1 // SPDX-License-Identifier: GPL-2.0-or-later
5 * Extent allocs and frees
7 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
11 #include <linux/types.h>
12 #include <linux/slab.h>
13 #include <linux/highmem.h>
14 #include <linux/swap.h>
15 #include <linux/quotaops.h>
16 #include <linux/blkdev.h>
17 #include <linux/sched/signal.h>
19 #include <cluster/masklog.h>
25 #include "blockcheck.h"
27 #include "extent_map.h"
30 #include "localalloc.h"
37 #include "refcounttree.h"
38 #include "ocfs2_trace.h"
40 #include "buffer_head_io.h"
42 enum ocfs2_contig_type
{
49 static enum ocfs2_contig_type
50 ocfs2_extent_rec_contig(struct super_block
*sb
,
51 struct ocfs2_extent_rec
*ext
,
52 struct ocfs2_extent_rec
*insert_rec
);
54 * Operations for a specific extent tree type.
56 * To implement an on-disk btree (extent tree) type in ocfs2, add
57 * an ocfs2_extent_tree_operations structure and the matching
58 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
59 * for the allocation portion of the extent tree.
61 struct ocfs2_extent_tree_operations
{
63 * last_eb_blk is the block number of the right most leaf extent
64 * block. Most on-disk structures containing an extent tree store
65 * this value for fast access. The ->eo_set_last_eb_blk() and
66 * ->eo_get_last_eb_blk() operations access this value. They are
69 void (*eo_set_last_eb_blk
)(struct ocfs2_extent_tree
*et
,
71 u64 (*eo_get_last_eb_blk
)(struct ocfs2_extent_tree
*et
);
74 * The on-disk structure usually keeps track of how many total
75 * clusters are stored in this extent tree. This function updates
76 * that value. new_clusters is the delta, and must be
77 * added to the total. Required.
79 void (*eo_update_clusters
)(struct ocfs2_extent_tree
*et
,
83 * If this extent tree is supported by an extent map, insert
84 * a record into the map.
86 void (*eo_extent_map_insert
)(struct ocfs2_extent_tree
*et
,
87 struct ocfs2_extent_rec
*rec
);
90 * If this extent tree is supported by an extent map, truncate the
93 void (*eo_extent_map_truncate
)(struct ocfs2_extent_tree
*et
,
97 * If ->eo_insert_check() exists, it is called before rec is
98 * inserted into the extent tree. It is optional.
100 int (*eo_insert_check
)(struct ocfs2_extent_tree
*et
,
101 struct ocfs2_extent_rec
*rec
);
102 int (*eo_sanity_check
)(struct ocfs2_extent_tree
*et
);
105 * --------------------------------------------------------------
106 * The remaining are internal to ocfs2_extent_tree and don't have
111 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
114 void (*eo_fill_root_el
)(struct ocfs2_extent_tree
*et
);
117 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
118 * it exists. If it does not, et->et_max_leaf_clusters is set
119 * to 0 (unlimited). Optional.
121 void (*eo_fill_max_leaf_clusters
)(struct ocfs2_extent_tree
*et
);
124 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
125 * are contiguous or not. Optional. Don't need to set it if use
126 * ocfs2_extent_rec as the tree leaf.
128 enum ocfs2_contig_type
129 (*eo_extent_contig
)(struct ocfs2_extent_tree
*et
,
130 struct ocfs2_extent_rec
*ext
,
131 struct ocfs2_extent_rec
*insert_rec
);
136 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
139 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
);
140 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
142 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree
*et
,
144 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree
*et
,
145 struct ocfs2_extent_rec
*rec
);
146 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree
*et
,
148 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree
*et
,
149 struct ocfs2_extent_rec
*rec
);
150 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree
*et
);
151 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
);
153 static int ocfs2_reuse_blk_from_dealloc(handle_t
*handle
,
154 struct ocfs2_extent_tree
*et
,
155 struct buffer_head
**new_eb_bh
,
156 int blk_wanted
, int *blk_given
);
157 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree
*et
);
159 static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops
= {
160 .eo_set_last_eb_blk
= ocfs2_dinode_set_last_eb_blk
,
161 .eo_get_last_eb_blk
= ocfs2_dinode_get_last_eb_blk
,
162 .eo_update_clusters
= ocfs2_dinode_update_clusters
,
163 .eo_extent_map_insert
= ocfs2_dinode_extent_map_insert
,
164 .eo_extent_map_truncate
= ocfs2_dinode_extent_map_truncate
,
165 .eo_insert_check
= ocfs2_dinode_insert_check
,
166 .eo_sanity_check
= ocfs2_dinode_sanity_check
,
167 .eo_fill_root_el
= ocfs2_dinode_fill_root_el
,
170 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
173 struct ocfs2_dinode
*di
= et
->et_object
;
175 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
176 di
->i_last_eb_blk
= cpu_to_le64(blkno
);
179 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
181 struct ocfs2_dinode
*di
= et
->et_object
;
183 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
184 return le64_to_cpu(di
->i_last_eb_blk
);
187 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree
*et
,
190 struct ocfs2_inode_info
*oi
= cache_info_to_inode(et
->et_ci
);
191 struct ocfs2_dinode
*di
= et
->et_object
;
193 le32_add_cpu(&di
->i_clusters
, clusters
);
194 spin_lock(&oi
->ip_lock
);
195 oi
->ip_clusters
= le32_to_cpu(di
->i_clusters
);
196 spin_unlock(&oi
->ip_lock
);
199 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree
*et
,
200 struct ocfs2_extent_rec
*rec
)
202 struct inode
*inode
= &cache_info_to_inode(et
->et_ci
)->vfs_inode
;
204 ocfs2_extent_map_insert_rec(inode
, rec
);
207 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree
*et
,
210 struct inode
*inode
= &cache_info_to_inode(et
->et_ci
)->vfs_inode
;
212 ocfs2_extent_map_trunc(inode
, clusters
);
215 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree
*et
,
216 struct ocfs2_extent_rec
*rec
)
218 struct ocfs2_inode_info
*oi
= cache_info_to_inode(et
->et_ci
);
219 struct ocfs2_super
*osb
= OCFS2_SB(oi
->vfs_inode
.i_sb
);
221 BUG_ON(oi
->ip_dyn_features
& OCFS2_INLINE_DATA_FL
);
222 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb
) &&
223 (oi
->ip_clusters
!= le32_to_cpu(rec
->e_cpos
)),
224 "Device %s, asking for sparse allocation: inode %llu, "
225 "cpos %u, clusters %u\n",
227 (unsigned long long)oi
->ip_blkno
,
228 rec
->e_cpos
, oi
->ip_clusters
);
233 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree
*et
)
235 struct ocfs2_dinode
*di
= et
->et_object
;
237 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
238 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
243 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
)
245 struct ocfs2_dinode
*di
= et
->et_object
;
247 et
->et_root_el
= &di
->id2
.i_list
;
251 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree
*et
)
253 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
255 et
->et_root_el
= &vb
->vb_xv
->xr_list
;
258 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
261 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
263 vb
->vb_xv
->xr_last_eb_blk
= cpu_to_le64(blkno
);
266 static u64
ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
268 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
270 return le64_to_cpu(vb
->vb_xv
->xr_last_eb_blk
);
273 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree
*et
,
276 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
278 le32_add_cpu(&vb
->vb_xv
->xr_clusters
, clusters
);
281 static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops
= {
282 .eo_set_last_eb_blk
= ocfs2_xattr_value_set_last_eb_blk
,
283 .eo_get_last_eb_blk
= ocfs2_xattr_value_get_last_eb_blk
,
284 .eo_update_clusters
= ocfs2_xattr_value_update_clusters
,
285 .eo_fill_root_el
= ocfs2_xattr_value_fill_root_el
,
288 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree
*et
)
290 struct ocfs2_xattr_block
*xb
= et
->et_object
;
292 et
->et_root_el
= &xb
->xb_attrs
.xb_root
.xt_list
;
295 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree
*et
)
297 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
298 et
->et_max_leaf_clusters
=
299 ocfs2_clusters_for_bytes(sb
, OCFS2_MAX_XATTR_TREE_LEAF_SIZE
);
302 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
305 struct ocfs2_xattr_block
*xb
= et
->et_object
;
306 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
308 xt
->xt_last_eb_blk
= cpu_to_le64(blkno
);
311 static u64
ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
313 struct ocfs2_xattr_block
*xb
= et
->et_object
;
314 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
316 return le64_to_cpu(xt
->xt_last_eb_blk
);
319 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree
*et
,
322 struct ocfs2_xattr_block
*xb
= et
->et_object
;
324 le32_add_cpu(&xb
->xb_attrs
.xb_root
.xt_clusters
, clusters
);
327 static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops
= {
328 .eo_set_last_eb_blk
= ocfs2_xattr_tree_set_last_eb_blk
,
329 .eo_get_last_eb_blk
= ocfs2_xattr_tree_get_last_eb_blk
,
330 .eo_update_clusters
= ocfs2_xattr_tree_update_clusters
,
331 .eo_fill_root_el
= ocfs2_xattr_tree_fill_root_el
,
332 .eo_fill_max_leaf_clusters
= ocfs2_xattr_tree_fill_max_leaf_clusters
,
335 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
338 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
340 dx_root
->dr_last_eb_blk
= cpu_to_le64(blkno
);
343 static u64
ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
345 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
347 return le64_to_cpu(dx_root
->dr_last_eb_blk
);
350 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree
*et
,
353 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
355 le32_add_cpu(&dx_root
->dr_clusters
, clusters
);
358 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree
*et
)
360 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
362 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root
));
367 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree
*et
)
369 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
371 et
->et_root_el
= &dx_root
->dr_list
;
374 static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops
= {
375 .eo_set_last_eb_blk
= ocfs2_dx_root_set_last_eb_blk
,
376 .eo_get_last_eb_blk
= ocfs2_dx_root_get_last_eb_blk
,
377 .eo_update_clusters
= ocfs2_dx_root_update_clusters
,
378 .eo_sanity_check
= ocfs2_dx_root_sanity_check
,
379 .eo_fill_root_el
= ocfs2_dx_root_fill_root_el
,
382 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree
*et
)
384 struct ocfs2_refcount_block
*rb
= et
->et_object
;
386 et
->et_root_el
= &rb
->rf_list
;
389 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
392 struct ocfs2_refcount_block
*rb
= et
->et_object
;
394 rb
->rf_last_eb_blk
= cpu_to_le64(blkno
);
397 static u64
ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
399 struct ocfs2_refcount_block
*rb
= et
->et_object
;
401 return le64_to_cpu(rb
->rf_last_eb_blk
);
404 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree
*et
,
407 struct ocfs2_refcount_block
*rb
= et
->et_object
;
409 le32_add_cpu(&rb
->rf_clusters
, clusters
);
412 static enum ocfs2_contig_type
413 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree
*et
,
414 struct ocfs2_extent_rec
*ext
,
415 struct ocfs2_extent_rec
*insert_rec
)
420 static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops
= {
421 .eo_set_last_eb_blk
= ocfs2_refcount_tree_set_last_eb_blk
,
422 .eo_get_last_eb_blk
= ocfs2_refcount_tree_get_last_eb_blk
,
423 .eo_update_clusters
= ocfs2_refcount_tree_update_clusters
,
424 .eo_fill_root_el
= ocfs2_refcount_tree_fill_root_el
,
425 .eo_extent_contig
= ocfs2_refcount_tree_extent_contig
,
428 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree
*et
,
429 struct ocfs2_caching_info
*ci
,
430 struct buffer_head
*bh
,
431 ocfs2_journal_access_func access
,
433 const struct ocfs2_extent_tree_operations
*ops
)
438 et
->et_root_journal_access
= access
;
440 obj
= (void *)bh
->b_data
;
442 et
->et_dealloc
= NULL
;
444 et
->et_ops
->eo_fill_root_el(et
);
445 if (!et
->et_ops
->eo_fill_max_leaf_clusters
)
446 et
->et_max_leaf_clusters
= 0;
448 et
->et_ops
->eo_fill_max_leaf_clusters(et
);
451 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree
*et
,
452 struct ocfs2_caching_info
*ci
,
453 struct buffer_head
*bh
)
455 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_di
,
456 NULL
, &ocfs2_dinode_et_ops
);
459 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree
*et
,
460 struct ocfs2_caching_info
*ci
,
461 struct buffer_head
*bh
)
463 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_xb
,
464 NULL
, &ocfs2_xattr_tree_et_ops
);
467 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree
*et
,
468 struct ocfs2_caching_info
*ci
,
469 struct ocfs2_xattr_value_buf
*vb
)
471 __ocfs2_init_extent_tree(et
, ci
, vb
->vb_bh
, vb
->vb_access
, vb
,
472 &ocfs2_xattr_value_et_ops
);
475 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree
*et
,
476 struct ocfs2_caching_info
*ci
,
477 struct buffer_head
*bh
)
479 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_dr
,
480 NULL
, &ocfs2_dx_root_et_ops
);
483 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree
*et
,
484 struct ocfs2_caching_info
*ci
,
485 struct buffer_head
*bh
)
487 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_rb
,
488 NULL
, &ocfs2_refcount_tree_et_ops
);
491 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
494 et
->et_ops
->eo_set_last_eb_blk(et
, new_last_eb_blk
);
497 static inline u64
ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
499 return et
->et_ops
->eo_get_last_eb_blk(et
);
502 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree
*et
,
505 et
->et_ops
->eo_update_clusters(et
, clusters
);
508 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree
*et
,
509 struct ocfs2_extent_rec
*rec
)
511 if (et
->et_ops
->eo_extent_map_insert
)
512 et
->et_ops
->eo_extent_map_insert(et
, rec
);
515 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree
*et
,
518 if (et
->et_ops
->eo_extent_map_truncate
)
519 et
->et_ops
->eo_extent_map_truncate(et
, clusters
);
522 static inline int ocfs2_et_root_journal_access(handle_t
*handle
,
523 struct ocfs2_extent_tree
*et
,
526 return et
->et_root_journal_access(handle
, et
->et_ci
, et
->et_root_bh
,
530 static inline enum ocfs2_contig_type
531 ocfs2_et_extent_contig(struct ocfs2_extent_tree
*et
,
532 struct ocfs2_extent_rec
*rec
,
533 struct ocfs2_extent_rec
*insert_rec
)
535 if (et
->et_ops
->eo_extent_contig
)
536 return et
->et_ops
->eo_extent_contig(et
, rec
, insert_rec
);
538 return ocfs2_extent_rec_contig(
539 ocfs2_metadata_cache_get_super(et
->et_ci
),
543 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree
*et
,
544 struct ocfs2_extent_rec
*rec
)
548 if (et
->et_ops
->eo_insert_check
)
549 ret
= et
->et_ops
->eo_insert_check(et
, rec
);
553 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree
*et
)
557 if (et
->et_ops
->eo_sanity_check
)
558 ret
= et
->et_ops
->eo_sanity_check(et
);
562 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
563 struct ocfs2_extent_block
*eb
);
564 static void ocfs2_adjust_rightmost_records(handle_t
*handle
,
565 struct ocfs2_extent_tree
*et
,
566 struct ocfs2_path
*path
,
567 struct ocfs2_extent_rec
*insert_rec
);
569 * Reset the actual path elements so that we can re-use the structure
570 * to build another path. Generally, this involves freeing the buffer
573 void ocfs2_reinit_path(struct ocfs2_path
*path
, int keep_root
)
575 int i
, start
= 0, depth
= 0;
576 struct ocfs2_path_item
*node
;
581 for(i
= start
; i
< path_num_items(path
); i
++) {
582 node
= &path
->p_node
[i
];
590 * Tree depth may change during truncate, or insert. If we're
591 * keeping the root extent list, then make sure that our path
592 * structure reflects the proper depth.
595 depth
= le16_to_cpu(path_root_el(path
)->l_tree_depth
);
597 path_root_access(path
) = NULL
;
599 path
->p_tree_depth
= depth
;
602 void ocfs2_free_path(struct ocfs2_path
*path
)
605 ocfs2_reinit_path(path
, 0);
611 * All the elements of src into dest. After this call, src could be freed
612 * without affecting dest.
614 * Both paths should have the same root. Any non-root elements of dest
617 static void ocfs2_cp_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
621 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
622 BUG_ON(path_root_el(dest
) != path_root_el(src
));
623 BUG_ON(path_root_access(dest
) != path_root_access(src
));
625 ocfs2_reinit_path(dest
, 1);
627 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
628 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
629 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
631 if (dest
->p_node
[i
].bh
)
632 get_bh(dest
->p_node
[i
].bh
);
637 * Make the *dest path the same as src and re-initialize src path to
640 static void ocfs2_mv_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
644 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
645 BUG_ON(path_root_access(dest
) != path_root_access(src
));
647 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
648 brelse(dest
->p_node
[i
].bh
);
650 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
651 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
653 src
->p_node
[i
].bh
= NULL
;
654 src
->p_node
[i
].el
= NULL
;
659 * Insert an extent block at given index.
661 * This will not take an additional reference on eb_bh.
663 static inline void ocfs2_path_insert_eb(struct ocfs2_path
*path
, int index
,
664 struct buffer_head
*eb_bh
)
666 struct ocfs2_extent_block
*eb
= (struct ocfs2_extent_block
*)eb_bh
->b_data
;
669 * Right now, no root bh is an extent block, so this helps
670 * catch code errors with dinode trees. The assertion can be
671 * safely removed if we ever need to insert extent block
672 * structures at the root.
676 path
->p_node
[index
].bh
= eb_bh
;
677 path
->p_node
[index
].el
= &eb
->h_list
;
680 static struct ocfs2_path
*ocfs2_new_path(struct buffer_head
*root_bh
,
681 struct ocfs2_extent_list
*root_el
,
682 ocfs2_journal_access_func access
)
684 struct ocfs2_path
*path
;
686 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) >= OCFS2_MAX_PATH_DEPTH
);
688 path
= kzalloc(sizeof(*path
), GFP_NOFS
);
690 path
->p_tree_depth
= le16_to_cpu(root_el
->l_tree_depth
);
692 path_root_bh(path
) = root_bh
;
693 path_root_el(path
) = root_el
;
694 path_root_access(path
) = access
;
700 struct ocfs2_path
*ocfs2_new_path_from_path(struct ocfs2_path
*path
)
702 return ocfs2_new_path(path_root_bh(path
), path_root_el(path
),
703 path_root_access(path
));
706 struct ocfs2_path
*ocfs2_new_path_from_et(struct ocfs2_extent_tree
*et
)
708 return ocfs2_new_path(et
->et_root_bh
, et
->et_root_el
,
709 et
->et_root_journal_access
);
713 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
714 * otherwise it's the root_access function.
716 * I don't like the way this function's name looks next to
717 * ocfs2_journal_access_path(), but I don't have a better one.
719 int ocfs2_path_bh_journal_access(handle_t
*handle
,
720 struct ocfs2_caching_info
*ci
,
721 struct ocfs2_path
*path
,
724 ocfs2_journal_access_func access
= path_root_access(path
);
727 access
= ocfs2_journal_access
;
730 access
= ocfs2_journal_access_eb
;
732 return access(handle
, ci
, path
->p_node
[idx
].bh
,
733 OCFS2_JOURNAL_ACCESS_WRITE
);
737 * Convenience function to journal all components in a path.
739 int ocfs2_journal_access_path(struct ocfs2_caching_info
*ci
,
741 struct ocfs2_path
*path
)
748 for(i
= 0; i
< path_num_items(path
); i
++) {
749 ret
= ocfs2_path_bh_journal_access(handle
, ci
, path
, i
);
761 * Return the index of the extent record which contains cluster #v_cluster.
762 * -1 is returned if it was not found.
764 * Should work fine on interior and exterior nodes.
766 int ocfs2_search_extent_list(struct ocfs2_extent_list
*el
, u32 v_cluster
)
770 struct ocfs2_extent_rec
*rec
;
771 u32 rec_end
, rec_start
, clusters
;
773 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
774 rec
= &el
->l_recs
[i
];
776 rec_start
= le32_to_cpu(rec
->e_cpos
);
777 clusters
= ocfs2_rec_clusters(el
, rec
);
779 rec_end
= rec_start
+ clusters
;
781 if (v_cluster
>= rec_start
&& v_cluster
< rec_end
) {
791 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
792 * ocfs2_extent_rec_contig only work properly against leaf nodes!
794 static int ocfs2_block_extent_contig(struct super_block
*sb
,
795 struct ocfs2_extent_rec
*ext
,
798 u64 blk_end
= le64_to_cpu(ext
->e_blkno
);
800 blk_end
+= ocfs2_clusters_to_blocks(sb
,
801 le16_to_cpu(ext
->e_leaf_clusters
));
803 return blkno
== blk_end
;
806 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec
*left
,
807 struct ocfs2_extent_rec
*right
)
811 left_range
= le32_to_cpu(left
->e_cpos
) +
812 le16_to_cpu(left
->e_leaf_clusters
);
814 return (left_range
== le32_to_cpu(right
->e_cpos
));
817 static enum ocfs2_contig_type
818 ocfs2_extent_rec_contig(struct super_block
*sb
,
819 struct ocfs2_extent_rec
*ext
,
820 struct ocfs2_extent_rec
*insert_rec
)
822 u64 blkno
= le64_to_cpu(insert_rec
->e_blkno
);
825 * Refuse to coalesce extent records with different flag
826 * fields - we don't want to mix unwritten extents with user
829 if (ext
->e_flags
!= insert_rec
->e_flags
)
832 if (ocfs2_extents_adjacent(ext
, insert_rec
) &&
833 ocfs2_block_extent_contig(sb
, ext
, blkno
))
836 blkno
= le64_to_cpu(ext
->e_blkno
);
837 if (ocfs2_extents_adjacent(insert_rec
, ext
) &&
838 ocfs2_block_extent_contig(sb
, insert_rec
, blkno
))
845 * NOTE: We can have pretty much any combination of contiguousness and
848 * The usefulness of APPEND_TAIL is more in that it lets us know that
849 * we'll have to update the path to that leaf.
851 enum ocfs2_append_type
{
856 enum ocfs2_split_type
{
862 struct ocfs2_insert_type
{
863 enum ocfs2_split_type ins_split
;
864 enum ocfs2_append_type ins_appending
;
865 enum ocfs2_contig_type ins_contig
;
866 int ins_contig_index
;
870 struct ocfs2_merge_ctxt
{
871 enum ocfs2_contig_type c_contig_type
;
872 int c_has_empty_extent
;
873 int c_split_covers_rec
;
876 static int ocfs2_validate_extent_block(struct super_block
*sb
,
877 struct buffer_head
*bh
)
880 struct ocfs2_extent_block
*eb
=
881 (struct ocfs2_extent_block
*)bh
->b_data
;
883 trace_ocfs2_validate_extent_block((unsigned long long)bh
->b_blocknr
);
885 BUG_ON(!buffer_uptodate(bh
));
888 * If the ecc fails, we return the error but otherwise
889 * leave the filesystem running. We know any error is
890 * local to this block.
892 rc
= ocfs2_validate_meta_ecc(sb
, bh
->b_data
, &eb
->h_check
);
894 mlog(ML_ERROR
, "Checksum failed for extent block %llu\n",
895 (unsigned long long)bh
->b_blocknr
);
900 * Errors after here are fatal.
903 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
905 "Extent block #%llu has bad signature %.*s\n",
906 (unsigned long long)bh
->b_blocknr
, 7,
911 if (le64_to_cpu(eb
->h_blkno
) != bh
->b_blocknr
) {
913 "Extent block #%llu has an invalid h_blkno of %llu\n",
914 (unsigned long long)bh
->b_blocknr
,
915 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
919 if (le32_to_cpu(eb
->h_fs_generation
) != OCFS2_SB(sb
)->fs_generation
)
921 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
922 (unsigned long long)bh
->b_blocknr
,
923 le32_to_cpu(eb
->h_fs_generation
));
928 int ocfs2_read_extent_block(struct ocfs2_caching_info
*ci
, u64 eb_blkno
,
929 struct buffer_head
**bh
)
932 struct buffer_head
*tmp
= *bh
;
934 rc
= ocfs2_read_block(ci
, eb_blkno
, &tmp
,
935 ocfs2_validate_extent_block
);
937 /* If ocfs2_read_block() got us a new bh, pass it up. */
946 * How many free extents have we got before we need more meta data?
948 int ocfs2_num_free_extents(struct ocfs2_extent_tree
*et
)
951 struct ocfs2_extent_list
*el
= NULL
;
952 struct ocfs2_extent_block
*eb
;
953 struct buffer_head
*eb_bh
= NULL
;
957 last_eb_blk
= ocfs2_et_get_last_eb_blk(et
);
960 retval
= ocfs2_read_extent_block(et
->et_ci
, last_eb_blk
,
966 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
970 BUG_ON(el
->l_tree_depth
!= 0);
972 retval
= le16_to_cpu(el
->l_count
) - le16_to_cpu(el
->l_next_free_rec
);
976 trace_ocfs2_num_free_extents(retval
);
980 /* expects array to already be allocated
982 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
985 static int ocfs2_create_new_meta_bhs(handle_t
*handle
,
986 struct ocfs2_extent_tree
*et
,
988 struct ocfs2_alloc_context
*meta_ac
,
989 struct buffer_head
*bhs
[])
991 int count
, status
, i
;
992 u16 suballoc_bit_start
;
994 u64 suballoc_loc
, first_blkno
;
995 struct ocfs2_super
*osb
=
996 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
997 struct ocfs2_extent_block
*eb
;
1000 while (count
< wanted
) {
1001 status
= ocfs2_claim_metadata(handle
,
1005 &suballoc_bit_start
,
1013 for(i
= count
; i
< (num_got
+ count
); i
++) {
1014 bhs
[i
] = sb_getblk(osb
->sb
, first_blkno
);
1015 if (bhs
[i
] == NULL
) {
1020 ocfs2_set_new_buffer_uptodate(et
->et_ci
, bhs
[i
]);
1022 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
1024 OCFS2_JOURNAL_ACCESS_CREATE
);
1030 memset(bhs
[i
]->b_data
, 0, osb
->sb
->s_blocksize
);
1031 eb
= (struct ocfs2_extent_block
*) bhs
[i
]->b_data
;
1032 /* Ok, setup the minimal stuff here. */
1033 strcpy(eb
->h_signature
, OCFS2_EXTENT_BLOCK_SIGNATURE
);
1034 eb
->h_blkno
= cpu_to_le64(first_blkno
);
1035 eb
->h_fs_generation
= cpu_to_le32(osb
->fs_generation
);
1036 eb
->h_suballoc_slot
=
1037 cpu_to_le16(meta_ac
->ac_alloc_slot
);
1038 eb
->h_suballoc_loc
= cpu_to_le64(suballoc_loc
);
1039 eb
->h_suballoc_bit
= cpu_to_le16(suballoc_bit_start
);
1040 eb
->h_list
.l_count
=
1041 cpu_to_le16(ocfs2_extent_recs_per_eb(osb
->sb
));
1043 suballoc_bit_start
++;
1046 /* We'll also be dirtied by the caller, so
1047 * this isn't absolutely necessary. */
1048 ocfs2_journal_dirty(handle
, bhs
[i
]);
1057 for(i
= 0; i
< wanted
; i
++) {
1066 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1068 * Returns the sum of the rightmost extent rec logical offset and
1071 * ocfs2_add_branch() uses this to determine what logical cluster
1072 * value should be populated into the leftmost new branch records.
1074 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1075 * value for the new topmost tree record.
1077 static inline u32
ocfs2_sum_rightmost_rec(struct ocfs2_extent_list
*el
)
1081 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1083 return le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
1084 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
1088 * Change range of the branches in the right most path according to the leaf
1089 * extent block's rightmost record.
1091 static int ocfs2_adjust_rightmost_branch(handle_t
*handle
,
1092 struct ocfs2_extent_tree
*et
)
1095 struct ocfs2_path
*path
= NULL
;
1096 struct ocfs2_extent_list
*el
;
1097 struct ocfs2_extent_rec
*rec
;
1099 path
= ocfs2_new_path_from_et(et
);
1105 status
= ocfs2_find_path(et
->et_ci
, path
, UINT_MAX
);
1111 status
= ocfs2_extend_trans(handle
, path_num_items(path
));
1117 status
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
1123 el
= path_leaf_el(path
);
1124 rec
= &el
->l_recs
[le16_to_cpu(el
->l_next_free_rec
) - 1];
1126 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
1129 ocfs2_free_path(path
);
1134 * Add an entire tree branch to our inode. eb_bh is the extent block
1135 * to start at, if we don't want to start the branch at the root
1138 * last_eb_bh is required as we have to update it's next_leaf pointer
1139 * for the new last extent block.
1141 * the new branch will be 'empty' in the sense that every block will
1142 * contain a single record with cluster count == 0.
1144 static int ocfs2_add_branch(handle_t
*handle
,
1145 struct ocfs2_extent_tree
*et
,
1146 struct buffer_head
*eb_bh
,
1147 struct buffer_head
**last_eb_bh
,
1148 struct ocfs2_alloc_context
*meta_ac
)
1150 int status
, new_blocks
, i
, block_given
= 0;
1151 u64 next_blkno
, new_last_eb_blk
;
1152 struct buffer_head
*bh
;
1153 struct buffer_head
**new_eb_bhs
= NULL
;
1154 struct ocfs2_extent_block
*eb
;
1155 struct ocfs2_extent_list
*eb_el
;
1156 struct ocfs2_extent_list
*el
;
1157 u32 new_cpos
, root_end
;
1159 BUG_ON(!last_eb_bh
|| !*last_eb_bh
);
1162 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
1165 el
= et
->et_root_el
;
1167 /* we never add a branch to a leaf. */
1168 BUG_ON(!el
->l_tree_depth
);
1170 new_blocks
= le16_to_cpu(el
->l_tree_depth
);
1172 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
1173 new_cpos
= ocfs2_sum_rightmost_rec(&eb
->h_list
);
1174 root_end
= ocfs2_sum_rightmost_rec(et
->et_root_el
);
1177 * If there is a gap before the root end and the real end
1178 * of the righmost leaf block, we need to remove the gap
1179 * between new_cpos and root_end first so that the tree
1180 * is consistent after we add a new branch(it will start
1183 if (root_end
> new_cpos
) {
1184 trace_ocfs2_adjust_rightmost_branch(
1185 (unsigned long long)
1186 ocfs2_metadata_cache_owner(et
->et_ci
),
1187 root_end
, new_cpos
);
1189 status
= ocfs2_adjust_rightmost_branch(handle
, et
);
1196 /* allocate the number of new eb blocks we need */
1197 new_eb_bhs
= kcalloc(new_blocks
, sizeof(struct buffer_head
*),
1205 /* Firstyly, try to reuse dealloc since we have already estimated how
1206 * many extent blocks we may use.
1208 if (!ocfs2_is_dealloc_empty(et
)) {
1209 status
= ocfs2_reuse_blk_from_dealloc(handle
, et
,
1210 new_eb_bhs
, new_blocks
,
1218 BUG_ON(block_given
> new_blocks
);
1220 if (block_given
< new_blocks
) {
1222 status
= ocfs2_create_new_meta_bhs(handle
, et
,
1223 new_blocks
- block_given
,
1225 &new_eb_bhs
[block_given
]);
1232 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1233 * linked with the rest of the tree.
1234 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1236 * when we leave the loop, new_last_eb_blk will point to the
1237 * newest leaf, and next_blkno will point to the topmost extent
1239 next_blkno
= new_last_eb_blk
= 0;
1240 for(i
= 0; i
< new_blocks
; i
++) {
1242 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1243 /* ocfs2_create_new_meta_bhs() should create it right! */
1244 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1245 eb_el
= &eb
->h_list
;
1247 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, bh
,
1248 OCFS2_JOURNAL_ACCESS_CREATE
);
1254 eb
->h_next_leaf_blk
= 0;
1255 eb_el
->l_tree_depth
= cpu_to_le16(i
);
1256 eb_el
->l_next_free_rec
= cpu_to_le16(1);
1258 * This actually counts as an empty extent as
1261 eb_el
->l_recs
[0].e_cpos
= cpu_to_le32(new_cpos
);
1262 eb_el
->l_recs
[0].e_blkno
= cpu_to_le64(next_blkno
);
1264 * eb_el isn't always an interior node, but even leaf
1265 * nodes want a zero'd flags and reserved field so
1266 * this gets the whole 32 bits regardless of use.
1268 eb_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(0);
1269 if (!eb_el
->l_tree_depth
)
1270 new_last_eb_blk
= le64_to_cpu(eb
->h_blkno
);
1272 ocfs2_journal_dirty(handle
, bh
);
1273 next_blkno
= le64_to_cpu(eb
->h_blkno
);
1276 /* This is a bit hairy. We want to update up to three blocks
1277 * here without leaving any of them in an inconsistent state
1278 * in case of error. We don't have to worry about
1279 * journal_dirty erroring as it won't unless we've aborted the
1280 * handle (in which case we would never be here) so reserving
1281 * the write with journal_access is all we need to do. */
1282 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, *last_eb_bh
,
1283 OCFS2_JOURNAL_ACCESS_WRITE
);
1288 status
= ocfs2_et_root_journal_access(handle
, et
,
1289 OCFS2_JOURNAL_ACCESS_WRITE
);
1295 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, eb_bh
,
1296 OCFS2_JOURNAL_ACCESS_WRITE
);
1303 /* Link the new branch into the rest of the tree (el will
1304 * either be on the root_bh, or the extent block passed in. */
1305 i
= le16_to_cpu(el
->l_next_free_rec
);
1306 el
->l_recs
[i
].e_blkno
= cpu_to_le64(next_blkno
);
1307 el
->l_recs
[i
].e_cpos
= cpu_to_le32(new_cpos
);
1308 el
->l_recs
[i
].e_int_clusters
= 0;
1309 le16_add_cpu(&el
->l_next_free_rec
, 1);
1311 /* fe needs a new last extent block pointer, as does the
1312 * next_leaf on the previously last-extent-block. */
1313 ocfs2_et_set_last_eb_blk(et
, new_last_eb_blk
);
1315 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
1316 eb
->h_next_leaf_blk
= cpu_to_le64(new_last_eb_blk
);
1318 ocfs2_journal_dirty(handle
, *last_eb_bh
);
1319 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1321 ocfs2_journal_dirty(handle
, eb_bh
);
1324 * Some callers want to track the rightmost leaf so pass it
1327 brelse(*last_eb_bh
);
1328 get_bh(new_eb_bhs
[0]);
1329 *last_eb_bh
= new_eb_bhs
[0];
1334 for (i
= 0; i
< new_blocks
; i
++)
1335 brelse(new_eb_bhs
[i
]);
1343 * adds another level to the allocation tree.
1344 * returns back the new extent block so you can add a branch to it
1347 static int ocfs2_shift_tree_depth(handle_t
*handle
,
1348 struct ocfs2_extent_tree
*et
,
1349 struct ocfs2_alloc_context
*meta_ac
,
1350 struct buffer_head
**ret_new_eb_bh
)
1352 int status
, i
, block_given
= 0;
1354 struct buffer_head
*new_eb_bh
= NULL
;
1355 struct ocfs2_extent_block
*eb
;
1356 struct ocfs2_extent_list
*root_el
;
1357 struct ocfs2_extent_list
*eb_el
;
1359 if (!ocfs2_is_dealloc_empty(et
)) {
1360 status
= ocfs2_reuse_blk_from_dealloc(handle
, et
,
1363 } else if (meta_ac
) {
1364 status
= ocfs2_create_new_meta_bhs(handle
, et
, 1, meta_ac
,
1376 eb
= (struct ocfs2_extent_block
*) new_eb_bh
->b_data
;
1377 /* ocfs2_create_new_meta_bhs() should create it right! */
1378 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1380 eb_el
= &eb
->h_list
;
1381 root_el
= et
->et_root_el
;
1383 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, new_eb_bh
,
1384 OCFS2_JOURNAL_ACCESS_CREATE
);
1390 /* copy the root extent list data into the new extent block */
1391 eb_el
->l_tree_depth
= root_el
->l_tree_depth
;
1392 eb_el
->l_next_free_rec
= root_el
->l_next_free_rec
;
1393 for (i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1394 eb_el
->l_recs
[i
] = root_el
->l_recs
[i
];
1396 ocfs2_journal_dirty(handle
, new_eb_bh
);
1398 status
= ocfs2_et_root_journal_access(handle
, et
,
1399 OCFS2_JOURNAL_ACCESS_WRITE
);
1405 new_clusters
= ocfs2_sum_rightmost_rec(eb_el
);
1407 /* update root_bh now */
1408 le16_add_cpu(&root_el
->l_tree_depth
, 1);
1409 root_el
->l_recs
[0].e_cpos
= 0;
1410 root_el
->l_recs
[0].e_blkno
= eb
->h_blkno
;
1411 root_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(new_clusters
);
1412 for (i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1413 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
1414 root_el
->l_next_free_rec
= cpu_to_le16(1);
1416 /* If this is our 1st tree depth shift, then last_eb_blk
1417 * becomes the allocated extent block */
1418 if (root_el
->l_tree_depth
== cpu_to_le16(1))
1419 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
1421 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1423 *ret_new_eb_bh
= new_eb_bh
;
1433 * Should only be called when there is no space left in any of the
1434 * leaf nodes. What we want to do is find the lowest tree depth
1435 * non-leaf extent block with room for new records. There are three
1436 * valid results of this search:
1438 * 1) a lowest extent block is found, then we pass it back in
1439 * *lowest_eb_bh and return '0'
1441 * 2) the search fails to find anything, but the root_el has room. We
1442 * pass NULL back in *lowest_eb_bh, but still return '0'
1444 * 3) the search fails to find anything AND the root_el is full, in
1445 * which case we return > 0
1447 * return status < 0 indicates an error.
1449 static int ocfs2_find_branch_target(struct ocfs2_extent_tree
*et
,
1450 struct buffer_head
**target_bh
)
1454 struct ocfs2_extent_block
*eb
;
1455 struct ocfs2_extent_list
*el
;
1456 struct buffer_head
*bh
= NULL
;
1457 struct buffer_head
*lowest_bh
= NULL
;
1461 el
= et
->et_root_el
;
1463 while(le16_to_cpu(el
->l_tree_depth
) > 1) {
1464 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1465 status
= ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1466 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1467 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
1470 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1471 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1473 status
= ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1474 "Owner %llu has extent list where extent # %d has no physical block start\n",
1475 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
), i
);
1482 status
= ocfs2_read_extent_block(et
->et_ci
, blkno
, &bh
);
1488 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1491 if (le16_to_cpu(el
->l_next_free_rec
) <
1492 le16_to_cpu(el
->l_count
)) {
1499 /* If we didn't find one and the fe doesn't have any room,
1500 * then return '1' */
1501 el
= et
->et_root_el
;
1502 if (!lowest_bh
&& (el
->l_next_free_rec
== el
->l_count
))
1505 *target_bh
= lowest_bh
;
1513 * Grow a b-tree so that it has more records.
1515 * We might shift the tree depth in which case existing paths should
1516 * be considered invalid.
1518 * Tree depth after the grow is returned via *final_depth.
1520 * *last_eb_bh will be updated by ocfs2_add_branch().
1522 static int ocfs2_grow_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
1523 int *final_depth
, struct buffer_head
**last_eb_bh
,
1524 struct ocfs2_alloc_context
*meta_ac
)
1527 struct ocfs2_extent_list
*el
= et
->et_root_el
;
1528 int depth
= le16_to_cpu(el
->l_tree_depth
);
1529 struct buffer_head
*bh
= NULL
;
1531 BUG_ON(meta_ac
== NULL
&& ocfs2_is_dealloc_empty(et
));
1533 shift
= ocfs2_find_branch_target(et
, &bh
);
1540 /* We traveled all the way to the bottom of the allocation tree
1541 * and didn't find room for any more extents - we need to add
1542 * another tree level */
1545 trace_ocfs2_grow_tree(
1546 (unsigned long long)
1547 ocfs2_metadata_cache_owner(et
->et_ci
),
1550 /* ocfs2_shift_tree_depth will return us a buffer with
1551 * the new extent block (so we can pass that to
1552 * ocfs2_add_branch). */
1553 ret
= ocfs2_shift_tree_depth(handle
, et
, meta_ac
, &bh
);
1561 * Special case: we have room now if we shifted from
1562 * tree_depth 0, so no more work needs to be done.
1564 * We won't be calling add_branch, so pass
1565 * back *last_eb_bh as the new leaf. At depth
1566 * zero, it should always be null so there's
1567 * no reason to brelse.
1569 BUG_ON(*last_eb_bh
);
1576 /* call ocfs2_add_branch to add the final part of the tree with
1578 ret
= ocfs2_add_branch(handle
, et
, bh
, last_eb_bh
,
1585 *final_depth
= depth
;
1591 * This function will discard the rightmost extent record.
1593 static void ocfs2_shift_records_right(struct ocfs2_extent_list
*el
)
1595 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1596 int count
= le16_to_cpu(el
->l_count
);
1597 unsigned int num_bytes
;
1600 /* This will cause us to go off the end of our extent list. */
1601 BUG_ON(next_free
>= count
);
1603 num_bytes
= sizeof(struct ocfs2_extent_rec
) * next_free
;
1605 memmove(&el
->l_recs
[1], &el
->l_recs
[0], num_bytes
);
1608 static void ocfs2_rotate_leaf(struct ocfs2_extent_list
*el
,
1609 struct ocfs2_extent_rec
*insert_rec
)
1611 int i
, insert_index
, next_free
, has_empty
, num_bytes
;
1612 u32 insert_cpos
= le32_to_cpu(insert_rec
->e_cpos
);
1613 struct ocfs2_extent_rec
*rec
;
1615 next_free
= le16_to_cpu(el
->l_next_free_rec
);
1616 has_empty
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
1620 /* The tree code before us didn't allow enough room in the leaf. */
1621 BUG_ON(el
->l_next_free_rec
== el
->l_count
&& !has_empty
);
1624 * The easiest way to approach this is to just remove the
1625 * empty extent and temporarily decrement next_free.
1629 * If next_free was 1 (only an empty extent), this
1630 * loop won't execute, which is fine. We still want
1631 * the decrement above to happen.
1633 for(i
= 0; i
< (next_free
- 1); i
++)
1634 el
->l_recs
[i
] = el
->l_recs
[i
+1];
1640 * Figure out what the new record index should be.
1642 for(i
= 0; i
< next_free
; i
++) {
1643 rec
= &el
->l_recs
[i
];
1645 if (insert_cpos
< le32_to_cpu(rec
->e_cpos
))
1650 trace_ocfs2_rotate_leaf(insert_cpos
, insert_index
,
1651 has_empty
, next_free
,
1652 le16_to_cpu(el
->l_count
));
1654 BUG_ON(insert_index
< 0);
1655 BUG_ON(insert_index
>= le16_to_cpu(el
->l_count
));
1656 BUG_ON(insert_index
> next_free
);
1659 * No need to memmove if we're just adding to the tail.
1661 if (insert_index
!= next_free
) {
1662 BUG_ON(next_free
>= le16_to_cpu(el
->l_count
));
1664 num_bytes
= next_free
- insert_index
;
1665 num_bytes
*= sizeof(struct ocfs2_extent_rec
);
1666 memmove(&el
->l_recs
[insert_index
+ 1],
1667 &el
->l_recs
[insert_index
],
1672 * Either we had an empty extent, and need to re-increment or
1673 * there was no empty extent on a non full rightmost leaf node,
1674 * in which case we still need to increment.
1677 el
->l_next_free_rec
= cpu_to_le16(next_free
);
1679 * Make sure none of the math above just messed up our tree.
1681 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) > le16_to_cpu(el
->l_count
));
1683 el
->l_recs
[insert_index
] = *insert_rec
;
1687 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list
*el
)
1689 int size
, num_recs
= le16_to_cpu(el
->l_next_free_rec
);
1691 BUG_ON(num_recs
== 0);
1693 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
1695 size
= num_recs
* sizeof(struct ocfs2_extent_rec
);
1696 memmove(&el
->l_recs
[0], &el
->l_recs
[1], size
);
1697 memset(&el
->l_recs
[num_recs
], 0,
1698 sizeof(struct ocfs2_extent_rec
));
1699 el
->l_next_free_rec
= cpu_to_le16(num_recs
);
1704 * Create an empty extent record .
1706 * l_next_free_rec may be updated.
1708 * If an empty extent already exists do nothing.
1710 static void ocfs2_create_empty_extent(struct ocfs2_extent_list
*el
)
1712 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1714 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
1719 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
1722 mlog_bug_on_msg(el
->l_count
== el
->l_next_free_rec
,
1723 "Asked to create an empty extent in a full list:\n"
1724 "count = %u, tree depth = %u",
1725 le16_to_cpu(el
->l_count
),
1726 le16_to_cpu(el
->l_tree_depth
));
1728 ocfs2_shift_records_right(el
);
1731 le16_add_cpu(&el
->l_next_free_rec
, 1);
1732 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1736 * For a rotation which involves two leaf nodes, the "root node" is
1737 * the lowest level tree node which contains a path to both leafs. This
1738 * resulting set of information can be used to form a complete "subtree"
1740 * This function is passed two full paths from the dinode down to a
1741 * pair of adjacent leaves. It's task is to figure out which path
1742 * index contains the subtree root - this can be the root index itself
1743 * in a worst-case rotation.
1745 * The array index of the subtree root is passed back.
1747 int ocfs2_find_subtree_root(struct ocfs2_extent_tree
*et
,
1748 struct ocfs2_path
*left
,
1749 struct ocfs2_path
*right
)
1754 * Check that the caller passed in two paths from the same tree.
1756 BUG_ON(path_root_bh(left
) != path_root_bh(right
));
1762 * The caller didn't pass two adjacent paths.
1764 mlog_bug_on_msg(i
> left
->p_tree_depth
,
1765 "Owner %llu, left depth %u, right depth %u\n"
1766 "left leaf blk %llu, right leaf blk %llu\n",
1767 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
1768 left
->p_tree_depth
, right
->p_tree_depth
,
1769 (unsigned long long)path_leaf_bh(left
)->b_blocknr
,
1770 (unsigned long long)path_leaf_bh(right
)->b_blocknr
);
1771 } while (left
->p_node
[i
].bh
->b_blocknr
==
1772 right
->p_node
[i
].bh
->b_blocknr
);
1777 typedef void (path_insert_t
)(void *, struct buffer_head
*);
1780 * Traverse a btree path in search of cpos, starting at root_el.
1782 * This code can be called with a cpos larger than the tree, in which
1783 * case it will return the rightmost path.
1785 static int __ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1786 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1787 path_insert_t
*func
, void *data
)
1792 struct buffer_head
*bh
= NULL
;
1793 struct ocfs2_extent_block
*eb
;
1794 struct ocfs2_extent_list
*el
;
1795 struct ocfs2_extent_rec
*rec
;
1798 while (el
->l_tree_depth
) {
1799 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1800 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1801 "Owner %llu has empty extent list at depth %u\n",
1802 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1803 le16_to_cpu(el
->l_tree_depth
));
1809 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
) - 1; i
++) {
1810 rec
= &el
->l_recs
[i
];
1813 * In the case that cpos is off the allocation
1814 * tree, this should just wind up returning the
1817 range
= le32_to_cpu(rec
->e_cpos
) +
1818 ocfs2_rec_clusters(el
, rec
);
1819 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1823 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1825 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1826 "Owner %llu has bad blkno in extent list 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 at block %llu (next free=%u, count=%u)\n",
1848 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1849 (unsigned long long)bh
->b_blocknr
,
1850 le16_to_cpu(el
->l_next_free_rec
),
1851 le16_to_cpu(el
->l_count
));
1862 * Catch any trailing bh that the loop didn't handle.
1870 * Given an initialized path (that is, it has a valid root extent
1871 * list), this function will traverse the btree in search of the path
1872 * which would contain cpos.
1874 * The path traveled is recorded in the path structure.
1876 * Note that this will not do any comparisons on leaf node extent
1877 * records, so it will work fine in the case that we just added a tree
1880 struct find_path_data
{
1882 struct ocfs2_path
*path
;
1884 static void find_path_ins(void *data
, struct buffer_head
*bh
)
1886 struct find_path_data
*fp
= data
;
1889 ocfs2_path_insert_eb(fp
->path
, fp
->index
, bh
);
1892 int ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1893 struct ocfs2_path
*path
, u32 cpos
)
1895 struct find_path_data data
;
1899 return __ocfs2_find_path(ci
, path_root_el(path
), cpos
,
1900 find_path_ins
, &data
);
1903 static void find_leaf_ins(void *data
, struct buffer_head
*bh
)
1905 struct ocfs2_extent_block
*eb
=(struct ocfs2_extent_block
*)bh
->b_data
;
1906 struct ocfs2_extent_list
*el
= &eb
->h_list
;
1907 struct buffer_head
**ret
= data
;
1909 /* We want to retain only the leaf block. */
1910 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
1916 * Find the leaf block in the tree which would contain cpos. No
1917 * checking of the actual leaf is done.
1919 * Some paths want to call this instead of allocating a path structure
1920 * and calling ocfs2_find_path().
1922 * This function doesn't handle non btree extent lists.
1924 int ocfs2_find_leaf(struct ocfs2_caching_info
*ci
,
1925 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1926 struct buffer_head
**leaf_bh
)
1929 struct buffer_head
*bh
= NULL
;
1931 ret
= __ocfs2_find_path(ci
, root_el
, cpos
, find_leaf_ins
, &bh
);
1943 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1945 * Basically, we've moved stuff around at the bottom of the tree and
1946 * we need to fix up the extent records above the changes to reflect
1949 * left_rec: the record on the left.
1950 * right_rec: the record to the right of left_rec
1951 * right_child_el: is the child list pointed to by right_rec
1953 * By definition, this only works on interior nodes.
1955 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec
*left_rec
,
1956 struct ocfs2_extent_rec
*right_rec
,
1957 struct ocfs2_extent_list
*right_child_el
)
1959 u32 left_clusters
, right_end
;
1962 * Interior nodes never have holes. Their cpos is the cpos of
1963 * the leftmost record in their child list. Their cluster
1964 * count covers the full theoretical range of their child list
1965 * - the range between their cpos and the cpos of the record
1966 * immediately to their right.
1968 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[0].e_cpos
);
1969 if (!ocfs2_rec_clusters(right_child_el
, &right_child_el
->l_recs
[0])) {
1970 BUG_ON(right_child_el
->l_tree_depth
);
1971 BUG_ON(le16_to_cpu(right_child_el
->l_next_free_rec
) <= 1);
1972 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[1].e_cpos
);
1974 left_clusters
-= le32_to_cpu(left_rec
->e_cpos
);
1975 left_rec
->e_int_clusters
= cpu_to_le32(left_clusters
);
1978 * Calculate the rightmost cluster count boundary before
1979 * moving cpos - we will need to adjust clusters after
1980 * updating e_cpos to keep the same highest cluster count.
1982 right_end
= le32_to_cpu(right_rec
->e_cpos
);
1983 right_end
+= le32_to_cpu(right_rec
->e_int_clusters
);
1985 right_rec
->e_cpos
= left_rec
->e_cpos
;
1986 le32_add_cpu(&right_rec
->e_cpos
, left_clusters
);
1988 right_end
-= le32_to_cpu(right_rec
->e_cpos
);
1989 right_rec
->e_int_clusters
= cpu_to_le32(right_end
);
1993 * Adjust the adjacent root node records involved in a
1994 * rotation. left_el_blkno is passed in as a key so that we can easily
1995 * find it's index in the root list.
1997 static void ocfs2_adjust_root_records(struct ocfs2_extent_list
*root_el
,
1998 struct ocfs2_extent_list
*left_el
,
1999 struct ocfs2_extent_list
*right_el
,
2004 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) <=
2005 le16_to_cpu(left_el
->l_tree_depth
));
2007 for(i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
) - 1; i
++) {
2008 if (le64_to_cpu(root_el
->l_recs
[i
].e_blkno
) == left_el_blkno
)
2013 * The path walking code should have never returned a root and
2014 * two paths which are not adjacent.
2016 BUG_ON(i
>= (le16_to_cpu(root_el
->l_next_free_rec
) - 1));
2018 ocfs2_adjust_adjacent_records(&root_el
->l_recs
[i
],
2019 &root_el
->l_recs
[i
+ 1], right_el
);
2023 * We've changed a leaf block (in right_path) and need to reflect that
2024 * change back up the subtree.
2026 * This happens in multiple places:
2027 * - When we've moved an extent record from the left path leaf to the right
2028 * path leaf to make room for an empty extent in the left path leaf.
2029 * - When our insert into the right path leaf is at the leftmost edge
2030 * and requires an update of the path immediately to it's left. This
2031 * can occur at the end of some types of rotation and appending inserts.
2032 * - When we've adjusted the last extent record in the left path leaf and the
2033 * 1st extent record in the right path leaf during cross extent block merge.
2035 static void ocfs2_complete_edge_insert(handle_t
*handle
,
2036 struct ocfs2_path
*left_path
,
2037 struct ocfs2_path
*right_path
,
2041 struct ocfs2_extent_list
*el
, *left_el
, *right_el
;
2042 struct ocfs2_extent_rec
*left_rec
, *right_rec
;
2043 struct buffer_head
*root_bh
;
2046 * Update the counts and position values within all the
2047 * interior nodes to reflect the leaf rotation we just did.
2049 * The root node is handled below the loop.
2051 * We begin the loop with right_el and left_el pointing to the
2052 * leaf lists and work our way up.
2054 * NOTE: within this loop, left_el and right_el always refer
2055 * to the *child* lists.
2057 left_el
= path_leaf_el(left_path
);
2058 right_el
= path_leaf_el(right_path
);
2059 for(i
= left_path
->p_tree_depth
- 1; i
> subtree_index
; i
--) {
2060 trace_ocfs2_complete_edge_insert(i
);
2063 * One nice property of knowing that all of these
2064 * nodes are below the root is that we only deal with
2065 * the leftmost right node record and the rightmost
2068 el
= left_path
->p_node
[i
].el
;
2069 idx
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2070 left_rec
= &el
->l_recs
[idx
];
2072 el
= right_path
->p_node
[i
].el
;
2073 right_rec
= &el
->l_recs
[0];
2075 ocfs2_adjust_adjacent_records(left_rec
, right_rec
, right_el
);
2077 ocfs2_journal_dirty(handle
, left_path
->p_node
[i
].bh
);
2078 ocfs2_journal_dirty(handle
, right_path
->p_node
[i
].bh
);
2081 * Setup our list pointers now so that the current
2082 * parents become children in the next iteration.
2084 left_el
= left_path
->p_node
[i
].el
;
2085 right_el
= right_path
->p_node
[i
].el
;
2089 * At the root node, adjust the two adjacent records which
2090 * begin our path to the leaves.
2093 el
= left_path
->p_node
[subtree_index
].el
;
2094 left_el
= left_path
->p_node
[subtree_index
+ 1].el
;
2095 right_el
= right_path
->p_node
[subtree_index
+ 1].el
;
2097 ocfs2_adjust_root_records(el
, left_el
, right_el
,
2098 left_path
->p_node
[subtree_index
+ 1].bh
->b_blocknr
);
2100 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2102 ocfs2_journal_dirty(handle
, root_bh
);
2105 static int ocfs2_rotate_subtree_right(handle_t
*handle
,
2106 struct ocfs2_extent_tree
*et
,
2107 struct ocfs2_path
*left_path
,
2108 struct ocfs2_path
*right_path
,
2112 struct buffer_head
*right_leaf_bh
;
2113 struct buffer_head
*left_leaf_bh
= NULL
;
2114 struct buffer_head
*root_bh
;
2115 struct ocfs2_extent_list
*right_el
, *left_el
;
2116 struct ocfs2_extent_rec move_rec
;
2118 left_leaf_bh
= path_leaf_bh(left_path
);
2119 left_el
= path_leaf_el(left_path
);
2121 if (left_el
->l_next_free_rec
!= left_el
->l_count
) {
2122 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
2123 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2124 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2125 (unsigned long long)left_leaf_bh
->b_blocknr
,
2126 le16_to_cpu(left_el
->l_next_free_rec
));
2131 * This extent block may already have an empty record, so we
2132 * return early if so.
2134 if (ocfs2_is_empty_extent(&left_el
->l_recs
[0]))
2137 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2138 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2140 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2147 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2148 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2155 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2163 right_leaf_bh
= path_leaf_bh(right_path
);
2164 right_el
= path_leaf_el(right_path
);
2166 /* This is a code error, not a disk corruption. */
2167 mlog_bug_on_msg(!right_el
->l_next_free_rec
, "Inode %llu: Rotate fails "
2168 "because rightmost leaf block %llu is empty\n",
2169 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2170 (unsigned long long)right_leaf_bh
->b_blocknr
);
2172 ocfs2_create_empty_extent(right_el
);
2174 ocfs2_journal_dirty(handle
, right_leaf_bh
);
2176 /* Do the copy now. */
2177 i
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2178 move_rec
= left_el
->l_recs
[i
];
2179 right_el
->l_recs
[0] = move_rec
;
2182 * Clear out the record we just copied and shift everything
2183 * over, leaving an empty extent in the left leaf.
2185 * We temporarily subtract from next_free_rec so that the
2186 * shift will lose the tail record (which is now defunct).
2188 le16_add_cpu(&left_el
->l_next_free_rec
, -1);
2189 ocfs2_shift_records_right(left_el
);
2190 memset(&left_el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2191 le16_add_cpu(&left_el
->l_next_free_rec
, 1);
2193 ocfs2_journal_dirty(handle
, left_leaf_bh
);
2195 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2203 * Given a full path, determine what cpos value would return us a path
2204 * containing the leaf immediately to the left of the current one.
2206 * Will return zero if the path passed in is already the leftmost path.
2208 int ocfs2_find_cpos_for_left_leaf(struct super_block
*sb
,
2209 struct ocfs2_path
*path
, u32
*cpos
)
2213 struct ocfs2_extent_list
*el
;
2215 BUG_ON(path
->p_tree_depth
== 0);
2219 blkno
= path_leaf_bh(path
)->b_blocknr
;
2221 /* Start at the tree node just above the leaf and work our way up. */
2222 i
= path
->p_tree_depth
- 1;
2224 el
= path
->p_node
[i
].el
;
2227 * Find the extent record just before the one in our
2230 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2231 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2235 * We've determined that the
2236 * path specified is already
2237 * the leftmost one - return a
2243 * The leftmost record points to our
2244 * leaf - we need to travel up the
2250 *cpos
= le32_to_cpu(el
->l_recs
[j
- 1].e_cpos
);
2251 *cpos
= *cpos
+ ocfs2_rec_clusters(el
,
2252 &el
->l_recs
[j
- 1]);
2259 * If we got here, we never found a valid node where
2260 * the tree indicated one should be.
2262 ocfs2_error(sb
, "Invalid extent tree at extent block %llu\n",
2263 (unsigned long long)blkno
);
2268 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2277 * Extend the transaction by enough credits to complete the rotation,
2278 * and still leave at least the original number of credits allocated
2279 * to this transaction.
2281 static int ocfs2_extend_rotate_transaction(handle_t
*handle
, int subtree_depth
,
2283 struct ocfs2_path
*path
)
2286 int credits
= (path
->p_tree_depth
- subtree_depth
) * 2 + 1 + op_credits
;
2288 if (jbd2_handle_buffer_credits(handle
) < credits
)
2289 ret
= ocfs2_extend_trans(handle
,
2290 credits
- jbd2_handle_buffer_credits(handle
));
2296 * Trap the case where we're inserting into the theoretical range past
2297 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2298 * whose cpos is less than ours into the right leaf.
2300 * It's only necessary to look at the rightmost record of the left
2301 * leaf because the logic that calls us should ensure that the
2302 * theoretical ranges in the path components above the leaves are
2305 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path
*left_path
,
2308 struct ocfs2_extent_list
*left_el
;
2309 struct ocfs2_extent_rec
*rec
;
2312 left_el
= path_leaf_el(left_path
);
2313 next_free
= le16_to_cpu(left_el
->l_next_free_rec
);
2314 rec
= &left_el
->l_recs
[next_free
- 1];
2316 if (insert_cpos
> le32_to_cpu(rec
->e_cpos
))
2321 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list
*el
, u32 cpos
)
2323 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
2325 struct ocfs2_extent_rec
*rec
;
2330 rec
= &el
->l_recs
[0];
2331 if (ocfs2_is_empty_extent(rec
)) {
2335 rec
= &el
->l_recs
[1];
2338 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2339 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
2345 * Rotate all the records in a btree right one record, starting at insert_cpos.
2347 * The path to the rightmost leaf should be passed in.
2349 * The array is assumed to be large enough to hold an entire path (tree depth).
2351 * Upon successful return from this function:
2353 * - The 'right_path' array will contain a path to the leaf block
2354 * whose range contains e_cpos.
2355 * - That leaf block will have a single empty extent in list index 0.
2356 * - In the case that the rotation requires a post-insert update,
2357 * *ret_left_path will contain a valid path which can be passed to
2358 * ocfs2_insert_path().
2360 static int ocfs2_rotate_tree_right(handle_t
*handle
,
2361 struct ocfs2_extent_tree
*et
,
2362 enum ocfs2_split_type split
,
2364 struct ocfs2_path
*right_path
,
2365 struct ocfs2_path
**ret_left_path
)
2367 int ret
, start
, orig_credits
= jbd2_handle_buffer_credits(handle
);
2369 struct ocfs2_path
*left_path
= NULL
;
2370 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2372 *ret_left_path
= NULL
;
2374 left_path
= ocfs2_new_path_from_path(right_path
);
2381 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2387 trace_ocfs2_rotate_tree_right(
2388 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2392 * What we want to do here is:
2394 * 1) Start with the rightmost path.
2396 * 2) Determine a path to the leaf block directly to the left
2399 * 3) Determine the 'subtree root' - the lowest level tree node
2400 * which contains a path to both leaves.
2402 * 4) Rotate the subtree.
2404 * 5) Find the next subtree by considering the left path to be
2405 * the new right path.
2407 * The check at the top of this while loop also accepts
2408 * insert_cpos == cpos because cpos is only a _theoretical_
2409 * value to get us the left path - insert_cpos might very well
2410 * be filling that hole.
2412 * Stop at a cpos of '0' because we either started at the
2413 * leftmost branch (i.e., a tree with one branch and a
2414 * rotation inside of it), or we've gone as far as we can in
2415 * rotating subtrees.
2417 while (cpos
&& insert_cpos
<= cpos
) {
2418 trace_ocfs2_rotate_tree_right(
2419 (unsigned long long)
2420 ocfs2_metadata_cache_owner(et
->et_ci
),
2423 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
2429 mlog_bug_on_msg(path_leaf_bh(left_path
) ==
2430 path_leaf_bh(right_path
),
2431 "Owner %llu: error during insert of %u "
2432 "(left path cpos %u) results in two identical "
2433 "paths ending at %llu\n",
2434 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2436 (unsigned long long)
2437 path_leaf_bh(left_path
)->b_blocknr
);
2439 if (split
== SPLIT_NONE
&&
2440 ocfs2_rotate_requires_path_adjustment(left_path
,
2444 * We've rotated the tree as much as we
2445 * should. The rest is up to
2446 * ocfs2_insert_path() to complete, after the
2447 * record insertion. We indicate this
2448 * situation by returning the left path.
2450 * The reason we don't adjust the records here
2451 * before the record insert is that an error
2452 * later might break the rule where a parent
2453 * record e_cpos will reflect the actual
2454 * e_cpos of the 1st nonempty record of the
2457 *ret_left_path
= left_path
;
2461 start
= ocfs2_find_subtree_root(et
, left_path
, right_path
);
2463 trace_ocfs2_rotate_subtree(start
,
2464 (unsigned long long)
2465 right_path
->p_node
[start
].bh
->b_blocknr
,
2466 right_path
->p_tree_depth
);
2468 ret
= ocfs2_extend_rotate_transaction(handle
, start
,
2469 orig_credits
, right_path
);
2475 ret
= ocfs2_rotate_subtree_right(handle
, et
, left_path
,
2482 if (split
!= SPLIT_NONE
&&
2483 ocfs2_leftmost_rec_contains(path_leaf_el(right_path
),
2486 * A rotate moves the rightmost left leaf
2487 * record over to the leftmost right leaf
2488 * slot. If we're doing an extent split
2489 * instead of a real insert, then we have to
2490 * check that the extent to be split wasn't
2491 * just moved over. If it was, then we can
2492 * exit here, passing left_path back -
2493 * ocfs2_split_extent() is smart enough to
2494 * search both leaves.
2496 *ret_left_path
= left_path
;
2501 * There is no need to re-read the next right path
2502 * as we know that it'll be our current left
2503 * path. Optimize by copying values instead.
2505 ocfs2_mv_path(right_path
, left_path
);
2507 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2515 ocfs2_free_path(left_path
);
2521 static int ocfs2_update_edge_lengths(handle_t
*handle
,
2522 struct ocfs2_extent_tree
*et
,
2523 struct ocfs2_path
*path
)
2526 struct ocfs2_extent_rec
*rec
;
2527 struct ocfs2_extent_list
*el
;
2528 struct ocfs2_extent_block
*eb
;
2531 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
2537 /* Path should always be rightmost. */
2538 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2539 BUG_ON(eb
->h_next_leaf_blk
!= 0ULL);
2542 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
2543 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2544 rec
= &el
->l_recs
[idx
];
2545 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2547 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
2548 el
= path
->p_node
[i
].el
;
2549 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2550 rec
= &el
->l_recs
[idx
];
2552 rec
->e_int_clusters
= cpu_to_le32(range
);
2553 le32_add_cpu(&rec
->e_int_clusters
, -le32_to_cpu(rec
->e_cpos
));
2555 ocfs2_journal_dirty(handle
, path
->p_node
[i
].bh
);
2561 static void ocfs2_unlink_path(handle_t
*handle
,
2562 struct ocfs2_extent_tree
*et
,
2563 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2564 struct ocfs2_path
*path
, int unlink_start
)
2567 struct ocfs2_extent_block
*eb
;
2568 struct ocfs2_extent_list
*el
;
2569 struct buffer_head
*bh
;
2571 for(i
= unlink_start
; i
< path_num_items(path
); i
++) {
2572 bh
= path
->p_node
[i
].bh
;
2574 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
2576 * Not all nodes might have had their final count
2577 * decremented by the caller - handle this here.
2580 if (le16_to_cpu(el
->l_next_free_rec
) > 1) {
2582 "Inode %llu, attempted to remove extent block "
2583 "%llu with %u records\n",
2584 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2585 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
2586 le16_to_cpu(el
->l_next_free_rec
));
2588 ocfs2_journal_dirty(handle
, bh
);
2589 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2593 el
->l_next_free_rec
= 0;
2594 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2596 ocfs2_journal_dirty(handle
, bh
);
2598 ret
= ocfs2_cache_extent_block_free(dealloc
, eb
);
2602 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2606 static void ocfs2_unlink_subtree(handle_t
*handle
,
2607 struct ocfs2_extent_tree
*et
,
2608 struct ocfs2_path
*left_path
,
2609 struct ocfs2_path
*right_path
,
2611 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2614 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2615 struct ocfs2_extent_list
*root_el
= left_path
->p_node
[subtree_index
].el
;
2616 struct ocfs2_extent_block
*eb
;
2618 eb
= (struct ocfs2_extent_block
*)right_path
->p_node
[subtree_index
+ 1].bh
->b_data
;
2620 for(i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
2621 if (root_el
->l_recs
[i
].e_blkno
== eb
->h_blkno
)
2624 BUG_ON(i
>= le16_to_cpu(root_el
->l_next_free_rec
));
2626 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
2627 le16_add_cpu(&root_el
->l_next_free_rec
, -1);
2629 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2630 eb
->h_next_leaf_blk
= 0;
2632 ocfs2_journal_dirty(handle
, root_bh
);
2633 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2635 ocfs2_unlink_path(handle
, et
, dealloc
, right_path
,
2639 static int ocfs2_rotate_subtree_left(handle_t
*handle
,
2640 struct ocfs2_extent_tree
*et
,
2641 struct ocfs2_path
*left_path
,
2642 struct ocfs2_path
*right_path
,
2644 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2647 int ret
, i
, del_right_subtree
= 0, right_has_empty
= 0;
2648 struct buffer_head
*root_bh
, *et_root_bh
= path_root_bh(right_path
);
2649 struct ocfs2_extent_list
*right_leaf_el
, *left_leaf_el
;
2650 struct ocfs2_extent_block
*eb
;
2654 right_leaf_el
= path_leaf_el(right_path
);
2655 left_leaf_el
= path_leaf_el(left_path
);
2656 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2657 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2659 if (!ocfs2_is_empty_extent(&left_leaf_el
->l_recs
[0]))
2662 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(right_path
)->b_data
;
2663 if (ocfs2_is_empty_extent(&right_leaf_el
->l_recs
[0])) {
2665 * It's legal for us to proceed if the right leaf is
2666 * the rightmost one and it has an empty extent. There
2667 * are two cases to handle - whether the leaf will be
2668 * empty after removal or not. If the leaf isn't empty
2669 * then just remove the empty extent up front. The
2670 * next block will handle empty leaves by flagging
2673 * Non rightmost leaves will throw -EAGAIN and the
2674 * caller can manually move the subtree and retry.
2677 if (eb
->h_next_leaf_blk
!= 0ULL)
2680 if (le16_to_cpu(right_leaf_el
->l_next_free_rec
) > 1) {
2681 ret
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
2682 path_leaf_bh(right_path
),
2683 OCFS2_JOURNAL_ACCESS_WRITE
);
2689 ocfs2_remove_empty_extent(right_leaf_el
);
2691 right_has_empty
= 1;
2694 if (eb
->h_next_leaf_blk
== 0ULL &&
2695 le16_to_cpu(right_leaf_el
->l_next_free_rec
) == 1) {
2697 * We have to update i_last_eb_blk during the meta
2700 ret
= ocfs2_et_root_journal_access(handle
, et
,
2701 OCFS2_JOURNAL_ACCESS_WRITE
);
2707 del_right_subtree
= 1;
2711 * Getting here with an empty extent in the right path implies
2712 * that it's the rightmost path and will be deleted.
2714 BUG_ON(right_has_empty
&& !del_right_subtree
);
2716 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2723 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2724 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2731 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2739 if (!right_has_empty
) {
2741 * Only do this if we're moving a real
2742 * record. Otherwise, the action is delayed until
2743 * after removal of the right path in which case we
2744 * can do a simple shift to remove the empty extent.
2746 ocfs2_rotate_leaf(left_leaf_el
, &right_leaf_el
->l_recs
[0]);
2747 memset(&right_leaf_el
->l_recs
[0], 0,
2748 sizeof(struct ocfs2_extent_rec
));
2750 if (eb
->h_next_leaf_blk
== 0ULL) {
2752 * Move recs over to get rid of empty extent, decrease
2753 * next_free. This is allowed to remove the last
2754 * extent in our leaf (setting l_next_free_rec to
2755 * zero) - the delete code below won't care.
2757 ocfs2_remove_empty_extent(right_leaf_el
);
2760 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2761 ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2763 if (del_right_subtree
) {
2764 ocfs2_unlink_subtree(handle
, et
, left_path
, right_path
,
2765 subtree_index
, dealloc
);
2766 ret
= ocfs2_update_edge_lengths(handle
, et
, left_path
);
2772 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2773 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
2776 * Removal of the extent in the left leaf was skipped
2777 * above so we could delete the right path
2780 if (right_has_empty
)
2781 ocfs2_remove_empty_extent(left_leaf_el
);
2783 ocfs2_journal_dirty(handle
, et_root_bh
);
2787 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2795 * Given a full path, determine what cpos value would return us a path
2796 * containing the leaf immediately to the right of the current one.
2798 * Will return zero if the path passed in is already the rightmost path.
2800 * This looks similar, but is subtly different to
2801 * ocfs2_find_cpos_for_left_leaf().
2803 int ocfs2_find_cpos_for_right_leaf(struct super_block
*sb
,
2804 struct ocfs2_path
*path
, u32
*cpos
)
2808 struct ocfs2_extent_list
*el
;
2812 if (path
->p_tree_depth
== 0)
2815 blkno
= path_leaf_bh(path
)->b_blocknr
;
2817 /* Start at the tree node just above the leaf and work our way up. */
2818 i
= path
->p_tree_depth
- 1;
2822 el
= path
->p_node
[i
].el
;
2825 * Find the extent record just after the one in our
2828 next_free
= le16_to_cpu(el
->l_next_free_rec
);
2829 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2830 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2831 if (j
== (next_free
- 1)) {
2834 * We've determined that the
2835 * path specified is already
2836 * the rightmost one - return a
2842 * The rightmost record points to our
2843 * leaf - we need to travel up the
2849 *cpos
= le32_to_cpu(el
->l_recs
[j
+ 1].e_cpos
);
2855 * If we got here, we never found a valid node where
2856 * the tree indicated one should be.
2858 ocfs2_error(sb
, "Invalid extent tree at extent block %llu\n",
2859 (unsigned long long)blkno
);
2864 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2872 static int ocfs2_rotate_rightmost_leaf_left(handle_t
*handle
,
2873 struct ocfs2_extent_tree
*et
,
2874 struct ocfs2_path
*path
)
2877 struct buffer_head
*bh
= path_leaf_bh(path
);
2878 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
2880 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2883 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
2884 path_num_items(path
) - 1);
2890 ocfs2_remove_empty_extent(el
);
2891 ocfs2_journal_dirty(handle
, bh
);
2897 static int __ocfs2_rotate_tree_left(handle_t
*handle
,
2898 struct ocfs2_extent_tree
*et
,
2900 struct ocfs2_path
*path
,
2901 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2902 struct ocfs2_path
**empty_extent_path
)
2904 int ret
, subtree_root
, deleted
;
2906 struct ocfs2_path
*left_path
= NULL
;
2907 struct ocfs2_path
*right_path
= NULL
;
2908 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2910 if (!ocfs2_is_empty_extent(&(path_leaf_el(path
)->l_recs
[0])))
2913 *empty_extent_path
= NULL
;
2915 ret
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
2921 left_path
= ocfs2_new_path_from_path(path
);
2928 ocfs2_cp_path(left_path
, path
);
2930 right_path
= ocfs2_new_path_from_path(path
);
2937 while (right_cpos
) {
2938 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
2944 subtree_root
= ocfs2_find_subtree_root(et
, left_path
,
2947 trace_ocfs2_rotate_subtree(subtree_root
,
2948 (unsigned long long)
2949 right_path
->p_node
[subtree_root
].bh
->b_blocknr
,
2950 right_path
->p_tree_depth
);
2952 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
2953 orig_credits
, left_path
);
2960 * Caller might still want to make changes to the
2961 * tree root, so re-add it to the journal here.
2963 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2970 ret
= ocfs2_rotate_subtree_left(handle
, et
, left_path
,
2971 right_path
, subtree_root
,
2973 if (ret
== -EAGAIN
) {
2975 * The rotation has to temporarily stop due to
2976 * the right subtree having an empty
2977 * extent. Pass it back to the caller for a
2980 *empty_extent_path
= right_path
;
2990 * The subtree rotate might have removed records on
2991 * the rightmost edge. If so, then rotation is
2997 ocfs2_mv_path(left_path
, right_path
);
2999 ret
= ocfs2_find_cpos_for_right_leaf(sb
, left_path
,
3008 ocfs2_free_path(right_path
);
3009 ocfs2_free_path(left_path
);
3014 static int ocfs2_remove_rightmost_path(handle_t
*handle
,
3015 struct ocfs2_extent_tree
*et
,
3016 struct ocfs2_path
*path
,
3017 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3019 int ret
, subtree_index
;
3021 struct ocfs2_path
*left_path
= NULL
;
3022 struct ocfs2_extent_block
*eb
;
3023 struct ocfs2_extent_list
*el
;
3025 ret
= ocfs2_et_sanity_check(et
);
3029 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
3035 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3044 * We have a path to the left of this one - it needs
3047 left_path
= ocfs2_new_path_from_path(path
);
3054 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
3060 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
3066 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
3068 ocfs2_unlink_subtree(handle
, et
, left_path
, path
,
3069 subtree_index
, dealloc
);
3070 ret
= ocfs2_update_edge_lengths(handle
, et
, left_path
);
3076 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
3077 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
3080 * 'path' is also the leftmost path which
3081 * means it must be the only one. This gets
3082 * handled differently because we want to
3083 * revert the root back to having extents
3086 ocfs2_unlink_path(handle
, et
, dealloc
, path
, 1);
3088 el
= et
->et_root_el
;
3089 el
->l_tree_depth
= 0;
3090 el
->l_next_free_rec
= 0;
3091 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3093 ocfs2_et_set_last_eb_blk(et
, 0);
3096 ocfs2_journal_dirty(handle
, path_root_bh(path
));
3099 ocfs2_free_path(left_path
);
3103 static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super
*osb
,
3104 struct ocfs2_extent_tree
*et
,
3105 struct ocfs2_path
*path
,
3106 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3110 int credits
= path
->p_tree_depth
* 2 + 1;
3112 handle
= ocfs2_start_trans(osb
, credits
);
3113 if (IS_ERR(handle
)) {
3114 ret
= PTR_ERR(handle
);
3119 ret
= ocfs2_remove_rightmost_path(handle
, et
, path
, dealloc
);
3123 ocfs2_commit_trans(osb
, handle
);
3128 * Left rotation of btree records.
3130 * In many ways, this is (unsurprisingly) the opposite of right
3131 * rotation. We start at some non-rightmost path containing an empty
3132 * extent in the leaf block. The code works its way to the rightmost
3133 * path by rotating records to the left in every subtree.
3135 * This is used by any code which reduces the number of extent records
3136 * in a leaf. After removal, an empty record should be placed in the
3137 * leftmost list position.
3139 * This won't handle a length update of the rightmost path records if
3140 * the rightmost tree leaf record is removed so the caller is
3141 * responsible for detecting and correcting that.
3143 static int ocfs2_rotate_tree_left(handle_t
*handle
,
3144 struct ocfs2_extent_tree
*et
,
3145 struct ocfs2_path
*path
,
3146 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3148 int ret
, orig_credits
= jbd2_handle_buffer_credits(handle
);
3149 struct ocfs2_path
*tmp_path
= NULL
, *restart_path
= NULL
;
3150 struct ocfs2_extent_block
*eb
;
3151 struct ocfs2_extent_list
*el
;
3153 el
= path_leaf_el(path
);
3154 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
3157 if (path
->p_tree_depth
== 0) {
3158 rightmost_no_delete
:
3160 * Inline extents. This is trivially handled, so do
3163 ret
= ocfs2_rotate_rightmost_leaf_left(handle
, et
, path
);
3170 * Handle rightmost branch now. There's several cases:
3171 * 1) simple rotation leaving records in there. That's trivial.
3172 * 2) rotation requiring a branch delete - there's no more
3173 * records left. Two cases of this:
3174 * a) There are branches to the left.
3175 * b) This is also the leftmost (the only) branch.
3177 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3178 * 2a) we need the left branch so that we can update it with the unlink
3179 * 2b) we need to bring the root back to inline extents.
3182 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
3184 if (eb
->h_next_leaf_blk
== 0) {
3186 * This gets a bit tricky if we're going to delete the
3187 * rightmost path. Get the other cases out of the way
3190 if (le16_to_cpu(el
->l_next_free_rec
) > 1)
3191 goto rightmost_no_delete
;
3193 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
3194 ret
= ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
3195 "Owner %llu has empty extent block at %llu\n",
3196 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
3197 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
3202 * XXX: The caller can not trust "path" any more after
3203 * this as it will have been deleted. What do we do?
3205 * In theory the rotate-for-merge code will never get
3206 * here because it'll always ask for a rotate in a
3210 ret
= ocfs2_remove_rightmost_path(handle
, et
, path
,
3218 * Now we can loop, remembering the path we get from -EAGAIN
3219 * and restarting from there.
3222 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
, path
,
3223 dealloc
, &restart_path
);
3224 if (ret
&& ret
!= -EAGAIN
) {
3229 while (ret
== -EAGAIN
) {
3230 tmp_path
= restart_path
;
3231 restart_path
= NULL
;
3233 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
,
3236 if (ret
&& ret
!= -EAGAIN
) {
3241 ocfs2_free_path(tmp_path
);
3249 ocfs2_free_path(tmp_path
);
3250 ocfs2_free_path(restart_path
);
3254 static void ocfs2_cleanup_merge(struct ocfs2_extent_list
*el
,
3257 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[index
];
3260 if (rec
->e_leaf_clusters
== 0) {
3262 * We consumed all of the merged-from record. An empty
3263 * extent cannot exist anywhere but the 1st array
3264 * position, so move things over if the merged-from
3265 * record doesn't occupy that position.
3267 * This creates a new empty extent so the caller
3268 * should be smart enough to have removed any existing
3272 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3273 size
= index
* sizeof(struct ocfs2_extent_rec
);
3274 memmove(&el
->l_recs
[1], &el
->l_recs
[0], size
);
3278 * Always memset - the caller doesn't check whether it
3279 * created an empty extent, so there could be junk in
3282 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3286 static int ocfs2_get_right_path(struct ocfs2_extent_tree
*et
,
3287 struct ocfs2_path
*left_path
,
3288 struct ocfs2_path
**ret_right_path
)
3292 struct ocfs2_path
*right_path
= NULL
;
3293 struct ocfs2_extent_list
*left_el
;
3295 *ret_right_path
= NULL
;
3297 /* This function shouldn't be called for non-trees. */
3298 BUG_ON(left_path
->p_tree_depth
== 0);
3300 left_el
= path_leaf_el(left_path
);
3301 BUG_ON(left_el
->l_next_free_rec
!= left_el
->l_count
);
3303 ret
= ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3304 left_path
, &right_cpos
);
3310 /* This function shouldn't be called for the rightmost leaf. */
3311 BUG_ON(right_cpos
== 0);
3313 right_path
= ocfs2_new_path_from_path(left_path
);
3320 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
3326 *ret_right_path
= right_path
;
3329 ocfs2_free_path(right_path
);
3334 * Remove split_rec clusters from the record at index and merge them
3335 * onto the beginning of the record "next" to it.
3336 * For index < l_count - 1, the next means the extent rec at index + 1.
3337 * For index == l_count - 1, the "next" means the 1st extent rec of the
3338 * next extent block.
3340 static int ocfs2_merge_rec_right(struct ocfs2_path
*left_path
,
3342 struct ocfs2_extent_tree
*et
,
3343 struct ocfs2_extent_rec
*split_rec
,
3346 int ret
, next_free
, i
;
3347 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3348 struct ocfs2_extent_rec
*left_rec
;
3349 struct ocfs2_extent_rec
*right_rec
;
3350 struct ocfs2_extent_list
*right_el
;
3351 struct ocfs2_path
*right_path
= NULL
;
3352 int subtree_index
= 0;
3353 struct ocfs2_extent_list
*el
= path_leaf_el(left_path
);
3354 struct buffer_head
*bh
= path_leaf_bh(left_path
);
3355 struct buffer_head
*root_bh
= NULL
;
3357 BUG_ON(index
>= le16_to_cpu(el
->l_next_free_rec
));
3358 left_rec
= &el
->l_recs
[index
];
3360 if (index
== le16_to_cpu(el
->l_next_free_rec
) - 1 &&
3361 le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
)) {
3362 /* we meet with a cross extent block merge. */
3363 ret
= ocfs2_get_right_path(et
, left_path
, &right_path
);
3369 right_el
= path_leaf_el(right_path
);
3370 next_free
= le16_to_cpu(right_el
->l_next_free_rec
);
3371 BUG_ON(next_free
<= 0);
3372 right_rec
= &right_el
->l_recs
[0];
3373 if (ocfs2_is_empty_extent(right_rec
)) {
3374 BUG_ON(next_free
<= 1);
3375 right_rec
= &right_el
->l_recs
[1];
3378 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3379 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3380 le32_to_cpu(right_rec
->e_cpos
));
3382 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3385 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3386 jbd2_handle_buffer_credits(handle
),
3393 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3394 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3396 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3403 for (i
= subtree_index
+ 1;
3404 i
< path_num_items(right_path
); i
++) {
3405 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3412 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3421 BUG_ON(index
== le16_to_cpu(el
->l_next_free_rec
) - 1);
3422 right_rec
= &el
->l_recs
[index
+ 1];
3425 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, left_path
,
3426 path_num_items(left_path
) - 1);
3432 le16_add_cpu(&left_rec
->e_leaf_clusters
, -split_clusters
);
3434 le32_add_cpu(&right_rec
->e_cpos
, -split_clusters
);
3435 le64_add_cpu(&right_rec
->e_blkno
,
3436 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3438 le16_add_cpu(&right_rec
->e_leaf_clusters
, split_clusters
);
3440 ocfs2_cleanup_merge(el
, index
);
3442 ocfs2_journal_dirty(handle
, bh
);
3444 ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
3445 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
3449 ocfs2_free_path(right_path
);
3453 static int ocfs2_get_left_path(struct ocfs2_extent_tree
*et
,
3454 struct ocfs2_path
*right_path
,
3455 struct ocfs2_path
**ret_left_path
)
3459 struct ocfs2_path
*left_path
= NULL
;
3461 *ret_left_path
= NULL
;
3463 /* This function shouldn't be called for non-trees. */
3464 BUG_ON(right_path
->p_tree_depth
== 0);
3466 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3467 right_path
, &left_cpos
);
3473 /* This function shouldn't be called for the leftmost leaf. */
3474 BUG_ON(left_cpos
== 0);
3476 left_path
= ocfs2_new_path_from_path(right_path
);
3483 ret
= ocfs2_find_path(et
->et_ci
, left_path
, left_cpos
);
3489 *ret_left_path
= left_path
;
3492 ocfs2_free_path(left_path
);
3497 * Remove split_rec clusters from the record at index and merge them
3498 * onto the tail of the record "before" it.
3499 * For index > 0, the "before" means the extent rec at index - 1.
3501 * For index == 0, the "before" means the last record of the previous
3502 * extent block. And there is also a situation that we may need to
3503 * remove the rightmost leaf extent block in the right_path and change
3504 * the right path to indicate the new rightmost path.
3506 static int ocfs2_merge_rec_left(struct ocfs2_path
*right_path
,
3508 struct ocfs2_extent_tree
*et
,
3509 struct ocfs2_extent_rec
*split_rec
,
3510 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3513 int ret
, i
, subtree_index
= 0, has_empty_extent
= 0;
3514 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3515 struct ocfs2_extent_rec
*left_rec
;
3516 struct ocfs2_extent_rec
*right_rec
;
3517 struct ocfs2_extent_list
*el
= path_leaf_el(right_path
);
3518 struct buffer_head
*bh
= path_leaf_bh(right_path
);
3519 struct buffer_head
*root_bh
= NULL
;
3520 struct ocfs2_path
*left_path
= NULL
;
3521 struct ocfs2_extent_list
*left_el
;
3525 right_rec
= &el
->l_recs
[index
];
3527 /* we meet with a cross extent block merge. */
3528 ret
= ocfs2_get_left_path(et
, right_path
, &left_path
);
3534 left_el
= path_leaf_el(left_path
);
3535 BUG_ON(le16_to_cpu(left_el
->l_next_free_rec
) !=
3536 le16_to_cpu(left_el
->l_count
));
3538 left_rec
= &left_el
->l_recs
[
3539 le16_to_cpu(left_el
->l_next_free_rec
) - 1];
3540 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3541 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3542 le32_to_cpu(split_rec
->e_cpos
));
3544 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3547 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3548 jbd2_handle_buffer_credits(handle
),
3555 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3556 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3558 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3565 for (i
= subtree_index
+ 1;
3566 i
< path_num_items(right_path
); i
++) {
3567 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3574 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3582 left_rec
= &el
->l_recs
[index
- 1];
3583 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
3584 has_empty_extent
= 1;
3587 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3588 path_num_items(right_path
) - 1);
3594 if (has_empty_extent
&& index
== 1) {
3596 * The easy case - we can just plop the record right in.
3598 *left_rec
= *split_rec
;
3600 le16_add_cpu(&left_rec
->e_leaf_clusters
, split_clusters
);
3602 le32_add_cpu(&right_rec
->e_cpos
, split_clusters
);
3603 le64_add_cpu(&right_rec
->e_blkno
,
3604 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3606 le16_add_cpu(&right_rec
->e_leaf_clusters
, -split_clusters
);
3608 ocfs2_cleanup_merge(el
, index
);
3610 ocfs2_journal_dirty(handle
, bh
);
3612 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
3615 * In the situation that the right_rec is empty and the extent
3616 * block is empty also, ocfs2_complete_edge_insert can't handle
3617 * it and we need to delete the right extent block.
3619 if (le16_to_cpu(right_rec
->e_leaf_clusters
) == 0 &&
3620 le16_to_cpu(el
->l_next_free_rec
) == 1) {
3621 /* extend credit for ocfs2_remove_rightmost_path */
3622 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3623 jbd2_handle_buffer_credits(handle
),
3630 ret
= ocfs2_remove_rightmost_path(handle
, et
,
3638 /* Now the rightmost extent block has been deleted.
3639 * So we use the new rightmost path.
3641 ocfs2_mv_path(right_path
, left_path
);
3644 ocfs2_complete_edge_insert(handle
, left_path
,
3645 right_path
, subtree_index
);
3648 ocfs2_free_path(left_path
);
3652 static int ocfs2_try_to_merge_extent(handle_t
*handle
,
3653 struct ocfs2_extent_tree
*et
,
3654 struct ocfs2_path
*path
,
3656 struct ocfs2_extent_rec
*split_rec
,
3657 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3658 struct ocfs2_merge_ctxt
*ctxt
)
3661 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
3662 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
3664 BUG_ON(ctxt
->c_contig_type
== CONTIG_NONE
);
3666 if (ctxt
->c_split_covers_rec
&& ctxt
->c_has_empty_extent
) {
3667 /* extend credit for ocfs2_remove_rightmost_path */
3668 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3669 jbd2_handle_buffer_credits(handle
),
3676 * The merge code will need to create an empty
3677 * extent to take the place of the newly
3678 * emptied slot. Remove any pre-existing empty
3679 * extents - having more than one in a leaf is
3682 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3688 rec
= &el
->l_recs
[split_index
];
3691 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
) {
3693 * Left-right contig implies this.
3695 BUG_ON(!ctxt
->c_split_covers_rec
);
3698 * Since the leftright insert always covers the entire
3699 * extent, this call will delete the insert record
3700 * entirely, resulting in an empty extent record added to
3703 * Since the adding of an empty extent shifts
3704 * everything back to the right, there's no need to
3705 * update split_index here.
3707 * When the split_index is zero, we need to merge it to the
3708 * prevoius extent block. It is more efficient and easier
3709 * if we do merge_right first and merge_left later.
3711 ret
= ocfs2_merge_rec_right(path
, handle
, et
, split_rec
,
3719 * We can only get this from logic error above.
3721 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
3723 /* extend credit for ocfs2_remove_rightmost_path */
3724 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3725 jbd2_handle_buffer_credits(handle
),
3732 /* The merge left us with an empty extent, remove it. */
3733 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3739 rec
= &el
->l_recs
[split_index
];
3742 * Note that we don't pass split_rec here on purpose -
3743 * we've merged it into the rec already.
3745 ret
= ocfs2_merge_rec_left(path
, handle
, et
, rec
,
3746 dealloc
, split_index
);
3753 /* extend credit for ocfs2_remove_rightmost_path */
3754 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3755 jbd2_handle_buffer_credits(handle
),
3762 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3764 * Error from this last rotate is not critical, so
3765 * print but don't bubble it up.
3772 * Merge a record to the left or right.
3774 * 'contig_type' is relative to the existing record,
3775 * so for example, if we're "right contig", it's to
3776 * the record on the left (hence the left merge).
3778 if (ctxt
->c_contig_type
== CONTIG_RIGHT
) {
3779 ret
= ocfs2_merge_rec_left(path
, handle
, et
,
3787 ret
= ocfs2_merge_rec_right(path
, handle
,
3796 if (ctxt
->c_split_covers_rec
) {
3797 /* extend credit for ocfs2_remove_rightmost_path */
3798 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3799 jbd2_handle_buffer_credits(handle
),
3808 * The merge may have left an empty extent in
3809 * our leaf. Try to rotate it away.
3811 ret
= ocfs2_rotate_tree_left(handle
, et
, path
,
3823 static void ocfs2_subtract_from_rec(struct super_block
*sb
,
3824 enum ocfs2_split_type split
,
3825 struct ocfs2_extent_rec
*rec
,
3826 struct ocfs2_extent_rec
*split_rec
)
3830 len_blocks
= ocfs2_clusters_to_blocks(sb
,
3831 le16_to_cpu(split_rec
->e_leaf_clusters
));
3833 if (split
== SPLIT_LEFT
) {
3835 * Region is on the left edge of the existing
3838 le32_add_cpu(&rec
->e_cpos
,
3839 le16_to_cpu(split_rec
->e_leaf_clusters
));
3840 le64_add_cpu(&rec
->e_blkno
, len_blocks
);
3841 le16_add_cpu(&rec
->e_leaf_clusters
,
3842 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3845 * Region is on the right edge of the existing
3848 le16_add_cpu(&rec
->e_leaf_clusters
,
3849 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3854 * Do the final bits of extent record insertion at the target leaf
3855 * list. If this leaf is part of an allocation tree, it is assumed
3856 * that the tree above has been prepared.
3858 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree
*et
,
3859 struct ocfs2_extent_rec
*insert_rec
,
3860 struct ocfs2_extent_list
*el
,
3861 struct ocfs2_insert_type
*insert
)
3863 int i
= insert
->ins_contig_index
;
3865 struct ocfs2_extent_rec
*rec
;
3867 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3869 if (insert
->ins_split
!= SPLIT_NONE
) {
3870 i
= ocfs2_search_extent_list(el
, le32_to_cpu(insert_rec
->e_cpos
));
3872 rec
= &el
->l_recs
[i
];
3873 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
3874 insert
->ins_split
, rec
,
3880 * Contiguous insert - either left or right.
3882 if (insert
->ins_contig
!= CONTIG_NONE
) {
3883 rec
= &el
->l_recs
[i
];
3884 if (insert
->ins_contig
== CONTIG_LEFT
) {
3885 rec
->e_blkno
= insert_rec
->e_blkno
;
3886 rec
->e_cpos
= insert_rec
->e_cpos
;
3888 le16_add_cpu(&rec
->e_leaf_clusters
,
3889 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3894 * Handle insert into an empty leaf.
3896 if (le16_to_cpu(el
->l_next_free_rec
) == 0 ||
3897 ((le16_to_cpu(el
->l_next_free_rec
) == 1) &&
3898 ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3899 el
->l_recs
[0] = *insert_rec
;
3900 el
->l_next_free_rec
= cpu_to_le16(1);
3907 if (insert
->ins_appending
== APPEND_TAIL
) {
3908 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3909 rec
= &el
->l_recs
[i
];
3910 range
= le32_to_cpu(rec
->e_cpos
)
3911 + le16_to_cpu(rec
->e_leaf_clusters
);
3912 BUG_ON(le32_to_cpu(insert_rec
->e_cpos
) < range
);
3914 mlog_bug_on_msg(le16_to_cpu(el
->l_next_free_rec
) >=
3915 le16_to_cpu(el
->l_count
),
3916 "owner %llu, depth %u, count %u, next free %u, "
3917 "rec.cpos %u, rec.clusters %u, "
3918 "insert.cpos %u, insert.clusters %u\n",
3919 ocfs2_metadata_cache_owner(et
->et_ci
),
3920 le16_to_cpu(el
->l_tree_depth
),
3921 le16_to_cpu(el
->l_count
),
3922 le16_to_cpu(el
->l_next_free_rec
),
3923 le32_to_cpu(el
->l_recs
[i
].e_cpos
),
3924 le16_to_cpu(el
->l_recs
[i
].e_leaf_clusters
),
3925 le32_to_cpu(insert_rec
->e_cpos
),
3926 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3928 el
->l_recs
[i
] = *insert_rec
;
3929 le16_add_cpu(&el
->l_next_free_rec
, 1);
3935 * Ok, we have to rotate.
3937 * At this point, it is safe to assume that inserting into an
3938 * empty leaf and appending to a leaf have both been handled
3941 * This leaf needs to have space, either by the empty 1st
3942 * extent record, or by virtue of an l_next_free_rec < l_count.
3944 ocfs2_rotate_leaf(el
, insert_rec
);
3947 static void ocfs2_adjust_rightmost_records(handle_t
*handle
,
3948 struct ocfs2_extent_tree
*et
,
3949 struct ocfs2_path
*path
,
3950 struct ocfs2_extent_rec
*insert_rec
)
3953 struct buffer_head
*bh
;
3954 struct ocfs2_extent_list
*el
;
3955 struct ocfs2_extent_rec
*rec
;
3958 * Update everything except the leaf block.
3960 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
3961 bh
= path
->p_node
[i
].bh
;
3962 el
= path
->p_node
[i
].el
;
3964 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3965 if (next_free
== 0) {
3966 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
3967 "Owner %llu has a bad extent list\n",
3968 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
3972 rec
= &el
->l_recs
[next_free
- 1];
3974 rec
->e_int_clusters
= insert_rec
->e_cpos
;
3975 le32_add_cpu(&rec
->e_int_clusters
,
3976 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3977 le32_add_cpu(&rec
->e_int_clusters
,
3978 -le32_to_cpu(rec
->e_cpos
));
3980 ocfs2_journal_dirty(handle
, bh
);
3984 static int ocfs2_append_rec_to_path(handle_t
*handle
,
3985 struct ocfs2_extent_tree
*et
,
3986 struct ocfs2_extent_rec
*insert_rec
,
3987 struct ocfs2_path
*right_path
,
3988 struct ocfs2_path
**ret_left_path
)
3991 struct ocfs2_extent_list
*el
;
3992 struct ocfs2_path
*left_path
= NULL
;
3994 *ret_left_path
= NULL
;
3997 * This shouldn't happen for non-trees. The extent rec cluster
3998 * count manipulation below only works for interior nodes.
4000 BUG_ON(right_path
->p_tree_depth
== 0);
4003 * If our appending insert is at the leftmost edge of a leaf,
4004 * then we might need to update the rightmost records of the
4007 el
= path_leaf_el(right_path
);
4008 next_free
= le16_to_cpu(el
->l_next_free_rec
);
4009 if (next_free
== 0 ||
4010 (next_free
== 1 && ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
4013 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
4014 right_path
, &left_cpos
);
4020 trace_ocfs2_append_rec_to_path(
4021 (unsigned long long)
4022 ocfs2_metadata_cache_owner(et
->et_ci
),
4023 le32_to_cpu(insert_rec
->e_cpos
),
4027 * No need to worry if the append is already in the
4031 left_path
= ocfs2_new_path_from_path(right_path
);
4038 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
4046 * ocfs2_insert_path() will pass the left_path to the
4052 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4058 ocfs2_adjust_rightmost_records(handle
, et
, right_path
, insert_rec
);
4060 *ret_left_path
= left_path
;
4064 ocfs2_free_path(left_path
);
4069 static void ocfs2_split_record(struct ocfs2_extent_tree
*et
,
4070 struct ocfs2_path
*left_path
,
4071 struct ocfs2_path
*right_path
,
4072 struct ocfs2_extent_rec
*split_rec
,
4073 enum ocfs2_split_type split
)
4076 u32 cpos
= le32_to_cpu(split_rec
->e_cpos
);
4077 struct ocfs2_extent_list
*left_el
= NULL
, *right_el
, *insert_el
, *el
;
4078 struct ocfs2_extent_rec
*rec
, *tmprec
;
4080 right_el
= path_leaf_el(right_path
);
4082 left_el
= path_leaf_el(left_path
);
4085 insert_el
= right_el
;
4086 index
= ocfs2_search_extent_list(el
, cpos
);
4088 if (index
== 0 && left_path
) {
4089 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
4092 * This typically means that the record
4093 * started in the left path but moved to the
4094 * right as a result of rotation. We either
4095 * move the existing record to the left, or we
4096 * do the later insert there.
4098 * In this case, the left path should always
4099 * exist as the rotate code will have passed
4100 * it back for a post-insert update.
4103 if (split
== SPLIT_LEFT
) {
4105 * It's a left split. Since we know
4106 * that the rotate code gave us an
4107 * empty extent in the left path, we
4108 * can just do the insert there.
4110 insert_el
= left_el
;
4113 * Right split - we have to move the
4114 * existing record over to the left
4115 * leaf. The insert will be into the
4116 * newly created empty extent in the
4119 tmprec
= &right_el
->l_recs
[index
];
4120 ocfs2_rotate_leaf(left_el
, tmprec
);
4123 memset(tmprec
, 0, sizeof(*tmprec
));
4124 index
= ocfs2_search_extent_list(left_el
, cpos
);
4125 BUG_ON(index
== -1);
4130 BUG_ON(!ocfs2_is_empty_extent(&left_el
->l_recs
[0]));
4132 * Left path is easy - we can just allow the insert to
4136 insert_el
= left_el
;
4137 index
= ocfs2_search_extent_list(el
, cpos
);
4138 BUG_ON(index
== -1);
4141 rec
= &el
->l_recs
[index
];
4142 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
4143 split
, rec
, split_rec
);
4144 ocfs2_rotate_leaf(insert_el
, split_rec
);
4148 * This function only does inserts on an allocation b-tree. For tree
4149 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4151 * right_path is the path we want to do the actual insert
4152 * in. left_path should only be passed in if we need to update that
4153 * portion of the tree after an edge insert.
4155 static int ocfs2_insert_path(handle_t
*handle
,
4156 struct ocfs2_extent_tree
*et
,
4157 struct ocfs2_path
*left_path
,
4158 struct ocfs2_path
*right_path
,
4159 struct ocfs2_extent_rec
*insert_rec
,
4160 struct ocfs2_insert_type
*insert
)
4162 int ret
, subtree_index
;
4163 struct buffer_head
*leaf_bh
= path_leaf_bh(right_path
);
4167 * There's a chance that left_path got passed back to
4168 * us without being accounted for in the
4169 * journal. Extend our transaction here to be sure we
4170 * can change those blocks.
4172 ret
= ocfs2_extend_trans(handle
, left_path
->p_tree_depth
);
4178 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
4186 * Pass both paths to the journal. The majority of inserts
4187 * will be touching all components anyway.
4189 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4195 if (insert
->ins_split
!= SPLIT_NONE
) {
4197 * We could call ocfs2_insert_at_leaf() for some types
4198 * of splits, but it's easier to just let one separate
4199 * function sort it all out.
4201 ocfs2_split_record(et
, left_path
, right_path
,
4202 insert_rec
, insert
->ins_split
);
4205 * Split might have modified either leaf and we don't
4206 * have a guarantee that the later edge insert will
4207 * dirty this for us.
4210 ocfs2_journal_dirty(handle
,
4211 path_leaf_bh(left_path
));
4213 ocfs2_insert_at_leaf(et
, insert_rec
, path_leaf_el(right_path
),
4216 ocfs2_journal_dirty(handle
, leaf_bh
);
4220 * The rotate code has indicated that we need to fix
4221 * up portions of the tree after the insert.
4223 * XXX: Should we extend the transaction here?
4225 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
4227 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
4236 static int ocfs2_do_insert_extent(handle_t
*handle
,
4237 struct ocfs2_extent_tree
*et
,
4238 struct ocfs2_extent_rec
*insert_rec
,
4239 struct ocfs2_insert_type
*type
)
4241 int ret
, rotate
= 0;
4243 struct ocfs2_path
*right_path
= NULL
;
4244 struct ocfs2_path
*left_path
= NULL
;
4245 struct ocfs2_extent_list
*el
;
4247 el
= et
->et_root_el
;
4249 ret
= ocfs2_et_root_journal_access(handle
, et
,
4250 OCFS2_JOURNAL_ACCESS_WRITE
);
4256 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
4257 ocfs2_insert_at_leaf(et
, insert_rec
, el
, type
);
4258 goto out_update_clusters
;
4261 right_path
= ocfs2_new_path_from_et(et
);
4269 * Determine the path to start with. Rotations need the
4270 * rightmost path, everything else can go directly to the
4273 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4274 if (type
->ins_appending
== APPEND_NONE
&&
4275 type
->ins_contig
== CONTIG_NONE
) {
4280 ret
= ocfs2_find_path(et
->et_ci
, right_path
, cpos
);
4287 * Rotations and appends need special treatment - they modify
4288 * parts of the tree's above them.
4290 * Both might pass back a path immediate to the left of the
4291 * one being inserted to. This will be cause
4292 * ocfs2_insert_path() to modify the rightmost records of
4293 * left_path to account for an edge insert.
4295 * XXX: When modifying this code, keep in mind that an insert
4296 * can wind up skipping both of these two special cases...
4299 ret
= ocfs2_rotate_tree_right(handle
, et
, type
->ins_split
,
4300 le32_to_cpu(insert_rec
->e_cpos
),
4301 right_path
, &left_path
);
4308 * ocfs2_rotate_tree_right() might have extended the
4309 * transaction without re-journaling our tree root.
4311 ret
= ocfs2_et_root_journal_access(handle
, et
,
4312 OCFS2_JOURNAL_ACCESS_WRITE
);
4317 } else if (type
->ins_appending
== APPEND_TAIL
4318 && type
->ins_contig
!= CONTIG_LEFT
) {
4319 ret
= ocfs2_append_rec_to_path(handle
, et
, insert_rec
,
4320 right_path
, &left_path
);
4327 ret
= ocfs2_insert_path(handle
, et
, left_path
, right_path
,
4334 out_update_clusters
:
4335 if (type
->ins_split
== SPLIT_NONE
)
4336 ocfs2_et_update_clusters(et
,
4337 le16_to_cpu(insert_rec
->e_leaf_clusters
));
4339 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4342 ocfs2_free_path(left_path
);
4343 ocfs2_free_path(right_path
);
4348 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree
*et
,
4349 struct ocfs2_path
*path
,
4350 struct ocfs2_extent_list
*el
, int index
,
4351 struct ocfs2_extent_rec
*split_rec
,
4352 struct ocfs2_merge_ctxt
*ctxt
)
4355 enum ocfs2_contig_type ret
= CONTIG_NONE
;
4356 u32 left_cpos
, right_cpos
;
4357 struct ocfs2_extent_rec
*rec
= NULL
;
4358 struct ocfs2_extent_list
*new_el
;
4359 struct ocfs2_path
*left_path
= NULL
, *right_path
= NULL
;
4360 struct buffer_head
*bh
;
4361 struct ocfs2_extent_block
*eb
;
4362 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
4365 rec
= &el
->l_recs
[index
- 1];
4366 } else if (path
->p_tree_depth
> 0) {
4367 status
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
4371 if (left_cpos
!= 0) {
4372 left_path
= ocfs2_new_path_from_path(path
);
4379 status
= ocfs2_find_path(et
->et_ci
, left_path
,
4382 goto free_left_path
;
4384 new_el
= path_leaf_el(left_path
);
4386 if (le16_to_cpu(new_el
->l_next_free_rec
) !=
4387 le16_to_cpu(new_el
->l_count
)) {
4388 bh
= path_leaf_bh(left_path
);
4389 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4390 status
= ocfs2_error(sb
,
4391 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4392 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4393 le16_to_cpu(new_el
->l_next_free_rec
),
4394 le16_to_cpu(new_el
->l_count
));
4395 goto free_left_path
;
4397 rec
= &new_el
->l_recs
[
4398 le16_to_cpu(new_el
->l_next_free_rec
) - 1];
4403 * We're careful to check for an empty extent record here -
4404 * the merge code will know what to do if it sees one.
4407 if (index
== 1 && ocfs2_is_empty_extent(rec
)) {
4408 if (split_rec
->e_cpos
== el
->l_recs
[index
].e_cpos
)
4411 ret
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4416 if (index
< (le16_to_cpu(el
->l_next_free_rec
) - 1))
4417 rec
= &el
->l_recs
[index
+ 1];
4418 else if (le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
) &&
4419 path
->p_tree_depth
> 0) {
4420 status
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
4422 goto free_left_path
;
4424 if (right_cpos
== 0)
4425 goto free_left_path
;
4427 right_path
= ocfs2_new_path_from_path(path
);
4431 goto free_left_path
;
4434 status
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
4436 goto free_right_path
;
4438 new_el
= path_leaf_el(right_path
);
4439 rec
= &new_el
->l_recs
[0];
4440 if (ocfs2_is_empty_extent(rec
)) {
4441 if (le16_to_cpu(new_el
->l_next_free_rec
) <= 1) {
4442 bh
= path_leaf_bh(right_path
);
4443 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4444 status
= ocfs2_error(sb
,
4445 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4446 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4447 le16_to_cpu(new_el
->l_next_free_rec
));
4448 goto free_right_path
;
4450 rec
= &new_el
->l_recs
[1];
4455 enum ocfs2_contig_type contig_type
;
4457 contig_type
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4459 if (contig_type
== CONTIG_LEFT
&& ret
== CONTIG_RIGHT
)
4460 ret
= CONTIG_LEFTRIGHT
;
4461 else if (ret
== CONTIG_NONE
)
4466 ocfs2_free_path(right_path
);
4468 ocfs2_free_path(left_path
);
4471 ctxt
->c_contig_type
= ret
;
4476 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree
*et
,
4477 struct ocfs2_insert_type
*insert
,
4478 struct ocfs2_extent_list
*el
,
4479 struct ocfs2_extent_rec
*insert_rec
)
4482 enum ocfs2_contig_type contig_type
= CONTIG_NONE
;
4484 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4486 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
4487 contig_type
= ocfs2_et_extent_contig(et
, &el
->l_recs
[i
],
4489 if (contig_type
!= CONTIG_NONE
) {
4490 insert
->ins_contig_index
= i
;
4494 insert
->ins_contig
= contig_type
;
4496 if (insert
->ins_contig
!= CONTIG_NONE
) {
4497 struct ocfs2_extent_rec
*rec
=
4498 &el
->l_recs
[insert
->ins_contig_index
];
4499 unsigned int len
= le16_to_cpu(rec
->e_leaf_clusters
) +
4500 le16_to_cpu(insert_rec
->e_leaf_clusters
);
4503 * Caller might want us to limit the size of extents, don't
4504 * calculate contiguousness if we might exceed that limit.
4506 if (et
->et_max_leaf_clusters
&&
4507 (len
> et
->et_max_leaf_clusters
))
4508 insert
->ins_contig
= CONTIG_NONE
;
4513 * This should only be called against the righmost leaf extent list.
4515 * ocfs2_figure_appending_type() will figure out whether we'll have to
4516 * insert at the tail of the rightmost leaf.
4518 * This should also work against the root extent list for tree's with 0
4519 * depth. If we consider the root extent list to be the rightmost leaf node
4520 * then the logic here makes sense.
4522 static void ocfs2_figure_appending_type(struct ocfs2_insert_type
*insert
,
4523 struct ocfs2_extent_list
*el
,
4524 struct ocfs2_extent_rec
*insert_rec
)
4527 u32 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4528 struct ocfs2_extent_rec
*rec
;
4530 insert
->ins_appending
= APPEND_NONE
;
4532 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4534 if (!el
->l_next_free_rec
)
4535 goto set_tail_append
;
4537 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
4538 /* Were all records empty? */
4539 if (le16_to_cpu(el
->l_next_free_rec
) == 1)
4540 goto set_tail_append
;
4543 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
4544 rec
= &el
->l_recs
[i
];
4547 (le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)))
4548 goto set_tail_append
;
4553 insert
->ins_appending
= APPEND_TAIL
;
4557 * Helper function called at the beginning of an insert.
4559 * This computes a few things that are commonly used in the process of
4560 * inserting into the btree:
4561 * - Whether the new extent is contiguous with an existing one.
4562 * - The current tree depth.
4563 * - Whether the insert is an appending one.
4564 * - The total # of free records in the tree.
4566 * All of the information is stored on the ocfs2_insert_type
4569 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree
*et
,
4570 struct buffer_head
**last_eb_bh
,
4571 struct ocfs2_extent_rec
*insert_rec
,
4573 struct ocfs2_insert_type
*insert
)
4576 struct ocfs2_extent_block
*eb
;
4577 struct ocfs2_extent_list
*el
;
4578 struct ocfs2_path
*path
= NULL
;
4579 struct buffer_head
*bh
= NULL
;
4581 insert
->ins_split
= SPLIT_NONE
;
4583 el
= et
->et_root_el
;
4584 insert
->ins_tree_depth
= le16_to_cpu(el
->l_tree_depth
);
4586 if (el
->l_tree_depth
) {
4588 * If we have tree depth, we read in the
4589 * rightmost extent block ahead of time as
4590 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4591 * may want it later.
4593 ret
= ocfs2_read_extent_block(et
->et_ci
,
4594 ocfs2_et_get_last_eb_blk(et
),
4600 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
4605 * Unless we have a contiguous insert, we'll need to know if
4606 * there is room left in our allocation tree for another
4609 * XXX: This test is simplistic, we can search for empty
4610 * extent records too.
4612 *free_records
= le16_to_cpu(el
->l_count
) -
4613 le16_to_cpu(el
->l_next_free_rec
);
4615 if (!insert
->ins_tree_depth
) {
4616 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4617 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4621 path
= ocfs2_new_path_from_et(et
);
4629 * In the case that we're inserting past what the tree
4630 * currently accounts for, ocfs2_find_path() will return for
4631 * us the rightmost tree path. This is accounted for below in
4632 * the appending code.
4634 ret
= ocfs2_find_path(et
->et_ci
, path
, le32_to_cpu(insert_rec
->e_cpos
));
4640 el
= path_leaf_el(path
);
4643 * Now that we have the path, there's two things we want to determine:
4644 * 1) Contiguousness (also set contig_index if this is so)
4646 * 2) Are we doing an append? We can trivially break this up
4647 * into two types of appends: simple record append, or a
4648 * rotate inside the tail leaf.
4650 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4653 * The insert code isn't quite ready to deal with all cases of
4654 * left contiguousness. Specifically, if it's an insert into
4655 * the 1st record in a leaf, it will require the adjustment of
4656 * cluster count on the last record of the path directly to it's
4657 * left. For now, just catch that case and fool the layers
4658 * above us. This works just fine for tree_depth == 0, which
4659 * is why we allow that above.
4661 if (insert
->ins_contig
== CONTIG_LEFT
&&
4662 insert
->ins_contig_index
== 0)
4663 insert
->ins_contig
= CONTIG_NONE
;
4666 * Ok, so we can simply compare against last_eb to figure out
4667 * whether the path doesn't exist. This will only happen in
4668 * the case that we're doing a tail append, so maybe we can
4669 * take advantage of that information somehow.
4671 if (ocfs2_et_get_last_eb_blk(et
) ==
4672 path_leaf_bh(path
)->b_blocknr
) {
4674 * Ok, ocfs2_find_path() returned us the rightmost
4675 * tree path. This might be an appending insert. There are
4677 * 1) We're doing a true append at the tail:
4678 * -This might even be off the end of the leaf
4679 * 2) We're "appending" by rotating in the tail
4681 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4685 ocfs2_free_path(path
);
4695 * Insert an extent into a btree.
4697 * The caller needs to update the owning btree's cluster count.
4699 int ocfs2_insert_extent(handle_t
*handle
,
4700 struct ocfs2_extent_tree
*et
,
4705 struct ocfs2_alloc_context
*meta_ac
)
4709 struct buffer_head
*last_eb_bh
= NULL
;
4710 struct ocfs2_insert_type insert
= {0, };
4711 struct ocfs2_extent_rec rec
;
4713 trace_ocfs2_insert_extent_start(
4714 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
4715 cpos
, new_clusters
);
4717 memset(&rec
, 0, sizeof(rec
));
4718 rec
.e_cpos
= cpu_to_le32(cpos
);
4719 rec
.e_blkno
= cpu_to_le64(start_blk
);
4720 rec
.e_leaf_clusters
= cpu_to_le16(new_clusters
);
4721 rec
.e_flags
= flags
;
4722 status
= ocfs2_et_insert_check(et
, &rec
);
4728 status
= ocfs2_figure_insert_type(et
, &last_eb_bh
, &rec
,
4729 &free_records
, &insert
);
4735 trace_ocfs2_insert_extent(insert
.ins_appending
, insert
.ins_contig
,
4736 insert
.ins_contig_index
, free_records
,
4737 insert
.ins_tree_depth
);
4739 if (insert
.ins_contig
== CONTIG_NONE
&& free_records
== 0) {
4740 status
= ocfs2_grow_tree(handle
, et
,
4741 &insert
.ins_tree_depth
, &last_eb_bh
,
4749 /* Finally, we can add clusters. This might rotate the tree for us. */
4750 status
= ocfs2_do_insert_extent(handle
, et
, &rec
, &insert
);
4754 ocfs2_et_extent_map_insert(et
, &rec
);
4763 * Allcate and add clusters into the extent b-tree.
4764 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4765 * The extent b-tree's root is specified by et, and
4766 * it is not limited to the file storage. Any extent tree can use this
4767 * function if it implements the proper ocfs2_extent_tree.
4769 int ocfs2_add_clusters_in_btree(handle_t
*handle
,
4770 struct ocfs2_extent_tree
*et
,
4771 u32
*logical_offset
,
4772 u32 clusters_to_add
,
4774 struct ocfs2_alloc_context
*data_ac
,
4775 struct ocfs2_alloc_context
*meta_ac
,
4776 enum ocfs2_alloc_restarted
*reason_ret
)
4778 int status
= 0, err
= 0;
4781 enum ocfs2_alloc_restarted reason
= RESTART_NONE
;
4782 u32 bit_off
, num_bits
;
4785 struct ocfs2_super
*osb
=
4786 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
4788 BUG_ON(!clusters_to_add
);
4791 flags
= OCFS2_EXT_UNWRITTEN
;
4793 free_extents
= ocfs2_num_free_extents(et
);
4794 if (free_extents
< 0) {
4795 status
= free_extents
;
4800 /* there are two cases which could cause us to EAGAIN in the
4801 * we-need-more-metadata case:
4802 * 1) we haven't reserved *any*
4803 * 2) we are so fragmented, we've needed to add metadata too
4805 if (!free_extents
&& !meta_ac
) {
4808 reason
= RESTART_META
;
4810 } else if ((!free_extents
)
4811 && (ocfs2_alloc_context_bits_left(meta_ac
)
4812 < ocfs2_extend_meta_needed(et
->et_root_el
))) {
4815 reason
= RESTART_META
;
4819 status
= __ocfs2_claim_clusters(handle
, data_ac
, 1,
4820 clusters_to_add
, &bit_off
, &num_bits
);
4822 if (status
!= -ENOSPC
)
4827 BUG_ON(num_bits
> clusters_to_add
);
4829 /* reserve our write early -- insert_extent may update the tree root */
4830 status
= ocfs2_et_root_journal_access(handle
, et
,
4831 OCFS2_JOURNAL_ACCESS_WRITE
);
4838 block
= ocfs2_clusters_to_blocks(osb
->sb
, bit_off
);
4839 trace_ocfs2_add_clusters_in_btree(
4840 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
4842 status
= ocfs2_insert_extent(handle
, et
, *logical_offset
, block
,
4843 num_bits
, flags
, meta_ac
);
4850 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4852 clusters_to_add
-= num_bits
;
4853 *logical_offset
+= num_bits
;
4855 if (clusters_to_add
) {
4856 err
= clusters_to_add
;
4858 reason
= RESTART_TRANS
;
4863 if (data_ac
->ac_which
== OCFS2_AC_USE_LOCAL
)
4864 ocfs2_free_local_alloc_bits(osb
, handle
, data_ac
,
4867 ocfs2_free_clusters(handle
,
4870 ocfs2_clusters_to_blocks(osb
->sb
, bit_off
),
4876 *reason_ret
= reason
;
4877 trace_ocfs2_add_clusters_in_btree_ret(status
, reason
, err
);
4881 static void ocfs2_make_right_split_rec(struct super_block
*sb
,
4882 struct ocfs2_extent_rec
*split_rec
,
4884 struct ocfs2_extent_rec
*rec
)
4886 u32 rec_cpos
= le32_to_cpu(rec
->e_cpos
);
4887 u32 rec_range
= rec_cpos
+ le16_to_cpu(rec
->e_leaf_clusters
);
4889 memset(split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4891 split_rec
->e_cpos
= cpu_to_le32(cpos
);
4892 split_rec
->e_leaf_clusters
= cpu_to_le16(rec_range
- cpos
);
4894 split_rec
->e_blkno
= rec
->e_blkno
;
4895 le64_add_cpu(&split_rec
->e_blkno
,
4896 ocfs2_clusters_to_blocks(sb
, cpos
- rec_cpos
));
4898 split_rec
->e_flags
= rec
->e_flags
;
4901 static int ocfs2_split_and_insert(handle_t
*handle
,
4902 struct ocfs2_extent_tree
*et
,
4903 struct ocfs2_path
*path
,
4904 struct buffer_head
**last_eb_bh
,
4906 struct ocfs2_extent_rec
*orig_split_rec
,
4907 struct ocfs2_alloc_context
*meta_ac
)
4910 unsigned int insert_range
, rec_range
, do_leftright
= 0;
4911 struct ocfs2_extent_rec tmprec
;
4912 struct ocfs2_extent_list
*rightmost_el
;
4913 struct ocfs2_extent_rec rec
;
4914 struct ocfs2_extent_rec split_rec
= *orig_split_rec
;
4915 struct ocfs2_insert_type insert
;
4916 struct ocfs2_extent_block
*eb
;
4920 * Store a copy of the record on the stack - it might move
4921 * around as the tree is manipulated below.
4923 rec
= path_leaf_el(path
)->l_recs
[split_index
];
4925 rightmost_el
= et
->et_root_el
;
4927 depth
= le16_to_cpu(rightmost_el
->l_tree_depth
);
4929 BUG_ON(!(*last_eb_bh
));
4930 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
4931 rightmost_el
= &eb
->h_list
;
4934 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4935 le16_to_cpu(rightmost_el
->l_count
)) {
4936 ret
= ocfs2_grow_tree(handle
, et
,
4937 &depth
, last_eb_bh
, meta_ac
);
4944 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4945 insert
.ins_appending
= APPEND_NONE
;
4946 insert
.ins_contig
= CONTIG_NONE
;
4947 insert
.ins_tree_depth
= depth
;
4949 insert_range
= le32_to_cpu(split_rec
.e_cpos
) +
4950 le16_to_cpu(split_rec
.e_leaf_clusters
);
4951 rec_range
= le32_to_cpu(rec
.e_cpos
) +
4952 le16_to_cpu(rec
.e_leaf_clusters
);
4954 if (split_rec
.e_cpos
== rec
.e_cpos
) {
4955 insert
.ins_split
= SPLIT_LEFT
;
4956 } else if (insert_range
== rec_range
) {
4957 insert
.ins_split
= SPLIT_RIGHT
;
4960 * Left/right split. We fake this as a right split
4961 * first and then make a second pass as a left split.
4963 insert
.ins_split
= SPLIT_RIGHT
;
4965 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
4966 &tmprec
, insert_range
, &rec
);
4970 BUG_ON(do_leftright
);
4974 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
4980 if (do_leftright
== 1) {
4982 struct ocfs2_extent_list
*el
;
4985 split_rec
= *orig_split_rec
;
4987 ocfs2_reinit_path(path
, 1);
4989 cpos
= le32_to_cpu(split_rec
.e_cpos
);
4990 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
4996 el
= path_leaf_el(path
);
4997 split_index
= ocfs2_search_extent_list(el
, cpos
);
4998 if (split_index
== -1) {
4999 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5000 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5001 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5013 static int ocfs2_replace_extent_rec(handle_t
*handle
,
5014 struct ocfs2_extent_tree
*et
,
5015 struct ocfs2_path
*path
,
5016 struct ocfs2_extent_list
*el
,
5018 struct ocfs2_extent_rec
*split_rec
)
5022 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
5023 path_num_items(path
) - 1);
5029 el
->l_recs
[split_index
] = *split_rec
;
5031 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5037 * Split part or all of the extent record at split_index in the leaf
5038 * pointed to by path. Merge with the contiguous extent record if needed.
5040 * Care is taken to handle contiguousness so as to not grow the tree.
5042 * meta_ac is not strictly necessary - we only truly need it if growth
5043 * of the tree is required. All other cases will degrade into a less
5044 * optimal tree layout.
5046 * last_eb_bh should be the rightmost leaf block for any extent
5047 * btree. Since a split may grow the tree or a merge might shrink it,
5048 * the caller cannot trust the contents of that buffer after this call.
5050 * This code is optimized for readability - several passes might be
5051 * made over certain portions of the tree. All of those blocks will
5052 * have been brought into cache (and pinned via the journal), so the
5053 * extra overhead is not expressed in terms of disk reads.
5055 int ocfs2_split_extent(handle_t
*handle
,
5056 struct ocfs2_extent_tree
*et
,
5057 struct ocfs2_path
*path
,
5059 struct ocfs2_extent_rec
*split_rec
,
5060 struct ocfs2_alloc_context
*meta_ac
,
5061 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5064 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5065 struct buffer_head
*last_eb_bh
= NULL
;
5066 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
5067 struct ocfs2_merge_ctxt ctxt
;
5069 if (le32_to_cpu(rec
->e_cpos
) > le32_to_cpu(split_rec
->e_cpos
) ||
5070 ((le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)) <
5071 (le32_to_cpu(split_rec
->e_cpos
) + le16_to_cpu(split_rec
->e_leaf_clusters
)))) {
5077 ret
= ocfs2_figure_merge_contig_type(et
, path
, el
,
5087 * The core merge / split code wants to know how much room is
5088 * left in this allocation tree, so we pass the
5089 * rightmost extent list.
5091 if (path
->p_tree_depth
) {
5092 ret
= ocfs2_read_extent_block(et
->et_ci
,
5093 ocfs2_et_get_last_eb_blk(et
),
5101 if (rec
->e_cpos
== split_rec
->e_cpos
&&
5102 rec
->e_leaf_clusters
== split_rec
->e_leaf_clusters
)
5103 ctxt
.c_split_covers_rec
= 1;
5105 ctxt
.c_split_covers_rec
= 0;
5107 ctxt
.c_has_empty_extent
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
5109 trace_ocfs2_split_extent(split_index
, ctxt
.c_contig_type
,
5110 ctxt
.c_has_empty_extent
,
5111 ctxt
.c_split_covers_rec
);
5113 if (ctxt
.c_contig_type
== CONTIG_NONE
) {
5114 if (ctxt
.c_split_covers_rec
)
5115 ret
= ocfs2_replace_extent_rec(handle
, et
, path
, el
,
5116 split_index
, split_rec
);
5118 ret
= ocfs2_split_and_insert(handle
, et
, path
,
5119 &last_eb_bh
, split_index
,
5120 split_rec
, meta_ac
);
5124 ret
= ocfs2_try_to_merge_extent(handle
, et
, path
,
5125 split_index
, split_rec
,
5137 * Change the flags of the already-existing extent at cpos for len clusters.
5139 * new_flags: the flags we want to set.
5140 * clear_flags: the flags we want to clear.
5141 * phys: the new physical offset we want this new extent starts from.
5143 * If the existing extent is larger than the request, initiate a
5144 * split. An attempt will be made at merging with adjacent extents.
5146 * The caller is responsible for passing down meta_ac if we'll need it.
5148 int ocfs2_change_extent_flag(handle_t
*handle
,
5149 struct ocfs2_extent_tree
*et
,
5150 u32 cpos
, u32 len
, u32 phys
,
5151 struct ocfs2_alloc_context
*meta_ac
,
5152 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5153 int new_flags
, int clear_flags
)
5156 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5157 u64 start_blkno
= ocfs2_clusters_to_blocks(sb
, phys
);
5158 struct ocfs2_extent_rec split_rec
;
5159 struct ocfs2_path
*left_path
= NULL
;
5160 struct ocfs2_extent_list
*el
;
5161 struct ocfs2_extent_rec
*rec
;
5163 left_path
= ocfs2_new_path_from_et(et
);
5170 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
5175 el
= path_leaf_el(left_path
);
5177 index
= ocfs2_search_extent_list(el
, cpos
);
5180 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5181 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5188 rec
= &el
->l_recs
[index
];
5189 if (new_flags
&& (rec
->e_flags
& new_flags
)) {
5190 mlog(ML_ERROR
, "Owner %llu tried to set %d flags on an "
5191 "extent that already had them\n",
5192 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5197 if (clear_flags
&& !(rec
->e_flags
& clear_flags
)) {
5198 mlog(ML_ERROR
, "Owner %llu tried to clear %d flags on an "
5199 "extent that didn't have them\n",
5200 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5205 memset(&split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
5206 split_rec
.e_cpos
= cpu_to_le32(cpos
);
5207 split_rec
.e_leaf_clusters
= cpu_to_le16(len
);
5208 split_rec
.e_blkno
= cpu_to_le64(start_blkno
);
5209 split_rec
.e_flags
= rec
->e_flags
;
5211 split_rec
.e_flags
|= new_flags
;
5213 split_rec
.e_flags
&= ~clear_flags
;
5215 ret
= ocfs2_split_extent(handle
, et
, left_path
,
5216 index
, &split_rec
, meta_ac
,
5222 ocfs2_free_path(left_path
);
5228 * Mark the already-existing extent at cpos as written for len clusters.
5229 * This removes the unwritten extent flag.
5231 * If the existing extent is larger than the request, initiate a
5232 * split. An attempt will be made at merging with adjacent extents.
5234 * The caller is responsible for passing down meta_ac if we'll need it.
5236 int ocfs2_mark_extent_written(struct inode
*inode
,
5237 struct ocfs2_extent_tree
*et
,
5238 handle_t
*handle
, u32 cpos
, u32 len
, u32 phys
,
5239 struct ocfs2_alloc_context
*meta_ac
,
5240 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5244 trace_ocfs2_mark_extent_written(
5245 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
5248 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode
->i_sb
))) {
5249 ocfs2_error(inode
->i_sb
, "Inode %llu has unwritten extents that are being written to, but the feature bit is not set in the super block\n",
5250 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
5256 * XXX: This should be fixed up so that we just re-insert the
5257 * next extent records.
5259 ocfs2_et_extent_map_truncate(et
, 0);
5261 ret
= ocfs2_change_extent_flag(handle
, et
, cpos
,
5262 len
, phys
, meta_ac
, dealloc
,
5263 0, OCFS2_EXT_UNWRITTEN
);
5271 static int ocfs2_split_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
5272 struct ocfs2_path
*path
,
5273 int index
, u32 new_range
,
5274 struct ocfs2_alloc_context
*meta_ac
)
5276 int ret
, depth
, credits
;
5277 struct buffer_head
*last_eb_bh
= NULL
;
5278 struct ocfs2_extent_block
*eb
;
5279 struct ocfs2_extent_list
*rightmost_el
, *el
;
5280 struct ocfs2_extent_rec split_rec
;
5281 struct ocfs2_extent_rec
*rec
;
5282 struct ocfs2_insert_type insert
;
5285 * Setup the record to split before we grow the tree.
5287 el
= path_leaf_el(path
);
5288 rec
= &el
->l_recs
[index
];
5289 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
5290 &split_rec
, new_range
, rec
);
5292 depth
= path
->p_tree_depth
;
5294 ret
= ocfs2_read_extent_block(et
->et_ci
,
5295 ocfs2_et_get_last_eb_blk(et
),
5302 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5303 rightmost_el
= &eb
->h_list
;
5305 rightmost_el
= path_leaf_el(path
);
5307 credits
= path
->p_tree_depth
+
5308 ocfs2_extend_meta_needed(et
->et_root_el
);
5309 ret
= ocfs2_extend_trans(handle
, credits
);
5315 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
5316 le16_to_cpu(rightmost_el
->l_count
)) {
5317 ret
= ocfs2_grow_tree(handle
, et
, &depth
, &last_eb_bh
,
5325 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
5326 insert
.ins_appending
= APPEND_NONE
;
5327 insert
.ins_contig
= CONTIG_NONE
;
5328 insert
.ins_split
= SPLIT_RIGHT
;
5329 insert
.ins_tree_depth
= depth
;
5331 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
5340 static int ocfs2_truncate_rec(handle_t
*handle
,
5341 struct ocfs2_extent_tree
*et
,
5342 struct ocfs2_path
*path
, int index
,
5343 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5347 u32 left_cpos
, rec_range
, trunc_range
;
5348 int is_rightmost_tree_rec
= 0;
5349 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5350 struct ocfs2_path
*left_path
= NULL
;
5351 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5352 struct ocfs2_extent_rec
*rec
;
5353 struct ocfs2_extent_block
*eb
;
5355 if (ocfs2_is_empty_extent(&el
->l_recs
[0]) && index
> 0) {
5356 /* extend credit for ocfs2_remove_rightmost_path */
5357 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
5358 jbd2_handle_buffer_credits(handle
),
5365 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5374 if (index
== (le16_to_cpu(el
->l_next_free_rec
) - 1) &&
5375 path
->p_tree_depth
) {
5377 * Check whether this is the rightmost tree record. If
5378 * we remove all of this record or part of its right
5379 * edge then an update of the record lengths above it
5382 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
5383 if (eb
->h_next_leaf_blk
== 0)
5384 is_rightmost_tree_rec
= 1;
5387 rec
= &el
->l_recs
[index
];
5388 if (index
== 0 && path
->p_tree_depth
&&
5389 le32_to_cpu(rec
->e_cpos
) == cpos
) {
5391 * Changing the leftmost offset (via partial or whole
5392 * record truncate) of an interior (or rightmost) path
5393 * means we have to update the subtree that is formed
5394 * by this leaf and the one to it's left.
5396 * There are two cases we can skip:
5397 * 1) Path is the leftmost one in our btree.
5398 * 2) The leaf is rightmost and will be empty after
5399 * we remove the extent record - the rotate code
5400 * knows how to update the newly formed edge.
5403 ret
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
5409 if (left_cpos
&& le16_to_cpu(el
->l_next_free_rec
) > 1) {
5410 left_path
= ocfs2_new_path_from_path(path
);
5417 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
5426 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
5427 jbd2_handle_buffer_credits(handle
),
5434 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
5440 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
5446 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5447 trunc_range
= cpos
+ len
;
5449 if (le32_to_cpu(rec
->e_cpos
) == cpos
&& rec_range
== trunc_range
) {
5452 memset(rec
, 0, sizeof(*rec
));
5453 ocfs2_cleanup_merge(el
, index
);
5455 next_free
= le16_to_cpu(el
->l_next_free_rec
);
5456 if (is_rightmost_tree_rec
&& next_free
> 1) {
5458 * We skip the edge update if this path will
5459 * be deleted by the rotate code.
5461 rec
= &el
->l_recs
[next_free
- 1];
5462 ocfs2_adjust_rightmost_records(handle
, et
, path
,
5465 } else if (le32_to_cpu(rec
->e_cpos
) == cpos
) {
5466 /* Remove leftmost portion of the record. */
5467 le32_add_cpu(&rec
->e_cpos
, len
);
5468 le64_add_cpu(&rec
->e_blkno
, ocfs2_clusters_to_blocks(sb
, len
));
5469 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5470 } else if (rec_range
== trunc_range
) {
5471 /* Remove rightmost portion of the record */
5472 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5473 if (is_rightmost_tree_rec
)
5474 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
5476 /* Caller should have trapped this. */
5477 mlog(ML_ERROR
, "Owner %llu: Invalid record truncate: (%u, %u) "
5479 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5480 le32_to_cpu(rec
->e_cpos
),
5481 le16_to_cpu(rec
->e_leaf_clusters
), cpos
, len
);
5488 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
5489 ocfs2_complete_edge_insert(handle
, left_path
, path
,
5493 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5495 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5500 ocfs2_free_path(left_path
);
5504 int ocfs2_remove_extent(handle_t
*handle
,
5505 struct ocfs2_extent_tree
*et
,
5507 struct ocfs2_alloc_context
*meta_ac
,
5508 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5511 u32 rec_range
, trunc_range
;
5512 struct ocfs2_extent_rec
*rec
;
5513 struct ocfs2_extent_list
*el
;
5514 struct ocfs2_path
*path
= NULL
;
5517 * XXX: Why are we truncating to 0 instead of wherever this
5520 ocfs2_et_extent_map_truncate(et
, 0);
5522 path
= ocfs2_new_path_from_et(et
);
5529 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5535 el
= path_leaf_el(path
);
5536 index
= ocfs2_search_extent_list(el
, cpos
);
5538 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5539 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5540 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5547 * We have 3 cases of extent removal:
5548 * 1) Range covers the entire extent rec
5549 * 2) Range begins or ends on one edge of the extent rec
5550 * 3) Range is in the middle of the extent rec (no shared edges)
5552 * For case 1 we remove the extent rec and left rotate to
5555 * For case 2 we just shrink the existing extent rec, with a
5556 * tree update if the shrinking edge is also the edge of an
5559 * For case 3 we do a right split to turn the extent rec into
5560 * something case 2 can handle.
5562 rec
= &el
->l_recs
[index
];
5563 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5564 trunc_range
= cpos
+ len
;
5566 BUG_ON(cpos
< le32_to_cpu(rec
->e_cpos
) || trunc_range
> rec_range
);
5568 trace_ocfs2_remove_extent(
5569 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5570 cpos
, len
, index
, le32_to_cpu(rec
->e_cpos
),
5571 ocfs2_rec_clusters(el
, rec
));
5573 if (le32_to_cpu(rec
->e_cpos
) == cpos
|| rec_range
== trunc_range
) {
5574 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5581 ret
= ocfs2_split_tree(handle
, et
, path
, index
,
5582 trunc_range
, meta_ac
);
5589 * The split could have manipulated the tree enough to
5590 * move the record location, so we have to look for it again.
5592 ocfs2_reinit_path(path
, 1);
5594 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5600 el
= path_leaf_el(path
);
5601 index
= ocfs2_search_extent_list(el
, cpos
);
5603 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5604 "Owner %llu: split at cpos %u lost record\n",
5605 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5612 * Double check our values here. If anything is fishy,
5613 * it's easier to catch it at the top level.
5615 rec
= &el
->l_recs
[index
];
5616 rec_range
= le32_to_cpu(rec
->e_cpos
) +
5617 ocfs2_rec_clusters(el
, rec
);
5618 if (rec_range
!= trunc_range
) {
5619 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5620 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5621 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5622 cpos
, len
, le32_to_cpu(rec
->e_cpos
),
5623 ocfs2_rec_clusters(el
, rec
));
5628 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5635 ocfs2_free_path(path
);
5640 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5641 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5642 * number to reserve some extra blocks, and it only handles meta
5645 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5646 * and punching holes.
5648 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode
*inode
,
5649 struct ocfs2_extent_tree
*et
,
5650 u32 extents_to_split
,
5651 struct ocfs2_alloc_context
**ac
,
5654 int ret
= 0, num_free_extents
;
5655 unsigned int max_recs_needed
= 2 * extents_to_split
;
5656 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
5660 num_free_extents
= ocfs2_num_free_extents(et
);
5661 if (num_free_extents
< 0) {
5662 ret
= num_free_extents
;
5667 if (!num_free_extents
||
5668 (ocfs2_sparse_alloc(osb
) && num_free_extents
< max_recs_needed
))
5669 extra_blocks
+= ocfs2_extend_meta_needed(et
->et_root_el
);
5672 ret
= ocfs2_reserve_new_metadata_blocks(osb
, extra_blocks
, ac
);
5682 ocfs2_free_alloc_context(*ac
);
5690 int ocfs2_remove_btree_range(struct inode
*inode
,
5691 struct ocfs2_extent_tree
*et
,
5692 u32 cpos
, u32 phys_cpos
, u32 len
, int flags
,
5693 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5694 u64 refcount_loc
, bool refcount_tree_locked
)
5696 int ret
, credits
= 0, extra_blocks
= 0;
5697 u64 phys_blkno
= ocfs2_clusters_to_blocks(inode
->i_sb
, phys_cpos
);
5698 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
5699 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5701 struct ocfs2_alloc_context
*meta_ac
= NULL
;
5702 struct ocfs2_refcount_tree
*ref_tree
= NULL
;
5704 if ((flags
& OCFS2_EXT_REFCOUNTED
) && len
) {
5705 BUG_ON(!ocfs2_is_refcount_inode(inode
));
5707 if (!refcount_tree_locked
) {
5708 ret
= ocfs2_lock_refcount_tree(osb
, refcount_loc
, 1,
5716 ret
= ocfs2_prepare_refcount_change_for_del(inode
,
5728 ret
= ocfs2_reserve_blocks_for_rec_trunc(inode
, et
, 1, &meta_ac
,
5735 inode_lock(tl_inode
);
5737 if (ocfs2_truncate_log_needs_flush(osb
)) {
5738 ret
= __ocfs2_flush_truncate_log(osb
);
5745 handle
= ocfs2_start_trans(osb
,
5746 ocfs2_remove_extent_credits(osb
->sb
) + credits
);
5747 if (IS_ERR(handle
)) {
5748 ret
= PTR_ERR(handle
);
5753 ret
= ocfs2_et_root_journal_access(handle
, et
,
5754 OCFS2_JOURNAL_ACCESS_WRITE
);
5760 dquot_free_space_nodirty(inode
,
5761 ocfs2_clusters_to_bytes(inode
->i_sb
, len
));
5763 ret
= ocfs2_remove_extent(handle
, et
, cpos
, len
, meta_ac
, dealloc
);
5769 ocfs2_et_update_clusters(et
, -len
);
5770 ocfs2_update_inode_fsync_trans(handle
, inode
, 1);
5772 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
5775 if (flags
& OCFS2_EXT_REFCOUNTED
)
5776 ret
= ocfs2_decrease_refcount(inode
, handle
,
5777 ocfs2_blocks_to_clusters(osb
->sb
,
5782 ret
= ocfs2_truncate_log_append(osb
, handle
,
5790 ocfs2_commit_trans(osb
, handle
);
5792 inode_unlock(tl_inode
);
5795 ocfs2_free_alloc_context(meta_ac
);
5798 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
5803 int ocfs2_truncate_log_needs_flush(struct ocfs2_super
*osb
)
5805 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5806 struct ocfs2_dinode
*di
;
5807 struct ocfs2_truncate_log
*tl
;
5809 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5810 tl
= &di
->id2
.i_dealloc
;
5812 mlog_bug_on_msg(le16_to_cpu(tl
->tl_used
) > le16_to_cpu(tl
->tl_count
),
5813 "slot %d, invalid truncate log parameters: used = "
5814 "%u, count = %u\n", osb
->slot_num
,
5815 le16_to_cpu(tl
->tl_used
), le16_to_cpu(tl
->tl_count
));
5816 return le16_to_cpu(tl
->tl_used
) == le16_to_cpu(tl
->tl_count
);
5819 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log
*tl
,
5820 unsigned int new_start
)
5822 unsigned int tail_index
;
5823 unsigned int current_tail
;
5825 /* No records, nothing to coalesce */
5826 if (!le16_to_cpu(tl
->tl_used
))
5829 tail_index
= le16_to_cpu(tl
->tl_used
) - 1;
5830 current_tail
= le32_to_cpu(tl
->tl_recs
[tail_index
].t_start
);
5831 current_tail
+= le32_to_cpu(tl
->tl_recs
[tail_index
].t_clusters
);
5833 return current_tail
== new_start
;
5836 int ocfs2_truncate_log_append(struct ocfs2_super
*osb
,
5839 unsigned int num_clusters
)
5842 unsigned int start_cluster
, tl_count
;
5843 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5844 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5845 struct ocfs2_dinode
*di
;
5846 struct ocfs2_truncate_log
*tl
;
5848 BUG_ON(inode_trylock(tl_inode
));
5850 start_cluster
= ocfs2_blocks_to_clusters(osb
->sb
, start_blk
);
5852 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5854 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5855 * by the underlying call to ocfs2_read_inode_block(), so any
5856 * corruption is a code bug */
5857 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
5859 tl
= &di
->id2
.i_dealloc
;
5860 tl_count
= le16_to_cpu(tl
->tl_count
);
5861 mlog_bug_on_msg(tl_count
> ocfs2_truncate_recs_per_inode(osb
->sb
) ||
5863 "Truncate record count on #%llu invalid "
5864 "wanted %u, actual %u\n",
5865 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5866 ocfs2_truncate_recs_per_inode(osb
->sb
),
5867 le16_to_cpu(tl
->tl_count
));
5869 /* Caller should have known to flush before calling us. */
5870 index
= le16_to_cpu(tl
->tl_used
);
5871 if (index
>= tl_count
) {
5877 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5878 OCFS2_JOURNAL_ACCESS_WRITE
);
5884 trace_ocfs2_truncate_log_append(
5885 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
, index
,
5886 start_cluster
, num_clusters
);
5887 if (ocfs2_truncate_log_can_coalesce(tl
, start_cluster
)) {
5889 * Move index back to the record we are coalescing with.
5890 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5894 num_clusters
+= le32_to_cpu(tl
->tl_recs
[index
].t_clusters
);
5895 trace_ocfs2_truncate_log_append(
5896 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5897 index
, le32_to_cpu(tl
->tl_recs
[index
].t_start
),
5900 tl
->tl_recs
[index
].t_start
= cpu_to_le32(start_cluster
);
5901 tl
->tl_used
= cpu_to_le16(index
+ 1);
5903 tl
->tl_recs
[index
].t_clusters
= cpu_to_le32(num_clusters
);
5905 ocfs2_journal_dirty(handle
, tl_bh
);
5907 osb
->truncated_clusters
+= num_clusters
;
5912 static int ocfs2_replay_truncate_records(struct ocfs2_super
*osb
,
5913 struct inode
*data_alloc_inode
,
5914 struct buffer_head
*data_alloc_bh
)
5918 unsigned int num_clusters
;
5920 struct ocfs2_truncate_rec rec
;
5921 struct ocfs2_dinode
*di
;
5922 struct ocfs2_truncate_log
*tl
;
5923 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5924 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5927 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5928 tl
= &di
->id2
.i_dealloc
;
5929 i
= le16_to_cpu(tl
->tl_used
) - 1;
5931 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC
);
5932 if (IS_ERR(handle
)) {
5933 status
= PTR_ERR(handle
);
5938 /* Caller has given us at least enough credits to
5939 * update the truncate log dinode */
5940 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5941 OCFS2_JOURNAL_ACCESS_WRITE
);
5943 ocfs2_commit_trans(osb
, handle
);
5948 tl
->tl_used
= cpu_to_le16(i
);
5950 ocfs2_journal_dirty(handle
, tl_bh
);
5952 rec
= tl
->tl_recs
[i
];
5953 start_blk
= ocfs2_clusters_to_blocks(data_alloc_inode
->i_sb
,
5954 le32_to_cpu(rec
.t_start
));
5955 num_clusters
= le32_to_cpu(rec
.t_clusters
);
5957 /* if start_blk is not set, we ignore the record as
5960 trace_ocfs2_replay_truncate_records(
5961 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5962 i
, le32_to_cpu(rec
.t_start
), num_clusters
);
5964 status
= ocfs2_free_clusters(handle
, data_alloc_inode
,
5965 data_alloc_bh
, start_blk
,
5968 ocfs2_commit_trans(osb
, handle
);
5974 ocfs2_commit_trans(osb
, handle
);
5978 osb
->truncated_clusters
= 0;
5984 /* Expects you to already be holding tl_inode->i_rwsem */
5985 int __ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
5988 unsigned int num_to_flush
;
5989 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5990 struct inode
*data_alloc_inode
= NULL
;
5991 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5992 struct buffer_head
*data_alloc_bh
= NULL
;
5993 struct ocfs2_dinode
*di
;
5994 struct ocfs2_truncate_log
*tl
;
5995 struct ocfs2_journal
*journal
= osb
->journal
;
5997 BUG_ON(inode_trylock(tl_inode
));
5999 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
6001 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
6002 * by the underlying call to ocfs2_read_inode_block(), so any
6003 * corruption is a code bug */
6004 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
6006 tl
= &di
->id2
.i_dealloc
;
6007 num_to_flush
= le16_to_cpu(tl
->tl_used
);
6008 trace_ocfs2_flush_truncate_log(
6009 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
6011 if (!num_to_flush
) {
6016 /* Appending truncate log(TA) and flushing truncate log(TF) are
6017 * two separated transactions. They can be both committed but not
6018 * checkpointed. If crash occurs then, both two transaction will be
6019 * replayed with several already released to global bitmap clusters.
6020 * Then truncate log will be replayed resulting in cluster double free.
6022 jbd2_journal_lock_updates(journal
->j_journal
);
6023 status
= jbd2_journal_flush(journal
->j_journal
, 0);
6024 jbd2_journal_unlock_updates(journal
->j_journal
);
6030 data_alloc_inode
= ocfs2_get_system_file_inode(osb
,
6031 GLOBAL_BITMAP_SYSTEM_INODE
,
6032 OCFS2_INVALID_SLOT
);
6033 if (!data_alloc_inode
) {
6035 mlog(ML_ERROR
, "Could not get bitmap inode!\n");
6039 inode_lock(data_alloc_inode
);
6041 status
= ocfs2_inode_lock(data_alloc_inode
, &data_alloc_bh
, 1);
6047 status
= ocfs2_replay_truncate_records(osb
, data_alloc_inode
,
6052 brelse(data_alloc_bh
);
6053 ocfs2_inode_unlock(data_alloc_inode
, 1);
6056 inode_unlock(data_alloc_inode
);
6057 iput(data_alloc_inode
);
6063 int ocfs2_flush_truncate_log(struct ocfs2_super
*osb
)
6066 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6068 inode_lock(tl_inode
);
6069 status
= __ocfs2_flush_truncate_log(osb
);
6070 inode_unlock(tl_inode
);
6075 static void ocfs2_truncate_log_worker(struct work_struct
*work
)
6078 struct ocfs2_super
*osb
=
6079 container_of(work
, struct ocfs2_super
,
6080 osb_truncate_log_wq
.work
);
6082 status
= ocfs2_flush_truncate_log(osb
);
6086 ocfs2_init_steal_slots(osb
);
6089 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6090 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super
*osb
,
6093 if (osb
->osb_tl_inode
&&
6094 atomic_read(&osb
->osb_tl_disable
) == 0) {
6095 /* We want to push off log flushes while truncates are
6098 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
6100 queue_delayed_work(osb
->ocfs2_wq
, &osb
->osb_truncate_log_wq
,
6101 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL
);
6106 * Try to flush truncate logs if we can free enough clusters from it.
6107 * As for return value, "< 0" means error, "0" no space and "1" means
6108 * we have freed enough spaces and let the caller try to allocate again.
6110 int ocfs2_try_to_free_truncate_log(struct ocfs2_super
*osb
,
6111 unsigned int needed
)
6115 unsigned int truncated_clusters
;
6117 inode_lock(osb
->osb_tl_inode
);
6118 truncated_clusters
= osb
->truncated_clusters
;
6119 inode_unlock(osb
->osb_tl_inode
);
6122 * Check whether we can succeed in allocating if we free
6125 if (truncated_clusters
< needed
)
6128 ret
= ocfs2_flush_truncate_log(osb
);
6134 if (jbd2_journal_start_commit(osb
->journal
->j_journal
, &target
)) {
6135 jbd2_log_wait_commit(osb
->journal
->j_journal
, target
);
6142 static int ocfs2_get_truncate_log_info(struct ocfs2_super
*osb
,
6144 struct inode
**tl_inode
,
6145 struct buffer_head
**tl_bh
)
6148 struct inode
*inode
= NULL
;
6149 struct buffer_head
*bh
= NULL
;
6151 inode
= ocfs2_get_system_file_inode(osb
,
6152 TRUNCATE_LOG_SYSTEM_INODE
,
6156 mlog(ML_ERROR
, "Could not get load truncate log inode!\n");
6160 status
= ocfs2_read_inode_block(inode
, &bh
);
6173 /* called during the 1st stage of node recovery. we stamp a clean
6174 * truncate log and pass back a copy for processing later. if the
6175 * truncate log does not require processing, a *tl_copy is set to
6177 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super
*osb
,
6179 struct ocfs2_dinode
**tl_copy
)
6182 struct inode
*tl_inode
= NULL
;
6183 struct buffer_head
*tl_bh
= NULL
;
6184 struct ocfs2_dinode
*di
;
6185 struct ocfs2_truncate_log
*tl
;
6189 trace_ocfs2_begin_truncate_log_recovery(slot_num
);
6191 status
= ocfs2_get_truncate_log_info(osb
, slot_num
, &tl_inode
, &tl_bh
);
6197 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
6199 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6200 * validated by the underlying call to ocfs2_read_inode_block(),
6201 * so any corruption is a code bug */
6202 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
6204 tl
= &di
->id2
.i_dealloc
;
6205 if (le16_to_cpu(tl
->tl_used
)) {
6206 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl
->tl_used
));
6209 * Assuming the write-out below goes well, this copy will be
6210 * passed back to recovery for processing.
6212 *tl_copy
= kmemdup(tl_bh
->b_data
, tl_bh
->b_size
, GFP_KERNEL
);
6219 /* All we need to do to clear the truncate log is set
6223 ocfs2_compute_meta_ecc(osb
->sb
, tl_bh
->b_data
, &di
->i_check
);
6224 status
= ocfs2_write_block(osb
, tl_bh
, INODE_CACHE(tl_inode
));
6244 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super
*osb
,
6245 struct ocfs2_dinode
*tl_copy
)
6249 unsigned int clusters
, num_recs
, start_cluster
;
6252 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6253 struct ocfs2_truncate_log
*tl
;
6255 if (OCFS2_I(tl_inode
)->ip_blkno
== le64_to_cpu(tl_copy
->i_blkno
)) {
6256 mlog(ML_ERROR
, "Asked to recover my own truncate log!\n");
6260 tl
= &tl_copy
->id2
.i_dealloc
;
6261 num_recs
= le16_to_cpu(tl
->tl_used
);
6262 trace_ocfs2_complete_truncate_log_recovery(
6263 (unsigned long long)le64_to_cpu(tl_copy
->i_blkno
),
6266 inode_lock(tl_inode
);
6267 for(i
= 0; i
< num_recs
; i
++) {
6268 if (ocfs2_truncate_log_needs_flush(osb
)) {
6269 status
= __ocfs2_flush_truncate_log(osb
);
6276 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
6277 if (IS_ERR(handle
)) {
6278 status
= PTR_ERR(handle
);
6283 clusters
= le32_to_cpu(tl
->tl_recs
[i
].t_clusters
);
6284 start_cluster
= le32_to_cpu(tl
->tl_recs
[i
].t_start
);
6285 start_blk
= ocfs2_clusters_to_blocks(osb
->sb
, start_cluster
);
6287 status
= ocfs2_truncate_log_append(osb
, handle
,
6288 start_blk
, clusters
);
6289 ocfs2_commit_trans(osb
, handle
);
6297 inode_unlock(tl_inode
);
6302 void ocfs2_truncate_log_shutdown(struct ocfs2_super
*osb
)
6305 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6307 atomic_set(&osb
->osb_tl_disable
, 1);
6310 cancel_delayed_work(&osb
->osb_truncate_log_wq
);
6311 flush_workqueue(osb
->ocfs2_wq
);
6313 status
= ocfs2_flush_truncate_log(osb
);
6317 brelse(osb
->osb_tl_bh
);
6318 iput(osb
->osb_tl_inode
);
6322 int ocfs2_truncate_log_init(struct ocfs2_super
*osb
)
6325 struct inode
*tl_inode
= NULL
;
6326 struct buffer_head
*tl_bh
= NULL
;
6328 status
= ocfs2_get_truncate_log_info(osb
,
6335 /* ocfs2_truncate_log_shutdown keys on the existence of
6336 * osb->osb_tl_inode so we don't set any of the osb variables
6337 * until we're sure all is well. */
6338 INIT_DELAYED_WORK(&osb
->osb_truncate_log_wq
,
6339 ocfs2_truncate_log_worker
);
6340 atomic_set(&osb
->osb_tl_disable
, 0);
6341 osb
->osb_tl_bh
= tl_bh
;
6342 osb
->osb_tl_inode
= tl_inode
;
6348 * Delayed de-allocation of suballocator blocks.
6350 * Some sets of block de-allocations might involve multiple suballocator inodes.
6352 * The locking for this can get extremely complicated, especially when
6353 * the suballocator inodes to delete from aren't known until deep
6354 * within an unrelated codepath.
6356 * ocfs2_extent_block structures are a good example of this - an inode
6357 * btree could have been grown by any number of nodes each allocating
6358 * out of their own suballoc inode.
6360 * These structures allow the delay of block de-allocation until a
6361 * later time, when locking of multiple cluster inodes won't cause
6366 * Describe a single bit freed from a suballocator. For the block
6367 * suballocators, it represents one block. For the global cluster
6368 * allocator, it represents some clusters and free_bit indicates
6371 struct ocfs2_cached_block_free
{
6372 struct ocfs2_cached_block_free
*free_next
;
6375 unsigned int free_bit
;
6378 struct ocfs2_per_slot_free_list
{
6379 struct ocfs2_per_slot_free_list
*f_next_suballocator
;
6382 struct ocfs2_cached_block_free
*f_first
;
6385 static int ocfs2_free_cached_blocks(struct ocfs2_super
*osb
,
6388 struct ocfs2_cached_block_free
*head
)
6393 struct inode
*inode
;
6394 struct buffer_head
*di_bh
= NULL
;
6395 struct ocfs2_cached_block_free
*tmp
;
6397 inode
= ocfs2_get_system_file_inode(osb
, sysfile_type
, slot
);
6406 ret
= ocfs2_inode_lock(inode
, &di_bh
, 1);
6414 bg_blkno
= head
->free_bg
;
6416 bg_blkno
= ocfs2_which_suballoc_group(head
->free_blk
,
6418 handle
= ocfs2_start_trans(osb
, OCFS2_SUBALLOC_FREE
);
6419 if (IS_ERR(handle
)) {
6420 ret
= PTR_ERR(handle
);
6425 trace_ocfs2_free_cached_blocks(
6426 (unsigned long long)head
->free_blk
, head
->free_bit
);
6428 ret
= ocfs2_free_suballoc_bits(handle
, inode
, di_bh
,
6429 head
->free_bit
, bg_blkno
, 1);
6433 ocfs2_commit_trans(osb
, handle
);
6436 head
= head
->free_next
;
6441 ocfs2_inode_unlock(inode
, 1);
6444 inode_unlock(inode
);
6448 /* Premature exit may have left some dangling items. */
6450 head
= head
->free_next
;
6457 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6458 u64 blkno
, unsigned int bit
)
6461 struct ocfs2_cached_block_free
*item
;
6463 item
= kzalloc(sizeof(*item
), GFP_NOFS
);
6470 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno
, bit
);
6472 item
->free_blk
= blkno
;
6473 item
->free_bit
= bit
;
6474 item
->free_next
= ctxt
->c_global_allocator
;
6476 ctxt
->c_global_allocator
= item
;
6480 static int ocfs2_free_cached_clusters(struct ocfs2_super
*osb
,
6481 struct ocfs2_cached_block_free
*head
)
6483 struct ocfs2_cached_block_free
*tmp
;
6484 struct inode
*tl_inode
= osb
->osb_tl_inode
;
6488 inode_lock(tl_inode
);
6491 if (ocfs2_truncate_log_needs_flush(osb
)) {
6492 ret
= __ocfs2_flush_truncate_log(osb
);
6499 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_UPDATE
);
6500 if (IS_ERR(handle
)) {
6501 ret
= PTR_ERR(handle
);
6506 ret
= ocfs2_truncate_log_append(osb
, handle
, head
->free_blk
,
6509 ocfs2_commit_trans(osb
, handle
);
6511 head
= head
->free_next
;
6520 inode_unlock(tl_inode
);
6523 /* Premature exit may have left some dangling items. */
6525 head
= head
->free_next
;
6532 int ocfs2_run_deallocs(struct ocfs2_super
*osb
,
6533 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
6536 struct ocfs2_per_slot_free_list
*fl
;
6541 while (ctxt
->c_first_suballocator
) {
6542 fl
= ctxt
->c_first_suballocator
;
6545 trace_ocfs2_run_deallocs(fl
->f_inode_type
,
6547 ret2
= ocfs2_free_cached_blocks(osb
,
6557 ctxt
->c_first_suballocator
= fl
->f_next_suballocator
;
6561 if (ctxt
->c_global_allocator
) {
6562 ret2
= ocfs2_free_cached_clusters(osb
,
6563 ctxt
->c_global_allocator
);
6569 ctxt
->c_global_allocator
= NULL
;
6575 static struct ocfs2_per_slot_free_list
*
6576 ocfs2_find_per_slot_free_list(int type
,
6578 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
6580 struct ocfs2_per_slot_free_list
*fl
= ctxt
->c_first_suballocator
;
6583 if (fl
->f_inode_type
== type
&& fl
->f_slot
== slot
)
6586 fl
= fl
->f_next_suballocator
;
6589 fl
= kmalloc(sizeof(*fl
), GFP_NOFS
);
6591 fl
->f_inode_type
= type
;
6594 fl
->f_next_suballocator
= ctxt
->c_first_suballocator
;
6596 ctxt
->c_first_suballocator
= fl
;
6601 static struct ocfs2_per_slot_free_list
*
6602 ocfs2_find_preferred_free_list(int type
,
6605 struct ocfs2_cached_dealloc_ctxt
*ctxt
)
6607 struct ocfs2_per_slot_free_list
*fl
= ctxt
->c_first_suballocator
;
6610 if (fl
->f_inode_type
== type
&& fl
->f_slot
== preferred_slot
) {
6611 *real_slot
= fl
->f_slot
;
6615 fl
= fl
->f_next_suballocator
;
6618 /* If we can't find any free list matching preferred slot, just use
6621 fl
= ctxt
->c_first_suballocator
;
6622 *real_slot
= fl
->f_slot
;
6627 /* Return Value 1 indicates empty */
6628 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree
*et
)
6630 struct ocfs2_per_slot_free_list
*fl
= NULL
;
6632 if (!et
->et_dealloc
)
6635 fl
= et
->et_dealloc
->c_first_suballocator
;
6645 /* If extent was deleted from tree due to extent rotation and merging, and
6646 * no metadata is reserved ahead of time. Try to reuse some extents
6647 * just deleted. This is only used to reuse extent blocks.
6648 * It is supposed to find enough extent blocks in dealloc if our estimation
6649 * on metadata is accurate.
6651 static int ocfs2_reuse_blk_from_dealloc(handle_t
*handle
,
6652 struct ocfs2_extent_tree
*et
,
6653 struct buffer_head
**new_eb_bh
,
6654 int blk_wanted
, int *blk_given
)
6656 int i
, status
= 0, real_slot
;
6657 struct ocfs2_cached_dealloc_ctxt
*dealloc
;
6658 struct ocfs2_per_slot_free_list
*fl
;
6659 struct ocfs2_cached_block_free
*bf
;
6660 struct ocfs2_extent_block
*eb
;
6661 struct ocfs2_super
*osb
=
6662 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
6666 /* If extent tree doesn't have a dealloc, this is not faulty. Just
6667 * tell upper caller dealloc can't provide any block and it should
6668 * ask for alloc to claim more space.
6670 dealloc
= et
->et_dealloc
;
6674 for (i
= 0; i
< blk_wanted
; i
++) {
6675 /* Prefer to use local slot */
6676 fl
= ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE
,
6677 osb
->slot_num
, &real_slot
,
6679 /* If no more block can be reused, we should claim more
6680 * from alloc. Just return here normally.
6688 fl
->f_first
= bf
->free_next
;
6690 new_eb_bh
[i
] = sb_getblk(osb
->sb
, bf
->free_blk
);
6691 if (new_eb_bh
[i
] == NULL
) {
6697 mlog(0, "Reusing block(%llu) from "
6698 "dealloc(local slot:%d, real slot:%d)\n",
6699 bf
->free_blk
, osb
->slot_num
, real_slot
);
6701 ocfs2_set_new_buffer_uptodate(et
->et_ci
, new_eb_bh
[i
]);
6703 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
6705 OCFS2_JOURNAL_ACCESS_CREATE
);
6711 memset(new_eb_bh
[i
]->b_data
, 0, osb
->sb
->s_blocksize
);
6712 eb
= (struct ocfs2_extent_block
*) new_eb_bh
[i
]->b_data
;
6714 /* We can't guarantee that buffer head is still cached, so
6715 * polutlate the extent block again.
6717 strcpy(eb
->h_signature
, OCFS2_EXTENT_BLOCK_SIGNATURE
);
6718 eb
->h_blkno
= cpu_to_le64(bf
->free_blk
);
6719 eb
->h_fs_generation
= cpu_to_le32(osb
->fs_generation
);
6720 eb
->h_suballoc_slot
= cpu_to_le16(real_slot
);
6721 eb
->h_suballoc_loc
= cpu_to_le64(bf
->free_bg
);
6722 eb
->h_suballoc_bit
= cpu_to_le16(bf
->free_bit
);
6723 eb
->h_list
.l_count
=
6724 cpu_to_le16(ocfs2_extent_recs_per_eb(osb
->sb
));
6726 /* We'll also be dirtied by the caller, so
6727 * this isn't absolutely necessary.
6729 ocfs2_journal_dirty(handle
, new_eb_bh
[i
]);
6732 dealloc
->c_first_suballocator
= fl
->f_next_suballocator
;
6741 if (unlikely(status
< 0)) {
6742 for (i
= 0; i
< blk_wanted
; i
++)
6743 brelse(new_eb_bh
[i
]);
6749 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6750 int type
, int slot
, u64 suballoc
,
6751 u64 blkno
, unsigned int bit
)
6754 struct ocfs2_per_slot_free_list
*fl
;
6755 struct ocfs2_cached_block_free
*item
;
6757 fl
= ocfs2_find_per_slot_free_list(type
, slot
, ctxt
);
6764 item
= kzalloc(sizeof(*item
), GFP_NOFS
);
6771 trace_ocfs2_cache_block_dealloc(type
, slot
,
6772 (unsigned long long)suballoc
,
6773 (unsigned long long)blkno
, bit
);
6775 item
->free_bg
= suballoc
;
6776 item
->free_blk
= blkno
;
6777 item
->free_bit
= bit
;
6778 item
->free_next
= fl
->f_first
;
6787 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
6788 struct ocfs2_extent_block
*eb
)
6790 return ocfs2_cache_block_dealloc(ctxt
, EXTENT_ALLOC_SYSTEM_INODE
,
6791 le16_to_cpu(eb
->h_suballoc_slot
),
6792 le64_to_cpu(eb
->h_suballoc_loc
),
6793 le64_to_cpu(eb
->h_blkno
),
6794 le16_to_cpu(eb
->h_suballoc_bit
));
6797 static int ocfs2_zero_func(handle_t
*handle
, struct buffer_head
*bh
)
6799 set_buffer_uptodate(bh
);
6800 mark_buffer_dirty(bh
);
6804 void ocfs2_map_and_dirty_page(struct inode
*inode
, handle_t
*handle
,
6805 unsigned int from
, unsigned int to
,
6806 struct page
*page
, int zero
, u64
*phys
)
6808 int ret
, partial
= 0;
6809 loff_t start_byte
= ((loff_t
)page
->index
<< PAGE_SHIFT
) + from
;
6810 loff_t length
= to
- from
;
6812 ret
= ocfs2_map_page_blocks(page
, phys
, inode
, from
, to
, 0);
6817 zero_user_segment(page
, from
, to
);
6820 * Need to set the buffers we zero'd into uptodate
6821 * here if they aren't - ocfs2_map_page_blocks()
6822 * might've skipped some
6824 ret
= walk_page_buffers(handle
, page_buffers(page
),
6829 else if (ocfs2_should_order_data(inode
)) {
6830 ret
= ocfs2_jbd2_inode_add_write(handle
, inode
,
6831 start_byte
, length
);
6837 SetPageUptodate(page
);
6839 flush_dcache_page(page
);
6842 static void ocfs2_zero_cluster_pages(struct inode
*inode
, loff_t start
,
6843 loff_t end
, struct page
**pages
,
6844 int numpages
, u64 phys
, handle_t
*handle
)
6848 unsigned int from
, to
= PAGE_SIZE
;
6849 struct super_block
*sb
= inode
->i_sb
;
6851 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb
)));
6857 for(i
= 0; i
< numpages
; i
++) {
6860 from
= start
& (PAGE_SIZE
- 1);
6861 if ((end
>> PAGE_SHIFT
) == page
->index
)
6862 to
= end
& (PAGE_SIZE
- 1);
6864 BUG_ON(from
> PAGE_SIZE
);
6865 BUG_ON(to
> PAGE_SIZE
);
6867 ocfs2_map_and_dirty_page(inode
, handle
, from
, to
, page
, 1,
6870 start
= (page
->index
+ 1) << PAGE_SHIFT
;
6874 ocfs2_unlock_and_free_pages(pages
, numpages
);
6877 int ocfs2_grab_pages(struct inode
*inode
, loff_t start
, loff_t end
,
6878 struct page
**pages
, int *num
)
6880 int numpages
, ret
= 0;
6881 struct address_space
*mapping
= inode
->i_mapping
;
6882 unsigned long index
;
6883 loff_t last_page_bytes
;
6885 BUG_ON(start
> end
);
6888 last_page_bytes
= PAGE_ALIGN(end
);
6889 index
= start
>> PAGE_SHIFT
;
6891 pages
[numpages
] = find_or_create_page(mapping
, index
, GFP_NOFS
);
6892 if (!pages
[numpages
]) {
6900 } while (index
< (last_page_bytes
>> PAGE_SHIFT
));
6905 ocfs2_unlock_and_free_pages(pages
, numpages
);
6914 static int ocfs2_grab_eof_pages(struct inode
*inode
, loff_t start
, loff_t end
,
6915 struct page
**pages
, int *num
)
6917 struct super_block
*sb
= inode
->i_sb
;
6919 BUG_ON(start
>> OCFS2_SB(sb
)->s_clustersize_bits
!=
6920 (end
- 1) >> OCFS2_SB(sb
)->s_clustersize_bits
);
6922 return ocfs2_grab_pages(inode
, start
, end
, pages
, num
);
6926 * Zero partial cluster for a hole punch or truncate. This avoids exposing
6927 * nonzero data on subsequent file extends.
6929 * We need to call this before i_size is updated on the inode because
6930 * otherwise block_write_full_page() will skip writeout of pages past
6933 int ocfs2_zero_range_for_truncate(struct inode
*inode
, handle_t
*handle
,
6934 u64 range_start
, u64 range_end
)
6936 int ret
= 0, numpages
;
6937 struct page
**pages
= NULL
;
6939 unsigned int ext_flags
;
6940 struct super_block
*sb
= inode
->i_sb
;
6943 * File systems which don't support sparse files zero on every
6946 if (!ocfs2_sparse_alloc(OCFS2_SB(sb
)))
6950 * Avoid zeroing pages fully beyond current i_size. It is pointless as
6951 * underlying blocks of those pages should be already zeroed out and
6952 * page writeback will skip them anyway.
6954 range_end
= min_t(u64
, range_end
, i_size_read(inode
));
6955 if (range_start
>= range_end
)
6958 pages
= kcalloc(ocfs2_pages_per_cluster(sb
),
6959 sizeof(struct page
*), GFP_NOFS
);
6960 if (pages
== NULL
) {
6966 ret
= ocfs2_extent_map_get_blocks(inode
,
6967 range_start
>> sb
->s_blocksize_bits
,
6968 &phys
, NULL
, &ext_flags
);
6975 * Tail is a hole, or is marked unwritten. In either case, we
6976 * can count on read and write to return/push zero's.
6978 if (phys
== 0 || ext_flags
& OCFS2_EXT_UNWRITTEN
)
6981 ret
= ocfs2_grab_eof_pages(inode
, range_start
, range_end
, pages
,
6988 ocfs2_zero_cluster_pages(inode
, range_start
, range_end
, pages
,
6989 numpages
, phys
, handle
);
6992 * Initiate writeout of the pages we zero'd here. We don't
6993 * wait on them - the truncate_inode_pages() call later will
6996 ret
= filemap_fdatawrite_range(inode
->i_mapping
, range_start
,
7007 static void ocfs2_zero_dinode_id2_with_xattr(struct inode
*inode
,
7008 struct ocfs2_dinode
*di
)
7010 unsigned int blocksize
= 1 << inode
->i_sb
->s_blocksize_bits
;
7011 unsigned int xattrsize
= le16_to_cpu(di
->i_xattr_inline_size
);
7013 if (le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_XATTR_FL
)
7014 memset(&di
->id2
, 0, blocksize
-
7015 offsetof(struct ocfs2_dinode
, id2
) -
7018 memset(&di
->id2
, 0, blocksize
-
7019 offsetof(struct ocfs2_dinode
, id2
));
7022 void ocfs2_dinode_new_extent_list(struct inode
*inode
,
7023 struct ocfs2_dinode
*di
)
7025 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7026 di
->id2
.i_list
.l_tree_depth
= 0;
7027 di
->id2
.i_list
.l_next_free_rec
= 0;
7028 di
->id2
.i_list
.l_count
= cpu_to_le16(
7029 ocfs2_extent_recs_per_inode_with_xattr(inode
->i_sb
, di
));
7032 void ocfs2_set_inode_data_inline(struct inode
*inode
, struct ocfs2_dinode
*di
)
7034 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7035 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7037 spin_lock(&oi
->ip_lock
);
7038 oi
->ip_dyn_features
|= OCFS2_INLINE_DATA_FL
;
7039 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7040 spin_unlock(&oi
->ip_lock
);
7043 * We clear the entire i_data structure here so that all
7044 * fields can be properly initialized.
7046 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7048 idata
->id_count
= cpu_to_le16(
7049 ocfs2_max_inline_data_with_xattr(inode
->i_sb
, di
));
7052 int ocfs2_convert_inline_data_to_extents(struct inode
*inode
,
7053 struct buffer_head
*di_bh
)
7055 int ret
, has_data
, num_pages
= 0;
7060 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7061 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7062 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7063 struct ocfs2_alloc_context
*data_ac
= NULL
;
7064 struct page
*page
= NULL
;
7065 struct ocfs2_extent_tree et
;
7068 has_data
= i_size_read(inode
) ? 1 : 0;
7071 ret
= ocfs2_reserve_clusters(osb
, 1, &data_ac
);
7078 handle
= ocfs2_start_trans(osb
,
7079 ocfs2_inline_to_extents_credits(osb
->sb
));
7080 if (IS_ERR(handle
)) {
7081 ret
= PTR_ERR(handle
);
7086 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7087 OCFS2_JOURNAL_ACCESS_WRITE
);
7094 unsigned int page_end
= min_t(unsigned, PAGE_SIZE
,
7095 osb
->s_clustersize
);
7098 ret
= dquot_alloc_space_nodirty(inode
,
7099 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7104 data_ac
->ac_resv
= &oi
->ip_la_data_resv
;
7106 ret
= ocfs2_claim_clusters(handle
, data_ac
, 1, &bit_off
,
7114 * Save two copies, one for insert, and one that can
7115 * be changed by ocfs2_map_and_dirty_page() below.
7117 block
= phys
= ocfs2_clusters_to_blocks(inode
->i_sb
, bit_off
);
7119 ret
= ocfs2_grab_eof_pages(inode
, 0, page_end
, &page
,
7128 * This should populate the 1st page for us and mark
7131 ret
= ocfs2_read_inline_data(inode
, page
, di_bh
);
7138 ocfs2_map_and_dirty_page(inode
, handle
, 0, page_end
, page
, 0,
7142 spin_lock(&oi
->ip_lock
);
7143 oi
->ip_dyn_features
&= ~OCFS2_INLINE_DATA_FL
;
7144 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7145 spin_unlock(&oi
->ip_lock
);
7147 ocfs2_update_inode_fsync_trans(handle
, inode
, 1);
7148 ocfs2_dinode_new_extent_list(inode
, di
);
7150 ocfs2_journal_dirty(handle
, di_bh
);
7154 * An error at this point should be extremely rare. If
7155 * this proves to be false, we could always re-build
7156 * the in-inode data from our pages.
7158 ocfs2_init_dinode_extent_tree(&et
, INODE_CACHE(inode
), di_bh
);
7159 ret
= ocfs2_insert_extent(handle
, &et
, 0, block
, 1, 0, NULL
);
7166 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7171 ocfs2_unlock_and_free_pages(&page
, num_pages
);
7174 if (ret
< 0 && did_quota
)
7175 dquot_free_space_nodirty(inode
,
7176 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7179 if (data_ac
->ac_which
== OCFS2_AC_USE_LOCAL
)
7180 ocfs2_free_local_alloc_bits(osb
, handle
, data_ac
,
7183 ocfs2_free_clusters(handle
,
7186 ocfs2_clusters_to_blocks(osb
->sb
, bit_off
),
7190 ocfs2_commit_trans(osb
, handle
);
7194 ocfs2_free_alloc_context(data_ac
);
7199 * It is expected, that by the time you call this function,
7200 * inode->i_size and fe->i_size have been adjusted.
7202 * WARNING: This will kfree the truncate context
7204 int ocfs2_commit_truncate(struct ocfs2_super
*osb
,
7205 struct inode
*inode
,
7206 struct buffer_head
*di_bh
)
7208 int status
= 0, i
, flags
= 0;
7209 u32 new_highest_cpos
, range
, trunc_cpos
, trunc_len
, phys_cpos
, coff
;
7211 struct ocfs2_extent_list
*el
;
7212 struct ocfs2_extent_rec
*rec
;
7213 struct ocfs2_path
*path
= NULL
;
7214 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7215 struct ocfs2_extent_list
*root_el
= &(di
->id2
.i_list
);
7216 u64 refcount_loc
= le64_to_cpu(di
->i_refcount_loc
);
7217 struct ocfs2_extent_tree et
;
7218 struct ocfs2_cached_dealloc_ctxt dealloc
;
7219 struct ocfs2_refcount_tree
*ref_tree
= NULL
;
7221 ocfs2_init_dinode_extent_tree(&et
, INODE_CACHE(inode
), di_bh
);
7222 ocfs2_init_dealloc_ctxt(&dealloc
);
7224 new_highest_cpos
= ocfs2_clusters_for_bytes(osb
->sb
,
7225 i_size_read(inode
));
7227 path
= ocfs2_new_path(di_bh
, &di
->id2
.i_list
,
7228 ocfs2_journal_access_di
);
7235 ocfs2_extent_map_trunc(inode
, new_highest_cpos
);
7239 * Check that we still have allocation to delete.
7241 if (OCFS2_I(inode
)->ip_clusters
== 0) {
7247 * Truncate always works against the rightmost tree branch.
7249 status
= ocfs2_find_path(INODE_CACHE(inode
), path
, UINT_MAX
);
7255 trace_ocfs2_commit_truncate(
7256 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7258 OCFS2_I(inode
)->ip_clusters
,
7259 path
->p_tree_depth
);
7262 * By now, el will point to the extent list on the bottom most
7263 * portion of this tree. Only the tail record is considered in
7266 * We handle the following cases, in order:
7267 * - empty extent: delete the remaining branch
7268 * - remove the entire record
7269 * - remove a partial record
7270 * - no record needs to be removed (truncate has completed)
7272 el
= path_leaf_el(path
);
7273 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
7274 ocfs2_error(inode
->i_sb
,
7275 "Inode %llu has empty extent block at %llu\n",
7276 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7277 (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
7282 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
7283 rec
= &el
->l_recs
[i
];
7284 flags
= rec
->e_flags
;
7285 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
7287 if (i
== 0 && ocfs2_is_empty_extent(rec
)) {
7289 * Lower levels depend on this never happening, but it's best
7290 * to check it up here before changing the tree.
7292 if (root_el
->l_tree_depth
&& rec
->e_int_clusters
== 0) {
7293 mlog(ML_ERROR
, "Inode %lu has an empty "
7294 "extent record, depth %u\n", inode
->i_ino
,
7295 le16_to_cpu(root_el
->l_tree_depth
));
7296 status
= ocfs2_remove_rightmost_empty_extent(osb
,
7297 &et
, path
, &dealloc
);
7303 ocfs2_reinit_path(path
, 1);
7306 trunc_cpos
= le32_to_cpu(rec
->e_cpos
);
7310 } else if (le32_to_cpu(rec
->e_cpos
) >= new_highest_cpos
) {
7312 * Truncate entire record.
7314 trunc_cpos
= le32_to_cpu(rec
->e_cpos
);
7315 trunc_len
= ocfs2_rec_clusters(el
, rec
);
7316 blkno
= le64_to_cpu(rec
->e_blkno
);
7317 } else if (range
> new_highest_cpos
) {
7319 * Partial truncate. it also should be
7320 * the last truncate we're doing.
7322 trunc_cpos
= new_highest_cpos
;
7323 trunc_len
= range
- new_highest_cpos
;
7324 coff
= new_highest_cpos
- le32_to_cpu(rec
->e_cpos
);
7325 blkno
= le64_to_cpu(rec
->e_blkno
) +
7326 ocfs2_clusters_to_blocks(inode
->i_sb
, coff
);
7329 * Truncate completed, leave happily.
7335 phys_cpos
= ocfs2_blocks_to_clusters(inode
->i_sb
, blkno
);
7337 if ((flags
& OCFS2_EXT_REFCOUNTED
) && trunc_len
&& !ref_tree
) {
7338 status
= ocfs2_lock_refcount_tree(osb
, refcount_loc
, 1,
7346 status
= ocfs2_remove_btree_range(inode
, &et
, trunc_cpos
,
7347 phys_cpos
, trunc_len
, flags
, &dealloc
,
7348 refcount_loc
, true);
7354 ocfs2_reinit_path(path
, 1);
7357 * The check above will catch the case where we've truncated
7358 * away all allocation.
7364 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
7366 ocfs2_schedule_truncate_log_flush(osb
, 1);
7368 ocfs2_run_deallocs(osb
, &dealloc
);
7370 ocfs2_free_path(path
);
7376 * 'start' is inclusive, 'end' is not.
7378 int ocfs2_truncate_inline(struct inode
*inode
, struct buffer_head
*di_bh
,
7379 unsigned int start
, unsigned int end
, int trunc
)
7382 unsigned int numbytes
;
7384 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7385 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7386 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7388 /* No need to punch hole beyond i_size. */
7389 if (start
>= i_size_read(inode
))
7392 if (end
> i_size_read(inode
))
7393 end
= i_size_read(inode
);
7395 BUG_ON(start
> end
);
7397 if (!(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
) ||
7398 !(le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_DATA_FL
) ||
7399 !ocfs2_supports_inline_data(osb
)) {
7400 ocfs2_error(inode
->i_sb
,
7401 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7402 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7403 le16_to_cpu(di
->i_dyn_features
),
7404 OCFS2_I(inode
)->ip_dyn_features
,
7405 osb
->s_feature_incompat
);
7410 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
7411 if (IS_ERR(handle
)) {
7412 ret
= PTR_ERR(handle
);
7417 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7418 OCFS2_JOURNAL_ACCESS_WRITE
);
7424 numbytes
= end
- start
;
7425 memset(idata
->id_data
+ start
, 0, numbytes
);
7428 * No need to worry about the data page here - it's been
7429 * truncated already and inline data doesn't need it for
7430 * pushing zero's to disk, so we'll let read_folio pick it up
7434 i_size_write(inode
, start
);
7435 di
->i_size
= cpu_to_le64(start
);
7438 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7439 inode
->i_ctime
= inode
->i_mtime
= current_time(inode
);
7441 di
->i_ctime
= di
->i_mtime
= cpu_to_le64(inode
->i_ctime
.tv_sec
);
7442 di
->i_ctime_nsec
= di
->i_mtime_nsec
= cpu_to_le32(inode
->i_ctime
.tv_nsec
);
7444 ocfs2_update_inode_fsync_trans(handle
, inode
, 1);
7445 ocfs2_journal_dirty(handle
, di_bh
);
7448 ocfs2_commit_trans(osb
, handle
);
7454 static int ocfs2_trim_extent(struct super_block
*sb
,
7455 struct ocfs2_group_desc
*gd
,
7456 u64 group
, u32 start
, u32 count
)
7458 u64 discard
, bcount
;
7459 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
7461 bcount
= ocfs2_clusters_to_blocks(sb
, count
);
7462 discard
= ocfs2_clusters_to_blocks(sb
, start
);
7465 * For the first cluster group, the gd->bg_blkno is not at the start
7466 * of the group, but at an offset from the start. If we add it while
7467 * calculating discard for first group, we will wrongly start fstrim a
7468 * few blocks after the desried start block and the range can cross
7469 * over into the next cluster group. So, add it only if this is not
7470 * the first cluster group.
7472 if (group
!= osb
->first_cluster_group_blkno
)
7473 discard
+= le64_to_cpu(gd
->bg_blkno
);
7475 trace_ocfs2_trim_extent(sb
, (unsigned long long)discard
, bcount
);
7477 return sb_issue_discard(sb
, discard
, bcount
, GFP_NOFS
, 0);
7480 static int ocfs2_trim_group(struct super_block
*sb
,
7481 struct ocfs2_group_desc
*gd
, u64 group
,
7482 u32 start
, u32 max
, u32 minbits
)
7484 int ret
= 0, count
= 0, next
;
7485 void *bitmap
= gd
->bg_bitmap
;
7487 if (le16_to_cpu(gd
->bg_free_bits_count
) < minbits
)
7490 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd
->bg_blkno
),
7491 start
, max
, minbits
);
7493 while (start
< max
) {
7494 start
= ocfs2_find_next_zero_bit(bitmap
, max
, start
);
7497 next
= ocfs2_find_next_bit(bitmap
, max
, start
);
7499 if ((next
- start
) >= minbits
) {
7500 ret
= ocfs2_trim_extent(sb
, gd
, group
,
7501 start
, next
- start
);
7506 count
+= next
- start
;
7510 if (fatal_signal_pending(current
)) {
7511 count
= -ERESTARTSYS
;
7515 if ((le16_to_cpu(gd
->bg_free_bits_count
) - count
) < minbits
)
7526 int ocfs2_trim_mainbm(struct super_block
*sb
, struct fstrim_range
*range
)
7528 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
7529 u64 start
, len
, trimmed
= 0, first_group
, last_group
= 0, group
= 0;
7531 u32 first_bit
, last_bit
, minlen
;
7532 struct buffer_head
*main_bm_bh
= NULL
;
7533 struct inode
*main_bm_inode
= NULL
;
7534 struct buffer_head
*gd_bh
= NULL
;
7535 struct ocfs2_dinode
*main_bm
;
7536 struct ocfs2_group_desc
*gd
= NULL
;
7538 start
= range
->start
>> osb
->s_clustersize_bits
;
7539 len
= range
->len
>> osb
->s_clustersize_bits
;
7540 minlen
= range
->minlen
>> osb
->s_clustersize_bits
;
7542 if (minlen
>= osb
->bitmap_cpg
|| range
->len
< sb
->s_blocksize
)
7545 trace_ocfs2_trim_mainbm(start
, len
, minlen
);
7548 main_bm_inode
= ocfs2_get_system_file_inode(osb
,
7549 GLOBAL_BITMAP_SYSTEM_INODE
,
7550 OCFS2_INVALID_SLOT
);
7551 if (!main_bm_inode
) {
7557 inode_lock(main_bm_inode
);
7559 ret
= ocfs2_inode_lock(main_bm_inode
, &main_bm_bh
, 0);
7564 main_bm
= (struct ocfs2_dinode
*)main_bm_bh
->b_data
;
7567 * Do some check before trim the first group.
7570 if (start
>= le32_to_cpu(main_bm
->i_clusters
)) {
7575 if (start
+ len
> le32_to_cpu(main_bm
->i_clusters
))
7576 len
= le32_to_cpu(main_bm
->i_clusters
) - start
;
7579 * Determine first and last group to examine based on
7582 first_group
= ocfs2_which_cluster_group(main_bm_inode
, start
);
7583 if (first_group
== osb
->first_cluster_group_blkno
)
7586 first_bit
= start
- ocfs2_blocks_to_clusters(sb
,
7588 last_group
= ocfs2_which_cluster_group(main_bm_inode
,
7590 group
= first_group
;
7594 if (first_bit
+ len
>= osb
->bitmap_cpg
)
7595 last_bit
= osb
->bitmap_cpg
;
7597 last_bit
= first_bit
+ len
;
7599 ret
= ocfs2_read_group_descriptor(main_bm_inode
,
7607 gd
= (struct ocfs2_group_desc
*)gd_bh
->b_data
;
7608 cnt
= ocfs2_trim_group(sb
, gd
, group
,
7609 first_bit
, last_bit
, minlen
);
7619 len
-= osb
->bitmap_cpg
- first_bit
;
7621 if (group
== osb
->first_cluster_group_blkno
)
7622 group
= ocfs2_clusters_to_blocks(sb
, osb
->bitmap_cpg
);
7624 group
+= ocfs2_clusters_to_blocks(sb
, osb
->bitmap_cpg
);
7628 ocfs2_inode_unlock(main_bm_inode
, 0);
7632 inode_unlock(main_bm_inode
);
7633 iput(main_bm_inode
);
7636 * If all the groups trim are not done or failed, but we should release
7637 * main_bm related locks for avoiding the current IO starve, then go to
7638 * trim the next group
7640 if (ret
>= 0 && group
<= last_group
) {
7645 range
->len
= trimmed
* sb
->s_blocksize
;
7649 int ocfs2_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
7652 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
7653 struct ocfs2_trim_fs_info info
, *pinfo
= NULL
;
7655 ocfs2_trim_fs_lock_res_init(osb
);
7657 trace_ocfs2_trim_fs(range
->start
, range
->len
, range
->minlen
);
7659 ret
= ocfs2_trim_fs_lock(osb
, NULL
, 1);
7661 if (ret
!= -EAGAIN
) {
7663 ocfs2_trim_fs_lock_res_uninit(osb
);
7667 mlog(ML_NOTICE
, "Wait for trim on device (%s) to "
7668 "finish, which is running from another node.\n",
7670 ret
= ocfs2_trim_fs_lock(osb
, &info
, 0);
7673 ocfs2_trim_fs_lock_res_uninit(osb
);
7677 if (info
.tf_valid
&& info
.tf_success
&&
7678 info
.tf_start
== range
->start
&&
7679 info
.tf_len
== range
->len
&&
7680 info
.tf_minlen
== range
->minlen
) {
7681 /* Avoid sending duplicated trim to a shared device */
7682 mlog(ML_NOTICE
, "The same trim on device (%s) was "
7683 "just done from node (%u), return.\n",
7684 osb
->dev_str
, info
.tf_nodenum
);
7685 range
->len
= info
.tf_trimlen
;
7690 info
.tf_nodenum
= osb
->node_num
;
7691 info
.tf_start
= range
->start
;
7692 info
.tf_len
= range
->len
;
7693 info
.tf_minlen
= range
->minlen
;
7695 ret
= ocfs2_trim_mainbm(sb
, range
);
7697 info
.tf_trimlen
= range
->len
;
7698 info
.tf_success
= (ret
< 0 ? 0 : 1);
7701 ocfs2_trim_fs_unlock(osb
, pinfo
);
7702 ocfs2_trim_fs_lock_res_uninit(osb
);