1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
7 * Extent allocs and frees
9 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/highmem.h>
16 #include <linux/swap.h>
17 #include <linux/quotaops.h>
18 #include <linux/blkdev.h>
19 #include <linux/sched/signal.h>
21 #include <cluster/masklog.h>
27 #include "blockcheck.h"
29 #include "extent_map.h"
32 #include "localalloc.h"
39 #include "refcounttree.h"
40 #include "ocfs2_trace.h"
42 #include "buffer_head_io.h"
44 enum ocfs2_contig_type
{
51 static enum ocfs2_contig_type
52 ocfs2_extent_rec_contig(struct super_block
*sb
,
53 struct ocfs2_extent_rec
*ext
,
54 struct ocfs2_extent_rec
*insert_rec
);
56 * Operations for a specific extent tree type.
58 * To implement an on-disk btree (extent tree) type in ocfs2, add
59 * an ocfs2_extent_tree_operations structure and the matching
60 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
61 * for the allocation portion of the extent tree.
63 struct ocfs2_extent_tree_operations
{
65 * last_eb_blk is the block number of the right most leaf extent
66 * block. Most on-disk structures containing an extent tree store
67 * this value for fast access. The ->eo_set_last_eb_blk() and
68 * ->eo_get_last_eb_blk() operations access this value. They are
71 void (*eo_set_last_eb_blk
)(struct ocfs2_extent_tree
*et
,
73 u64 (*eo_get_last_eb_blk
)(struct ocfs2_extent_tree
*et
);
76 * The on-disk structure usually keeps track of how many total
77 * clusters are stored in this extent tree. This function updates
78 * that value. new_clusters is the delta, and must be
79 * added to the total. Required.
81 void (*eo_update_clusters
)(struct ocfs2_extent_tree
*et
,
85 * If this extent tree is supported by an extent map, insert
86 * a record into the map.
88 void (*eo_extent_map_insert
)(struct ocfs2_extent_tree
*et
,
89 struct ocfs2_extent_rec
*rec
);
92 * If this extent tree is supported by an extent map, truncate the
95 void (*eo_extent_map_truncate
)(struct ocfs2_extent_tree
*et
,
99 * If ->eo_insert_check() exists, it is called before rec is
100 * inserted into the extent tree. It is optional.
102 int (*eo_insert_check
)(struct ocfs2_extent_tree
*et
,
103 struct ocfs2_extent_rec
*rec
);
104 int (*eo_sanity_check
)(struct ocfs2_extent_tree
*et
);
107 * --------------------------------------------------------------
108 * The remaining are internal to ocfs2_extent_tree and don't have
113 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
116 void (*eo_fill_root_el
)(struct ocfs2_extent_tree
*et
);
119 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
120 * it exists. If it does not, et->et_max_leaf_clusters is set
121 * to 0 (unlimited). Optional.
123 void (*eo_fill_max_leaf_clusters
)(struct ocfs2_extent_tree
*et
);
126 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
127 * are contiguous or not. Optional. Don't need to set it if use
128 * ocfs2_extent_rec as the tree leaf.
130 enum ocfs2_contig_type
131 (*eo_extent_contig
)(struct ocfs2_extent_tree
*et
,
132 struct ocfs2_extent_rec
*ext
,
133 struct ocfs2_extent_rec
*insert_rec
);
138 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
141 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
);
142 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
144 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree
*et
,
146 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree
*et
,
147 struct ocfs2_extent_rec
*rec
);
148 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree
*et
,
150 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree
*et
,
151 struct ocfs2_extent_rec
*rec
);
152 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree
*et
);
153 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
);
155 static int ocfs2_reuse_blk_from_dealloc(handle_t
*handle
,
156 struct ocfs2_extent_tree
*et
,
157 struct buffer_head
**new_eb_bh
,
158 int blk_wanted
, int *blk_given
);
159 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree
*et
);
161 static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops
= {
162 .eo_set_last_eb_blk
= ocfs2_dinode_set_last_eb_blk
,
163 .eo_get_last_eb_blk
= ocfs2_dinode_get_last_eb_blk
,
164 .eo_update_clusters
= ocfs2_dinode_update_clusters
,
165 .eo_extent_map_insert
= ocfs2_dinode_extent_map_insert
,
166 .eo_extent_map_truncate
= ocfs2_dinode_extent_map_truncate
,
167 .eo_insert_check
= ocfs2_dinode_insert_check
,
168 .eo_sanity_check
= ocfs2_dinode_sanity_check
,
169 .eo_fill_root_el
= ocfs2_dinode_fill_root_el
,
172 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
175 struct ocfs2_dinode
*di
= et
->et_object
;
177 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
178 di
->i_last_eb_blk
= cpu_to_le64(blkno
);
181 static u64
ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
183 struct ocfs2_dinode
*di
= et
->et_object
;
185 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
186 return le64_to_cpu(di
->i_last_eb_blk
);
189 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree
*et
,
192 struct ocfs2_inode_info
*oi
= cache_info_to_inode(et
->et_ci
);
193 struct ocfs2_dinode
*di
= et
->et_object
;
195 le32_add_cpu(&di
->i_clusters
, clusters
);
196 spin_lock(&oi
->ip_lock
);
197 oi
->ip_clusters
= le32_to_cpu(di
->i_clusters
);
198 spin_unlock(&oi
->ip_lock
);
201 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree
*et
,
202 struct ocfs2_extent_rec
*rec
)
204 struct inode
*inode
= &cache_info_to_inode(et
->et_ci
)->vfs_inode
;
206 ocfs2_extent_map_insert_rec(inode
, rec
);
209 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree
*et
,
212 struct inode
*inode
= &cache_info_to_inode(et
->et_ci
)->vfs_inode
;
214 ocfs2_extent_map_trunc(inode
, clusters
);
217 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree
*et
,
218 struct ocfs2_extent_rec
*rec
)
220 struct ocfs2_inode_info
*oi
= cache_info_to_inode(et
->et_ci
);
221 struct ocfs2_super
*osb
= OCFS2_SB(oi
->vfs_inode
.i_sb
);
223 BUG_ON(oi
->ip_dyn_features
& OCFS2_INLINE_DATA_FL
);
224 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb
) &&
225 (oi
->ip_clusters
!= le32_to_cpu(rec
->e_cpos
)),
226 "Device %s, asking for sparse allocation: inode %llu, "
227 "cpos %u, clusters %u\n",
229 (unsigned long long)oi
->ip_blkno
,
230 rec
->e_cpos
, oi
->ip_clusters
);
235 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree
*et
)
237 struct ocfs2_dinode
*di
= et
->et_object
;
239 BUG_ON(et
->et_ops
!= &ocfs2_dinode_et_ops
);
240 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
245 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree
*et
)
247 struct ocfs2_dinode
*di
= et
->et_object
;
249 et
->et_root_el
= &di
->id2
.i_list
;
253 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree
*et
)
255 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
257 et
->et_root_el
= &vb
->vb_xv
->xr_list
;
260 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
263 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
265 vb
->vb_xv
->xr_last_eb_blk
= cpu_to_le64(blkno
);
268 static u64
ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
270 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
272 return le64_to_cpu(vb
->vb_xv
->xr_last_eb_blk
);
275 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree
*et
,
278 struct ocfs2_xattr_value_buf
*vb
= et
->et_object
;
280 le32_add_cpu(&vb
->vb_xv
->xr_clusters
, clusters
);
283 static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops
= {
284 .eo_set_last_eb_blk
= ocfs2_xattr_value_set_last_eb_blk
,
285 .eo_get_last_eb_blk
= ocfs2_xattr_value_get_last_eb_blk
,
286 .eo_update_clusters
= ocfs2_xattr_value_update_clusters
,
287 .eo_fill_root_el
= ocfs2_xattr_value_fill_root_el
,
290 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree
*et
)
292 struct ocfs2_xattr_block
*xb
= et
->et_object
;
294 et
->et_root_el
= &xb
->xb_attrs
.xb_root
.xt_list
;
297 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree
*et
)
299 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
300 et
->et_max_leaf_clusters
=
301 ocfs2_clusters_for_bytes(sb
, OCFS2_MAX_XATTR_TREE_LEAF_SIZE
);
304 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
307 struct ocfs2_xattr_block
*xb
= et
->et_object
;
308 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
310 xt
->xt_last_eb_blk
= cpu_to_le64(blkno
);
313 static u64
ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
315 struct ocfs2_xattr_block
*xb
= et
->et_object
;
316 struct ocfs2_xattr_tree_root
*xt
= &xb
->xb_attrs
.xb_root
;
318 return le64_to_cpu(xt
->xt_last_eb_blk
);
321 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree
*et
,
324 struct ocfs2_xattr_block
*xb
= et
->et_object
;
326 le32_add_cpu(&xb
->xb_attrs
.xb_root
.xt_clusters
, clusters
);
329 static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops
= {
330 .eo_set_last_eb_blk
= ocfs2_xattr_tree_set_last_eb_blk
,
331 .eo_get_last_eb_blk
= ocfs2_xattr_tree_get_last_eb_blk
,
332 .eo_update_clusters
= ocfs2_xattr_tree_update_clusters
,
333 .eo_fill_root_el
= ocfs2_xattr_tree_fill_root_el
,
334 .eo_fill_max_leaf_clusters
= ocfs2_xattr_tree_fill_max_leaf_clusters
,
337 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
340 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
342 dx_root
->dr_last_eb_blk
= cpu_to_le64(blkno
);
345 static u64
ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
347 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
349 return le64_to_cpu(dx_root
->dr_last_eb_blk
);
352 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree
*et
,
355 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
357 le32_add_cpu(&dx_root
->dr_clusters
, clusters
);
360 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree
*et
)
362 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
364 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root
));
369 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree
*et
)
371 struct ocfs2_dx_root_block
*dx_root
= et
->et_object
;
373 et
->et_root_el
= &dx_root
->dr_list
;
376 static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops
= {
377 .eo_set_last_eb_blk
= ocfs2_dx_root_set_last_eb_blk
,
378 .eo_get_last_eb_blk
= ocfs2_dx_root_get_last_eb_blk
,
379 .eo_update_clusters
= ocfs2_dx_root_update_clusters
,
380 .eo_sanity_check
= ocfs2_dx_root_sanity_check
,
381 .eo_fill_root_el
= ocfs2_dx_root_fill_root_el
,
384 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree
*et
)
386 struct ocfs2_refcount_block
*rb
= et
->et_object
;
388 et
->et_root_el
= &rb
->rf_list
;
391 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
394 struct ocfs2_refcount_block
*rb
= et
->et_object
;
396 rb
->rf_last_eb_blk
= cpu_to_le64(blkno
);
399 static u64
ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
401 struct ocfs2_refcount_block
*rb
= et
->et_object
;
403 return le64_to_cpu(rb
->rf_last_eb_blk
);
406 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree
*et
,
409 struct ocfs2_refcount_block
*rb
= et
->et_object
;
411 le32_add_cpu(&rb
->rf_clusters
, clusters
);
414 static enum ocfs2_contig_type
415 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree
*et
,
416 struct ocfs2_extent_rec
*ext
,
417 struct ocfs2_extent_rec
*insert_rec
)
422 static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops
= {
423 .eo_set_last_eb_blk
= ocfs2_refcount_tree_set_last_eb_blk
,
424 .eo_get_last_eb_blk
= ocfs2_refcount_tree_get_last_eb_blk
,
425 .eo_update_clusters
= ocfs2_refcount_tree_update_clusters
,
426 .eo_fill_root_el
= ocfs2_refcount_tree_fill_root_el
,
427 .eo_extent_contig
= ocfs2_refcount_tree_extent_contig
,
430 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree
*et
,
431 struct ocfs2_caching_info
*ci
,
432 struct buffer_head
*bh
,
433 ocfs2_journal_access_func access
,
435 const struct ocfs2_extent_tree_operations
*ops
)
440 et
->et_root_journal_access
= access
;
442 obj
= (void *)bh
->b_data
;
444 et
->et_dealloc
= NULL
;
446 et
->et_ops
->eo_fill_root_el(et
);
447 if (!et
->et_ops
->eo_fill_max_leaf_clusters
)
448 et
->et_max_leaf_clusters
= 0;
450 et
->et_ops
->eo_fill_max_leaf_clusters(et
);
453 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree
*et
,
454 struct ocfs2_caching_info
*ci
,
455 struct buffer_head
*bh
)
457 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_di
,
458 NULL
, &ocfs2_dinode_et_ops
);
461 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree
*et
,
462 struct ocfs2_caching_info
*ci
,
463 struct buffer_head
*bh
)
465 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_xb
,
466 NULL
, &ocfs2_xattr_tree_et_ops
);
469 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree
*et
,
470 struct ocfs2_caching_info
*ci
,
471 struct ocfs2_xattr_value_buf
*vb
)
473 __ocfs2_init_extent_tree(et
, ci
, vb
->vb_bh
, vb
->vb_access
, vb
,
474 &ocfs2_xattr_value_et_ops
);
477 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree
*et
,
478 struct ocfs2_caching_info
*ci
,
479 struct buffer_head
*bh
)
481 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_dr
,
482 NULL
, &ocfs2_dx_root_et_ops
);
485 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree
*et
,
486 struct ocfs2_caching_info
*ci
,
487 struct buffer_head
*bh
)
489 __ocfs2_init_extent_tree(et
, ci
, bh
, ocfs2_journal_access_rb
,
490 NULL
, &ocfs2_refcount_tree_et_ops
);
493 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree
*et
,
496 et
->et_ops
->eo_set_last_eb_blk(et
, new_last_eb_blk
);
499 static inline u64
ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree
*et
)
501 return et
->et_ops
->eo_get_last_eb_blk(et
);
504 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree
*et
,
507 et
->et_ops
->eo_update_clusters(et
, clusters
);
510 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree
*et
,
511 struct ocfs2_extent_rec
*rec
)
513 if (et
->et_ops
->eo_extent_map_insert
)
514 et
->et_ops
->eo_extent_map_insert(et
, rec
);
517 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree
*et
,
520 if (et
->et_ops
->eo_extent_map_truncate
)
521 et
->et_ops
->eo_extent_map_truncate(et
, clusters
);
524 static inline int ocfs2_et_root_journal_access(handle_t
*handle
,
525 struct ocfs2_extent_tree
*et
,
528 return et
->et_root_journal_access(handle
, et
->et_ci
, et
->et_root_bh
,
532 static inline enum ocfs2_contig_type
533 ocfs2_et_extent_contig(struct ocfs2_extent_tree
*et
,
534 struct ocfs2_extent_rec
*rec
,
535 struct ocfs2_extent_rec
*insert_rec
)
537 if (et
->et_ops
->eo_extent_contig
)
538 return et
->et_ops
->eo_extent_contig(et
, rec
, insert_rec
);
540 return ocfs2_extent_rec_contig(
541 ocfs2_metadata_cache_get_super(et
->et_ci
),
545 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree
*et
,
546 struct ocfs2_extent_rec
*rec
)
550 if (et
->et_ops
->eo_insert_check
)
551 ret
= et
->et_ops
->eo_insert_check(et
, rec
);
555 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree
*et
)
559 if (et
->et_ops
->eo_sanity_check
)
560 ret
= et
->et_ops
->eo_sanity_check(et
);
564 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt
*ctxt
,
565 struct ocfs2_extent_block
*eb
);
566 static void ocfs2_adjust_rightmost_records(handle_t
*handle
,
567 struct ocfs2_extent_tree
*et
,
568 struct ocfs2_path
*path
,
569 struct ocfs2_extent_rec
*insert_rec
);
571 * Reset the actual path elements so that we can re-use the structure
572 * to build another path. Generally, this involves freeing the buffer
575 void ocfs2_reinit_path(struct ocfs2_path
*path
, int keep_root
)
577 int i
, start
= 0, depth
= 0;
578 struct ocfs2_path_item
*node
;
583 for(i
= start
; i
< path_num_items(path
); i
++) {
584 node
= &path
->p_node
[i
];
592 * Tree depth may change during truncate, or insert. If we're
593 * keeping the root extent list, then make sure that our path
594 * structure reflects the proper depth.
597 depth
= le16_to_cpu(path_root_el(path
)->l_tree_depth
);
599 path_root_access(path
) = NULL
;
601 path
->p_tree_depth
= depth
;
604 void ocfs2_free_path(struct ocfs2_path
*path
)
607 ocfs2_reinit_path(path
, 0);
613 * All the elements of src into dest. After this call, src could be freed
614 * without affecting dest.
616 * Both paths should have the same root. Any non-root elements of dest
619 static void ocfs2_cp_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
623 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
624 BUG_ON(path_root_el(dest
) != path_root_el(src
));
625 BUG_ON(path_root_access(dest
) != path_root_access(src
));
627 ocfs2_reinit_path(dest
, 1);
629 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
630 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
631 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
633 if (dest
->p_node
[i
].bh
)
634 get_bh(dest
->p_node
[i
].bh
);
639 * Make the *dest path the same as src and re-initialize src path to
642 static void ocfs2_mv_path(struct ocfs2_path
*dest
, struct ocfs2_path
*src
)
646 BUG_ON(path_root_bh(dest
) != path_root_bh(src
));
647 BUG_ON(path_root_access(dest
) != path_root_access(src
));
649 for(i
= 1; i
< OCFS2_MAX_PATH_DEPTH
; i
++) {
650 brelse(dest
->p_node
[i
].bh
);
652 dest
->p_node
[i
].bh
= src
->p_node
[i
].bh
;
653 dest
->p_node
[i
].el
= src
->p_node
[i
].el
;
655 src
->p_node
[i
].bh
= NULL
;
656 src
->p_node
[i
].el
= NULL
;
661 * Insert an extent block at given index.
663 * This will not take an additional reference on eb_bh.
665 static inline void ocfs2_path_insert_eb(struct ocfs2_path
*path
, int index
,
666 struct buffer_head
*eb_bh
)
668 struct ocfs2_extent_block
*eb
= (struct ocfs2_extent_block
*)eb_bh
->b_data
;
671 * Right now, no root bh is an extent block, so this helps
672 * catch code errors with dinode trees. The assertion can be
673 * safely removed if we ever need to insert extent block
674 * structures at the root.
678 path
->p_node
[index
].bh
= eb_bh
;
679 path
->p_node
[index
].el
= &eb
->h_list
;
682 static struct ocfs2_path
*ocfs2_new_path(struct buffer_head
*root_bh
,
683 struct ocfs2_extent_list
*root_el
,
684 ocfs2_journal_access_func access
)
686 struct ocfs2_path
*path
;
688 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) >= OCFS2_MAX_PATH_DEPTH
);
690 path
= kzalloc(sizeof(*path
), GFP_NOFS
);
692 path
->p_tree_depth
= le16_to_cpu(root_el
->l_tree_depth
);
694 path_root_bh(path
) = root_bh
;
695 path_root_el(path
) = root_el
;
696 path_root_access(path
) = access
;
702 struct ocfs2_path
*ocfs2_new_path_from_path(struct ocfs2_path
*path
)
704 return ocfs2_new_path(path_root_bh(path
), path_root_el(path
),
705 path_root_access(path
));
708 struct ocfs2_path
*ocfs2_new_path_from_et(struct ocfs2_extent_tree
*et
)
710 return ocfs2_new_path(et
->et_root_bh
, et
->et_root_el
,
711 et
->et_root_journal_access
);
715 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
716 * otherwise it's the root_access function.
718 * I don't like the way this function's name looks next to
719 * ocfs2_journal_access_path(), but I don't have a better one.
721 int ocfs2_path_bh_journal_access(handle_t
*handle
,
722 struct ocfs2_caching_info
*ci
,
723 struct ocfs2_path
*path
,
726 ocfs2_journal_access_func access
= path_root_access(path
);
729 access
= ocfs2_journal_access
;
732 access
= ocfs2_journal_access_eb
;
734 return access(handle
, ci
, path
->p_node
[idx
].bh
,
735 OCFS2_JOURNAL_ACCESS_WRITE
);
739 * Convenience function to journal all components in a path.
741 int ocfs2_journal_access_path(struct ocfs2_caching_info
*ci
,
743 struct ocfs2_path
*path
)
750 for(i
= 0; i
< path_num_items(path
); i
++) {
751 ret
= ocfs2_path_bh_journal_access(handle
, ci
, path
, i
);
763 * Return the index of the extent record which contains cluster #v_cluster.
764 * -1 is returned if it was not found.
766 * Should work fine on interior and exterior nodes.
768 int ocfs2_search_extent_list(struct ocfs2_extent_list
*el
, u32 v_cluster
)
772 struct ocfs2_extent_rec
*rec
;
773 u32 rec_end
, rec_start
, clusters
;
775 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
776 rec
= &el
->l_recs
[i
];
778 rec_start
= le32_to_cpu(rec
->e_cpos
);
779 clusters
= ocfs2_rec_clusters(el
, rec
);
781 rec_end
= rec_start
+ clusters
;
783 if (v_cluster
>= rec_start
&& v_cluster
< rec_end
) {
793 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
794 * ocfs2_extent_rec_contig only work properly against leaf nodes!
796 static int ocfs2_block_extent_contig(struct super_block
*sb
,
797 struct ocfs2_extent_rec
*ext
,
800 u64 blk_end
= le64_to_cpu(ext
->e_blkno
);
802 blk_end
+= ocfs2_clusters_to_blocks(sb
,
803 le16_to_cpu(ext
->e_leaf_clusters
));
805 return blkno
== blk_end
;
808 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec
*left
,
809 struct ocfs2_extent_rec
*right
)
813 left_range
= le32_to_cpu(left
->e_cpos
) +
814 le16_to_cpu(left
->e_leaf_clusters
);
816 return (left_range
== le32_to_cpu(right
->e_cpos
));
819 static enum ocfs2_contig_type
820 ocfs2_extent_rec_contig(struct super_block
*sb
,
821 struct ocfs2_extent_rec
*ext
,
822 struct ocfs2_extent_rec
*insert_rec
)
824 u64 blkno
= le64_to_cpu(insert_rec
->e_blkno
);
827 * Refuse to coalesce extent records with different flag
828 * fields - we don't want to mix unwritten extents with user
831 if (ext
->e_flags
!= insert_rec
->e_flags
)
834 if (ocfs2_extents_adjacent(ext
, insert_rec
) &&
835 ocfs2_block_extent_contig(sb
, ext
, blkno
))
838 blkno
= le64_to_cpu(ext
->e_blkno
);
839 if (ocfs2_extents_adjacent(insert_rec
, ext
) &&
840 ocfs2_block_extent_contig(sb
, insert_rec
, blkno
))
847 * NOTE: We can have pretty much any combination of contiguousness and
850 * The usefulness of APPEND_TAIL is more in that it lets us know that
851 * we'll have to update the path to that leaf.
853 enum ocfs2_append_type
{
858 enum ocfs2_split_type
{
864 struct ocfs2_insert_type
{
865 enum ocfs2_split_type ins_split
;
866 enum ocfs2_append_type ins_appending
;
867 enum ocfs2_contig_type ins_contig
;
868 int ins_contig_index
;
872 struct ocfs2_merge_ctxt
{
873 enum ocfs2_contig_type c_contig_type
;
874 int c_has_empty_extent
;
875 int c_split_covers_rec
;
878 static int ocfs2_validate_extent_block(struct super_block
*sb
,
879 struct buffer_head
*bh
)
882 struct ocfs2_extent_block
*eb
=
883 (struct ocfs2_extent_block
*)bh
->b_data
;
885 trace_ocfs2_validate_extent_block((unsigned long long)bh
->b_blocknr
);
887 BUG_ON(!buffer_uptodate(bh
));
890 * If the ecc fails, we return the error but otherwise
891 * leave the filesystem running. We know any error is
892 * local to this block.
894 rc
= ocfs2_validate_meta_ecc(sb
, bh
->b_data
, &eb
->h_check
);
896 mlog(ML_ERROR
, "Checksum failed for extent block %llu\n",
897 (unsigned long long)bh
->b_blocknr
);
902 * Errors after here are fatal.
905 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb
)) {
907 "Extent block #%llu has bad signature %.*s\n",
908 (unsigned long long)bh
->b_blocknr
, 7,
913 if (le64_to_cpu(eb
->h_blkno
) != bh
->b_blocknr
) {
915 "Extent block #%llu has an invalid h_blkno of %llu\n",
916 (unsigned long long)bh
->b_blocknr
,
917 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
921 if (le32_to_cpu(eb
->h_fs_generation
) != OCFS2_SB(sb
)->fs_generation
)
923 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
924 (unsigned long long)bh
->b_blocknr
,
925 le32_to_cpu(eb
->h_fs_generation
));
930 int ocfs2_read_extent_block(struct ocfs2_caching_info
*ci
, u64 eb_blkno
,
931 struct buffer_head
**bh
)
934 struct buffer_head
*tmp
= *bh
;
936 rc
= ocfs2_read_block(ci
, eb_blkno
, &tmp
,
937 ocfs2_validate_extent_block
);
939 /* If ocfs2_read_block() got us a new bh, pass it up. */
948 * How many free extents have we got before we need more meta data?
950 int ocfs2_num_free_extents(struct ocfs2_extent_tree
*et
)
953 struct ocfs2_extent_list
*el
= NULL
;
954 struct ocfs2_extent_block
*eb
;
955 struct buffer_head
*eb_bh
= NULL
;
959 last_eb_blk
= ocfs2_et_get_last_eb_blk(et
);
962 retval
= ocfs2_read_extent_block(et
->et_ci
, last_eb_blk
,
968 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
972 BUG_ON(el
->l_tree_depth
!= 0);
974 retval
= le16_to_cpu(el
->l_count
) - le16_to_cpu(el
->l_next_free_rec
);
978 trace_ocfs2_num_free_extents(retval
);
982 /* expects array to already be allocated
984 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
987 static int ocfs2_create_new_meta_bhs(handle_t
*handle
,
988 struct ocfs2_extent_tree
*et
,
990 struct ocfs2_alloc_context
*meta_ac
,
991 struct buffer_head
*bhs
[])
993 int count
, status
, i
;
994 u16 suballoc_bit_start
;
996 u64 suballoc_loc
, first_blkno
;
997 struct ocfs2_super
*osb
=
998 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
999 struct ocfs2_extent_block
*eb
;
1002 while (count
< wanted
) {
1003 status
= ocfs2_claim_metadata(handle
,
1007 &suballoc_bit_start
,
1015 for(i
= count
; i
< (num_got
+ count
); i
++) {
1016 bhs
[i
] = sb_getblk(osb
->sb
, first_blkno
);
1017 if (bhs
[i
] == NULL
) {
1022 ocfs2_set_new_buffer_uptodate(et
->et_ci
, bhs
[i
]);
1024 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
1026 OCFS2_JOURNAL_ACCESS_CREATE
);
1032 memset(bhs
[i
]->b_data
, 0, osb
->sb
->s_blocksize
);
1033 eb
= (struct ocfs2_extent_block
*) bhs
[i
]->b_data
;
1034 /* Ok, setup the minimal stuff here. */
1035 strcpy(eb
->h_signature
, OCFS2_EXTENT_BLOCK_SIGNATURE
);
1036 eb
->h_blkno
= cpu_to_le64(first_blkno
);
1037 eb
->h_fs_generation
= cpu_to_le32(osb
->fs_generation
);
1038 eb
->h_suballoc_slot
=
1039 cpu_to_le16(meta_ac
->ac_alloc_slot
);
1040 eb
->h_suballoc_loc
= cpu_to_le64(suballoc_loc
);
1041 eb
->h_suballoc_bit
= cpu_to_le16(suballoc_bit_start
);
1042 eb
->h_list
.l_count
=
1043 cpu_to_le16(ocfs2_extent_recs_per_eb(osb
->sb
));
1045 suballoc_bit_start
++;
1048 /* We'll also be dirtied by the caller, so
1049 * this isn't absolutely necessary. */
1050 ocfs2_journal_dirty(handle
, bhs
[i
]);
1059 for(i
= 0; i
< wanted
; i
++) {
1068 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1070 * Returns the sum of the rightmost extent rec logical offset and
1073 * ocfs2_add_branch() uses this to determine what logical cluster
1074 * value should be populated into the leftmost new branch records.
1076 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1077 * value for the new topmost tree record.
1079 static inline u32
ocfs2_sum_rightmost_rec(struct ocfs2_extent_list
*el
)
1083 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1085 return le32_to_cpu(el
->l_recs
[i
].e_cpos
) +
1086 ocfs2_rec_clusters(el
, &el
->l_recs
[i
]);
1090 * Change range of the branches in the right most path according to the leaf
1091 * extent block's rightmost record.
1093 static int ocfs2_adjust_rightmost_branch(handle_t
*handle
,
1094 struct ocfs2_extent_tree
*et
)
1097 struct ocfs2_path
*path
= NULL
;
1098 struct ocfs2_extent_list
*el
;
1099 struct ocfs2_extent_rec
*rec
;
1101 path
= ocfs2_new_path_from_et(et
);
1107 status
= ocfs2_find_path(et
->et_ci
, path
, UINT_MAX
);
1113 status
= ocfs2_extend_trans(handle
, path_num_items(path
));
1119 status
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
1125 el
= path_leaf_el(path
);
1126 rec
= &el
->l_recs
[le16_to_cpu(el
->l_next_free_rec
) - 1];
1128 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
1131 ocfs2_free_path(path
);
1136 * Add an entire tree branch to our inode. eb_bh is the extent block
1137 * to start at, if we don't want to start the branch at the root
1140 * last_eb_bh is required as we have to update it's next_leaf pointer
1141 * for the new last extent block.
1143 * the new branch will be 'empty' in the sense that every block will
1144 * contain a single record with cluster count == 0.
1146 static int ocfs2_add_branch(handle_t
*handle
,
1147 struct ocfs2_extent_tree
*et
,
1148 struct buffer_head
*eb_bh
,
1149 struct buffer_head
**last_eb_bh
,
1150 struct ocfs2_alloc_context
*meta_ac
)
1152 int status
, new_blocks
, i
, block_given
= 0;
1153 u64 next_blkno
, new_last_eb_blk
;
1154 struct buffer_head
*bh
;
1155 struct buffer_head
**new_eb_bhs
= NULL
;
1156 struct ocfs2_extent_block
*eb
;
1157 struct ocfs2_extent_list
*eb_el
;
1158 struct ocfs2_extent_list
*el
;
1159 u32 new_cpos
, root_end
;
1161 BUG_ON(!last_eb_bh
|| !*last_eb_bh
);
1164 eb
= (struct ocfs2_extent_block
*) eb_bh
->b_data
;
1167 el
= et
->et_root_el
;
1169 /* we never add a branch to a leaf. */
1170 BUG_ON(!el
->l_tree_depth
);
1172 new_blocks
= le16_to_cpu(el
->l_tree_depth
);
1174 eb
= (struct ocfs2_extent_block
*)(*last_eb_bh
)->b_data
;
1175 new_cpos
= ocfs2_sum_rightmost_rec(&eb
->h_list
);
1176 root_end
= ocfs2_sum_rightmost_rec(et
->et_root_el
);
1179 * If there is a gap before the root end and the real end
1180 * of the righmost leaf block, we need to remove the gap
1181 * between new_cpos and root_end first so that the tree
1182 * is consistent after we add a new branch(it will start
1185 if (root_end
> new_cpos
) {
1186 trace_ocfs2_adjust_rightmost_branch(
1187 (unsigned long long)
1188 ocfs2_metadata_cache_owner(et
->et_ci
),
1189 root_end
, new_cpos
);
1191 status
= ocfs2_adjust_rightmost_branch(handle
, et
);
1198 /* allocate the number of new eb blocks we need */
1199 new_eb_bhs
= kcalloc(new_blocks
, sizeof(struct buffer_head
*),
1207 /* Firstyly, try to reuse dealloc since we have already estimated how
1208 * many extent blocks we may use.
1210 if (!ocfs2_is_dealloc_empty(et
)) {
1211 status
= ocfs2_reuse_blk_from_dealloc(handle
, et
,
1212 new_eb_bhs
, new_blocks
,
1220 BUG_ON(block_given
> new_blocks
);
1222 if (block_given
< new_blocks
) {
1224 status
= ocfs2_create_new_meta_bhs(handle
, et
,
1225 new_blocks
- block_given
,
1227 &new_eb_bhs
[block_given
]);
1234 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1235 * linked with the rest of the tree.
1236 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1238 * when we leave the loop, new_last_eb_blk will point to the
1239 * newest leaf, and next_blkno will point to the topmost extent
1241 next_blkno
= new_last_eb_blk
= 0;
1242 for(i
= 0; i
< new_blocks
; i
++) {
1244 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1245 /* ocfs2_create_new_meta_bhs() should create it right! */
1246 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1247 eb_el
= &eb
->h_list
;
1249 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, bh
,
1250 OCFS2_JOURNAL_ACCESS_CREATE
);
1256 eb
->h_next_leaf_blk
= 0;
1257 eb_el
->l_tree_depth
= cpu_to_le16(i
);
1258 eb_el
->l_next_free_rec
= cpu_to_le16(1);
1260 * This actually counts as an empty extent as
1263 eb_el
->l_recs
[0].e_cpos
= cpu_to_le32(new_cpos
);
1264 eb_el
->l_recs
[0].e_blkno
= cpu_to_le64(next_blkno
);
1266 * eb_el isn't always an interior node, but even leaf
1267 * nodes want a zero'd flags and reserved field so
1268 * this gets the whole 32 bits regardless of use.
1270 eb_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(0);
1271 if (!eb_el
->l_tree_depth
)
1272 new_last_eb_blk
= le64_to_cpu(eb
->h_blkno
);
1274 ocfs2_journal_dirty(handle
, bh
);
1275 next_blkno
= le64_to_cpu(eb
->h_blkno
);
1278 /* This is a bit hairy. We want to update up to three blocks
1279 * here without leaving any of them in an inconsistent state
1280 * in case of error. We don't have to worry about
1281 * journal_dirty erroring as it won't unless we've aborted the
1282 * handle (in which case we would never be here) so reserving
1283 * the write with journal_access is all we need to do. */
1284 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, *last_eb_bh
,
1285 OCFS2_JOURNAL_ACCESS_WRITE
);
1290 status
= ocfs2_et_root_journal_access(handle
, et
,
1291 OCFS2_JOURNAL_ACCESS_WRITE
);
1297 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, eb_bh
,
1298 OCFS2_JOURNAL_ACCESS_WRITE
);
1305 /* Link the new branch into the rest of the tree (el will
1306 * either be on the root_bh, or the extent block passed in. */
1307 i
= le16_to_cpu(el
->l_next_free_rec
);
1308 el
->l_recs
[i
].e_blkno
= cpu_to_le64(next_blkno
);
1309 el
->l_recs
[i
].e_cpos
= cpu_to_le32(new_cpos
);
1310 el
->l_recs
[i
].e_int_clusters
= 0;
1311 le16_add_cpu(&el
->l_next_free_rec
, 1);
1313 /* fe needs a new last extent block pointer, as does the
1314 * next_leaf on the previously last-extent-block. */
1315 ocfs2_et_set_last_eb_blk(et
, new_last_eb_blk
);
1317 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
1318 eb
->h_next_leaf_blk
= cpu_to_le64(new_last_eb_blk
);
1320 ocfs2_journal_dirty(handle
, *last_eb_bh
);
1321 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1323 ocfs2_journal_dirty(handle
, eb_bh
);
1326 * Some callers want to track the rightmost leaf so pass it
1329 brelse(*last_eb_bh
);
1330 get_bh(new_eb_bhs
[0]);
1331 *last_eb_bh
= new_eb_bhs
[0];
1336 for (i
= 0; i
< new_blocks
; i
++)
1337 brelse(new_eb_bhs
[i
]);
1345 * adds another level to the allocation tree.
1346 * returns back the new extent block so you can add a branch to it
1349 static int ocfs2_shift_tree_depth(handle_t
*handle
,
1350 struct ocfs2_extent_tree
*et
,
1351 struct ocfs2_alloc_context
*meta_ac
,
1352 struct buffer_head
**ret_new_eb_bh
)
1354 int status
, i
, block_given
= 0;
1356 struct buffer_head
*new_eb_bh
= NULL
;
1357 struct ocfs2_extent_block
*eb
;
1358 struct ocfs2_extent_list
*root_el
;
1359 struct ocfs2_extent_list
*eb_el
;
1361 if (!ocfs2_is_dealloc_empty(et
)) {
1362 status
= ocfs2_reuse_blk_from_dealloc(handle
, et
,
1365 } else if (meta_ac
) {
1366 status
= ocfs2_create_new_meta_bhs(handle
, et
, 1, meta_ac
,
1378 eb
= (struct ocfs2_extent_block
*) new_eb_bh
->b_data
;
1379 /* ocfs2_create_new_meta_bhs() should create it right! */
1380 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb
));
1382 eb_el
= &eb
->h_list
;
1383 root_el
= et
->et_root_el
;
1385 status
= ocfs2_journal_access_eb(handle
, et
->et_ci
, new_eb_bh
,
1386 OCFS2_JOURNAL_ACCESS_CREATE
);
1392 /* copy the root extent list data into the new extent block */
1393 eb_el
->l_tree_depth
= root_el
->l_tree_depth
;
1394 eb_el
->l_next_free_rec
= root_el
->l_next_free_rec
;
1395 for (i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1396 eb_el
->l_recs
[i
] = root_el
->l_recs
[i
];
1398 ocfs2_journal_dirty(handle
, new_eb_bh
);
1400 status
= ocfs2_et_root_journal_access(handle
, et
,
1401 OCFS2_JOURNAL_ACCESS_WRITE
);
1407 new_clusters
= ocfs2_sum_rightmost_rec(eb_el
);
1409 /* update root_bh now */
1410 le16_add_cpu(&root_el
->l_tree_depth
, 1);
1411 root_el
->l_recs
[0].e_cpos
= 0;
1412 root_el
->l_recs
[0].e_blkno
= eb
->h_blkno
;
1413 root_el
->l_recs
[0].e_int_clusters
= cpu_to_le32(new_clusters
);
1414 for (i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
1415 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
1416 root_el
->l_next_free_rec
= cpu_to_le16(1);
1418 /* If this is our 1st tree depth shift, then last_eb_blk
1419 * becomes the allocated extent block */
1420 if (root_el
->l_tree_depth
== cpu_to_le16(1))
1421 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
1423 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
1425 *ret_new_eb_bh
= new_eb_bh
;
1435 * Should only be called when there is no space left in any of the
1436 * leaf nodes. What we want to do is find the lowest tree depth
1437 * non-leaf extent block with room for new records. There are three
1438 * valid results of this search:
1440 * 1) a lowest extent block is found, then we pass it back in
1441 * *lowest_eb_bh and return '0'
1443 * 2) the search fails to find anything, but the root_el has room. We
1444 * pass NULL back in *lowest_eb_bh, but still return '0'
1446 * 3) the search fails to find anything AND the root_el is full, in
1447 * which case we return > 0
1449 * return status < 0 indicates an error.
1451 static int ocfs2_find_branch_target(struct ocfs2_extent_tree
*et
,
1452 struct buffer_head
**target_bh
)
1456 struct ocfs2_extent_block
*eb
;
1457 struct ocfs2_extent_list
*el
;
1458 struct buffer_head
*bh
= NULL
;
1459 struct buffer_head
*lowest_bh
= NULL
;
1463 el
= et
->et_root_el
;
1465 while(le16_to_cpu(el
->l_tree_depth
) > 1) {
1466 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1467 status
= ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1468 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1469 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
1472 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
1473 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1475 status
= ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
1476 "Owner %llu has extent list where extent # %d has no physical block start\n",
1477 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
), i
);
1484 status
= ocfs2_read_extent_block(et
->et_ci
, blkno
, &bh
);
1490 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1493 if (le16_to_cpu(el
->l_next_free_rec
) <
1494 le16_to_cpu(el
->l_count
)) {
1501 /* If we didn't find one and the fe doesn't have any room,
1502 * then return '1' */
1503 el
= et
->et_root_el
;
1504 if (!lowest_bh
&& (el
->l_next_free_rec
== el
->l_count
))
1507 *target_bh
= lowest_bh
;
1515 * Grow a b-tree so that it has more records.
1517 * We might shift the tree depth in which case existing paths should
1518 * be considered invalid.
1520 * Tree depth after the grow is returned via *final_depth.
1522 * *last_eb_bh will be updated by ocfs2_add_branch().
1524 static int ocfs2_grow_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
1525 int *final_depth
, struct buffer_head
**last_eb_bh
,
1526 struct ocfs2_alloc_context
*meta_ac
)
1529 struct ocfs2_extent_list
*el
= et
->et_root_el
;
1530 int depth
= le16_to_cpu(el
->l_tree_depth
);
1531 struct buffer_head
*bh
= NULL
;
1533 BUG_ON(meta_ac
== NULL
&& ocfs2_is_dealloc_empty(et
));
1535 shift
= ocfs2_find_branch_target(et
, &bh
);
1542 /* We traveled all the way to the bottom of the allocation tree
1543 * and didn't find room for any more extents - we need to add
1544 * another tree level */
1547 trace_ocfs2_grow_tree(
1548 (unsigned long long)
1549 ocfs2_metadata_cache_owner(et
->et_ci
),
1552 /* ocfs2_shift_tree_depth will return us a buffer with
1553 * the new extent block (so we can pass that to
1554 * ocfs2_add_branch). */
1555 ret
= ocfs2_shift_tree_depth(handle
, et
, meta_ac
, &bh
);
1563 * Special case: we have room now if we shifted from
1564 * tree_depth 0, so no more work needs to be done.
1566 * We won't be calling add_branch, so pass
1567 * back *last_eb_bh as the new leaf. At depth
1568 * zero, it should always be null so there's
1569 * no reason to brelse.
1571 BUG_ON(*last_eb_bh
);
1578 /* call ocfs2_add_branch to add the final part of the tree with
1580 ret
= ocfs2_add_branch(handle
, et
, bh
, last_eb_bh
,
1587 *final_depth
= depth
;
1593 * This function will discard the rightmost extent record.
1595 static void ocfs2_shift_records_right(struct ocfs2_extent_list
*el
)
1597 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1598 int count
= le16_to_cpu(el
->l_count
);
1599 unsigned int num_bytes
;
1602 /* This will cause us to go off the end of our extent list. */
1603 BUG_ON(next_free
>= count
);
1605 num_bytes
= sizeof(struct ocfs2_extent_rec
) * next_free
;
1607 memmove(&el
->l_recs
[1], &el
->l_recs
[0], num_bytes
);
1610 static void ocfs2_rotate_leaf(struct ocfs2_extent_list
*el
,
1611 struct ocfs2_extent_rec
*insert_rec
)
1613 int i
, insert_index
, next_free
, has_empty
, num_bytes
;
1614 u32 insert_cpos
= le32_to_cpu(insert_rec
->e_cpos
);
1615 struct ocfs2_extent_rec
*rec
;
1617 next_free
= le16_to_cpu(el
->l_next_free_rec
);
1618 has_empty
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
1622 /* The tree code before us didn't allow enough room in the leaf. */
1623 BUG_ON(el
->l_next_free_rec
== el
->l_count
&& !has_empty
);
1626 * The easiest way to approach this is to just remove the
1627 * empty extent and temporarily decrement next_free.
1631 * If next_free was 1 (only an empty extent), this
1632 * loop won't execute, which is fine. We still want
1633 * the decrement above to happen.
1635 for(i
= 0; i
< (next_free
- 1); i
++)
1636 el
->l_recs
[i
] = el
->l_recs
[i
+1];
1642 * Figure out what the new record index should be.
1644 for(i
= 0; i
< next_free
; i
++) {
1645 rec
= &el
->l_recs
[i
];
1647 if (insert_cpos
< le32_to_cpu(rec
->e_cpos
))
1652 trace_ocfs2_rotate_leaf(insert_cpos
, insert_index
,
1653 has_empty
, next_free
,
1654 le16_to_cpu(el
->l_count
));
1656 BUG_ON(insert_index
< 0);
1657 BUG_ON(insert_index
>= le16_to_cpu(el
->l_count
));
1658 BUG_ON(insert_index
> next_free
);
1661 * No need to memmove if we're just adding to the tail.
1663 if (insert_index
!= next_free
) {
1664 BUG_ON(next_free
>= le16_to_cpu(el
->l_count
));
1666 num_bytes
= next_free
- insert_index
;
1667 num_bytes
*= sizeof(struct ocfs2_extent_rec
);
1668 memmove(&el
->l_recs
[insert_index
+ 1],
1669 &el
->l_recs
[insert_index
],
1674 * Either we had an empty extent, and need to re-increment or
1675 * there was no empty extent on a non full rightmost leaf node,
1676 * in which case we still need to increment.
1679 el
->l_next_free_rec
= cpu_to_le16(next_free
);
1681 * Make sure none of the math above just messed up our tree.
1683 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) > le16_to_cpu(el
->l_count
));
1685 el
->l_recs
[insert_index
] = *insert_rec
;
1689 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list
*el
)
1691 int size
, num_recs
= le16_to_cpu(el
->l_next_free_rec
);
1693 BUG_ON(num_recs
== 0);
1695 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
1697 size
= num_recs
* sizeof(struct ocfs2_extent_rec
);
1698 memmove(&el
->l_recs
[0], &el
->l_recs
[1], size
);
1699 memset(&el
->l_recs
[num_recs
], 0,
1700 sizeof(struct ocfs2_extent_rec
));
1701 el
->l_next_free_rec
= cpu_to_le16(num_recs
);
1706 * Create an empty extent record .
1708 * l_next_free_rec may be updated.
1710 * If an empty extent already exists do nothing.
1712 static void ocfs2_create_empty_extent(struct ocfs2_extent_list
*el
)
1714 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
1716 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
1721 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
1724 mlog_bug_on_msg(el
->l_count
== el
->l_next_free_rec
,
1725 "Asked to create an empty extent in a full list:\n"
1726 "count = %u, tree depth = %u",
1727 le16_to_cpu(el
->l_count
),
1728 le16_to_cpu(el
->l_tree_depth
));
1730 ocfs2_shift_records_right(el
);
1733 le16_add_cpu(&el
->l_next_free_rec
, 1);
1734 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
1738 * For a rotation which involves two leaf nodes, the "root node" is
1739 * the lowest level tree node which contains a path to both leafs. This
1740 * resulting set of information can be used to form a complete "subtree"
1742 * This function is passed two full paths from the dinode down to a
1743 * pair of adjacent leaves. It's task is to figure out which path
1744 * index contains the subtree root - this can be the root index itself
1745 * in a worst-case rotation.
1747 * The array index of the subtree root is passed back.
1749 int ocfs2_find_subtree_root(struct ocfs2_extent_tree
*et
,
1750 struct ocfs2_path
*left
,
1751 struct ocfs2_path
*right
)
1756 * Check that the caller passed in two paths from the same tree.
1758 BUG_ON(path_root_bh(left
) != path_root_bh(right
));
1764 * The caller didn't pass two adjacent paths.
1766 mlog_bug_on_msg(i
> left
->p_tree_depth
,
1767 "Owner %llu, left depth %u, right depth %u\n"
1768 "left leaf blk %llu, right leaf blk %llu\n",
1769 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
1770 left
->p_tree_depth
, right
->p_tree_depth
,
1771 (unsigned long long)path_leaf_bh(left
)->b_blocknr
,
1772 (unsigned long long)path_leaf_bh(right
)->b_blocknr
);
1773 } while (left
->p_node
[i
].bh
->b_blocknr
==
1774 right
->p_node
[i
].bh
->b_blocknr
);
1779 typedef void (path_insert_t
)(void *, struct buffer_head
*);
1782 * Traverse a btree path in search of cpos, starting at root_el.
1784 * This code can be called with a cpos larger than the tree, in which
1785 * case it will return the rightmost path.
1787 static int __ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1788 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1789 path_insert_t
*func
, void *data
)
1794 struct buffer_head
*bh
= NULL
;
1795 struct ocfs2_extent_block
*eb
;
1796 struct ocfs2_extent_list
*el
;
1797 struct ocfs2_extent_rec
*rec
;
1800 while (el
->l_tree_depth
) {
1801 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
1802 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1803 "Owner %llu has empty extent list at depth %u\n",
1804 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1805 le16_to_cpu(el
->l_tree_depth
));
1811 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
) - 1; i
++) {
1812 rec
= &el
->l_recs
[i
];
1815 * In the case that cpos is off the allocation
1816 * tree, this should just wind up returning the
1819 range
= le32_to_cpu(rec
->e_cpos
) +
1820 ocfs2_rec_clusters(el
, rec
);
1821 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
1825 blkno
= le64_to_cpu(el
->l_recs
[i
].e_blkno
);
1827 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1828 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1829 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1830 le16_to_cpu(el
->l_tree_depth
), i
);
1837 ret
= ocfs2_read_extent_block(ci
, blkno
, &bh
);
1843 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
1846 if (le16_to_cpu(el
->l_next_free_rec
) >
1847 le16_to_cpu(el
->l_count
)) {
1848 ocfs2_error(ocfs2_metadata_cache_get_super(ci
),
1849 "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1850 (unsigned long long)ocfs2_metadata_cache_owner(ci
),
1851 (unsigned long long)bh
->b_blocknr
,
1852 le16_to_cpu(el
->l_next_free_rec
),
1853 le16_to_cpu(el
->l_count
));
1864 * Catch any trailing bh that the loop didn't handle.
1872 * Given an initialized path (that is, it has a valid root extent
1873 * list), this function will traverse the btree in search of the path
1874 * which would contain cpos.
1876 * The path traveled is recorded in the path structure.
1878 * Note that this will not do any comparisons on leaf node extent
1879 * records, so it will work fine in the case that we just added a tree
1882 struct find_path_data
{
1884 struct ocfs2_path
*path
;
1886 static void find_path_ins(void *data
, struct buffer_head
*bh
)
1888 struct find_path_data
*fp
= data
;
1891 ocfs2_path_insert_eb(fp
->path
, fp
->index
, bh
);
1894 int ocfs2_find_path(struct ocfs2_caching_info
*ci
,
1895 struct ocfs2_path
*path
, u32 cpos
)
1897 struct find_path_data data
;
1901 return __ocfs2_find_path(ci
, path_root_el(path
), cpos
,
1902 find_path_ins
, &data
);
1905 static void find_leaf_ins(void *data
, struct buffer_head
*bh
)
1907 struct ocfs2_extent_block
*eb
=(struct ocfs2_extent_block
*)bh
->b_data
;
1908 struct ocfs2_extent_list
*el
= &eb
->h_list
;
1909 struct buffer_head
**ret
= data
;
1911 /* We want to retain only the leaf block. */
1912 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
1918 * Find the leaf block in the tree which would contain cpos. No
1919 * checking of the actual leaf is done.
1921 * Some paths want to call this instead of allocating a path structure
1922 * and calling ocfs2_find_path().
1924 * This function doesn't handle non btree extent lists.
1926 int ocfs2_find_leaf(struct ocfs2_caching_info
*ci
,
1927 struct ocfs2_extent_list
*root_el
, u32 cpos
,
1928 struct buffer_head
**leaf_bh
)
1931 struct buffer_head
*bh
= NULL
;
1933 ret
= __ocfs2_find_path(ci
, root_el
, cpos
, find_leaf_ins
, &bh
);
1945 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1947 * Basically, we've moved stuff around at the bottom of the tree and
1948 * we need to fix up the extent records above the changes to reflect
1951 * left_rec: the record on the left.
1952 * right_rec: the record to the right of left_rec
1953 * right_child_el: is the child list pointed to by right_rec
1955 * By definition, this only works on interior nodes.
1957 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec
*left_rec
,
1958 struct ocfs2_extent_rec
*right_rec
,
1959 struct ocfs2_extent_list
*right_child_el
)
1961 u32 left_clusters
, right_end
;
1964 * Interior nodes never have holes. Their cpos is the cpos of
1965 * the leftmost record in their child list. Their cluster
1966 * count covers the full theoretical range of their child list
1967 * - the range between their cpos and the cpos of the record
1968 * immediately to their right.
1970 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[0].e_cpos
);
1971 if (!ocfs2_rec_clusters(right_child_el
, &right_child_el
->l_recs
[0])) {
1972 BUG_ON(right_child_el
->l_tree_depth
);
1973 BUG_ON(le16_to_cpu(right_child_el
->l_next_free_rec
) <= 1);
1974 left_clusters
= le32_to_cpu(right_child_el
->l_recs
[1].e_cpos
);
1976 left_clusters
-= le32_to_cpu(left_rec
->e_cpos
);
1977 left_rec
->e_int_clusters
= cpu_to_le32(left_clusters
);
1980 * Calculate the rightmost cluster count boundary before
1981 * moving cpos - we will need to adjust clusters after
1982 * updating e_cpos to keep the same highest cluster count.
1984 right_end
= le32_to_cpu(right_rec
->e_cpos
);
1985 right_end
+= le32_to_cpu(right_rec
->e_int_clusters
);
1987 right_rec
->e_cpos
= left_rec
->e_cpos
;
1988 le32_add_cpu(&right_rec
->e_cpos
, left_clusters
);
1990 right_end
-= le32_to_cpu(right_rec
->e_cpos
);
1991 right_rec
->e_int_clusters
= cpu_to_le32(right_end
);
1995 * Adjust the adjacent root node records involved in a
1996 * rotation. left_el_blkno is passed in as a key so that we can easily
1997 * find it's index in the root list.
1999 static void ocfs2_adjust_root_records(struct ocfs2_extent_list
*root_el
,
2000 struct ocfs2_extent_list
*left_el
,
2001 struct ocfs2_extent_list
*right_el
,
2006 BUG_ON(le16_to_cpu(root_el
->l_tree_depth
) <=
2007 le16_to_cpu(left_el
->l_tree_depth
));
2009 for(i
= 0; i
< le16_to_cpu(root_el
->l_next_free_rec
) - 1; i
++) {
2010 if (le64_to_cpu(root_el
->l_recs
[i
].e_blkno
) == left_el_blkno
)
2015 * The path walking code should have never returned a root and
2016 * two paths which are not adjacent.
2018 BUG_ON(i
>= (le16_to_cpu(root_el
->l_next_free_rec
) - 1));
2020 ocfs2_adjust_adjacent_records(&root_el
->l_recs
[i
],
2021 &root_el
->l_recs
[i
+ 1], right_el
);
2025 * We've changed a leaf block (in right_path) and need to reflect that
2026 * change back up the subtree.
2028 * This happens in multiple places:
2029 * - When we've moved an extent record from the left path leaf to the right
2030 * path leaf to make room for an empty extent in the left path leaf.
2031 * - When our insert into the right path leaf is at the leftmost edge
2032 * and requires an update of the path immediately to it's left. This
2033 * can occur at the end of some types of rotation and appending inserts.
2034 * - When we've adjusted the last extent record in the left path leaf and the
2035 * 1st extent record in the right path leaf during cross extent block merge.
2037 static void ocfs2_complete_edge_insert(handle_t
*handle
,
2038 struct ocfs2_path
*left_path
,
2039 struct ocfs2_path
*right_path
,
2043 struct ocfs2_extent_list
*el
, *left_el
, *right_el
;
2044 struct ocfs2_extent_rec
*left_rec
, *right_rec
;
2045 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2048 * Update the counts and position values within all the
2049 * interior nodes to reflect the leaf rotation we just did.
2051 * The root node is handled below the loop.
2053 * We begin the loop with right_el and left_el pointing to the
2054 * leaf lists and work our way up.
2056 * NOTE: within this loop, left_el and right_el always refer
2057 * to the *child* lists.
2059 left_el
= path_leaf_el(left_path
);
2060 right_el
= path_leaf_el(right_path
);
2061 for(i
= left_path
->p_tree_depth
- 1; i
> subtree_index
; i
--) {
2062 trace_ocfs2_complete_edge_insert(i
);
2065 * One nice property of knowing that all of these
2066 * nodes are below the root is that we only deal with
2067 * the leftmost right node record and the rightmost
2070 el
= left_path
->p_node
[i
].el
;
2071 idx
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2072 left_rec
= &el
->l_recs
[idx
];
2074 el
= right_path
->p_node
[i
].el
;
2075 right_rec
= &el
->l_recs
[0];
2077 ocfs2_adjust_adjacent_records(left_rec
, right_rec
, right_el
);
2079 ocfs2_journal_dirty(handle
, left_path
->p_node
[i
].bh
);
2080 ocfs2_journal_dirty(handle
, right_path
->p_node
[i
].bh
);
2083 * Setup our list pointers now so that the current
2084 * parents become children in the next iteration.
2086 left_el
= left_path
->p_node
[i
].el
;
2087 right_el
= right_path
->p_node
[i
].el
;
2091 * At the root node, adjust the two adjacent records which
2092 * begin our path to the leaves.
2095 el
= left_path
->p_node
[subtree_index
].el
;
2096 left_el
= left_path
->p_node
[subtree_index
+ 1].el
;
2097 right_el
= right_path
->p_node
[subtree_index
+ 1].el
;
2099 ocfs2_adjust_root_records(el
, left_el
, right_el
,
2100 left_path
->p_node
[subtree_index
+ 1].bh
->b_blocknr
);
2102 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2104 ocfs2_journal_dirty(handle
, root_bh
);
2107 static int ocfs2_rotate_subtree_right(handle_t
*handle
,
2108 struct ocfs2_extent_tree
*et
,
2109 struct ocfs2_path
*left_path
,
2110 struct ocfs2_path
*right_path
,
2114 struct buffer_head
*right_leaf_bh
;
2115 struct buffer_head
*left_leaf_bh
= NULL
;
2116 struct buffer_head
*root_bh
;
2117 struct ocfs2_extent_list
*right_el
, *left_el
;
2118 struct ocfs2_extent_rec move_rec
;
2120 left_leaf_bh
= path_leaf_bh(left_path
);
2121 left_el
= path_leaf_el(left_path
);
2123 if (left_el
->l_next_free_rec
!= left_el
->l_count
) {
2124 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
2125 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2126 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2127 (unsigned long long)left_leaf_bh
->b_blocknr
,
2128 le16_to_cpu(left_el
->l_next_free_rec
));
2133 * This extent block may already have an empty record, so we
2134 * return early if so.
2136 if (ocfs2_is_empty_extent(&left_el
->l_recs
[0]))
2139 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2140 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2142 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2149 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2150 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2157 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2165 right_leaf_bh
= path_leaf_bh(right_path
);
2166 right_el
= path_leaf_el(right_path
);
2168 /* This is a code error, not a disk corruption. */
2169 mlog_bug_on_msg(!right_el
->l_next_free_rec
, "Inode %llu: Rotate fails "
2170 "because rightmost leaf block %llu is empty\n",
2171 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2172 (unsigned long long)right_leaf_bh
->b_blocknr
);
2174 ocfs2_create_empty_extent(right_el
);
2176 ocfs2_journal_dirty(handle
, right_leaf_bh
);
2178 /* Do the copy now. */
2179 i
= le16_to_cpu(left_el
->l_next_free_rec
) - 1;
2180 move_rec
= left_el
->l_recs
[i
];
2181 right_el
->l_recs
[0] = move_rec
;
2184 * Clear out the record we just copied and shift everything
2185 * over, leaving an empty extent in the left leaf.
2187 * We temporarily subtract from next_free_rec so that the
2188 * shift will lose the tail record (which is now defunct).
2190 le16_add_cpu(&left_el
->l_next_free_rec
, -1);
2191 ocfs2_shift_records_right(left_el
);
2192 memset(&left_el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2193 le16_add_cpu(&left_el
->l_next_free_rec
, 1);
2195 ocfs2_journal_dirty(handle
, left_leaf_bh
);
2197 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2205 * Given a full path, determine what cpos value would return us a path
2206 * containing the leaf immediately to the left of the current one.
2208 * Will return zero if the path passed in is already the leftmost path.
2210 int ocfs2_find_cpos_for_left_leaf(struct super_block
*sb
,
2211 struct ocfs2_path
*path
, u32
*cpos
)
2215 struct ocfs2_extent_list
*el
;
2217 BUG_ON(path
->p_tree_depth
== 0);
2221 blkno
= path_leaf_bh(path
)->b_blocknr
;
2223 /* Start at the tree node just above the leaf and work our way up. */
2224 i
= path
->p_tree_depth
- 1;
2226 el
= path
->p_node
[i
].el
;
2229 * Find the extent record just before the one in our
2232 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2233 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2237 * We've determined that the
2238 * path specified is already
2239 * the leftmost one - return a
2245 * The leftmost record points to our
2246 * leaf - we need to travel up the
2252 *cpos
= le32_to_cpu(el
->l_recs
[j
- 1].e_cpos
);
2253 *cpos
= *cpos
+ ocfs2_rec_clusters(el
,
2254 &el
->l_recs
[j
- 1]);
2261 * If we got here, we never found a valid node where
2262 * the tree indicated one should be.
2264 ocfs2_error(sb
, "Invalid extent tree at extent block %llu\n",
2265 (unsigned long long)blkno
);
2270 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2279 * Extend the transaction by enough credits to complete the rotation,
2280 * and still leave at least the original number of credits allocated
2281 * to this transaction.
2283 static int ocfs2_extend_rotate_transaction(handle_t
*handle
, int subtree_depth
,
2285 struct ocfs2_path
*path
)
2288 int credits
= (path
->p_tree_depth
- subtree_depth
) * 2 + 1 + op_credits
;
2290 if (jbd2_handle_buffer_credits(handle
) < credits
)
2291 ret
= ocfs2_extend_trans(handle
,
2292 credits
- jbd2_handle_buffer_credits(handle
));
2298 * Trap the case where we're inserting into the theoretical range past
2299 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2300 * whose cpos is less than ours into the right leaf.
2302 * It's only necessary to look at the rightmost record of the left
2303 * leaf because the logic that calls us should ensure that the
2304 * theoretical ranges in the path components above the leaves are
2307 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path
*left_path
,
2310 struct ocfs2_extent_list
*left_el
;
2311 struct ocfs2_extent_rec
*rec
;
2314 left_el
= path_leaf_el(left_path
);
2315 next_free
= le16_to_cpu(left_el
->l_next_free_rec
);
2316 rec
= &left_el
->l_recs
[next_free
- 1];
2318 if (insert_cpos
> le32_to_cpu(rec
->e_cpos
))
2323 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list
*el
, u32 cpos
)
2325 int next_free
= le16_to_cpu(el
->l_next_free_rec
);
2327 struct ocfs2_extent_rec
*rec
;
2332 rec
= &el
->l_recs
[0];
2333 if (ocfs2_is_empty_extent(rec
)) {
2337 rec
= &el
->l_recs
[1];
2340 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2341 if (cpos
>= le32_to_cpu(rec
->e_cpos
) && cpos
< range
)
2347 * Rotate all the records in a btree right one record, starting at insert_cpos.
2349 * The path to the rightmost leaf should be passed in.
2351 * The array is assumed to be large enough to hold an entire path (tree depth).
2353 * Upon successful return from this function:
2355 * - The 'right_path' array will contain a path to the leaf block
2356 * whose range contains e_cpos.
2357 * - That leaf block will have a single empty extent in list index 0.
2358 * - In the case that the rotation requires a post-insert update,
2359 * *ret_left_path will contain a valid path which can be passed to
2360 * ocfs2_insert_path().
2362 static int ocfs2_rotate_tree_right(handle_t
*handle
,
2363 struct ocfs2_extent_tree
*et
,
2364 enum ocfs2_split_type split
,
2366 struct ocfs2_path
*right_path
,
2367 struct ocfs2_path
**ret_left_path
)
2369 int ret
, start
, orig_credits
= jbd2_handle_buffer_credits(handle
);
2371 struct ocfs2_path
*left_path
= NULL
;
2372 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2374 *ret_left_path
= NULL
;
2376 left_path
= ocfs2_new_path_from_path(right_path
);
2383 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2389 trace_ocfs2_rotate_tree_right(
2390 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2394 * What we want to do here is:
2396 * 1) Start with the rightmost path.
2398 * 2) Determine a path to the leaf block directly to the left
2401 * 3) Determine the 'subtree root' - the lowest level tree node
2402 * which contains a path to both leaves.
2404 * 4) Rotate the subtree.
2406 * 5) Find the next subtree by considering the left path to be
2407 * the new right path.
2409 * The check at the top of this while loop also accepts
2410 * insert_cpos == cpos because cpos is only a _theoretical_
2411 * value to get us the left path - insert_cpos might very well
2412 * be filling that hole.
2414 * Stop at a cpos of '0' because we either started at the
2415 * leftmost branch (i.e., a tree with one branch and a
2416 * rotation inside of it), or we've gone as far as we can in
2417 * rotating subtrees.
2419 while (cpos
&& insert_cpos
<= cpos
) {
2420 trace_ocfs2_rotate_tree_right(
2421 (unsigned long long)
2422 ocfs2_metadata_cache_owner(et
->et_ci
),
2425 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
2431 mlog_bug_on_msg(path_leaf_bh(left_path
) ==
2432 path_leaf_bh(right_path
),
2433 "Owner %llu: error during insert of %u "
2434 "(left path cpos %u) results in two identical "
2435 "paths ending at %llu\n",
2436 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2438 (unsigned long long)
2439 path_leaf_bh(left_path
)->b_blocknr
);
2441 if (split
== SPLIT_NONE
&&
2442 ocfs2_rotate_requires_path_adjustment(left_path
,
2446 * We've rotated the tree as much as we
2447 * should. The rest is up to
2448 * ocfs2_insert_path() to complete, after the
2449 * record insertion. We indicate this
2450 * situation by returning the left path.
2452 * The reason we don't adjust the records here
2453 * before the record insert is that an error
2454 * later might break the rule where a parent
2455 * record e_cpos will reflect the actual
2456 * e_cpos of the 1st nonempty record of the
2459 *ret_left_path
= left_path
;
2463 start
= ocfs2_find_subtree_root(et
, left_path
, right_path
);
2465 trace_ocfs2_rotate_subtree(start
,
2466 (unsigned long long)
2467 right_path
->p_node
[start
].bh
->b_blocknr
,
2468 right_path
->p_tree_depth
);
2470 ret
= ocfs2_extend_rotate_transaction(handle
, start
,
2471 orig_credits
, right_path
);
2477 ret
= ocfs2_rotate_subtree_right(handle
, et
, left_path
,
2484 if (split
!= SPLIT_NONE
&&
2485 ocfs2_leftmost_rec_contains(path_leaf_el(right_path
),
2488 * A rotate moves the rightmost left leaf
2489 * record over to the leftmost right leaf
2490 * slot. If we're doing an extent split
2491 * instead of a real insert, then we have to
2492 * check that the extent to be split wasn't
2493 * just moved over. If it was, then we can
2494 * exit here, passing left_path back -
2495 * ocfs2_split_extent() is smart enough to
2496 * search both leaves.
2498 *ret_left_path
= left_path
;
2503 * There is no need to re-read the next right path
2504 * as we know that it'll be our current left
2505 * path. Optimize by copying values instead.
2507 ocfs2_mv_path(right_path
, left_path
);
2509 ret
= ocfs2_find_cpos_for_left_leaf(sb
, right_path
, &cpos
);
2517 ocfs2_free_path(left_path
);
2523 static int ocfs2_update_edge_lengths(handle_t
*handle
,
2524 struct ocfs2_extent_tree
*et
,
2525 struct ocfs2_path
*path
)
2528 struct ocfs2_extent_rec
*rec
;
2529 struct ocfs2_extent_list
*el
;
2530 struct ocfs2_extent_block
*eb
;
2533 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
2539 /* Path should always be rightmost. */
2540 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
2541 BUG_ON(eb
->h_next_leaf_blk
!= 0ULL);
2544 BUG_ON(le16_to_cpu(el
->l_next_free_rec
) == 0);
2545 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2546 rec
= &el
->l_recs
[idx
];
2547 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
2549 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
2550 el
= path
->p_node
[i
].el
;
2551 idx
= le16_to_cpu(el
->l_next_free_rec
) - 1;
2552 rec
= &el
->l_recs
[idx
];
2554 rec
->e_int_clusters
= cpu_to_le32(range
);
2555 le32_add_cpu(&rec
->e_int_clusters
, -le32_to_cpu(rec
->e_cpos
));
2557 ocfs2_journal_dirty(handle
, path
->p_node
[i
].bh
);
2563 static void ocfs2_unlink_path(handle_t
*handle
,
2564 struct ocfs2_extent_tree
*et
,
2565 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2566 struct ocfs2_path
*path
, int unlink_start
)
2569 struct ocfs2_extent_block
*eb
;
2570 struct ocfs2_extent_list
*el
;
2571 struct buffer_head
*bh
;
2573 for(i
= unlink_start
; i
< path_num_items(path
); i
++) {
2574 bh
= path
->p_node
[i
].bh
;
2576 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
2578 * Not all nodes might have had their final count
2579 * decremented by the caller - handle this here.
2582 if (le16_to_cpu(el
->l_next_free_rec
) > 1) {
2584 "Inode %llu, attempted to remove extent block "
2585 "%llu with %u records\n",
2586 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
2587 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
2588 le16_to_cpu(el
->l_next_free_rec
));
2590 ocfs2_journal_dirty(handle
, bh
);
2591 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2595 el
->l_next_free_rec
= 0;
2596 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
2598 ocfs2_journal_dirty(handle
, bh
);
2600 ret
= ocfs2_cache_extent_block_free(dealloc
, eb
);
2604 ocfs2_remove_from_cache(et
->et_ci
, bh
);
2608 static void ocfs2_unlink_subtree(handle_t
*handle
,
2609 struct ocfs2_extent_tree
*et
,
2610 struct ocfs2_path
*left_path
,
2611 struct ocfs2_path
*right_path
,
2613 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
2616 struct buffer_head
*root_bh
= left_path
->p_node
[subtree_index
].bh
;
2617 struct ocfs2_extent_list
*root_el
= left_path
->p_node
[subtree_index
].el
;
2618 struct ocfs2_extent_block
*eb
;
2620 eb
= (struct ocfs2_extent_block
*)right_path
->p_node
[subtree_index
+ 1].bh
->b_data
;
2622 for(i
= 1; i
< le16_to_cpu(root_el
->l_next_free_rec
); i
++)
2623 if (root_el
->l_recs
[i
].e_blkno
== eb
->h_blkno
)
2626 BUG_ON(i
>= le16_to_cpu(root_el
->l_next_free_rec
));
2628 memset(&root_el
->l_recs
[i
], 0, sizeof(struct ocfs2_extent_rec
));
2629 le16_add_cpu(&root_el
->l_next_free_rec
, -1);
2631 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2632 eb
->h_next_leaf_blk
= 0;
2634 ocfs2_journal_dirty(handle
, root_bh
);
2635 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2637 ocfs2_unlink_path(handle
, et
, dealloc
, right_path
,
2641 static int ocfs2_rotate_subtree_left(handle_t
*handle
,
2642 struct ocfs2_extent_tree
*et
,
2643 struct ocfs2_path
*left_path
,
2644 struct ocfs2_path
*right_path
,
2646 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2649 int ret
, i
, del_right_subtree
= 0, right_has_empty
= 0;
2650 struct buffer_head
*root_bh
, *et_root_bh
= path_root_bh(right_path
);
2651 struct ocfs2_extent_list
*right_leaf_el
, *left_leaf_el
;
2652 struct ocfs2_extent_block
*eb
;
2656 right_leaf_el
= path_leaf_el(right_path
);
2657 left_leaf_el
= path_leaf_el(left_path
);
2658 root_bh
= left_path
->p_node
[subtree_index
].bh
;
2659 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
2661 if (!ocfs2_is_empty_extent(&left_leaf_el
->l_recs
[0]))
2664 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(right_path
)->b_data
;
2665 if (ocfs2_is_empty_extent(&right_leaf_el
->l_recs
[0])) {
2667 * It's legal for us to proceed if the right leaf is
2668 * the rightmost one and it has an empty extent. There
2669 * are two cases to handle - whether the leaf will be
2670 * empty after removal or not. If the leaf isn't empty
2671 * then just remove the empty extent up front. The
2672 * next block will handle empty leaves by flagging
2675 * Non rightmost leaves will throw -EAGAIN and the
2676 * caller can manually move the subtree and retry.
2679 if (eb
->h_next_leaf_blk
!= 0ULL)
2682 if (le16_to_cpu(right_leaf_el
->l_next_free_rec
) > 1) {
2683 ret
= ocfs2_journal_access_eb(handle
, et
->et_ci
,
2684 path_leaf_bh(right_path
),
2685 OCFS2_JOURNAL_ACCESS_WRITE
);
2691 ocfs2_remove_empty_extent(right_leaf_el
);
2693 right_has_empty
= 1;
2696 if (eb
->h_next_leaf_blk
== 0ULL &&
2697 le16_to_cpu(right_leaf_el
->l_next_free_rec
) == 1) {
2699 * We have to update i_last_eb_blk during the meta
2702 ret
= ocfs2_et_root_journal_access(handle
, et
,
2703 OCFS2_JOURNAL_ACCESS_WRITE
);
2709 del_right_subtree
= 1;
2713 * Getting here with an empty extent in the right path implies
2714 * that it's the rightmost path and will be deleted.
2716 BUG_ON(right_has_empty
&& !del_right_subtree
);
2718 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
2725 for(i
= subtree_index
+ 1; i
< path_num_items(right_path
); i
++) {
2726 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2733 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2741 if (!right_has_empty
) {
2743 * Only do this if we're moving a real
2744 * record. Otherwise, the action is delayed until
2745 * after removal of the right path in which case we
2746 * can do a simple shift to remove the empty extent.
2748 ocfs2_rotate_leaf(left_leaf_el
, &right_leaf_el
->l_recs
[0]);
2749 memset(&right_leaf_el
->l_recs
[0], 0,
2750 sizeof(struct ocfs2_extent_rec
));
2752 if (eb
->h_next_leaf_blk
== 0ULL) {
2754 * Move recs over to get rid of empty extent, decrease
2755 * next_free. This is allowed to remove the last
2756 * extent in our leaf (setting l_next_free_rec to
2757 * zero) - the delete code below won't care.
2759 ocfs2_remove_empty_extent(right_leaf_el
);
2762 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
2763 ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
2765 if (del_right_subtree
) {
2766 ocfs2_unlink_subtree(handle
, et
, left_path
, right_path
,
2767 subtree_index
, dealloc
);
2768 ret
= ocfs2_update_edge_lengths(handle
, et
, left_path
);
2774 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
2775 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
2778 * Removal of the extent in the left leaf was skipped
2779 * above so we could delete the right path
2782 if (right_has_empty
)
2783 ocfs2_remove_empty_extent(left_leaf_el
);
2785 ocfs2_journal_dirty(handle
, et_root_bh
);
2789 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
2797 * Given a full path, determine what cpos value would return us a path
2798 * containing the leaf immediately to the right of the current one.
2800 * Will return zero if the path passed in is already the rightmost path.
2802 * This looks similar, but is subtly different to
2803 * ocfs2_find_cpos_for_left_leaf().
2805 int ocfs2_find_cpos_for_right_leaf(struct super_block
*sb
,
2806 struct ocfs2_path
*path
, u32
*cpos
)
2810 struct ocfs2_extent_list
*el
;
2814 if (path
->p_tree_depth
== 0)
2817 blkno
= path_leaf_bh(path
)->b_blocknr
;
2819 /* Start at the tree node just above the leaf and work our way up. */
2820 i
= path
->p_tree_depth
- 1;
2824 el
= path
->p_node
[i
].el
;
2827 * Find the extent record just after the one in our
2830 next_free
= le16_to_cpu(el
->l_next_free_rec
);
2831 for(j
= 0; j
< le16_to_cpu(el
->l_next_free_rec
); j
++) {
2832 if (le64_to_cpu(el
->l_recs
[j
].e_blkno
) == blkno
) {
2833 if (j
== (next_free
- 1)) {
2836 * We've determined that the
2837 * path specified is already
2838 * the rightmost one - return a
2844 * The rightmost record points to our
2845 * leaf - we need to travel up the
2851 *cpos
= le32_to_cpu(el
->l_recs
[j
+ 1].e_cpos
);
2857 * If we got here, we never found a valid node where
2858 * the tree indicated one should be.
2860 ocfs2_error(sb
, "Invalid extent tree at extent block %llu\n",
2861 (unsigned long long)blkno
);
2866 blkno
= path
->p_node
[i
].bh
->b_blocknr
;
2874 static int ocfs2_rotate_rightmost_leaf_left(handle_t
*handle
,
2875 struct ocfs2_extent_tree
*et
,
2876 struct ocfs2_path
*path
)
2879 struct buffer_head
*bh
= path_leaf_bh(path
);
2880 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
2882 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
2885 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
2886 path_num_items(path
) - 1);
2892 ocfs2_remove_empty_extent(el
);
2893 ocfs2_journal_dirty(handle
, bh
);
2899 static int __ocfs2_rotate_tree_left(handle_t
*handle
,
2900 struct ocfs2_extent_tree
*et
,
2902 struct ocfs2_path
*path
,
2903 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
2904 struct ocfs2_path
**empty_extent_path
)
2906 int ret
, subtree_root
, deleted
;
2908 struct ocfs2_path
*left_path
= NULL
;
2909 struct ocfs2_path
*right_path
= NULL
;
2910 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
2912 if (!ocfs2_is_empty_extent(&(path_leaf_el(path
)->l_recs
[0])))
2915 *empty_extent_path
= NULL
;
2917 ret
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
2923 left_path
= ocfs2_new_path_from_path(path
);
2930 ocfs2_cp_path(left_path
, path
);
2932 right_path
= ocfs2_new_path_from_path(path
);
2939 while (right_cpos
) {
2940 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
2946 subtree_root
= ocfs2_find_subtree_root(et
, left_path
,
2949 trace_ocfs2_rotate_subtree(subtree_root
,
2950 (unsigned long long)
2951 right_path
->p_node
[subtree_root
].bh
->b_blocknr
,
2952 right_path
->p_tree_depth
);
2954 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
2955 orig_credits
, left_path
);
2962 * Caller might still want to make changes to the
2963 * tree root, so re-add it to the journal here.
2965 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
2972 ret
= ocfs2_rotate_subtree_left(handle
, et
, left_path
,
2973 right_path
, subtree_root
,
2975 if (ret
== -EAGAIN
) {
2977 * The rotation has to temporarily stop due to
2978 * the right subtree having an empty
2979 * extent. Pass it back to the caller for a
2982 *empty_extent_path
= right_path
;
2992 * The subtree rotate might have removed records on
2993 * the rightmost edge. If so, then rotation is
2999 ocfs2_mv_path(left_path
, right_path
);
3001 ret
= ocfs2_find_cpos_for_right_leaf(sb
, left_path
,
3010 ocfs2_free_path(right_path
);
3011 ocfs2_free_path(left_path
);
3016 static int ocfs2_remove_rightmost_path(handle_t
*handle
,
3017 struct ocfs2_extent_tree
*et
,
3018 struct ocfs2_path
*path
,
3019 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3021 int ret
, subtree_index
;
3023 struct ocfs2_path
*left_path
= NULL
;
3024 struct ocfs2_extent_block
*eb
;
3025 struct ocfs2_extent_list
*el
;
3027 ret
= ocfs2_et_sanity_check(et
);
3031 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
3037 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3046 * We have a path to the left of this one - it needs
3049 left_path
= ocfs2_new_path_from_path(path
);
3056 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
3062 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
3068 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
3070 ocfs2_unlink_subtree(handle
, et
, left_path
, path
,
3071 subtree_index
, dealloc
);
3072 ret
= ocfs2_update_edge_lengths(handle
, et
, left_path
);
3078 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(left_path
)->b_data
;
3079 ocfs2_et_set_last_eb_blk(et
, le64_to_cpu(eb
->h_blkno
));
3082 * 'path' is also the leftmost path which
3083 * means it must be the only one. This gets
3084 * handled differently because we want to
3085 * revert the root back to having extents
3088 ocfs2_unlink_path(handle
, et
, dealloc
, path
, 1);
3090 el
= et
->et_root_el
;
3091 el
->l_tree_depth
= 0;
3092 el
->l_next_free_rec
= 0;
3093 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3095 ocfs2_et_set_last_eb_blk(et
, 0);
3098 ocfs2_journal_dirty(handle
, path_root_bh(path
));
3101 ocfs2_free_path(left_path
);
3105 static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super
*osb
,
3106 struct ocfs2_extent_tree
*et
,
3107 struct ocfs2_path
*path
,
3108 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3112 int credits
= path
->p_tree_depth
* 2 + 1;
3114 handle
= ocfs2_start_trans(osb
, credits
);
3115 if (IS_ERR(handle
)) {
3116 ret
= PTR_ERR(handle
);
3121 ret
= ocfs2_remove_rightmost_path(handle
, et
, path
, dealloc
);
3125 ocfs2_commit_trans(osb
, handle
);
3130 * Left rotation of btree records.
3132 * In many ways, this is (unsurprisingly) the opposite of right
3133 * rotation. We start at some non-rightmost path containing an empty
3134 * extent in the leaf block. The code works its way to the rightmost
3135 * path by rotating records to the left in every subtree.
3137 * This is used by any code which reduces the number of extent records
3138 * in a leaf. After removal, an empty record should be placed in the
3139 * leftmost list position.
3141 * This won't handle a length update of the rightmost path records if
3142 * the rightmost tree leaf record is removed so the caller is
3143 * responsible for detecting and correcting that.
3145 static int ocfs2_rotate_tree_left(handle_t
*handle
,
3146 struct ocfs2_extent_tree
*et
,
3147 struct ocfs2_path
*path
,
3148 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
3150 int ret
, orig_credits
= jbd2_handle_buffer_credits(handle
);
3151 struct ocfs2_path
*tmp_path
= NULL
, *restart_path
= NULL
;
3152 struct ocfs2_extent_block
*eb
;
3153 struct ocfs2_extent_list
*el
;
3155 el
= path_leaf_el(path
);
3156 if (!ocfs2_is_empty_extent(&el
->l_recs
[0]))
3159 if (path
->p_tree_depth
== 0) {
3160 rightmost_no_delete
:
3162 * Inline extents. This is trivially handled, so do
3165 ret
= ocfs2_rotate_rightmost_leaf_left(handle
, et
, path
);
3172 * Handle rightmost branch now. There's several cases:
3173 * 1) simple rotation leaving records in there. That's trivial.
3174 * 2) rotation requiring a branch delete - there's no more
3175 * records left. Two cases of this:
3176 * a) There are branches to the left.
3177 * b) This is also the leftmost (the only) branch.
3179 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3180 * 2a) we need the left branch so that we can update it with the unlink
3181 * 2b) we need to bring the root back to inline extents.
3184 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
3186 if (eb
->h_next_leaf_blk
== 0) {
3188 * This gets a bit tricky if we're going to delete the
3189 * rightmost path. Get the other cases out of the way
3192 if (le16_to_cpu(el
->l_next_free_rec
) > 1)
3193 goto rightmost_no_delete
;
3195 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
3196 ret
= ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
3197 "Owner %llu has empty extent block at %llu\n",
3198 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
3199 (unsigned long long)le64_to_cpu(eb
->h_blkno
));
3204 * XXX: The caller can not trust "path" any more after
3205 * this as it will have been deleted. What do we do?
3207 * In theory the rotate-for-merge code will never get
3208 * here because it'll always ask for a rotate in a
3212 ret
= ocfs2_remove_rightmost_path(handle
, et
, path
,
3220 * Now we can loop, remembering the path we get from -EAGAIN
3221 * and restarting from there.
3224 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
, path
,
3225 dealloc
, &restart_path
);
3226 if (ret
&& ret
!= -EAGAIN
) {
3231 while (ret
== -EAGAIN
) {
3232 tmp_path
= restart_path
;
3233 restart_path
= NULL
;
3235 ret
= __ocfs2_rotate_tree_left(handle
, et
, orig_credits
,
3238 if (ret
&& ret
!= -EAGAIN
) {
3243 ocfs2_free_path(tmp_path
);
3251 ocfs2_free_path(tmp_path
);
3252 ocfs2_free_path(restart_path
);
3256 static void ocfs2_cleanup_merge(struct ocfs2_extent_list
*el
,
3259 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[index
];
3262 if (rec
->e_leaf_clusters
== 0) {
3264 * We consumed all of the merged-from record. An empty
3265 * extent cannot exist anywhere but the 1st array
3266 * position, so move things over if the merged-from
3267 * record doesn't occupy that position.
3269 * This creates a new empty extent so the caller
3270 * should be smart enough to have removed any existing
3274 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
3275 size
= index
* sizeof(struct ocfs2_extent_rec
);
3276 memmove(&el
->l_recs
[1], &el
->l_recs
[0], size
);
3280 * Always memset - the caller doesn't check whether it
3281 * created an empty extent, so there could be junk in
3284 memset(&el
->l_recs
[0], 0, sizeof(struct ocfs2_extent_rec
));
3288 static int ocfs2_get_right_path(struct ocfs2_extent_tree
*et
,
3289 struct ocfs2_path
*left_path
,
3290 struct ocfs2_path
**ret_right_path
)
3294 struct ocfs2_path
*right_path
= NULL
;
3295 struct ocfs2_extent_list
*left_el
;
3297 *ret_right_path
= NULL
;
3299 /* This function shouldn't be called for non-trees. */
3300 BUG_ON(left_path
->p_tree_depth
== 0);
3302 left_el
= path_leaf_el(left_path
);
3303 BUG_ON(left_el
->l_next_free_rec
!= left_el
->l_count
);
3305 ret
= ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3306 left_path
, &right_cpos
);
3312 /* This function shouldn't be called for the rightmost leaf. */
3313 BUG_ON(right_cpos
== 0);
3315 right_path
= ocfs2_new_path_from_path(left_path
);
3322 ret
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
3328 *ret_right_path
= right_path
;
3331 ocfs2_free_path(right_path
);
3336 * Remove split_rec clusters from the record at index and merge them
3337 * onto the beginning of the record "next" to it.
3338 * For index < l_count - 1, the next means the extent rec at index + 1.
3339 * For index == l_count - 1, the "next" means the 1st extent rec of the
3340 * next extent block.
3342 static int ocfs2_merge_rec_right(struct ocfs2_path
*left_path
,
3344 struct ocfs2_extent_tree
*et
,
3345 struct ocfs2_extent_rec
*split_rec
,
3348 int ret
, next_free
, i
;
3349 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3350 struct ocfs2_extent_rec
*left_rec
;
3351 struct ocfs2_extent_rec
*right_rec
;
3352 struct ocfs2_extent_list
*right_el
;
3353 struct ocfs2_path
*right_path
= NULL
;
3354 int subtree_index
= 0;
3355 struct ocfs2_extent_list
*el
= path_leaf_el(left_path
);
3356 struct buffer_head
*bh
= path_leaf_bh(left_path
);
3357 struct buffer_head
*root_bh
= NULL
;
3359 BUG_ON(index
>= le16_to_cpu(el
->l_next_free_rec
));
3360 left_rec
= &el
->l_recs
[index
];
3362 if (index
== le16_to_cpu(el
->l_next_free_rec
) - 1 &&
3363 le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
)) {
3364 /* we meet with a cross extent block merge. */
3365 ret
= ocfs2_get_right_path(et
, left_path
, &right_path
);
3371 right_el
= path_leaf_el(right_path
);
3372 next_free
= le16_to_cpu(right_el
->l_next_free_rec
);
3373 BUG_ON(next_free
<= 0);
3374 right_rec
= &right_el
->l_recs
[0];
3375 if (ocfs2_is_empty_extent(right_rec
)) {
3376 BUG_ON(next_free
<= 1);
3377 right_rec
= &right_el
->l_recs
[1];
3380 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3381 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3382 le32_to_cpu(right_rec
->e_cpos
));
3384 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3387 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3388 jbd2_handle_buffer_credits(handle
),
3395 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3396 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3398 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3405 for (i
= subtree_index
+ 1;
3406 i
< path_num_items(right_path
); i
++) {
3407 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3414 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3423 BUG_ON(index
== le16_to_cpu(el
->l_next_free_rec
) - 1);
3424 right_rec
= &el
->l_recs
[index
+ 1];
3427 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, left_path
,
3428 path_num_items(left_path
) - 1);
3434 le16_add_cpu(&left_rec
->e_leaf_clusters
, -split_clusters
);
3436 le32_add_cpu(&right_rec
->e_cpos
, -split_clusters
);
3437 le64_add_cpu(&right_rec
->e_blkno
,
3438 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3440 le16_add_cpu(&right_rec
->e_leaf_clusters
, split_clusters
);
3442 ocfs2_cleanup_merge(el
, index
);
3444 ocfs2_journal_dirty(handle
, bh
);
3446 ocfs2_journal_dirty(handle
, path_leaf_bh(right_path
));
3447 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
3451 ocfs2_free_path(right_path
);
3455 static int ocfs2_get_left_path(struct ocfs2_extent_tree
*et
,
3456 struct ocfs2_path
*right_path
,
3457 struct ocfs2_path
**ret_left_path
)
3461 struct ocfs2_path
*left_path
= NULL
;
3463 *ret_left_path
= NULL
;
3465 /* This function shouldn't be called for non-trees. */
3466 BUG_ON(right_path
->p_tree_depth
== 0);
3468 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
3469 right_path
, &left_cpos
);
3475 /* This function shouldn't be called for the leftmost leaf. */
3476 BUG_ON(left_cpos
== 0);
3478 left_path
= ocfs2_new_path_from_path(right_path
);
3485 ret
= ocfs2_find_path(et
->et_ci
, left_path
, left_cpos
);
3491 *ret_left_path
= left_path
;
3494 ocfs2_free_path(left_path
);
3499 * Remove split_rec clusters from the record at index and merge them
3500 * onto the tail of the record "before" it.
3501 * For index > 0, the "before" means the extent rec at index - 1.
3503 * For index == 0, the "before" means the last record of the previous
3504 * extent block. And there is also a situation that we may need to
3505 * remove the rightmost leaf extent block in the right_path and change
3506 * the right path to indicate the new rightmost path.
3508 static int ocfs2_merge_rec_left(struct ocfs2_path
*right_path
,
3510 struct ocfs2_extent_tree
*et
,
3511 struct ocfs2_extent_rec
*split_rec
,
3512 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3515 int ret
, i
, subtree_index
= 0, has_empty_extent
= 0;
3516 unsigned int split_clusters
= le16_to_cpu(split_rec
->e_leaf_clusters
);
3517 struct ocfs2_extent_rec
*left_rec
;
3518 struct ocfs2_extent_rec
*right_rec
;
3519 struct ocfs2_extent_list
*el
= path_leaf_el(right_path
);
3520 struct buffer_head
*bh
= path_leaf_bh(right_path
);
3521 struct buffer_head
*root_bh
= NULL
;
3522 struct ocfs2_path
*left_path
= NULL
;
3523 struct ocfs2_extent_list
*left_el
;
3527 right_rec
= &el
->l_recs
[index
];
3529 /* we meet with a cross extent block merge. */
3530 ret
= ocfs2_get_left_path(et
, right_path
, &left_path
);
3536 left_el
= path_leaf_el(left_path
);
3537 BUG_ON(le16_to_cpu(left_el
->l_next_free_rec
) !=
3538 le16_to_cpu(left_el
->l_count
));
3540 left_rec
= &left_el
->l_recs
[
3541 le16_to_cpu(left_el
->l_next_free_rec
) - 1];
3542 BUG_ON(le32_to_cpu(left_rec
->e_cpos
) +
3543 le16_to_cpu(left_rec
->e_leaf_clusters
) !=
3544 le32_to_cpu(split_rec
->e_cpos
));
3546 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
3549 ret
= ocfs2_extend_rotate_transaction(handle
, subtree_index
,
3550 jbd2_handle_buffer_credits(handle
),
3557 root_bh
= left_path
->p_node
[subtree_index
].bh
;
3558 BUG_ON(root_bh
!= right_path
->p_node
[subtree_index
].bh
);
3560 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3567 for (i
= subtree_index
+ 1;
3568 i
< path_num_items(right_path
); i
++) {
3569 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3576 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
,
3584 left_rec
= &el
->l_recs
[index
- 1];
3585 if (ocfs2_is_empty_extent(&el
->l_recs
[0]))
3586 has_empty_extent
= 1;
3589 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, right_path
,
3590 path_num_items(right_path
) - 1);
3596 if (has_empty_extent
&& index
== 1) {
3598 * The easy case - we can just plop the record right in.
3600 *left_rec
= *split_rec
;
3602 le16_add_cpu(&left_rec
->e_leaf_clusters
, split_clusters
);
3604 le32_add_cpu(&right_rec
->e_cpos
, split_clusters
);
3605 le64_add_cpu(&right_rec
->e_blkno
,
3606 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et
->et_ci
),
3608 le16_add_cpu(&right_rec
->e_leaf_clusters
, -split_clusters
);
3610 ocfs2_cleanup_merge(el
, index
);
3612 ocfs2_journal_dirty(handle
, bh
);
3614 ocfs2_journal_dirty(handle
, path_leaf_bh(left_path
));
3617 * In the situation that the right_rec is empty and the extent
3618 * block is empty also, ocfs2_complete_edge_insert can't handle
3619 * it and we need to delete the right extent block.
3621 if (le16_to_cpu(right_rec
->e_leaf_clusters
) == 0 &&
3622 le16_to_cpu(el
->l_next_free_rec
) == 1) {
3623 /* extend credit for ocfs2_remove_rightmost_path */
3624 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3625 jbd2_handle_buffer_credits(handle
),
3632 ret
= ocfs2_remove_rightmost_path(handle
, et
,
3640 /* Now the rightmost extent block has been deleted.
3641 * So we use the new rightmost path.
3643 ocfs2_mv_path(right_path
, left_path
);
3646 ocfs2_complete_edge_insert(handle
, left_path
,
3647 right_path
, subtree_index
);
3650 ocfs2_free_path(left_path
);
3654 static int ocfs2_try_to_merge_extent(handle_t
*handle
,
3655 struct ocfs2_extent_tree
*et
,
3656 struct ocfs2_path
*path
,
3658 struct ocfs2_extent_rec
*split_rec
,
3659 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
3660 struct ocfs2_merge_ctxt
*ctxt
)
3663 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
3664 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
3666 BUG_ON(ctxt
->c_contig_type
== CONTIG_NONE
);
3668 if (ctxt
->c_split_covers_rec
&& ctxt
->c_has_empty_extent
) {
3669 /* extend credit for ocfs2_remove_rightmost_path */
3670 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3671 jbd2_handle_buffer_credits(handle
),
3678 * The merge code will need to create an empty
3679 * extent to take the place of the newly
3680 * emptied slot. Remove any pre-existing empty
3681 * extents - having more than one in a leaf is
3684 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3690 rec
= &el
->l_recs
[split_index
];
3693 if (ctxt
->c_contig_type
== CONTIG_LEFTRIGHT
) {
3695 * Left-right contig implies this.
3697 BUG_ON(!ctxt
->c_split_covers_rec
);
3700 * Since the leftright insert always covers the entire
3701 * extent, this call will delete the insert record
3702 * entirely, resulting in an empty extent record added to
3705 * Since the adding of an empty extent shifts
3706 * everything back to the right, there's no need to
3707 * update split_index here.
3709 * When the split_index is zero, we need to merge it to the
3710 * prevoius extent block. It is more efficient and easier
3711 * if we do merge_right first and merge_left later.
3713 ret
= ocfs2_merge_rec_right(path
, handle
, et
, split_rec
,
3721 * We can only get this from logic error above.
3723 BUG_ON(!ocfs2_is_empty_extent(&el
->l_recs
[0]));
3725 /* extend credit for ocfs2_remove_rightmost_path */
3726 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3727 jbd2_handle_buffer_credits(handle
),
3734 /* The merge left us with an empty extent, remove it. */
3735 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3741 rec
= &el
->l_recs
[split_index
];
3744 * Note that we don't pass split_rec here on purpose -
3745 * we've merged it into the rec already.
3747 ret
= ocfs2_merge_rec_left(path
, handle
, et
, rec
,
3748 dealloc
, split_index
);
3755 /* extend credit for ocfs2_remove_rightmost_path */
3756 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3757 jbd2_handle_buffer_credits(handle
),
3764 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
3766 * Error from this last rotate is not critical, so
3767 * print but don't bubble it up.
3774 * Merge a record to the left or right.
3776 * 'contig_type' is relative to the existing record,
3777 * so for example, if we're "right contig", it's to
3778 * the record on the left (hence the left merge).
3780 if (ctxt
->c_contig_type
== CONTIG_RIGHT
) {
3781 ret
= ocfs2_merge_rec_left(path
, handle
, et
,
3789 ret
= ocfs2_merge_rec_right(path
, handle
,
3798 if (ctxt
->c_split_covers_rec
) {
3799 /* extend credit for ocfs2_remove_rightmost_path */
3800 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
3801 jbd2_handle_buffer_credits(handle
),
3810 * The merge may have left an empty extent in
3811 * our leaf. Try to rotate it away.
3813 ret
= ocfs2_rotate_tree_left(handle
, et
, path
,
3825 static void ocfs2_subtract_from_rec(struct super_block
*sb
,
3826 enum ocfs2_split_type split
,
3827 struct ocfs2_extent_rec
*rec
,
3828 struct ocfs2_extent_rec
*split_rec
)
3832 len_blocks
= ocfs2_clusters_to_blocks(sb
,
3833 le16_to_cpu(split_rec
->e_leaf_clusters
));
3835 if (split
== SPLIT_LEFT
) {
3837 * Region is on the left edge of the existing
3840 le32_add_cpu(&rec
->e_cpos
,
3841 le16_to_cpu(split_rec
->e_leaf_clusters
));
3842 le64_add_cpu(&rec
->e_blkno
, len_blocks
);
3843 le16_add_cpu(&rec
->e_leaf_clusters
,
3844 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3847 * Region is on the right edge of the existing
3850 le16_add_cpu(&rec
->e_leaf_clusters
,
3851 -le16_to_cpu(split_rec
->e_leaf_clusters
));
3856 * Do the final bits of extent record insertion at the target leaf
3857 * list. If this leaf is part of an allocation tree, it is assumed
3858 * that the tree above has been prepared.
3860 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree
*et
,
3861 struct ocfs2_extent_rec
*insert_rec
,
3862 struct ocfs2_extent_list
*el
,
3863 struct ocfs2_insert_type
*insert
)
3865 int i
= insert
->ins_contig_index
;
3867 struct ocfs2_extent_rec
*rec
;
3869 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
3871 if (insert
->ins_split
!= SPLIT_NONE
) {
3872 i
= ocfs2_search_extent_list(el
, le32_to_cpu(insert_rec
->e_cpos
));
3874 rec
= &el
->l_recs
[i
];
3875 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
3876 insert
->ins_split
, rec
,
3882 * Contiguous insert - either left or right.
3884 if (insert
->ins_contig
!= CONTIG_NONE
) {
3885 rec
= &el
->l_recs
[i
];
3886 if (insert
->ins_contig
== CONTIG_LEFT
) {
3887 rec
->e_blkno
= insert_rec
->e_blkno
;
3888 rec
->e_cpos
= insert_rec
->e_cpos
;
3890 le16_add_cpu(&rec
->e_leaf_clusters
,
3891 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3896 * Handle insert into an empty leaf.
3898 if (le16_to_cpu(el
->l_next_free_rec
) == 0 ||
3899 ((le16_to_cpu(el
->l_next_free_rec
) == 1) &&
3900 ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
3901 el
->l_recs
[0] = *insert_rec
;
3902 el
->l_next_free_rec
= cpu_to_le16(1);
3909 if (insert
->ins_appending
== APPEND_TAIL
) {
3910 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
3911 rec
= &el
->l_recs
[i
];
3912 range
= le32_to_cpu(rec
->e_cpos
)
3913 + le16_to_cpu(rec
->e_leaf_clusters
);
3914 BUG_ON(le32_to_cpu(insert_rec
->e_cpos
) < range
);
3916 mlog_bug_on_msg(le16_to_cpu(el
->l_next_free_rec
) >=
3917 le16_to_cpu(el
->l_count
),
3918 "owner %llu, depth %u, count %u, next free %u, "
3919 "rec.cpos %u, rec.clusters %u, "
3920 "insert.cpos %u, insert.clusters %u\n",
3921 ocfs2_metadata_cache_owner(et
->et_ci
),
3922 le16_to_cpu(el
->l_tree_depth
),
3923 le16_to_cpu(el
->l_count
),
3924 le16_to_cpu(el
->l_next_free_rec
),
3925 le32_to_cpu(el
->l_recs
[i
].e_cpos
),
3926 le16_to_cpu(el
->l_recs
[i
].e_leaf_clusters
),
3927 le32_to_cpu(insert_rec
->e_cpos
),
3928 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3930 el
->l_recs
[i
] = *insert_rec
;
3931 le16_add_cpu(&el
->l_next_free_rec
, 1);
3937 * Ok, we have to rotate.
3939 * At this point, it is safe to assume that inserting into an
3940 * empty leaf and appending to a leaf have both been handled
3943 * This leaf needs to have space, either by the empty 1st
3944 * extent record, or by virtue of an l_next_free_rec < l_count.
3946 ocfs2_rotate_leaf(el
, insert_rec
);
3949 static void ocfs2_adjust_rightmost_records(handle_t
*handle
,
3950 struct ocfs2_extent_tree
*et
,
3951 struct ocfs2_path
*path
,
3952 struct ocfs2_extent_rec
*insert_rec
)
3955 struct buffer_head
*bh
;
3956 struct ocfs2_extent_list
*el
;
3957 struct ocfs2_extent_rec
*rec
;
3960 * Update everything except the leaf block.
3962 for (i
= 0; i
< path
->p_tree_depth
; i
++) {
3963 bh
= path
->p_node
[i
].bh
;
3964 el
= path
->p_node
[i
].el
;
3966 next_free
= le16_to_cpu(el
->l_next_free_rec
);
3967 if (next_free
== 0) {
3968 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
3969 "Owner %llu has a bad extent list\n",
3970 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
));
3974 rec
= &el
->l_recs
[next_free
- 1];
3976 rec
->e_int_clusters
= insert_rec
->e_cpos
;
3977 le32_add_cpu(&rec
->e_int_clusters
,
3978 le16_to_cpu(insert_rec
->e_leaf_clusters
));
3979 le32_add_cpu(&rec
->e_int_clusters
,
3980 -le32_to_cpu(rec
->e_cpos
));
3982 ocfs2_journal_dirty(handle
, bh
);
3986 static int ocfs2_append_rec_to_path(handle_t
*handle
,
3987 struct ocfs2_extent_tree
*et
,
3988 struct ocfs2_extent_rec
*insert_rec
,
3989 struct ocfs2_path
*right_path
,
3990 struct ocfs2_path
**ret_left_path
)
3993 struct ocfs2_extent_list
*el
;
3994 struct ocfs2_path
*left_path
= NULL
;
3996 *ret_left_path
= NULL
;
3999 * This shouldn't happen for non-trees. The extent rec cluster
4000 * count manipulation below only works for interior nodes.
4002 BUG_ON(right_path
->p_tree_depth
== 0);
4005 * If our appending insert is at the leftmost edge of a leaf,
4006 * then we might need to update the rightmost records of the
4009 el
= path_leaf_el(right_path
);
4010 next_free
= le16_to_cpu(el
->l_next_free_rec
);
4011 if (next_free
== 0 ||
4012 (next_free
== 1 && ocfs2_is_empty_extent(&el
->l_recs
[0]))) {
4015 ret
= ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et
->et_ci
),
4016 right_path
, &left_cpos
);
4022 trace_ocfs2_append_rec_to_path(
4023 (unsigned long long)
4024 ocfs2_metadata_cache_owner(et
->et_ci
),
4025 le32_to_cpu(insert_rec
->e_cpos
),
4029 * No need to worry if the append is already in the
4033 left_path
= ocfs2_new_path_from_path(right_path
);
4040 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
4048 * ocfs2_insert_path() will pass the left_path to the
4054 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4060 ocfs2_adjust_rightmost_records(handle
, et
, right_path
, insert_rec
);
4062 *ret_left_path
= left_path
;
4066 ocfs2_free_path(left_path
);
4071 static void ocfs2_split_record(struct ocfs2_extent_tree
*et
,
4072 struct ocfs2_path
*left_path
,
4073 struct ocfs2_path
*right_path
,
4074 struct ocfs2_extent_rec
*split_rec
,
4075 enum ocfs2_split_type split
)
4078 u32 cpos
= le32_to_cpu(split_rec
->e_cpos
);
4079 struct ocfs2_extent_list
*left_el
= NULL
, *right_el
, *insert_el
, *el
;
4080 struct ocfs2_extent_rec
*rec
, *tmprec
;
4082 right_el
= path_leaf_el(right_path
);
4084 left_el
= path_leaf_el(left_path
);
4087 insert_el
= right_el
;
4088 index
= ocfs2_search_extent_list(el
, cpos
);
4090 if (index
== 0 && left_path
) {
4091 BUG_ON(ocfs2_is_empty_extent(&el
->l_recs
[0]));
4094 * This typically means that the record
4095 * started in the left path but moved to the
4096 * right as a result of rotation. We either
4097 * move the existing record to the left, or we
4098 * do the later insert there.
4100 * In this case, the left path should always
4101 * exist as the rotate code will have passed
4102 * it back for a post-insert update.
4105 if (split
== SPLIT_LEFT
) {
4107 * It's a left split. Since we know
4108 * that the rotate code gave us an
4109 * empty extent in the left path, we
4110 * can just do the insert there.
4112 insert_el
= left_el
;
4115 * Right split - we have to move the
4116 * existing record over to the left
4117 * leaf. The insert will be into the
4118 * newly created empty extent in the
4121 tmprec
= &right_el
->l_recs
[index
];
4122 ocfs2_rotate_leaf(left_el
, tmprec
);
4125 memset(tmprec
, 0, sizeof(*tmprec
));
4126 index
= ocfs2_search_extent_list(left_el
, cpos
);
4127 BUG_ON(index
== -1);
4132 BUG_ON(!ocfs2_is_empty_extent(&left_el
->l_recs
[0]));
4134 * Left path is easy - we can just allow the insert to
4138 insert_el
= left_el
;
4139 index
= ocfs2_search_extent_list(el
, cpos
);
4140 BUG_ON(index
== -1);
4143 rec
= &el
->l_recs
[index
];
4144 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
4145 split
, rec
, split_rec
);
4146 ocfs2_rotate_leaf(insert_el
, split_rec
);
4150 * This function only does inserts on an allocation b-tree. For tree
4151 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4153 * right_path is the path we want to do the actual insert
4154 * in. left_path should only be passed in if we need to update that
4155 * portion of the tree after an edge insert.
4157 static int ocfs2_insert_path(handle_t
*handle
,
4158 struct ocfs2_extent_tree
*et
,
4159 struct ocfs2_path
*left_path
,
4160 struct ocfs2_path
*right_path
,
4161 struct ocfs2_extent_rec
*insert_rec
,
4162 struct ocfs2_insert_type
*insert
)
4164 int ret
, subtree_index
;
4165 struct buffer_head
*leaf_bh
= path_leaf_bh(right_path
);
4169 * There's a chance that left_path got passed back to
4170 * us without being accounted for in the
4171 * journal. Extend our transaction here to be sure we
4172 * can change those blocks.
4174 ret
= ocfs2_extend_trans(handle
, left_path
->p_tree_depth
);
4180 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
4188 * Pass both paths to the journal. The majority of inserts
4189 * will be touching all components anyway.
4191 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, right_path
);
4197 if (insert
->ins_split
!= SPLIT_NONE
) {
4199 * We could call ocfs2_insert_at_leaf() for some types
4200 * of splits, but it's easier to just let one separate
4201 * function sort it all out.
4203 ocfs2_split_record(et
, left_path
, right_path
,
4204 insert_rec
, insert
->ins_split
);
4207 * Split might have modified either leaf and we don't
4208 * have a guarantee that the later edge insert will
4209 * dirty this for us.
4212 ocfs2_journal_dirty(handle
,
4213 path_leaf_bh(left_path
));
4215 ocfs2_insert_at_leaf(et
, insert_rec
, path_leaf_el(right_path
),
4218 ocfs2_journal_dirty(handle
, leaf_bh
);
4222 * The rotate code has indicated that we need to fix
4223 * up portions of the tree after the insert.
4225 * XXX: Should we extend the transaction here?
4227 subtree_index
= ocfs2_find_subtree_root(et
, left_path
,
4229 ocfs2_complete_edge_insert(handle
, left_path
, right_path
,
4238 static int ocfs2_do_insert_extent(handle_t
*handle
,
4239 struct ocfs2_extent_tree
*et
,
4240 struct ocfs2_extent_rec
*insert_rec
,
4241 struct ocfs2_insert_type
*type
)
4243 int ret
, rotate
= 0;
4245 struct ocfs2_path
*right_path
= NULL
;
4246 struct ocfs2_path
*left_path
= NULL
;
4247 struct ocfs2_extent_list
*el
;
4249 el
= et
->et_root_el
;
4251 ret
= ocfs2_et_root_journal_access(handle
, et
,
4252 OCFS2_JOURNAL_ACCESS_WRITE
);
4258 if (le16_to_cpu(el
->l_tree_depth
) == 0) {
4259 ocfs2_insert_at_leaf(et
, insert_rec
, el
, type
);
4260 goto out_update_clusters
;
4263 right_path
= ocfs2_new_path_from_et(et
);
4271 * Determine the path to start with. Rotations need the
4272 * rightmost path, everything else can go directly to the
4275 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4276 if (type
->ins_appending
== APPEND_NONE
&&
4277 type
->ins_contig
== CONTIG_NONE
) {
4282 ret
= ocfs2_find_path(et
->et_ci
, right_path
, cpos
);
4289 * Rotations and appends need special treatment - they modify
4290 * parts of the tree's above them.
4292 * Both might pass back a path immediate to the left of the
4293 * one being inserted to. This will be cause
4294 * ocfs2_insert_path() to modify the rightmost records of
4295 * left_path to account for an edge insert.
4297 * XXX: When modifying this code, keep in mind that an insert
4298 * can wind up skipping both of these two special cases...
4301 ret
= ocfs2_rotate_tree_right(handle
, et
, type
->ins_split
,
4302 le32_to_cpu(insert_rec
->e_cpos
),
4303 right_path
, &left_path
);
4310 * ocfs2_rotate_tree_right() might have extended the
4311 * transaction without re-journaling our tree root.
4313 ret
= ocfs2_et_root_journal_access(handle
, et
,
4314 OCFS2_JOURNAL_ACCESS_WRITE
);
4319 } else if (type
->ins_appending
== APPEND_TAIL
4320 && type
->ins_contig
!= CONTIG_LEFT
) {
4321 ret
= ocfs2_append_rec_to_path(handle
, et
, insert_rec
,
4322 right_path
, &left_path
);
4329 ret
= ocfs2_insert_path(handle
, et
, left_path
, right_path
,
4336 out_update_clusters
:
4337 if (type
->ins_split
== SPLIT_NONE
)
4338 ocfs2_et_update_clusters(et
,
4339 le16_to_cpu(insert_rec
->e_leaf_clusters
));
4341 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4344 ocfs2_free_path(left_path
);
4345 ocfs2_free_path(right_path
);
4350 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree
*et
,
4351 struct ocfs2_path
*path
,
4352 struct ocfs2_extent_list
*el
, int index
,
4353 struct ocfs2_extent_rec
*split_rec
,
4354 struct ocfs2_merge_ctxt
*ctxt
)
4357 enum ocfs2_contig_type ret
= CONTIG_NONE
;
4358 u32 left_cpos
, right_cpos
;
4359 struct ocfs2_extent_rec
*rec
= NULL
;
4360 struct ocfs2_extent_list
*new_el
;
4361 struct ocfs2_path
*left_path
= NULL
, *right_path
= NULL
;
4362 struct buffer_head
*bh
;
4363 struct ocfs2_extent_block
*eb
;
4364 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
4367 rec
= &el
->l_recs
[index
- 1];
4368 } else if (path
->p_tree_depth
> 0) {
4369 status
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
4373 if (left_cpos
!= 0) {
4374 left_path
= ocfs2_new_path_from_path(path
);
4381 status
= ocfs2_find_path(et
->et_ci
, left_path
,
4384 goto free_left_path
;
4386 new_el
= path_leaf_el(left_path
);
4388 if (le16_to_cpu(new_el
->l_next_free_rec
) !=
4389 le16_to_cpu(new_el
->l_count
)) {
4390 bh
= path_leaf_bh(left_path
);
4391 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4392 status
= ocfs2_error(sb
,
4393 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4394 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4395 le16_to_cpu(new_el
->l_next_free_rec
),
4396 le16_to_cpu(new_el
->l_count
));
4397 goto free_left_path
;
4399 rec
= &new_el
->l_recs
[
4400 le16_to_cpu(new_el
->l_next_free_rec
) - 1];
4405 * We're careful to check for an empty extent record here -
4406 * the merge code will know what to do if it sees one.
4409 if (index
== 1 && ocfs2_is_empty_extent(rec
)) {
4410 if (split_rec
->e_cpos
== el
->l_recs
[index
].e_cpos
)
4413 ret
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4418 if (index
< (le16_to_cpu(el
->l_next_free_rec
) - 1))
4419 rec
= &el
->l_recs
[index
+ 1];
4420 else if (le16_to_cpu(el
->l_next_free_rec
) == le16_to_cpu(el
->l_count
) &&
4421 path
->p_tree_depth
> 0) {
4422 status
= ocfs2_find_cpos_for_right_leaf(sb
, path
, &right_cpos
);
4424 goto free_left_path
;
4426 if (right_cpos
== 0)
4427 goto free_left_path
;
4429 right_path
= ocfs2_new_path_from_path(path
);
4433 goto free_left_path
;
4436 status
= ocfs2_find_path(et
->et_ci
, right_path
, right_cpos
);
4438 goto free_right_path
;
4440 new_el
= path_leaf_el(right_path
);
4441 rec
= &new_el
->l_recs
[0];
4442 if (ocfs2_is_empty_extent(rec
)) {
4443 if (le16_to_cpu(new_el
->l_next_free_rec
) <= 1) {
4444 bh
= path_leaf_bh(right_path
);
4445 eb
= (struct ocfs2_extent_block
*)bh
->b_data
;
4446 status
= ocfs2_error(sb
,
4447 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4448 (unsigned long long)le64_to_cpu(eb
->h_blkno
),
4449 le16_to_cpu(new_el
->l_next_free_rec
));
4450 goto free_right_path
;
4452 rec
= &new_el
->l_recs
[1];
4457 enum ocfs2_contig_type contig_type
;
4459 contig_type
= ocfs2_et_extent_contig(et
, rec
, split_rec
);
4461 if (contig_type
== CONTIG_LEFT
&& ret
== CONTIG_RIGHT
)
4462 ret
= CONTIG_LEFTRIGHT
;
4463 else if (ret
== CONTIG_NONE
)
4468 ocfs2_free_path(right_path
);
4470 ocfs2_free_path(left_path
);
4473 ctxt
->c_contig_type
= ret
;
4478 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree
*et
,
4479 struct ocfs2_insert_type
*insert
,
4480 struct ocfs2_extent_list
*el
,
4481 struct ocfs2_extent_rec
*insert_rec
)
4484 enum ocfs2_contig_type contig_type
= CONTIG_NONE
;
4486 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4488 for(i
= 0; i
< le16_to_cpu(el
->l_next_free_rec
); i
++) {
4489 contig_type
= ocfs2_et_extent_contig(et
, &el
->l_recs
[i
],
4491 if (contig_type
!= CONTIG_NONE
) {
4492 insert
->ins_contig_index
= i
;
4496 insert
->ins_contig
= contig_type
;
4498 if (insert
->ins_contig
!= CONTIG_NONE
) {
4499 struct ocfs2_extent_rec
*rec
=
4500 &el
->l_recs
[insert
->ins_contig_index
];
4501 unsigned int len
= le16_to_cpu(rec
->e_leaf_clusters
) +
4502 le16_to_cpu(insert_rec
->e_leaf_clusters
);
4505 * Caller might want us to limit the size of extents, don't
4506 * calculate contiguousness if we might exceed that limit.
4508 if (et
->et_max_leaf_clusters
&&
4509 (len
> et
->et_max_leaf_clusters
))
4510 insert
->ins_contig
= CONTIG_NONE
;
4515 * This should only be called against the righmost leaf extent list.
4517 * ocfs2_figure_appending_type() will figure out whether we'll have to
4518 * insert at the tail of the rightmost leaf.
4520 * This should also work against the root extent list for tree's with 0
4521 * depth. If we consider the root extent list to be the rightmost leaf node
4522 * then the logic here makes sense.
4524 static void ocfs2_figure_appending_type(struct ocfs2_insert_type
*insert
,
4525 struct ocfs2_extent_list
*el
,
4526 struct ocfs2_extent_rec
*insert_rec
)
4529 u32 cpos
= le32_to_cpu(insert_rec
->e_cpos
);
4530 struct ocfs2_extent_rec
*rec
;
4532 insert
->ins_appending
= APPEND_NONE
;
4534 BUG_ON(le16_to_cpu(el
->l_tree_depth
) != 0);
4536 if (!el
->l_next_free_rec
)
4537 goto set_tail_append
;
4539 if (ocfs2_is_empty_extent(&el
->l_recs
[0])) {
4540 /* Were all records empty? */
4541 if (le16_to_cpu(el
->l_next_free_rec
) == 1)
4542 goto set_tail_append
;
4545 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
4546 rec
= &el
->l_recs
[i
];
4549 (le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)))
4550 goto set_tail_append
;
4555 insert
->ins_appending
= APPEND_TAIL
;
4559 * Helper function called at the beginning of an insert.
4561 * This computes a few things that are commonly used in the process of
4562 * inserting into the btree:
4563 * - Whether the new extent is contiguous with an existing one.
4564 * - The current tree depth.
4565 * - Whether the insert is an appending one.
4566 * - The total # of free records in the tree.
4568 * All of the information is stored on the ocfs2_insert_type
4571 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree
*et
,
4572 struct buffer_head
**last_eb_bh
,
4573 struct ocfs2_extent_rec
*insert_rec
,
4575 struct ocfs2_insert_type
*insert
)
4578 struct ocfs2_extent_block
*eb
;
4579 struct ocfs2_extent_list
*el
;
4580 struct ocfs2_path
*path
= NULL
;
4581 struct buffer_head
*bh
= NULL
;
4583 insert
->ins_split
= SPLIT_NONE
;
4585 el
= et
->et_root_el
;
4586 insert
->ins_tree_depth
= le16_to_cpu(el
->l_tree_depth
);
4588 if (el
->l_tree_depth
) {
4590 * If we have tree depth, we read in the
4591 * rightmost extent block ahead of time as
4592 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4593 * may want it later.
4595 ret
= ocfs2_read_extent_block(et
->et_ci
,
4596 ocfs2_et_get_last_eb_blk(et
),
4602 eb
= (struct ocfs2_extent_block
*) bh
->b_data
;
4607 * Unless we have a contiguous insert, we'll need to know if
4608 * there is room left in our allocation tree for another
4611 * XXX: This test is simplistic, we can search for empty
4612 * extent records too.
4614 *free_records
= le16_to_cpu(el
->l_count
) -
4615 le16_to_cpu(el
->l_next_free_rec
);
4617 if (!insert
->ins_tree_depth
) {
4618 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4619 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4623 path
= ocfs2_new_path_from_et(et
);
4631 * In the case that we're inserting past what the tree
4632 * currently accounts for, ocfs2_find_path() will return for
4633 * us the rightmost tree path. This is accounted for below in
4634 * the appending code.
4636 ret
= ocfs2_find_path(et
->et_ci
, path
, le32_to_cpu(insert_rec
->e_cpos
));
4642 el
= path_leaf_el(path
);
4645 * Now that we have the path, there's two things we want to determine:
4646 * 1) Contiguousness (also set contig_index if this is so)
4648 * 2) Are we doing an append? We can trivially break this up
4649 * into two types of appends: simple record append, or a
4650 * rotate inside the tail leaf.
4652 ocfs2_figure_contig_type(et
, insert
, el
, insert_rec
);
4655 * The insert code isn't quite ready to deal with all cases of
4656 * left contiguousness. Specifically, if it's an insert into
4657 * the 1st record in a leaf, it will require the adjustment of
4658 * cluster count on the last record of the path directly to it's
4659 * left. For now, just catch that case and fool the layers
4660 * above us. This works just fine for tree_depth == 0, which
4661 * is why we allow that above.
4663 if (insert
->ins_contig
== CONTIG_LEFT
&&
4664 insert
->ins_contig_index
== 0)
4665 insert
->ins_contig
= CONTIG_NONE
;
4668 * Ok, so we can simply compare against last_eb to figure out
4669 * whether the path doesn't exist. This will only happen in
4670 * the case that we're doing a tail append, so maybe we can
4671 * take advantage of that information somehow.
4673 if (ocfs2_et_get_last_eb_blk(et
) ==
4674 path_leaf_bh(path
)->b_blocknr
) {
4676 * Ok, ocfs2_find_path() returned us the rightmost
4677 * tree path. This might be an appending insert. There are
4679 * 1) We're doing a true append at the tail:
4680 * -This might even be off the end of the leaf
4681 * 2) We're "appending" by rotating in the tail
4683 ocfs2_figure_appending_type(insert
, el
, insert_rec
);
4687 ocfs2_free_path(path
);
4697 * Insert an extent into a btree.
4699 * The caller needs to update the owning btree's cluster count.
4701 int ocfs2_insert_extent(handle_t
*handle
,
4702 struct ocfs2_extent_tree
*et
,
4707 struct ocfs2_alloc_context
*meta_ac
)
4711 struct buffer_head
*last_eb_bh
= NULL
;
4712 struct ocfs2_insert_type insert
= {0, };
4713 struct ocfs2_extent_rec rec
;
4715 trace_ocfs2_insert_extent_start(
4716 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
4717 cpos
, new_clusters
);
4719 memset(&rec
, 0, sizeof(rec
));
4720 rec
.e_cpos
= cpu_to_le32(cpos
);
4721 rec
.e_blkno
= cpu_to_le64(start_blk
);
4722 rec
.e_leaf_clusters
= cpu_to_le16(new_clusters
);
4723 rec
.e_flags
= flags
;
4724 status
= ocfs2_et_insert_check(et
, &rec
);
4730 status
= ocfs2_figure_insert_type(et
, &last_eb_bh
, &rec
,
4731 &free_records
, &insert
);
4737 trace_ocfs2_insert_extent(insert
.ins_appending
, insert
.ins_contig
,
4738 insert
.ins_contig_index
, free_records
,
4739 insert
.ins_tree_depth
);
4741 if (insert
.ins_contig
== CONTIG_NONE
&& free_records
== 0) {
4742 status
= ocfs2_grow_tree(handle
, et
,
4743 &insert
.ins_tree_depth
, &last_eb_bh
,
4751 /* Finally, we can add clusters. This might rotate the tree for us. */
4752 status
= ocfs2_do_insert_extent(handle
, et
, &rec
, &insert
);
4756 ocfs2_et_extent_map_insert(et
, &rec
);
4765 * Allcate and add clusters into the extent b-tree.
4766 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4767 * The extent b-tree's root is specified by et, and
4768 * it is not limited to the file storage. Any extent tree can use this
4769 * function if it implements the proper ocfs2_extent_tree.
4771 int ocfs2_add_clusters_in_btree(handle_t
*handle
,
4772 struct ocfs2_extent_tree
*et
,
4773 u32
*logical_offset
,
4774 u32 clusters_to_add
,
4776 struct ocfs2_alloc_context
*data_ac
,
4777 struct ocfs2_alloc_context
*meta_ac
,
4778 enum ocfs2_alloc_restarted
*reason_ret
)
4780 int status
= 0, err
= 0;
4783 enum ocfs2_alloc_restarted reason
= RESTART_NONE
;
4784 u32 bit_off
, num_bits
;
4787 struct ocfs2_super
*osb
=
4788 OCFS2_SB(ocfs2_metadata_cache_get_super(et
->et_ci
));
4790 BUG_ON(!clusters_to_add
);
4793 flags
= OCFS2_EXT_UNWRITTEN
;
4795 free_extents
= ocfs2_num_free_extents(et
);
4796 if (free_extents
< 0) {
4797 status
= free_extents
;
4802 /* there are two cases which could cause us to EAGAIN in the
4803 * we-need-more-metadata case:
4804 * 1) we haven't reserved *any*
4805 * 2) we are so fragmented, we've needed to add metadata too
4807 if (!free_extents
&& !meta_ac
) {
4810 reason
= RESTART_META
;
4812 } else if ((!free_extents
)
4813 && (ocfs2_alloc_context_bits_left(meta_ac
)
4814 < ocfs2_extend_meta_needed(et
->et_root_el
))) {
4817 reason
= RESTART_META
;
4821 status
= __ocfs2_claim_clusters(handle
, data_ac
, 1,
4822 clusters_to_add
, &bit_off
, &num_bits
);
4824 if (status
!= -ENOSPC
)
4829 BUG_ON(num_bits
> clusters_to_add
);
4831 /* reserve our write early -- insert_extent may update the tree root */
4832 status
= ocfs2_et_root_journal_access(handle
, et
,
4833 OCFS2_JOURNAL_ACCESS_WRITE
);
4840 block
= ocfs2_clusters_to_blocks(osb
->sb
, bit_off
);
4841 trace_ocfs2_add_clusters_in_btree(
4842 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
4844 status
= ocfs2_insert_extent(handle
, et
, *logical_offset
, block
,
4845 num_bits
, flags
, meta_ac
);
4852 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
4854 clusters_to_add
-= num_bits
;
4855 *logical_offset
+= num_bits
;
4857 if (clusters_to_add
) {
4858 err
= clusters_to_add
;
4860 reason
= RESTART_TRANS
;
4865 if (data_ac
->ac_which
== OCFS2_AC_USE_LOCAL
)
4866 ocfs2_free_local_alloc_bits(osb
, handle
, data_ac
,
4869 ocfs2_free_clusters(handle
,
4872 ocfs2_clusters_to_blocks(osb
->sb
, bit_off
),
4878 *reason_ret
= reason
;
4879 trace_ocfs2_add_clusters_in_btree_ret(status
, reason
, err
);
4883 static void ocfs2_make_right_split_rec(struct super_block
*sb
,
4884 struct ocfs2_extent_rec
*split_rec
,
4886 struct ocfs2_extent_rec
*rec
)
4888 u32 rec_cpos
= le32_to_cpu(rec
->e_cpos
);
4889 u32 rec_range
= rec_cpos
+ le16_to_cpu(rec
->e_leaf_clusters
);
4891 memset(split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
4893 split_rec
->e_cpos
= cpu_to_le32(cpos
);
4894 split_rec
->e_leaf_clusters
= cpu_to_le16(rec_range
- cpos
);
4896 split_rec
->e_blkno
= rec
->e_blkno
;
4897 le64_add_cpu(&split_rec
->e_blkno
,
4898 ocfs2_clusters_to_blocks(sb
, cpos
- rec_cpos
));
4900 split_rec
->e_flags
= rec
->e_flags
;
4903 static int ocfs2_split_and_insert(handle_t
*handle
,
4904 struct ocfs2_extent_tree
*et
,
4905 struct ocfs2_path
*path
,
4906 struct buffer_head
**last_eb_bh
,
4908 struct ocfs2_extent_rec
*orig_split_rec
,
4909 struct ocfs2_alloc_context
*meta_ac
)
4912 unsigned int insert_range
, rec_range
, do_leftright
= 0;
4913 struct ocfs2_extent_rec tmprec
;
4914 struct ocfs2_extent_list
*rightmost_el
;
4915 struct ocfs2_extent_rec rec
;
4916 struct ocfs2_extent_rec split_rec
= *orig_split_rec
;
4917 struct ocfs2_insert_type insert
;
4918 struct ocfs2_extent_block
*eb
;
4922 * Store a copy of the record on the stack - it might move
4923 * around as the tree is manipulated below.
4925 rec
= path_leaf_el(path
)->l_recs
[split_index
];
4927 rightmost_el
= et
->et_root_el
;
4929 depth
= le16_to_cpu(rightmost_el
->l_tree_depth
);
4931 BUG_ON(!(*last_eb_bh
));
4932 eb
= (struct ocfs2_extent_block
*) (*last_eb_bh
)->b_data
;
4933 rightmost_el
= &eb
->h_list
;
4936 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
4937 le16_to_cpu(rightmost_el
->l_count
)) {
4938 ret
= ocfs2_grow_tree(handle
, et
,
4939 &depth
, last_eb_bh
, meta_ac
);
4946 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
4947 insert
.ins_appending
= APPEND_NONE
;
4948 insert
.ins_contig
= CONTIG_NONE
;
4949 insert
.ins_tree_depth
= depth
;
4951 insert_range
= le32_to_cpu(split_rec
.e_cpos
) +
4952 le16_to_cpu(split_rec
.e_leaf_clusters
);
4953 rec_range
= le32_to_cpu(rec
.e_cpos
) +
4954 le16_to_cpu(rec
.e_leaf_clusters
);
4956 if (split_rec
.e_cpos
== rec
.e_cpos
) {
4957 insert
.ins_split
= SPLIT_LEFT
;
4958 } else if (insert_range
== rec_range
) {
4959 insert
.ins_split
= SPLIT_RIGHT
;
4962 * Left/right split. We fake this as a right split
4963 * first and then make a second pass as a left split.
4965 insert
.ins_split
= SPLIT_RIGHT
;
4967 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
4968 &tmprec
, insert_range
, &rec
);
4972 BUG_ON(do_leftright
);
4976 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
4982 if (do_leftright
== 1) {
4984 struct ocfs2_extent_list
*el
;
4987 split_rec
= *orig_split_rec
;
4989 ocfs2_reinit_path(path
, 1);
4991 cpos
= le32_to_cpu(split_rec
.e_cpos
);
4992 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
4998 el
= path_leaf_el(path
);
4999 split_index
= ocfs2_search_extent_list(el
, cpos
);
5000 if (split_index
== -1) {
5001 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5002 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5003 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5015 static int ocfs2_replace_extent_rec(handle_t
*handle
,
5016 struct ocfs2_extent_tree
*et
,
5017 struct ocfs2_path
*path
,
5018 struct ocfs2_extent_list
*el
,
5020 struct ocfs2_extent_rec
*split_rec
)
5024 ret
= ocfs2_path_bh_journal_access(handle
, et
->et_ci
, path
,
5025 path_num_items(path
) - 1);
5031 el
->l_recs
[split_index
] = *split_rec
;
5033 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5039 * Split part or all of the extent record at split_index in the leaf
5040 * pointed to by path. Merge with the contiguous extent record if needed.
5042 * Care is taken to handle contiguousness so as to not grow the tree.
5044 * meta_ac is not strictly necessary - we only truly need it if growth
5045 * of the tree is required. All other cases will degrade into a less
5046 * optimal tree layout.
5048 * last_eb_bh should be the rightmost leaf block for any extent
5049 * btree. Since a split may grow the tree or a merge might shrink it,
5050 * the caller cannot trust the contents of that buffer after this call.
5052 * This code is optimized for readability - several passes might be
5053 * made over certain portions of the tree. All of those blocks will
5054 * have been brought into cache (and pinned via the journal), so the
5055 * extra overhead is not expressed in terms of disk reads.
5057 int ocfs2_split_extent(handle_t
*handle
,
5058 struct ocfs2_extent_tree
*et
,
5059 struct ocfs2_path
*path
,
5061 struct ocfs2_extent_rec
*split_rec
,
5062 struct ocfs2_alloc_context
*meta_ac
,
5063 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5066 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5067 struct buffer_head
*last_eb_bh
= NULL
;
5068 struct ocfs2_extent_rec
*rec
= &el
->l_recs
[split_index
];
5069 struct ocfs2_merge_ctxt ctxt
;
5071 if (le32_to_cpu(rec
->e_cpos
) > le32_to_cpu(split_rec
->e_cpos
) ||
5072 ((le32_to_cpu(rec
->e_cpos
) + le16_to_cpu(rec
->e_leaf_clusters
)) <
5073 (le32_to_cpu(split_rec
->e_cpos
) + le16_to_cpu(split_rec
->e_leaf_clusters
)))) {
5079 ret
= ocfs2_figure_merge_contig_type(et
, path
, el
,
5089 * The core merge / split code wants to know how much room is
5090 * left in this allocation tree, so we pass the
5091 * rightmost extent list.
5093 if (path
->p_tree_depth
) {
5094 ret
= ocfs2_read_extent_block(et
->et_ci
,
5095 ocfs2_et_get_last_eb_blk(et
),
5103 if (rec
->e_cpos
== split_rec
->e_cpos
&&
5104 rec
->e_leaf_clusters
== split_rec
->e_leaf_clusters
)
5105 ctxt
.c_split_covers_rec
= 1;
5107 ctxt
.c_split_covers_rec
= 0;
5109 ctxt
.c_has_empty_extent
= ocfs2_is_empty_extent(&el
->l_recs
[0]);
5111 trace_ocfs2_split_extent(split_index
, ctxt
.c_contig_type
,
5112 ctxt
.c_has_empty_extent
,
5113 ctxt
.c_split_covers_rec
);
5115 if (ctxt
.c_contig_type
== CONTIG_NONE
) {
5116 if (ctxt
.c_split_covers_rec
)
5117 ret
= ocfs2_replace_extent_rec(handle
, et
, path
, el
,
5118 split_index
, split_rec
);
5120 ret
= ocfs2_split_and_insert(handle
, et
, path
,
5121 &last_eb_bh
, split_index
,
5122 split_rec
, meta_ac
);
5126 ret
= ocfs2_try_to_merge_extent(handle
, et
, path
,
5127 split_index
, split_rec
,
5139 * Change the flags of the already-existing extent at cpos for len clusters.
5141 * new_flags: the flags we want to set.
5142 * clear_flags: the flags we want to clear.
5143 * phys: the new physical offset we want this new extent starts from.
5145 * If the existing extent is larger than the request, initiate a
5146 * split. An attempt will be made at merging with adjacent extents.
5148 * The caller is responsible for passing down meta_ac if we'll need it.
5150 int ocfs2_change_extent_flag(handle_t
*handle
,
5151 struct ocfs2_extent_tree
*et
,
5152 u32 cpos
, u32 len
, u32 phys
,
5153 struct ocfs2_alloc_context
*meta_ac
,
5154 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5155 int new_flags
, int clear_flags
)
5158 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5159 u64 start_blkno
= ocfs2_clusters_to_blocks(sb
, phys
);
5160 struct ocfs2_extent_rec split_rec
;
5161 struct ocfs2_path
*left_path
= NULL
;
5162 struct ocfs2_extent_list
*el
;
5163 struct ocfs2_extent_rec
*rec
;
5165 left_path
= ocfs2_new_path_from_et(et
);
5172 ret
= ocfs2_find_path(et
->et_ci
, left_path
, cpos
);
5177 el
= path_leaf_el(left_path
);
5179 index
= ocfs2_search_extent_list(el
, cpos
);
5182 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5183 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5190 rec
= &el
->l_recs
[index
];
5191 if (new_flags
&& (rec
->e_flags
& new_flags
)) {
5192 mlog(ML_ERROR
, "Owner %llu tried to set %d flags on an "
5193 "extent that already had them\n",
5194 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5199 if (clear_flags
&& !(rec
->e_flags
& clear_flags
)) {
5200 mlog(ML_ERROR
, "Owner %llu tried to clear %d flags on an "
5201 "extent that didn't have them\n",
5202 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5207 memset(&split_rec
, 0, sizeof(struct ocfs2_extent_rec
));
5208 split_rec
.e_cpos
= cpu_to_le32(cpos
);
5209 split_rec
.e_leaf_clusters
= cpu_to_le16(len
);
5210 split_rec
.e_blkno
= cpu_to_le64(start_blkno
);
5211 split_rec
.e_flags
= rec
->e_flags
;
5213 split_rec
.e_flags
|= new_flags
;
5215 split_rec
.e_flags
&= ~clear_flags
;
5217 ret
= ocfs2_split_extent(handle
, et
, left_path
,
5218 index
, &split_rec
, meta_ac
,
5224 ocfs2_free_path(left_path
);
5230 * Mark the already-existing extent at cpos as written for len clusters.
5231 * This removes the unwritten extent flag.
5233 * If the existing extent is larger than the request, initiate a
5234 * split. An attempt will be made at merging with adjacent extents.
5236 * The caller is responsible for passing down meta_ac if we'll need it.
5238 int ocfs2_mark_extent_written(struct inode
*inode
,
5239 struct ocfs2_extent_tree
*et
,
5240 handle_t
*handle
, u32 cpos
, u32 len
, u32 phys
,
5241 struct ocfs2_alloc_context
*meta_ac
,
5242 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5246 trace_ocfs2_mark_extent_written(
5247 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
5250 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode
->i_sb
))) {
5251 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",
5252 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
5258 * XXX: This should be fixed up so that we just re-insert the
5259 * next extent records.
5261 ocfs2_et_extent_map_truncate(et
, 0);
5263 ret
= ocfs2_change_extent_flag(handle
, et
, cpos
,
5264 len
, phys
, meta_ac
, dealloc
,
5265 0, OCFS2_EXT_UNWRITTEN
);
5273 static int ocfs2_split_tree(handle_t
*handle
, struct ocfs2_extent_tree
*et
,
5274 struct ocfs2_path
*path
,
5275 int index
, u32 new_range
,
5276 struct ocfs2_alloc_context
*meta_ac
)
5278 int ret
, depth
, credits
;
5279 struct buffer_head
*last_eb_bh
= NULL
;
5280 struct ocfs2_extent_block
*eb
;
5281 struct ocfs2_extent_list
*rightmost_el
, *el
;
5282 struct ocfs2_extent_rec split_rec
;
5283 struct ocfs2_extent_rec
*rec
;
5284 struct ocfs2_insert_type insert
;
5287 * Setup the record to split before we grow the tree.
5289 el
= path_leaf_el(path
);
5290 rec
= &el
->l_recs
[index
];
5291 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et
->et_ci
),
5292 &split_rec
, new_range
, rec
);
5294 depth
= path
->p_tree_depth
;
5296 ret
= ocfs2_read_extent_block(et
->et_ci
,
5297 ocfs2_et_get_last_eb_blk(et
),
5304 eb
= (struct ocfs2_extent_block
*) last_eb_bh
->b_data
;
5305 rightmost_el
= &eb
->h_list
;
5307 rightmost_el
= path_leaf_el(path
);
5309 credits
= path
->p_tree_depth
+
5310 ocfs2_extend_meta_needed(et
->et_root_el
);
5311 ret
= ocfs2_extend_trans(handle
, credits
);
5317 if (le16_to_cpu(rightmost_el
->l_next_free_rec
) ==
5318 le16_to_cpu(rightmost_el
->l_count
)) {
5319 ret
= ocfs2_grow_tree(handle
, et
, &depth
, &last_eb_bh
,
5327 memset(&insert
, 0, sizeof(struct ocfs2_insert_type
));
5328 insert
.ins_appending
= APPEND_NONE
;
5329 insert
.ins_contig
= CONTIG_NONE
;
5330 insert
.ins_split
= SPLIT_RIGHT
;
5331 insert
.ins_tree_depth
= depth
;
5333 ret
= ocfs2_do_insert_extent(handle
, et
, &split_rec
, &insert
);
5342 static int ocfs2_truncate_rec(handle_t
*handle
,
5343 struct ocfs2_extent_tree
*et
,
5344 struct ocfs2_path
*path
, int index
,
5345 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5349 u32 left_cpos
, rec_range
, trunc_range
;
5350 int is_rightmost_tree_rec
= 0;
5351 struct super_block
*sb
= ocfs2_metadata_cache_get_super(et
->et_ci
);
5352 struct ocfs2_path
*left_path
= NULL
;
5353 struct ocfs2_extent_list
*el
= path_leaf_el(path
);
5354 struct ocfs2_extent_rec
*rec
;
5355 struct ocfs2_extent_block
*eb
;
5357 if (ocfs2_is_empty_extent(&el
->l_recs
[0]) && index
> 0) {
5358 /* extend credit for ocfs2_remove_rightmost_path */
5359 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
5360 jbd2_handle_buffer_credits(handle
),
5367 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5376 if (index
== (le16_to_cpu(el
->l_next_free_rec
) - 1) &&
5377 path
->p_tree_depth
) {
5379 * Check whether this is the rightmost tree record. If
5380 * we remove all of this record or part of its right
5381 * edge then an update of the record lengths above it
5384 eb
= (struct ocfs2_extent_block
*)path_leaf_bh(path
)->b_data
;
5385 if (eb
->h_next_leaf_blk
== 0)
5386 is_rightmost_tree_rec
= 1;
5389 rec
= &el
->l_recs
[index
];
5390 if (index
== 0 && path
->p_tree_depth
&&
5391 le32_to_cpu(rec
->e_cpos
) == cpos
) {
5393 * Changing the leftmost offset (via partial or whole
5394 * record truncate) of an interior (or rightmost) path
5395 * means we have to update the subtree that is formed
5396 * by this leaf and the one to it's left.
5398 * There are two cases we can skip:
5399 * 1) Path is the leftmost one in our btree.
5400 * 2) The leaf is rightmost and will be empty after
5401 * we remove the extent record - the rotate code
5402 * knows how to update the newly formed edge.
5405 ret
= ocfs2_find_cpos_for_left_leaf(sb
, path
, &left_cpos
);
5411 if (left_cpos
&& le16_to_cpu(el
->l_next_free_rec
) > 1) {
5412 left_path
= ocfs2_new_path_from_path(path
);
5419 ret
= ocfs2_find_path(et
->et_ci
, left_path
,
5428 ret
= ocfs2_extend_rotate_transaction(handle
, 0,
5429 jbd2_handle_buffer_credits(handle
),
5436 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, path
);
5442 ret
= ocfs2_journal_access_path(et
->et_ci
, handle
, left_path
);
5448 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5449 trunc_range
= cpos
+ len
;
5451 if (le32_to_cpu(rec
->e_cpos
) == cpos
&& rec_range
== trunc_range
) {
5454 memset(rec
, 0, sizeof(*rec
));
5455 ocfs2_cleanup_merge(el
, index
);
5457 next_free
= le16_to_cpu(el
->l_next_free_rec
);
5458 if (is_rightmost_tree_rec
&& next_free
> 1) {
5460 * We skip the edge update if this path will
5461 * be deleted by the rotate code.
5463 rec
= &el
->l_recs
[next_free
- 1];
5464 ocfs2_adjust_rightmost_records(handle
, et
, path
,
5467 } else if (le32_to_cpu(rec
->e_cpos
) == cpos
) {
5468 /* Remove leftmost portion of the record. */
5469 le32_add_cpu(&rec
->e_cpos
, len
);
5470 le64_add_cpu(&rec
->e_blkno
, ocfs2_clusters_to_blocks(sb
, len
));
5471 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5472 } else if (rec_range
== trunc_range
) {
5473 /* Remove rightmost portion of the record */
5474 le16_add_cpu(&rec
->e_leaf_clusters
, -len
);
5475 if (is_rightmost_tree_rec
)
5476 ocfs2_adjust_rightmost_records(handle
, et
, path
, rec
);
5478 /* Caller should have trapped this. */
5479 mlog(ML_ERROR
, "Owner %llu: Invalid record truncate: (%u, %u) "
5481 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5482 le32_to_cpu(rec
->e_cpos
),
5483 le16_to_cpu(rec
->e_leaf_clusters
), cpos
, len
);
5490 subtree_index
= ocfs2_find_subtree_root(et
, left_path
, path
);
5491 ocfs2_complete_edge_insert(handle
, left_path
, path
,
5495 ocfs2_journal_dirty(handle
, path_leaf_bh(path
));
5497 ret
= ocfs2_rotate_tree_left(handle
, et
, path
, dealloc
);
5502 ocfs2_free_path(left_path
);
5506 int ocfs2_remove_extent(handle_t
*handle
,
5507 struct ocfs2_extent_tree
*et
,
5509 struct ocfs2_alloc_context
*meta_ac
,
5510 struct ocfs2_cached_dealloc_ctxt
*dealloc
)
5513 u32 rec_range
, trunc_range
;
5514 struct ocfs2_extent_rec
*rec
;
5515 struct ocfs2_extent_list
*el
;
5516 struct ocfs2_path
*path
= NULL
;
5519 * XXX: Why are we truncating to 0 instead of wherever this
5522 ocfs2_et_extent_map_truncate(et
, 0);
5524 path
= ocfs2_new_path_from_et(et
);
5531 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5537 el
= path_leaf_el(path
);
5538 index
= ocfs2_search_extent_list(el
, cpos
);
5540 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5541 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5542 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5549 * We have 3 cases of extent removal:
5550 * 1) Range covers the entire extent rec
5551 * 2) Range begins or ends on one edge of the extent rec
5552 * 3) Range is in the middle of the extent rec (no shared edges)
5554 * For case 1 we remove the extent rec and left rotate to
5557 * For case 2 we just shrink the existing extent rec, with a
5558 * tree update if the shrinking edge is also the edge of an
5561 * For case 3 we do a right split to turn the extent rec into
5562 * something case 2 can handle.
5564 rec
= &el
->l_recs
[index
];
5565 rec_range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
5566 trunc_range
= cpos
+ len
;
5568 BUG_ON(cpos
< le32_to_cpu(rec
->e_cpos
) || trunc_range
> rec_range
);
5570 trace_ocfs2_remove_extent(
5571 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5572 cpos
, len
, index
, le32_to_cpu(rec
->e_cpos
),
5573 ocfs2_rec_clusters(el
, rec
));
5575 if (le32_to_cpu(rec
->e_cpos
) == cpos
|| rec_range
== trunc_range
) {
5576 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5583 ret
= ocfs2_split_tree(handle
, et
, path
, index
,
5584 trunc_range
, meta_ac
);
5591 * The split could have manipulated the tree enough to
5592 * move the record location, so we have to look for it again.
5594 ocfs2_reinit_path(path
, 1);
5596 ret
= ocfs2_find_path(et
->et_ci
, path
, cpos
);
5602 el
= path_leaf_el(path
);
5603 index
= ocfs2_search_extent_list(el
, cpos
);
5605 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5606 "Owner %llu: split at cpos %u lost record\n",
5607 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5614 * Double check our values here. If anything is fishy,
5615 * it's easier to catch it at the top level.
5617 rec
= &el
->l_recs
[index
];
5618 rec_range
= le32_to_cpu(rec
->e_cpos
) +
5619 ocfs2_rec_clusters(el
, rec
);
5620 if (rec_range
!= trunc_range
) {
5621 ocfs2_error(ocfs2_metadata_cache_get_super(et
->et_ci
),
5622 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5623 (unsigned long long)ocfs2_metadata_cache_owner(et
->et_ci
),
5624 cpos
, len
, le32_to_cpu(rec
->e_cpos
),
5625 ocfs2_rec_clusters(el
, rec
));
5630 ret
= ocfs2_truncate_rec(handle
, et
, path
, index
, dealloc
,
5637 ocfs2_free_path(path
);
5642 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5643 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5644 * number to reserve some extra blocks, and it only handles meta
5647 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5648 * and punching holes.
5650 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode
*inode
,
5651 struct ocfs2_extent_tree
*et
,
5652 u32 extents_to_split
,
5653 struct ocfs2_alloc_context
**ac
,
5656 int ret
= 0, num_free_extents
;
5657 unsigned int max_recs_needed
= 2 * extents_to_split
;
5658 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
5662 num_free_extents
= ocfs2_num_free_extents(et
);
5663 if (num_free_extents
< 0) {
5664 ret
= num_free_extents
;
5669 if (!num_free_extents
||
5670 (ocfs2_sparse_alloc(osb
) && num_free_extents
< max_recs_needed
))
5671 extra_blocks
+= ocfs2_extend_meta_needed(et
->et_root_el
);
5674 ret
= ocfs2_reserve_new_metadata_blocks(osb
, extra_blocks
, ac
);
5684 ocfs2_free_alloc_context(*ac
);
5692 int ocfs2_remove_btree_range(struct inode
*inode
,
5693 struct ocfs2_extent_tree
*et
,
5694 u32 cpos
, u32 phys_cpos
, u32 len
, int flags
,
5695 struct ocfs2_cached_dealloc_ctxt
*dealloc
,
5696 u64 refcount_loc
, bool refcount_tree_locked
)
5698 int ret
, credits
= 0, extra_blocks
= 0;
5699 u64 phys_blkno
= ocfs2_clusters_to_blocks(inode
->i_sb
, phys_cpos
);
5700 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
5701 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5703 struct ocfs2_alloc_context
*meta_ac
= NULL
;
5704 struct ocfs2_refcount_tree
*ref_tree
= NULL
;
5706 if ((flags
& OCFS2_EXT_REFCOUNTED
) && len
) {
5707 BUG_ON(!ocfs2_is_refcount_inode(inode
));
5709 if (!refcount_tree_locked
) {
5710 ret
= ocfs2_lock_refcount_tree(osb
, refcount_loc
, 1,
5718 ret
= ocfs2_prepare_refcount_change_for_del(inode
,
5730 ret
= ocfs2_reserve_blocks_for_rec_trunc(inode
, et
, 1, &meta_ac
,
5737 inode_lock(tl_inode
);
5739 if (ocfs2_truncate_log_needs_flush(osb
)) {
5740 ret
= __ocfs2_flush_truncate_log(osb
);
5747 handle
= ocfs2_start_trans(osb
,
5748 ocfs2_remove_extent_credits(osb
->sb
) + credits
);
5749 if (IS_ERR(handle
)) {
5750 ret
= PTR_ERR(handle
);
5755 ret
= ocfs2_et_root_journal_access(handle
, et
,
5756 OCFS2_JOURNAL_ACCESS_WRITE
);
5762 dquot_free_space_nodirty(inode
,
5763 ocfs2_clusters_to_bytes(inode
->i_sb
, len
));
5765 ret
= ocfs2_remove_extent(handle
, et
, cpos
, len
, meta_ac
, dealloc
);
5771 ocfs2_et_update_clusters(et
, -len
);
5772 ocfs2_update_inode_fsync_trans(handle
, inode
, 1);
5774 ocfs2_journal_dirty(handle
, et
->et_root_bh
);
5777 if (flags
& OCFS2_EXT_REFCOUNTED
)
5778 ret
= ocfs2_decrease_refcount(inode
, handle
,
5779 ocfs2_blocks_to_clusters(osb
->sb
,
5784 ret
= ocfs2_truncate_log_append(osb
, handle
,
5792 ocfs2_commit_trans(osb
, handle
);
5794 inode_unlock(tl_inode
);
5797 ocfs2_free_alloc_context(meta_ac
);
5800 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
5805 int ocfs2_truncate_log_needs_flush(struct ocfs2_super
*osb
)
5807 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5808 struct ocfs2_dinode
*di
;
5809 struct ocfs2_truncate_log
*tl
;
5811 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5812 tl
= &di
->id2
.i_dealloc
;
5814 mlog_bug_on_msg(le16_to_cpu(tl
->tl_used
) > le16_to_cpu(tl
->tl_count
),
5815 "slot %d, invalid truncate log parameters: used = "
5816 "%u, count = %u\n", osb
->slot_num
,
5817 le16_to_cpu(tl
->tl_used
), le16_to_cpu(tl
->tl_count
));
5818 return le16_to_cpu(tl
->tl_used
) == le16_to_cpu(tl
->tl_count
);
5821 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log
*tl
,
5822 unsigned int new_start
)
5824 unsigned int tail_index
;
5825 unsigned int current_tail
;
5827 /* No records, nothing to coalesce */
5828 if (!le16_to_cpu(tl
->tl_used
))
5831 tail_index
= le16_to_cpu(tl
->tl_used
) - 1;
5832 current_tail
= le32_to_cpu(tl
->tl_recs
[tail_index
].t_start
);
5833 current_tail
+= le32_to_cpu(tl
->tl_recs
[tail_index
].t_clusters
);
5835 return current_tail
== new_start
;
5838 int ocfs2_truncate_log_append(struct ocfs2_super
*osb
,
5841 unsigned int num_clusters
)
5844 unsigned int start_cluster
, tl_count
;
5845 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5846 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5847 struct ocfs2_dinode
*di
;
5848 struct ocfs2_truncate_log
*tl
;
5850 BUG_ON(inode_trylock(tl_inode
));
5852 start_cluster
= ocfs2_blocks_to_clusters(osb
->sb
, start_blk
);
5854 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5856 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5857 * by the underlying call to ocfs2_read_inode_block(), so any
5858 * corruption is a code bug */
5859 BUG_ON(!OCFS2_IS_VALID_DINODE(di
));
5861 tl
= &di
->id2
.i_dealloc
;
5862 tl_count
= le16_to_cpu(tl
->tl_count
);
5863 mlog_bug_on_msg(tl_count
> ocfs2_truncate_recs_per_inode(osb
->sb
) ||
5865 "Truncate record count on #%llu invalid "
5866 "wanted %u, actual %u\n",
5867 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5868 ocfs2_truncate_recs_per_inode(osb
->sb
),
5869 le16_to_cpu(tl
->tl_count
));
5871 /* Caller should have known to flush before calling us. */
5872 index
= le16_to_cpu(tl
->tl_used
);
5873 if (index
>= tl_count
) {
5879 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5880 OCFS2_JOURNAL_ACCESS_WRITE
);
5886 trace_ocfs2_truncate_log_append(
5887 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
, index
,
5888 start_cluster
, num_clusters
);
5889 if (ocfs2_truncate_log_can_coalesce(tl
, start_cluster
)) {
5891 * Move index back to the record we are coalescing with.
5892 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5896 num_clusters
+= le32_to_cpu(tl
->tl_recs
[index
].t_clusters
);
5897 trace_ocfs2_truncate_log_append(
5898 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5899 index
, le32_to_cpu(tl
->tl_recs
[index
].t_start
),
5902 tl
->tl_recs
[index
].t_start
= cpu_to_le32(start_cluster
);
5903 tl
->tl_used
= cpu_to_le16(index
+ 1);
5905 tl
->tl_recs
[index
].t_clusters
= cpu_to_le32(num_clusters
);
5907 ocfs2_journal_dirty(handle
, tl_bh
);
5909 osb
->truncated_clusters
+= num_clusters
;
5914 static int ocfs2_replay_truncate_records(struct ocfs2_super
*osb
,
5915 struct inode
*data_alloc_inode
,
5916 struct buffer_head
*data_alloc_bh
)
5920 unsigned int num_clusters
;
5922 struct ocfs2_truncate_rec rec
;
5923 struct ocfs2_dinode
*di
;
5924 struct ocfs2_truncate_log
*tl
;
5925 struct inode
*tl_inode
= osb
->osb_tl_inode
;
5926 struct buffer_head
*tl_bh
= osb
->osb_tl_bh
;
5929 di
= (struct ocfs2_dinode
*) tl_bh
->b_data
;
5930 tl
= &di
->id2
.i_dealloc
;
5931 i
= le16_to_cpu(tl
->tl_used
) - 1;
5933 handle
= ocfs2_start_trans(osb
, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC
);
5934 if (IS_ERR(handle
)) {
5935 status
= PTR_ERR(handle
);
5940 /* Caller has given us at least enough credits to
5941 * update the truncate log dinode */
5942 status
= ocfs2_journal_access_di(handle
, INODE_CACHE(tl_inode
), tl_bh
,
5943 OCFS2_JOURNAL_ACCESS_WRITE
);
5949 tl
->tl_used
= cpu_to_le16(i
);
5951 ocfs2_journal_dirty(handle
, tl_bh
);
5953 rec
= tl
->tl_recs
[i
];
5954 start_blk
= ocfs2_clusters_to_blocks(data_alloc_inode
->i_sb
,
5955 le32_to_cpu(rec
.t_start
));
5956 num_clusters
= le32_to_cpu(rec
.t_clusters
);
5958 /* if start_blk is not set, we ignore the record as
5961 trace_ocfs2_replay_truncate_records(
5962 (unsigned long long)OCFS2_I(tl_inode
)->ip_blkno
,
5963 i
, le32_to_cpu(rec
.t_start
), num_clusters
);
5965 status
= ocfs2_free_clusters(handle
, data_alloc_inode
,
5966 data_alloc_bh
, start_blk
,
5974 ocfs2_commit_trans(osb
, handle
);
5978 osb
->truncated_clusters
= 0;
5984 /* Expects you to already be holding tl_inode->i_mutex */
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
);
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 the area past i_size but still within an allocated
6927 * cluster. This avoids exposing nonzero data on subsequent file
6930 * We need to call this before i_size is updated on the inode because
6931 * otherwise block_write_full_page() will skip writeout of pages past
6932 * i_size. The new_i_size parameter is passed for this reason.
6934 int ocfs2_zero_range_for_truncate(struct inode
*inode
, handle_t
*handle
,
6935 u64 range_start
, u64 range_end
)
6937 int ret
= 0, numpages
;
6938 struct page
**pages
= NULL
;
6940 unsigned int ext_flags
;
6941 struct super_block
*sb
= inode
->i_sb
;
6944 * File systems which don't support sparse files zero on every
6947 if (!ocfs2_sparse_alloc(OCFS2_SB(sb
)))
6950 pages
= kcalloc(ocfs2_pages_per_cluster(sb
),
6951 sizeof(struct page
*), GFP_NOFS
);
6952 if (pages
== NULL
) {
6958 if (range_start
== range_end
)
6961 ret
= ocfs2_extent_map_get_blocks(inode
,
6962 range_start
>> sb
->s_blocksize_bits
,
6963 &phys
, NULL
, &ext_flags
);
6970 * Tail is a hole, or is marked unwritten. In either case, we
6971 * can count on read and write to return/push zero's.
6973 if (phys
== 0 || ext_flags
& OCFS2_EXT_UNWRITTEN
)
6976 ret
= ocfs2_grab_eof_pages(inode
, range_start
, range_end
, pages
,
6983 ocfs2_zero_cluster_pages(inode
, range_start
, range_end
, pages
,
6984 numpages
, phys
, handle
);
6987 * Initiate writeout of the pages we zero'd here. We don't
6988 * wait on them - the truncate_inode_pages() call later will
6991 ret
= filemap_fdatawrite_range(inode
->i_mapping
, range_start
,
7002 static void ocfs2_zero_dinode_id2_with_xattr(struct inode
*inode
,
7003 struct ocfs2_dinode
*di
)
7005 unsigned int blocksize
= 1 << inode
->i_sb
->s_blocksize_bits
;
7006 unsigned int xattrsize
= le16_to_cpu(di
->i_xattr_inline_size
);
7008 if (le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_XATTR_FL
)
7009 memset(&di
->id2
, 0, blocksize
-
7010 offsetof(struct ocfs2_dinode
, id2
) -
7013 memset(&di
->id2
, 0, blocksize
-
7014 offsetof(struct ocfs2_dinode
, id2
));
7017 void ocfs2_dinode_new_extent_list(struct inode
*inode
,
7018 struct ocfs2_dinode
*di
)
7020 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7021 di
->id2
.i_list
.l_tree_depth
= 0;
7022 di
->id2
.i_list
.l_next_free_rec
= 0;
7023 di
->id2
.i_list
.l_count
= cpu_to_le16(
7024 ocfs2_extent_recs_per_inode_with_xattr(inode
->i_sb
, di
));
7027 void ocfs2_set_inode_data_inline(struct inode
*inode
, struct ocfs2_dinode
*di
)
7029 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7030 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7032 spin_lock(&oi
->ip_lock
);
7033 oi
->ip_dyn_features
|= OCFS2_INLINE_DATA_FL
;
7034 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7035 spin_unlock(&oi
->ip_lock
);
7038 * We clear the entire i_data structure here so that all
7039 * fields can be properly initialized.
7041 ocfs2_zero_dinode_id2_with_xattr(inode
, di
);
7043 idata
->id_count
= cpu_to_le16(
7044 ocfs2_max_inline_data_with_xattr(inode
->i_sb
, di
));
7047 int ocfs2_convert_inline_data_to_extents(struct inode
*inode
,
7048 struct buffer_head
*di_bh
)
7050 int ret
, i
, has_data
, num_pages
= 0;
7055 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
7056 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7057 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7058 struct ocfs2_alloc_context
*data_ac
= NULL
;
7059 struct page
**pages
= NULL
;
7060 loff_t end
= osb
->s_clustersize
;
7061 struct ocfs2_extent_tree et
;
7064 has_data
= i_size_read(inode
) ? 1 : 0;
7067 pages
= kcalloc(ocfs2_pages_per_cluster(osb
->sb
),
7068 sizeof(struct page
*), GFP_NOFS
);
7069 if (pages
== NULL
) {
7075 ret
= ocfs2_reserve_clusters(osb
, 1, &data_ac
);
7082 handle
= ocfs2_start_trans(osb
,
7083 ocfs2_inline_to_extents_credits(osb
->sb
));
7084 if (IS_ERR(handle
)) {
7085 ret
= PTR_ERR(handle
);
7090 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7091 OCFS2_JOURNAL_ACCESS_WRITE
);
7098 unsigned int page_end
;
7101 ret
= dquot_alloc_space_nodirty(inode
,
7102 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7107 data_ac
->ac_resv
= &oi
->ip_la_data_resv
;
7109 ret
= ocfs2_claim_clusters(handle
, data_ac
, 1, &bit_off
,
7117 * Save two copies, one for insert, and one that can
7118 * be changed by ocfs2_map_and_dirty_page() below.
7120 block
= phys
= ocfs2_clusters_to_blocks(inode
->i_sb
, bit_off
);
7123 * Non sparse file systems zero on extend, so no need
7126 if (!ocfs2_sparse_alloc(osb
) &&
7127 PAGE_SIZE
< osb
->s_clustersize
)
7130 ret
= ocfs2_grab_eof_pages(inode
, 0, end
, pages
, &num_pages
);
7138 * This should populate the 1st page for us and mark
7141 ret
= ocfs2_read_inline_data(inode
, pages
[0], di_bh
);
7148 page_end
= PAGE_SIZE
;
7149 if (PAGE_SIZE
> osb
->s_clustersize
)
7150 page_end
= osb
->s_clustersize
;
7152 for (i
= 0; i
< num_pages
; i
++)
7153 ocfs2_map_and_dirty_page(inode
, handle
, 0, page_end
,
7154 pages
[i
], i
> 0, &phys
);
7157 spin_lock(&oi
->ip_lock
);
7158 oi
->ip_dyn_features
&= ~OCFS2_INLINE_DATA_FL
;
7159 di
->i_dyn_features
= cpu_to_le16(oi
->ip_dyn_features
);
7160 spin_unlock(&oi
->ip_lock
);
7162 ocfs2_update_inode_fsync_trans(handle
, inode
, 1);
7163 ocfs2_dinode_new_extent_list(inode
, di
);
7165 ocfs2_journal_dirty(handle
, di_bh
);
7169 * An error at this point should be extremely rare. If
7170 * this proves to be false, we could always re-build
7171 * the in-inode data from our pages.
7173 ocfs2_init_dinode_extent_tree(&et
, INODE_CACHE(inode
), di_bh
);
7174 ret
= ocfs2_insert_extent(handle
, &et
, 0, block
, 1, 0, NULL
);
7181 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7186 ocfs2_unlock_and_free_pages(pages
, num_pages
);
7189 if (ret
< 0 && did_quota
)
7190 dquot_free_space_nodirty(inode
,
7191 ocfs2_clusters_to_bytes(osb
->sb
, 1));
7194 if (data_ac
->ac_which
== OCFS2_AC_USE_LOCAL
)
7195 ocfs2_free_local_alloc_bits(osb
, handle
, data_ac
,
7198 ocfs2_free_clusters(handle
,
7201 ocfs2_clusters_to_blocks(osb
->sb
, bit_off
),
7205 ocfs2_commit_trans(osb
, handle
);
7209 ocfs2_free_alloc_context(data_ac
);
7216 * It is expected, that by the time you call this function,
7217 * inode->i_size and fe->i_size have been adjusted.
7219 * WARNING: This will kfree the truncate context
7221 int ocfs2_commit_truncate(struct ocfs2_super
*osb
,
7222 struct inode
*inode
,
7223 struct buffer_head
*di_bh
)
7225 int status
= 0, i
, flags
= 0;
7226 u32 new_highest_cpos
, range
, trunc_cpos
, trunc_len
, phys_cpos
, coff
;
7228 struct ocfs2_extent_list
*el
;
7229 struct ocfs2_extent_rec
*rec
;
7230 struct ocfs2_path
*path
= NULL
;
7231 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7232 struct ocfs2_extent_list
*root_el
= &(di
->id2
.i_list
);
7233 u64 refcount_loc
= le64_to_cpu(di
->i_refcount_loc
);
7234 struct ocfs2_extent_tree et
;
7235 struct ocfs2_cached_dealloc_ctxt dealloc
;
7236 struct ocfs2_refcount_tree
*ref_tree
= NULL
;
7238 ocfs2_init_dinode_extent_tree(&et
, INODE_CACHE(inode
), di_bh
);
7239 ocfs2_init_dealloc_ctxt(&dealloc
);
7241 new_highest_cpos
= ocfs2_clusters_for_bytes(osb
->sb
,
7242 i_size_read(inode
));
7244 path
= ocfs2_new_path(di_bh
, &di
->id2
.i_list
,
7245 ocfs2_journal_access_di
);
7252 ocfs2_extent_map_trunc(inode
, new_highest_cpos
);
7256 * Check that we still have allocation to delete.
7258 if (OCFS2_I(inode
)->ip_clusters
== 0) {
7264 * Truncate always works against the rightmost tree branch.
7266 status
= ocfs2_find_path(INODE_CACHE(inode
), path
, UINT_MAX
);
7272 trace_ocfs2_commit_truncate(
7273 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7275 OCFS2_I(inode
)->ip_clusters
,
7276 path
->p_tree_depth
);
7279 * By now, el will point to the extent list on the bottom most
7280 * portion of this tree. Only the tail record is considered in
7283 * We handle the following cases, in order:
7284 * - empty extent: delete the remaining branch
7285 * - remove the entire record
7286 * - remove a partial record
7287 * - no record needs to be removed (truncate has completed)
7289 el
= path_leaf_el(path
);
7290 if (le16_to_cpu(el
->l_next_free_rec
) == 0) {
7291 ocfs2_error(inode
->i_sb
,
7292 "Inode %llu has empty extent block at %llu\n",
7293 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7294 (unsigned long long)path_leaf_bh(path
)->b_blocknr
);
7299 i
= le16_to_cpu(el
->l_next_free_rec
) - 1;
7300 rec
= &el
->l_recs
[i
];
7301 flags
= rec
->e_flags
;
7302 range
= le32_to_cpu(rec
->e_cpos
) + ocfs2_rec_clusters(el
, rec
);
7304 if (i
== 0 && ocfs2_is_empty_extent(rec
)) {
7306 * Lower levels depend on this never happening, but it's best
7307 * to check it up here before changing the tree.
7309 if (root_el
->l_tree_depth
&& rec
->e_int_clusters
== 0) {
7310 mlog(ML_ERROR
, "Inode %lu has an empty "
7311 "extent record, depth %u\n", inode
->i_ino
,
7312 le16_to_cpu(root_el
->l_tree_depth
));
7313 status
= ocfs2_remove_rightmost_empty_extent(osb
,
7314 &et
, path
, &dealloc
);
7320 ocfs2_reinit_path(path
, 1);
7323 trunc_cpos
= le32_to_cpu(rec
->e_cpos
);
7327 } else if (le32_to_cpu(rec
->e_cpos
) >= new_highest_cpos
) {
7329 * Truncate entire record.
7331 trunc_cpos
= le32_to_cpu(rec
->e_cpos
);
7332 trunc_len
= ocfs2_rec_clusters(el
, rec
);
7333 blkno
= le64_to_cpu(rec
->e_blkno
);
7334 } else if (range
> new_highest_cpos
) {
7336 * Partial truncate. it also should be
7337 * the last truncate we're doing.
7339 trunc_cpos
= new_highest_cpos
;
7340 trunc_len
= range
- new_highest_cpos
;
7341 coff
= new_highest_cpos
- le32_to_cpu(rec
->e_cpos
);
7342 blkno
= le64_to_cpu(rec
->e_blkno
) +
7343 ocfs2_clusters_to_blocks(inode
->i_sb
, coff
);
7346 * Truncate completed, leave happily.
7352 phys_cpos
= ocfs2_blocks_to_clusters(inode
->i_sb
, blkno
);
7354 if ((flags
& OCFS2_EXT_REFCOUNTED
) && trunc_len
&& !ref_tree
) {
7355 status
= ocfs2_lock_refcount_tree(osb
, refcount_loc
, 1,
7363 status
= ocfs2_remove_btree_range(inode
, &et
, trunc_cpos
,
7364 phys_cpos
, trunc_len
, flags
, &dealloc
,
7365 refcount_loc
, true);
7371 ocfs2_reinit_path(path
, 1);
7374 * The check above will catch the case where we've truncated
7375 * away all allocation.
7381 ocfs2_unlock_refcount_tree(osb
, ref_tree
, 1);
7383 ocfs2_schedule_truncate_log_flush(osb
, 1);
7385 ocfs2_run_deallocs(osb
, &dealloc
);
7387 ocfs2_free_path(path
);
7393 * 'start' is inclusive, 'end' is not.
7395 int ocfs2_truncate_inline(struct inode
*inode
, struct buffer_head
*di_bh
,
7396 unsigned int start
, unsigned int end
, int trunc
)
7399 unsigned int numbytes
;
7401 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
7402 struct ocfs2_dinode
*di
= (struct ocfs2_dinode
*)di_bh
->b_data
;
7403 struct ocfs2_inline_data
*idata
= &di
->id2
.i_data
;
7405 /* No need to punch hole beyond i_size. */
7406 if (start
>= i_size_read(inode
))
7409 if (end
> i_size_read(inode
))
7410 end
= i_size_read(inode
);
7412 BUG_ON(start
> end
);
7414 if (!(OCFS2_I(inode
)->ip_dyn_features
& OCFS2_INLINE_DATA_FL
) ||
7415 !(le16_to_cpu(di
->i_dyn_features
) & OCFS2_INLINE_DATA_FL
) ||
7416 !ocfs2_supports_inline_data(osb
)) {
7417 ocfs2_error(inode
->i_sb
,
7418 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7419 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
7420 le16_to_cpu(di
->i_dyn_features
),
7421 OCFS2_I(inode
)->ip_dyn_features
,
7422 osb
->s_feature_incompat
);
7427 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
7428 if (IS_ERR(handle
)) {
7429 ret
= PTR_ERR(handle
);
7434 ret
= ocfs2_journal_access_di(handle
, INODE_CACHE(inode
), di_bh
,
7435 OCFS2_JOURNAL_ACCESS_WRITE
);
7441 numbytes
= end
- start
;
7442 memset(idata
->id_data
+ start
, 0, numbytes
);
7445 * No need to worry about the data page here - it's been
7446 * truncated already and inline data doesn't need it for
7447 * pushing zero's to disk, so we'll let readpage pick it up
7451 i_size_write(inode
, start
);
7452 di
->i_size
= cpu_to_le64(start
);
7455 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
7456 inode
->i_ctime
= inode
->i_mtime
= current_time(inode
);
7458 di
->i_ctime
= di
->i_mtime
= cpu_to_le64(inode
->i_ctime
.tv_sec
);
7459 di
->i_ctime_nsec
= di
->i_mtime_nsec
= cpu_to_le32(inode
->i_ctime
.tv_nsec
);
7461 ocfs2_update_inode_fsync_trans(handle
, inode
, 1);
7462 ocfs2_journal_dirty(handle
, di_bh
);
7465 ocfs2_commit_trans(osb
, handle
);
7471 static int ocfs2_trim_extent(struct super_block
*sb
,
7472 struct ocfs2_group_desc
*gd
,
7473 u64 group
, u32 start
, u32 count
)
7475 u64 discard
, bcount
;
7476 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
7478 bcount
= ocfs2_clusters_to_blocks(sb
, count
);
7479 discard
= ocfs2_clusters_to_blocks(sb
, start
);
7482 * For the first cluster group, the gd->bg_blkno is not at the start
7483 * of the group, but at an offset from the start. If we add it while
7484 * calculating discard for first group, we will wrongly start fstrim a
7485 * few blocks after the desried start block and the range can cross
7486 * over into the next cluster group. So, add it only if this is not
7487 * the first cluster group.
7489 if (group
!= osb
->first_cluster_group_blkno
)
7490 discard
+= le64_to_cpu(gd
->bg_blkno
);
7492 trace_ocfs2_trim_extent(sb
, (unsigned long long)discard
, bcount
);
7494 return sb_issue_discard(sb
, discard
, bcount
, GFP_NOFS
, 0);
7497 static int ocfs2_trim_group(struct super_block
*sb
,
7498 struct ocfs2_group_desc
*gd
, u64 group
,
7499 u32 start
, u32 max
, u32 minbits
)
7501 int ret
= 0, count
= 0, next
;
7502 void *bitmap
= gd
->bg_bitmap
;
7504 if (le16_to_cpu(gd
->bg_free_bits_count
) < minbits
)
7507 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd
->bg_blkno
),
7508 start
, max
, minbits
);
7510 while (start
< max
) {
7511 start
= ocfs2_find_next_zero_bit(bitmap
, max
, start
);
7514 next
= ocfs2_find_next_bit(bitmap
, max
, start
);
7516 if ((next
- start
) >= minbits
) {
7517 ret
= ocfs2_trim_extent(sb
, gd
, group
,
7518 start
, next
- start
);
7523 count
+= next
- start
;
7527 if (fatal_signal_pending(current
)) {
7528 count
= -ERESTARTSYS
;
7532 if ((le16_to_cpu(gd
->bg_free_bits_count
) - count
) < minbits
)
7543 int ocfs2_trim_mainbm(struct super_block
*sb
, struct fstrim_range
*range
)
7545 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
7546 u64 start
, len
, trimmed
= 0, first_group
, last_group
= 0, group
= 0;
7548 u32 first_bit
, last_bit
, minlen
;
7549 struct buffer_head
*main_bm_bh
= NULL
;
7550 struct inode
*main_bm_inode
= NULL
;
7551 struct buffer_head
*gd_bh
= NULL
;
7552 struct ocfs2_dinode
*main_bm
;
7553 struct ocfs2_group_desc
*gd
= NULL
;
7555 start
= range
->start
>> osb
->s_clustersize_bits
;
7556 len
= range
->len
>> osb
->s_clustersize_bits
;
7557 minlen
= range
->minlen
>> osb
->s_clustersize_bits
;
7559 if (minlen
>= osb
->bitmap_cpg
|| range
->len
< sb
->s_blocksize
)
7562 trace_ocfs2_trim_mainbm(start
, len
, minlen
);
7565 main_bm_inode
= ocfs2_get_system_file_inode(osb
,
7566 GLOBAL_BITMAP_SYSTEM_INODE
,
7567 OCFS2_INVALID_SLOT
);
7568 if (!main_bm_inode
) {
7574 inode_lock(main_bm_inode
);
7576 ret
= ocfs2_inode_lock(main_bm_inode
, &main_bm_bh
, 0);
7581 main_bm
= (struct ocfs2_dinode
*)main_bm_bh
->b_data
;
7584 * Do some check before trim the first group.
7587 if (start
>= le32_to_cpu(main_bm
->i_clusters
)) {
7592 if (start
+ len
> le32_to_cpu(main_bm
->i_clusters
))
7593 len
= le32_to_cpu(main_bm
->i_clusters
) - start
;
7596 * Determine first and last group to examine based on
7599 first_group
= ocfs2_which_cluster_group(main_bm_inode
, start
);
7600 if (first_group
== osb
->first_cluster_group_blkno
)
7603 first_bit
= start
- ocfs2_blocks_to_clusters(sb
,
7605 last_group
= ocfs2_which_cluster_group(main_bm_inode
,
7607 group
= first_group
;
7611 if (first_bit
+ len
>= osb
->bitmap_cpg
)
7612 last_bit
= osb
->bitmap_cpg
;
7614 last_bit
= first_bit
+ len
;
7616 ret
= ocfs2_read_group_descriptor(main_bm_inode
,
7624 gd
= (struct ocfs2_group_desc
*)gd_bh
->b_data
;
7625 cnt
= ocfs2_trim_group(sb
, gd
, group
,
7626 first_bit
, last_bit
, minlen
);
7636 len
-= osb
->bitmap_cpg
- first_bit
;
7638 if (group
== osb
->first_cluster_group_blkno
)
7639 group
= ocfs2_clusters_to_blocks(sb
, osb
->bitmap_cpg
);
7641 group
+= ocfs2_clusters_to_blocks(sb
, osb
->bitmap_cpg
);
7645 ocfs2_inode_unlock(main_bm_inode
, 0);
7649 inode_unlock(main_bm_inode
);
7650 iput(main_bm_inode
);
7653 * If all the groups trim are not done or failed, but we should release
7654 * main_bm related locks for avoiding the current IO starve, then go to
7655 * trim the next group
7657 if (ret
>= 0 && group
<= last_group
) {
7662 range
->len
= trimmed
* sb
->s_blocksize
;
7666 int ocfs2_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
7669 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
7670 struct ocfs2_trim_fs_info info
, *pinfo
= NULL
;
7672 ocfs2_trim_fs_lock_res_init(osb
);
7674 trace_ocfs2_trim_fs(range
->start
, range
->len
, range
->minlen
);
7676 ret
= ocfs2_trim_fs_lock(osb
, NULL
, 1);
7678 if (ret
!= -EAGAIN
) {
7680 ocfs2_trim_fs_lock_res_uninit(osb
);
7684 mlog(ML_NOTICE
, "Wait for trim on device (%s) to "
7685 "finish, which is running from another node.\n",
7687 ret
= ocfs2_trim_fs_lock(osb
, &info
, 0);
7690 ocfs2_trim_fs_lock_res_uninit(osb
);
7694 if (info
.tf_valid
&& info
.tf_success
&&
7695 info
.tf_start
== range
->start
&&
7696 info
.tf_len
== range
->len
&&
7697 info
.tf_minlen
== range
->minlen
) {
7698 /* Avoid sending duplicated trim to a shared device */
7699 mlog(ML_NOTICE
, "The same trim on device (%s) was "
7700 "just done from node (%u), return.\n",
7701 osb
->dev_str
, info
.tf_nodenum
);
7702 range
->len
= info
.tf_trimlen
;
7707 info
.tf_nodenum
= osb
->node_num
;
7708 info
.tf_start
= range
->start
;
7709 info
.tf_len
= range
->len
;
7710 info
.tf_minlen
= range
->minlen
;
7712 ret
= ocfs2_trim_mainbm(sb
, range
);
7714 info
.tf_trimlen
= range
->len
;
7715 info
.tf_success
= (ret
< 0 ? 0 : 1);
7718 ocfs2_trim_fs_unlock(osb
, pinfo
);
7719 ocfs2_trim_fs_lock_res_uninit(osb
);