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++) {
1069 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1071 * Returns the sum of the rightmost extent rec logical offset and
1074 * ocfs2_add_branch() uses this to determine what logical cluster
1075 * value should be populated into the leftmost new branch records.
1077 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1078 * value for the new topmost tree record.
1080 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1084 i = le16_to_cpu(el->l_next_free_rec) - 1;
1086 return le32_to_cpu(el->l_recs[i].e_cpos) +
1087 ocfs2_rec_clusters(el, &el->l_recs[i]);
1091 * Change range of the branches in the right most path according to the leaf
1092 * extent block's rightmost record.
1094 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1095 struct ocfs2_extent_tree *et)
1098 struct ocfs2_path *path = NULL;
1099 struct ocfs2_extent_list *el;
1100 struct ocfs2_extent_rec *rec;
1102 path = ocfs2_new_path_from_et(et);
1108 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1114 status = ocfs2_extend_trans(handle, path_num_items(path));
1120 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1126 el = path_leaf_el(path);
1127 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1129 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1132 ocfs2_free_path(path);
1137 * Add an entire tree branch to our inode. eb_bh is the extent block
1138 * to start at, if we don't want to start the branch at the root
1141 * last_eb_bh is required as we have to update it's next_leaf pointer
1142 * for the new last extent block.
1144 * the new branch will be 'empty' in the sense that every block will
1145 * contain a single record with cluster count == 0.
1147 static int ocfs2_add_branch(handle_t *handle,
1148 struct ocfs2_extent_tree *et,
1149 struct buffer_head *eb_bh,
1150 struct buffer_head **last_eb_bh,
1151 struct ocfs2_alloc_context *meta_ac)
1153 int status, new_blocks, i, block_given = 0;
1154 u64 next_blkno, new_last_eb_blk;
1155 struct buffer_head *bh;
1156 struct buffer_head **new_eb_bhs = NULL;
1157 struct ocfs2_extent_block *eb;
1158 struct ocfs2_extent_list *eb_el;
1159 struct ocfs2_extent_list *el;
1160 u32 new_cpos, root_end;
1162 BUG_ON(!last_eb_bh || !*last_eb_bh);
1165 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1168 el = et->et_root_el;
1170 /* we never add a branch to a leaf. */
1171 BUG_ON(!el->l_tree_depth);
1173 new_blocks = le16_to_cpu(el->l_tree_depth);
1175 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1176 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1177 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1180 * If there is a gap before the root end and the real end
1181 * of the righmost leaf block, we need to remove the gap
1182 * between new_cpos and root_end first so that the tree
1183 * is consistent after we add a new branch(it will start
1186 if (root_end > new_cpos) {
1187 trace_ocfs2_adjust_rightmost_branch(
1188 (unsigned long long)
1189 ocfs2_metadata_cache_owner(et->et_ci),
1190 root_end, new_cpos);
1192 status = ocfs2_adjust_rightmost_branch(handle, et);
1199 /* allocate the number of new eb blocks we need */
1200 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1208 /* Firstyly, try to reuse dealloc since we have already estimated how
1209 * many extent blocks we may use.
1211 if (!ocfs2_is_dealloc_empty(et)) {
1212 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1213 new_eb_bhs, new_blocks,
1221 BUG_ON(block_given > new_blocks);
1223 if (block_given < new_blocks) {
1225 status = ocfs2_create_new_meta_bhs(handle, et,
1226 new_blocks - block_given,
1228 &new_eb_bhs[block_given]);
1235 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1236 * linked with the rest of the tree.
1237 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1239 * when we leave the loop, new_last_eb_blk will point to the
1240 * newest leaf, and next_blkno will point to the topmost extent
1242 next_blkno = new_last_eb_blk = 0;
1243 for(i = 0; i < new_blocks; i++) {
1245 eb = (struct ocfs2_extent_block *) bh->b_data;
1246 /* ocfs2_create_new_meta_bhs() should create it right! */
1247 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1248 eb_el = &eb->h_list;
1250 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1251 OCFS2_JOURNAL_ACCESS_CREATE);
1257 eb->h_next_leaf_blk = 0;
1258 eb_el->l_tree_depth = cpu_to_le16(i);
1259 eb_el->l_next_free_rec = cpu_to_le16(1);
1261 * This actually counts as an empty extent as
1264 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1265 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1267 * eb_el isn't always an interior node, but even leaf
1268 * nodes want a zero'd flags and reserved field so
1269 * this gets the whole 32 bits regardless of use.
1271 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1272 if (!eb_el->l_tree_depth)
1273 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1275 ocfs2_journal_dirty(handle, bh);
1276 next_blkno = le64_to_cpu(eb->h_blkno);
1279 /* This is a bit hairy. We want to update up to three blocks
1280 * here without leaving any of them in an inconsistent state
1281 * in case of error. We don't have to worry about
1282 * journal_dirty erroring as it won't unless we've aborted the
1283 * handle (in which case we would never be here) so reserving
1284 * the write with journal_access is all we need to do. */
1285 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1286 OCFS2_JOURNAL_ACCESS_WRITE);
1291 status = ocfs2_et_root_journal_access(handle, et,
1292 OCFS2_JOURNAL_ACCESS_WRITE);
1298 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1299 OCFS2_JOURNAL_ACCESS_WRITE);
1306 /* Link the new branch into the rest of the tree (el will
1307 * either be on the root_bh, or the extent block passed in. */
1308 i = le16_to_cpu(el->l_next_free_rec);
1309 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1310 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1311 el->l_recs[i].e_int_clusters = 0;
1312 le16_add_cpu(&el->l_next_free_rec, 1);
1314 /* fe needs a new last extent block pointer, as does the
1315 * next_leaf on the previously last-extent-block. */
1316 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1318 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1319 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1321 ocfs2_journal_dirty(handle, *last_eb_bh);
1322 ocfs2_journal_dirty(handle, et->et_root_bh);
1324 ocfs2_journal_dirty(handle, eb_bh);
1327 * Some callers want to track the rightmost leaf so pass it
1330 brelse(*last_eb_bh);
1331 get_bh(new_eb_bhs[0]);
1332 *last_eb_bh = new_eb_bhs[0];
1337 for (i = 0; i < new_blocks; i++)
1338 brelse(new_eb_bhs[i]);
1346 * adds another level to the allocation tree.
1347 * returns back the new extent block so you can add a branch to it
1350 static int ocfs2_shift_tree_depth(handle_t *handle,
1351 struct ocfs2_extent_tree *et,
1352 struct ocfs2_alloc_context *meta_ac,
1353 struct buffer_head **ret_new_eb_bh)
1355 int status, i, block_given = 0;
1357 struct buffer_head *new_eb_bh = NULL;
1358 struct ocfs2_extent_block *eb;
1359 struct ocfs2_extent_list *root_el;
1360 struct ocfs2_extent_list *eb_el;
1362 if (!ocfs2_is_dealloc_empty(et)) {
1363 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1366 } else if (meta_ac) {
1367 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1379 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1380 /* ocfs2_create_new_meta_bhs() should create it right! */
1381 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1383 eb_el = &eb->h_list;
1384 root_el = et->et_root_el;
1386 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1387 OCFS2_JOURNAL_ACCESS_CREATE);
1393 /* copy the root extent list data into the new extent block */
1394 eb_el->l_tree_depth = root_el->l_tree_depth;
1395 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1396 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1397 eb_el->l_recs[i] = root_el->l_recs[i];
1399 ocfs2_journal_dirty(handle, new_eb_bh);
1401 status = ocfs2_et_root_journal_access(handle, et,
1402 OCFS2_JOURNAL_ACCESS_WRITE);
1408 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1410 /* update root_bh now */
1411 le16_add_cpu(&root_el->l_tree_depth, 1);
1412 root_el->l_recs[0].e_cpos = 0;
1413 root_el->l_recs[0].e_blkno = eb->h_blkno;
1414 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1415 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1416 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1417 root_el->l_next_free_rec = cpu_to_le16(1);
1419 /* If this is our 1st tree depth shift, then last_eb_blk
1420 * becomes the allocated extent block */
1421 if (root_el->l_tree_depth == cpu_to_le16(1))
1422 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1424 ocfs2_journal_dirty(handle, et->et_root_bh);
1426 *ret_new_eb_bh = new_eb_bh;
1436 * Should only be called when there is no space left in any of the
1437 * leaf nodes. What we want to do is find the lowest tree depth
1438 * non-leaf extent block with room for new records. There are three
1439 * valid results of this search:
1441 * 1) a lowest extent block is found, then we pass it back in
1442 * *lowest_eb_bh and return '0'
1444 * 2) the search fails to find anything, but the root_el has room. We
1445 * pass NULL back in *lowest_eb_bh, but still return '0'
1447 * 3) the search fails to find anything AND the root_el is full, in
1448 * which case we return > 0
1450 * return status < 0 indicates an error.
1452 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1453 struct buffer_head **target_bh)
1457 struct ocfs2_extent_block *eb;
1458 struct ocfs2_extent_list *el;
1459 struct buffer_head *bh = NULL;
1460 struct buffer_head *lowest_bh = NULL;
1464 el = et->et_root_el;
1466 while(le16_to_cpu(el->l_tree_depth) > 1) {
1467 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1468 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1469 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1470 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1473 i = le16_to_cpu(el->l_next_free_rec) - 1;
1474 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1476 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1477 "Owner %llu has extent list where extent # %d has no physical block start\n",
1478 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1485 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1491 eb = (struct ocfs2_extent_block *) bh->b_data;
1494 if (le16_to_cpu(el->l_next_free_rec) <
1495 le16_to_cpu(el->l_count)) {
1502 /* If we didn't find one and the fe doesn't have any room,
1503 * then return '1' */
1504 el = et->et_root_el;
1505 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1508 *target_bh = lowest_bh;
1516 * Grow a b-tree so that it has more records.
1518 * We might shift the tree depth in which case existing paths should
1519 * be considered invalid.
1521 * Tree depth after the grow is returned via *final_depth.
1523 * *last_eb_bh will be updated by ocfs2_add_branch().
1525 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1526 int *final_depth, struct buffer_head **last_eb_bh,
1527 struct ocfs2_alloc_context *meta_ac)
1530 struct ocfs2_extent_list *el = et->et_root_el;
1531 int depth = le16_to_cpu(el->l_tree_depth);
1532 struct buffer_head *bh = NULL;
1534 BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));
1536 shift = ocfs2_find_branch_target(et, &bh);
1543 /* We traveled all the way to the bottom of the allocation tree
1544 * and didn't find room for any more extents - we need to add
1545 * another tree level */
1548 trace_ocfs2_grow_tree(
1549 (unsigned long long)
1550 ocfs2_metadata_cache_owner(et->et_ci),
1553 /* ocfs2_shift_tree_depth will return us a buffer with
1554 * the new extent block (so we can pass that to
1555 * ocfs2_add_branch). */
1556 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1564 * Special case: we have room now if we shifted from
1565 * tree_depth 0, so no more work needs to be done.
1567 * We won't be calling add_branch, so pass
1568 * back *last_eb_bh as the new leaf. At depth
1569 * zero, it should always be null so there's
1570 * no reason to brelse.
1572 BUG_ON(*last_eb_bh);
1579 /* call ocfs2_add_branch to add the final part of the tree with
1581 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1588 *final_depth = depth;
1594 * This function will discard the rightmost extent record.
1596 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1598 int next_free = le16_to_cpu(el->l_next_free_rec);
1599 int count = le16_to_cpu(el->l_count);
1600 unsigned int num_bytes;
1603 /* This will cause us to go off the end of our extent list. */
1604 BUG_ON(next_free >= count);
1606 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1608 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1611 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1612 struct ocfs2_extent_rec *insert_rec)
1614 int i, insert_index, next_free, has_empty, num_bytes;
1615 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1616 struct ocfs2_extent_rec *rec;
1618 next_free = le16_to_cpu(el->l_next_free_rec);
1619 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1623 /* The tree code before us didn't allow enough room in the leaf. */
1624 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1627 * The easiest way to approach this is to just remove the
1628 * empty extent and temporarily decrement next_free.
1632 * If next_free was 1 (only an empty extent), this
1633 * loop won't execute, which is fine. We still want
1634 * the decrement above to happen.
1636 for(i = 0; i < (next_free - 1); i++)
1637 el->l_recs[i] = el->l_recs[i+1];
1643 * Figure out what the new record index should be.
1645 for(i = 0; i < next_free; i++) {
1646 rec = &el->l_recs[i];
1648 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1653 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1654 has_empty, next_free,
1655 le16_to_cpu(el->l_count));
1657 BUG_ON(insert_index < 0);
1658 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1659 BUG_ON(insert_index > next_free);
1662 * No need to memmove if we're just adding to the tail.
1664 if (insert_index != next_free) {
1665 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1667 num_bytes = next_free - insert_index;
1668 num_bytes *= sizeof(struct ocfs2_extent_rec);
1669 memmove(&el->l_recs[insert_index + 1],
1670 &el->l_recs[insert_index],
1675 * Either we had an empty extent, and need to re-increment or
1676 * there was no empty extent on a non full rightmost leaf node,
1677 * in which case we still need to increment.
1680 el->l_next_free_rec = cpu_to_le16(next_free);
1682 * Make sure none of the math above just messed up our tree.
1684 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1686 el->l_recs[insert_index] = *insert_rec;
1690 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1692 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1694 BUG_ON(num_recs == 0);
1696 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1698 size = num_recs * sizeof(struct ocfs2_extent_rec);
1699 memmove(&el->l_recs[0], &el->l_recs[1], size);
1700 memset(&el->l_recs[num_recs], 0,
1701 sizeof(struct ocfs2_extent_rec));
1702 el->l_next_free_rec = cpu_to_le16(num_recs);
1707 * Create an empty extent record .
1709 * l_next_free_rec may be updated.
1711 * If an empty extent already exists do nothing.
1713 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1715 int next_free = le16_to_cpu(el->l_next_free_rec);
1717 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1722 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1725 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1726 "Asked to create an empty extent in a full list:\n"
1727 "count = %u, tree depth = %u",
1728 le16_to_cpu(el->l_count),
1729 le16_to_cpu(el->l_tree_depth));
1731 ocfs2_shift_records_right(el);
1734 le16_add_cpu(&el->l_next_free_rec, 1);
1735 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1739 * For a rotation which involves two leaf nodes, the "root node" is
1740 * the lowest level tree node which contains a path to both leafs. This
1741 * resulting set of information can be used to form a complete "subtree"
1743 * This function is passed two full paths from the dinode down to a
1744 * pair of adjacent leaves. It's task is to figure out which path
1745 * index contains the subtree root - this can be the root index itself
1746 * in a worst-case rotation.
1748 * The array index of the subtree root is passed back.
1750 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1751 struct ocfs2_path *left,
1752 struct ocfs2_path *right)
1757 * Check that the caller passed in two paths from the same tree.
1759 BUG_ON(path_root_bh(left) != path_root_bh(right));
1765 * The caller didn't pass two adjacent paths.
1767 mlog_bug_on_msg(i > left->p_tree_depth,
1768 "Owner %llu, left depth %u, right depth %u\n"
1769 "left leaf blk %llu, right leaf blk %llu\n",
1770 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1771 left->p_tree_depth, right->p_tree_depth,
1772 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1773 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1774 } while (left->p_node[i].bh->b_blocknr ==
1775 right->p_node[i].bh->b_blocknr);
1780 typedef void (path_insert_t)(void *, struct buffer_head *);
1783 * Traverse a btree path in search of cpos, starting at root_el.
1785 * This code can be called with a cpos larger than the tree, in which
1786 * case it will return the rightmost path.
1788 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1789 struct ocfs2_extent_list *root_el, u32 cpos,
1790 path_insert_t *func, void *data)
1795 struct buffer_head *bh = NULL;
1796 struct ocfs2_extent_block *eb;
1797 struct ocfs2_extent_list *el;
1798 struct ocfs2_extent_rec *rec;
1801 while (el->l_tree_depth) {
1802 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1803 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1804 "Owner %llu has empty extent list at depth %u\n",
1805 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1806 le16_to_cpu(el->l_tree_depth));
1812 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1813 rec = &el->l_recs[i];
1816 * In the case that cpos is off the allocation
1817 * tree, this should just wind up returning the
1820 range = le32_to_cpu(rec->e_cpos) +
1821 ocfs2_rec_clusters(el, rec);
1822 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1826 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1828 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1829 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1830 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1831 le16_to_cpu(el->l_tree_depth), i);
1838 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1844 eb = (struct ocfs2_extent_block *) bh->b_data;
1847 if (le16_to_cpu(el->l_next_free_rec) >
1848 le16_to_cpu(el->l_count)) {
1849 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1850 "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1851 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1852 (unsigned long long)bh->b_blocknr,
1853 le16_to_cpu(el->l_next_free_rec),
1854 le16_to_cpu(el->l_count));
1865 * Catch any trailing bh that the loop didn't handle.
1873 * Given an initialized path (that is, it has a valid root extent
1874 * list), this function will traverse the btree in search of the path
1875 * which would contain cpos.
1877 * The path traveled is recorded in the path structure.
1879 * Note that this will not do any comparisons on leaf node extent
1880 * records, so it will work fine in the case that we just added a tree
1883 struct find_path_data {
1885 struct ocfs2_path *path;
1887 static void find_path_ins(void *data, struct buffer_head *bh)
1889 struct find_path_data *fp = data;
1892 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1895 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1896 struct ocfs2_path *path, u32 cpos)
1898 struct find_path_data data;
1902 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1903 find_path_ins, &data);
1906 static void find_leaf_ins(void *data, struct buffer_head *bh)
1908 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1909 struct ocfs2_extent_list *el = &eb->h_list;
1910 struct buffer_head **ret = data;
1912 /* We want to retain only the leaf block. */
1913 if (le16_to_cpu(el->l_tree_depth) == 0) {
1919 * Find the leaf block in the tree which would contain cpos. No
1920 * checking of the actual leaf is done.
1922 * Some paths want to call this instead of allocating a path structure
1923 * and calling ocfs2_find_path().
1925 * This function doesn't handle non btree extent lists.
1927 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1928 struct ocfs2_extent_list *root_el, u32 cpos,
1929 struct buffer_head **leaf_bh)
1932 struct buffer_head *bh = NULL;
1934 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1946 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1948 * Basically, we've moved stuff around at the bottom of the tree and
1949 * we need to fix up the extent records above the changes to reflect
1952 * left_rec: the record on the left.
1953 * right_rec: the record to the right of left_rec
1954 * right_child_el: is the child list pointed to by right_rec
1956 * By definition, this only works on interior nodes.
1958 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1959 struct ocfs2_extent_rec *right_rec,
1960 struct ocfs2_extent_list *right_child_el)
1962 u32 left_clusters, right_end;
1965 * Interior nodes never have holes. Their cpos is the cpos of
1966 * the leftmost record in their child list. Their cluster
1967 * count covers the full theoretical range of their child list
1968 * - the range between their cpos and the cpos of the record
1969 * immediately to their right.
1971 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1972 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1973 BUG_ON(right_child_el->l_tree_depth);
1974 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1975 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1977 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1978 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1981 * Calculate the rightmost cluster count boundary before
1982 * moving cpos - we will need to adjust clusters after
1983 * updating e_cpos to keep the same highest cluster count.
1985 right_end = le32_to_cpu(right_rec->e_cpos);
1986 right_end += le32_to_cpu(right_rec->e_int_clusters);
1988 right_rec->e_cpos = left_rec->e_cpos;
1989 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1991 right_end -= le32_to_cpu(right_rec->e_cpos);
1992 right_rec->e_int_clusters = cpu_to_le32(right_end);
1996 * Adjust the adjacent root node records involved in a
1997 * rotation. left_el_blkno is passed in as a key so that we can easily
1998 * find it's index in the root list.
2000 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
2001 struct ocfs2_extent_list *left_el,
2002 struct ocfs2_extent_list *right_el,
2007 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2008 le16_to_cpu(left_el->l_tree_depth));
2010 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2011 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2016 * The path walking code should have never returned a root and
2017 * two paths which are not adjacent.
2019 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2021 ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
2022 &root_el->l_recs[i + 1], right_el);
2026 * We've changed a leaf block (in right_path) and need to reflect that
2027 * change back up the subtree.
2029 * This happens in multiple places:
2030 * - When we've moved an extent record from the left path leaf to the right
2031 * path leaf to make room for an empty extent in the left path leaf.
2032 * - When our insert into the right path leaf is at the leftmost edge
2033 * and requires an update of the path immediately to it's left. This
2034 * can occur at the end of some types of rotation and appending inserts.
2035 * - When we've adjusted the last extent record in the left path leaf and the
2036 * 1st extent record in the right path leaf during cross extent block merge.
2038 static void ocfs2_complete_edge_insert(handle_t *handle,
2039 struct ocfs2_path *left_path,
2040 struct ocfs2_path *right_path,
2044 struct ocfs2_extent_list *el, *left_el, *right_el;
2045 struct ocfs2_extent_rec *left_rec, *right_rec;
2046 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2049 * Update the counts and position values within all the
2050 * interior nodes to reflect the leaf rotation we just did.
2052 * The root node is handled below the loop.
2054 * We begin the loop with right_el and left_el pointing to the
2055 * leaf lists and work our way up.
2057 * NOTE: within this loop, left_el and right_el always refer
2058 * to the *child* lists.
2060 left_el = path_leaf_el(left_path);
2061 right_el = path_leaf_el(right_path);
2062 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2063 trace_ocfs2_complete_edge_insert(i);
2066 * One nice property of knowing that all of these
2067 * nodes are below the root is that we only deal with
2068 * the leftmost right node record and the rightmost
2071 el = left_path->p_node[i].el;
2072 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2073 left_rec = &el->l_recs[idx];
2075 el = right_path->p_node[i].el;
2076 right_rec = &el->l_recs[0];
2078 ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);
2080 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2081 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2084 * Setup our list pointers now so that the current
2085 * parents become children in the next iteration.
2087 left_el = left_path->p_node[i].el;
2088 right_el = right_path->p_node[i].el;
2092 * At the root node, adjust the two adjacent records which
2093 * begin our path to the leaves.
2096 el = left_path->p_node[subtree_index].el;
2097 left_el = left_path->p_node[subtree_index + 1].el;
2098 right_el = right_path->p_node[subtree_index + 1].el;
2100 ocfs2_adjust_root_records(el, left_el, right_el,
2101 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2103 root_bh = left_path->p_node[subtree_index].bh;
2105 ocfs2_journal_dirty(handle, root_bh);
2108 static int ocfs2_rotate_subtree_right(handle_t *handle,
2109 struct ocfs2_extent_tree *et,
2110 struct ocfs2_path *left_path,
2111 struct ocfs2_path *right_path,
2115 struct buffer_head *right_leaf_bh;
2116 struct buffer_head *left_leaf_bh = NULL;
2117 struct buffer_head *root_bh;
2118 struct ocfs2_extent_list *right_el, *left_el;
2119 struct ocfs2_extent_rec move_rec;
2121 left_leaf_bh = path_leaf_bh(left_path);
2122 left_el = path_leaf_el(left_path);
2124 if (left_el->l_next_free_rec != left_el->l_count) {
2125 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2126 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2127 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2128 (unsigned long long)left_leaf_bh->b_blocknr,
2129 le16_to_cpu(left_el->l_next_free_rec));
2134 * This extent block may already have an empty record, so we
2135 * return early if so.
2137 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2140 root_bh = left_path->p_node[subtree_index].bh;
2141 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2143 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2150 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2151 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2158 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2166 right_leaf_bh = path_leaf_bh(right_path);
2167 right_el = path_leaf_el(right_path);
2169 /* This is a code error, not a disk corruption. */
2170 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2171 "because rightmost leaf block %llu is empty\n",
2172 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2173 (unsigned long long)right_leaf_bh->b_blocknr);
2175 ocfs2_create_empty_extent(right_el);
2177 ocfs2_journal_dirty(handle, right_leaf_bh);
2179 /* Do the copy now. */
2180 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2181 move_rec = left_el->l_recs[i];
2182 right_el->l_recs[0] = move_rec;
2185 * Clear out the record we just copied and shift everything
2186 * over, leaving an empty extent in the left leaf.
2188 * We temporarily subtract from next_free_rec so that the
2189 * shift will lose the tail record (which is now defunct).
2191 le16_add_cpu(&left_el->l_next_free_rec, -1);
2192 ocfs2_shift_records_right(left_el);
2193 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2194 le16_add_cpu(&left_el->l_next_free_rec, 1);
2196 ocfs2_journal_dirty(handle, left_leaf_bh);
2198 ocfs2_complete_edge_insert(handle, left_path, right_path,
2206 * Given a full path, determine what cpos value would return us a path
2207 * containing the leaf immediately to the left of the current one.
2209 * Will return zero if the path passed in is already the leftmost path.
2211 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2212 struct ocfs2_path *path, u32 *cpos)
2216 struct ocfs2_extent_list *el;
2218 BUG_ON(path->p_tree_depth == 0);
2222 blkno = path_leaf_bh(path)->b_blocknr;
2224 /* Start at the tree node just above the leaf and work our way up. */
2225 i = path->p_tree_depth - 1;
2227 el = path->p_node[i].el;
2230 * Find the extent record just before the one in our
2233 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2234 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2238 * We've determined that the
2239 * path specified is already
2240 * the leftmost one - return a
2246 * The leftmost record points to our
2247 * leaf - we need to travel up the
2253 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2254 *cpos = *cpos + ocfs2_rec_clusters(el,
2255 &el->l_recs[j - 1]);
2262 * If we got here, we never found a valid node where
2263 * the tree indicated one should be.
2265 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2266 (unsigned long long)blkno);
2271 blkno = path->p_node[i].bh->b_blocknr;
2280 * Extend the transaction by enough credits to complete the rotation,
2281 * and still leave at least the original number of credits allocated
2282 * to this transaction.
2284 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2286 struct ocfs2_path *path)
2289 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2291 if (handle->h_buffer_credits < credits)
2292 ret = ocfs2_extend_trans(handle,
2293 credits - handle->h_buffer_credits);
2299 * Trap the case where we're inserting into the theoretical range past
2300 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2301 * whose cpos is less than ours into the right leaf.
2303 * It's only necessary to look at the rightmost record of the left
2304 * leaf because the logic that calls us should ensure that the
2305 * theoretical ranges in the path components above the leaves are
2308 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2311 struct ocfs2_extent_list *left_el;
2312 struct ocfs2_extent_rec *rec;
2315 left_el = path_leaf_el(left_path);
2316 next_free = le16_to_cpu(left_el->l_next_free_rec);
2317 rec = &left_el->l_recs[next_free - 1];
2319 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2324 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2326 int next_free = le16_to_cpu(el->l_next_free_rec);
2328 struct ocfs2_extent_rec *rec;
2333 rec = &el->l_recs[0];
2334 if (ocfs2_is_empty_extent(rec)) {
2338 rec = &el->l_recs[1];
2341 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2342 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2348 * Rotate all the records in a btree right one record, starting at insert_cpos.
2350 * The path to the rightmost leaf should be passed in.
2352 * The array is assumed to be large enough to hold an entire path (tree depth).
2354 * Upon successful return from this function:
2356 * - The 'right_path' array will contain a path to the leaf block
2357 * whose range contains e_cpos.
2358 * - That leaf block will have a single empty extent in list index 0.
2359 * - In the case that the rotation requires a post-insert update,
2360 * *ret_left_path will contain a valid path which can be passed to
2361 * ocfs2_insert_path().
2363 static int ocfs2_rotate_tree_right(handle_t *handle,
2364 struct ocfs2_extent_tree *et,
2365 enum ocfs2_split_type split,
2367 struct ocfs2_path *right_path,
2368 struct ocfs2_path **ret_left_path)
2370 int ret, start, orig_credits = handle->h_buffer_credits;
2372 struct ocfs2_path *left_path = NULL;
2373 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2375 *ret_left_path = NULL;
2377 left_path = ocfs2_new_path_from_path(right_path);
2384 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2390 trace_ocfs2_rotate_tree_right(
2391 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2395 * What we want to do here is:
2397 * 1) Start with the rightmost path.
2399 * 2) Determine a path to the leaf block directly to the left
2402 * 3) Determine the 'subtree root' - the lowest level tree node
2403 * which contains a path to both leaves.
2405 * 4) Rotate the subtree.
2407 * 5) Find the next subtree by considering the left path to be
2408 * the new right path.
2410 * The check at the top of this while loop also accepts
2411 * insert_cpos == cpos because cpos is only a _theoretical_
2412 * value to get us the left path - insert_cpos might very well
2413 * be filling that hole.
2415 * Stop at a cpos of '0' because we either started at the
2416 * leftmost branch (i.e., a tree with one branch and a
2417 * rotation inside of it), or we've gone as far as we can in
2418 * rotating subtrees.
2420 while (cpos && insert_cpos <= cpos) {
2421 trace_ocfs2_rotate_tree_right(
2422 (unsigned long long)
2423 ocfs2_metadata_cache_owner(et->et_ci),
2426 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2432 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2433 path_leaf_bh(right_path),
2434 "Owner %llu: error during insert of %u "
2435 "(left path cpos %u) results in two identical "
2436 "paths ending at %llu\n",
2437 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2439 (unsigned long long)
2440 path_leaf_bh(left_path)->b_blocknr);
2442 if (split == SPLIT_NONE &&
2443 ocfs2_rotate_requires_path_adjustment(left_path,
2447 * We've rotated the tree as much as we
2448 * should. The rest is up to
2449 * ocfs2_insert_path() to complete, after the
2450 * record insertion. We indicate this
2451 * situation by returning the left path.
2453 * The reason we don't adjust the records here
2454 * before the record insert is that an error
2455 * later might break the rule where a parent
2456 * record e_cpos will reflect the actual
2457 * e_cpos of the 1st nonempty record of the
2460 *ret_left_path = left_path;
2464 start = ocfs2_find_subtree_root(et, left_path, right_path);
2466 trace_ocfs2_rotate_subtree(start,
2467 (unsigned long long)
2468 right_path->p_node[start].bh->b_blocknr,
2469 right_path->p_tree_depth);
2471 ret = ocfs2_extend_rotate_transaction(handle, start,
2472 orig_credits, right_path);
2478 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2485 if (split != SPLIT_NONE &&
2486 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2489 * A rotate moves the rightmost left leaf
2490 * record over to the leftmost right leaf
2491 * slot. If we're doing an extent split
2492 * instead of a real insert, then we have to
2493 * check that the extent to be split wasn't
2494 * just moved over. If it was, then we can
2495 * exit here, passing left_path back -
2496 * ocfs2_split_extent() is smart enough to
2497 * search both leaves.
2499 *ret_left_path = left_path;
2504 * There is no need to re-read the next right path
2505 * as we know that it'll be our current left
2506 * path. Optimize by copying values instead.
2508 ocfs2_mv_path(right_path, left_path);
2510 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2518 ocfs2_free_path(left_path);
2524 static int ocfs2_update_edge_lengths(handle_t *handle,
2525 struct ocfs2_extent_tree *et,
2526 struct ocfs2_path *path)
2529 struct ocfs2_extent_rec *rec;
2530 struct ocfs2_extent_list *el;
2531 struct ocfs2_extent_block *eb;
2534 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2540 /* Path should always be rightmost. */
2541 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2542 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2545 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2546 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2547 rec = &el->l_recs[idx];
2548 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2550 for (i = 0; i < path->p_tree_depth; i++) {
2551 el = path->p_node[i].el;
2552 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2553 rec = &el->l_recs[idx];
2555 rec->e_int_clusters = cpu_to_le32(range);
2556 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2558 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2564 static void ocfs2_unlink_path(handle_t *handle,
2565 struct ocfs2_extent_tree *et,
2566 struct ocfs2_cached_dealloc_ctxt *dealloc,
2567 struct ocfs2_path *path, int unlink_start)
2570 struct ocfs2_extent_block *eb;
2571 struct ocfs2_extent_list *el;
2572 struct buffer_head *bh;
2574 for(i = unlink_start; i < path_num_items(path); i++) {
2575 bh = path->p_node[i].bh;
2577 eb = (struct ocfs2_extent_block *)bh->b_data;
2579 * Not all nodes might have had their final count
2580 * decremented by the caller - handle this here.
2583 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2585 "Inode %llu, attempted to remove extent block "
2586 "%llu with %u records\n",
2587 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2588 (unsigned long long)le64_to_cpu(eb->h_blkno),
2589 le16_to_cpu(el->l_next_free_rec));
2591 ocfs2_journal_dirty(handle, bh);
2592 ocfs2_remove_from_cache(et->et_ci, bh);
2596 el->l_next_free_rec = 0;
2597 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2599 ocfs2_journal_dirty(handle, bh);
2601 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2605 ocfs2_remove_from_cache(et->et_ci, bh);
2609 static void ocfs2_unlink_subtree(handle_t *handle,
2610 struct ocfs2_extent_tree *et,
2611 struct ocfs2_path *left_path,
2612 struct ocfs2_path *right_path,
2614 struct ocfs2_cached_dealloc_ctxt *dealloc)
2617 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2618 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2619 struct ocfs2_extent_block *eb;
2621 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2623 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2624 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2627 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2629 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2630 le16_add_cpu(&root_el->l_next_free_rec, -1);
2632 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2633 eb->h_next_leaf_blk = 0;
2635 ocfs2_journal_dirty(handle, root_bh);
2636 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2638 ocfs2_unlink_path(handle, et, dealloc, right_path,
2642 static int ocfs2_rotate_subtree_left(handle_t *handle,
2643 struct ocfs2_extent_tree *et,
2644 struct ocfs2_path *left_path,
2645 struct ocfs2_path *right_path,
2647 struct ocfs2_cached_dealloc_ctxt *dealloc,
2650 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2651 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2652 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2653 struct ocfs2_extent_block *eb;
2657 right_leaf_el = path_leaf_el(right_path);
2658 left_leaf_el = path_leaf_el(left_path);
2659 root_bh = left_path->p_node[subtree_index].bh;
2660 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2662 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2665 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2666 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2668 * It's legal for us to proceed if the right leaf is
2669 * the rightmost one and it has an empty extent. There
2670 * are two cases to handle - whether the leaf will be
2671 * empty after removal or not. If the leaf isn't empty
2672 * then just remove the empty extent up front. The
2673 * next block will handle empty leaves by flagging
2676 * Non rightmost leaves will throw -EAGAIN and the
2677 * caller can manually move the subtree and retry.
2680 if (eb->h_next_leaf_blk != 0ULL)
2683 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2684 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2685 path_leaf_bh(right_path),
2686 OCFS2_JOURNAL_ACCESS_WRITE);
2692 ocfs2_remove_empty_extent(right_leaf_el);
2694 right_has_empty = 1;
2697 if (eb->h_next_leaf_blk == 0ULL &&
2698 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2700 * We have to update i_last_eb_blk during the meta
2703 ret = ocfs2_et_root_journal_access(handle, et,
2704 OCFS2_JOURNAL_ACCESS_WRITE);
2710 del_right_subtree = 1;
2714 * Getting here with an empty extent in the right path implies
2715 * that it's the rightmost path and will be deleted.
2717 BUG_ON(right_has_empty && !del_right_subtree);
2719 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2726 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2727 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2734 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2742 if (!right_has_empty) {
2744 * Only do this if we're moving a real
2745 * record. Otherwise, the action is delayed until
2746 * after removal of the right path in which case we
2747 * can do a simple shift to remove the empty extent.
2749 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2750 memset(&right_leaf_el->l_recs[0], 0,
2751 sizeof(struct ocfs2_extent_rec));
2753 if (eb->h_next_leaf_blk == 0ULL) {
2755 * Move recs over to get rid of empty extent, decrease
2756 * next_free. This is allowed to remove the last
2757 * extent in our leaf (setting l_next_free_rec to
2758 * zero) - the delete code below won't care.
2760 ocfs2_remove_empty_extent(right_leaf_el);
2763 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2764 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2766 if (del_right_subtree) {
2767 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2768 subtree_index, dealloc);
2769 ret = ocfs2_update_edge_lengths(handle, et, left_path);
2775 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2776 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2779 * Removal of the extent in the left leaf was skipped
2780 * above so we could delete the right path
2783 if (right_has_empty)
2784 ocfs2_remove_empty_extent(left_leaf_el);
2786 ocfs2_journal_dirty(handle, et_root_bh);
2790 ocfs2_complete_edge_insert(handle, left_path, right_path,
2798 * Given a full path, determine what cpos value would return us a path
2799 * containing the leaf immediately to the right of the current one.
2801 * Will return zero if the path passed in is already the rightmost path.
2803 * This looks similar, but is subtly different to
2804 * ocfs2_find_cpos_for_left_leaf().
2806 int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2807 struct ocfs2_path *path, u32 *cpos)
2811 struct ocfs2_extent_list *el;
2815 if (path->p_tree_depth == 0)
2818 blkno = path_leaf_bh(path)->b_blocknr;
2820 /* Start at the tree node just above the leaf and work our way up. */
2821 i = path->p_tree_depth - 1;
2825 el = path->p_node[i].el;
2828 * Find the extent record just after the one in our
2831 next_free = le16_to_cpu(el->l_next_free_rec);
2832 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2833 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2834 if (j == (next_free - 1)) {
2837 * We've determined that the
2838 * path specified is already
2839 * the rightmost one - return a
2845 * The rightmost record points to our
2846 * leaf - we need to travel up the
2852 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2858 * If we got here, we never found a valid node where
2859 * the tree indicated one should be.
2861 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2862 (unsigned long long)blkno);
2867 blkno = path->p_node[i].bh->b_blocknr;
2875 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2876 struct ocfs2_extent_tree *et,
2877 struct ocfs2_path *path)
2880 struct buffer_head *bh = path_leaf_bh(path);
2881 struct ocfs2_extent_list *el = path_leaf_el(path);
2883 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2886 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2887 path_num_items(path) - 1);
2893 ocfs2_remove_empty_extent(el);
2894 ocfs2_journal_dirty(handle, bh);
2900 static int __ocfs2_rotate_tree_left(handle_t *handle,
2901 struct ocfs2_extent_tree *et,
2903 struct ocfs2_path *path,
2904 struct ocfs2_cached_dealloc_ctxt *dealloc,
2905 struct ocfs2_path **empty_extent_path)
2907 int ret, subtree_root, deleted;
2909 struct ocfs2_path *left_path = NULL;
2910 struct ocfs2_path *right_path = NULL;
2911 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2913 if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2916 *empty_extent_path = NULL;
2918 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2924 left_path = ocfs2_new_path_from_path(path);
2931 ocfs2_cp_path(left_path, path);
2933 right_path = ocfs2_new_path_from_path(path);
2940 while (right_cpos) {
2941 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2947 subtree_root = ocfs2_find_subtree_root(et, left_path,
2950 trace_ocfs2_rotate_subtree(subtree_root,
2951 (unsigned long long)
2952 right_path->p_node[subtree_root].bh->b_blocknr,
2953 right_path->p_tree_depth);
2955 ret = ocfs2_extend_rotate_transaction(handle, 0,
2956 orig_credits, left_path);
2963 * Caller might still want to make changes to the
2964 * tree root, so re-add it to the journal here.
2966 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2973 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2974 right_path, subtree_root,
2976 if (ret == -EAGAIN) {
2978 * The rotation has to temporarily stop due to
2979 * the right subtree having an empty
2980 * extent. Pass it back to the caller for a
2983 *empty_extent_path = right_path;
2993 * The subtree rotate might have removed records on
2994 * the rightmost edge. If so, then rotation is
3000 ocfs2_mv_path(left_path, right_path);
3002 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3011 ocfs2_free_path(right_path);
3012 ocfs2_free_path(left_path);
3017 static int ocfs2_remove_rightmost_path(handle_t *handle,
3018 struct ocfs2_extent_tree *et,
3019 struct ocfs2_path *path,
3020 struct ocfs2_cached_dealloc_ctxt *dealloc)
3022 int ret, subtree_index;
3024 struct ocfs2_path *left_path = NULL;
3025 struct ocfs2_extent_block *eb;
3026 struct ocfs2_extent_list *el;
3028 ret = ocfs2_et_sanity_check(et);
3032 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3038 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3047 * We have a path to the left of this one - it needs
3050 left_path = ocfs2_new_path_from_path(path);
3057 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3063 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3069 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3071 ocfs2_unlink_subtree(handle, et, left_path, path,
3072 subtree_index, dealloc);
3073 ret = ocfs2_update_edge_lengths(handle, et, left_path);
3079 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3080 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3083 * 'path' is also the leftmost path which
3084 * means it must be the only one. This gets
3085 * handled differently because we want to
3086 * revert the root back to having extents
3089 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3091 el = et->et_root_el;
3092 el->l_tree_depth = 0;
3093 el->l_next_free_rec = 0;
3094 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3096 ocfs2_et_set_last_eb_blk(et, 0);
3099 ocfs2_journal_dirty(handle, path_root_bh(path));
3102 ocfs2_free_path(left_path);
3106 static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3107 struct ocfs2_extent_tree *et,
3108 struct ocfs2_path *path,
3109 struct ocfs2_cached_dealloc_ctxt *dealloc)
3113 int credits = path->p_tree_depth * 2 + 1;
3115 handle = ocfs2_start_trans(osb, credits);
3116 if (IS_ERR(handle)) {
3117 ret = PTR_ERR(handle);
3122 ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3126 ocfs2_commit_trans(osb, handle);
3131 * Left rotation of btree records.
3133 * In many ways, this is (unsurprisingly) the opposite of right
3134 * rotation. We start at some non-rightmost path containing an empty
3135 * extent in the leaf block. The code works its way to the rightmost
3136 * path by rotating records to the left in every subtree.
3138 * This is used by any code which reduces the number of extent records
3139 * in a leaf. After removal, an empty record should be placed in the
3140 * leftmost list position.
3142 * This won't handle a length update of the rightmost path records if
3143 * the rightmost tree leaf record is removed so the caller is
3144 * responsible for detecting and correcting that.
3146 static int ocfs2_rotate_tree_left(handle_t *handle,
3147 struct ocfs2_extent_tree *et,
3148 struct ocfs2_path *path,
3149 struct ocfs2_cached_dealloc_ctxt *dealloc)
3151 int ret, orig_credits = handle->h_buffer_credits;
3152 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3153 struct ocfs2_extent_block *eb;
3154 struct ocfs2_extent_list *el;
3156 el = path_leaf_el(path);
3157 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3160 if (path->p_tree_depth == 0) {
3161 rightmost_no_delete:
3163 * Inline extents. This is trivially handled, so do
3166 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3173 * Handle rightmost branch now. There's several cases:
3174 * 1) simple rotation leaving records in there. That's trivial.
3175 * 2) rotation requiring a branch delete - there's no more
3176 * records left. Two cases of this:
3177 * a) There are branches to the left.
3178 * b) This is also the leftmost (the only) branch.
3180 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3181 * 2a) we need the left branch so that we can update it with the unlink
3182 * 2b) we need to bring the root back to inline extents.
3185 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3187 if (eb->h_next_leaf_blk == 0) {
3189 * This gets a bit tricky if we're going to delete the
3190 * rightmost path. Get the other cases out of the way
3193 if (le16_to_cpu(el->l_next_free_rec) > 1)
3194 goto rightmost_no_delete;
3196 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3197 ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3198 "Owner %llu has empty extent block at %llu\n",
3199 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3200 (unsigned long long)le64_to_cpu(eb->h_blkno));
3205 * XXX: The caller can not trust "path" any more after
3206 * this as it will have been deleted. What do we do?
3208 * In theory the rotate-for-merge code will never get
3209 * here because it'll always ask for a rotate in a
3213 ret = ocfs2_remove_rightmost_path(handle, et, path,
3221 * Now we can loop, remembering the path we get from -EAGAIN
3222 * and restarting from there.
3225 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3226 dealloc, &restart_path);
3227 if (ret && ret != -EAGAIN) {
3232 while (ret == -EAGAIN) {
3233 tmp_path = restart_path;
3234 restart_path = NULL;
3236 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3239 if (ret && ret != -EAGAIN) {
3244 ocfs2_free_path(tmp_path);
3252 ocfs2_free_path(tmp_path);
3253 ocfs2_free_path(restart_path);
3257 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3260 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3263 if (rec->e_leaf_clusters == 0) {
3265 * We consumed all of the merged-from record. An empty
3266 * extent cannot exist anywhere but the 1st array
3267 * position, so move things over if the merged-from
3268 * record doesn't occupy that position.
3270 * This creates a new empty extent so the caller
3271 * should be smart enough to have removed any existing
3275 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3276 size = index * sizeof(struct ocfs2_extent_rec);
3277 memmove(&el->l_recs[1], &el->l_recs[0], size);
3281 * Always memset - the caller doesn't check whether it
3282 * created an empty extent, so there could be junk in
3285 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3289 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3290 struct ocfs2_path *left_path,
3291 struct ocfs2_path **ret_right_path)
3295 struct ocfs2_path *right_path = NULL;
3296 struct ocfs2_extent_list *left_el;
3298 *ret_right_path = NULL;
3300 /* This function shouldn't be called for non-trees. */
3301 BUG_ON(left_path->p_tree_depth == 0);
3303 left_el = path_leaf_el(left_path);
3304 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3306 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3307 left_path, &right_cpos);
3313 /* This function shouldn't be called for the rightmost leaf. */
3314 BUG_ON(right_cpos == 0);
3316 right_path = ocfs2_new_path_from_path(left_path);
3323 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3329 *ret_right_path = right_path;
3332 ocfs2_free_path(right_path);
3337 * Remove split_rec clusters from the record at index and merge them
3338 * onto the beginning of the record "next" to it.
3339 * For index < l_count - 1, the next means the extent rec at index + 1.
3340 * For index == l_count - 1, the "next" means the 1st extent rec of the
3341 * next extent block.
3343 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3345 struct ocfs2_extent_tree *et,
3346 struct ocfs2_extent_rec *split_rec,
3349 int ret, next_free, i;
3350 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3351 struct ocfs2_extent_rec *left_rec;
3352 struct ocfs2_extent_rec *right_rec;
3353 struct ocfs2_extent_list *right_el;
3354 struct ocfs2_path *right_path = NULL;
3355 int subtree_index = 0;
3356 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3357 struct buffer_head *bh = path_leaf_bh(left_path);
3358 struct buffer_head *root_bh = NULL;
3360 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3361 left_rec = &el->l_recs[index];
3363 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3364 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3365 /* we meet with a cross extent block merge. */
3366 ret = ocfs2_get_right_path(et, left_path, &right_path);
3372 right_el = path_leaf_el(right_path);
3373 next_free = le16_to_cpu(right_el->l_next_free_rec);
3374 BUG_ON(next_free <= 0);
3375 right_rec = &right_el->l_recs[0];
3376 if (ocfs2_is_empty_extent(right_rec)) {
3377 BUG_ON(next_free <= 1);
3378 right_rec = &right_el->l_recs[1];
3381 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3382 le16_to_cpu(left_rec->e_leaf_clusters) !=
3383 le32_to_cpu(right_rec->e_cpos));
3385 subtree_index = ocfs2_find_subtree_root(et, left_path,
3388 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3389 handle->h_buffer_credits,
3396 root_bh = left_path->p_node[subtree_index].bh;
3397 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3399 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3406 for (i = subtree_index + 1;
3407 i < path_num_items(right_path); i++) {
3408 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3415 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3424 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3425 right_rec = &el->l_recs[index + 1];
3428 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3429 path_num_items(left_path) - 1);
3435 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3437 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3438 le64_add_cpu(&right_rec->e_blkno,
3439 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3441 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3443 ocfs2_cleanup_merge(el, index);
3445 ocfs2_journal_dirty(handle, bh);
3447 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3448 ocfs2_complete_edge_insert(handle, left_path, right_path,
3452 ocfs2_free_path(right_path);
3456 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3457 struct ocfs2_path *right_path,
3458 struct ocfs2_path **ret_left_path)
3462 struct ocfs2_path *left_path = NULL;
3464 *ret_left_path = NULL;
3466 /* This function shouldn't be called for non-trees. */
3467 BUG_ON(right_path->p_tree_depth == 0);
3469 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3470 right_path, &left_cpos);
3476 /* This function shouldn't be called for the leftmost leaf. */
3477 BUG_ON(left_cpos == 0);
3479 left_path = ocfs2_new_path_from_path(right_path);
3486 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3492 *ret_left_path = left_path;
3495 ocfs2_free_path(left_path);
3500 * Remove split_rec clusters from the record at index and merge them
3501 * onto the tail of the record "before" it.
3502 * For index > 0, the "before" means the extent rec at index - 1.
3504 * For index == 0, the "before" means the last record of the previous
3505 * extent block. And there is also a situation that we may need to
3506 * remove the rightmost leaf extent block in the right_path and change
3507 * the right path to indicate the new rightmost path.
3509 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3511 struct ocfs2_extent_tree *et,
3512 struct ocfs2_extent_rec *split_rec,
3513 struct ocfs2_cached_dealloc_ctxt *dealloc,
3516 int ret, i, subtree_index = 0, has_empty_extent = 0;
3517 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3518 struct ocfs2_extent_rec *left_rec;
3519 struct ocfs2_extent_rec *right_rec;
3520 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3521 struct buffer_head *bh = path_leaf_bh(right_path);
3522 struct buffer_head *root_bh = NULL;
3523 struct ocfs2_path *left_path = NULL;
3524 struct ocfs2_extent_list *left_el;
3528 right_rec = &el->l_recs[index];
3530 /* we meet with a cross extent block merge. */
3531 ret = ocfs2_get_left_path(et, right_path, &left_path);
3537 left_el = path_leaf_el(left_path);
3538 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3539 le16_to_cpu(left_el->l_count));
3541 left_rec = &left_el->l_recs[
3542 le16_to_cpu(left_el->l_next_free_rec) - 1];
3543 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3544 le16_to_cpu(left_rec->e_leaf_clusters) !=
3545 le32_to_cpu(split_rec->e_cpos));
3547 subtree_index = ocfs2_find_subtree_root(et, left_path,
3550 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3551 handle->h_buffer_credits,
3558 root_bh = left_path->p_node[subtree_index].bh;
3559 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3561 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3568 for (i = subtree_index + 1;
3569 i < path_num_items(right_path); i++) {
3570 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3577 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3585 left_rec = &el->l_recs[index - 1];
3586 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3587 has_empty_extent = 1;
3590 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3591 path_num_items(right_path) - 1);
3597 if (has_empty_extent && index == 1) {
3599 * The easy case - we can just plop the record right in.
3601 *left_rec = *split_rec;
3603 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3605 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3606 le64_add_cpu(&right_rec->e_blkno,
3607 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3609 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3611 ocfs2_cleanup_merge(el, index);
3613 ocfs2_journal_dirty(handle, bh);
3615 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3618 * In the situation that the right_rec is empty and the extent
3619 * block is empty also, ocfs2_complete_edge_insert can't handle
3620 * it and we need to delete the right extent block.
3622 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3623 le16_to_cpu(el->l_next_free_rec) == 1) {
3624 /* extend credit for ocfs2_remove_rightmost_path */
3625 ret = ocfs2_extend_rotate_transaction(handle, 0,
3626 handle->h_buffer_credits,
3633 ret = ocfs2_remove_rightmost_path(handle, et,
3641 /* Now the rightmost extent block has been deleted.
3642 * So we use the new rightmost path.
3644 ocfs2_mv_path(right_path, left_path);
3647 ocfs2_complete_edge_insert(handle, left_path,
3648 right_path, subtree_index);
3651 ocfs2_free_path(left_path);
3655 static int ocfs2_try_to_merge_extent(handle_t *handle,
3656 struct ocfs2_extent_tree *et,
3657 struct ocfs2_path *path,
3659 struct ocfs2_extent_rec *split_rec,
3660 struct ocfs2_cached_dealloc_ctxt *dealloc,
3661 struct ocfs2_merge_ctxt *ctxt)
3664 struct ocfs2_extent_list *el = path_leaf_el(path);
3665 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3667 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3669 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3670 /* extend credit for ocfs2_remove_rightmost_path */
3671 ret = ocfs2_extend_rotate_transaction(handle, 0,
3672 handle->h_buffer_credits,
3679 * The merge code will need to create an empty
3680 * extent to take the place of the newly
3681 * emptied slot. Remove any pre-existing empty
3682 * extents - having more than one in a leaf is
3685 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3691 rec = &el->l_recs[split_index];
3694 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3696 * Left-right contig implies this.
3698 BUG_ON(!ctxt->c_split_covers_rec);
3701 * Since the leftright insert always covers the entire
3702 * extent, this call will delete the insert record
3703 * entirely, resulting in an empty extent record added to
3706 * Since the adding of an empty extent shifts
3707 * everything back to the right, there's no need to
3708 * update split_index here.
3710 * When the split_index is zero, we need to merge it to the
3711 * prevoius extent block. It is more efficient and easier
3712 * if we do merge_right first and merge_left later.
3714 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3722 * We can only get this from logic error above.
3724 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3726 /* extend credit for ocfs2_remove_rightmost_path */
3727 ret = ocfs2_extend_rotate_transaction(handle, 0,
3728 handle->h_buffer_credits,
3735 /* The merge left us with an empty extent, remove it. */
3736 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3742 rec = &el->l_recs[split_index];
3745 * Note that we don't pass split_rec here on purpose -
3746 * we've merged it into the rec already.
3748 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3749 dealloc, split_index);
3756 /* extend credit for ocfs2_remove_rightmost_path */
3757 ret = ocfs2_extend_rotate_transaction(handle, 0,
3758 handle->h_buffer_credits,
3765 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3767 * Error from this last rotate is not critical, so
3768 * print but don't bubble it up.
3775 * Merge a record to the left or right.
3777 * 'contig_type' is relative to the existing record,
3778 * so for example, if we're "right contig", it's to
3779 * the record on the left (hence the left merge).
3781 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3782 ret = ocfs2_merge_rec_left(path, handle, et,
3790 ret = ocfs2_merge_rec_right(path, handle,
3799 if (ctxt->c_split_covers_rec) {
3800 /* extend credit for ocfs2_remove_rightmost_path */
3801 ret = ocfs2_extend_rotate_transaction(handle, 0,
3802 handle->h_buffer_credits,
3811 * The merge may have left an empty extent in
3812 * our leaf. Try to rotate it away.
3814 ret = ocfs2_rotate_tree_left(handle, et, path,
3826 static void ocfs2_subtract_from_rec(struct super_block *sb,
3827 enum ocfs2_split_type split,
3828 struct ocfs2_extent_rec *rec,
3829 struct ocfs2_extent_rec *split_rec)
3833 len_blocks = ocfs2_clusters_to_blocks(sb,
3834 le16_to_cpu(split_rec->e_leaf_clusters));
3836 if (split == SPLIT_LEFT) {
3838 * Region is on the left edge of the existing
3841 le32_add_cpu(&rec->e_cpos,
3842 le16_to_cpu(split_rec->e_leaf_clusters));
3843 le64_add_cpu(&rec->e_blkno, len_blocks);
3844 le16_add_cpu(&rec->e_leaf_clusters,
3845 -le16_to_cpu(split_rec->e_leaf_clusters));
3848 * Region is on the right edge of the existing
3851 le16_add_cpu(&rec->e_leaf_clusters,
3852 -le16_to_cpu(split_rec->e_leaf_clusters));
3857 * Do the final bits of extent record insertion at the target leaf
3858 * list. If this leaf is part of an allocation tree, it is assumed
3859 * that the tree above has been prepared.
3861 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3862 struct ocfs2_extent_rec *insert_rec,
3863 struct ocfs2_extent_list *el,
3864 struct ocfs2_insert_type *insert)
3866 int i = insert->ins_contig_index;
3868 struct ocfs2_extent_rec *rec;
3870 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3872 if (insert->ins_split != SPLIT_NONE) {
3873 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3875 rec = &el->l_recs[i];
3876 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3877 insert->ins_split, rec,
3883 * Contiguous insert - either left or right.
3885 if (insert->ins_contig != CONTIG_NONE) {
3886 rec = &el->l_recs[i];
3887 if (insert->ins_contig == CONTIG_LEFT) {
3888 rec->e_blkno = insert_rec->e_blkno;
3889 rec->e_cpos = insert_rec->e_cpos;
3891 le16_add_cpu(&rec->e_leaf_clusters,
3892 le16_to_cpu(insert_rec->e_leaf_clusters));
3897 * Handle insert into an empty leaf.
3899 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3900 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3901 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3902 el->l_recs[0] = *insert_rec;
3903 el->l_next_free_rec = cpu_to_le16(1);
3910 if (insert->ins_appending == APPEND_TAIL) {
3911 i = le16_to_cpu(el->l_next_free_rec) - 1;
3912 rec = &el->l_recs[i];
3913 range = le32_to_cpu(rec->e_cpos)
3914 + le16_to_cpu(rec->e_leaf_clusters);
3915 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3917 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3918 le16_to_cpu(el->l_count),
3919 "owner %llu, depth %u, count %u, next free %u, "
3920 "rec.cpos %u, rec.clusters %u, "
3921 "insert.cpos %u, insert.clusters %u\n",
3922 ocfs2_metadata_cache_owner(et->et_ci),
3923 le16_to_cpu(el->l_tree_depth),
3924 le16_to_cpu(el->l_count),
3925 le16_to_cpu(el->l_next_free_rec),
3926 le32_to_cpu(el->l_recs[i].e_cpos),
3927 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3928 le32_to_cpu(insert_rec->e_cpos),
3929 le16_to_cpu(insert_rec->e_leaf_clusters));
3931 el->l_recs[i] = *insert_rec;
3932 le16_add_cpu(&el->l_next_free_rec, 1);
3938 * Ok, we have to rotate.
3940 * At this point, it is safe to assume that inserting into an
3941 * empty leaf and appending to a leaf have both been handled
3944 * This leaf needs to have space, either by the empty 1st
3945 * extent record, or by virtue of an l_next_rec < l_count.
3947 ocfs2_rotate_leaf(el, insert_rec);
3950 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3951 struct ocfs2_extent_tree *et,
3952 struct ocfs2_path *path,
3953 struct ocfs2_extent_rec *insert_rec)
3956 struct buffer_head *bh;
3957 struct ocfs2_extent_list *el;
3958 struct ocfs2_extent_rec *rec;
3961 * Update everything except the leaf block.
3963 for (i = 0; i < path->p_tree_depth; i++) {
3964 bh = path->p_node[i].bh;
3965 el = path->p_node[i].el;
3967 next_free = le16_to_cpu(el->l_next_free_rec);
3968 if (next_free == 0) {
3969 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3970 "Owner %llu has a bad extent list\n",
3971 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3975 rec = &el->l_recs[next_free - 1];
3977 rec->e_int_clusters = insert_rec->e_cpos;
3978 le32_add_cpu(&rec->e_int_clusters,
3979 le16_to_cpu(insert_rec->e_leaf_clusters));
3980 le32_add_cpu(&rec->e_int_clusters,
3981 -le32_to_cpu(rec->e_cpos));
3983 ocfs2_journal_dirty(handle, bh);
3987 static int ocfs2_append_rec_to_path(handle_t *handle,
3988 struct ocfs2_extent_tree *et,
3989 struct ocfs2_extent_rec *insert_rec,
3990 struct ocfs2_path *right_path,
3991 struct ocfs2_path **ret_left_path)
3994 struct ocfs2_extent_list *el;
3995 struct ocfs2_path *left_path = NULL;
3997 *ret_left_path = NULL;
4000 * This shouldn't happen for non-trees. The extent rec cluster
4001 * count manipulation below only works for interior nodes.
4003 BUG_ON(right_path->p_tree_depth == 0);
4006 * If our appending insert is at the leftmost edge of a leaf,
4007 * then we might need to update the rightmost records of the
4010 el = path_leaf_el(right_path);
4011 next_free = le16_to_cpu(el->l_next_free_rec);
4012 if (next_free == 0 ||
4013 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4016 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4017 right_path, &left_cpos);
4023 trace_ocfs2_append_rec_to_path(
4024 (unsigned long long)
4025 ocfs2_metadata_cache_owner(et->et_ci),
4026 le32_to_cpu(insert_rec->e_cpos),
4030 * No need to worry if the append is already in the
4034 left_path = ocfs2_new_path_from_path(right_path);
4041 ret = ocfs2_find_path(et->et_ci, left_path,
4049 * ocfs2_insert_path() will pass the left_path to the
4055 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4061 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4063 *ret_left_path = left_path;
4067 ocfs2_free_path(left_path);
4072 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4073 struct ocfs2_path *left_path,
4074 struct ocfs2_path *right_path,
4075 struct ocfs2_extent_rec *split_rec,
4076 enum ocfs2_split_type split)
4079 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4080 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4081 struct ocfs2_extent_rec *rec, *tmprec;
4083 right_el = path_leaf_el(right_path);
4085 left_el = path_leaf_el(left_path);
4088 insert_el = right_el;
4089 index = ocfs2_search_extent_list(el, cpos);
4091 if (index == 0 && left_path) {
4092 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4095 * This typically means that the record
4096 * started in the left path but moved to the
4097 * right as a result of rotation. We either
4098 * move the existing record to the left, or we
4099 * do the later insert there.
4101 * In this case, the left path should always
4102 * exist as the rotate code will have passed
4103 * it back for a post-insert update.
4106 if (split == SPLIT_LEFT) {
4108 * It's a left split. Since we know
4109 * that the rotate code gave us an
4110 * empty extent in the left path, we
4111 * can just do the insert there.
4113 insert_el = left_el;
4116 * Right split - we have to move the
4117 * existing record over to the left
4118 * leaf. The insert will be into the
4119 * newly created empty extent in the
4122 tmprec = &right_el->l_recs[index];
4123 ocfs2_rotate_leaf(left_el, tmprec);
4126 memset(tmprec, 0, sizeof(*tmprec));
4127 index = ocfs2_search_extent_list(left_el, cpos);
4128 BUG_ON(index == -1);
4133 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4135 * Left path is easy - we can just allow the insert to
4139 insert_el = left_el;
4140 index = ocfs2_search_extent_list(el, cpos);
4141 BUG_ON(index == -1);
4144 rec = &el->l_recs[index];
4145 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4146 split, rec, split_rec);
4147 ocfs2_rotate_leaf(insert_el, split_rec);
4151 * This function only does inserts on an allocation b-tree. For tree
4152 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4154 * right_path is the path we want to do the actual insert
4155 * in. left_path should only be passed in if we need to update that
4156 * portion of the tree after an edge insert.
4158 static int ocfs2_insert_path(handle_t *handle,
4159 struct ocfs2_extent_tree *et,
4160 struct ocfs2_path *left_path,
4161 struct ocfs2_path *right_path,
4162 struct ocfs2_extent_rec *insert_rec,
4163 struct ocfs2_insert_type *insert)
4165 int ret, subtree_index;
4166 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4170 * There's a chance that left_path got passed back to
4171 * us without being accounted for in the
4172 * journal. Extend our transaction here to be sure we
4173 * can change those blocks.
4175 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4181 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4189 * Pass both paths to the journal. The majority of inserts
4190 * will be touching all components anyway.
4192 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4198 if (insert->ins_split != SPLIT_NONE) {
4200 * We could call ocfs2_insert_at_leaf() for some types
4201 * of splits, but it's easier to just let one separate
4202 * function sort it all out.
4204 ocfs2_split_record(et, left_path, right_path,
4205 insert_rec, insert->ins_split);
4208 * Split might have modified either leaf and we don't
4209 * have a guarantee that the later edge insert will
4210 * dirty this for us.
4213 ocfs2_journal_dirty(handle,
4214 path_leaf_bh(left_path));
4216 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4219 ocfs2_journal_dirty(handle, leaf_bh);
4223 * The rotate code has indicated that we need to fix
4224 * up portions of the tree after the insert.
4226 * XXX: Should we extend the transaction here?
4228 subtree_index = ocfs2_find_subtree_root(et, left_path,
4230 ocfs2_complete_edge_insert(handle, left_path, right_path,
4239 static int ocfs2_do_insert_extent(handle_t *handle,
4240 struct ocfs2_extent_tree *et,
4241 struct ocfs2_extent_rec *insert_rec,
4242 struct ocfs2_insert_type *type)
4244 int ret, rotate = 0;
4246 struct ocfs2_path *right_path = NULL;
4247 struct ocfs2_path *left_path = NULL;
4248 struct ocfs2_extent_list *el;
4250 el = et->et_root_el;
4252 ret = ocfs2_et_root_journal_access(handle, et,
4253 OCFS2_JOURNAL_ACCESS_WRITE);
4259 if (le16_to_cpu(el->l_tree_depth) == 0) {
4260 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4261 goto out_update_clusters;
4264 right_path = ocfs2_new_path_from_et(et);
4272 * Determine the path to start with. Rotations need the
4273 * rightmost path, everything else can go directly to the
4276 cpos = le32_to_cpu(insert_rec->e_cpos);
4277 if (type->ins_appending == APPEND_NONE &&
4278 type->ins_contig == CONTIG_NONE) {
4283 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4290 * Rotations and appends need special treatment - they modify
4291 * parts of the tree's above them.
4293 * Both might pass back a path immediate to the left of the
4294 * one being inserted to. This will be cause
4295 * ocfs2_insert_path() to modify the rightmost records of
4296 * left_path to account for an edge insert.
4298 * XXX: When modifying this code, keep in mind that an insert
4299 * can wind up skipping both of these two special cases...
4302 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4303 le32_to_cpu(insert_rec->e_cpos),
4304 right_path, &left_path);
4311 * ocfs2_rotate_tree_right() might have extended the
4312 * transaction without re-journaling our tree root.
4314 ret = ocfs2_et_root_journal_access(handle, et,
4315 OCFS2_JOURNAL_ACCESS_WRITE);
4320 } else if (type->ins_appending == APPEND_TAIL
4321 && type->ins_contig != CONTIG_LEFT) {
4322 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4323 right_path, &left_path);
4330 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4337 out_update_clusters:
4338 if (type->ins_split == SPLIT_NONE)
4339 ocfs2_et_update_clusters(et,
4340 le16_to_cpu(insert_rec->e_leaf_clusters));
4342 ocfs2_journal_dirty(handle, et->et_root_bh);
4345 ocfs2_free_path(left_path);
4346 ocfs2_free_path(right_path);
4351 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4352 struct ocfs2_path *path,
4353 struct ocfs2_extent_list *el, int index,
4354 struct ocfs2_extent_rec *split_rec,
4355 struct ocfs2_merge_ctxt *ctxt)
4358 enum ocfs2_contig_type ret = CONTIG_NONE;
4359 u32 left_cpos, right_cpos;
4360 struct ocfs2_extent_rec *rec = NULL;
4361 struct ocfs2_extent_list *new_el;
4362 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4363 struct buffer_head *bh;
4364 struct ocfs2_extent_block *eb;
4365 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4368 rec = &el->l_recs[index - 1];
4369 } else if (path->p_tree_depth > 0) {
4370 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4374 if (left_cpos != 0) {
4375 left_path = ocfs2_new_path_from_path(path);
4382 status = ocfs2_find_path(et->et_ci, left_path,
4385 goto free_left_path;
4387 new_el = path_leaf_el(left_path);
4389 if (le16_to_cpu(new_el->l_next_free_rec) !=
4390 le16_to_cpu(new_el->l_count)) {
4391 bh = path_leaf_bh(left_path);
4392 eb = (struct ocfs2_extent_block *)bh->b_data;
4393 status = ocfs2_error(sb,
4394 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4395 (unsigned long long)le64_to_cpu(eb->h_blkno),
4396 le16_to_cpu(new_el->l_next_free_rec),
4397 le16_to_cpu(new_el->l_count));
4398 goto free_left_path;
4400 rec = &new_el->l_recs[
4401 le16_to_cpu(new_el->l_next_free_rec) - 1];
4406 * We're careful to check for an empty extent record here -
4407 * the merge code will know what to do if it sees one.
4410 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4411 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4414 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4419 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4420 rec = &el->l_recs[index + 1];
4421 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4422 path->p_tree_depth > 0) {
4423 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4425 goto free_left_path;
4427 if (right_cpos == 0)
4428 goto free_left_path;
4430 right_path = ocfs2_new_path_from_path(path);
4434 goto free_left_path;
4437 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4439 goto free_right_path;
4441 new_el = path_leaf_el(right_path);
4442 rec = &new_el->l_recs[0];
4443 if (ocfs2_is_empty_extent(rec)) {
4444 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4445 bh = path_leaf_bh(right_path);
4446 eb = (struct ocfs2_extent_block *)bh->b_data;
4447 status = ocfs2_error(sb,
4448 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4449 (unsigned long long)le64_to_cpu(eb->h_blkno),
4450 le16_to_cpu(new_el->l_next_free_rec));
4451 goto free_right_path;
4453 rec = &new_el->l_recs[1];
4458 enum ocfs2_contig_type contig_type;
4460 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4462 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4463 ret = CONTIG_LEFTRIGHT;
4464 else if (ret == CONTIG_NONE)
4469 ocfs2_free_path(right_path);
4471 ocfs2_free_path(left_path);
4474 ctxt->c_contig_type = ret;
4479 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4480 struct ocfs2_insert_type *insert,
4481 struct ocfs2_extent_list *el,
4482 struct ocfs2_extent_rec *insert_rec)
4485 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4487 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4489 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4490 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4492 if (contig_type != CONTIG_NONE) {
4493 insert->ins_contig_index = i;
4497 insert->ins_contig = contig_type;
4499 if (insert->ins_contig != CONTIG_NONE) {
4500 struct ocfs2_extent_rec *rec =
4501 &el->l_recs[insert->ins_contig_index];
4502 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4503 le16_to_cpu(insert_rec->e_leaf_clusters);
4506 * Caller might want us to limit the size of extents, don't
4507 * calculate contiguousness if we might exceed that limit.
4509 if (et->et_max_leaf_clusters &&
4510 (len > et->et_max_leaf_clusters))
4511 insert->ins_contig = CONTIG_NONE;
4516 * This should only be called against the righmost leaf extent list.
4518 * ocfs2_figure_appending_type() will figure out whether we'll have to
4519 * insert at the tail of the rightmost leaf.
4521 * This should also work against the root extent list for tree's with 0
4522 * depth. If we consider the root extent list to be the rightmost leaf node
4523 * then the logic here makes sense.
4525 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4526 struct ocfs2_extent_list *el,
4527 struct ocfs2_extent_rec *insert_rec)
4530 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4531 struct ocfs2_extent_rec *rec;
4533 insert->ins_appending = APPEND_NONE;
4535 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4537 if (!el->l_next_free_rec)
4538 goto set_tail_append;
4540 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4541 /* Were all records empty? */
4542 if (le16_to_cpu(el->l_next_free_rec) == 1)
4543 goto set_tail_append;
4546 i = le16_to_cpu(el->l_next_free_rec) - 1;
4547 rec = &el->l_recs[i];
4550 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4551 goto set_tail_append;
4556 insert->ins_appending = APPEND_TAIL;
4560 * Helper function called at the beginning of an insert.
4562 * This computes a few things that are commonly used in the process of
4563 * inserting into the btree:
4564 * - Whether the new extent is contiguous with an existing one.
4565 * - The current tree depth.
4566 * - Whether the insert is an appending one.
4567 * - The total # of free records in the tree.
4569 * All of the information is stored on the ocfs2_insert_type
4572 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4573 struct buffer_head **last_eb_bh,
4574 struct ocfs2_extent_rec *insert_rec,
4576 struct ocfs2_insert_type *insert)
4579 struct ocfs2_extent_block *eb;
4580 struct ocfs2_extent_list *el;
4581 struct ocfs2_path *path = NULL;
4582 struct buffer_head *bh = NULL;
4584 insert->ins_split = SPLIT_NONE;
4586 el = et->et_root_el;
4587 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4589 if (el->l_tree_depth) {
4591 * If we have tree depth, we read in the
4592 * rightmost extent block ahead of time as
4593 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4594 * may want it later.
4596 ret = ocfs2_read_extent_block(et->et_ci,
4597 ocfs2_et_get_last_eb_blk(et),
4603 eb = (struct ocfs2_extent_block *) bh->b_data;
4608 * Unless we have a contiguous insert, we'll need to know if
4609 * there is room left in our allocation tree for another
4612 * XXX: This test is simplistic, we can search for empty
4613 * extent records too.
4615 *free_records = le16_to_cpu(el->l_count) -
4616 le16_to_cpu(el->l_next_free_rec);
4618 if (!insert->ins_tree_depth) {
4619 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4620 ocfs2_figure_appending_type(insert, el, insert_rec);
4624 path = ocfs2_new_path_from_et(et);
4632 * In the case that we're inserting past what the tree
4633 * currently accounts for, ocfs2_find_path() will return for
4634 * us the rightmost tree path. This is accounted for below in
4635 * the appending code.
4637 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4643 el = path_leaf_el(path);
4646 * Now that we have the path, there's two things we want to determine:
4647 * 1) Contiguousness (also set contig_index if this is so)
4649 * 2) Are we doing an append? We can trivially break this up
4650 * into two types of appends: simple record append, or a
4651 * rotate inside the tail leaf.
4653 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4656 * The insert code isn't quite ready to deal with all cases of
4657 * left contiguousness. Specifically, if it's an insert into
4658 * the 1st record in a leaf, it will require the adjustment of
4659 * cluster count on the last record of the path directly to it's
4660 * left. For now, just catch that case and fool the layers
4661 * above us. This works just fine for tree_depth == 0, which
4662 * is why we allow that above.
4664 if (insert->ins_contig == CONTIG_LEFT &&
4665 insert->ins_contig_index == 0)
4666 insert->ins_contig = CONTIG_NONE;
4669 * Ok, so we can simply compare against last_eb to figure out
4670 * whether the path doesn't exist. This will only happen in
4671 * the case that we're doing a tail append, so maybe we can
4672 * take advantage of that information somehow.
4674 if (ocfs2_et_get_last_eb_blk(et) ==
4675 path_leaf_bh(path)->b_blocknr) {
4677 * Ok, ocfs2_find_path() returned us the rightmost
4678 * tree path. This might be an appending insert. There are
4680 * 1) We're doing a true append at the tail:
4681 * -This might even be off the end of the leaf
4682 * 2) We're "appending" by rotating in the tail
4684 ocfs2_figure_appending_type(insert, el, insert_rec);
4688 ocfs2_free_path(path);
4698 * Insert an extent into a btree.
4700 * The caller needs to update the owning btree's cluster count.
4702 int ocfs2_insert_extent(handle_t *handle,
4703 struct ocfs2_extent_tree *et,
4708 struct ocfs2_alloc_context *meta_ac)
4711 int uninitialized_var(free_records);
4712 struct buffer_head *last_eb_bh = NULL;
4713 struct ocfs2_insert_type insert = {0, };
4714 struct ocfs2_extent_rec rec;
4716 trace_ocfs2_insert_extent_start(
4717 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4718 cpos, new_clusters);
4720 memset(&rec, 0, sizeof(rec));
4721 rec.e_cpos = cpu_to_le32(cpos);
4722 rec.e_blkno = cpu_to_le64(start_blk);
4723 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4724 rec.e_flags = flags;
4725 status = ocfs2_et_insert_check(et, &rec);
4731 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4732 &free_records, &insert);
4738 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4739 insert.ins_contig_index, free_records,
4740 insert.ins_tree_depth);
4742 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4743 status = ocfs2_grow_tree(handle, et,
4744 &insert.ins_tree_depth, &last_eb_bh,
4752 /* Finally, we can add clusters. This might rotate the tree for us. */
4753 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4757 ocfs2_et_extent_map_insert(et, &rec);
4766 * Allcate and add clusters into the extent b-tree.
4767 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4768 * The extent b-tree's root is specified by et, and
4769 * it is not limited to the file storage. Any extent tree can use this
4770 * function if it implements the proper ocfs2_extent_tree.
4772 int ocfs2_add_clusters_in_btree(handle_t *handle,
4773 struct ocfs2_extent_tree *et,
4774 u32 *logical_offset,
4775 u32 clusters_to_add,
4777 struct ocfs2_alloc_context *data_ac,
4778 struct ocfs2_alloc_context *meta_ac,
4779 enum ocfs2_alloc_restarted *reason_ret)
4781 int status = 0, err = 0;
4784 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4785 u32 bit_off, num_bits;
4788 struct ocfs2_super *osb =
4789 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4791 BUG_ON(!clusters_to_add);
4794 flags = OCFS2_EXT_UNWRITTEN;
4796 free_extents = ocfs2_num_free_extents(et);
4797 if (free_extents < 0) {
4798 status = free_extents;
4803 /* there are two cases which could cause us to EAGAIN in the
4804 * we-need-more-metadata case:
4805 * 1) we haven't reserved *any*
4806 * 2) we are so fragmented, we've needed to add metadata too
4808 if (!free_extents && !meta_ac) {
4811 reason = RESTART_META;
4813 } else if ((!free_extents)
4814 && (ocfs2_alloc_context_bits_left(meta_ac)
4815 < ocfs2_extend_meta_needed(et->et_root_el))) {
4818 reason = RESTART_META;
4822 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4823 clusters_to_add, &bit_off, &num_bits);
4825 if (status != -ENOSPC)
4830 BUG_ON(num_bits > clusters_to_add);
4832 /* reserve our write early -- insert_extent may update the tree root */
4833 status = ocfs2_et_root_journal_access(handle, et,
4834 OCFS2_JOURNAL_ACCESS_WRITE);
4841 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4842 trace_ocfs2_add_clusters_in_btree(
4843 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4845 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4846 num_bits, flags, meta_ac);
4853 ocfs2_journal_dirty(handle, et->et_root_bh);
4855 clusters_to_add -= num_bits;
4856 *logical_offset += num_bits;
4858 if (clusters_to_add) {
4859 err = clusters_to_add;
4861 reason = RESTART_TRANS;
4866 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4867 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4870 ocfs2_free_clusters(handle,
4873 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4879 *reason_ret = reason;
4880 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4884 static void ocfs2_make_right_split_rec(struct super_block *sb,
4885 struct ocfs2_extent_rec *split_rec,
4887 struct ocfs2_extent_rec *rec)
4889 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4890 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4892 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4894 split_rec->e_cpos = cpu_to_le32(cpos);
4895 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4897 split_rec->e_blkno = rec->e_blkno;
4898 le64_add_cpu(&split_rec->e_blkno,
4899 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4901 split_rec->e_flags = rec->e_flags;
4904 static int ocfs2_split_and_insert(handle_t *handle,
4905 struct ocfs2_extent_tree *et,
4906 struct ocfs2_path *path,
4907 struct buffer_head **last_eb_bh,
4909 struct ocfs2_extent_rec *orig_split_rec,
4910 struct ocfs2_alloc_context *meta_ac)
4913 unsigned int insert_range, rec_range, do_leftright = 0;
4914 struct ocfs2_extent_rec tmprec;
4915 struct ocfs2_extent_list *rightmost_el;
4916 struct ocfs2_extent_rec rec;
4917 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4918 struct ocfs2_insert_type insert;
4919 struct ocfs2_extent_block *eb;
4923 * Store a copy of the record on the stack - it might move
4924 * around as the tree is manipulated below.
4926 rec = path_leaf_el(path)->l_recs[split_index];
4928 rightmost_el = et->et_root_el;
4930 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4932 BUG_ON(!(*last_eb_bh));
4933 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4934 rightmost_el = &eb->h_list;
4937 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4938 le16_to_cpu(rightmost_el->l_count)) {
4939 ret = ocfs2_grow_tree(handle, et,
4940 &depth, last_eb_bh, meta_ac);
4947 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4948 insert.ins_appending = APPEND_NONE;
4949 insert.ins_contig = CONTIG_NONE;
4950 insert.ins_tree_depth = depth;
4952 insert_range = le32_to_cpu(split_rec.e_cpos) +
4953 le16_to_cpu(split_rec.e_leaf_clusters);
4954 rec_range = le32_to_cpu(rec.e_cpos) +
4955 le16_to_cpu(rec.e_leaf_clusters);
4957 if (split_rec.e_cpos == rec.e_cpos) {
4958 insert.ins_split = SPLIT_LEFT;
4959 } else if (insert_range == rec_range) {
4960 insert.ins_split = SPLIT_RIGHT;
4963 * Left/right split. We fake this as a right split
4964 * first and then make a second pass as a left split.
4966 insert.ins_split = SPLIT_RIGHT;
4968 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4969 &tmprec, insert_range, &rec);
4973 BUG_ON(do_leftright);
4977 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4983 if (do_leftright == 1) {
4985 struct ocfs2_extent_list *el;
4988 split_rec = *orig_split_rec;
4990 ocfs2_reinit_path(path, 1);
4992 cpos = le32_to_cpu(split_rec.e_cpos);
4993 ret = ocfs2_find_path(et->et_ci, path, cpos);
4999 el = path_leaf_el(path);
5000 split_index = ocfs2_search_extent_list(el, cpos);
5001 if (split_index == -1) {
5002 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5003 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5004 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5016 static int ocfs2_replace_extent_rec(handle_t *handle,
5017 struct ocfs2_extent_tree *et,
5018 struct ocfs2_path *path,
5019 struct ocfs2_extent_list *el,
5021 struct ocfs2_extent_rec *split_rec)
5025 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5026 path_num_items(path) - 1);
5032 el->l_recs[split_index] = *split_rec;
5034 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5040 * Split part or all of the extent record at split_index in the leaf
5041 * pointed to by path. Merge with the contiguous extent record if needed.
5043 * Care is taken to handle contiguousness so as to not grow the tree.
5045 * meta_ac is not strictly necessary - we only truly need it if growth
5046 * of the tree is required. All other cases will degrade into a less
5047 * optimal tree layout.
5049 * last_eb_bh should be the rightmost leaf block for any extent
5050 * btree. Since a split may grow the tree or a merge might shrink it,
5051 * the caller cannot trust the contents of that buffer after this call.
5053 * This code is optimized for readability - several passes might be
5054 * made over certain portions of the tree. All of those blocks will
5055 * have been brought into cache (and pinned via the journal), so the
5056 * extra overhead is not expressed in terms of disk reads.
5058 int ocfs2_split_extent(handle_t *handle,
5059 struct ocfs2_extent_tree *et,
5060 struct ocfs2_path *path,
5062 struct ocfs2_extent_rec *split_rec,
5063 struct ocfs2_alloc_context *meta_ac,
5064 struct ocfs2_cached_dealloc_ctxt *dealloc)
5067 struct ocfs2_extent_list *el = path_leaf_el(path);
5068 struct buffer_head *last_eb_bh = NULL;
5069 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5070 struct ocfs2_merge_ctxt ctxt;
5072 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5073 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5074 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5080 ret = ocfs2_figure_merge_contig_type(et, path, el,
5090 * The core merge / split code wants to know how much room is
5091 * left in this allocation tree, so we pass the
5092 * rightmost extent list.
5094 if (path->p_tree_depth) {
5095 ret = ocfs2_read_extent_block(et->et_ci,
5096 ocfs2_et_get_last_eb_blk(et),
5104 if (rec->e_cpos == split_rec->e_cpos &&
5105 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5106 ctxt.c_split_covers_rec = 1;
5108 ctxt.c_split_covers_rec = 0;
5110 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5112 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5113 ctxt.c_has_empty_extent,
5114 ctxt.c_split_covers_rec);
5116 if (ctxt.c_contig_type == CONTIG_NONE) {
5117 if (ctxt.c_split_covers_rec)
5118 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5119 split_index, split_rec);
5121 ret = ocfs2_split_and_insert(handle, et, path,
5122 &last_eb_bh, split_index,
5123 split_rec, meta_ac);
5127 ret = ocfs2_try_to_merge_extent(handle, et, path,
5128 split_index, split_rec,
5140 * Change the flags of the already-existing extent at cpos for len clusters.
5142 * new_flags: the flags we want to set.
5143 * clear_flags: the flags we want to clear.
5144 * phys: the new physical offset we want this new extent starts from.
5146 * If the existing extent is larger than the request, initiate a
5147 * split. An attempt will be made at merging with adjacent extents.
5149 * The caller is responsible for passing down meta_ac if we'll need it.
5151 int ocfs2_change_extent_flag(handle_t *handle,
5152 struct ocfs2_extent_tree *et,
5153 u32 cpos, u32 len, u32 phys,
5154 struct ocfs2_alloc_context *meta_ac,
5155 struct ocfs2_cached_dealloc_ctxt *dealloc,
5156 int new_flags, int clear_flags)
5159 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5160 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5161 struct ocfs2_extent_rec split_rec;
5162 struct ocfs2_path *left_path = NULL;
5163 struct ocfs2_extent_list *el;
5164 struct ocfs2_extent_rec *rec;
5166 left_path = ocfs2_new_path_from_et(et);
5173 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5178 el = path_leaf_el(left_path);
5180 index = ocfs2_search_extent_list(el, cpos);
5183 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5184 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5191 rec = &el->l_recs[index];
5192 if (new_flags && (rec->e_flags & new_flags)) {
5193 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5194 "extent that already had them\n",
5195 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5200 if (clear_flags && !(rec->e_flags & clear_flags)) {
5201 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5202 "extent that didn't have them\n",
5203 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5208 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5209 split_rec.e_cpos = cpu_to_le32(cpos);
5210 split_rec.e_leaf_clusters = cpu_to_le16(len);
5211 split_rec.e_blkno = cpu_to_le64(start_blkno);
5212 split_rec.e_flags = rec->e_flags;
5214 split_rec.e_flags |= new_flags;
5216 split_rec.e_flags &= ~clear_flags;
5218 ret = ocfs2_split_extent(handle, et, left_path,
5219 index, &split_rec, meta_ac,
5225 ocfs2_free_path(left_path);
5231 * Mark the already-existing extent at cpos as written for len clusters.
5232 * This removes the unwritten extent flag.
5234 * If the existing extent is larger than the request, initiate a
5235 * split. An attempt will be made at merging with adjacent extents.
5237 * The caller is responsible for passing down meta_ac if we'll need it.
5239 int ocfs2_mark_extent_written(struct inode *inode,
5240 struct ocfs2_extent_tree *et,
5241 handle_t *handle, u32 cpos, u32 len, u32 phys,
5242 struct ocfs2_alloc_context *meta_ac,
5243 struct ocfs2_cached_dealloc_ctxt *dealloc)
5247 trace_ocfs2_mark_extent_written(
5248 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5251 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5252 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",
5253 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5259 * XXX: This should be fixed up so that we just re-insert the
5260 * next extent records.
5262 ocfs2_et_extent_map_truncate(et, 0);
5264 ret = ocfs2_change_extent_flag(handle, et, cpos,
5265 len, phys, meta_ac, dealloc,
5266 0, OCFS2_EXT_UNWRITTEN);
5274 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5275 struct ocfs2_path *path,
5276 int index, u32 new_range,
5277 struct ocfs2_alloc_context *meta_ac)
5279 int ret, depth, credits;
5280 struct buffer_head *last_eb_bh = NULL;
5281 struct ocfs2_extent_block *eb;
5282 struct ocfs2_extent_list *rightmost_el, *el;
5283 struct ocfs2_extent_rec split_rec;
5284 struct ocfs2_extent_rec *rec;
5285 struct ocfs2_insert_type insert;
5288 * Setup the record to split before we grow the tree.
5290 el = path_leaf_el(path);
5291 rec = &el->l_recs[index];
5292 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5293 &split_rec, new_range, rec);
5295 depth = path->p_tree_depth;
5297 ret = ocfs2_read_extent_block(et->et_ci,
5298 ocfs2_et_get_last_eb_blk(et),
5305 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5306 rightmost_el = &eb->h_list;
5308 rightmost_el = path_leaf_el(path);
5310 credits = path->p_tree_depth +
5311 ocfs2_extend_meta_needed(et->et_root_el);
5312 ret = ocfs2_extend_trans(handle, credits);
5318 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5319 le16_to_cpu(rightmost_el->l_count)) {
5320 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5328 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5329 insert.ins_appending = APPEND_NONE;
5330 insert.ins_contig = CONTIG_NONE;
5331 insert.ins_split = SPLIT_RIGHT;
5332 insert.ins_tree_depth = depth;
5334 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5343 static int ocfs2_truncate_rec(handle_t *handle,
5344 struct ocfs2_extent_tree *et,
5345 struct ocfs2_path *path, int index,
5346 struct ocfs2_cached_dealloc_ctxt *dealloc,
5350 u32 left_cpos, rec_range, trunc_range;
5351 int is_rightmost_tree_rec = 0;
5352 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5353 struct ocfs2_path *left_path = NULL;
5354 struct ocfs2_extent_list *el = path_leaf_el(path);
5355 struct ocfs2_extent_rec *rec;
5356 struct ocfs2_extent_block *eb;
5358 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5359 /* extend credit for ocfs2_remove_rightmost_path */
5360 ret = ocfs2_extend_rotate_transaction(handle, 0,
5361 handle->h_buffer_credits,
5368 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5377 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5378 path->p_tree_depth) {
5380 * Check whether this is the rightmost tree record. If
5381 * we remove all of this record or part of its right
5382 * edge then an update of the record lengths above it
5385 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5386 if (eb->h_next_leaf_blk == 0)
5387 is_rightmost_tree_rec = 1;
5390 rec = &el->l_recs[index];
5391 if (index == 0 && path->p_tree_depth &&
5392 le32_to_cpu(rec->e_cpos) == cpos) {
5394 * Changing the leftmost offset (via partial or whole
5395 * record truncate) of an interior (or rightmost) path
5396 * means we have to update the subtree that is formed
5397 * by this leaf and the one to it's left.
5399 * There are two cases we can skip:
5400 * 1) Path is the leftmost one in our btree.
5401 * 2) The leaf is rightmost and will be empty after
5402 * we remove the extent record - the rotate code
5403 * knows how to update the newly formed edge.
5406 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5412 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5413 left_path = ocfs2_new_path_from_path(path);
5420 ret = ocfs2_find_path(et->et_ci, left_path,
5429 ret = ocfs2_extend_rotate_transaction(handle, 0,
5430 handle->h_buffer_credits,
5437 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5443 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5449 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5450 trunc_range = cpos + len;
5452 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5455 memset(rec, 0, sizeof(*rec));
5456 ocfs2_cleanup_merge(el, index);
5458 next_free = le16_to_cpu(el->l_next_free_rec);
5459 if (is_rightmost_tree_rec && next_free > 1) {
5461 * We skip the edge update if this path will
5462 * be deleted by the rotate code.
5464 rec = &el->l_recs[next_free - 1];
5465 ocfs2_adjust_rightmost_records(handle, et, path,
5468 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5469 /* Remove leftmost portion of the record. */
5470 le32_add_cpu(&rec->e_cpos, len);
5471 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5472 le16_add_cpu(&rec->e_leaf_clusters, -len);
5473 } else if (rec_range == trunc_range) {
5474 /* Remove rightmost portion of the record */
5475 le16_add_cpu(&rec->e_leaf_clusters, -len);
5476 if (is_rightmost_tree_rec)
5477 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5479 /* Caller should have trapped this. */
5480 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5482 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5483 le32_to_cpu(rec->e_cpos),
5484 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5491 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5492 ocfs2_complete_edge_insert(handle, left_path, path,
5496 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5498 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5503 ocfs2_free_path(left_path);
5507 int ocfs2_remove_extent(handle_t *handle,
5508 struct ocfs2_extent_tree *et,
5510 struct ocfs2_alloc_context *meta_ac,
5511 struct ocfs2_cached_dealloc_ctxt *dealloc)
5514 u32 rec_range, trunc_range;
5515 struct ocfs2_extent_rec *rec;
5516 struct ocfs2_extent_list *el;
5517 struct ocfs2_path *path = NULL;
5520 * XXX: Why are we truncating to 0 instead of wherever this
5523 ocfs2_et_extent_map_truncate(et, 0);
5525 path = ocfs2_new_path_from_et(et);
5532 ret = ocfs2_find_path(et->et_ci, path, cpos);
5538 el = path_leaf_el(path);
5539 index = ocfs2_search_extent_list(el, cpos);
5541 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5542 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5543 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5550 * We have 3 cases of extent removal:
5551 * 1) Range covers the entire extent rec
5552 * 2) Range begins or ends on one edge of the extent rec
5553 * 3) Range is in the middle of the extent rec (no shared edges)
5555 * For case 1 we remove the extent rec and left rotate to
5558 * For case 2 we just shrink the existing extent rec, with a
5559 * tree update if the shrinking edge is also the edge of an
5562 * For case 3 we do a right split to turn the extent rec into
5563 * something case 2 can handle.
5565 rec = &el->l_recs[index];
5566 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5567 trunc_range = cpos + len;
5569 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5571 trace_ocfs2_remove_extent(
5572 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5573 cpos, len, index, le32_to_cpu(rec->e_cpos),
5574 ocfs2_rec_clusters(el, rec));
5576 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5577 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5584 ret = ocfs2_split_tree(handle, et, path, index,
5585 trunc_range, meta_ac);
5592 * The split could have manipulated the tree enough to
5593 * move the record location, so we have to look for it again.
5595 ocfs2_reinit_path(path, 1);
5597 ret = ocfs2_find_path(et->et_ci, path, cpos);
5603 el = path_leaf_el(path);
5604 index = ocfs2_search_extent_list(el, cpos);
5606 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5607 "Owner %llu: split at cpos %u lost record\n",
5608 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5615 * Double check our values here. If anything is fishy,
5616 * it's easier to catch it at the top level.
5618 rec = &el->l_recs[index];
5619 rec_range = le32_to_cpu(rec->e_cpos) +
5620 ocfs2_rec_clusters(el, rec);
5621 if (rec_range != trunc_range) {
5622 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5623 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5624 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5625 cpos, len, le32_to_cpu(rec->e_cpos),
5626 ocfs2_rec_clusters(el, rec));
5631 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5638 ocfs2_free_path(path);
5643 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5644 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5645 * number to reserve some extra blocks, and it only handles meta
5648 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5649 * and punching holes.
5651 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5652 struct ocfs2_extent_tree *et,
5653 u32 extents_to_split,
5654 struct ocfs2_alloc_context **ac,
5657 int ret = 0, num_free_extents;
5658 unsigned int max_recs_needed = 2 * extents_to_split;
5659 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5663 num_free_extents = ocfs2_num_free_extents(et);
5664 if (num_free_extents < 0) {
5665 ret = num_free_extents;
5670 if (!num_free_extents ||
5671 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5672 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5675 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5685 ocfs2_free_alloc_context(*ac);
5693 int ocfs2_remove_btree_range(struct inode *inode,
5694 struct ocfs2_extent_tree *et,
5695 u32 cpos, u32 phys_cpos, u32 len, int flags,
5696 struct ocfs2_cached_dealloc_ctxt *dealloc,
5697 u64 refcount_loc, bool refcount_tree_locked)
5699 int ret, credits = 0, extra_blocks = 0;
5700 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5701 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5702 struct inode *tl_inode = osb->osb_tl_inode;
5704 struct ocfs2_alloc_context *meta_ac = NULL;
5705 struct ocfs2_refcount_tree *ref_tree = NULL;
5707 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5708 BUG_ON(!ocfs2_is_refcount_inode(inode));
5710 if (!refcount_tree_locked) {
5711 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5719 ret = ocfs2_prepare_refcount_change_for_del(inode,
5731 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5738 inode_lock(tl_inode);
5740 if (ocfs2_truncate_log_needs_flush(osb)) {
5741 ret = __ocfs2_flush_truncate_log(osb);
5748 handle = ocfs2_start_trans(osb,
5749 ocfs2_remove_extent_credits(osb->sb) + credits);
5750 if (IS_ERR(handle)) {
5751 ret = PTR_ERR(handle);
5756 ret = ocfs2_et_root_journal_access(handle, et,
5757 OCFS2_JOURNAL_ACCESS_WRITE);
5763 dquot_free_space_nodirty(inode,
5764 ocfs2_clusters_to_bytes(inode->i_sb, len));
5766 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5772 ocfs2_et_update_clusters(et, -len);
5773 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5775 ocfs2_journal_dirty(handle, et->et_root_bh);
5778 if (flags & OCFS2_EXT_REFCOUNTED)
5779 ret = ocfs2_decrease_refcount(inode, handle,
5780 ocfs2_blocks_to_clusters(osb->sb,
5785 ret = ocfs2_truncate_log_append(osb, handle,
5793 ocfs2_commit_trans(osb, handle);
5795 inode_unlock(tl_inode);
5798 ocfs2_free_alloc_context(meta_ac);
5801 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5806 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5808 struct buffer_head *tl_bh = osb->osb_tl_bh;
5809 struct ocfs2_dinode *di;
5810 struct ocfs2_truncate_log *tl;
5812 di = (struct ocfs2_dinode *) tl_bh->b_data;
5813 tl = &di->id2.i_dealloc;
5815 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5816 "slot %d, invalid truncate log parameters: used = "
5817 "%u, count = %u\n", osb->slot_num,
5818 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5819 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5822 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5823 unsigned int new_start)
5825 unsigned int tail_index;
5826 unsigned int current_tail;
5828 /* No records, nothing to coalesce */
5829 if (!le16_to_cpu(tl->tl_used))
5832 tail_index = le16_to_cpu(tl->tl_used) - 1;
5833 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5834 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5836 return current_tail == new_start;
5839 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5842 unsigned int num_clusters)
5845 unsigned int start_cluster, tl_count;
5846 struct inode *tl_inode = osb->osb_tl_inode;
5847 struct buffer_head *tl_bh = osb->osb_tl_bh;
5848 struct ocfs2_dinode *di;
5849 struct ocfs2_truncate_log *tl;
5851 BUG_ON(inode_trylock(tl_inode));
5853 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5855 di = (struct ocfs2_dinode *) tl_bh->b_data;
5857 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5858 * by the underlying call to ocfs2_read_inode_block(), so any
5859 * corruption is a code bug */
5860 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5862 tl = &di->id2.i_dealloc;
5863 tl_count = le16_to_cpu(tl->tl_count);
5864 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5866 "Truncate record count on #%llu invalid "
5867 "wanted %u, actual %u\n",
5868 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5869 ocfs2_truncate_recs_per_inode(osb->sb),
5870 le16_to_cpu(tl->tl_count));
5872 /* Caller should have known to flush before calling us. */
5873 index = le16_to_cpu(tl->tl_used);
5874 if (index >= tl_count) {
5880 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5881 OCFS2_JOURNAL_ACCESS_WRITE);
5887 trace_ocfs2_truncate_log_append(
5888 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5889 start_cluster, num_clusters);
5890 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5892 * Move index back to the record we are coalescing with.
5893 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5897 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5898 trace_ocfs2_truncate_log_append(
5899 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5900 index, le32_to_cpu(tl->tl_recs[index].t_start),
5903 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5904 tl->tl_used = cpu_to_le16(index + 1);
5906 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5908 ocfs2_journal_dirty(handle, tl_bh);
5910 osb->truncated_clusters += num_clusters;
5915 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5916 struct inode *data_alloc_inode,
5917 struct buffer_head *data_alloc_bh)
5921 unsigned int num_clusters;
5923 struct ocfs2_truncate_rec rec;
5924 struct ocfs2_dinode *di;
5925 struct ocfs2_truncate_log *tl;
5926 struct inode *tl_inode = osb->osb_tl_inode;
5927 struct buffer_head *tl_bh = osb->osb_tl_bh;
5930 di = (struct ocfs2_dinode *) tl_bh->b_data;
5931 tl = &di->id2.i_dealloc;
5932 i = le16_to_cpu(tl->tl_used) - 1;
5934 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5935 if (IS_ERR(handle)) {
5936 status = PTR_ERR(handle);
5941 /* Caller has given us at least enough credits to
5942 * update the truncate log dinode */
5943 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5944 OCFS2_JOURNAL_ACCESS_WRITE);
5950 tl->tl_used = cpu_to_le16(i);
5952 ocfs2_journal_dirty(handle, tl_bh);
5954 rec = tl->tl_recs[i];
5955 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5956 le32_to_cpu(rec.t_start));
5957 num_clusters = le32_to_cpu(rec.t_clusters);
5959 /* if start_blk is not set, we ignore the record as
5962 trace_ocfs2_replay_truncate_records(
5963 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5964 i, le32_to_cpu(rec.t_start), num_clusters);
5966 status = ocfs2_free_clusters(handle, data_alloc_inode,
5967 data_alloc_bh, start_blk,
5975 ocfs2_commit_trans(osb, handle);
5979 osb->truncated_clusters = 0;
5985 /* Expects you to already be holding tl_inode->i_mutex */
5986 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5989 unsigned int num_to_flush;
5990 struct inode *tl_inode = osb->osb_tl_inode;
5991 struct inode *data_alloc_inode = NULL;
5992 struct buffer_head *tl_bh = osb->osb_tl_bh;
5993 struct buffer_head *data_alloc_bh = NULL;
5994 struct ocfs2_dinode *di;
5995 struct ocfs2_truncate_log *tl;
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 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6017 GLOBAL_BITMAP_SYSTEM_INODE,
6018 OCFS2_INVALID_SLOT);
6019 if (!data_alloc_inode) {
6021 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6025 inode_lock(data_alloc_inode);
6027 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6033 status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6038 brelse(data_alloc_bh);
6039 ocfs2_inode_unlock(data_alloc_inode, 1);
6042 inode_unlock(data_alloc_inode);
6043 iput(data_alloc_inode);
6049 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6052 struct inode *tl_inode = osb->osb_tl_inode;
6054 inode_lock(tl_inode);
6055 status = __ocfs2_flush_truncate_log(osb);
6056 inode_unlock(tl_inode);
6061 static void ocfs2_truncate_log_worker(struct work_struct *work)
6064 struct ocfs2_super *osb =
6065 container_of(work, struct ocfs2_super,
6066 osb_truncate_log_wq.work);
6068 status = ocfs2_flush_truncate_log(osb);
6072 ocfs2_init_steal_slots(osb);
6075 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6076 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6079 if (osb->osb_tl_inode &&
6080 atomic_read(&osb->osb_tl_disable) == 0) {
6081 /* We want to push off log flushes while truncates are
6084 cancel_delayed_work(&osb->osb_truncate_log_wq);
6086 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6087 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6092 * Try to flush truncate logs if we can free enough clusters from it.
6093 * As for return value, "< 0" means error, "0" no space and "1" means
6094 * we have freed enough spaces and let the caller try to allocate again.
6096 int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6097 unsigned int needed)
6101 unsigned int truncated_clusters;
6103 inode_lock(osb->osb_tl_inode);
6104 truncated_clusters = osb->truncated_clusters;
6105 inode_unlock(osb->osb_tl_inode);
6108 * Check whether we can succeed in allocating if we free
6111 if (truncated_clusters < needed)
6114 ret = ocfs2_flush_truncate_log(osb);
6120 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6121 jbd2_log_wait_commit(osb->journal->j_journal, target);
6128 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6130 struct inode **tl_inode,
6131 struct buffer_head **tl_bh)
6134 struct inode *inode = NULL;
6135 struct buffer_head *bh = NULL;
6137 inode = ocfs2_get_system_file_inode(osb,
6138 TRUNCATE_LOG_SYSTEM_INODE,
6142 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6146 status = ocfs2_read_inode_block(inode, &bh);
6159 /* called during the 1st stage of node recovery. we stamp a clean
6160 * truncate log and pass back a copy for processing later. if the
6161 * truncate log does not require processing, a *tl_copy is set to
6163 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6165 struct ocfs2_dinode **tl_copy)
6168 struct inode *tl_inode = NULL;
6169 struct buffer_head *tl_bh = NULL;
6170 struct ocfs2_dinode *di;
6171 struct ocfs2_truncate_log *tl;
6175 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6177 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6183 di = (struct ocfs2_dinode *) tl_bh->b_data;
6185 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6186 * validated by the underlying call to ocfs2_read_inode_block(),
6187 * so any corruption is a code bug */
6188 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6190 tl = &di->id2.i_dealloc;
6191 if (le16_to_cpu(tl->tl_used)) {
6192 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6195 * Assuming the write-out below goes well, this copy will be
6196 * passed back to recovery for processing.
6198 *tl_copy = kmemdup(tl_bh->b_data, tl_bh->b_size, GFP_KERNEL);
6205 /* All we need to do to clear the truncate log is set
6209 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6210 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6230 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6231 struct ocfs2_dinode *tl_copy)
6235 unsigned int clusters, num_recs, start_cluster;
6238 struct inode *tl_inode = osb->osb_tl_inode;
6239 struct ocfs2_truncate_log *tl;
6241 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6242 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6246 tl = &tl_copy->id2.i_dealloc;
6247 num_recs = le16_to_cpu(tl->tl_used);
6248 trace_ocfs2_complete_truncate_log_recovery(
6249 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6252 inode_lock(tl_inode);
6253 for(i = 0; i < num_recs; i++) {
6254 if (ocfs2_truncate_log_needs_flush(osb)) {
6255 status = __ocfs2_flush_truncate_log(osb);
6262 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6263 if (IS_ERR(handle)) {
6264 status = PTR_ERR(handle);
6269 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6270 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6271 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6273 status = ocfs2_truncate_log_append(osb, handle,
6274 start_blk, clusters);
6275 ocfs2_commit_trans(osb, handle);
6283 inode_unlock(tl_inode);
6288 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6291 struct inode *tl_inode = osb->osb_tl_inode;
6293 atomic_set(&osb->osb_tl_disable, 1);
6296 cancel_delayed_work(&osb->osb_truncate_log_wq);
6297 flush_workqueue(osb->ocfs2_wq);
6299 status = ocfs2_flush_truncate_log(osb);
6303 brelse(osb->osb_tl_bh);
6304 iput(osb->osb_tl_inode);
6308 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6311 struct inode *tl_inode = NULL;
6312 struct buffer_head *tl_bh = NULL;
6314 status = ocfs2_get_truncate_log_info(osb,
6321 /* ocfs2_truncate_log_shutdown keys on the existence of
6322 * osb->osb_tl_inode so we don't set any of the osb variables
6323 * until we're sure all is well. */
6324 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6325 ocfs2_truncate_log_worker);
6326 atomic_set(&osb->osb_tl_disable, 0);
6327 osb->osb_tl_bh = tl_bh;
6328 osb->osb_tl_inode = tl_inode;
6334 * Delayed de-allocation of suballocator blocks.
6336 * Some sets of block de-allocations might involve multiple suballocator inodes.
6338 * The locking for this can get extremely complicated, especially when
6339 * the suballocator inodes to delete from aren't known until deep
6340 * within an unrelated codepath.
6342 * ocfs2_extent_block structures are a good example of this - an inode
6343 * btree could have been grown by any number of nodes each allocating
6344 * out of their own suballoc inode.
6346 * These structures allow the delay of block de-allocation until a
6347 * later time, when locking of multiple cluster inodes won't cause
6352 * Describe a single bit freed from a suballocator. For the block
6353 * suballocators, it represents one block. For the global cluster
6354 * allocator, it represents some clusters and free_bit indicates
6357 struct ocfs2_cached_block_free {
6358 struct ocfs2_cached_block_free *free_next;
6361 unsigned int free_bit;
6364 struct ocfs2_per_slot_free_list {
6365 struct ocfs2_per_slot_free_list *f_next_suballocator;
6368 struct ocfs2_cached_block_free *f_first;
6371 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6374 struct ocfs2_cached_block_free *head)
6379 struct inode *inode;
6380 struct buffer_head *di_bh = NULL;
6381 struct ocfs2_cached_block_free *tmp;
6383 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6392 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6400 bg_blkno = head->free_bg;
6402 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6404 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6405 if (IS_ERR(handle)) {
6406 ret = PTR_ERR(handle);
6411 trace_ocfs2_free_cached_blocks(
6412 (unsigned long long)head->free_blk, head->free_bit);
6414 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6415 head->free_bit, bg_blkno, 1);
6419 ocfs2_commit_trans(osb, handle);
6422 head = head->free_next;
6427 ocfs2_inode_unlock(inode, 1);
6430 inode_unlock(inode);
6434 /* Premature exit may have left some dangling items. */
6436 head = head->free_next;
6443 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6444 u64 blkno, unsigned int bit)
6447 struct ocfs2_cached_block_free *item;
6449 item = kzalloc(sizeof(*item), GFP_NOFS);
6456 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6458 item->free_blk = blkno;
6459 item->free_bit = bit;
6460 item->free_next = ctxt->c_global_allocator;
6462 ctxt->c_global_allocator = item;
6466 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6467 struct ocfs2_cached_block_free *head)
6469 struct ocfs2_cached_block_free *tmp;
6470 struct inode *tl_inode = osb->osb_tl_inode;
6474 inode_lock(tl_inode);
6477 if (ocfs2_truncate_log_needs_flush(osb)) {
6478 ret = __ocfs2_flush_truncate_log(osb);
6485 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6486 if (IS_ERR(handle)) {
6487 ret = PTR_ERR(handle);
6492 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6495 ocfs2_commit_trans(osb, handle);
6497 head = head->free_next;
6506 inode_unlock(tl_inode);
6509 /* Premature exit may have left some dangling items. */
6511 head = head->free_next;
6518 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6519 struct ocfs2_cached_dealloc_ctxt *ctxt)
6522 struct ocfs2_per_slot_free_list *fl;
6527 while (ctxt->c_first_suballocator) {
6528 fl = ctxt->c_first_suballocator;
6531 trace_ocfs2_run_deallocs(fl->f_inode_type,
6533 ret2 = ocfs2_free_cached_blocks(osb,
6543 ctxt->c_first_suballocator = fl->f_next_suballocator;
6547 if (ctxt->c_global_allocator) {
6548 ret2 = ocfs2_free_cached_clusters(osb,
6549 ctxt->c_global_allocator);
6555 ctxt->c_global_allocator = NULL;
6561 static struct ocfs2_per_slot_free_list *
6562 ocfs2_find_per_slot_free_list(int type,
6564 struct ocfs2_cached_dealloc_ctxt *ctxt)
6566 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6569 if (fl->f_inode_type == type && fl->f_slot == slot)
6572 fl = fl->f_next_suballocator;
6575 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6577 fl->f_inode_type = type;
6580 fl->f_next_suballocator = ctxt->c_first_suballocator;
6582 ctxt->c_first_suballocator = fl;
6587 static struct ocfs2_per_slot_free_list *
6588 ocfs2_find_preferred_free_list(int type,
6591 struct ocfs2_cached_dealloc_ctxt *ctxt)
6593 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6596 if (fl->f_inode_type == type && fl->f_slot == preferred_slot) {
6597 *real_slot = fl->f_slot;
6601 fl = fl->f_next_suballocator;
6604 /* If we can't find any free list matching preferred slot, just use
6607 fl = ctxt->c_first_suballocator;
6608 *real_slot = fl->f_slot;
6613 /* Return Value 1 indicates empty */
6614 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et)
6616 struct ocfs2_per_slot_free_list *fl = NULL;
6618 if (!et->et_dealloc)
6621 fl = et->et_dealloc->c_first_suballocator;
6631 /* If extent was deleted from tree due to extent rotation and merging, and
6632 * no metadata is reserved ahead of time. Try to reuse some extents
6633 * just deleted. This is only used to reuse extent blocks.
6634 * It is supposed to find enough extent blocks in dealloc if our estimation
6635 * on metadata is accurate.
6637 static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
6638 struct ocfs2_extent_tree *et,
6639 struct buffer_head **new_eb_bh,
6640 int blk_wanted, int *blk_given)
6642 int i, status = 0, real_slot;
6643 struct ocfs2_cached_dealloc_ctxt *dealloc;
6644 struct ocfs2_per_slot_free_list *fl;
6645 struct ocfs2_cached_block_free *bf;
6646 struct ocfs2_extent_block *eb;
6647 struct ocfs2_super *osb =
6648 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
6652 /* If extent tree doesn't have a dealloc, this is not faulty. Just
6653 * tell upper caller dealloc can't provide any block and it should
6654 * ask for alloc to claim more space.
6656 dealloc = et->et_dealloc;
6660 for (i = 0; i < blk_wanted; i++) {
6661 /* Prefer to use local slot */
6662 fl = ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE,
6663 osb->slot_num, &real_slot,
6665 /* If no more block can be reused, we should claim more
6666 * from alloc. Just return here normally.
6674 fl->f_first = bf->free_next;
6676 new_eb_bh[i] = sb_getblk(osb->sb, bf->free_blk);
6677 if (new_eb_bh[i] == NULL) {
6683 mlog(0, "Reusing block(%llu) from "
6684 "dealloc(local slot:%d, real slot:%d)\n",
6685 bf->free_blk, osb->slot_num, real_slot);
6687 ocfs2_set_new_buffer_uptodate(et->et_ci, new_eb_bh[i]);
6689 status = ocfs2_journal_access_eb(handle, et->et_ci,
6691 OCFS2_JOURNAL_ACCESS_CREATE);
6697 memset(new_eb_bh[i]->b_data, 0, osb->sb->s_blocksize);
6698 eb = (struct ocfs2_extent_block *) new_eb_bh[i]->b_data;
6700 /* We can't guarantee that buffer head is still cached, so
6701 * polutlate the extent block again.
6703 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
6704 eb->h_blkno = cpu_to_le64(bf->free_blk);
6705 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
6706 eb->h_suballoc_slot = cpu_to_le16(real_slot);
6707 eb->h_suballoc_loc = cpu_to_le64(bf->free_bg);
6708 eb->h_suballoc_bit = cpu_to_le16(bf->free_bit);
6709 eb->h_list.l_count =
6710 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
6712 /* We'll also be dirtied by the caller, so
6713 * this isn't absolutely necessary.
6715 ocfs2_journal_dirty(handle, new_eb_bh[i]);
6718 dealloc->c_first_suballocator = fl->f_next_suballocator;
6727 if (unlikely(status < 0)) {
6728 for (i = 0; i < blk_wanted; i++)
6729 brelse(new_eb_bh[i]);
6735 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6736 int type, int slot, u64 suballoc,
6737 u64 blkno, unsigned int bit)
6740 struct ocfs2_per_slot_free_list *fl;
6741 struct ocfs2_cached_block_free *item;
6743 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6750 item = kzalloc(sizeof(*item), GFP_NOFS);
6757 trace_ocfs2_cache_block_dealloc(type, slot,
6758 (unsigned long long)suballoc,
6759 (unsigned long long)blkno, bit);
6761 item->free_bg = suballoc;
6762 item->free_blk = blkno;
6763 item->free_bit = bit;
6764 item->free_next = fl->f_first;
6773 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6774 struct ocfs2_extent_block *eb)
6776 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6777 le16_to_cpu(eb->h_suballoc_slot),
6778 le64_to_cpu(eb->h_suballoc_loc),
6779 le64_to_cpu(eb->h_blkno),
6780 le16_to_cpu(eb->h_suballoc_bit));
6783 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6785 set_buffer_uptodate(bh);
6786 mark_buffer_dirty(bh);
6790 void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6791 unsigned int from, unsigned int to,
6792 struct page *page, int zero, u64 *phys)
6794 int ret, partial = 0;
6796 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6801 zero_user_segment(page, from, to);
6804 * Need to set the buffers we zero'd into uptodate
6805 * here if they aren't - ocfs2_map_page_blocks()
6806 * might've skipped some
6808 ret = walk_page_buffers(handle, page_buffers(page),
6813 else if (ocfs2_should_order_data(inode)) {
6814 ret = ocfs2_jbd2_file_inode(handle, inode);
6820 SetPageUptodate(page);
6822 flush_dcache_page(page);
6825 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6826 loff_t end, struct page **pages,
6827 int numpages, u64 phys, handle_t *handle)
6831 unsigned int from, to = PAGE_SIZE;
6832 struct super_block *sb = inode->i_sb;
6834 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6840 for(i = 0; i < numpages; i++) {
6843 from = start & (PAGE_SIZE - 1);
6844 if ((end >> PAGE_SHIFT) == page->index)
6845 to = end & (PAGE_SIZE - 1);
6847 BUG_ON(from > PAGE_SIZE);
6848 BUG_ON(to > PAGE_SIZE);
6850 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6853 start = (page->index + 1) << PAGE_SHIFT;
6857 ocfs2_unlock_and_free_pages(pages, numpages);
6860 int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6861 struct page **pages, int *num)
6863 int numpages, ret = 0;
6864 struct address_space *mapping = inode->i_mapping;
6865 unsigned long index;
6866 loff_t last_page_bytes;
6868 BUG_ON(start > end);
6871 last_page_bytes = PAGE_ALIGN(end);
6872 index = start >> PAGE_SHIFT;
6874 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6875 if (!pages[numpages]) {
6883 } while (index < (last_page_bytes >> PAGE_SHIFT));
6888 ocfs2_unlock_and_free_pages(pages, numpages);
6897 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6898 struct page **pages, int *num)
6900 struct super_block *sb = inode->i_sb;
6902 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6903 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6905 return ocfs2_grab_pages(inode, start, end, pages, num);
6909 * Zero the area past i_size but still within an allocated
6910 * cluster. This avoids exposing nonzero data on subsequent file
6913 * We need to call this before i_size is updated on the inode because
6914 * otherwise block_write_full_page() will skip writeout of pages past
6915 * i_size. The new_i_size parameter is passed for this reason.
6917 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6918 u64 range_start, u64 range_end)
6920 int ret = 0, numpages;
6921 struct page **pages = NULL;
6923 unsigned int ext_flags;
6924 struct super_block *sb = inode->i_sb;
6927 * File systems which don't support sparse files zero on every
6930 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6933 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6934 sizeof(struct page *), GFP_NOFS);
6935 if (pages == NULL) {
6941 if (range_start == range_end)
6944 ret = ocfs2_extent_map_get_blocks(inode,
6945 range_start >> sb->s_blocksize_bits,
6946 &phys, NULL, &ext_flags);
6953 * Tail is a hole, or is marked unwritten. In either case, we
6954 * can count on read and write to return/push zero's.
6956 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6959 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6966 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6967 numpages, phys, handle);
6970 * Initiate writeout of the pages we zero'd here. We don't
6971 * wait on them - the truncate_inode_pages() call later will
6974 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6985 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6986 struct ocfs2_dinode *di)
6988 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6989 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6991 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6992 memset(&di->id2, 0, blocksize -
6993 offsetof(struct ocfs2_dinode, id2) -
6996 memset(&di->id2, 0, blocksize -
6997 offsetof(struct ocfs2_dinode, id2));
7000 void ocfs2_dinode_new_extent_list(struct inode *inode,
7001 struct ocfs2_dinode *di)
7003 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7004 di->id2.i_list.l_tree_depth = 0;
7005 di->id2.i_list.l_next_free_rec = 0;
7006 di->id2.i_list.l_count = cpu_to_le16(
7007 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7010 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7012 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7013 struct ocfs2_inline_data *idata = &di->id2.i_data;
7015 spin_lock(&oi->ip_lock);
7016 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7017 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7018 spin_unlock(&oi->ip_lock);
7021 * We clear the entire i_data structure here so that all
7022 * fields can be properly initialized.
7024 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7026 idata->id_count = cpu_to_le16(
7027 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7030 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7031 struct buffer_head *di_bh)
7033 int ret, i, has_data, num_pages = 0;
7037 u64 uninitialized_var(block);
7038 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7039 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7040 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7041 struct ocfs2_alloc_context *data_ac = NULL;
7042 struct page **pages = NULL;
7043 loff_t end = osb->s_clustersize;
7044 struct ocfs2_extent_tree et;
7047 has_data = i_size_read(inode) ? 1 : 0;
7050 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
7051 sizeof(struct page *), GFP_NOFS);
7052 if (pages == NULL) {
7058 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7065 handle = ocfs2_start_trans(osb,
7066 ocfs2_inline_to_extents_credits(osb->sb));
7067 if (IS_ERR(handle)) {
7068 ret = PTR_ERR(handle);
7073 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7074 OCFS2_JOURNAL_ACCESS_WRITE);
7081 unsigned int page_end;
7084 ret = dquot_alloc_space_nodirty(inode,
7085 ocfs2_clusters_to_bytes(osb->sb, 1));
7090 data_ac->ac_resv = &oi->ip_la_data_resv;
7092 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
7100 * Save two copies, one for insert, and one that can
7101 * be changed by ocfs2_map_and_dirty_page() below.
7103 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7106 * Non sparse file systems zero on extend, so no need
7109 if (!ocfs2_sparse_alloc(osb) &&
7110 PAGE_SIZE < osb->s_clustersize)
7113 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
7121 * This should populate the 1st page for us and mark
7124 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
7131 page_end = PAGE_SIZE;
7132 if (PAGE_SIZE > osb->s_clustersize)
7133 page_end = osb->s_clustersize;
7135 for (i = 0; i < num_pages; i++)
7136 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
7137 pages[i], i > 0, &phys);
7140 spin_lock(&oi->ip_lock);
7141 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7142 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7143 spin_unlock(&oi->ip_lock);
7145 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7146 ocfs2_dinode_new_extent_list(inode, di);
7148 ocfs2_journal_dirty(handle, di_bh);
7152 * An error at this point should be extremely rare. If
7153 * this proves to be false, we could always re-build
7154 * the in-inode data from our pages.
7156 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7157 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7164 inode->i_blocks = ocfs2_inode_sector_count(inode);
7169 ocfs2_unlock_and_free_pages(pages, num_pages);
7172 if (ret < 0 && did_quota)
7173 dquot_free_space_nodirty(inode,
7174 ocfs2_clusters_to_bytes(osb->sb, 1));
7177 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7178 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7181 ocfs2_free_clusters(handle,
7184 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7188 ocfs2_commit_trans(osb, handle);
7192 ocfs2_free_alloc_context(data_ac);
7199 * It is expected, that by the time you call this function,
7200 * inode->i_size and fe->i_size have been adjusted.
7202 * WARNING: This will kfree the truncate context
7204 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7205 struct inode *inode,
7206 struct buffer_head *di_bh)
7208 int status = 0, i, flags = 0;
7209 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7211 struct ocfs2_extent_list *el;
7212 struct ocfs2_extent_rec *rec;
7213 struct ocfs2_path *path = NULL;
7214 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7215 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7216 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7217 struct ocfs2_extent_tree et;
7218 struct ocfs2_cached_dealloc_ctxt dealloc;
7219 struct ocfs2_refcount_tree *ref_tree = NULL;
7221 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7222 ocfs2_init_dealloc_ctxt(&dealloc);
7224 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7225 i_size_read(inode));
7227 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7228 ocfs2_journal_access_di);
7235 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7239 * Check that we still have allocation to delete.
7241 if (OCFS2_I(inode)->ip_clusters == 0) {
7247 * Truncate always works against the rightmost tree branch.
7249 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7255 trace_ocfs2_commit_truncate(
7256 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7258 OCFS2_I(inode)->ip_clusters,
7259 path->p_tree_depth);
7262 * By now, el will point to the extent list on the bottom most
7263 * portion of this tree. Only the tail record is considered in
7266 * We handle the following cases, in order:
7267 * - empty extent: delete the remaining branch
7268 * - remove the entire record
7269 * - remove a partial record
7270 * - no record needs to be removed (truncate has completed)
7272 el = path_leaf_el(path);
7273 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7274 ocfs2_error(inode->i_sb,
7275 "Inode %llu has empty extent block at %llu\n",
7276 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7277 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7282 i = le16_to_cpu(el->l_next_free_rec) - 1;
7283 rec = &el->l_recs[i];
7284 flags = rec->e_flags;
7285 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7287 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7289 * Lower levels depend on this never happening, but it's best
7290 * to check it up here before changing the tree.
7292 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7293 mlog(ML_ERROR, "Inode %lu has an empty "
7294 "extent record, depth %u\n", inode->i_ino,
7295 le16_to_cpu(root_el->l_tree_depth));
7296 status = ocfs2_remove_rightmost_empty_extent(osb,
7297 &et, path, &dealloc);
7303 ocfs2_reinit_path(path, 1);
7306 trunc_cpos = le32_to_cpu(rec->e_cpos);
7310 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7312 * Truncate entire record.
7314 trunc_cpos = le32_to_cpu(rec->e_cpos);
7315 trunc_len = ocfs2_rec_clusters(el, rec);
7316 blkno = le64_to_cpu(rec->e_blkno);
7317 } else if (range > new_highest_cpos) {
7319 * Partial truncate. it also should be
7320 * the last truncate we're doing.
7322 trunc_cpos = new_highest_cpos;
7323 trunc_len = range - new_highest_cpos;
7324 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7325 blkno = le64_to_cpu(rec->e_blkno) +
7326 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7329 * Truncate completed, leave happily.
7335 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7337 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7338 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7346 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7347 phys_cpos, trunc_len, flags, &dealloc,
7348 refcount_loc, true);
7354 ocfs2_reinit_path(path, 1);
7357 * The check above will catch the case where we've truncated
7358 * away all allocation.
7364 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7366 ocfs2_schedule_truncate_log_flush(osb, 1);
7368 ocfs2_run_deallocs(osb, &dealloc);
7370 ocfs2_free_path(path);
7376 * 'start' is inclusive, 'end' is not.
7378 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7379 unsigned int start, unsigned int end, int trunc)
7382 unsigned int numbytes;
7384 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7385 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7386 struct ocfs2_inline_data *idata = &di->id2.i_data;
7388 if (end > i_size_read(inode))
7389 end = i_size_read(inode);
7391 BUG_ON(start > end);
7393 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7394 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7395 !ocfs2_supports_inline_data(osb)) {
7396 ocfs2_error(inode->i_sb,
7397 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7398 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7399 le16_to_cpu(di->i_dyn_features),
7400 OCFS2_I(inode)->ip_dyn_features,
7401 osb->s_feature_incompat);
7406 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7407 if (IS_ERR(handle)) {
7408 ret = PTR_ERR(handle);
7413 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7414 OCFS2_JOURNAL_ACCESS_WRITE);
7420 numbytes = end - start;
7421 memset(idata->id_data + start, 0, numbytes);
7424 * No need to worry about the data page here - it's been
7425 * truncated already and inline data doesn't need it for
7426 * pushing zero's to disk, so we'll let readpage pick it up
7430 i_size_write(inode, start);
7431 di->i_size = cpu_to_le64(start);
7434 inode->i_blocks = ocfs2_inode_sector_count(inode);
7435 inode->i_ctime = inode->i_mtime = current_time(inode);
7437 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7438 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7440 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7441 ocfs2_journal_dirty(handle, di_bh);
7444 ocfs2_commit_trans(osb, handle);
7450 static int ocfs2_trim_extent(struct super_block *sb,
7451 struct ocfs2_group_desc *gd,
7452 u64 group, u32 start, u32 count)
7454 u64 discard, bcount;
7455 struct ocfs2_super *osb = OCFS2_SB(sb);
7457 bcount = ocfs2_clusters_to_blocks(sb, count);
7458 discard = ocfs2_clusters_to_blocks(sb, start);
7461 * For the first cluster group, the gd->bg_blkno is not at the start
7462 * of the group, but at an offset from the start. If we add it while
7463 * calculating discard for first group, we will wrongly start fstrim a
7464 * few blocks after the desried start block and the range can cross
7465 * over into the next cluster group. So, add it only if this is not
7466 * the first cluster group.
7468 if (group != osb->first_cluster_group_blkno)
7469 discard += le64_to_cpu(gd->bg_blkno);
7471 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7473 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7476 static int ocfs2_trim_group(struct super_block *sb,
7477 struct ocfs2_group_desc *gd, u64 group,
7478 u32 start, u32 max, u32 minbits)
7480 int ret = 0, count = 0, next;
7481 void *bitmap = gd->bg_bitmap;
7483 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7486 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7487 start, max, minbits);
7489 while (start < max) {
7490 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7493 next = ocfs2_find_next_bit(bitmap, max, start);
7495 if ((next - start) >= minbits) {
7496 ret = ocfs2_trim_extent(sb, gd, group,
7497 start, next - start);
7502 count += next - start;
7506 if (fatal_signal_pending(current)) {
7507 count = -ERESTARTSYS;
7511 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7522 int ocfs2_trim_mainbm(struct super_block *sb, struct fstrim_range *range)
7524 struct ocfs2_super *osb = OCFS2_SB(sb);
7525 u64 start, len, trimmed = 0, first_group, last_group = 0, group = 0;
7527 u32 first_bit, last_bit, minlen;
7528 struct buffer_head *main_bm_bh = NULL;
7529 struct inode *main_bm_inode = NULL;
7530 struct buffer_head *gd_bh = NULL;
7531 struct ocfs2_dinode *main_bm;
7532 struct ocfs2_group_desc *gd = NULL;
7534 start = range->start >> osb->s_clustersize_bits;
7535 len = range->len >> osb->s_clustersize_bits;
7536 minlen = range->minlen >> osb->s_clustersize_bits;
7538 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7541 trace_ocfs2_trim_mainbm(start, len, minlen);
7544 main_bm_inode = ocfs2_get_system_file_inode(osb,
7545 GLOBAL_BITMAP_SYSTEM_INODE,
7546 OCFS2_INVALID_SLOT);
7547 if (!main_bm_inode) {
7553 inode_lock(main_bm_inode);
7555 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7560 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7563 * Do some check before trim the first group.
7566 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7571 if (start + len > le32_to_cpu(main_bm->i_clusters))
7572 len = le32_to_cpu(main_bm->i_clusters) - start;
7575 * Determine first and last group to examine based on
7578 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7579 if (first_group == osb->first_cluster_group_blkno)
7582 first_bit = start - ocfs2_blocks_to_clusters(sb,
7584 last_group = ocfs2_which_cluster_group(main_bm_inode,
7586 group = first_group;
7590 if (first_bit + len >= osb->bitmap_cpg)
7591 last_bit = osb->bitmap_cpg;
7593 last_bit = first_bit + len;
7595 ret = ocfs2_read_group_descriptor(main_bm_inode,
7603 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7604 cnt = ocfs2_trim_group(sb, gd, group,
7605 first_bit, last_bit, minlen);
7615 len -= osb->bitmap_cpg - first_bit;
7617 if (group == osb->first_cluster_group_blkno)
7618 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7620 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7624 ocfs2_inode_unlock(main_bm_inode, 0);
7628 inode_unlock(main_bm_inode);
7629 iput(main_bm_inode);
7632 * If all the groups trim are not done or failed, but we should release
7633 * main_bm related locks for avoiding the current IO starve, then go to
7634 * trim the next group
7636 if (ret >= 0 && group <= last_group)
7639 range->len = trimmed * sb->s_blocksize;
7643 int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7646 struct ocfs2_super *osb = OCFS2_SB(sb);
7647 struct ocfs2_trim_fs_info info, *pinfo = NULL;
7649 ocfs2_trim_fs_lock_res_init(osb);
7651 trace_ocfs2_trim_fs(range->start, range->len, range->minlen);
7653 ret = ocfs2_trim_fs_lock(osb, NULL, 1);
7655 if (ret != -EAGAIN) {
7657 ocfs2_trim_fs_lock_res_uninit(osb);
7661 mlog(ML_NOTICE, "Wait for trim on device (%s) to "
7662 "finish, which is running from another node.\n",
7664 ret = ocfs2_trim_fs_lock(osb, &info, 0);
7667 ocfs2_trim_fs_lock_res_uninit(osb);
7671 if (info.tf_valid && info.tf_success &&
7672 info.tf_start == range->start &&
7673 info.tf_len == range->len &&
7674 info.tf_minlen == range->minlen) {
7675 /* Avoid sending duplicated trim to a shared device */
7676 mlog(ML_NOTICE, "The same trim on device (%s) was "
7677 "just done from node (%u), return.\n",
7678 osb->dev_str, info.tf_nodenum);
7679 range->len = info.tf_trimlen;
7684 info.tf_nodenum = osb->node_num;
7685 info.tf_start = range->start;
7686 info.tf_len = range->len;
7687 info.tf_minlen = range->minlen;
7689 ret = ocfs2_trim_mainbm(sb, range);
7691 info.tf_trimlen = range->len;
7692 info.tf_success = (ret < 0 ? 0 : 1);
7695 ocfs2_trim_fs_unlock(osb, pinfo);
7696 ocfs2_trim_fs_lock_res_uninit(osb);