpos
);
1206 status
= ocfs2_adjust_rightmost_branch(handle
, et
);
1213 /* allocate the number of new eb blocks we need */
1214 new_eb_bhs
= kcalloc(new_blocks
, sizeof(struct buffer_head
*),
1222 status
= ocfs2_create_new_meta_bhs(handle
, et
, new_blocks
,
1223 meta_ac
, new_eb_bhs
);
1229 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1230 * linked with the rest of the tree.
1231 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1233 * when we leave the loop, new_last_eb_blk will point to the
1234 * newest leaf, and next_blkno will point to the topmost extent
1236 next_blkno
= new_last_eb_blk
= 0;
1237 for(i
= 0; i
< new_blocks
; i
++) {
1239 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1240 /* ocfs2_create_new_meta_bhs() should create it right! */
1241 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1242 eb_el
= &eb
->h_list
;
1244 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, bh
,
1245 OCFS2_JOURNAL_ACCESS_CREATE
);
1251 eb
->h_next_leaf_blk
= 0;
1252 eb_el
->l_tree_depth
= cpu_to_le16(i
);
1253 eb_el
->l_next_free_rec
= cpu_to_le16(1);
1255 * This actually counts as an empty extent as
1258 eb_el
->l_recs
[0].e_cpos
= cpu_to_le32(new_cpos
);
1259 eb_el
->l_recs
[0].e_blkno
= cpu_to_le64(next_blkno
);
1261 * eb_el isn't always an interior node, but even leaf
1262 * nodes want a zero'd flags and reserved field so
1263 * this gets the whole 32 bits regardless of use.
1265 eb_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(0);
1266 if (!eb_el
->l_tree_depth
)
1267 new_last_eb_blk
= le64_to_cpu(eb
->h_blkno
);
1269 ocfs2_journal_dirty(handle
, bh
);
1270 next_blkno
= le64_to_cpu(eb
->h_blkno
);
1273 /* This is a bit hairy. We want to update up to three blocks
1274 * here without leaving any of them in an inconsistent state
1275 * in case of error. We don't have to worry about
1276 * journal_dirty erroring as it won't unless we've aborted the
1277 * handle (in which case we would never be here) so reserving
1278 * the write with journal_access is all we need to do. */
1279 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, *last_eb_bh
,
1280 OCFS2_JOURNAL_ACCESS_WRITE
);
1285 status
= ocfs2_et_root_journal_access(handle
, et
,
1286 OCFS2_JOURNAL_ACCESS_WRITE
);
1292 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, eb_bh
,
1293 OCFS2_JOURNAL_ACCESS_WRITE
);
1300 /* Link the new branch into the rest of the tree (el will
1301 * either be on the root_bh, or the extent block passed in. */
1302 i
= le16_to_cpu(el
->l_next_free_rec
);
1303 el
->l_recs
[i
].e_blkno
= cpu_to_le64(next_blkno
);
1304 el
->l_recs
[i
].e_cpos
= cpu_to_le32(new_cpos
);
1305 el
->l_recs
[i
].e_int_clusters
= 0;
1306 le16_add_cpu(&el
->l_next_free_rec
, 1);
1308 /* fe needs a new last extent block pointer, as does the
1309 * next_leaf on the previously last-extent-block. */
1310 ocfs2_et_set_last_eb_blk(et
, new_last_eb_blk
);
1312 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
1313 eb
->h_next_leaf_blk
= cpu_to_le64(new_last_eb_blk
);
1315 ocfs2_journal_dirty(handle
, *last_eb_bh
);
1316 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1318 ocfs2_journal_dirty(handle
, eb_bh
);
1321 * Some callers want to track the rightmost leaf so pass it
1324 brelse(*last_eb_bh
);
1325 get_bh(new_eb_bhs
[0]);
1326 *last_eb_bh
= new_eb_bhs
[0];
1331 for (i
= 0; i
< new_blocks
; i
++)
1332 brelse(new_eb_bhs
[i
]);
1341 * adds another level to the allocation tree.
1342 * returns back the new extent block so you can add a branch to it
1345 static int ocfs2_shift_tree_depth(handle_t
*handle
,
1346 struct ocfs2_extent_tree
*et
,
1347 struct ocfs2_alloc_context
*meta_ac
,
1348 struct buffer_head
**ret_new_eb_bh
)
1352 struct buffer_head
*new_eb_bh
= NULL
;
1353 struct ocfs2_extent_block
*eb
;
1354 struct ocfs2_extent_list
*root_el
;
1355 struct ocfs2_extent_list
*eb_el
;
1359 status
= ocfs2_create_new_meta_bhs(handle
, et
, 1, meta_ac
,
1366 eb
= (struct ocfs2_extent_block
*) new_eb_bh
->b_data
;
1367 /* ocfs2_create_new_meta_bhs() should create it right! */
1368 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1370 eb_el
= &eb
->h_list
;
1371 root_el
= et
->et_root_el
;
1373 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, new_eb_bh
,
1374 OCFS2_JOURNAL_ACCESS_CREATE
);
1380 /* copy the root extent list data into the new extent block */
1381 eb_el
->l_tree_depth
= root_el
->l_tree_depth
;
1382 eb_el
->l_next_free_rec
= root_el
->l_next_free_rec
;
1383 for (i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1384 eb_el
->l_recs
[i
] = root_el
->l_recs
[i
];
1386 ocfs2_journal_dirty(handle
, new_eb_bh
);
1388 status
= ocfs2_et_root_journal_access(handle
, et
,
1389 OCFS2_JOURNAL_ACCESS_WRITE
);
1395 new_clusters
= ocfs2_sum_rightmost_rec(eb_el
);
1397 /* update root_bh now */
1398 le16_add_cpu(&root_el
->l_tree_depth
, 1);
1399 root_el
->l_recs
[0].e_cpos
= 0;
1400 root_el
->l_recs
[0].e_blkno
= eb
->h_blkno
;
1401 root_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(new_clusters
);
1402 for (i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1403 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
1404 root_el
->l_next_free_rec
= cpu_to_le16(1);
1406 /* If this is our 1st tree depth shift, then last_eb_blk
1407 * becomes the allocated extent block */
1408 if (root_el
->l_tree_depth
== cpu_to_le16(1))
1409 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
1411 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1413 *ret_new_eb_bh
= new_eb_bh
;
1424 * Should only be called when there is no space left in any of the
1425 * leaf nodes. What we want to do is find the lowest tree depth
1426 * non-leaf extent block with room for new records. There are three
1427 * valid results of this search:
1429 * 1) a lowest extent block is found, then we pass it back in
1430 * *lowest_eb_bh and return '0'
1432 * 2) the search fails to find anything, but the root_el has room. We
1433 * pass NULL back in *lowest_eb_bh, but still return '0'
1435 * 3) the search fails to find anything AND the root_el is full, in
1436 * which case we return > 0
1438 * return status < 0 indicates an error.
1440 static int ocfs2_find_branch_target(struct ocfs2_extent_tree
*et
,
1441 struct buffer_head
**target_bh
)
1445 struct ocfs2_extent_block
*eb
;
1446 struct ocfs2_extent_list
*el
;
1447 struct buffer_head
*bh
= NULL
;
1448 struct buffer_head
*lowest_bh
= NULL
;
1454 el
= et
->et_root_el
;
1456 while(le16_to_cpu(el
->l_tree_depth
) > 1) {
1457 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1458 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1459 "Owner %llu has empty "
1460 "extent list (next_free_rec == 0)",
1461 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
1465 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1466 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1468 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1469 "Owner %llu has extent "
1470 "list where extent # %d has no physical "
1472 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
), i
);
1480 status
= ocfs2_read_extent_block(et
->et_ci
, blkno
, &bh
);
1486 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1489 if (le16_to_cpu(el
->l_next_free_rec
) <
1490 le16_to_cpu(el
->l_count
)) {
1497 /* If we didn't find one and the fe doesn't have any room,
1498 * then return '1' */
1499 el
= et
->et_root_el
;
1500 if (!lowest_bh
&& (el
->l_next_free_rec
== el
->l_count
))
1503 *target_bh
= lowest_bh
;
1512 * Grow a b-tree so that it has more records.
1514 * We might shift the tree depth in which case existing paths should
1515 * be considered invalid.
1517 * Tree depth after the grow is returned via *final_depth.
1519 * *last_eb_bh will be updated by ocfs2_add_branch().
1521 static int ocfs2_grow_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
1522 int *final_depth
, struct buffer_head
**last_eb_bh
,
1523 struct ocfs2_alloc_context
*meta_ac
)
1526 struct ocfs2_extent_list
*el
= et
->et_root_el
;
1527 int depth
= le16_to_cpu(el
->l_tree_depth
);
1528 struct buffer_head
*bh
= NULL
;
1530 BUG_ON(meta_ac
== NULL
);
1532 shift
= ocfs2_find_branch_target(et
, &bh
);
1539 /* We traveled all the way to the bottom of the allocation tree
1540 * and didn't find room for any more extents - we need to add
1541 * another tree level */
1544 mlog(0, "need to shift tree depth (current = %d)\n", depth
);
1546 /* ocfs2_shift_tree_depth will return us a buffer with
1547 * the new extent block (so we can pass that to
1548 * ocfs2_add_branch). */
1549 ret
= ocfs2_shift_tree_depth(handle
, et
, meta_ac
, &bh
);
1557 * Special case: we have room now if we shifted from
1558 * tree_depth 0, so no more work needs to be done.
1560 * We won't be calling add_branch, so pass
1561 * back *last_eb_bh as the new leaf. At depth
1562 * zero, it should always be null so there's
1563 * no reason to brelse.
1565 BUG_ON(*last_eb_bh
);
1572 /* call ocfs2_add_branch to add the final part of the tree with
1574 mlog(0, "add branch. bh = %p\n", bh
);
1575 ret
= ocfs2_add_branch(handle
, et
, bh
, last_eb_bh
,
1584 *final_depth
= depth
;
1590 * This function will discard the rightmost extent record.
1592 static void ocfs2_shift_records_right(struct ocfs2_extent_list
*el
)
1594 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1595 int count
= le16_to_cpu(el
->l_count
);
1596 unsigned int num_bytes
;
1599 /* This will cause us to go off the end of our extent list. */
1600 BUG_ON(next_free
>= count
);
1602 num_bytes
= sizeof(struct ocfs2_extent_rec
) * next_free
;
1604 memmove(&el
->l_recs
[1], &el
->l_recs
[0], num_bytes
);
1607 static void ocfs2_rotate_leaf(struct ocfs2_extent_list
*el
,
1608 struct ocfs2_extent_rec
*insert_rec
)
1610 int i
, insert_index
, next_free
, has_empty
, num_bytes
;
1611 u32 insert_cpos
= le32_to_cpu(insert_rec
->e_cpos
);
1612 struct ocfs2_extent_rec
*rec
;
1614 next_free
= le16_to_cpu(el
->l_next_free_rec
);
1615 has_empty
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
1619 /* The tree code before us didn't allow enough room in the leaf. */
1620 BUG_ON(el
->l_next_free_rec
== el
->l_count
&& !has_empty
);
1623 * The easiest way to approach this is to just remove the
1624 * empty extent and temporarily decrement next_free.
1628 * If next_free was 1 (only an empty extent), this
1629 * loop won't execute, which is fine. We still want
1630 * the decrement above to happen.
1632 for(i
= 0; i
< (next_free
- 1); i
++)
1633 el
->l_recs
[i
] = el
->l_recs
[i
+1];
1639 * Figure out what the new record index should be.
1641 for(i
= 0; i
< next_free
; i
++) {
1642 rec
= &el
->l_recs
[i
];
1644 if (insert_cpos
< le32_to_cpu(rec
->e_cpos
))
1649 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1650 insert_cpos
, insert_index
, has_empty
, next_free
, le16_to_cpu(el
->l_count
));
1652 BUG_ON(insert_index
< 0);
1653 BUG_ON(insert_index
>= le16_to_cpu(el
->l_count
));
1654 BUG_ON(insert_index
> next_free
);
1657 * No need to memmove if we're just adding to the tail.
1659 if (insert_index
!= next_free
) {
1660 BUG_ON(next_free
>= le16_to_cpu(el
->l_count
));
1662 num_bytes
= next_free
- insert_index
;
1663 num_bytes
*= sizeof(struct ocfs2_extent_rec
);
1664 memmove(&el
->l_recs
[insert_index
+ 1],
1665 &el
->l_recs
[insert_index
],
1670 * Either we had an empty extent, and need to re-increment or
1671 * there was no empty extent on a non full rightmost leaf node,
1672 * in which case we still need to increment.
1675 el
->l_next_free_rec
= cpu_to_le16(next_free
);
1677 * Make sure none of the math above just messed up our tree.
1679 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) > le16_to_cpu(el
->l_count
));
1681 el
->l_recs
[insert_index
] = *insert_rec
;
1685 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list
*el
)
1687 int size
, num_recs
= le16_to_cpu(el
->l_next_free_rec
);
1689 BUG_ON(num_recs
== 0);
1691 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
1693 size
= num_recs
* sizeof(struct ocfs2_extent_rec
);
1694 memmove(&el
->l_recs
[0], &el
->l_recs
[1], size
);
1695 memset(&el
->l_recs
[num_recs
], 0,
1696 sizeof(struct ocfs2_extent_rec
));
1697 el
->l_next_free_rec
= cpu_to_le16(num_recs
);
1702 * Create an empty extent record .
1704 * l_next_free_rec may be updated.
1706 * If an empty extent already exists do nothing.
1708 static void ocfs2_create_empty_extent(struct ocfs2_extent_list
*el
)
1710 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1712 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
1717 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
1720 mlog_bug_on_msg(el
->l_count
== el
->l_next_free_rec
,
1721 "Asked to create an empty extent in a full list:\n"
1722 "count = %u, tree depth = %u",
1723 le16_to_cpu(el
->l_count
),
1724 le16_to_cpu(el
->l_tree_depth
));
1726 ocfs2_shift_records_right(el
);
1729 le16_add_cpu(&el
->l_next_free_rec
, 1);
1730 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1734 * For a rotation which involves two leaf nodes, the "root node" is
1735 * the lowest level tree node which contains a path to both leafs. This
1736 * resulting set of information can be used to form a complete "subtree"
1738 * This function is passed two full paths from the dinode down to a
1739 * pair of adjacent leaves. It's task is to figure out which path
1740 * index contains the subtree root - this can be the root index itself
1741 * in a worst-case rotation.
1743 * The array index of the subtree root is passed back.
1745 int ocfs2_find_subtree_root(struct ocfs2_extent_tree
*et
,
1746 struct ocfs2_path
*left
,
1747 struct ocfs2_path
*right
)
1752 * Check that the caller passed in two paths from the same tree.
1754 BUG_ON(path_root_bh(left
) != path_root_bh(right
));
1760 * The caller didn't pass two adjacent paths.
1762 mlog_bug_on_msg(i
> left
->p_tree_depth
,
1763 "Owner %llu, left depth %u, right depth %u\n"
1764 "left leaf blk %llu, right leaf blk %llu\n",
1765 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
1766 left
->p_tree_depth
, right
->p_tree_depth
,
1767 (unsigned long long)path_leaf_bh(left
)->b_blocknr
,
1768 (unsigned long long)path_leaf_bh(right
)->b_blocknr
);
1769 } while (left
->p_node
[i
].bh
->b_blocknr
==
1770 right
->p_node
[i
].bh
->b_blocknr
);
1775 typedef void (path_insert_t
)(void *, struct buffer_head
*);
1778 * Traverse a btree path in search of cpos, starting at root_el.
1780 * This code can be called with a cpos larger than the tree, in which
1781 * case it will return the rightmost path.
1783 static int __ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1784 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1785 path_insert_t
*func
, void *data
)
1790 struct buffer_head
*bh
= NULL
;
1791 struct ocfs2_extent_block
*eb
;
1792 struct ocfs2_extent_list
*el
;
1793 struct ocfs2_extent_rec
*rec
;
1796 while (el
->l_tree_depth
) {
1797 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1798 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1799 "Owner %llu has empty extent list at "
1801 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1802 le16_to_cpu(el
->l_tree_depth
));
1808 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
) - 1; i
++) {
1809 rec
= &el
->l_recs
[i
];
1812 * In the case that cpos is off the allocation
1813 * tree, this should just wind up returning the
1816 range
= le32_to_cpu(rec
->e_cpos
) +
1817 ocfs2_rec_clusters(el
, rec
);
1818 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1822 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1824 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1825 "Owner %llu has bad blkno in extent list "
1826 "at depth %u (index %d)\n",
1827 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1828 le16_to_cpu(el
->l_tree_depth
), i
);
1835 ret
= ocfs2_read_extent_block(ci
, blkno
, &bh
);
1841 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1844 if (le16_to_cpu(el
->l_next_free_rec
) >
1845 le16_to_cpu(el
->l_count
)) {
1846 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1847 "Owner %llu has bad count in extent list "
1848 "at block %llu (next free=%u, count=%u)\n",
1849 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1850 (unsigned long long)bh
->b_blocknr
,
1851 le16_to_cpu(el
->l_next_free_rec
),
1852 le16_to_cpu(el
->l_count
));
1863 * Catch any trailing bh that the loop didn't handle.
1871 * Given an initialized path (that is, it has a valid root extent
1872 * list), this function will traverse the btree in search of the path
1873 * which would contain cpos.
1875 * The path traveled is recorded in the path structure.
1877 * Note that this will not do any comparisons on leaf node extent
1878 * records, so it will work fine in the case that we just added a tree
1881 struct find_path_data
{
1883 struct ocfs2_path
*path
;
1885 static void find_path_ins(void *data
, struct buffer_head
*bh
)
1887 struct find_path_data
*fp
= data
;
1890 ocfs2_path_insert_eb(fp
->path
, fp
->index
, bh
);
1893 int ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1894 struct ocfs2_path
*path
, u32 cpos
)
1896 struct find_path_data data
;
1900 return __ocfs2_find_path(ci
, path_root_el(path
), cpos
,
1901 find_path_ins
, &data
);
1904 static void find_leaf_ins(void *data
, struct buffer_head
*bh
)
1906 struct ocfs2_extent_block
*eb
=(struct ocfs2_extent_block
*)bh
->b_data
;
1907 struct ocfs2_extent_list
*el
= &eb
->h_list
;
1908 struct buffer_head
**ret
= data
;
1910 /* We want to retain only the leaf block. */
1911 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
1917 * Find the leaf block in the tree which would contain cpos. No
1918 * checking of the actual leaf is done.
1920 * Some paths want to call this instead of allocating a path structure
1921 * and calling ocfs2_find_path().
1923 * This function doesn't handle non btree extent lists.
1925 int ocfs2_find_leaf(struct ocfs2_caching_info
*ci
,
1926 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1927 struct buffer_head
**leaf_bh
)
1930 struct buffer_head
*bh
= NULL
;
1932 ret
= __ocfs2_find_path(ci
, root_el
, cpos
, find_leaf_ins
, &bh
);
1944 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1946 * Basically, we've moved stuff around at the bottom of the tree and
1947 * we need to fix up the extent records above the changes to reflect
1950 * left_rec: the record on the left.
1951 * left_child_el: is the child list pointed to by left_rec
1952 * right_rec: the record to the right of left_rec
1953 * right_child_el: is the child list pointed to by right_rec
1955 * By definition, this only works on interior nodes.
1957 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec
*left_rec
,
1958 struct ocfs2_extent_list
*left_child_el
,
1959 struct ocfs2_extent_rec
*right_rec
,
1960 struct ocfs2_extent_list
*right_child_el
)
1962 u32 left_clusters
, right_end
;
1965 * Interior nodes never have holes. Their cpos is the cpos of
1966 * the leftmost record in their child list. Their cluster
1967 * count covers the full theoretical range of their child list
1968 * - the range between their cpos and the cpos of the record
1969 * immediately to their right.
1971 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[0].e_cpos
);
1972 if (!ocfs2_rec_clusters(right_child_el
, &right_child_el
->l_recs
[0])) {
1973 BUG_ON(right_child_el
->l_tree_depth
);
1974 BUG_ON(le16_to_cpu(right_child_el
->l_next_free_rec
) <= 1);
1975 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[1].e_cpos
);
1977 left_clusters
-= le32_to_cpu(left_rec
->e_cpos
);
1978 left_rec
->e_int_clusters
= cpu_to_le32(left_clusters
);
1981 * Calculate the rightmost cluster count boundary before
1982 * moving cpos - we will need to adjust clusters after
1983 * updating e_cpos to keep the same highest cluster count.
1985 right_end
= le32_to_cpu(right_rec
->e_cpos
);
1986 right_end
+= le32_to_cpu(right_rec
->e_int_clusters
);
1988 right_rec
->e_cpos
= left_rec
->e_cpos
;
1989 le32_add_cpu(&right_rec
->e_cpos
, left_clusters
);
1991 right_end
-= le32_to_cpu(right_rec
->e_cpos
);
1992 right_rec
->e_int_clusters
= cpu_to_le32(right_end
);
1996 * Adjust the adjacent root node records involved in a
1997 * rotation. left_el_blkno is passed in as a key so that we can easily
1998 * find it's index in the root list.
2000 static void ocfs2_adjust_root_records(struct ocfs2_extent_list
*root_el
,
2001 struct ocfs2_extent_list
*left_el
,
2002 struct ocfs2_extent_list
*right_el
,
2007 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) <=
2008 le16_to_cpu(left_el
->l_tree_depth
));
2010 for(i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
) - 1; i
++) {
2011 if (le64_to_cpu(root_el
->l_recs
[i
].e_blkno
) == left_el_blkno
)
2016 * The path walking code should have never returned a root and
2017 * two paths which are not adjacent.
2019 BUG_ON(i
>= (le16_to_cpu(root_el
->l_next_free_rec
) - 1));
2021 ocfs2_adjust_adjacent_records(&root_el
->l_recs
[i
], left_el
,
2022 &root_el
->l_recs
[i
+ 1], right_el
);
2026 * We've changed a leaf block (in right_path) and need to reflect that
2027 * change back up the subtree.
2029 * This happens in multiple places:
2030 * - When we've moved an extent record from the left path leaf to the right
2031 * path leaf to make room for an empty extent in the left path leaf.
2032 * - When our insert into the right path leaf is at the leftmost edge
2033 * and requires an update of the path immediately to it's left. This
2034 * can occur at the end of some types of rotation and appending inserts.
2035 * - When we've adjusted the last extent record in the left path leaf and the
2036 * 1st extent record in the right path leaf during cross extent block merge.
2038 static void ocfs2_complete_edge_insert(handle_t
*handle
,
2039 struct ocfs2_path
*left_path
,
2040 struct ocfs2_path
*right_path
,
2044 struct ocfs2_extent_list
*el
, *left_el
, *right_el
;
2045 struct ocfs2_extent_rec
*left_rec
, *right_rec
;
2046 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2049 * Update the counts and position values within all the
2050 * interior nodes to reflect the leaf rotation we just did.
2052 * The root node is handled below the loop.
2054 * We begin the loop with right_el and left_el pointing to the
2055 * leaf lists and work our way up.
2057 * NOTE: within this loop, left_el and right_el always refer
2058 * to the *child* lists.
2060 left_el
= path_leaf_el(left_path
);
2061 right_el
= path_leaf_el(right_path
);
2062 for(i
= left_path
->p_tree_depth
- 1; i
> subtree_index
; i
--) {
2063 mlog(0, "Adjust records at index %u\n", i
);
2066 * One nice property of knowing that all of these
2067 * nodes are below the root is that we only deal with
2068 * the leftmost right node record and the rightmost
2071 el
= left_path
->p_node
[i
].el
;
2072 idx
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2073 left_rec
= &el
->l_recs
[idx
];
2075 el
= right_path
->p_node
[i
].el
;
2076 right_rec
= &el
->l_recs
[0];
2078 ocfs2_adjust_adjacent_records(left_rec
, left_el
, right_rec
,
2081 ocfs2_journal_dirty(handle
, left_path
->p_node
[i
].bh
);
2082 ocfs2_journal_dirty(handle
, right_path
->p_node
[i
].bh
);
2085 * Setup our list pointers now so that the current
2086 * parents become children in the next iteration.
2088 left_el
= left_path
->p_node
[i
].el
;
2089 right_el
= right_path
->p_node
[i
].el
;
2093 * At the root node, adjust the two adjacent records which
2094 * begin our path to the leaves.
2097 el
= left_path
->p_node
[subtree_index
].el
;
2098 left_el
= left_path
->p_node
[subtree_index
+ 1].el
;
2099 right_el
= right_path
->p_node
[subtree_index
+ 1].el
;
2101 ocfs2_adjust_root_records(el
, left_el
, right_el
,
2102 left_path
->p_node
[subtree_index
+ 1].bh
->b_blocknr
);
2104 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2106 ocfs2_journal_dirty(handle
, root_bh
);
2109 static int ocfs2_rotate_subtree_right(handle_t
*handle
,
2110 struct ocfs2_extent_tree
*et
,
2111 struct ocfs2_path
*left_path
,
2112 struct ocfs2_path
*right_path
,
2116 struct buffer_head
*right_leaf_bh
;
2117 struct buffer_head
*left_leaf_bh
= NULL
;
2118 struct buffer_head
*root_bh
;
2119 struct ocfs2_extent_list
*right_el
, *left_el
;
2120 struct ocfs2_extent_rec move_rec
;
2122 left_leaf_bh
= path_leaf_bh(left_path
);
2123 left_el
= path_leaf_el(left_path
);
2125 if (left_el
->l_next_free_rec
!= left_el
->l_count
) {
2126 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
2127 "Inode %llu has non-full interior leaf node %llu"
2129 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2130 (unsigned long long)left_leaf_bh
->b_blocknr
,
2131 le16_to_cpu(left_el
->l_next_free_rec
));
2136 * This extent block may already have an empty record, so we
2137 * return early if so.
2139 if (ocfs2_is_empty_extent(&left_el
->l_recs
[0]))
2142 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2143 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2145 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2152 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2153 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2160 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2168 right_leaf_bh
= path_leaf_bh(right_path
);
2169 right_el
= path_leaf_el(right_path
);
2171 /* This is a code error, not a disk corruption. */
2172 mlog_bug_on_msg(!right_el
->l_next_free_rec
, "Inode %llu: Rotate fails "
2173 "because rightmost leaf block %llu is empty\n",
2174 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2175 (unsigned long long)right_leaf_bh
->b_blocknr
);
2177 ocfs2_create_empty_extent(right_el
);
2179 ocfs2_journal_dirty(handle
, right_leaf_bh
);
2181 /* Do the copy now. */
2182 i
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2183 move_rec
= left_el
->l_recs
[i
];
2184 right_el
->l_recs
[0] = move_rec
;
2187 * Clear out the record we just copied and shift everything
2188 * over, leaving an empty extent in the left leaf.
2190 * We temporarily subtract from next_free_rec so that the
2191 * shift will lose the tail record (which is now defunct).
2193 le16_add_cpu(&left_el
->l_next_free_rec
, -1);
2194 ocfs2_shift_records_right(left_el
);
2195 memset(&left_el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2196 le16_add_cpu(&left_el
->l_next_free_rec
, 1);
2198 ocfs2_journal_dirty(handle
, left_leaf_bh
);
2200 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2208 * Given a full path, determine what cpos value would return us a path
2209 * containing the leaf immediately to the left of the current one.
2211 * Will return zero if the path passed in is already the leftmost path.
2213 static int ocfs2_find_cpos_for_left_leaf(struct super_block
*sb
,
2214 struct ocfs2_path
*path
, u32
*cpos
)
2218 struct ocfs2_extent_list
*el
;
2220 BUG_ON(path
->p_tree_depth
== 0);
2224 blkno
= path_leaf_bh(path
)->b_blocknr
;
2226 /* Start at the tree node just above the leaf and work our way up. */
2227 i
= path
->p_tree_depth
- 1;
2229 el
= path
->p_node
[i
].el
;
2232 * Find the extent record just before the one in our
2235 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2236 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2240 * We've determined that the
2241 * path specified is already
2242 * the leftmost one - return a
2248 * The leftmost record points to our
2249 * leaf - we need to travel up the
2255 *cpos
= le32_to_cpu(el
->l_recs
[j
- 1].e_cpos
);
2256 *cpos
= *cpos
+ ocfs2_rec_clusters(el
,
2257 &el
->l_recs
[j
- 1]);
2264 * If we got here, we never found a valid node where
2265 * the tree indicated one should be.
2268 "Invalid extent tree at extent block %llu\n",
2269 (unsigned long long)blkno
);
2274 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2283 * Extend the transaction by enough credits to complete the rotation,
2284 * and still leave at least the original number of credits allocated
2285 * to this transaction.
2287 static int ocfs2_extend_rotate_transaction(handle_t
*handle
, int subtree_depth
,
2289 struct ocfs2_path
*path
)
2292 int credits
= (path
->p_tree_depth
- subtree_depth
) * 2 + 1 + op_credits
;
2294 if (handle
->h_buffer_credits
< credits
) {
2295 ret
= ocfs2_extend_trans(handle
,
2296 credits
- handle
->h_buffer_credits
);
2300 if (unlikely(handle
->h_buffer_credits
< credits
))
2301 return ocfs2_extend_trans(handle
, credits
);
2308 * Trap the case where we're inserting into the theoretical range past
2309 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2310 * whose cpos is less than ours into the right leaf.
2312 * It's only necessary to look at the rightmost record of the left
2313 * leaf because the logic that calls us should ensure that the
2314 * theoretical ranges in the path components above the leaves are
2317 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path
*left_path
,
2320 struct ocfs2_extent_list
*left_el
;
2321 struct ocfs2_extent_rec
*rec
;
2324 left_el
= path_leaf_el(left_path
);
2325 next_free
= le16_to_cpu(left_el
->l_next_free_rec
);
2326 rec
= &left_el
->l_recs
[next_free
- 1];
2328 if (insert_cpos
> le32_to_cpu(rec
->e_cpos
))
2333 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list
*el
, u32 cpos
)
2335 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
2337 struct ocfs2_extent_rec
*rec
;
2342 rec
= &el
->l_recs
[0];
2343 if (ocfs2_is_empty_extent(rec
)) {
2347 rec
= &el
->l_recs
[1];
2350 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2351 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
2357 * Rotate all the records in a btree right one record, starting at insert_cpos.
2359 * The path to the rightmost leaf should be passed in.
2361 * The array is assumed to be large enough to hold an entire path (tree depth).
2363 * Upon successful return from this function:
2365 * - The 'right_path' array will contain a path to the leaf block
2366 * whose range contains e_cpos.
2367 * - That leaf block will have a single empty extent in list index 0.
2368 * - In the case that the rotation requires a post-insert update,
2369 * *ret_left_path will contain a valid path which can be passed to
2370 * ocfs2_insert_path().
2372 static int ocfs2_rotate_tree_right(handle_t
*handle
,
2373 struct ocfs2_extent_tree
*et
,
2374 enum ocfs2_split_type split
,
2376 struct ocfs2_path
*right_path
,
2377 struct ocfs2_path
**ret_left_path
)
2379 int ret
, start
, orig_credits
= handle
->h_buffer_credits
;
2381 struct ocfs2_path
*left_path
= NULL
;
2382 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2384 *ret_left_path
= NULL
;
2386 left_path
= ocfs2_new_path_from_path(right_path
);
2393 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2399 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos
, cpos
);
2402 * What we want to do here is:
2404 * 1) Start with the rightmost path.
2406 * 2) Determine a path to the leaf block directly to the left
2409 * 3) Determine the 'subtree root' - the lowest level tree node
2410 * which contains a path to both leaves.
2412 * 4) Rotate the subtree.
2414 * 5) Find the next subtree by considering the left path to be
2415 * the new right path.
2417 * The check at the top of this while loop also accepts
2418 * insert_cpos == cpos because cpos is only a _theoretical_
2419 * value to get us the left path - insert_cpos might very well
2420 * be filling that hole.
2422 * Stop at a cpos of '0' because we either started at the
2423 * leftmost branch (i.e., a tree with one branch and a
2424 * rotation inside of it), or we've gone as far as we can in
2425 * rotating subtrees.
2427 while (cpos
&& insert_cpos
<= cpos
) {
2428 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
2431 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
2437 mlog_bug_on_msg(path_leaf_bh(left_path
) ==
2438 path_leaf_bh(right_path
),
2439 "Owner %llu: error during insert of %u "
2440 "(left path cpos %u) results in two identical "
2441 "paths ending at %llu\n",
2442 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2444 (unsigned long long)
2445 path_leaf_bh(left_path
)->b_blocknr
);
2447 if (split
== SPLIT_NONE
&&
2448 ocfs2_rotate_requires_path_adjustment(left_path
,
2452 * We've rotated the tree as much as we
2453 * should. The rest is up to
2454 * ocfs2_insert_path() to complete, after the
2455 * record insertion. We indicate this
2456 * situation by returning the left path.
2458 * The reason we don't adjust the records here
2459 * before the record insert is that an error
2460 * later might break the rule where a parent
2461 * record e_cpos will reflect the actual
2462 * e_cpos of the 1st nonempty record of the
2465 *ret_left_path
= left_path
;
2469 start
= ocfs2_find_subtree_root(et
, left_path
, right_path
);
2471 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2473 (unsigned long long) right_path
->p_node
[start
].bh
->b_blocknr
,
2474 right_path
->p_tree_depth
);
2476 ret
= ocfs2_extend_rotate_transaction(handle
, start
,
2477 orig_credits
, right_path
);
2483 ret
= ocfs2_rotate_subtree_right(handle
, et
, left_path
,
2490 if (split
!= SPLIT_NONE
&&
2491 ocfs2_leftmost_rec_contains(path_leaf_el(right_path
),
2494 * A rotate moves the rightmost left leaf
2495 * record over to the leftmost right leaf
2496 * slot. If we're doing an extent split
2497 * instead of a real insert, then we have to
2498 * check that the extent to be split wasn't
2499 * just moved over. If it was, then we can
2500 * exit here, passing left_path back -
2501 * ocfs2_split_extent() is smart enough to
2502 * search both leaves.
2504 *ret_left_path
= left_path
;
2509 * There is no need to re-read the next right path
2510 * as we know that it'll be our current left
2511 * path. Optimize by copying values instead.
2513 ocfs2_mv_path(right_path
, left_path
);
2515 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2523 ocfs2_free_path(left_path
);
2529 static int ocfs2_update_edge_lengths(handle_t
*handle
,
2530 struct ocfs2_extent_tree
*et
,
2531 int subtree_index
, struct ocfs2_path
*path
)
2534 struct ocfs2_extent_rec
*rec
;
2535 struct ocfs2_extent_list
*el
;
2536 struct ocfs2_extent_block
*eb
;
2540 * In normal tree rotation process, we will never touch the
2541 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2542 * doesn't reserve the credits for them either.
2544 * But we do have a special case here which will update the rightmost
2545 * records for all the bh in the path.
2546 * So we have to allocate extra credits and access them.
2548 ret
= ocfs2_extend_trans(handle
,
2549 handle
->h_buffer_credits
+ subtree_index
);
2555 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
2561 /* Path should always be rightmost. */
2562 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2563 BUG_ON(eb
->h_next_leaf_blk
!= 0ULL);
2566 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
2567 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2568 rec
= &el
->l_recs
[idx
];
2569 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2571 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
2572 el
= path
->p_node
[i
].el
;
2573 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2574 rec
= &el
->l_recs
[idx
];
2576 rec
->e_int_clusters
= cpu_to_le32(range
);
2577 le32_add_cpu(&rec
->e_int_clusters
, -le32_to_cpu(rec
->e_cpos
));
2579 ocfs2_journal_dirty(handle
, path
->p_node
[i
].bh
);
2585 static void ocfs2_unlink_path(handle_t
*handle
,
2586 struct ocfs2_extent_tree
*et
,
2587 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2588 struct ocfs2_path
*path
, int unlink_start
)
2591 struct ocfs2_extent_block
*eb
;
2592 struct ocfs2_extent_list
*el
;
2593 struct buffer_head
*bh
;
2595 for(i
= unlink_start
; i
< path_num_items(path
); i
++) {
2596 bh
= path
->p_node
[i
].bh
;
2598 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
2600 * Not all nodes might have had their final count
2601 * decremented by the caller - handle this here.
2604 if (le16_to_cpu(el
->l_next_free_rec
) > 1) {
2606 "Inode %llu, attempted to remove extent block "
2607 "%llu with %u records\n",
2608 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2609 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
2610 le16_to_cpu(el
->l_next_free_rec
));
2612 ocfs2_journal_dirty(handle
, bh
);
2613 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2617 el
->l_next_free_rec
= 0;
2618 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2620 ocfs2_journal_dirty(handle
, bh
);
2622 ret
= ocfs2_cache_extent_block_free(dealloc
, eb
);
2626 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2630 static void ocfs2_unlink_subtree(handle_t
*handle
,
2631 struct ocfs2_extent_tree
*et
,
2632 struct ocfs2_path
*left_path
,
2633 struct ocfs2_path
*right_path
,
2635 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2638 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2639 struct ocfs2_extent_list
*root_el
= left_path
->p_node
[subtree_index
].el
;
2640 struct ocfs2_extent_list
*el
;
2641 struct ocfs2_extent_block
*eb
;
2643 el
= path_leaf_el(left_path
);
2645 eb
= (struct ocfs2_extent_block
*)right_path
->p_node
[subtree_index
+ 1].bh
->b_data
;
2647 for(i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
2648 if (root_el
->l_recs
[i
].e_blkno
== eb
->h_blkno
)
2651 BUG_ON(i
>= le16_to_cpu(root_el
->l_next_free_rec
));
2653 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
2654 le16_add_cpu(&root_el
->l_next_free_rec
, -1);
2656 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2657 eb
->h_next_leaf_blk
= 0;
2659 ocfs2_journal_dirty(handle
, root_bh
);
2660 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2662 ocfs2_unlink_path(handle
, et
, dealloc
, right_path
,
2666 static int ocfs2_rotate_subtree_left(handle_t
*handle
,
2667 struct ocfs2_extent_tree
*et
,
2668 struct ocfs2_path
*left_path
,
2669 struct ocfs2_path
*right_path
,
2671 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2674 int ret
, i
, del_right_subtree
= 0, right_has_empty
= 0;
2675 struct buffer_head
*root_bh
, *et_root_bh
= path_root_bh(right_path
);
2676 struct ocfs2_extent_list
*right_leaf_el
, *left_leaf_el
;
2677 struct ocfs2_extent_block
*eb
;
2681 right_leaf_el
= path_leaf_el(right_path
);
2682 left_leaf_el
= path_leaf_el(left_path
);
2683 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2684 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2686 if (!ocfs2_is_empty_extent(&left_leaf_el
->l_recs
[0]))
2689 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(right_path
)->b_data
;
2690 if (ocfs2_is_empty_extent(&right_leaf_el
->l_recs
[0])) {
2692 * It's legal for us to proceed if the right leaf is
2693 * the rightmost one and it has an empty extent. There
2694 * are two cases to handle - whether the leaf will be
2695 * empty after removal or not. If the leaf isn't empty
2696 * then just remove the empty extent up front. The
2697 * next block will handle empty leaves by flagging
2700 * Non rightmost leaves will throw -EAGAIN and the
2701 * caller can manually move the subtree and retry.
2704 if (eb
->h_next_leaf_blk
!= 0ULL)
2707 if (le16_to_cpu(right_leaf_el
->l_next_free_rec
) > 1) {
2708 ret
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
2709 path_leaf_bh(right_path
),
2710 OCFS2_JOURNAL_ACCESS_WRITE
);
2716 ocfs2_remove_empty_extent(right_leaf_el
);
2718 right_has_empty
= 1;
2721 if (eb
->h_next_leaf_blk
== 0ULL &&
2722 le16_to_cpu(right_leaf_el
->l_next_free_rec
) == 1) {
2724 * We have to update i_last_eb_blk during the meta
2727 ret
= ocfs2_et_root_journal_access(handle
, et
,
2728 OCFS2_JOURNAL_ACCESS_WRITE
);
2734 del_right_subtree
= 1;
2738 * Getting here with an empty extent in the right path implies
2739 * that it's the rightmost path and will be deleted.
2741 BUG_ON(right_has_empty
&& !del_right_subtree
);
2743 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2750 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2751 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2758 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2766 if (!right_has_empty
) {
2768 * Only do this if we're moving a real
2769 * record. Otherwise, the action is delayed until
2770 * after removal of the right path in which case we
2771 * can do a simple shift to remove the empty extent.
2773 ocfs2_rotate_leaf(left_leaf_el
, &right_leaf_el
->l_recs
[0]);
2774 memset(&right_leaf_el
->l_recs
[0], 0,
2775 sizeof(struct ocfs2_extent_rec
));
2777 if (eb
->h_next_leaf_blk
== 0ULL) {
2779 * Move recs over to get rid of empty extent, decrease
2780 * next_free. This is allowed to remove the last
2781 * extent in our leaf (setting l_next_free_rec to
2782 * zero) - the delete code below won't care.
2784 ocfs2_remove_empty_extent(right_leaf_el
);
2787 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2788 ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2790 if (del_right_subtree
) {
2791 ocfs2_unlink_subtree(handle
, et
, left_path
, right_path
,
2792 subtree_index
, dealloc
);
2793 ret
= ocfs2_update_edge_lengths(handle
, et
, subtree_index
,
2800 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2801 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
2804 * Removal of the extent in the left leaf was skipped
2805 * above so we could delete the right path
2808 if (right_has_empty
)
2809 ocfs2_remove_empty_extent(left_leaf_el
);
2811 ocfs2_journal_dirty(handle
, et_root_bh
);
2815 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2823 * Given a full path, determine what cpos value would return us a path
2824 * containing the leaf immediately to the right of the current one.
2826 * Will return zero if the path passed in is already the rightmost path.
2828 * This looks similar, but is subtly different to
2829 * ocfs2_find_cpos_for_left_leaf().
2831 int ocfs2_find_cpos_for_right_leaf(struct super_block
*sb
,
2832 struct ocfs2_path
*path
, u32
*cpos
)
2836 struct ocfs2_extent_list
*el
;
2840 if (path
->p_tree_depth
== 0)
2843 blkno
= path_leaf_bh(path
)->b_blocknr
;
2845 /* Start at the tree node just above the leaf and work our way up. */
2846 i
= path
->p_tree_depth
- 1;
2850 el
= path
->p_node
[i
].el
;
2853 * Find the extent record just after the one in our
2856 next_free
= le16_to_cpu(el
->l_next_free_rec
);
2857 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2858 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2859 if (j
== (next_free
- 1)) {
2862 * We've determined that the
2863 * path specified is already
2864 * the rightmost one - return a
2870 * The rightmost record points to our
2871 * leaf - we need to travel up the
2877 *cpos
= le32_to_cpu(el
->l_recs
[j
+ 1].e_cpos
);
2883 * If we got here, we never found a valid node where
2884 * the tree indicated one should be.
2887 "Invalid extent tree at extent block %llu\n",
2888 (unsigned long long)blkno
);
2893 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2901 static int ocfs2_rotate_rightmost_leaf_left(handle_t
*handle
,
2902 struct ocfs2_extent_tree
*et
,
2903 struct ocfs2_path
*path
)
2906 struct buffer_head
*bh
= path_leaf_bh(path
);
2907 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
2909 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2912 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
2913 path_num_items(path
) - 1);
2919 ocfs2_remove_empty_extent(el
);
2920 ocfs2_journal_dirty(handle
, bh
);
2926 static int __ocfs2_rotate_tree_left(handle_t
*handle
,
2927 struct ocfs2_extent_tree
*et
,
2929 struct ocfs2_path
*path
,
2930 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2931 struct ocfs2_path
**empty_extent_path
)
2933 int ret
, subtree_root
, deleted
;
2935 struct ocfs2_path
*left_path
= NULL
;
2936 struct ocfs2_path
*right_path
= NULL
;
2937 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2939 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path
)->l_recs
[0])));
2941 *empty_extent_path
= NULL
;
2943 ret
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
2949 left_path
= ocfs2_new_path_from_path(path
);
2956 ocfs2_cp_path(left_path
, path
);
2958 right_path
= ocfs2_new_path_from_path(path
);
2965 while (right_cpos
) {
2966 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
2972 subtree_root
= ocfs2_find_subtree_root(et
, left_path
,
2975 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2977 (unsigned long long)
2978 right_path
->p_node
[subtree_root
].bh
->b_blocknr
,
2979 right_path
->p_tree_depth
);
2981 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_root
,
2982 orig_credits
, left_path
);
2989 * Caller might still want to make changes to the
2990 * tree root, so re-add it to the journal here.
2992 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2999 ret
= ocfs2_rotate_subtree_left(handle
, et
, left_path
,
3000 right_path
, subtree_root
,
3002 if (ret
== -EAGAIN
) {
3004 * The rotation has to temporarily stop due to
3005 * the right subtree having an empty
3006 * extent. Pass it back to the caller for a
3009 *empty_extent_path
= right_path
;
3019 * The subtree rotate might have removed records on
3020 * the rightmost edge. If so, then rotation is
3026 ocfs2_mv_path(left_path
, right_path
);
3028 ret
= ocfs2_find_cpos_for_right_leaf(sb
, left_path
,
3037 ocfs2_free_path(right_path
);
3038 ocfs2_free_path(left_path
);
3043 static int ocfs2_remove_rightmost_path(handle_t
*handle
,
3044 struct ocfs2_extent_tree
*et
,
3045 struct ocfs2_path
*path
,
3046 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3048 int ret
, subtree_index
;
3050 struct ocfs2_path
*left_path
= NULL
;
3051 struct ocfs2_extent_block
*eb
;
3052 struct ocfs2_extent_list
*el
;
3055 ret
= ocfs2_et_sanity_check(et
);
3059 * There's two ways we handle this depending on
3060 * whether path is the only existing one.
3062 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3063 handle
->h_buffer_credits
,
3070 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
3076 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3085 * We have a path to the left of this one - it needs
3088 left_path
= ocfs2_new_path_from_path(path
);
3095 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
3101 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
3107 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
3109 ocfs2_unlink_subtree(handle
, et
, left_path
, path
,
3110 subtree_index
, dealloc
);
3111 ret
= ocfs2_update_edge_lengths(handle
, et
, subtree_index
,
3118 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
3119 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
3122 * 'path' is also the leftmost path which
3123 * means it must be the only one. This gets
3124 * handled differently because we want to
3125 * revert the root back to having extents
3128 ocfs2_unlink_path(handle
, et
, dealloc
, path
, 1);
3130 el
= et
->et_root_el
;
3131 el
->l_tree_depth
= 0;
3132 el
->l_next_free_rec
= 0;
3133 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3135 ocfs2_et_set_last_eb_blk(et
, 0);
3138 ocfs2_journal_dirty(handle
, path_root_bh(path
));
3141 ocfs2_free_path(left_path
);
3146 * Left rotation of btree records.
3148 * In many ways, this is (unsurprisingly) the opposite of right
3149 * rotation. We start at some non-rightmost path containing an empty
3150 * extent in the leaf block. The code works its way to the rightmost
3151 * path by rotating records to the left in every subtree.
3153 * This is used by any code which reduces the number of extent records
3154 * in a leaf. After removal, an empty record should be placed in the
3155 * leftmost list position.
3157 * This won't handle a length update of the rightmost path records if
3158 * the rightmost tree leaf record is removed so the caller is
3159 * responsible for detecting and correcting that.
3161 static int ocfs2_rotate_tree_left(handle_t
*handle
,
3162 struct ocfs2_extent_tree
*et
,
3163 struct ocfs2_path
*path
,
3164 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3166 int ret
, orig_credits
= handle
->h_buffer_credits
;
3167 struct ocfs2_path
*tmp_path
= NULL
, *restart_path
= NULL
;
3168 struct ocfs2_extent_block
*eb
;
3169 struct ocfs2_extent_list
*el
;
3171 el
= path_leaf_el(path
);
3172 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
3175 if (path
->p_tree_depth
== 0) {
3176 rightmost_no_delete
:
3178 * Inline extents. This is trivially handled, so do
3181 ret
= ocfs2_rotate_rightmost_leaf_left(handle
, et
, path
);
3188 * Handle rightmost branch now. There's several cases:
3189 * 1) simple rotation leaving records in there. That's trivial.
3190 * 2) rotation requiring a branch delete - there's no more
3191 * records left. Two cases of this:
3192 * a) There are branches to the left.
3193 * b) This is also the leftmost (the only) branch.
3195 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3196 * 2a) we need the left branch so that we can update it with the unlink
3197 * 2b) we need to bring the root back to inline extents.
3200 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
3202 if (eb
->h_next_leaf_blk
== 0) {
3204 * This gets a bit tricky if we're going to delete the
3205 * rightmost path. Get the other cases out of the way
3208 if (le16_to_cpu(el
->l_next_free_rec
) > 1)
3209 goto rightmost_no_delete
;
3211 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
3213 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
3214 "Owner %llu has empty extent block at %llu",
3215 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
3216 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
3221 * XXX: The caller can not trust "path" any more after
3222 * this as it will have been deleted. What do we do?
3224 * In theory the rotate-for-merge code will never get
3225 * here because it'll always ask for a rotate in a
3229 ret
= ocfs2_remove_rightmost_path(handle
, et
, path
,
3237 * Now we can loop, remembering the path we get from -EAGAIN
3238 * and restarting from there.
3241 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
, path
,
3242 dealloc
, &restart_path
);
3243 if (ret
&& ret
!= -EAGAIN
) {
3248 while (ret
== -EAGAIN
) {
3249 tmp_path
= restart_path
;
3250 restart_path
= NULL
;
3252 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
,
3255 if (ret
&& ret
!= -EAGAIN
) {
3260 ocfs2_free_path(tmp_path
);
3268 ocfs2_free_path(tmp_path
);
3269 ocfs2_free_path(restart_path
);
3273 static void ocfs2_cleanup_merge(struct ocfs2_extent_list
*el
,
3276 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[index
];
3279 if (rec
->e_leaf_clusters
== 0) {
3281 * We consumed all of the merged-from record. An empty
3282 * extent cannot exist anywhere but the 1st array
3283 * position, so move things over if the merged-from
3284 * record doesn't occupy that position.
3286 * This creates a new empty extent so the caller
3287 * should be smart enough to have removed any existing
3291 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3292 size
= index
* sizeof(struct ocfs2_extent_rec
);
3293 memmove(&el
->l_recs
[1], &el
->l_recs
[0], size
);
3297 * Always memset - the caller doesn't check whether it
3298 * created an empty extent, so there could be junk in
3301 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3305 static int ocfs2_get_right_path(struct ocfs2_extent_tree
*et
,
3306 struct ocfs2_path
*left_path
,
3307 struct ocfs2_path
**ret_right_path
)
3311 struct ocfs2_path
*right_path
= NULL
;
3312 struct ocfs2_extent_list
*left_el
;
3314 *ret_right_path
= NULL
;
3316 /* This function shouldn't be called for non-trees. */
3317 BUG_ON(left_path
->p_tree_depth
== 0);
3319 left_el
= path_leaf_el(left_path
);
3320 BUG_ON(left_el
->l_next_free_rec
!= left_el
->l_count
);
3322 ret
= ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3323 left_path
, &right_cpos
);
3329 /* This function shouldn't be called for the rightmost leaf. */
3330 BUG_ON(right_cpos
== 0);
3332 right_path
= ocfs2_new_path_from_path(left_path
);
3339 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
3345 *ret_right_path
= right_path
;
3348 ocfs2_free_path(right_path
);
3353 * Remove split_rec clusters from the record at index and merge them
3354 * onto the beginning of the record "next" to it.
3355 * For index < l_count - 1, the next means the extent rec at index + 1.
3356 * For index == l_count - 1, the "next" means the 1st extent rec of the
3357 * next extent block.
3359 static int ocfs2_merge_rec_right(struct ocfs2_path
*left_path
,
3361 struct ocfs2_extent_tree
*et
,
3362 struct ocfs2_extent_rec
*split_rec
,
3365 int ret
, next_free
, i
;
3366 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3367 struct ocfs2_extent_rec
*left_rec
;
3368 struct ocfs2_extent_rec
*right_rec
;
3369 struct ocfs2_extent_list
*right_el
;
3370 struct ocfs2_path
*right_path
= NULL
;
3371 int subtree_index
= 0;
3372 struct ocfs2_extent_list
*el
= path_leaf_el(left_path
);
3373 struct buffer_head
*bh
= path_leaf_bh(left_path
);
3374 struct buffer_head
*root_bh
= NULL
;
3376 BUG_ON(index
>= le16_to_cpu(el
->l_next_free_rec
));
3377 left_rec
= &el
->l_recs
[index
];
3379 if (index
== le16_to_cpu(el
->l_next_free_rec
) - 1 &&
3380 le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
)) {
3381 /* we meet with a cross extent block merge. */
3382 ret
= ocfs2_get_right_path(et
, left_path
, &right_path
);
3388 right_el
= path_leaf_el(right_path
);
3389 next_free
= le16_to_cpu(right_el
->l_next_free_rec
);
3390 BUG_ON(next_free
<= 0);
3391 right_rec
= &right_el
->l_recs
[0];
3392 if (ocfs2_is_empty_extent(right_rec
)) {
3393 BUG_ON(next_free
<= 1);
3394 right_rec
= &right_el
->l_recs
[1];
3397 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3398 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3399 le32_to_cpu(right_rec
->e_cpos
));
3401 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3404 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3405 handle
->h_buffer_credits
,
3412 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3413 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3415 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3422 for (i
= subtree_index
+ 1;
3423 i
< path_num_items(right_path
); i
++) {
3424 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3431 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3440 BUG_ON(index
== le16_to_cpu(el
->l_next_free_rec
) - 1);
3441 right_rec
= &el
->l_recs
[index
+ 1];
3444 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, left_path
,
3445 path_num_items(left_path
) - 1);
3451 le16_add_cpu(&left_rec
->e_leaf_clusters
, -split_clusters
);
3453 le32_add_cpu(&right_rec
->e_cpos
, -split_clusters
);
3454 le64_add_cpu(&right_rec
->e_blkno
,
3455 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3457 le16_add_cpu(&right_rec
->e_leaf_clusters
, split_clusters
);
3459 ocfs2_cleanup_merge(el
, index
);
3461 ocfs2_journal_dirty(handle
, bh
);
3463 ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
3464 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
3469 ocfs2_free_path(right_path
);
3473 static int ocfs2_get_left_path(struct ocfs2_extent_tree
*et
,
3474 struct ocfs2_path
*right_path
,
3475 struct ocfs2_path
**ret_left_path
)
3479 struct ocfs2_path
*left_path
= NULL
;
3481 *ret_left_path
= NULL
;
3483 /* This function shouldn't be called for non-trees. */
3484 BUG_ON(right_path
->p_tree_depth
== 0);
3486 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3487 right_path
, &left_cpos
);
3493 /* This function shouldn't be called for the leftmost leaf. */
3494 BUG_ON(left_cpos
== 0);
3496 left_path
= ocfs2_new_path_from_path(right_path
);
3503 ret
= ocfs2_find_path(et
->et_ci
, left_path
, left_cpos
);
3509 *ret_left_path
= left_path
;
3512 ocfs2_free_path(left_path
);
3517 * Remove split_rec clusters from the record at index and merge them
3518 * onto the tail of the record "before" it.
3519 * For index > 0, the "before" means the extent rec at index - 1.
3521 * For index == 0, the "before" means the last record of the previous
3522 * extent block. And there is also a situation that we may need to
3523 * remove the rightmost leaf extent block in the right_path and change
3524 * the right path to indicate the new rightmost path.
3526 static int ocfs2_merge_rec_left(struct ocfs2_path
*right_path
,
3528 struct ocfs2_extent_tree
*et
,
3529 struct ocfs2_extent_rec
*split_rec
,
3530 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3533 int ret
, i
, subtree_index
= 0, has_empty_extent
= 0;
3534 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3535 struct ocfs2_extent_rec
*left_rec
;
3536 struct ocfs2_extent_rec
*right_rec
;
3537 struct ocfs2_extent_list
*el
= path_leaf_el(right_path
);
3538 struct buffer_head
*bh
= path_leaf_bh(right_path
);
3539 struct buffer_head
*root_bh
= NULL
;
3540 struct ocfs2_path
*left_path
= NULL
;
3541 struct ocfs2_extent_list
*left_el
;
3545 right_rec
= &el
->l_recs
[index
];
3547 /* we meet with a cross extent block merge. */
3548 ret
= ocfs2_get_left_path(et
, right_path
, &left_path
);
3554 left_el
= path_leaf_el(left_path
);
3555 BUG_ON(le16_to_cpu(left_el
->l_next_free_rec
) !=
3556 le16_to_cpu(left_el
->l_count
));
3558 left_rec
= &left_el
->l_recs
[
3559 le16_to_cpu(left_el
->l_next_free_rec
) - 1];
3560 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3561 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3562 le32_to_cpu(split_rec
->e_cpos
));
3564 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3567 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3568 handle
->h_buffer_credits
,
3575 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3576 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3578 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3585 for (i
= subtree_index
+ 1;
3586 i
< path_num_items(right_path
); i
++) {
3587 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3594 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3602 left_rec
= &el
->l_recs
[index
- 1];
3603 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
3604 has_empty_extent
= 1;
3607 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3608 path_num_items(right_path
) - 1);
3614 if (has_empty_extent
&& index
== 1) {
3616 * The easy case - we can just plop the record right in.
3618 *left_rec
= *split_rec
;
3620 has_empty_extent
= 0;
3622 le16_add_cpu(&left_rec
->e_leaf_clusters
, split_clusters
);
3624 le32_add_cpu(&right_rec
->e_cpos
, split_clusters
);
3625 le64_add_cpu(&right_rec
->e_blkno
,
3626 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3628 le16_add_cpu(&right_rec
->e_leaf_clusters
, -split_clusters
);
3630 ocfs2_cleanup_merge(el
, index
);
3632 ocfs2_journal_dirty(handle
, bh
);
3634 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
3637 * In the situation that the right_rec is empty and the extent
3638 * block is empty also, ocfs2_complete_edge_insert can't handle
3639 * it and we need to delete the right extent block.
3641 if (le16_to_cpu(right_rec
->e_leaf_clusters
) == 0 &&
3642 le16_to_cpu(el
->l_next_free_rec
) == 1) {
3644 ret
= ocfs2_remove_rightmost_path(handle
, et
,
3652 /* Now the rightmost extent block has been deleted.
3653 * So we use the new rightmost path.
3655 ocfs2_mv_path(right_path
, left_path
);
3658 ocfs2_complete_edge_insert(handle
, left_path
,
3659 right_path
, subtree_index
);
3663 ocfs2_free_path(left_path
);
3667 static int ocfs2_try_to_merge_extent(handle_t
*handle
,
3668 struct ocfs2_extent_tree
*et
,
3669 struct ocfs2_path
*path
,
3671 struct ocfs2_extent_rec
*split_rec
,
3672 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3673 struct ocfs2_merge_ctxt
*ctxt
)
3676 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
3677 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
3679 BUG_ON(ctxt
->c_contig_type
== CONTIG_NONE
);
3681 if (ctxt
->c_split_covers_rec
&& ctxt
->c_has_empty_extent
) {
3683 * The merge code will need to create an empty
3684 * extent to take the place of the newly
3685 * emptied slot. Remove any pre-existing empty
3686 * extents - having more than one in a leaf is
3689 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3695 rec
= &el
->l_recs
[split_index
];
3698 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
) {
3700 * Left-right contig implies this.
3702 BUG_ON(!ctxt
->c_split_covers_rec
);
3705 * Since the leftright insert always covers the entire
3706 * extent, this call will delete the insert record
3707 * entirely, resulting in an empty extent record added to
3710 * Since the adding of an empty extent shifts
3711 * everything back to the right, there's no need to
3712 * update split_index here.
3714 * When the split_index is zero, we need to merge it to the
3715 * prevoius extent block. It is more efficient and easier
3716 * if we do merge_right first and merge_left later.
3718 ret
= ocfs2_merge_rec_right(path
, handle
, et
, split_rec
,
3726 * We can only get this from logic error above.
3728 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
3730 /* The merge left us with an empty extent, remove it. */
3731 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3737 rec
= &el
->l_recs
[split_index
];
3740 * Note that we don't pass split_rec here on purpose -
3741 * we've merged it into the rec already.
3743 ret
= ocfs2_merge_rec_left(path
, handle
, et
, rec
,
3744 dealloc
, split_index
);
3751 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3753 * Error from this last rotate is not critical, so
3754 * print but don't bubble it up.
3761 * Merge a record to the left or right.
3763 * 'contig_type' is relative to the existing record,
3764 * so for example, if we're "right contig", it's to
3765 * the record on the left (hence the left merge).
3767 if (ctxt
->c_contig_type
== CONTIG_RIGHT
) {
3768 ret
= ocfs2_merge_rec_left(path
, handle
, et
,
3776 ret
= ocfs2_merge_rec_right(path
, handle
,
3785 if (ctxt
->c_split_covers_rec
) {
3787 * The merge may have left an empty extent in
3788 * our leaf. Try to rotate it away.
3790 ret
= ocfs2_rotate_tree_left(handle
, et
, path
,
3802 static void ocfs2_subtract_from_rec(struct super_block
*sb
,
3803 enum ocfs2_split_type split
,
3804 struct ocfs2_extent_rec
*rec
,
3805 struct ocfs2_extent_rec
*split_rec
)
3809 len_blocks
= ocfs2_clusters_to_blocks(sb
,
3810 le16_to_cpu(split_rec
->e_leaf_clusters
));
3812 if (split
== SPLIT_LEFT
) {
3814 * Region is on the left edge of the existing
3817 le32_add_cpu(&rec
->e_cpos
,
3818 le16_to_cpu(split_rec
->e_leaf_clusters
));
3819 le64_add_cpu(&rec
->e_blkno
, len_blocks
);
3820 le16_add_cpu(&rec
->e_leaf_clusters
,
3821 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3824 * Region is on the right edge of the existing
3827 le16_add_cpu(&rec
->e_leaf_clusters
,
3828 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3833 * Do the final bits of extent record insertion at the target leaf
3834 * list. If this leaf is part of an allocation tree, it is assumed
3835 * that the tree above has been prepared.
3837 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree
*et
,
3838 struct ocfs2_extent_rec
*insert_rec
,
3839 struct ocfs2_extent_list
*el
,
3840 struct ocfs2_insert_type
*insert
)
3842 int i
= insert
->ins_contig_index
;
3844 struct ocfs2_extent_rec
*rec
;
3846 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3848 if (insert
->ins_split
!= SPLIT_NONE
) {
3849 i
= ocfs2_search_extent_list(el
, le32_to_cpu(insert_rec
->e_cpos
));
3851 rec
= &el
->l_recs
[i
];
3852 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
3853 insert
->ins_split
, rec
,
3859 * Contiguous insert - either left or right.
3861 if (insert
->ins_contig
!= CONTIG_NONE
) {
3862 rec
= &el
->l_recs
[i
];
3863 if (insert
->ins_contig
== CONTIG_LEFT
) {
3864 rec
->e_blkno
= insert_rec
->e_blkno
;
3865 rec
->e_cpos
= insert_rec
->e_cpos
;
3867 le16_add_cpu(&rec
->e_leaf_clusters
,
3868 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3873 * Handle insert into an empty leaf.
3875 if (le16_to_cpu(el
->l_next_free_rec
) == 0 ||
3876 ((le16_to_cpu(el
->l_next_free_rec
) == 1) &&
3877 ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3878 el
->l_recs
[0] = *insert_rec
;
3879 el
->l_next_free_rec
= cpu_to_le16(1);
3886 if (insert
->ins_appending
== APPEND_TAIL
) {
3887 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3888 rec
= &el
->l_recs
[i
];
3889 range
= le32_to_cpu(rec
->e_cpos
)
3890 + le16_to_cpu(rec
->e_leaf_clusters
);
3891 BUG_ON(le32_to_cpu(insert_rec
->e_cpos
) < range
);
3893 mlog_bug_on_msg(le16_to_cpu(el
->l_next_free_rec
) >=
3894 le16_to_cpu(el
->l_count
),
3895 "owner %llu, depth %u, count %u, next free %u, "
3896 "rec.cpos %u, rec.clusters %u, "
3897 "insert.cpos %u, insert.clusters %u\n",
3898 ocfs2_metadata_cache_owner(et
->et_ci
),
3899 le16_to_cpu(el
->l_tree_depth
),
3900 le16_to_cpu(el
->l_count
),
3901 le16_to_cpu(el
->l_next_free_rec
),
3902 le32_to_cpu(el
->l_recs
[i
].e_cpos
),
3903 le16_to_cpu(el
->l_recs
[i
].e_leaf_clusters
),
3904 le32_to_cpu(insert_rec
->e_cpos
),
3905 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3907 el
->l_recs
[i
] = *insert_rec
;
3908 le16_add_cpu(&el
->l_next_free_rec
, 1);
3914 * Ok, we have to rotate.
3916 * At this point, it is safe to assume that inserting into an
3917 * empty leaf and appending to a leaf have both been handled
3920 * This leaf needs to have space, either by the empty 1st
3921 * extent record, or by virtue of an l_next_rec < l_count.
3923 ocfs2_rotate_leaf(el
, insert_rec
);
3926 static void ocfs2_adjust_rightmost_records(handle_t
*handle
,
3927 struct ocfs2_extent_tree
*et
,
3928 struct ocfs2_path
*path
,
3929 struct ocfs2_extent_rec
*insert_rec
)
3931 int ret
, i
, next_free
;
3932 struct buffer_head
*bh
;
3933 struct ocfs2_extent_list
*el
;
3934 struct ocfs2_extent_rec
*rec
;
3937 * Update everything except the leaf block.
3939 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
3940 bh
= path
->p_node
[i
].bh
;
3941 el
= path
->p_node
[i
].el
;
3943 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3944 if (next_free
== 0) {
3945 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
3946 "Owner %llu has a bad extent list",
3947 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
3952 rec
= &el
->l_recs
[next_free
- 1];
3954 rec
->e_int_clusters
= insert_rec
->e_cpos
;
3955 le32_add_cpu(&rec
->e_int_clusters
,
3956 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3957 le32_add_cpu(&rec
->e_int_clusters
,
3958 -le32_to_cpu(rec
->e_cpos
));
3960 ocfs2_journal_dirty(handle
, bh
);
3964 static int ocfs2_append_rec_to_path(handle_t
*handle
,
3965 struct ocfs2_extent_tree
*et
,
3966 struct ocfs2_extent_rec
*insert_rec
,
3967 struct ocfs2_path
*right_path
,
3968 struct ocfs2_path
**ret_left_path
)
3971 struct ocfs2_extent_list
*el
;
3972 struct ocfs2_path
*left_path
= NULL
;
3974 *ret_left_path
= NULL
;
3977 * This shouldn't happen for non-trees. The extent rec cluster
3978 * count manipulation below only works for interior nodes.
3980 BUG_ON(right_path
->p_tree_depth
== 0);
3983 * If our appending insert is at the leftmost edge of a leaf,
3984 * then we might need to update the rightmost records of the
3987 el
= path_leaf_el(right_path
);
3988 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3989 if (next_free
== 0 ||
3990 (next_free
== 1 && ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3993 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3994 right_path
, &left_cpos
);
4000 mlog(0, "Append may need a left path update. cpos: %u, "
4001 "left_cpos: %u\n", le32_to_cpu(insert_rec
->e_cpos
),
4005 * No need to worry if the append is already in the
4009 left_path
= ocfs2_new_path_from_path(right_path
);
4016 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
4024 * ocfs2_insert_path() will pass the left_path to the
4030 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4036 ocfs2_adjust_rightmost_records(handle
, et
, right_path
, insert_rec
);
4038 *ret_left_path
= left_path
;
4042 ocfs2_free_path(left_path
);
4047 static void ocfs2_split_record(struct ocfs2_extent_tree
*et
,
4048 struct ocfs2_path
*left_path
,
4049 struct ocfs2_path
*right_path
,
4050 struct ocfs2_extent_rec
*split_rec
,
4051 enum ocfs2_split_type split
)
4054 u32 cpos
= le32_to_cpu(split_rec
->e_cpos
);
4055 struct ocfs2_extent_list
*left_el
= NULL
, *right_el
, *insert_el
, *el
;
4056 struct ocfs2_extent_rec
*rec
, *tmprec
;
4058 right_el
= path_leaf_el(right_path
);
4060 left_el
= path_leaf_el(left_path
);
4063 insert_el
= right_el
;
4064 index
= ocfs2_search_extent_list(el
, cpos
);
4066 if (index
== 0 && left_path
) {
4067 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
4070 * This typically means that the record
4071 * started in the left path but moved to the
4072 * right as a result of rotation. We either
4073 * move the existing record to the left, or we
4074 * do the later insert there.
4076 * In this case, the left path should always
4077 * exist as the rotate code will have passed
4078 * it back for a post-insert update.
4081 if (split
== SPLIT_LEFT
) {
4083 * It's a left split. Since we know
4084 * that the rotate code gave us an
4085 * empty extent in the left path, we
4086 * can just do the insert there.
4088 insert_el
= left_el
;
4091 * Right split - we have to move the
4092 * existing record over to the left
4093 * leaf. The insert will be into the
4094 * newly created empty extent in the
4097 tmprec
= &right_el
->l_recs
[index
];
4098 ocfs2_rotate_leaf(left_el
, tmprec
);
4101 memset(tmprec
, 0, sizeof(*tmprec
));
4102 index
= ocfs2_search_extent_list(left_el
, cpos
);
4103 BUG_ON(index
== -1);
4108 BUG_ON(!ocfs2_is_empty_extent(&left_el
->l_recs
[0]));
4110 * Left path is easy - we can just allow the insert to
4114 insert_el
= left_el
;
4115 index
= ocfs2_search_extent_list(el
, cpos
);
4116 BUG_ON(index
== -1);
4119 rec
= &el
->l_recs
[index
];
4120 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
4121 split
, rec
, split_rec
);
4122 ocfs2_rotate_leaf(insert_el
, split_rec
);
4126 * This function only does inserts on an allocation b-tree. For tree
4127 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4129 * right_path is the path we want to do the actual insert
4130 * in. left_path should only be passed in if we need to update that
4131 * portion of the tree after an edge insert.
4133 static int ocfs2_insert_path(handle_t
*handle
,
4134 struct ocfs2_extent_tree
*et
,
4135 struct ocfs2_path
*left_path
,
4136 struct ocfs2_path
*right_path
,
4137 struct ocfs2_extent_rec
*insert_rec
,
4138 struct ocfs2_insert_type
*insert
)
4140 int ret
, subtree_index
;
4141 struct buffer_head
*leaf_bh
= path_leaf_bh(right_path
);
4144 int credits
= handle
->h_buffer_credits
;
4147 * There's a chance that left_path got passed back to
4148 * us without being accounted for in the
4149 * journal. Extend our transaction here to be sure we
4150 * can change those blocks.
4152 credits
+= left_path
->p_tree_depth
;
4154 ret
= ocfs2_extend_trans(handle
, credits
);
4160 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
4168 * Pass both paths to the journal. The majority of inserts
4169 * will be touching all components anyway.
4171 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4177 if (insert
->ins_split
!= SPLIT_NONE
) {
4179 * We could call ocfs2_insert_at_leaf() for some types
4180 * of splits, but it's easier to just let one separate
4181 * function sort it all out.
4183 ocfs2_split_record(et
, left_path
, right_path
,
4184 insert_rec
, insert
->ins_split
);
4187 * Split might have modified either leaf and we don't
4188 * have a guarantee that the later edge insert will
4189 * dirty this for us.
4192 ocfs2_journal_dirty(handle
,
4193 path_leaf_bh(left_path
));
4195 ocfs2_insert_at_leaf(et
, insert_rec
, path_leaf_el(right_path
),
4198 ocfs2_journal_dirty(handle
, leaf_bh
);
4202 * The rotate code has indicated that we need to fix
4203 * up portions of the tree after the insert.
4205 * XXX: Should we extend the transaction here?
4207 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
4209 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
4218 static int ocfs2_do_insert_extent(handle_t
*handle
,
4219 struct ocfs2_extent_tree
*et
,
4220 struct ocfs2_extent_rec
*insert_rec
,
4221 struct ocfs2_insert_type
*type
)
4223 int ret
, rotate
= 0;
4225 struct ocfs2_path
*right_path
= NULL
;
4226 struct ocfs2_path
*left_path
= NULL
;
4227 struct ocfs2_extent_list
*el
;
4229 el
= et
->et_root_el
;
4231 ret
= ocfs2_et_root_journal_access(handle
, et
,
4232 OCFS2_JOURNAL_ACCESS_WRITE
);
4238 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
4239 ocfs2_insert_at_leaf(et
, insert_rec
, el
, type
);
4240 goto out_update_clusters
;
4243 right_path
= ocfs2_new_path_from_et(et
);
4251 * Determine the path to start with. Rotations need the
4252 * rightmost path, everything else can go directly to the
4255 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4256 if (type
->ins_appending
== APPEND_NONE
&&
4257 type
->ins_contig
== CONTIG_NONE
) {
4262 ret
= ocfs2_find_path(et
->et_ci
, right_path
, cpos
);
4269 * Rotations and appends need special treatment - they modify
4270 * parts of the tree's above them.
4272 * Both might pass back a path immediate to the left of the
4273 * one being inserted to. This will be cause
4274 * ocfs2_insert_path() to modify the rightmost records of
4275 * left_path to account for an edge insert.
4277 * XXX: When modifying this code, keep in mind that an insert
4278 * can wind up skipping both of these two special cases...
4281 ret
= ocfs2_rotate_tree_right(handle
, et
, type
->ins_split
,
4282 le32_to_cpu(insert_rec
->e_cpos
),
4283 right_path
, &left_path
);
4290 * ocfs2_rotate_tree_right() might have extended the
4291 * transaction without re-journaling our tree root.
4293 ret
= ocfs2_et_root_journal_access(handle
, et
,
4294 OCFS2_JOURNAL_ACCESS_WRITE
);
4299 } else if (type
->ins_appending
== APPEND_TAIL
4300 && type
->ins_contig
!= CONTIG_LEFT
) {
4301 ret
= ocfs2_append_rec_to_path(handle
, et
, insert_rec
,
4302 right_path
, &left_path
);
4309 ret
= ocfs2_insert_path(handle
, et
, left_path
, right_path
,
4316 out_update_clusters
:
4317 if (type
->ins_split
== SPLIT_NONE
)
4318 ocfs2_et_update_clusters(et
,
4319 le16_to_cpu(insert_rec
->e_leaf_clusters
));
4321 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4324 ocfs2_free_path(left_path
);
4325 ocfs2_free_path(right_path
);
4330 static enum ocfs2_contig_type
4331 ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree
*et
,
4332 struct ocfs2_path
*path
,
4333 struct ocfs2_extent_list
*el
, int index
,
4334 struct ocfs2_extent_rec
*split_rec
)
4337 enum ocfs2_contig_type ret
= CONTIG_NONE
;
4338 u32 left_cpos
, right_cpos
;
4339 struct ocfs2_extent_rec
*rec
= NULL
;
4340 struct ocfs2_extent_list
*new_el
;
4341 struct ocfs2_path
*left_path
= NULL
, *right_path
= NULL
;
4342 struct buffer_head
*bh
;
4343 struct ocfs2_extent_block
*eb
;
4344 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
4347 rec
= &el
->l_recs
[index
- 1];
4348 } else if (path
->p_tree_depth
> 0) {
4349 status
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
4353 if (left_cpos
!= 0) {
4354 left_path
= ocfs2_new_path_from_path(path
);
4358 status
= ocfs2_find_path(et
->et_ci
, left_path
,
4363 new_el
= path_leaf_el(left_path
);
4365 if (le16_to_cpu(new_el
->l_next_free_rec
) !=
4366 le16_to_cpu(new_el
->l_count
)) {
4367 bh
= path_leaf_bh(left_path
);
4368 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4370 "Extent block #%llu has an "
4371 "invalid l_next_free_rec of "
4372 "%d. It should have "
4373 "matched the l_count of %d",
4374 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4375 le16_to_cpu(new_el
->l_next_free_rec
),
4376 le16_to_cpu(new_el
->l_count
));
4380 rec
= &new_el
->l_recs
[
4381 le16_to_cpu(new_el
->l_next_free_rec
) - 1];
4386 * We're careful to check for an empty extent record here -
4387 * the merge code will know what to do if it sees one.
4390 if (index
== 1 && ocfs2_is_empty_extent(rec
)) {
4391 if (split_rec
->e_cpos
== el
->l_recs
[index
].e_cpos
)
4394 ret
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4399 if (index
< (le16_to_cpu(el
->l_next_free_rec
) - 1))
4400 rec
= &el
->l_recs
[index
+ 1];
4401 else if (le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
) &&
4402 path
->p_tree_depth
> 0) {
4403 status
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
4407 if (right_cpos
== 0)
4410 right_path
= ocfs2_new_path_from_path(path
);
4414 status
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
4418 new_el
= path_leaf_el(right_path
);
4419 rec
= &new_el
->l_recs
[0];
4420 if (ocfs2_is_empty_extent(rec
)) {
4421 if (le16_to_cpu(new_el
->l_next_free_rec
) <= 1) {
4422 bh
= path_leaf_bh(right_path
);
4423 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4425 "Extent block #%llu has an "
4426 "invalid l_next_free_rec of %d",
4427 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4428 le16_to_cpu(new_el
->l_next_free_rec
));
4432 rec
= &new_el
->l_recs
[1];
4437 enum ocfs2_contig_type contig_type
;
4439 contig_type
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4441 if (contig_type
== CONTIG_LEFT
&& ret
== CONTIG_RIGHT
)
4442 ret
= CONTIG_LEFTRIGHT
;
4443 else if (ret
== CONTIG_NONE
)
4449 ocfs2_free_path(left_path
);
4451 ocfs2_free_path(right_path
);
4456 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree
*et
,
4457 struct ocfs2_insert_type
*insert
,
4458 struct ocfs2_extent_list
*el
,
4459 struct ocfs2_extent_rec
*insert_rec
)
4462 enum ocfs2_contig_type contig_type
= CONTIG_NONE
;
4464 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4466 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
4467 contig_type
= ocfs2_et_extent_contig(et
, &el
->l_recs
[i
],
4469 if (contig_type
!= CONTIG_NONE
) {
4470 insert
->ins_contig_index
= i
;
4474 insert
->ins_contig
= contig_type
;
4476 if (insert
->ins_contig
!= CONTIG_NONE
) {
4477 struct ocfs2_extent_rec
*rec
=
4478 &el
->l_recs
[insert
->ins_contig_index
];
4479 unsigned int len
= le16_to_cpu(rec
->e_leaf_clusters
) +
4480 le16_to_cpu(insert_rec
->e_leaf_clusters
);
4483 * Caller might want us to limit the size of extents, don't
4484 * calculate contiguousness if we might exceed that limit.
4486 if (et
->et_max_leaf_clusters
&&
4487 (len
> et
->et_max_leaf_clusters
))
4488 insert
->ins_contig
= CONTIG_NONE
;
4493 * This should only be called against the righmost leaf extent list.
4495 * ocfs2_figure_appending_type() will figure out whether we'll have to
4496 * insert at the tail of the rightmost leaf.
4498 * This should also work against the root extent list for tree's with 0
4499 * depth. If we consider the root extent list to be the rightmost leaf node
4500 * then the logic here makes sense.
4502 static void ocfs2_figure_appending_type(struct ocfs2_insert_type
*insert
,
4503 struct ocfs2_extent_list
*el
,
4504 struct ocfs2_extent_rec
*insert_rec
)
4507 u32 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4508 struct ocfs2_extent_rec
*rec
;
4510 insert
->ins_appending
= APPEND_NONE
;
4512 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4514 if (!el
->l_next_free_rec
)
4515 goto set_tail_append
;
4517 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
4518 /* Were all records empty? */
4519 if (le16_to_cpu(el
->l_next_free_rec
) == 1)
4520 goto set_tail_append
;
4523 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
4524 rec
= &el
->l_recs
[i
];
4527 (le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)))
4528 goto set_tail_append
;
4533 insert
->ins_appending
= APPEND_TAIL
;
4537 * Helper function called at the begining of an insert.
4539 * This computes a few things that are commonly used in the process of
4540 * inserting into the btree:
4541 * - Whether the new extent is contiguous with an existing one.
4542 * - The current tree depth.
4543 * - Whether the insert is an appending one.
4544 * - The total # of free records in the tree.
4546 * All of the information is stored on the ocfs2_insert_type
4549 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree
*et
,
4550 struct buffer_head
**last_eb_bh
,
4551 struct ocfs2_extent_rec
*insert_rec
,
4553 struct ocfs2_insert_type
*insert
)
4556 struct ocfs2_extent_block
*eb
;
4557 struct ocfs2_extent_list
*el
;
4558 struct ocfs2_path
*path
= NULL
;
4559 struct buffer_head
*bh
= NULL
;
4561 insert
->ins_split
= SPLIT_NONE
;
4563 el
= et
->et_root_el
;
4564 insert
->ins_tree_depth
= le16_to_cpu(el
->l_tree_depth
);
4566 if (el
->l_tree_depth
) {
4568 * If we have tree depth, we read in the
4569 * rightmost extent block ahead of time as
4570 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4571 * may want it later.
4573 ret
= ocfs2_read_extent_block(et
->et_ci
,
4574 ocfs2_et_get_last_eb_blk(et
),
4580 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
4585 * Unless we have a contiguous insert, we'll need to know if
4586 * there is room left in our allocation tree for another
4589 * XXX: This test is simplistic, we can search for empty
4590 * extent records too.
4592 *free_records
= le16_to_cpu(el
->l_count
) -
4593 le16_to_cpu(el
->l_next_free_rec
);
4595 if (!insert
->ins_tree_depth
) {
4596 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4597 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4601 path
= ocfs2_new_path_from_et(et
);
4609 * In the case that we're inserting past what the tree
4610 * currently accounts for, ocfs2_find_path() will return for
4611 * us the rightmost tree path. This is accounted for below in
4612 * the appending code.
4614 ret
= ocfs2_find_path(et
->et_ci
, path
, le32_to_cpu(insert_rec
->e_cpos
));
4620 el
= path_leaf_el(path
);
4623 * Now that we have the path, there's two things we want to determine:
4624 * 1) Contiguousness (also set contig_index if this is so)
4626 * 2) Are we doing an append? We can trivially break this up
4627 * into two types of appends: simple record append, or a
4628 * rotate inside the tail leaf.
4630 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4633 * The insert code isn't quite ready to deal with all cases of
4634 * left contiguousness. Specifically, if it's an insert into
4635 * the 1st record in a leaf, it will require the adjustment of
4636 * cluster count on the last record of the path directly to it's
4637 * left. For now, just catch that case and fool the layers
4638 * above us. This works just fine for tree_depth == 0, which
4639 * is why we allow that above.
4641 if (insert
->ins_contig
== CONTIG_LEFT
&&
4642 insert
->ins_contig_index
== 0)
4643 insert
->ins_contig
= CONTIG_NONE
;
4646 * Ok, so we can simply compare against last_eb to figure out
4647 * whether the path doesn't exist. This will only happen in
4648 * the case that we're doing a tail append, so maybe we can
4649 * take advantage of that information somehow.
4651 if (ocfs2_et_get_last_eb_blk(et
) ==
4652 path_leaf_bh(path
)->b_blocknr
) {
4654 * Ok, ocfs2_find_path() returned us the rightmost
4655 * tree path. This might be an appending insert. There are
4657 * 1) We're doing a true append at the tail:
4658 * -This might even be off the end of the leaf
4659 * 2) We're "appending" by rotating in the tail
4661 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4665 ocfs2_free_path(path
);
4675 * Insert an extent into a btree.
4677 * The caller needs to update the owning btree's cluster count.
4679 int ocfs2_insert_extent(handle_t
*handle
,
4680 struct ocfs2_extent_tree
*et
,
4685 struct ocfs2_alloc_context
*meta_ac
)
4688 int uninitialized_var(free_records
);
4689 struct buffer_head
*last_eb_bh
= NULL
;
4690 struct ocfs2_insert_type insert
= {0, };
4691 struct ocfs2_extent_rec rec
;
4693 mlog(0, "add %u clusters at position %u to owner %llu\n",
4695 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
4697 memset(&rec
, 0, sizeof(rec
));
4698 rec
.e_cpos
= cpu_to_le32(cpos
);
4699 rec
.e_blkno
= cpu_to_le64(start_blk
);
4700 rec
.e_leaf_clusters
= cpu_to_le16(new_clusters
);
4701 rec
.e_flags
= flags
;
4702 status
= ocfs2_et_insert_check(et
, &rec
);
4708 status
= ocfs2_figure_insert_type(et
, &last_eb_bh
, &rec
,
4709 &free_records
, &insert
);
4715 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
4716 "Insert.contig_index: %d, Insert.free_records: %d, "
4717 "Insert.tree_depth: %d\n",
4718 insert
.ins_appending
, insert
.ins_contig
, insert
.ins_contig_index
,
4719 free_records
, insert
.ins_tree_depth
);
4721 if (insert
.ins_contig
== CONTIG_NONE
&& free_records
== 0) {
4722 status
= ocfs2_grow_tree(handle
, et
,
4723 &insert
.ins_tree_depth
, &last_eb_bh
,
4731 /* Finally, we can add clusters. This might rotate the tree for us. */
4732 status
= ocfs2_do_insert_extent(handle
, et
, &rec
, &insert
);
4736 ocfs2_et_extent_map_insert(et
, &rec
);
4746 * Allcate and add clusters into the extent b-tree.
4747 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4748 * The extent b-tree's root is specified by et, and
4749 * it is not limited to the file storage. Any extent tree can use this
4750 * function if it implements the proper ocfs2_extent_tree.
4752 int ocfs2_add_clusters_in_btree(handle_t
*handle
,
4753 struct ocfs2_extent_tree
*et
,
4754 u32
*logical_offset
,
4755 u32 clusters_to_add
,
4757 struct ocfs2_alloc_context
*data_ac
,
4758 struct ocfs2_alloc_context
*meta_ac
,
4759 enum ocfs2_alloc_restarted
*reason_ret
)
4763 enum ocfs2_alloc_restarted reason
= RESTART_NONE
;
4764 u32 bit_off
, num_bits
;
4767 struct ocfs2_super
*osb
=
4768 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
4770 BUG_ON(!clusters_to_add
);
4773 flags
= OCFS2_EXT_UNWRITTEN
;
4775 free_extents
= ocfs2_num_free_extents(osb
, et
);
4776 if (free_extents
< 0) {
4777 status
= free_extents
;
4782 /* there are two cases which could cause us to EAGAIN in the
4783 * we-need-more-metadata case:
4784 * 1) we haven't reserved *any*
4785 * 2) we are so fragmented, we've needed to add metadata too
4787 if (!free_extents
&& !meta_ac
) {
4788 mlog(0, "we haven't reserved any metadata!\n");
4790 reason
= RESTART_META
;
4792 } else if ((!free_extents
)
4793 && (ocfs2_alloc_context_bits_left(meta_ac
)
4794 < ocfs2_extend_meta_needed(et
->et_root_el
))) {
4795 mlog(0, "filesystem is really fragmented...\n");
4797 reason
= RESTART_META
;
4801 status
= __ocfs2_claim_clusters(osb
, handle
, data_ac
, 1,
4802 clusters_to_add
, &bit_off
, &num_bits
);
4804 if (status
!= -ENOSPC
)
4809 BUG_ON(num_bits
> clusters_to_add
);
4811 /* reserve our write early -- insert_extent may update the tree root */
4812 status
= ocfs2_et_root_journal_access(handle
, et
,
4813 OCFS2_JOURNAL_ACCESS_WRITE
);
4819 block
= ocfs2_clusters_to_blocks(osb
->sb
, bit_off
);
4820 mlog(0, "Allocating %u clusters at block %u for owner %llu\n",
4822 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
4823 status
= ocfs2_insert_extent(handle
, et
, *logical_offset
, block
,
4824 num_bits
, flags
, meta_ac
);
4830 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4832 clusters_to_add
-= num_bits
;
4833 *logical_offset
+= num_bits
;
4835 if (clusters_to_add
) {
4836 mlog(0, "need to alloc once more, wanted = %u\n",
4839 reason
= RESTART_TRANS
;
4845 *reason_ret
= reason
;
4849 static void ocfs2_make_right_split_rec(struct super_block
*sb
,
4850 struct ocfs2_extent_rec
*split_rec
,
4852 struct ocfs2_extent_rec
*rec
)
4854 u32 rec_cpos
= le32_to_cpu(rec
->e_cpos
);
4855 u32 rec_range
= rec_cpos
+ le16_to_cpu(rec
->e_leaf_clusters
);
4857 memset(split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4859 split_rec
->e_cpos
= cpu_to_le32(cpos
);
4860 split_rec
->e_leaf_clusters
= cpu_to_le16(rec_range
- cpos
);
4862 split_rec
->e_blkno
= rec
->e_blkno
;
4863 le64_add_cpu(&split_rec
->e_blkno
,
4864 ocfs2_clusters_to_blocks(sb
, cpos
- rec_cpos
));
4866 split_rec
->e_flags
= rec
->e_flags
;
4869 static int ocfs2_split_and_insert(handle_t
*handle
,
4870 struct ocfs2_extent_tree
*et
,
4871 struct ocfs2_path
*path
,
4872 struct buffer_head
**last_eb_bh
,
4874 struct ocfs2_extent_rec
*orig_split_rec
,
4875 struct ocfs2_alloc_context
*meta_ac
)
4878 unsigned int insert_range
, rec_range
, do_leftright
= 0;
4879 struct ocfs2_extent_rec tmprec
;
4880 struct ocfs2_extent_list
*rightmost_el
;
4881 struct ocfs2_extent_rec rec
;
4882 struct ocfs2_extent_rec split_rec
= *orig_split_rec
;
4883 struct ocfs2_insert_type insert
;
4884 struct ocfs2_extent_block
*eb
;
4888 * Store a copy of the record on the stack - it might move
4889 * around as the tree is manipulated below.
4891 rec
= path_leaf_el(path
)->l_recs
[split_index
];
4893 rightmost_el
= et
->et_root_el
;
4895 depth
= le16_to_cpu(rightmost_el
->l_tree_depth
);
4897 BUG_ON(!(*last_eb_bh
));
4898 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
4899 rightmost_el
= &eb
->h_list
;
4902 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4903 le16_to_cpu(rightmost_el
->l_count
)) {
4904 ret
= ocfs2_grow_tree(handle
, et
,
4905 &depth
, last_eb_bh
, meta_ac
);
4912 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4913 insert
.ins_appending
= APPEND_NONE
;
4914 insert
.ins_contig
= CONTIG_NONE
;
4915 insert
.ins_tree_depth
= depth
;
4917 insert_range
= le32_to_cpu(split_rec
.e_cpos
) +
4918 le16_to_cpu(split_rec
.e_leaf_clusters
);
4919 rec_range
= le32_to_cpu(rec
.e_cpos
) +
4920 le16_to_cpu(rec
.e_leaf_clusters
);
4922 if (split_rec
.e_cpos
== rec
.e_cpos
) {
4923 insert
.ins_split
= SPLIT_LEFT
;
4924 } else if (insert_range
== rec_range
) {
4925 insert
.ins_split
= SPLIT_RIGHT
;
4928 * Left/right split. We fake this as a right split
4929 * first and then make a second pass as a left split.
4931 insert
.ins_split
= SPLIT_RIGHT
;
4933 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
4934 &tmprec
, insert_range
, &rec
);
4938 BUG_ON(do_leftright
);
4942 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
4948 if (do_leftright
== 1) {
4950 struct ocfs2_extent_list
*el
;
4953 split_rec
= *orig_split_rec
;
4955 ocfs2_reinit_path(path
, 1);
4957 cpos
= le32_to_cpu(split_rec
.e_cpos
);
4958 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
4964 el
= path_leaf_el(path
);
4965 split_index
= ocfs2_search_extent_list(el
, cpos
);
4973 static int ocfs2_replace_extent_rec(handle_t
*handle
,
4974 struct ocfs2_extent_tree
*et
,
4975 struct ocfs2_path
*path
,
4976 struct ocfs2_extent_list
*el
,
4978 struct ocfs2_extent_rec
*split_rec
)
4982 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
4983 path_num_items(path
) - 1);
4989 el
->l_recs
[split_index
] = *split_rec
;
4991 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
4997 * Split part or all of the extent record at split_index in the leaf
4998 * pointed to by path. Merge with the contiguous extent record if needed.
5000 * Care is taken to handle contiguousness so as to not grow the tree.
5002 * meta_ac is not strictly necessary - we only truly need it if growth
5003 * of the tree is required. All other cases will degrade into a less
5004 * optimal tree layout.
5006 * last_eb_bh should be the rightmost leaf block for any extent
5007 * btree. Since a split may grow the tree or a merge might shrink it,
5008 * the caller cannot trust the contents of that buffer after this call.
5010 * This code is optimized for readability - several passes might be
5011 * made over certain portions of the tree. All of those blocks will
5012 * have been brought into cache (and pinned via the journal), so the
5013 * extra overhead is not expressed in terms of disk reads.
5015 int ocfs2_split_extent(handle_t
*handle
,
5016 struct ocfs2_extent_tree
*et
,
5017 struct ocfs2_path
*path
,
5019 struct ocfs2_extent_rec
*split_rec
,
5020 struct ocfs2_alloc_context
*meta_ac
,
5021 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5024 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5025 struct buffer_head
*last_eb_bh
= NULL
;
5026 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
5027 struct ocfs2_merge_ctxt ctxt
;
5028 struct ocfs2_extent_list
*rightmost_el
;
5030 if (le32_to_cpu(rec
->e_cpos
) > le32_to_cpu(split_rec
->e_cpos
) ||
5031 ((le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)) <
5032 (le32_to_cpu(split_rec
->e_cpos
) + le16_to_cpu(split_rec
->e_leaf_clusters
)))) {
5038 ctxt
.c_contig_type
= ocfs2_figure_merge_contig_type(et
, path
, el
,
5043 * The core merge / split code wants to know how much room is
5044 * left in this allocation tree, so we pass the
5045 * rightmost extent list.
5047 if (path
->p_tree_depth
) {
5048 struct ocfs2_extent_block
*eb
;
5050 ret
= ocfs2_read_extent_block(et
->et_ci
,
5051 ocfs2_et_get_last_eb_blk(et
),
5058 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5059 rightmost_el
= &eb
->h_list
;
5061 rightmost_el
= path_root_el(path
);
5063 if (rec
->e_cpos
== split_rec
->e_cpos
&&
5064 rec
->e_leaf_clusters
== split_rec
->e_leaf_clusters
)
5065 ctxt
.c_split_covers_rec
= 1;
5067 ctxt
.c_split_covers_rec
= 0;
5069 ctxt
.c_has_empty_extent
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
5071 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
5072 split_index
, ctxt
.c_contig_type
, ctxt
.c_has_empty_extent
,
5073 ctxt
.c_split_covers_rec
);
5075 if (ctxt
.c_contig_type
== CONTIG_NONE
) {
5076 if (ctxt
.c_split_covers_rec
)
5077 ret
= ocfs2_replace_extent_rec(handle
, et
, path
, el
,
5078 split_index
, split_rec
);
5080 ret
= ocfs2_split_and_insert(handle
, et
, path
,
5081 &last_eb_bh
, split_index
,
5082 split_rec
, meta_ac
);
5086 ret
= ocfs2_try_to_merge_extent(handle
, et
, path
,
5087 split_index
, split_rec
,
5099 * Change the flags of the already-existing extent at cpos for len clusters.
5101 * new_flags: the flags we want to set.
5102 * clear_flags: the flags we want to clear.
5103 * phys: the new physical offset we want this new extent starts from.
5105 * If the existing extent is larger than the request, initiate a
5106 * split. An attempt will be made at merging with adjacent extents.
5108 * The caller is responsible for passing down meta_ac if we'll need it.
5110 int ocfs2_change_extent_flag(handle_t
*handle
,
5111 struct ocfs2_extent_tree
*et
,
5112 u32 cpos
, u32 len
, u32 phys
,
5113 struct ocfs2_alloc_context
*meta_ac
,
5114 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5115 int new_flags
, int clear_flags
)
5118 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5119 u64 start_blkno
= ocfs2_clusters_to_blocks(sb
, phys
);
5120 struct ocfs2_extent_rec split_rec
;
5121 struct ocfs2_path
*left_path
= NULL
;
5122 struct ocfs2_extent_list
*el
;
5123 struct ocfs2_extent_rec
*rec
;
5125 left_path
= ocfs2_new_path_from_et(et
);
5132 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
5137 el
= path_leaf_el(left_path
);
5139 index
= ocfs2_search_extent_list(el
, cpos
);
5140 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5142 "Owner %llu has an extent at cpos %u which can no "
5143 "longer be found.\n",
5144 (unsigned long long)
5145 ocfs2_metadata_cache_owner(et
->et_ci
), cpos
);
5151 rec
= &el
->l_recs
[index
];
5152 if (new_flags
&& (rec
->e_flags
& new_flags
)) {
5153 mlog(ML_ERROR
, "Owner %llu tried to set %d flags on an "
5154 "extent that already had them",
5155 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5160 if (clear_flags
&& !(rec
->e_flags
& clear_flags
)) {
5161 mlog(ML_ERROR
, "Owner %llu tried to clear %d flags on an "
5162 "extent that didn't have them",
5163 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5168 memset(&split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
5169 split_rec
.e_cpos
= cpu_to_le32(cpos
);
5170 split_rec
.e_leaf_clusters
= cpu_to_le16(len
);
5171 split_rec
.e_blkno
= cpu_to_le64(start_blkno
);
5172 split_rec
.e_flags
= rec
->e_flags
;
5174 split_rec
.e_flags
|= new_flags
;
5176 split_rec
.e_flags
&= ~clear_flags
;
5178 ret
= ocfs2_split_extent(handle
, et
, left_path
,
5179 index
, &split_rec
, meta_ac
,
5185 ocfs2_free_path(left_path
);
5191 * Mark the already-existing extent at cpos as written for len clusters.
5192 * This removes the unwritten extent flag.
5194 * If the existing extent is larger than the request, initiate a
5195 * split. An attempt will be made at merging with adjacent extents.
5197 * The caller is responsible for passing down meta_ac if we'll need it.
5199 int ocfs2_mark_extent_written(struct inode
*inode
,
5200 struct ocfs2_extent_tree
*et
,
5201 handle_t
*handle
, u32 cpos
, u32 len
, u32 phys
,
5202 struct ocfs2_alloc_context
*meta_ac
,
5203 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5207 mlog(0, "Inode %lu cpos %u, len %u, phys clusters %u\n",
5208 inode
->i_ino
, cpos
, len
, phys
);
5210 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode
->i_sb
))) {
5211 ocfs2_error(inode
->i_sb
, "Inode %llu has unwritten extents "
5212 "that are being written to, but the feature bit "
5213 "is not set in the super block.",
5214 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
5220 * XXX: This should be fixed up so that we just re-insert the
5221 * next extent records.
5223 ocfs2_et_extent_map_truncate(et
, 0);
5225 ret
= ocfs2_change_extent_flag(handle
, et
, cpos
,
5226 len
, phys
, meta_ac
, dealloc
,
5227 0, OCFS2_EXT_UNWRITTEN
);
5235 static int ocfs2_split_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
5236 struct ocfs2_path
*path
,
5237 int index
, u32 new_range
,
5238 struct ocfs2_alloc_context
*meta_ac
)
5240 int ret
, depth
, credits
= handle
->h_buffer_credits
;
5241 struct buffer_head
*last_eb_bh
= NULL
;
5242 struct ocfs2_extent_block
*eb
;
5243 struct ocfs2_extent_list
*rightmost_el
, *el
;
5244 struct ocfs2_extent_rec split_rec
;
5245 struct ocfs2_extent_rec
*rec
;
5246 struct ocfs2_insert_type insert
;
5249 * Setup the record to split before we grow the tree.
5251 el
= path_leaf_el(path
);
5252 rec
= &el
->l_recs
[index
];
5253 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
5254 &split_rec
, new_range
, rec
);
5256 depth
= path
->p_tree_depth
;
5258 ret
= ocfs2_read_extent_block(et
->et_ci
,
5259 ocfs2_et_get_last_eb_blk(et
),
5266 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5267 rightmost_el
= &eb
->h_list
;
5269 rightmost_el
= path_leaf_el(path
);
5271 credits
+= path
->p_tree_depth
+
5272 ocfs2_extend_meta_needed(et
->et_root_el
);
5273 ret
= ocfs2_extend_trans(handle
, credits
);
5279 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
5280 le16_to_cpu(rightmost_el
->l_count
)) {
5281 ret
= ocfs2_grow_tree(handle
, et
, &depth
, &last_eb_bh
,
5289 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
5290 insert
.ins_appending
= APPEND_NONE
;
5291 insert
.ins_contig
= CONTIG_NONE
;
5292 insert
.ins_split
= SPLIT_RIGHT
;
5293 insert
.ins_tree_depth
= depth
;
5295 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
5304 static int ocfs2_truncate_rec(handle_t
*handle
,
5305 struct ocfs2_extent_tree
*et
,
5306 struct ocfs2_path
*path
, int index
,
5307 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5311 u32 left_cpos
, rec_range
, trunc_range
;
5312 int wants_rotate
= 0, is_rightmost_tree_rec
= 0;
5313 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5314 struct ocfs2_path
*left_path
= NULL
;
5315 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5316 struct ocfs2_extent_rec
*rec
;
5317 struct ocfs2_extent_block
*eb
;
5319 if (ocfs2_is_empty_extent(&el
->l_recs
[0]) && index
> 0) {
5320 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5329 if (index
== (le16_to_cpu(el
->l_next_free_rec
) - 1) &&
5330 path
->p_tree_depth
) {
5332 * Check whether this is the rightmost tree record. If
5333 * we remove all of this record or part of its right
5334 * edge then an update of the record lengths above it
5337 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
5338 if (eb
->h_next_leaf_blk
== 0)
5339 is_rightmost_tree_rec
= 1;
5342 rec
= &el
->l_recs
[index
];
5343 if (index
== 0 && path
->p_tree_depth
&&
5344 le32_to_cpu(rec
->e_cpos
) == cpos
) {
5346 * Changing the leftmost offset (via partial or whole
5347 * record truncate) of an interior (or rightmost) path
5348 * means we have to update the subtree that is formed
5349 * by this leaf and the one to it's left.
5351 * There are two cases we can skip:
5352 * 1) Path is the leftmost one in our btree.
5353 * 2) The leaf is rightmost and will be empty after
5354 * we remove the extent record - the rotate code
5355 * knows how to update the newly formed edge.
5358 ret
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
5364 if (left_cpos
&& le16_to_cpu(el
->l_next_free_rec
) > 1) {
5365 left_path
= ocfs2_new_path_from_path(path
);
5372 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
5381 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
5382 handle
->h_buffer_credits
,
5389 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
5395 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
5401 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5402 trunc_range
= cpos
+ len
;
5404 if (le32_to_cpu(rec
->e_cpos
) == cpos
&& rec_range
== trunc_range
) {
5407 memset(rec
, 0, sizeof(*rec
));
5408 ocfs2_cleanup_merge(el
, index
);
5411 next_free
= le16_to_cpu(el
->l_next_free_rec
);
5412 if (is_rightmost_tree_rec
&& next_free
> 1) {
5414 * We skip the edge update if this path will
5415 * be deleted by the rotate code.
5417 rec
= &el
->l_recs
[next_free
- 1];
5418 ocfs2_adjust_rightmost_records(handle
, et
, path
,
5421 } else if (le32_to_cpu(rec
->e_cpos
) == cpos
) {
5422 /* Remove leftmost portion of the record. */
5423 le32_add_cpu(&rec
->e_cpos
, len
);
5424 le64_add_cpu(&rec
->e_blkno
, ocfs2_clusters_to_blocks(sb
, len
));
5425 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5426 } else if (rec_range
== trunc_range
) {
5427 /* Remove rightmost portion of the record */
5428 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5429 if (is_rightmost_tree_rec
)
5430 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
5432 /* Caller should have trapped this. */
5433 mlog(ML_ERROR
, "Owner %llu: Invalid record truncate: (%u, %u) "
5435 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5436 le32_to_cpu(rec
->e_cpos
),
5437 le16_to_cpu(rec
->e_leaf_clusters
), cpos
, len
);
5444 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
5445 ocfs2_complete_edge_insert(handle
, left_path
, path
,
5449 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5451 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5458 ocfs2_free_path(left_path
);
5462 int ocfs2_remove_extent(handle_t
*handle
,
5463 struct ocfs2_extent_tree
*et
,
5465 struct ocfs2_alloc_context
*meta_ac
,
5466 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5469 u32 rec_range
, trunc_range
;
5470 struct ocfs2_extent_rec
*rec
;
5471 struct ocfs2_extent_list
*el
;
5472 struct ocfs2_path
*path
= NULL
;
5475 * XXX: Why are we truncating to 0 instead of wherever this
5478 ocfs2_et_extent_map_truncate(et
, 0);
5480 path
= ocfs2_new_path_from_et(et
);
5487 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5493 el
= path_leaf_el(path
);
5494 index
= ocfs2_search_extent_list(el
, cpos
);
5495 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5496 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5497 "Owner %llu has an extent at cpos %u which can no "
5498 "longer be found.\n",
5499 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5506 * We have 3 cases of extent removal:
5507 * 1) Range covers the entire extent rec
5508 * 2) Range begins or ends on one edge of the extent rec
5509 * 3) Range is in the middle of the extent rec (no shared edges)
5511 * For case 1 we remove the extent rec and left rotate to
5514 * For case 2 we just shrink the existing extent rec, with a
5515 * tree update if the shrinking edge is also the edge of an
5518 * For case 3 we do a right split to turn the extent rec into
5519 * something case 2 can handle.
5521 rec
= &el
->l_recs
[index
];
5522 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5523 trunc_range
= cpos
+ len
;
5525 BUG_ON(cpos
< le32_to_cpu(rec
->e_cpos
) || trunc_range
> rec_range
);
5527 mlog(0, "Owner %llu, remove (cpos %u, len %u). Existing index %d "
5528 "(cpos %u, len %u)\n",
5529 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5531 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
));
5533 if (le32_to_cpu(rec
->e_cpos
) == cpos
|| rec_range
== trunc_range
) {
5534 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5541 ret
= ocfs2_split_tree(handle
, et
, path
, index
,
5542 trunc_range
, meta_ac
);
5549 * The split could have manipulated the tree enough to
5550 * move the record location, so we have to look for it again.
5552 ocfs2_reinit_path(path
, 1);
5554 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5560 el
= path_leaf_el(path
);
5561 index
= ocfs2_search_extent_list(el
, cpos
);
5562 if (index
== -1 || index
>= le16_to_cpu(el
->l_next_free_rec
)) {
5563 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5564 "Owner %llu: split at cpos %u lost record.",
5565 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5572 * Double check our values here. If anything is fishy,
5573 * it's easier to catch it at the top level.
5575 rec
= &el
->l_recs
[index
];
5576 rec_range
= le32_to_cpu(rec
->e_cpos
) +
5577 ocfs2_rec_clusters(el
, rec
);
5578 if (rec_range
!= trunc_range
) {
5579 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5580 "Owner %llu: error after split at cpos %u"
5581 "trunc len %u, existing record is (%u,%u)",
5582 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5583 cpos
, len
, le32_to_cpu(rec
->e_cpos
),
5584 ocfs2_rec_clusters(el
, rec
));
5589 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5598 ocfs2_free_path(path
);
5602 int ocfs2_remove_btree_range(struct inode
*inode
,
5603 struct ocfs2_extent_tree
*et
,
5604 u32 cpos
, u32 phys_cpos
, u32 len
,
5605 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5608 u64 phys_blkno
= ocfs2_clusters_to_blocks(inode
->i_sb
, phys_cpos
);
5609 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
5610 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5612 struct ocfs2_alloc_context
*meta_ac
= NULL
;
5614 ret
= ocfs2_lock_allocators(inode
, et
, 0, 1, NULL
, &meta_ac
);
5620 mutex_lock(&tl_inode
->i_mutex
);
5622 if (ocfs2_truncate_log_needs_flush(osb
)) {
5623 ret
= __ocfs2_flush_truncate_log(osb
);
5630 handle
= ocfs2_start_trans(osb
, ocfs2_remove_extent_credits(osb
->sb
));
5631 if (IS_ERR(handle
)) {
5632 ret
= PTR_ERR(handle
);
5637 ret
= ocfs2_et_root_journal_access(handle
, et
,
5638 OCFS2_JOURNAL_ACCESS_WRITE
);
5644 dquot_free_space_nodirty(inode
,
5645 ocfs2_clusters_to_bytes(inode
->i_sb
, len
));
5647 ret
= ocfs2_remove_extent(handle
, et
, cpos
, len
, meta_ac
, dealloc
);
5653 ocfs2_et_update_clusters(et
, -len
);
5655 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
5657 ret
= ocfs2_truncate_log_append(osb
, handle
, phys_blkno
, len
);
5662 ocfs2_commit_trans(osb
, handle
);
5664 mutex_unlock(&tl_inode
->i_mutex
);
5667 ocfs2_free_alloc_context(meta_ac
);
5672 int ocfs2_truncate_log_needs_flush(struct ocfs2_super
*osb
)
5674 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5675 struct ocfs2_dinode
*di
;
5676 struct ocfs2_truncate_log
*tl
;
5678 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5679 tl
= &di
->id2
.i_dealloc
;
5681 mlog_bug_on_msg(le16_to_cpu(tl
->tl_used
) > le16_to_cpu(tl
->tl_count
),
5682 "slot %d, invalid truncate log parameters: used = "
5683 "%u, count = %u\n", osb
->slot_num
,
5684 le16_to_cpu(tl
->tl_used
), le16_to_cpu(tl
->tl_count
));
5685 return le16_to_cpu(tl
->tl_used
) == le16_to_cpu(tl
->tl_count
);
5688 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log
*tl
,
5689 unsigned int new_start
)
5691 unsigned int tail_index
;
5692 unsigned int current_tail
;
5694 /* No records, nothing to coalesce */
5695 if (!le16_to_cpu(tl
->tl_used
))
5698 tail_index
= le16_to_cpu(tl
->tl_used
) - 1;
5699 current_tail
= le32_to_cpu(tl
->tl_recs
[tail_index
].t_start
);
5700 current_tail
+= le32_to_cpu(tl
->tl_recs
[tail_index
].t_clusters
);
5702 return current_tail
== new_start
;
5705 int ocfs2_truncate_log_append(struct ocfs2_super
*osb
,
5708 unsigned int num_clusters
)
5711 unsigned int start_cluster
, tl_count
;
5712 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5713 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5714 struct ocfs2_dinode
*di
;
5715 struct ocfs2_truncate_log
*tl
;
5717 mlog_entry("start_blk = %llu, num_clusters = %u\n",
5718 (unsigned long long)start_blk
, num_clusters
);
5720 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
5722 start_cluster
= ocfs2_blocks_to_clusters(osb
->sb
, start_blk
);
5724 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5726 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5727 * by the underlying call to ocfs2_read_inode_block(), so any
5728 * corruption is a code bug */
5729 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
5731 tl
= &di
->id2
.i_dealloc
;
5732 tl_count
= le16_to_cpu(tl
->tl_count
);
5733 mlog_bug_on_msg(tl_count
> ocfs2_truncate_recs_per_inode(osb
->sb
) ||
5735 "Truncate record count on #%llu invalid "
5736 "wanted %u, actual %u\n",
5737 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5738 ocfs2_truncate_recs_per_inode(osb
->sb
),
5739 le16_to_cpu(tl
->tl_count
));
5741 /* Caller should have known to flush before calling us. */
5742 index
= le16_to_cpu(tl
->tl_used
);
5743 if (index
>= tl_count
) {
5749 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5750 OCFS2_JOURNAL_ACCESS_WRITE
);
5756 mlog(0, "Log truncate of %u clusters starting at cluster %u to "
5757 "%llu (index = %d)\n", num_clusters
, start_cluster
,
5758 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
, index
);
5760 if (ocfs2_truncate_log_can_coalesce(tl
, start_cluster
)) {
5762 * Move index back to the record we are coalescing with.
5763 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5767 num_clusters
+= le32_to_cpu(tl
->tl_recs
[index
].t_clusters
);
5768 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
5769 index
, le32_to_cpu(tl
->tl_recs
[index
].t_start
),
5772 tl
->tl_recs
[index
].t_start
= cpu_to_le32(start_cluster
);
5773 tl
->tl_used
= cpu_to_le16(index
+ 1);
5775 tl
->tl_recs
[index
].t_clusters
= cpu_to_le32(num_clusters
);
5777 ocfs2_journal_dirty(handle
, tl_bh
);
5784 static int ocfs2_replay_truncate_records(struct ocfs2_super
*osb
,
5786 struct inode
*data_alloc_inode
,
5787 struct buffer_head
*data_alloc_bh
)
5791 unsigned int num_clusters
;
5793 struct ocfs2_truncate_rec rec
;
5794 struct ocfs2_dinode
*di
;
5795 struct ocfs2_truncate_log
*tl
;
5796 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5797 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5801 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5802 tl
= &di
->id2
.i_dealloc
;
5803 i
= le16_to_cpu(tl
->tl_used
) - 1;
5805 /* Caller has given us at least enough credits to
5806 * update the truncate log dinode */
5807 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5808 OCFS2_JOURNAL_ACCESS_WRITE
);
5814 tl
->tl_used
= cpu_to_le16(i
);
5816 ocfs2_journal_dirty(handle
, tl_bh
);
5818 /* TODO: Perhaps we can calculate the bulk of the
5819 * credits up front rather than extending like
5821 status
= ocfs2_extend_trans(handle
,
5822 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC
);
5828 rec
= tl
->tl_recs
[i
];
5829 start_blk
= ocfs2_clusters_to_blocks(data_alloc_inode
->i_sb
,
5830 le32_to_cpu(rec
.t_start
));
5831 num_clusters
= le32_to_cpu(rec
.t_clusters
);
5833 /* if start_blk is not set, we ignore the record as
5836 mlog(0, "free record %d, start = %u, clusters = %u\n",
5837 i
, le32_to_cpu(rec
.t_start
), num_clusters
);
5839 status
= ocfs2_free_clusters(handle
, data_alloc_inode
,
5840 data_alloc_bh
, start_blk
,
5855 /* Expects you to already be holding tl_inode->i_mutex */
5856 int __ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5859 unsigned int num_to_flush
;
5861 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5862 struct inode
*data_alloc_inode
= NULL
;
5863 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5864 struct buffer_head
*data_alloc_bh
= NULL
;
5865 struct ocfs2_dinode
*di
;
5866 struct ocfs2_truncate_log
*tl
;
5870 BUG_ON(mutex_trylock(&tl_inode
->i_mutex
));
5872 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5874 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5875 * by the underlying call to ocfs2_read_inode_block(), so any
5876 * corruption is a code bug */
5877 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
5879 tl
= &di
->id2
.i_dealloc
;
5880 num_to_flush
= le16_to_cpu(tl
->tl_used
);
5881 mlog(0, "Flush %u records from truncate log #%llu\n",
5882 num_to_flush
, (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
);
5883 if (!num_to_flush
) {
5888 data_alloc_inode
= ocfs2_get_system_file_inode(osb
,
5889 GLOBAL_BITMAP_SYSTEM_INODE
,
5890 OCFS2_INVALID_SLOT
);
5891 if (!data_alloc_inode
) {
5893 mlog(ML_ERROR
, "Could not get bitmap inode!\n");
5897 mutex_lock(&data_alloc_inode
->i_mutex
);
5899 status
= ocfs2_inode_lock(data_alloc_inode
, &data_alloc_bh
, 1);
5905 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
5906 if (IS_ERR(handle
)) {
5907 status
= PTR_ERR(handle
);
5912 status
= ocfs2_replay_truncate_records(osb
, handle
, data_alloc_inode
,
5917 ocfs2_commit_trans(osb
, handle
);
5920 brelse(data_alloc_bh
);
5921 ocfs2_inode_unlock(data_alloc_inode
, 1);
5924 mutex_unlock(&data_alloc_inode
->i_mutex
);
5925 iput(data_alloc_inode
);
5932 int ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5935 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5937 mutex_lock(&tl_inode
->i_mutex
);
5938 status
= __ocfs2_flush_truncate_log(osb
);
5939 mutex_unlock(&tl_inode
->i_mutex
);
5944 static void ocfs2_truncate_log_worker(struct work_struct
*work
)
5947 struct ocfs2_super
*osb
=
5948 container_of(work
, struct ocfs2_super
,
5949 osb_truncate_log_wq
.work
);
5953 status
= ocfs2_flush_truncate_log(osb
);
5957 ocfs2_init_steal_slots(osb
);
5962 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
5963 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super
*osb
,
5966 if (osb
->osb_tl_inode
) {
5967 /* We want to push off log flushes while truncates are
5970 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
5972 queue_delayed_work(ocfs2_wq
, &osb
->osb_truncate_log_wq
,
5973 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL
);
5977 static int ocfs2_get_truncate_log_info(struct ocfs2_super
*osb
,
5979 struct inode
**tl_inode
,
5980 struct buffer_head
**tl_bh
)
5983 struct inode
*inode
= NULL
;
5984 struct buffer_head
*bh
= NULL
;
5986 inode
= ocfs2_get_system_file_inode(osb
,
5987 TRUNCATE_LOG_SYSTEM_INODE
,
5991 mlog(ML_ERROR
, "Could not get load truncate log inode!\n");
5995 status
= ocfs2_read_inode_block(inode
, &bh
);
6009 /* called during the 1st stage of node recovery. we stamp a clean
6010 * truncate log and pass back a copy for processing later. if the
6011 * truncate log does not require processing, a *tl_copy is set to
6013 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super
*osb
,
6015 struct ocfs2_dinode
**tl_copy
)
6018 struct inode
*tl_inode
= NULL
;
6019 struct buffer_head
*tl_bh
= NULL
;
6020 struct ocfs2_dinode
*di
;
6021 struct ocfs2_truncate_log
*tl
;
6025 mlog(0, "recover truncate log from slot %d\n", slot_num
);
6027 status
= ocfs2_get_truncate_log_info(osb
, slot_num
, &tl_inode
, &tl_bh
);
6033 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
6035 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6036 * validated by the underlying call to ocfs2_read_inode_block(),
6037 * so any corruption is a code bug */
6038 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
6040 tl
= &di
->id2
.i_dealloc
;
6041 if (le16_to_cpu(tl
->tl_used
)) {
6042 mlog(0, "We'll have %u logs to recover\n",
6043 le16_to_cpu(tl
->tl_used
));
6045 *tl_copy
= kmalloc(tl_bh
->b_size
, GFP_KERNEL
);
6052 /* Assuming the write-out below goes well, this copy
6053 * will be passed back to recovery for processing. */
6054 memcpy(*tl_copy
, tl_bh
->b_data
, tl_bh
->b_size
);
6056 /* All we need to do to clear the truncate log is set
6060 ocfs2_compute_meta_ecc(osb
->sb
, tl_bh
->b_data
, &di
->i_check
);
6061 status
= ocfs2_write_block(osb
, tl_bh
, INODE_CACHE(tl_inode
));
6073 if (status
< 0 && (*tl_copy
)) {
6082 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super
*osb
,
6083 struct ocfs2_dinode
*tl_copy
)
6087 unsigned int clusters
, num_recs
, start_cluster
;
6090 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6091 struct ocfs2_truncate_log
*tl
;
6095 if (OCFS2_I(tl_inode
)->ip_blkno
== le64_to_cpu(tl_copy
->i_blkno
)) {
6096 mlog(ML_ERROR
, "Asked to recover my own truncate log!\n");
6100 tl
= &tl_copy
->id2
.i_dealloc
;
6101 num_recs
= le16_to_cpu(tl
->tl_used
);
6102 mlog(0, "cleanup %u records from %llu\n", num_recs
,
6103 (unsigned long long)le64_to_cpu(tl_copy
->i_blkno
));
6105 mutex_lock(&tl_inode
->i_mutex
);
6106 for(i
= 0; i
< num_recs
; i
++) {
6107 if (ocfs2_truncate_log_needs_flush(osb
)) {
6108 status
= __ocfs2_flush_truncate_log(osb
);
6115 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
6116 if (IS_ERR(handle
)) {
6117 status
= PTR_ERR(handle
);
6122 clusters
= le32_to_cpu(tl
->tl_recs
[i
].t_clusters
);
6123 start_cluster
= le32_to_cpu(tl
->tl_recs
[i
].t_start
);
6124 start_blk
= ocfs2_clusters_to_blocks(osb
->sb
, start_cluster
);
6126 status
= ocfs2_truncate_log_append(osb
, handle
,
6127 start_blk
, clusters
);
6128 ocfs2_commit_trans(osb
, handle
);
6136 mutex_unlock(&tl_inode
->i_mutex
);
6142 void ocfs2_truncate_log_shutdown(struct ocfs2_super
*osb
)
6145 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6150 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
6151 flush_workqueue(ocfs2_wq
);
6153 status
= ocfs2_flush_truncate_log(osb
);
6157 brelse(osb
->osb_tl_bh
);
6158 iput(osb
->osb_tl_inode
);
6164 int ocfs2_truncate_log_init(struct ocfs2_super
*osb
)
6167 struct inode
*tl_inode
= NULL
;
6168 struct buffer_head
*tl_bh
= NULL
;
6172 status
= ocfs2_get_truncate_log_info(osb
,
6179 /* ocfs2_truncate_log_shutdown keys on the existence of
6180 * osb->osb_tl_inode so we don't set any of the osb variables
6181 * until we're sure all is well. */
6182 INIT_DELAYED_WORK(&osb
->osb_truncate_log_wq
,
6183 ocfs2_truncate_log_worker
);
6184 osb
->osb_tl_bh
= tl_bh
;
6185 osb
->osb_tl_inode
= tl_inode
;
6192 * Delayed de-allocation of suballocator blocks.
6194 * Some sets of block de-allocations might involve multiple suballocator inodes.
6196 * The locking for this can get extremely complicated, especially when
6197 * the suballocator inodes to delete from aren't known until deep
6198 * within an unrelated codepath.
6200 * ocfs2_extent_block structures are a good example of this - an inode
6201 * btree could have been grown by any number of nodes each allocating
6202 * out of their own suballoc inode.
6204 * These structures allow the delay of block de-allocation until a
6205 * later time, when locking of multiple cluster inodes won't cause
6210 * Describe a single bit freed from a suballocator. For the block
6211 * suballocators, it represents one block. For the global cluster
6212 * allocator, it represents some clusters and free_bit indicates
6215 struct ocfs2_cached_block_free
{
6216 struct ocfs2_cached_block_free
*free_next
;
6218 unsigned int free_bit
;
6221 struct ocfs2_per_slot_free_list
{
6222 struct ocfs2_per_slot_free_list
*f_next_suballocator
;
6225 struct ocfs2_cached_block_free
*f_first
;
6228 static int ocfs2_free_cached_blocks(struct ocfs2_super
*osb
,
6231 struct ocfs2_cached_block_free
*head
)
6236 struct inode
*inode
;
6237 struct buffer_head
*di_bh
= NULL
;
6238 struct ocfs2_cached_block_free
*tmp
;
6240 inode
= ocfs2_get_system_file_inode(osb
, sysfile_type
, slot
);
6247 mutex_lock(&inode
->i_mutex
);
6249 ret
= ocfs2_inode_lock(inode
, &di_bh
, 1);
6255 handle
= ocfs2_start_trans(osb
, OCFS2_SUBALLOC_FREE
);
6256 if (IS_ERR(handle
)) {
6257 ret
= PTR_ERR(handle
);
6263 bg_blkno
= ocfs2_which_suballoc_group(head
->free_blk
,
6265 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
6266 head
->free_bit
, (unsigned long long)head
->free_blk
);
6268 ret
= ocfs2_free_suballoc_bits(handle
, inode
, di_bh
,
6269 head
->free_bit
, bg_blkno
, 1);
6275 ret
= ocfs2_extend_trans(handle
, OCFS2_SUBALLOC_FREE
);
6282 head
= head
->free_next
;
6287 ocfs2_commit_trans(osb
, handle
);
6290 ocfs2_inode_unlock(inode
, 1);
6293 mutex_unlock(&inode
->i_mutex
);
6297 /* Premature exit may have left some dangling items. */
6299 head
= head
->free_next
;
6306 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6307 u64 blkno
, unsigned int bit
)
6310 struct ocfs2_cached_block_free
*item
;
6312 item
= kmalloc(sizeof(*item
), GFP_NOFS
);
6319 mlog(0, "Insert clusters: (bit %u, blk %llu)\n",
6320 bit
, (unsigned long long)blkno
);
6322 item
->free_blk
= blkno
;
6323 item
->free_bit
= bit
;
6324 item
->free_next
= ctxt
->c_global_allocator
;
6326 ctxt
->c_global_allocator
= item
;
6330 static int ocfs2_free_cached_clusters(struct ocfs2_super
*osb
,
6331 struct ocfs2_cached_block_free
*head
)
6333 struct ocfs2_cached_block_free
*tmp
;
6334 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6338 mutex_lock(&tl_inode
->i_mutex
);
6341 if (ocfs2_truncate_log_needs_flush(osb
)) {
6342 ret
= __ocfs2_flush_truncate_log(osb
);
6349 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
6350 if (IS_ERR(handle
)) {
6351 ret
= PTR_ERR(handle
);
6356 ret
= ocfs2_truncate_log_append(osb
, handle
, head
->free_blk
,
6359 ocfs2_commit_trans(osb
, handle
);
6361 head
= head
->free_next
;
6370 mutex_unlock(&tl_inode
->i_mutex
);
6373 /* Premature exit may have left some dangling items. */
6375 head
= head
->free_next
;
6382 int ocfs2_run_deallocs(struct ocfs2_super
*osb
,
6383 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
6386 struct ocfs2_per_slot_free_list
*fl
;
6391 while (ctxt
->c_first_suballocator
) {
6392 fl
= ctxt
->c_first_suballocator
;
6395 mlog(0, "Free items: (type %u, slot %d)\n",
6396 fl
->f_inode_type
, fl
->f_slot
);
6397 ret2
= ocfs2_free_cached_blocks(osb
,
6407 ctxt
->c_first_suballocator
= fl
->f_next_suballocator
;
6411 if (ctxt
->c_global_allocator
) {
6412 ret2
= ocfs2_free_cached_clusters(osb
,
6413 ctxt
->c_global_allocator
);
6419 ctxt
->c_global_allocator
= NULL
;
6425 static struct ocfs2_per_slot_free_list
*
6426 ocfs2_find_per_slot_free_list(int type
,
6428 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
6430 struct ocfs2_per_slot_free_list
*fl
= ctxt
->c_first_suballocator
;
6433 if (fl
->f_inode_type
== type
&& fl
->f_slot
== slot
)
6436 fl
= fl
->f_next_suballocator
;
6439 fl
= kmalloc(sizeof(*fl
), GFP_NOFS
);
6441 fl
->f_inode_type
= type
;
6444 fl
->f_next_suballocator
= ctxt
->c_first_suballocator
;
6446 ctxt
->c_first_suballocator
= fl
;
6451 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6452 int type
, int slot
, u64 blkno
,
6456 struct ocfs2_per_slot_free_list
*fl
;
6457 struct ocfs2_cached_block_free
*item
;
6459 fl
= ocfs2_find_per_slot_free_list(type
, slot
, ctxt
);
6466 item
= kmalloc(sizeof(*item
), GFP_NOFS
);
6473 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
6474 type
, slot
, bit
, (unsigned long long)blkno
);
6476 item
->free_blk
= blkno
;
6477 item
->free_bit
= bit
;
6478 item
->free_next
= fl
->f_first
;
6487 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6488 struct ocfs2_extent_block
*eb
)
6490 return ocfs2_cache_block_dealloc(ctxt
, EXTENT_ALLOC_SYSTEM_INODE
,
6491 le16_to_cpu(eb
->h_suballoc_slot
),
6492 le64_to_cpu(eb
->h_blkno
),
6493 le16_to_cpu(eb
->h_suballoc_bit
));
6496 /* This function will figure out whether the currently last extent
6497 * block will be deleted, and if it will, what the new last extent
6498 * block will be so we can update his h_next_leaf_blk field, as well
6499 * as the dinodes i_last_eb_blk */
6500 static int ocfs2_find_new_last_ext_blk(struct inode
*inode
,
6501 unsigned int clusters_to_del
,
6502 struct ocfs2_path
*path
,
6503 struct buffer_head
**new_last_eb
)
6505 int next_free
, ret
= 0;
6507 struct ocfs2_extent_rec
*rec
;
6508 struct ocfs2_extent_block
*eb
;
6509 struct ocfs2_extent_list
*el
;
6510 struct buffer_head
*bh
= NULL
;
6512 *new_last_eb
= NULL
;
6514 /* we have no tree, so of course, no last_eb. */
6515 if (!path
->p_tree_depth
)
6518 /* trunc to zero special case - this makes tree_depth = 0
6519 * regardless of what it is. */
6520 if (OCFS2_I(inode
)->ip_clusters
== clusters_to_del
)
6523 el
= path_leaf_el(path
);
6524 BUG_ON(!el
->l_next_free_rec
);
6527 * Make sure that this extent list will actually be empty
6528 * after we clear away the data. We can shortcut out if
6529 * there's more than one non-empty extent in the
6530 * list. Otherwise, a check of the remaining extent is
6533 next_free
= le16_to_cpu(el
->l_next_free_rec
);
6535 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
6539 /* We may have a valid extent in index 1, check it. */
6541 rec
= &el
->l_recs
[1];
6544 * Fall through - no more nonempty extents, so we want
6545 * to delete this leaf.
6551 rec
= &el
->l_recs
[0];
6556 * Check it we'll only be trimming off the end of this
6559 if (le16_to_cpu(rec
->e_leaf_clusters
) > clusters_to_del
)
6563 ret
= ocfs2_find_cpos_for_left_leaf(inode
->i_sb
, path
, &cpos
);
6569 ret
= ocfs2_find_leaf(INODE_CACHE(inode
), path_root_el(path
), cpos
, &bh
);
6575 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
6578 /* ocfs2_find_leaf() gets the eb from ocfs2_read_extent_block().
6579 * Any corruption is a code bug. */
6580 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
6583 get_bh(*new_last_eb
);
6584 mlog(0, "returning block %llu, (cpos: %u)\n",
6585 (unsigned long long)le64_to_cpu(eb
->h_blkno
), cpos
);
6593 * Trim some clusters off the rightmost edge of a tree. Only called
6596 * The caller needs to:
6597 * - start journaling of each path component.
6598 * - compute and fully set up any new last ext block
6600 static int ocfs2_trim_tree(struct inode
*inode
, struct ocfs2_path
*path
,
6601 handle_t
*handle
, struct ocfs2_truncate_context
*tc
,
6602 u32 clusters_to_del
, u64
*delete_start
, u8
*flags
)
6604 int ret
, i
, index
= path
->p_tree_depth
;
6607 struct buffer_head
*bh
;
6608 struct ocfs2_extent_list
*el
;
6609 struct ocfs2_extent_rec
*rec
;
6614 while (index
>= 0) {
6615 bh
= path
->p_node
[index
].bh
;
6616 el
= path
->p_node
[index
].el
;
6618 mlog(0, "traveling tree (index = %d, block = %llu)\n",
6619 index
, (unsigned long long)bh
->b_blocknr
);
6621 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
6624 (path
->p_tree_depth
- le16_to_cpu(el
->l_tree_depth
))) {
6625 ocfs2_error(inode
->i_sb
,
6626 "Inode %lu has invalid ext. block %llu",
6628 (unsigned long long)bh
->b_blocknr
);
6634 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
6635 rec
= &el
->l_recs
[i
];
6637 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
6638 "next = %u\n", i
, le32_to_cpu(rec
->e_cpos
),
6639 ocfs2_rec_clusters(el
, rec
),
6640 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
6641 le16_to_cpu(el
->l_next_free_rec
));
6643 BUG_ON(ocfs2_rec_clusters(el
, rec
) < clusters_to_del
);
6645 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
6647 * If the leaf block contains a single empty
6648 * extent and no records, we can just remove
6651 if (i
== 0 && ocfs2_is_empty_extent(rec
)) {
6653 sizeof(struct ocfs2_extent_rec
));
6654 el
->l_next_free_rec
= cpu_to_le16(0);
6660 * Remove any empty extents by shifting things
6661 * left. That should make life much easier on
6662 * the code below. This condition is rare
6663 * enough that we shouldn't see a performance
6666 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
6667 le16_add_cpu(&el
->l_next_free_rec
, -1);
6670 i
< le16_to_cpu(el
->l_next_free_rec
); i
++)
6671 el
->l_recs
[i
] = el
->l_recs
[i
+ 1];
6673 memset(&el
->l_recs
[i
], 0,
6674 sizeof(struct ocfs2_extent_rec
));
6677 * We've modified our extent list. The
6678 * simplest way to handle this change
6679 * is to being the search from the
6682 goto find_tail_record
;
6685 le16_add_cpu(&rec
->e_leaf_clusters
, -clusters_to_del
);
6688 * We'll use "new_edge" on our way back up the
6689 * tree to know what our rightmost cpos is.
6691 new_edge
= le16_to_cpu(rec
->e_leaf_clusters
);
6692 new_edge
+= le32_to_cpu(rec
->e_cpos
);
6695 * The caller will use this to delete data blocks.
6697 *delete_start
= le64_to_cpu(rec
->e_blkno
)
6698 + ocfs2_clusters_to_blocks(inode
->i_sb
,
6699 le16_to_cpu(rec
->e_leaf_clusters
));
6700 *flags
= rec
->e_flags
;
6703 * If it's now empty, remove this record.
6705 if (le16_to_cpu(rec
->e_leaf_clusters
) == 0) {
6707 sizeof(struct ocfs2_extent_rec
));
6708 le16_add_cpu(&el
->l_next_free_rec
, -1);
6711 if (le64_to_cpu(rec
->e_blkno
) == deleted_eb
) {
6713 sizeof(struct ocfs2_extent_rec
));
6714 le16_add_cpu(&el
->l_next_free_rec
, -1);
6719 /* Can this actually happen? */
6720 if (le16_to_cpu(el
->l_next_free_rec
) == 0)
6724 * We never actually deleted any clusters
6725 * because our leaf was empty. There's no
6726 * reason to adjust the rightmost edge then.
6731 rec
->e_int_clusters
= cpu_to_le32(new_edge
);
6732 le32_add_cpu(&rec
->e_int_clusters
,
6733 -le32_to_cpu(rec
->e_cpos
));
6736 * A deleted child record should have been
6739 BUG_ON(le32_to_cpu(rec
->e_int_clusters
) == 0);
6743 ocfs2_journal_dirty(handle
, bh
);
6745 mlog(0, "extent list container %llu, after: record %d: "
6746 "(%u, %u, %llu), next = %u.\n",
6747 (unsigned long long)bh
->b_blocknr
, i
,
6748 le32_to_cpu(rec
->e_cpos
), ocfs2_rec_clusters(el
, rec
),
6749 (unsigned long long)le64_to_cpu(rec
->e_blkno
),
6750 le16_to_cpu(el
->l_next_free_rec
));
6753 * We must be careful to only attempt delete of an
6754 * extent block (and not the root inode block).
6756 if (index
> 0 && le16_to_cpu(el
->l_next_free_rec
) == 0) {
6757 struct ocfs2_extent_block
*eb
=
6758 (struct ocfs2_extent_block
*)bh
->b_data
;
6761 * Save this for use when processing the
6764 deleted_eb
= le64_to_cpu(eb
->h_blkno
);
6766 mlog(0, "deleting this extent block.\n");
6768 ocfs2_remove_from_cache(INODE_CACHE(inode
), bh
);
6770 BUG_ON(ocfs2_rec_clusters(el
, &el
->l_recs
[0]));
6771 BUG_ON(le32_to_cpu(el
->l_recs
[0].e_cpos
));
6772 BUG_ON(le64_to_cpu(el
->l_recs
[0].e_blkno
));
6774 ret
= ocfs2_cache_extent_block_free(&tc
->tc_dealloc
, eb
);
6775 /* An error here is not fatal. */
6790 static int ocfs2_do_truncate(struct ocfs2_super
*osb
,
6791 unsigned int clusters_to_del
,
6792 struct inode
*inode
,
6793 struct buffer_head
*fe_bh
,
6795 struct ocfs2_truncate_context
*tc
,
6796 struct ocfs2_path
*path
,
6797 struct ocfs2_alloc_context
*meta_ac
)
6800 struct ocfs2_dinode
*fe
;
6801 struct ocfs2_extent_block
*last_eb
= NULL
;
6802 struct ocfs2_extent_list
*el
;
6803 struct buffer_head
*last_eb_bh
= NULL
;
6807 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
6809 status
= ocfs2_find_new_last_ext_blk(inode
, clusters_to_del
,
6817 * Each component will be touched, so we might as well journal
6818 * here to avoid having to handle errors later.
6820 status
= ocfs2_journal_access_path(INODE_CACHE(inode
), handle
, path
);
6827 status
= ocfs2_journal_access_eb(handle
, INODE_CACHE(inode
), last_eb_bh
,
6828 OCFS2_JOURNAL_ACCESS_WRITE
);
6834 last_eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
6837 el
= &(fe
->id2
.i_list
);
6840 * Lower levels depend on this never happening, but it's best
6841 * to check it up here before changing the tree.
6843 if (el
->l_tree_depth
&& el
->l_recs
[0].e_int_clusters
== 0) {
6844 ocfs2_error(inode
->i_sb
,
6845 "Inode %lu has an empty extent record, depth %u\n",
6846 inode
->i_ino
, le16_to_cpu(el
->l_tree_depth
));
6851 dquot_free_space_nodirty(inode
,
6852 ocfs2_clusters_to_bytes(osb
->sb
, clusters_to_del
));
6853 spin_lock(&OCFS2_I(inode
)->ip_lock
);
6854 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
) -
6856 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
6857 le32_add_cpu(&fe
->i_clusters
, -clusters_to_del
);
6858 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
6860 status
= ocfs2_trim_tree(inode
, path
, handle
, tc
,
6861 clusters_to_del
, &delete_blk
, &rec_flags
);
6867 if (le32_to_cpu(fe
->i_clusters
) == 0) {
6868 /* trunc to zero is a special case. */
6869 el
->l_tree_depth
= 0;
6870 fe
->i_last_eb_blk
= 0;
6872 fe
->i_last_eb_blk
= last_eb
->h_blkno
;
6874 ocfs2_journal_dirty(handle
, fe_bh
);
6877 /* If there will be a new last extent block, then by
6878 * definition, there cannot be any leaves to the right of
6880 last_eb
->h_next_leaf_blk
= 0;
6881 ocfs2_journal_dirty(handle
, last_eb_bh
);
6885 if (rec_flags
& OCFS2_EXT_REFCOUNTED
)
6886 status
= ocfs2_decrease_refcount(inode
, handle
,
6887 ocfs2_blocks_to_clusters(osb
->sb
,
6889 clusters_to_del
, meta_ac
,
6890 &tc
->tc_dealloc
, 1);
6892 status
= ocfs2_truncate_log_append(osb
, handle
,
6907 static int ocfs2_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
6909 set_buffer_uptodate(bh
);
6910 mark_buffer_dirty(bh
);
6914 void ocfs2_map_and_dirty_page(struct inode
*inode
, handle_t
*handle
,
6915 unsigned int from
, unsigned int to
,
6916 struct page
*page
, int zero
, u64
*phys
)
6918 int ret
, partial
= 0;
6920 ret
= ocfs2_map_page_blocks(page
, phys
, inode
, from
, to
, 0);
6925 zero_user_segment(page
, from
, to
);
6928 * Need to set the buffers we zero'd into uptodate
6929 * here if they aren't - ocfs2_map_page_blocks()
6930 * might've skipped some
6932 ret
= walk_page_buffers(handle
, page_buffers(page
),
6937 else if (ocfs2_should_order_data(inode
)) {
6938 ret
= ocfs2_jbd2_file_inode(handle
, inode
);
6944 SetPageUptodate(page
);
6946 flush_dcache_page(page
);
6949 static void ocfs2_zero_cluster_pages(struct inode
*inode
, loff_t start
,
6950 loff_t end
, struct page
**pages
,
6951 int numpages
, u64 phys
, handle_t
*handle
)
6955 unsigned int from
, to
= PAGE_CACHE_SIZE
;
6956 struct super_block
*sb
= inode
->i_sb
;
6958 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb
)));
6963 to
= PAGE_CACHE_SIZE
;
6964 for(i
= 0; i
< numpages
; i
++) {
6967 from
= start
& (PAGE_CACHE_SIZE
- 1);
6968 if ((end
>> PAGE_CACHE_SHIFT
) == page
->index
)
6969 to
= end
& (PAGE_CACHE_SIZE
- 1);
6971 BUG_ON(from
> PAGE_CACHE_SIZE
);
6972 BUG_ON(to
> PAGE_CACHE_SIZE
);
6974 ocfs2_map_and_dirty_page(inode
, handle
, from
, to
, page
, 1,
6977 start
= (page
->index
+ 1) << PAGE_CACHE_SHIFT
;
6981 ocfs2_unlock_and_free_pages(pages
, numpages
);
6984 int ocfs2_grab_pages(struct inode
*inode
, loff_t start
, loff_t end
,
6985 struct page
**pages
, int *num
)
6987 int numpages
, ret
= 0;
6988 struct address_space
*mapping
= inode
->i_mapping
;
6989 unsigned long index
;
6990 loff_t last_page_bytes
;
6992 BUG_ON(start
> end
);
6995 last_page_bytes
= PAGE_ALIGN(end
);
6996 index
= start
>> PAGE_CACHE_SHIFT
;
6998 pages
[numpages
] = grab_cache_page(mapping
, index
);
6999 if (!pages
[numpages
]) {
7007 } while (index
< (last_page_bytes
>> PAGE_CACHE_SHIFT
));
7012 ocfs2_unlock_and_free_pages(pages
, numpages
);
7021 static int ocfs2_grab_eof_pages(struct inode
*inode
, loff_t start
, loff_t end
,
7022 struct page
**pages
, int *num
)
7024 struct super_block
*sb
= inode
->i_sb
;
7026 BUG_ON(start
>> OCFS2_SB(sb
)->s_clustersize_bits
!=
7027 (end
- 1) >> OCFS2_SB(sb
)->s_clustersize_bits
);
7029 return ocfs2_grab_pages(inode
, start
, end
, pages
, num
);
7033 * Zero the area past i_size but still within an allocated
7034 * cluster. This avoids exposing nonzero data on subsequent file
7037 * We need to call this before i_size is updated on the inode because
7038 * otherwise block_write_full_page() will skip writeout of pages past
7039 * i_size. The new_i_size parameter is passed for this reason.
7041 int ocfs2_zero_range_for_truncate(struct inode
*inode
, handle_t
*handle
,
7042 u64 range_start
, u64 range_end
)
7044 int ret
= 0, numpages
;
7045 struct page
**pages
= NULL
;
7047 unsigned int ext_flags
;
7048 struct super_block
*sb
= inode
->i_sb
;
7051 * File systems which don't support sparse files zero on every
7054 if (!ocfs2_sparse_alloc(OCFS2_SB(sb
)))
7057 pages
= kcalloc(ocfs2_pages_per_cluster(sb
),
7058 sizeof(struct page
*), GFP_NOFS
);
7059 if (pages
== NULL
) {
7065 if (range_start
== range_end
)
7068 ret
= ocfs2_extent_map_get_blocks(inode
,
7069 range_start
>> sb
->s_blocksize_bits
,
7070 &phys
, NULL
, &ext_flags
);
7077 * Tail is a hole, or is marked unwritten. In either case, we
7078 * can count on read and write to return/push zero's.
7080 if (phys
== 0 || ext_flags
& OCFS2_EXT_UNWRITTEN
)
7083 ret
= ocfs2_grab_eof_pages(inode
, range_start
, range_end
, pages
,
7090 ocfs2_zero_cluster_pages(inode
, range_start
, range_end
, pages
,
7091 numpages
, phys
, handle
);
7094 * Initiate writeout of the pages we zero'd here. We don't
7095 * wait on them - the truncate_inode_pages() call later will
7098 ret
= filemap_fdatawrite_range(inode
->i_mapping
, range_start
,
7110 static void ocfs2_zero_dinode_id2_with_xattr(struct inode
*inode
,
7111 struct ocfs2_dinode
*di
)
7113 unsigned int blocksize
= 1 << inode
->i_sb
->s_blocksize_bits
;
7114 unsigned int xattrsize
= le16_to_cpu(di
->i_xattr_inline_size
);
7116 if (le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_XATTR_FL
)
7117 memset(&di
->id2
, 0, blocksize
-
7118 offsetof(struct ocfs2_dinode
, id2
) -
7121 memset(&di
->id2
, 0, blocksize
-
7122 offsetof(struct ocfs2_dinode
, id2
));
7125 void ocfs2_dinode_new_extent_list(struct inode
*inode
,
7126 struct ocfs2_dinode
*di
)
7128 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7129 di
->id2
.i_list
.l_tree_depth
= 0;
7130 di
->id2
.i_list
.l_next_free_rec
= 0;
7131 di
->id2
.i_list
.l_count
= cpu_to_le16(
7132 ocfs2_extent_recs_per_inode_with_xattr(inode
->i_sb
, di
));
7135 void ocfs2_set_inode_data_inline(struct inode
*inode
, struct ocfs2_dinode
*di
)
7137 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7138 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7140 spin_lock(&oi
->ip_lock
);
7141 oi
->ip_dyn_features
|= OCFS2_INLINE_DATA_FL
;
7142 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7143 spin_unlock(&oi
->ip_lock
);
7146 * We clear the entire i_data structure here so that all
7147 * fields can be properly initialized.
7149 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7151 idata
->id_count
= cpu_to_le16(
7152 ocfs2_max_inline_data_with_xattr(inode
->i_sb
, di
));
7155 int ocfs2_convert_inline_data_to_extents(struct inode
*inode
,
7156 struct buffer_head
*di_bh
)
7158 int ret
, i
, has_data
, num_pages
= 0;
7160 u64
uninitialized_var(block
);
7161 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7162 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7163 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7164 struct ocfs2_alloc_context
*data_ac
= NULL
;
7165 struct page
**pages
= NULL
;
7166 loff_t end
= osb
->s_clustersize
;
7167 struct ocfs2_extent_tree et
;
7170 has_data
= i_size_read(inode
) ? 1 : 0;
7173 pages
= kcalloc(ocfs2_pages_per_cluster(osb
->sb
),
7174 sizeof(struct page
*), GFP_NOFS
);
7175 if (pages
== NULL
) {
7181 ret
= ocfs2_reserve_clusters(osb
, 1, &data_ac
);
7188 handle
= ocfs2_start_trans(osb
,
7189 ocfs2_inline_to_extents_credits(osb
->sb
));
7190 if (IS_ERR(handle
)) {
7191 ret
= PTR_ERR(handle
);
7196 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7197 OCFS2_JOURNAL_ACCESS_WRITE
);
7205 unsigned int page_end
;
7208 ret
= dquot_alloc_space_nodirty(inode
,
7209 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7214 data_ac
->ac_resv
= &OCFS2_I(inode
)->ip_la_data_resv
;
7216 ret
= ocfs2_claim_clusters(osb
, handle
, data_ac
, 1, &bit_off
,
7224 * Save two copies, one for insert, and one that can
7225 * be changed by ocfs2_map_and_dirty_page() below.
7227 block
= phys
= ocfs2_clusters_to_blocks(inode
->i_sb
, bit_off
);
7230 * Non sparse file systems zero on extend, so no need
7233 if (!ocfs2_sparse_alloc(osb
) &&
7234 PAGE_CACHE_SIZE
< osb
->s_clustersize
)
7235 end
= PAGE_CACHE_SIZE
;
7237 ret
= ocfs2_grab_eof_pages(inode
, 0, end
, pages
, &num_pages
);
7244 * This should populate the 1st page for us and mark
7247 ret
= ocfs2_read_inline_data(inode
, pages
[0], di_bh
);
7253 page_end
= PAGE_CACHE_SIZE
;
7254 if (PAGE_CACHE_SIZE
> osb
->s_clustersize
)
7255 page_end
= osb
->s_clustersize
;
7257 for (i
= 0; i
< num_pages
; i
++)
7258 ocfs2_map_and_dirty_page(inode
, handle
, 0, page_end
,
7259 pages
[i
], i
> 0, &phys
);
7262 spin_lock(&oi
->ip_lock
);
7263 oi
->ip_dyn_features
&= ~OCFS2_INLINE_DATA_FL
;
7264 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7265 spin_unlock(&oi
->ip_lock
);
7267 ocfs2_dinode_new_extent_list(inode
, di
);
7269 ocfs2_journal_dirty(handle
, di_bh
);
7273 * An error at this point should be extremely rare. If
7274 * this proves to be false, we could always re-build
7275 * the in-inode data from our pages.
7277 ocfs2_init_dinode_extent_tree(&et
, INODE_CACHE(inode
), di_bh
);
7278 ret
= ocfs2_insert_extent(handle
, &et
, 0, block
, 1, 0, NULL
);
7284 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7288 if (ret
< 0 && did_quota
)
7289 dquot_free_space_nodirty(inode
,
7290 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7292 ocfs2_commit_trans(osb
, handle
);
7296 ocfs2_free_alloc_context(data_ac
);
7300 ocfs2_unlock_and_free_pages(pages
, num_pages
);
7308 * It is expected, that by the time you call this function,
7309 * inode->i_size and fe->i_size have been adjusted.
7311 * WARNING: This will kfree the truncate context
7313 int ocfs2_commit_truncate(struct ocfs2_super
*osb
,
7314 struct inode
*inode
,
7315 struct buffer_head
*fe_bh
,
7316 struct ocfs2_truncate_context
*tc
)
7318 int status
, i
, credits
, tl_sem
= 0;
7319 u32 clusters_to_del
, new_highest_cpos
, range
;
7321 struct ocfs2_extent_list
*el
;
7322 handle_t
*handle
= NULL
;
7323 struct inode
*tl_inode
= osb
->osb_tl_inode
;
7324 struct ocfs2_path
*path
= NULL
;
7325 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)fe_bh
->b_data
;
7326 struct ocfs2_alloc_context
*meta_ac
= NULL
;
7327 struct ocfs2_refcount_tree
*ref_tree
= NULL
;
7331 new_highest_cpos
= ocfs2_clusters_for_bytes(osb
->sb
,
7332 i_size_read(inode
));
7334 path
= ocfs2_new_path(fe_bh
, &di
->id2
.i_list
,
7335 ocfs2_journal_access_di
);
7342 ocfs2_extent_map_trunc(inode
, new_highest_cpos
);
7346 * Check that we still have allocation to delete.
7348 if (OCFS2_I(inode
)->ip_clusters
== 0) {
7356 * Truncate always works against the rightmost tree branch.
7358 status
= ocfs2_find_path(INODE_CACHE(inode
), path
, UINT_MAX
);
7364 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
7365 OCFS2_I(inode
)->ip_clusters
, path
->p_tree_depth
);
7368 * By now, el will point to the extent list on the bottom most
7369 * portion of this tree. Only the tail record is considered in
7372 * We handle the following cases, in order:
7373 * - empty extent: delete the remaining branch
7374 * - remove the entire record
7375 * - remove a partial record
7376 * - no record needs to be removed (truncate has completed)
7378 el
= path_leaf_el(path
);
7379 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
7380 ocfs2_error(inode
->i_sb
,
7381 "Inode %llu has empty extent block at %llu\n",
7382 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7383 (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
7388 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
7389 range
= le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
7390 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
7391 if (i
== 0 && ocfs2_is_empty_extent(&el
->l_recs
[i
])) {
7392 clusters_to_del
= 0;
7393 } else if (le32_to_cpu(el
->l_recs
[i
].e_cpos
) >= new_highest_cpos
) {
7394 clusters_to_del
= ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
7395 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
7396 } else if (range
> new_highest_cpos
) {
7397 clusters_to_del
= (ocfs2_rec_clusters(el
, &el
->l_recs
[i
]) +
7398 le32_to_cpu(el
->l_recs
[i
].e_cpos
)) -
7400 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
) +
7401 ocfs2_clusters_to_blocks(inode
->i_sb
,
7402 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]) -
7409 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
7410 clusters_to_del
, (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
7412 if (el
->l_recs
[i
].e_flags
& OCFS2_EXT_REFCOUNTED
&& clusters_to_del
) {
7413 BUG_ON(!(OCFS2_I(inode
)->ip_dyn_features
&
7414 OCFS2_HAS_REFCOUNT_FL
));
7416 status
= ocfs2_lock_refcount_tree(osb
,
7417 le64_to_cpu(di
->i_refcount_loc
),
7418 1, &ref_tree
, NULL
);
7424 status
= ocfs2_prepare_refcount_change_for_del(inode
, fe_bh
,
7435 mutex_lock(&tl_inode
->i_mutex
);
7437 /* ocfs2_truncate_log_needs_flush guarantees us at least one
7438 * record is free for use. If there isn't any, we flush to get
7439 * an empty truncate log. */
7440 if (ocfs2_truncate_log_needs_flush(osb
)) {
7441 status
= __ocfs2_flush_truncate_log(osb
);
7448 credits
+= ocfs2_calc_tree_trunc_credits(osb
->sb
, clusters_to_del
,
7449 (struct ocfs2_dinode
*)fe_bh
->b_data
,
7451 handle
= ocfs2_start_trans(osb
, credits
);
7452 if (IS_ERR(handle
)) {
7453 status
= PTR_ERR(handle
);
7459 status
= ocfs2_do_truncate(osb
, clusters_to_del
, inode
, fe_bh
, handle
,
7466 mutex_unlock(&tl_inode
->i_mutex
);
7469 ocfs2_commit_trans(osb
, handle
);
7472 ocfs2_reinit_path(path
, 1);
7475 ocfs2_free_alloc_context(meta_ac
);
7480 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
7485 * The check above will catch the case where we've truncated
7486 * away all allocation.
7492 ocfs2_schedule_truncate_log_flush(osb
, 1);
7495 mutex_unlock(&tl_inode
->i_mutex
);
7498 ocfs2_commit_trans(osb
, handle
);
7501 ocfs2_free_alloc_context(meta_ac
);
7504 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
7506 ocfs2_run_deallocs(osb
, &tc
->tc_dealloc
);
7508 ocfs2_free_path(path
);
7510 /* This will drop the ext_alloc cluster lock for us */
7511 ocfs2_free_truncate_context(tc
);
7518 * Expects the inode to already be locked.
7520 int ocfs2_prepare_truncate(struct ocfs2_super
*osb
,
7521 struct inode
*inode
,
7522 struct buffer_head
*fe_bh
,
7523 struct ocfs2_truncate_context
**tc
)
7526 unsigned int new_i_clusters
;
7527 struct ocfs2_dinode
*fe
;
7528 struct ocfs2_extent_block
*eb
;
7529 struct buffer_head
*last_eb_bh
= NULL
;
7535 new_i_clusters
= ocfs2_clusters_for_bytes(osb
->sb
,
7536 i_size_read(inode
));
7537 fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
7539 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
7540 "%llu\n", le32_to_cpu(fe
->i_clusters
), new_i_clusters
,
7541 (unsigned long long)le64_to_cpu(fe
->i_size
));
7543 *tc
= kzalloc(sizeof(struct ocfs2_truncate_context
), GFP_KERNEL
);
7549 ocfs2_init_dealloc_ctxt(&(*tc
)->tc_dealloc
);
7551 if (fe
->id2
.i_list
.l_tree_depth
) {
7552 status
= ocfs2_read_extent_block(INODE_CACHE(inode
),
7553 le64_to_cpu(fe
->i_last_eb_blk
),
7559 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
7562 (*tc
)->tc_last_eb_bh
= last_eb_bh
;
7568 ocfs2_free_truncate_context(*tc
);
7576 * 'start' is inclusive, 'end' is not.
7578 int ocfs2_truncate_inline(struct inode
*inode
, struct buffer_head
*di_bh
,
7579 unsigned int start
, unsigned int end
, int trunc
)
7582 unsigned int numbytes
;
7584 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7585 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7586 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7588 if (end
> i_size_read(inode
))
7589 end
= i_size_read(inode
);
7591 BUG_ON(start
>= end
);
7593 if (!(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
) ||
7594 !(le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_DATA_FL
) ||
7595 !ocfs2_supports_inline_data(osb
)) {
7596 ocfs2_error(inode
->i_sb
,
7597 "Inline data flags for inode %llu don't agree! "
7598 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7599 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7600 le16_to_cpu(di
->i_dyn_features
),
7601 OCFS2_I(inode
)->ip_dyn_features
,
7602 osb
->s_feature_incompat
);
7607 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
7608 if (IS_ERR(handle
)) {
7609 ret
= PTR_ERR(handle
);
7614 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7615 OCFS2_JOURNAL_ACCESS_WRITE
);
7621 numbytes
= end
- start
;
7622 memset(idata
->id_data
+ start
, 0, numbytes
);
7625 * No need to worry about the data page here - it's been
7626 * truncated already and inline data doesn't need it for
7627 * pushing zero's to disk, so we'll let readpage pick it up
7631 i_size_write(inode
, start
);
7632 di
->i_size
= cpu_to_le64(start
);
7635 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7636 inode
->i_ctime
= inode
->i_mtime
= CURRENT_TIME
;
7638 di
->i_ctime
= di
->i_mtime
= cpu_to_le64(inode
->i_ctime
.tv_sec
);
7639 di
->i_ctime_nsec
= di
->i_mtime_nsec
= cpu_to_le32(inode
->i_ctime
.tv_nsec
);
7641 ocfs2_journal_dirty(handle
, di_bh
);
7644 ocfs2_commit_trans(osb
, handle
);
7650 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context
*tc
)
7653 * The caller is responsible for completing deallocation
7654 * before freeing the context.
7656 if (tc
->tc_dealloc
.c_first_suballocator
!= NULL
)
7658 "Truncate completion has non-empty dealloc context\n");
7660 brelse(tc
->tc_last_eb_bh
);