Pull cpuidle into release branch
[linux-2.6] / fs / ocfs2 / alloc.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * alloc.c
5  *
6  * Extent allocs and frees
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
31
32 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
33 #include <cluster/masklog.h>
34
35 #include "ocfs2.h"
36
37 #include "alloc.h"
38 #include "aops.h"
39 #include "dlmglue.h"
40 #include "extent_map.h"
41 #include "inode.h"
42 #include "journal.h"
43 #include "localalloc.h"
44 #include "suballoc.h"
45 #include "sysfile.h"
46 #include "file.h"
47 #include "super.h"
48 #include "uptodate.h"
49
50 #include "buffer_head_io.h"
51
52 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
53 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
54                                          struct ocfs2_extent_block *eb);
55
56 /*
57  * Structures which describe a path through a btree, and functions to
58  * manipulate them.
59  *
60  * The idea here is to be as generic as possible with the tree
61  * manipulation code.
62  */
63 struct ocfs2_path_item {
64         struct buffer_head              *bh;
65         struct ocfs2_extent_list        *el;
66 };
67
68 #define OCFS2_MAX_PATH_DEPTH    5
69
70 struct ocfs2_path {
71         int                     p_tree_depth;
72         struct ocfs2_path_item  p_node[OCFS2_MAX_PATH_DEPTH];
73 };
74
75 #define path_root_bh(_path) ((_path)->p_node[0].bh)
76 #define path_root_el(_path) ((_path)->p_node[0].el)
77 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
78 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
79 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
80
81 /*
82  * Reset the actual path elements so that we can re-use the structure
83  * to build another path. Generally, this involves freeing the buffer
84  * heads.
85  */
86 static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
87 {
88         int i, start = 0, depth = 0;
89         struct ocfs2_path_item *node;
90
91         if (keep_root)
92                 start = 1;
93
94         for(i = start; i < path_num_items(path); i++) {
95                 node = &path->p_node[i];
96
97                 brelse(node->bh);
98                 node->bh = NULL;
99                 node->el = NULL;
100         }
101
102         /*
103          * Tree depth may change during truncate, or insert. If we're
104          * keeping the root extent list, then make sure that our path
105          * structure reflects the proper depth.
106          */
107         if (keep_root)
108                 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
109
110         path->p_tree_depth = depth;
111 }
112
113 static void ocfs2_free_path(struct ocfs2_path *path)
114 {
115         if (path) {
116                 ocfs2_reinit_path(path, 0);
117                 kfree(path);
118         }
119 }
120
121 /*
122  * All the elements of src into dest. After this call, src could be freed
123  * without affecting dest.
124  *
125  * Both paths should have the same root. Any non-root elements of dest
126  * will be freed.
127  */
128 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
129 {
130         int i;
131
132         BUG_ON(path_root_bh(dest) != path_root_bh(src));
133         BUG_ON(path_root_el(dest) != path_root_el(src));
134
135         ocfs2_reinit_path(dest, 1);
136
137         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
138                 dest->p_node[i].bh = src->p_node[i].bh;
139                 dest->p_node[i].el = src->p_node[i].el;
140
141                 if (dest->p_node[i].bh)
142                         get_bh(dest->p_node[i].bh);
143         }
144 }
145
146 /*
147  * Make the *dest path the same as src and re-initialize src path to
148  * have a root only.
149  */
150 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
151 {
152         int i;
153
154         BUG_ON(path_root_bh(dest) != path_root_bh(src));
155
156         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
157                 brelse(dest->p_node[i].bh);
158
159                 dest->p_node[i].bh = src->p_node[i].bh;
160                 dest->p_node[i].el = src->p_node[i].el;
161
162                 src->p_node[i].bh = NULL;
163                 src->p_node[i].el = NULL;
164         }
165 }
166
167 /*
168  * Insert an extent block at given index.
169  *
170  * This will not take an additional reference on eb_bh.
171  */
172 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
173                                         struct buffer_head *eb_bh)
174 {
175         struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
176
177         /*
178          * Right now, no root bh is an extent block, so this helps
179          * catch code errors with dinode trees. The assertion can be
180          * safely removed if we ever need to insert extent block
181          * structures at the root.
182          */
183         BUG_ON(index == 0);
184
185         path->p_node[index].bh = eb_bh;
186         path->p_node[index].el = &eb->h_list;
187 }
188
189 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
190                                          struct ocfs2_extent_list *root_el)
191 {
192         struct ocfs2_path *path;
193
194         BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
195
196         path = kzalloc(sizeof(*path), GFP_NOFS);
197         if (path) {
198                 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
199                 get_bh(root_bh);
200                 path_root_bh(path) = root_bh;
201                 path_root_el(path) = root_el;
202         }
203
204         return path;
205 }
206
207 /*
208  * Allocate and initialize a new path based on a disk inode tree.
209  */
210 static struct ocfs2_path *ocfs2_new_inode_path(struct buffer_head *di_bh)
211 {
212         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
213         struct ocfs2_extent_list *el = &di->id2.i_list;
214
215         return ocfs2_new_path(di_bh, el);
216 }
217
218 /*
219  * Convenience function to journal all components in a path.
220  */
221 static int ocfs2_journal_access_path(struct inode *inode, handle_t *handle,
222                                      struct ocfs2_path *path)
223 {
224         int i, ret = 0;
225
226         if (!path)
227                 goto out;
228
229         for(i = 0; i < path_num_items(path); i++) {
230                 ret = ocfs2_journal_access(handle, inode, path->p_node[i].bh,
231                                            OCFS2_JOURNAL_ACCESS_WRITE);
232                 if (ret < 0) {
233                         mlog_errno(ret);
234                         goto out;
235                 }
236         }
237
238 out:
239         return ret;
240 }
241
242 /*
243  * Return the index of the extent record which contains cluster #v_cluster.
244  * -1 is returned if it was not found.
245  *
246  * Should work fine on interior and exterior nodes.
247  */
248 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
249 {
250         int ret = -1;
251         int i;
252         struct ocfs2_extent_rec *rec;
253         u32 rec_end, rec_start, clusters;
254
255         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
256                 rec = &el->l_recs[i];
257
258                 rec_start = le32_to_cpu(rec->e_cpos);
259                 clusters = ocfs2_rec_clusters(el, rec);
260
261                 rec_end = rec_start + clusters;
262
263                 if (v_cluster >= rec_start && v_cluster < rec_end) {
264                         ret = i;
265                         break;
266                 }
267         }
268
269         return ret;
270 }
271
272 enum ocfs2_contig_type {
273         CONTIG_NONE = 0,
274         CONTIG_LEFT,
275         CONTIG_RIGHT,
276         CONTIG_LEFTRIGHT,
277 };
278
279
280 /*
281  * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
282  * ocfs2_extent_contig only work properly against leaf nodes!
283  */
284 static int ocfs2_block_extent_contig(struct super_block *sb,
285                                      struct ocfs2_extent_rec *ext,
286                                      u64 blkno)
287 {
288         u64 blk_end = le64_to_cpu(ext->e_blkno);
289
290         blk_end += ocfs2_clusters_to_blocks(sb,
291                                     le16_to_cpu(ext->e_leaf_clusters));
292
293         return blkno == blk_end;
294 }
295
296 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
297                                   struct ocfs2_extent_rec *right)
298 {
299         u32 left_range;
300
301         left_range = le32_to_cpu(left->e_cpos) +
302                 le16_to_cpu(left->e_leaf_clusters);
303
304         return (left_range == le32_to_cpu(right->e_cpos));
305 }
306
307 static enum ocfs2_contig_type
308         ocfs2_extent_contig(struct inode *inode,
309                             struct ocfs2_extent_rec *ext,
310                             struct ocfs2_extent_rec *insert_rec)
311 {
312         u64 blkno = le64_to_cpu(insert_rec->e_blkno);
313
314         /*
315          * Refuse to coalesce extent records with different flag
316          * fields - we don't want to mix unwritten extents with user
317          * data.
318          */
319         if (ext->e_flags != insert_rec->e_flags)
320                 return CONTIG_NONE;
321
322         if (ocfs2_extents_adjacent(ext, insert_rec) &&
323             ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
324                         return CONTIG_RIGHT;
325
326         blkno = le64_to_cpu(ext->e_blkno);
327         if (ocfs2_extents_adjacent(insert_rec, ext) &&
328             ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno))
329                 return CONTIG_LEFT;
330
331         return CONTIG_NONE;
332 }
333
334 /*
335  * NOTE: We can have pretty much any combination of contiguousness and
336  * appending.
337  *
338  * The usefulness of APPEND_TAIL is more in that it lets us know that
339  * we'll have to update the path to that leaf.
340  */
341 enum ocfs2_append_type {
342         APPEND_NONE = 0,
343         APPEND_TAIL,
344 };
345
346 enum ocfs2_split_type {
347         SPLIT_NONE = 0,
348         SPLIT_LEFT,
349         SPLIT_RIGHT,
350 };
351
352 struct ocfs2_insert_type {
353         enum ocfs2_split_type   ins_split;
354         enum ocfs2_append_type  ins_appending;
355         enum ocfs2_contig_type  ins_contig;
356         int                     ins_contig_index;
357         int                     ins_tree_depth;
358 };
359
360 struct ocfs2_merge_ctxt {
361         enum ocfs2_contig_type  c_contig_type;
362         int                     c_has_empty_extent;
363         int                     c_split_covers_rec;
364 };
365
366 /*
367  * How many free extents have we got before we need more meta data?
368  */
369 int ocfs2_num_free_extents(struct ocfs2_super *osb,
370                            struct inode *inode,
371                            struct ocfs2_dinode *fe)
372 {
373         int retval;
374         struct ocfs2_extent_list *el;
375         struct ocfs2_extent_block *eb;
376         struct buffer_head *eb_bh = NULL;
377
378         mlog_entry_void();
379
380         if (!OCFS2_IS_VALID_DINODE(fe)) {
381                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
382                 retval = -EIO;
383                 goto bail;
384         }
385
386         if (fe->i_last_eb_blk) {
387                 retval = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
388                                           &eb_bh, OCFS2_BH_CACHED, inode);
389                 if (retval < 0) {
390                         mlog_errno(retval);
391                         goto bail;
392                 }
393                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
394                 el = &eb->h_list;
395         } else
396                 el = &fe->id2.i_list;
397
398         BUG_ON(el->l_tree_depth != 0);
399
400         retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
401 bail:
402         if (eb_bh)
403                 brelse(eb_bh);
404
405         mlog_exit(retval);
406         return retval;
407 }
408
409 /* expects array to already be allocated
410  *
411  * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
412  * l_count for you
413  */
414 static int ocfs2_create_new_meta_bhs(struct ocfs2_super *osb,
415                                      handle_t *handle,
416                                      struct inode *inode,
417                                      int wanted,
418                                      struct ocfs2_alloc_context *meta_ac,
419                                      struct buffer_head *bhs[])
420 {
421         int count, status, i;
422         u16 suballoc_bit_start;
423         u32 num_got;
424         u64 first_blkno;
425         struct ocfs2_extent_block *eb;
426
427         mlog_entry_void();
428
429         count = 0;
430         while (count < wanted) {
431                 status = ocfs2_claim_metadata(osb,
432                                               handle,
433                                               meta_ac,
434                                               wanted - count,
435                                               &suballoc_bit_start,
436                                               &num_got,
437                                               &first_blkno);
438                 if (status < 0) {
439                         mlog_errno(status);
440                         goto bail;
441                 }
442
443                 for(i = count;  i < (num_got + count); i++) {
444                         bhs[i] = sb_getblk(osb->sb, first_blkno);
445                         if (bhs[i] == NULL) {
446                                 status = -EIO;
447                                 mlog_errno(status);
448                                 goto bail;
449                         }
450                         ocfs2_set_new_buffer_uptodate(inode, bhs[i]);
451
452                         status = ocfs2_journal_access(handle, inode, bhs[i],
453                                                       OCFS2_JOURNAL_ACCESS_CREATE);
454                         if (status < 0) {
455                                 mlog_errno(status);
456                                 goto bail;
457                         }
458
459                         memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
460                         eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
461                         /* Ok, setup the minimal stuff here. */
462                         strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
463                         eb->h_blkno = cpu_to_le64(first_blkno);
464                         eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
465                         eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
466                         eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
467                         eb->h_list.l_count =
468                                 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
469
470                         suballoc_bit_start++;
471                         first_blkno++;
472
473                         /* We'll also be dirtied by the caller, so
474                          * this isn't absolutely necessary. */
475                         status = ocfs2_journal_dirty(handle, bhs[i]);
476                         if (status < 0) {
477                                 mlog_errno(status);
478                                 goto bail;
479                         }
480                 }
481
482                 count += num_got;
483         }
484
485         status = 0;
486 bail:
487         if (status < 0) {
488                 for(i = 0; i < wanted; i++) {
489                         if (bhs[i])
490                                 brelse(bhs[i]);
491                         bhs[i] = NULL;
492                 }
493         }
494         mlog_exit(status);
495         return status;
496 }
497
498 /*
499  * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
500  *
501  * Returns the sum of the rightmost extent rec logical offset and
502  * cluster count.
503  *
504  * ocfs2_add_branch() uses this to determine what logical cluster
505  * value should be populated into the leftmost new branch records.
506  *
507  * ocfs2_shift_tree_depth() uses this to determine the # clusters
508  * value for the new topmost tree record.
509  */
510 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
511 {
512         int i;
513
514         i = le16_to_cpu(el->l_next_free_rec) - 1;
515
516         return le32_to_cpu(el->l_recs[i].e_cpos) +
517                 ocfs2_rec_clusters(el, &el->l_recs[i]);
518 }
519
520 /*
521  * Add an entire tree branch to our inode. eb_bh is the extent block
522  * to start at, if we don't want to start the branch at the dinode
523  * structure.
524  *
525  * last_eb_bh is required as we have to update it's next_leaf pointer
526  * for the new last extent block.
527  *
528  * the new branch will be 'empty' in the sense that every block will
529  * contain a single record with cluster count == 0.
530  */
531 static int ocfs2_add_branch(struct ocfs2_super *osb,
532                             handle_t *handle,
533                             struct inode *inode,
534                             struct buffer_head *fe_bh,
535                             struct buffer_head *eb_bh,
536                             struct buffer_head **last_eb_bh,
537                             struct ocfs2_alloc_context *meta_ac)
538 {
539         int status, new_blocks, i;
540         u64 next_blkno, new_last_eb_blk;
541         struct buffer_head *bh;
542         struct buffer_head **new_eb_bhs = NULL;
543         struct ocfs2_dinode *fe;
544         struct ocfs2_extent_block *eb;
545         struct ocfs2_extent_list  *eb_el;
546         struct ocfs2_extent_list  *el;
547         u32 new_cpos;
548
549         mlog_entry_void();
550
551         BUG_ON(!last_eb_bh || !*last_eb_bh);
552
553         fe = (struct ocfs2_dinode *) fe_bh->b_data;
554
555         if (eb_bh) {
556                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
557                 el = &eb->h_list;
558         } else
559                 el = &fe->id2.i_list;
560
561         /* we never add a branch to a leaf. */
562         BUG_ON(!el->l_tree_depth);
563
564         new_blocks = le16_to_cpu(el->l_tree_depth);
565
566         /* allocate the number of new eb blocks we need */
567         new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
568                              GFP_KERNEL);
569         if (!new_eb_bhs) {
570                 status = -ENOMEM;
571                 mlog_errno(status);
572                 goto bail;
573         }
574
575         status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
576                                            meta_ac, new_eb_bhs);
577         if (status < 0) {
578                 mlog_errno(status);
579                 goto bail;
580         }
581
582         eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
583         new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
584
585         /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
586          * linked with the rest of the tree.
587          * conversly, new_eb_bhs[0] is the new bottommost leaf.
588          *
589          * when we leave the loop, new_last_eb_blk will point to the
590          * newest leaf, and next_blkno will point to the topmost extent
591          * block. */
592         next_blkno = new_last_eb_blk = 0;
593         for(i = 0; i < new_blocks; i++) {
594                 bh = new_eb_bhs[i];
595                 eb = (struct ocfs2_extent_block *) bh->b_data;
596                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
597                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
598                         status = -EIO;
599                         goto bail;
600                 }
601                 eb_el = &eb->h_list;
602
603                 status = ocfs2_journal_access(handle, inode, bh,
604                                               OCFS2_JOURNAL_ACCESS_CREATE);
605                 if (status < 0) {
606                         mlog_errno(status);
607                         goto bail;
608                 }
609
610                 eb->h_next_leaf_blk = 0;
611                 eb_el->l_tree_depth = cpu_to_le16(i);
612                 eb_el->l_next_free_rec = cpu_to_le16(1);
613                 /*
614                  * This actually counts as an empty extent as
615                  * c_clusters == 0
616                  */
617                 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
618                 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
619                 /*
620                  * eb_el isn't always an interior node, but even leaf
621                  * nodes want a zero'd flags and reserved field so
622                  * this gets the whole 32 bits regardless of use.
623                  */
624                 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
625                 if (!eb_el->l_tree_depth)
626                         new_last_eb_blk = le64_to_cpu(eb->h_blkno);
627
628                 status = ocfs2_journal_dirty(handle, bh);
629                 if (status < 0) {
630                         mlog_errno(status);
631                         goto bail;
632                 }
633
634                 next_blkno = le64_to_cpu(eb->h_blkno);
635         }
636
637         /* This is a bit hairy. We want to update up to three blocks
638          * here without leaving any of them in an inconsistent state
639          * in case of error. We don't have to worry about
640          * journal_dirty erroring as it won't unless we've aborted the
641          * handle (in which case we would never be here) so reserving
642          * the write with journal_access is all we need to do. */
643         status = ocfs2_journal_access(handle, inode, *last_eb_bh,
644                                       OCFS2_JOURNAL_ACCESS_WRITE);
645         if (status < 0) {
646                 mlog_errno(status);
647                 goto bail;
648         }
649         status = ocfs2_journal_access(handle, inode, fe_bh,
650                                       OCFS2_JOURNAL_ACCESS_WRITE);
651         if (status < 0) {
652                 mlog_errno(status);
653                 goto bail;
654         }
655         if (eb_bh) {
656                 status = ocfs2_journal_access(handle, inode, eb_bh,
657                                               OCFS2_JOURNAL_ACCESS_WRITE);
658                 if (status < 0) {
659                         mlog_errno(status);
660                         goto bail;
661                 }
662         }
663
664         /* Link the new branch into the rest of the tree (el will
665          * either be on the fe, or the extent block passed in. */
666         i = le16_to_cpu(el->l_next_free_rec);
667         el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
668         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
669         el->l_recs[i].e_int_clusters = 0;
670         le16_add_cpu(&el->l_next_free_rec, 1);
671
672         /* fe needs a new last extent block pointer, as does the
673          * next_leaf on the previously last-extent-block. */
674         fe->i_last_eb_blk = cpu_to_le64(new_last_eb_blk);
675
676         eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
677         eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
678
679         status = ocfs2_journal_dirty(handle, *last_eb_bh);
680         if (status < 0)
681                 mlog_errno(status);
682         status = ocfs2_journal_dirty(handle, fe_bh);
683         if (status < 0)
684                 mlog_errno(status);
685         if (eb_bh) {
686                 status = ocfs2_journal_dirty(handle, eb_bh);
687                 if (status < 0)
688                         mlog_errno(status);
689         }
690
691         /*
692          * Some callers want to track the rightmost leaf so pass it
693          * back here.
694          */
695         brelse(*last_eb_bh);
696         get_bh(new_eb_bhs[0]);
697         *last_eb_bh = new_eb_bhs[0];
698
699         status = 0;
700 bail:
701         if (new_eb_bhs) {
702                 for (i = 0; i < new_blocks; i++)
703                         if (new_eb_bhs[i])
704                                 brelse(new_eb_bhs[i]);
705                 kfree(new_eb_bhs);
706         }
707
708         mlog_exit(status);
709         return status;
710 }
711
712 /*
713  * adds another level to the allocation tree.
714  * returns back the new extent block so you can add a branch to it
715  * after this call.
716  */
717 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
718                                   handle_t *handle,
719                                   struct inode *inode,
720                                   struct buffer_head *fe_bh,
721                                   struct ocfs2_alloc_context *meta_ac,
722                                   struct buffer_head **ret_new_eb_bh)
723 {
724         int status, i;
725         u32 new_clusters;
726         struct buffer_head *new_eb_bh = NULL;
727         struct ocfs2_dinode *fe;
728         struct ocfs2_extent_block *eb;
729         struct ocfs2_extent_list  *fe_el;
730         struct ocfs2_extent_list  *eb_el;
731
732         mlog_entry_void();
733
734         status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
735                                            &new_eb_bh);
736         if (status < 0) {
737                 mlog_errno(status);
738                 goto bail;
739         }
740
741         eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
742         if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
743                 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
744                 status = -EIO;
745                 goto bail;
746         }
747
748         eb_el = &eb->h_list;
749         fe = (struct ocfs2_dinode *) fe_bh->b_data;
750         fe_el = &fe->id2.i_list;
751
752         status = ocfs2_journal_access(handle, inode, new_eb_bh,
753                                       OCFS2_JOURNAL_ACCESS_CREATE);
754         if (status < 0) {
755                 mlog_errno(status);
756                 goto bail;
757         }
758
759         /* copy the fe data into the new extent block */
760         eb_el->l_tree_depth = fe_el->l_tree_depth;
761         eb_el->l_next_free_rec = fe_el->l_next_free_rec;
762         for(i = 0; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
763                 eb_el->l_recs[i] = fe_el->l_recs[i];
764
765         status = ocfs2_journal_dirty(handle, new_eb_bh);
766         if (status < 0) {
767                 mlog_errno(status);
768                 goto bail;
769         }
770
771         status = ocfs2_journal_access(handle, inode, fe_bh,
772                                       OCFS2_JOURNAL_ACCESS_WRITE);
773         if (status < 0) {
774                 mlog_errno(status);
775                 goto bail;
776         }
777
778         new_clusters = ocfs2_sum_rightmost_rec(eb_el);
779
780         /* update fe now */
781         le16_add_cpu(&fe_el->l_tree_depth, 1);
782         fe_el->l_recs[0].e_cpos = 0;
783         fe_el->l_recs[0].e_blkno = eb->h_blkno;
784         fe_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
785         for(i = 1; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
786                 memset(&fe_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
787         fe_el->l_next_free_rec = cpu_to_le16(1);
788
789         /* If this is our 1st tree depth shift, then last_eb_blk
790          * becomes the allocated extent block */
791         if (fe_el->l_tree_depth == cpu_to_le16(1))
792                 fe->i_last_eb_blk = eb->h_blkno;
793
794         status = ocfs2_journal_dirty(handle, fe_bh);
795         if (status < 0) {
796                 mlog_errno(status);
797                 goto bail;
798         }
799
800         *ret_new_eb_bh = new_eb_bh;
801         new_eb_bh = NULL;
802         status = 0;
803 bail:
804         if (new_eb_bh)
805                 brelse(new_eb_bh);
806
807         mlog_exit(status);
808         return status;
809 }
810
811 /*
812  * Should only be called when there is no space left in any of the
813  * leaf nodes. What we want to do is find the lowest tree depth
814  * non-leaf extent block with room for new records. There are three
815  * valid results of this search:
816  *
817  * 1) a lowest extent block is found, then we pass it back in
818  *    *lowest_eb_bh and return '0'
819  *
820  * 2) the search fails to find anything, but the dinode has room. We
821  *    pass NULL back in *lowest_eb_bh, but still return '0'
822  *
823  * 3) the search fails to find anything AND the dinode is full, in
824  *    which case we return > 0
825  *
826  * return status < 0 indicates an error.
827  */
828 static int ocfs2_find_branch_target(struct ocfs2_super *osb,
829                                     struct inode *inode,
830                                     struct buffer_head *fe_bh,
831                                     struct buffer_head **target_bh)
832 {
833         int status = 0, i;
834         u64 blkno;
835         struct ocfs2_dinode *fe;
836         struct ocfs2_extent_block *eb;
837         struct ocfs2_extent_list  *el;
838         struct buffer_head *bh = NULL;
839         struct buffer_head *lowest_bh = NULL;
840
841         mlog_entry_void();
842
843         *target_bh = NULL;
844
845         fe = (struct ocfs2_dinode *) fe_bh->b_data;
846         el = &fe->id2.i_list;
847
848         while(le16_to_cpu(el->l_tree_depth) > 1) {
849                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
850                         ocfs2_error(inode->i_sb, "Dinode %llu has empty "
851                                     "extent list (next_free_rec == 0)",
852                                     (unsigned long long)OCFS2_I(inode)->ip_blkno);
853                         status = -EIO;
854                         goto bail;
855                 }
856                 i = le16_to_cpu(el->l_next_free_rec) - 1;
857                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
858                 if (!blkno) {
859                         ocfs2_error(inode->i_sb, "Dinode %llu has extent "
860                                     "list where extent # %d has no physical "
861                                     "block start",
862                                     (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
863                         status = -EIO;
864                         goto bail;
865                 }
866
867                 if (bh) {
868                         brelse(bh);
869                         bh = NULL;
870                 }
871
872                 status = ocfs2_read_block(osb, blkno, &bh, OCFS2_BH_CACHED,
873                                           inode);
874                 if (status < 0) {
875                         mlog_errno(status);
876                         goto bail;
877                 }
878
879                 eb = (struct ocfs2_extent_block *) bh->b_data;
880                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
881                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
882                         status = -EIO;
883                         goto bail;
884                 }
885                 el = &eb->h_list;
886
887                 if (le16_to_cpu(el->l_next_free_rec) <
888                     le16_to_cpu(el->l_count)) {
889                         if (lowest_bh)
890                                 brelse(lowest_bh);
891                         lowest_bh = bh;
892                         get_bh(lowest_bh);
893                 }
894         }
895
896         /* If we didn't find one and the fe doesn't have any room,
897          * then return '1' */
898         if (!lowest_bh
899             && (fe->id2.i_list.l_next_free_rec == fe->id2.i_list.l_count))
900                 status = 1;
901
902         *target_bh = lowest_bh;
903 bail:
904         if (bh)
905                 brelse(bh);
906
907         mlog_exit(status);
908         return status;
909 }
910
911 /*
912  * Grow a b-tree so that it has more records.
913  *
914  * We might shift the tree depth in which case existing paths should
915  * be considered invalid.
916  *
917  * Tree depth after the grow is returned via *final_depth.
918  *
919  * *last_eb_bh will be updated by ocfs2_add_branch().
920  */
921 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
922                            struct buffer_head *di_bh, int *final_depth,
923                            struct buffer_head **last_eb_bh,
924                            struct ocfs2_alloc_context *meta_ac)
925 {
926         int ret, shift;
927         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
928         int depth = le16_to_cpu(di->id2.i_list.l_tree_depth);
929         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
930         struct buffer_head *bh = NULL;
931
932         BUG_ON(meta_ac == NULL);
933
934         shift = ocfs2_find_branch_target(osb, inode, di_bh, &bh);
935         if (shift < 0) {
936                 ret = shift;
937                 mlog_errno(ret);
938                 goto out;
939         }
940
941         /* We traveled all the way to the bottom of the allocation tree
942          * and didn't find room for any more extents - we need to add
943          * another tree level */
944         if (shift) {
945                 BUG_ON(bh);
946                 mlog(0, "need to shift tree depth (current = %d)\n", depth);
947
948                 /* ocfs2_shift_tree_depth will return us a buffer with
949                  * the new extent block (so we can pass that to
950                  * ocfs2_add_branch). */
951                 ret = ocfs2_shift_tree_depth(osb, handle, inode, di_bh,
952                                              meta_ac, &bh);
953                 if (ret < 0) {
954                         mlog_errno(ret);
955                         goto out;
956                 }
957                 depth++;
958                 if (depth == 1) {
959                         /*
960                          * Special case: we have room now if we shifted from
961                          * tree_depth 0, so no more work needs to be done.
962                          *
963                          * We won't be calling add_branch, so pass
964                          * back *last_eb_bh as the new leaf. At depth
965                          * zero, it should always be null so there's
966                          * no reason to brelse.
967                          */
968                         BUG_ON(*last_eb_bh);
969                         get_bh(bh);
970                         *last_eb_bh = bh;
971                         goto out;
972                 }
973         }
974
975         /* call ocfs2_add_branch to add the final part of the tree with
976          * the new data. */
977         mlog(0, "add branch. bh = %p\n", bh);
978         ret = ocfs2_add_branch(osb, handle, inode, di_bh, bh, last_eb_bh,
979                                meta_ac);
980         if (ret < 0) {
981                 mlog_errno(ret);
982                 goto out;
983         }
984
985 out:
986         if (final_depth)
987                 *final_depth = depth;
988         brelse(bh);
989         return ret;
990 }
991
992 /*
993  * This is only valid for leaf nodes, which are the only ones that can
994  * have empty extents anyway.
995  */
996 static inline int ocfs2_is_empty_extent(struct ocfs2_extent_rec *rec)
997 {
998         return !rec->e_leaf_clusters;
999 }
1000
1001 /*
1002  * This function will discard the rightmost extent record.
1003  */
1004 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1005 {
1006         int next_free = le16_to_cpu(el->l_next_free_rec);
1007         int count = le16_to_cpu(el->l_count);
1008         unsigned int num_bytes;
1009
1010         BUG_ON(!next_free);
1011         /* This will cause us to go off the end of our extent list. */
1012         BUG_ON(next_free >= count);
1013
1014         num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1015
1016         memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1017 }
1018
1019 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1020                               struct ocfs2_extent_rec *insert_rec)
1021 {
1022         int i, insert_index, next_free, has_empty, num_bytes;
1023         u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1024         struct ocfs2_extent_rec *rec;
1025
1026         next_free = le16_to_cpu(el->l_next_free_rec);
1027         has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1028
1029         BUG_ON(!next_free);
1030
1031         /* The tree code before us didn't allow enough room in the leaf. */
1032         if (el->l_next_free_rec == el->l_count && !has_empty)
1033                 BUG();
1034
1035         /*
1036          * The easiest way to approach this is to just remove the
1037          * empty extent and temporarily decrement next_free.
1038          */
1039         if (has_empty) {
1040                 /*
1041                  * If next_free was 1 (only an empty extent), this
1042                  * loop won't execute, which is fine. We still want
1043                  * the decrement above to happen.
1044                  */
1045                 for(i = 0; i < (next_free - 1); i++)
1046                         el->l_recs[i] = el->l_recs[i+1];
1047
1048                 next_free--;
1049         }
1050
1051         /*
1052          * Figure out what the new record index should be.
1053          */
1054         for(i = 0; i < next_free; i++) {
1055                 rec = &el->l_recs[i];
1056
1057                 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1058                         break;
1059         }
1060         insert_index = i;
1061
1062         mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
1063              insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
1064
1065         BUG_ON(insert_index < 0);
1066         BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1067         BUG_ON(insert_index > next_free);
1068
1069         /*
1070          * No need to memmove if we're just adding to the tail.
1071          */
1072         if (insert_index != next_free) {
1073                 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1074
1075                 num_bytes = next_free - insert_index;
1076                 num_bytes *= sizeof(struct ocfs2_extent_rec);
1077                 memmove(&el->l_recs[insert_index + 1],
1078                         &el->l_recs[insert_index],
1079                         num_bytes);
1080         }
1081
1082         /*
1083          * Either we had an empty extent, and need to re-increment or
1084          * there was no empty extent on a non full rightmost leaf node,
1085          * in which case we still need to increment.
1086          */
1087         next_free++;
1088         el->l_next_free_rec = cpu_to_le16(next_free);
1089         /*
1090          * Make sure none of the math above just messed up our tree.
1091          */
1092         BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1093
1094         el->l_recs[insert_index] = *insert_rec;
1095
1096 }
1097
1098 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1099 {
1100         int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1101
1102         BUG_ON(num_recs == 0);
1103
1104         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1105                 num_recs--;
1106                 size = num_recs * sizeof(struct ocfs2_extent_rec);
1107                 memmove(&el->l_recs[0], &el->l_recs[1], size);
1108                 memset(&el->l_recs[num_recs], 0,
1109                        sizeof(struct ocfs2_extent_rec));
1110                 el->l_next_free_rec = cpu_to_le16(num_recs);
1111         }
1112 }
1113
1114 /*
1115  * Create an empty extent record .
1116  *
1117  * l_next_free_rec may be updated.
1118  *
1119  * If an empty extent already exists do nothing.
1120  */
1121 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1122 {
1123         int next_free = le16_to_cpu(el->l_next_free_rec);
1124
1125         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1126
1127         if (next_free == 0)
1128                 goto set_and_inc;
1129
1130         if (ocfs2_is_empty_extent(&el->l_recs[0]))
1131                 return;
1132
1133         mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1134                         "Asked to create an empty extent in a full list:\n"
1135                         "count = %u, tree depth = %u",
1136                         le16_to_cpu(el->l_count),
1137                         le16_to_cpu(el->l_tree_depth));
1138
1139         ocfs2_shift_records_right(el);
1140
1141 set_and_inc:
1142         le16_add_cpu(&el->l_next_free_rec, 1);
1143         memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1144 }
1145
1146 /*
1147  * For a rotation which involves two leaf nodes, the "root node" is
1148  * the lowest level tree node which contains a path to both leafs. This
1149  * resulting set of information can be used to form a complete "subtree"
1150  *
1151  * This function is passed two full paths from the dinode down to a
1152  * pair of adjacent leaves. It's task is to figure out which path
1153  * index contains the subtree root - this can be the root index itself
1154  * in a worst-case rotation.
1155  *
1156  * The array index of the subtree root is passed back.
1157  */
1158 static int ocfs2_find_subtree_root(struct inode *inode,
1159                                    struct ocfs2_path *left,
1160                                    struct ocfs2_path *right)
1161 {
1162         int i = 0;
1163
1164         /*
1165          * Check that the caller passed in two paths from the same tree.
1166          */
1167         BUG_ON(path_root_bh(left) != path_root_bh(right));
1168
1169         do {
1170                 i++;
1171
1172                 /*
1173                  * The caller didn't pass two adjacent paths.
1174                  */
1175                 mlog_bug_on_msg(i > left->p_tree_depth,
1176                                 "Inode %lu, left depth %u, right depth %u\n"
1177                                 "left leaf blk %llu, right leaf blk %llu\n",
1178                                 inode->i_ino, left->p_tree_depth,
1179                                 right->p_tree_depth,
1180                                 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1181                                 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1182         } while (left->p_node[i].bh->b_blocknr ==
1183                  right->p_node[i].bh->b_blocknr);
1184
1185         return i - 1;
1186 }
1187
1188 typedef void (path_insert_t)(void *, struct buffer_head *);
1189
1190 /*
1191  * Traverse a btree path in search of cpos, starting at root_el.
1192  *
1193  * This code can be called with a cpos larger than the tree, in which
1194  * case it will return the rightmost path.
1195  */
1196 static int __ocfs2_find_path(struct inode *inode,
1197                              struct ocfs2_extent_list *root_el, u32 cpos,
1198                              path_insert_t *func, void *data)
1199 {
1200         int i, ret = 0;
1201         u32 range;
1202         u64 blkno;
1203         struct buffer_head *bh = NULL;
1204         struct ocfs2_extent_block *eb;
1205         struct ocfs2_extent_list *el;
1206         struct ocfs2_extent_rec *rec;
1207         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1208
1209         el = root_el;
1210         while (el->l_tree_depth) {
1211                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1212                         ocfs2_error(inode->i_sb,
1213                                     "Inode %llu has empty extent list at "
1214                                     "depth %u\n",
1215                                     (unsigned long long)oi->ip_blkno,
1216                                     le16_to_cpu(el->l_tree_depth));
1217                         ret = -EROFS;
1218                         goto out;
1219
1220                 }
1221
1222                 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1223                         rec = &el->l_recs[i];
1224
1225                         /*
1226                          * In the case that cpos is off the allocation
1227                          * tree, this should just wind up returning the
1228                          * rightmost record.
1229                          */
1230                         range = le32_to_cpu(rec->e_cpos) +
1231                                 ocfs2_rec_clusters(el, rec);
1232                         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1233                             break;
1234                 }
1235
1236                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1237                 if (blkno == 0) {
1238                         ocfs2_error(inode->i_sb,
1239                                     "Inode %llu has bad blkno in extent list "
1240                                     "at depth %u (index %d)\n",
1241                                     (unsigned long long)oi->ip_blkno,
1242                                     le16_to_cpu(el->l_tree_depth), i);
1243                         ret = -EROFS;
1244                         goto out;
1245                 }
1246
1247                 brelse(bh);
1248                 bh = NULL;
1249                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno,
1250                                        &bh, OCFS2_BH_CACHED, inode);
1251                 if (ret) {
1252                         mlog_errno(ret);
1253                         goto out;
1254                 }
1255
1256                 eb = (struct ocfs2_extent_block *) bh->b_data;
1257                 el = &eb->h_list;
1258                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1259                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1260                         ret = -EIO;
1261                         goto out;
1262                 }
1263
1264                 if (le16_to_cpu(el->l_next_free_rec) >
1265                     le16_to_cpu(el->l_count)) {
1266                         ocfs2_error(inode->i_sb,
1267                                     "Inode %llu has bad count in extent list "
1268                                     "at block %llu (next free=%u, count=%u)\n",
1269                                     (unsigned long long)oi->ip_blkno,
1270                                     (unsigned long long)bh->b_blocknr,
1271                                     le16_to_cpu(el->l_next_free_rec),
1272                                     le16_to_cpu(el->l_count));
1273                         ret = -EROFS;
1274                         goto out;
1275                 }
1276
1277                 if (func)
1278                         func(data, bh);
1279         }
1280
1281 out:
1282         /*
1283          * Catch any trailing bh that the loop didn't handle.
1284          */
1285         brelse(bh);
1286
1287         return ret;
1288 }
1289
1290 /*
1291  * Given an initialized path (that is, it has a valid root extent
1292  * list), this function will traverse the btree in search of the path
1293  * which would contain cpos.
1294  *
1295  * The path traveled is recorded in the path structure.
1296  *
1297  * Note that this will not do any comparisons on leaf node extent
1298  * records, so it will work fine in the case that we just added a tree
1299  * branch.
1300  */
1301 struct find_path_data {
1302         int index;
1303         struct ocfs2_path *path;
1304 };
1305 static void find_path_ins(void *data, struct buffer_head *bh)
1306 {
1307         struct find_path_data *fp = data;
1308
1309         get_bh(bh);
1310         ocfs2_path_insert_eb(fp->path, fp->index, bh);
1311         fp->index++;
1312 }
1313 static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
1314                            u32 cpos)
1315 {
1316         struct find_path_data data;
1317
1318         data.index = 1;
1319         data.path = path;
1320         return __ocfs2_find_path(inode, path_root_el(path), cpos,
1321                                  find_path_ins, &data);
1322 }
1323
1324 static void find_leaf_ins(void *data, struct buffer_head *bh)
1325 {
1326         struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1327         struct ocfs2_extent_list *el = &eb->h_list;
1328         struct buffer_head **ret = data;
1329
1330         /* We want to retain only the leaf block. */
1331         if (le16_to_cpu(el->l_tree_depth) == 0) {
1332                 get_bh(bh);
1333                 *ret = bh;
1334         }
1335 }
1336 /*
1337  * Find the leaf block in the tree which would contain cpos. No
1338  * checking of the actual leaf is done.
1339  *
1340  * Some paths want to call this instead of allocating a path structure
1341  * and calling ocfs2_find_path().
1342  *
1343  * This function doesn't handle non btree extent lists.
1344  */
1345 int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
1346                     u32 cpos, struct buffer_head **leaf_bh)
1347 {
1348         int ret;
1349         struct buffer_head *bh = NULL;
1350
1351         ret = __ocfs2_find_path(inode, root_el, cpos, find_leaf_ins, &bh);
1352         if (ret) {
1353                 mlog_errno(ret);
1354                 goto out;
1355         }
1356
1357         *leaf_bh = bh;
1358 out:
1359         return ret;
1360 }
1361
1362 /*
1363  * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1364  *
1365  * Basically, we've moved stuff around at the bottom of the tree and
1366  * we need to fix up the extent records above the changes to reflect
1367  * the new changes.
1368  *
1369  * left_rec: the record on the left.
1370  * left_child_el: is the child list pointed to by left_rec
1371  * right_rec: the record to the right of left_rec
1372  * right_child_el: is the child list pointed to by right_rec
1373  *
1374  * By definition, this only works on interior nodes.
1375  */
1376 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1377                                   struct ocfs2_extent_list *left_child_el,
1378                                   struct ocfs2_extent_rec *right_rec,
1379                                   struct ocfs2_extent_list *right_child_el)
1380 {
1381         u32 left_clusters, right_end;
1382
1383         /*
1384          * Interior nodes never have holes. Their cpos is the cpos of
1385          * the leftmost record in their child list. Their cluster
1386          * count covers the full theoretical range of their child list
1387          * - the range between their cpos and the cpos of the record
1388          * immediately to their right.
1389          */
1390         left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1391         if (ocfs2_is_empty_extent(&right_child_el->l_recs[0])) {
1392                 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1393                 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1394         }
1395         left_clusters -= le32_to_cpu(left_rec->e_cpos);
1396         left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1397
1398         /*
1399          * Calculate the rightmost cluster count boundary before
1400          * moving cpos - we will need to adjust clusters after
1401          * updating e_cpos to keep the same highest cluster count.
1402          */
1403         right_end = le32_to_cpu(right_rec->e_cpos);
1404         right_end += le32_to_cpu(right_rec->e_int_clusters);
1405
1406         right_rec->e_cpos = left_rec->e_cpos;
1407         le32_add_cpu(&right_rec->e_cpos, left_clusters);
1408
1409         right_end -= le32_to_cpu(right_rec->e_cpos);
1410         right_rec->e_int_clusters = cpu_to_le32(right_end);
1411 }
1412
1413 /*
1414  * Adjust the adjacent root node records involved in a
1415  * rotation. left_el_blkno is passed in as a key so that we can easily
1416  * find it's index in the root list.
1417  */
1418 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1419                                       struct ocfs2_extent_list *left_el,
1420                                       struct ocfs2_extent_list *right_el,
1421                                       u64 left_el_blkno)
1422 {
1423         int i;
1424
1425         BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1426                le16_to_cpu(left_el->l_tree_depth));
1427
1428         for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1429                 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1430                         break;
1431         }
1432
1433         /*
1434          * The path walking code should have never returned a root and
1435          * two paths which are not adjacent.
1436          */
1437         BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1438
1439         ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1440                                       &root_el->l_recs[i + 1], right_el);
1441 }
1442
1443 /*
1444  * We've changed a leaf block (in right_path) and need to reflect that
1445  * change back up the subtree.
1446  *
1447  * This happens in multiple places:
1448  *   - When we've moved an extent record from the left path leaf to the right
1449  *     path leaf to make room for an empty extent in the left path leaf.
1450  *   - When our insert into the right path leaf is at the leftmost edge
1451  *     and requires an update of the path immediately to it's left. This
1452  *     can occur at the end of some types of rotation and appending inserts.
1453  */
1454 static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1455                                        struct ocfs2_path *left_path,
1456                                        struct ocfs2_path *right_path,
1457                                        int subtree_index)
1458 {
1459         int ret, i, idx;
1460         struct ocfs2_extent_list *el, *left_el, *right_el;
1461         struct ocfs2_extent_rec *left_rec, *right_rec;
1462         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1463
1464         /*
1465          * Update the counts and position values within all the
1466          * interior nodes to reflect the leaf rotation we just did.
1467          *
1468          * The root node is handled below the loop.
1469          *
1470          * We begin the loop with right_el and left_el pointing to the
1471          * leaf lists and work our way up.
1472          *
1473          * NOTE: within this loop, left_el and right_el always refer
1474          * to the *child* lists.
1475          */
1476         left_el = path_leaf_el(left_path);
1477         right_el = path_leaf_el(right_path);
1478         for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
1479                 mlog(0, "Adjust records at index %u\n", i);
1480
1481                 /*
1482                  * One nice property of knowing that all of these
1483                  * nodes are below the root is that we only deal with
1484                  * the leftmost right node record and the rightmost
1485                  * left node record.
1486                  */
1487                 el = left_path->p_node[i].el;
1488                 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
1489                 left_rec = &el->l_recs[idx];
1490
1491                 el = right_path->p_node[i].el;
1492                 right_rec = &el->l_recs[0];
1493
1494                 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
1495                                               right_el);
1496
1497                 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
1498                 if (ret)
1499                         mlog_errno(ret);
1500
1501                 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
1502                 if (ret)
1503                         mlog_errno(ret);
1504
1505                 /*
1506                  * Setup our list pointers now so that the current
1507                  * parents become children in the next iteration.
1508                  */
1509                 left_el = left_path->p_node[i].el;
1510                 right_el = right_path->p_node[i].el;
1511         }
1512
1513         /*
1514          * At the root node, adjust the two adjacent records which
1515          * begin our path to the leaves.
1516          */
1517
1518         el = left_path->p_node[subtree_index].el;
1519         left_el = left_path->p_node[subtree_index + 1].el;
1520         right_el = right_path->p_node[subtree_index + 1].el;
1521
1522         ocfs2_adjust_root_records(el, left_el, right_el,
1523                                   left_path->p_node[subtree_index + 1].bh->b_blocknr);
1524
1525         root_bh = left_path->p_node[subtree_index].bh;
1526
1527         ret = ocfs2_journal_dirty(handle, root_bh);
1528         if (ret)
1529                 mlog_errno(ret);
1530 }
1531
1532 static int ocfs2_rotate_subtree_right(struct inode *inode,
1533                                       handle_t *handle,
1534                                       struct ocfs2_path *left_path,
1535                                       struct ocfs2_path *right_path,
1536                                       int subtree_index)
1537 {
1538         int ret, i;
1539         struct buffer_head *right_leaf_bh;
1540         struct buffer_head *left_leaf_bh = NULL;
1541         struct buffer_head *root_bh;
1542         struct ocfs2_extent_list *right_el, *left_el;
1543         struct ocfs2_extent_rec move_rec;
1544
1545         left_leaf_bh = path_leaf_bh(left_path);
1546         left_el = path_leaf_el(left_path);
1547
1548         if (left_el->l_next_free_rec != left_el->l_count) {
1549                 ocfs2_error(inode->i_sb,
1550                             "Inode %llu has non-full interior leaf node %llu"
1551                             "(next free = %u)",
1552                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1553                             (unsigned long long)left_leaf_bh->b_blocknr,
1554                             le16_to_cpu(left_el->l_next_free_rec));
1555                 return -EROFS;
1556         }
1557
1558         /*
1559          * This extent block may already have an empty record, so we
1560          * return early if so.
1561          */
1562         if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
1563                 return 0;
1564
1565         root_bh = left_path->p_node[subtree_index].bh;
1566         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
1567
1568         ret = ocfs2_journal_access(handle, inode, root_bh,
1569                                    OCFS2_JOURNAL_ACCESS_WRITE);
1570         if (ret) {
1571                 mlog_errno(ret);
1572                 goto out;
1573         }
1574
1575         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
1576                 ret = ocfs2_journal_access(handle, inode,
1577                                            right_path->p_node[i].bh,
1578                                            OCFS2_JOURNAL_ACCESS_WRITE);
1579                 if (ret) {
1580                         mlog_errno(ret);
1581                         goto out;
1582                 }
1583
1584                 ret = ocfs2_journal_access(handle, inode,
1585                                            left_path->p_node[i].bh,
1586                                            OCFS2_JOURNAL_ACCESS_WRITE);
1587                 if (ret) {
1588                         mlog_errno(ret);
1589                         goto out;
1590                 }
1591         }
1592
1593         right_leaf_bh = path_leaf_bh(right_path);
1594         right_el = path_leaf_el(right_path);
1595
1596         /* This is a code error, not a disk corruption. */
1597         mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
1598                         "because rightmost leaf block %llu is empty\n",
1599                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1600                         (unsigned long long)right_leaf_bh->b_blocknr);
1601
1602         ocfs2_create_empty_extent(right_el);
1603
1604         ret = ocfs2_journal_dirty(handle, right_leaf_bh);
1605         if (ret) {
1606                 mlog_errno(ret);
1607                 goto out;
1608         }
1609
1610         /* Do the copy now. */
1611         i = le16_to_cpu(left_el->l_next_free_rec) - 1;
1612         move_rec = left_el->l_recs[i];
1613         right_el->l_recs[0] = move_rec;
1614
1615         /*
1616          * Clear out the record we just copied and shift everything
1617          * over, leaving an empty extent in the left leaf.
1618          *
1619          * We temporarily subtract from next_free_rec so that the
1620          * shift will lose the tail record (which is now defunct).
1621          */
1622         le16_add_cpu(&left_el->l_next_free_rec, -1);
1623         ocfs2_shift_records_right(left_el);
1624         memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1625         le16_add_cpu(&left_el->l_next_free_rec, 1);
1626
1627         ret = ocfs2_journal_dirty(handle, left_leaf_bh);
1628         if (ret) {
1629                 mlog_errno(ret);
1630                 goto out;
1631         }
1632
1633         ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
1634                                 subtree_index);
1635
1636 out:
1637         return ret;
1638 }
1639
1640 /*
1641  * Given a full path, determine what cpos value would return us a path
1642  * containing the leaf immediately to the left of the current one.
1643  *
1644  * Will return zero if the path passed in is already the leftmost path.
1645  */
1646 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
1647                                          struct ocfs2_path *path, u32 *cpos)
1648 {
1649         int i, j, ret = 0;
1650         u64 blkno;
1651         struct ocfs2_extent_list *el;
1652
1653         BUG_ON(path->p_tree_depth == 0);
1654
1655         *cpos = 0;
1656
1657         blkno = path_leaf_bh(path)->b_blocknr;
1658
1659         /* Start at the tree node just above the leaf and work our way up. */
1660         i = path->p_tree_depth - 1;
1661         while (i >= 0) {
1662                 el = path->p_node[i].el;
1663
1664                 /*
1665                  * Find the extent record just before the one in our
1666                  * path.
1667                  */
1668                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
1669                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
1670                                 if (j == 0) {
1671                                         if (i == 0) {
1672                                                 /*
1673                                                  * We've determined that the
1674                                                  * path specified is already
1675                                                  * the leftmost one - return a
1676                                                  * cpos of zero.
1677                                                  */
1678                                                 goto out;
1679                                         }
1680                                         /*
1681                                          * The leftmost record points to our
1682                                          * leaf - we need to travel up the
1683                                          * tree one level.
1684                                          */
1685                                         goto next_node;
1686                                 }
1687
1688                                 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
1689                                 *cpos = *cpos + ocfs2_rec_clusters(el,
1690                                                            &el->l_recs[j - 1]);
1691                                 *cpos = *cpos - 1;
1692                                 goto out;
1693                         }
1694                 }
1695
1696                 /*
1697                  * If we got here, we never found a valid node where
1698                  * the tree indicated one should be.
1699                  */
1700                 ocfs2_error(sb,
1701                             "Invalid extent tree at extent block %llu\n",
1702                             (unsigned long long)blkno);
1703                 ret = -EROFS;
1704                 goto out;
1705
1706 next_node:
1707                 blkno = path->p_node[i].bh->b_blocknr;
1708                 i--;
1709         }
1710
1711 out:
1712         return ret;
1713 }
1714
1715 /*
1716  * Extend the transaction by enough credits to complete the rotation,
1717  * and still leave at least the original number of credits allocated
1718  * to this transaction.
1719  */
1720 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
1721                                            int op_credits,
1722                                            struct ocfs2_path *path)
1723 {
1724         int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
1725
1726         if (handle->h_buffer_credits < credits)
1727                 return ocfs2_extend_trans(handle, credits);
1728
1729         return 0;
1730 }
1731
1732 /*
1733  * Trap the case where we're inserting into the theoretical range past
1734  * the _actual_ left leaf range. Otherwise, we'll rotate a record
1735  * whose cpos is less than ours into the right leaf.
1736  *
1737  * It's only necessary to look at the rightmost record of the left
1738  * leaf because the logic that calls us should ensure that the
1739  * theoretical ranges in the path components above the leaves are
1740  * correct.
1741  */
1742 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
1743                                                  u32 insert_cpos)
1744 {
1745         struct ocfs2_extent_list *left_el;
1746         struct ocfs2_extent_rec *rec;
1747         int next_free;
1748
1749         left_el = path_leaf_el(left_path);
1750         next_free = le16_to_cpu(left_el->l_next_free_rec);
1751         rec = &left_el->l_recs[next_free - 1];
1752
1753         if (insert_cpos > le32_to_cpu(rec->e_cpos))
1754                 return 1;
1755         return 0;
1756 }
1757
1758 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
1759 {
1760         int next_free = le16_to_cpu(el->l_next_free_rec);
1761         unsigned int range;
1762         struct ocfs2_extent_rec *rec;
1763
1764         if (next_free == 0)
1765                 return 0;
1766
1767         rec = &el->l_recs[0];
1768         if (ocfs2_is_empty_extent(rec)) {
1769                 /* Empty list. */
1770                 if (next_free == 1)
1771                         return 0;
1772                 rec = &el->l_recs[1];
1773         }
1774
1775         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1776         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1777                 return 1;
1778         return 0;
1779 }
1780
1781 /*
1782  * Rotate all the records in a btree right one record, starting at insert_cpos.
1783  *
1784  * The path to the rightmost leaf should be passed in.
1785  *
1786  * The array is assumed to be large enough to hold an entire path (tree depth).
1787  *
1788  * Upon succesful return from this function:
1789  *
1790  * - The 'right_path' array will contain a path to the leaf block
1791  *   whose range contains e_cpos.
1792  * - That leaf block will have a single empty extent in list index 0.
1793  * - In the case that the rotation requires a post-insert update,
1794  *   *ret_left_path will contain a valid path which can be passed to
1795  *   ocfs2_insert_path().
1796  */
1797 static int ocfs2_rotate_tree_right(struct inode *inode,
1798                                    handle_t *handle,
1799                                    enum ocfs2_split_type split,
1800                                    u32 insert_cpos,
1801                                    struct ocfs2_path *right_path,
1802                                    struct ocfs2_path **ret_left_path)
1803 {
1804         int ret, start, orig_credits = handle->h_buffer_credits;
1805         u32 cpos;
1806         struct ocfs2_path *left_path = NULL;
1807
1808         *ret_left_path = NULL;
1809
1810         left_path = ocfs2_new_path(path_root_bh(right_path),
1811                                    path_root_el(right_path));
1812         if (!left_path) {
1813                 ret = -ENOMEM;
1814                 mlog_errno(ret);
1815                 goto out;
1816         }
1817
1818         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
1819         if (ret) {
1820                 mlog_errno(ret);
1821                 goto out;
1822         }
1823
1824         mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
1825
1826         /*
1827          * What we want to do here is:
1828          *
1829          * 1) Start with the rightmost path.
1830          *
1831          * 2) Determine a path to the leaf block directly to the left
1832          *    of that leaf.
1833          *
1834          * 3) Determine the 'subtree root' - the lowest level tree node
1835          *    which contains a path to both leaves.
1836          *
1837          * 4) Rotate the subtree.
1838          *
1839          * 5) Find the next subtree by considering the left path to be
1840          *    the new right path.
1841          *
1842          * The check at the top of this while loop also accepts
1843          * insert_cpos == cpos because cpos is only a _theoretical_
1844          * value to get us the left path - insert_cpos might very well
1845          * be filling that hole.
1846          *
1847          * Stop at a cpos of '0' because we either started at the
1848          * leftmost branch (i.e., a tree with one branch and a
1849          * rotation inside of it), or we've gone as far as we can in
1850          * rotating subtrees.
1851          */
1852         while (cpos && insert_cpos <= cpos) {
1853                 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
1854                      insert_cpos, cpos);
1855
1856                 ret = ocfs2_find_path(inode, left_path, cpos);
1857                 if (ret) {
1858                         mlog_errno(ret);
1859                         goto out;
1860                 }
1861
1862                 mlog_bug_on_msg(path_leaf_bh(left_path) ==
1863                                 path_leaf_bh(right_path),
1864                                 "Inode %lu: error during insert of %u "
1865                                 "(left path cpos %u) results in two identical "
1866                                 "paths ending at %llu\n",
1867                                 inode->i_ino, insert_cpos, cpos,
1868                                 (unsigned long long)
1869                                 path_leaf_bh(left_path)->b_blocknr);
1870
1871                 if (split == SPLIT_NONE &&
1872                     ocfs2_rotate_requires_path_adjustment(left_path,
1873                                                           insert_cpos)) {
1874
1875                         /*
1876                          * We've rotated the tree as much as we
1877                          * should. The rest is up to
1878                          * ocfs2_insert_path() to complete, after the
1879                          * record insertion. We indicate this
1880                          * situation by returning the left path.
1881                          *
1882                          * The reason we don't adjust the records here
1883                          * before the record insert is that an error
1884                          * later might break the rule where a parent
1885                          * record e_cpos will reflect the actual
1886                          * e_cpos of the 1st nonempty record of the
1887                          * child list.
1888                          */
1889                         *ret_left_path = left_path;
1890                         goto out_ret_path;
1891                 }
1892
1893                 start = ocfs2_find_subtree_root(inode, left_path, right_path);
1894
1895                 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
1896                      start,
1897                      (unsigned long long) right_path->p_node[start].bh->b_blocknr,
1898                      right_path->p_tree_depth);
1899
1900                 ret = ocfs2_extend_rotate_transaction(handle, start,
1901                                                       orig_credits, right_path);
1902                 if (ret) {
1903                         mlog_errno(ret);
1904                         goto out;
1905                 }
1906
1907                 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
1908                                                  right_path, start);
1909                 if (ret) {
1910                         mlog_errno(ret);
1911                         goto out;
1912                 }
1913
1914                 if (split != SPLIT_NONE &&
1915                     ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
1916                                                 insert_cpos)) {
1917                         /*
1918                          * A rotate moves the rightmost left leaf
1919                          * record over to the leftmost right leaf
1920                          * slot. If we're doing an extent split
1921                          * instead of a real insert, then we have to
1922                          * check that the extent to be split wasn't
1923                          * just moved over. If it was, then we can
1924                          * exit here, passing left_path back -
1925                          * ocfs2_split_extent() is smart enough to
1926                          * search both leaves.
1927                          */
1928                         *ret_left_path = left_path;
1929                         goto out_ret_path;
1930                 }
1931
1932                 /*
1933                  * There is no need to re-read the next right path
1934                  * as we know that it'll be our current left
1935                  * path. Optimize by copying values instead.
1936                  */
1937                 ocfs2_mv_path(right_path, left_path);
1938
1939                 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
1940                                                     &cpos);
1941                 if (ret) {
1942                         mlog_errno(ret);
1943                         goto out;
1944                 }
1945         }
1946
1947 out:
1948         ocfs2_free_path(left_path);
1949
1950 out_ret_path:
1951         return ret;
1952 }
1953
1954 static void ocfs2_update_edge_lengths(struct inode *inode, handle_t *handle,
1955                                       struct ocfs2_path *path)
1956 {
1957         int i, idx;
1958         struct ocfs2_extent_rec *rec;
1959         struct ocfs2_extent_list *el;
1960         struct ocfs2_extent_block *eb;
1961         u32 range;
1962
1963         /* Path should always be rightmost. */
1964         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
1965         BUG_ON(eb->h_next_leaf_blk != 0ULL);
1966
1967         el = &eb->h_list;
1968         BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
1969         idx = le16_to_cpu(el->l_next_free_rec) - 1;
1970         rec = &el->l_recs[idx];
1971         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1972
1973         for (i = 0; i < path->p_tree_depth; i++) {
1974                 el = path->p_node[i].el;
1975                 idx = le16_to_cpu(el->l_next_free_rec) - 1;
1976                 rec = &el->l_recs[idx];
1977
1978                 rec->e_int_clusters = cpu_to_le32(range);
1979                 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
1980
1981                 ocfs2_journal_dirty(handle, path->p_node[i].bh);
1982         }
1983 }
1984
1985 static void ocfs2_unlink_path(struct inode *inode, handle_t *handle,
1986                               struct ocfs2_cached_dealloc_ctxt *dealloc,
1987                               struct ocfs2_path *path, int unlink_start)
1988 {
1989         int ret, i;
1990         struct ocfs2_extent_block *eb;
1991         struct ocfs2_extent_list *el;
1992         struct buffer_head *bh;
1993
1994         for(i = unlink_start; i < path_num_items(path); i++) {
1995                 bh = path->p_node[i].bh;
1996
1997                 eb = (struct ocfs2_extent_block *)bh->b_data;
1998                 /*
1999                  * Not all nodes might have had their final count
2000                  * decremented by the caller - handle this here.
2001                  */
2002                 el = &eb->h_list;
2003                 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2004                         mlog(ML_ERROR,
2005                              "Inode %llu, attempted to remove extent block "
2006                              "%llu with %u records\n",
2007                              (unsigned long long)OCFS2_I(inode)->ip_blkno,
2008                              (unsigned long long)le64_to_cpu(eb->h_blkno),
2009                              le16_to_cpu(el->l_next_free_rec));
2010
2011                         ocfs2_journal_dirty(handle, bh);
2012                         ocfs2_remove_from_cache(inode, bh);
2013                         continue;
2014                 }
2015
2016                 el->l_next_free_rec = 0;
2017                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2018
2019                 ocfs2_journal_dirty(handle, bh);
2020
2021                 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2022                 if (ret)
2023                         mlog_errno(ret);
2024
2025                 ocfs2_remove_from_cache(inode, bh);
2026         }
2027 }
2028
2029 static void ocfs2_unlink_subtree(struct inode *inode, handle_t *handle,
2030                                  struct ocfs2_path *left_path,
2031                                  struct ocfs2_path *right_path,
2032                                  int subtree_index,
2033                                  struct ocfs2_cached_dealloc_ctxt *dealloc)
2034 {
2035         int i;
2036         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2037         struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2038         struct ocfs2_extent_list *el;
2039         struct ocfs2_extent_block *eb;
2040
2041         el = path_leaf_el(left_path);
2042
2043         eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2044
2045         for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2046                 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2047                         break;
2048
2049         BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2050
2051         memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2052         le16_add_cpu(&root_el->l_next_free_rec, -1);
2053
2054         eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2055         eb->h_next_leaf_blk = 0;
2056
2057         ocfs2_journal_dirty(handle, root_bh);
2058         ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2059
2060         ocfs2_unlink_path(inode, handle, dealloc, right_path,
2061                           subtree_index + 1);
2062 }
2063
2064 static int ocfs2_rotate_subtree_left(struct inode *inode, handle_t *handle,
2065                                      struct ocfs2_path *left_path,
2066                                      struct ocfs2_path *right_path,
2067                                      int subtree_index,
2068                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2069                                      int *deleted)
2070 {
2071         int ret, i, del_right_subtree = 0, right_has_empty = 0;
2072         struct buffer_head *root_bh, *di_bh = path_root_bh(right_path);
2073         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2074         struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2075         struct ocfs2_extent_block *eb;
2076
2077         *deleted = 0;
2078
2079         right_leaf_el = path_leaf_el(right_path);
2080         left_leaf_el = path_leaf_el(left_path);
2081         root_bh = left_path->p_node[subtree_index].bh;
2082         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2083
2084         if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2085                 return 0;
2086
2087         eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2088         if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2089                 /*
2090                  * It's legal for us to proceed if the right leaf is
2091                  * the rightmost one and it has an empty extent. There
2092                  * are two cases to handle - whether the leaf will be
2093                  * empty after removal or not. If the leaf isn't empty
2094                  * then just remove the empty extent up front. The
2095                  * next block will handle empty leaves by flagging
2096                  * them for unlink.
2097                  *
2098                  * Non rightmost leaves will throw -EAGAIN and the
2099                  * caller can manually move the subtree and retry.
2100                  */
2101
2102                 if (eb->h_next_leaf_blk != 0ULL)
2103                         return -EAGAIN;
2104
2105                 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2106                         ret = ocfs2_journal_access(handle, inode,
2107                                                    path_leaf_bh(right_path),
2108                                                    OCFS2_JOURNAL_ACCESS_WRITE);
2109                         if (ret) {
2110                                 mlog_errno(ret);
2111                                 goto out;
2112                         }
2113
2114                         ocfs2_remove_empty_extent(right_leaf_el);
2115                 } else
2116                         right_has_empty = 1;
2117         }
2118
2119         if (eb->h_next_leaf_blk == 0ULL &&
2120             le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2121                 /*
2122                  * We have to update i_last_eb_blk during the meta
2123                  * data delete.
2124                  */
2125                 ret = ocfs2_journal_access(handle, inode, di_bh,
2126                                            OCFS2_JOURNAL_ACCESS_WRITE);
2127                 if (ret) {
2128                         mlog_errno(ret);
2129                         goto out;
2130                 }
2131
2132                 del_right_subtree = 1;
2133         }
2134
2135         /*
2136          * Getting here with an empty extent in the right path implies
2137          * that it's the rightmost path and will be deleted.
2138          */
2139         BUG_ON(right_has_empty && !del_right_subtree);
2140
2141         ret = ocfs2_journal_access(handle, inode, root_bh,
2142                                    OCFS2_JOURNAL_ACCESS_WRITE);
2143         if (ret) {
2144                 mlog_errno(ret);
2145                 goto out;
2146         }
2147
2148         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2149                 ret = ocfs2_journal_access(handle, inode,
2150                                            right_path->p_node[i].bh,
2151                                            OCFS2_JOURNAL_ACCESS_WRITE);
2152                 if (ret) {
2153                         mlog_errno(ret);
2154                         goto out;
2155                 }
2156
2157                 ret = ocfs2_journal_access(handle, inode,
2158                                            left_path->p_node[i].bh,
2159                                            OCFS2_JOURNAL_ACCESS_WRITE);
2160                 if (ret) {
2161                         mlog_errno(ret);
2162                         goto out;
2163                 }
2164         }
2165
2166         if (!right_has_empty) {
2167                 /*
2168                  * Only do this if we're moving a real
2169                  * record. Otherwise, the action is delayed until
2170                  * after removal of the right path in which case we
2171                  * can do a simple shift to remove the empty extent.
2172                  */
2173                 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2174                 memset(&right_leaf_el->l_recs[0], 0,
2175                        sizeof(struct ocfs2_extent_rec));
2176         }
2177         if (eb->h_next_leaf_blk == 0ULL) {
2178                 /*
2179                  * Move recs over to get rid of empty extent, decrease
2180                  * next_free. This is allowed to remove the last
2181                  * extent in our leaf (setting l_next_free_rec to
2182                  * zero) - the delete code below won't care.
2183                  */
2184                 ocfs2_remove_empty_extent(right_leaf_el);
2185         }
2186
2187         ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2188         if (ret)
2189                 mlog_errno(ret);
2190         ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2191         if (ret)
2192                 mlog_errno(ret);
2193
2194         if (del_right_subtree) {
2195                 ocfs2_unlink_subtree(inode, handle, left_path, right_path,
2196                                      subtree_index, dealloc);
2197                 ocfs2_update_edge_lengths(inode, handle, left_path);
2198
2199                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2200                 di->i_last_eb_blk = eb->h_blkno;
2201
2202                 /*
2203                  * Removal of the extent in the left leaf was skipped
2204                  * above so we could delete the right path
2205                  * 1st.
2206                  */
2207                 if (right_has_empty)
2208                         ocfs2_remove_empty_extent(left_leaf_el);
2209
2210                 ret = ocfs2_journal_dirty(handle, di_bh);
2211                 if (ret)
2212                         mlog_errno(ret);
2213
2214                 *deleted = 1;
2215         } else
2216                 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
2217                                            subtree_index);
2218
2219 out:
2220         return ret;
2221 }
2222
2223 /*
2224  * Given a full path, determine what cpos value would return us a path
2225  * containing the leaf immediately to the right of the current one.
2226  *
2227  * Will return zero if the path passed in is already the rightmost path.
2228  *
2229  * This looks similar, but is subtly different to
2230  * ocfs2_find_cpos_for_left_leaf().
2231  */
2232 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2233                                           struct ocfs2_path *path, u32 *cpos)
2234 {
2235         int i, j, ret = 0;
2236         u64 blkno;
2237         struct ocfs2_extent_list *el;
2238
2239         *cpos = 0;
2240
2241         if (path->p_tree_depth == 0)
2242                 return 0;
2243
2244         blkno = path_leaf_bh(path)->b_blocknr;
2245
2246         /* Start at the tree node just above the leaf and work our way up. */
2247         i = path->p_tree_depth - 1;
2248         while (i >= 0) {
2249                 int next_free;
2250
2251                 el = path->p_node[i].el;
2252
2253                 /*
2254                  * Find the extent record just after the one in our
2255                  * path.
2256                  */
2257                 next_free = le16_to_cpu(el->l_next_free_rec);
2258                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2259                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2260                                 if (j == (next_free - 1)) {
2261                                         if (i == 0) {
2262                                                 /*
2263                                                  * We've determined that the
2264                                                  * path specified is already
2265                                                  * the rightmost one - return a
2266                                                  * cpos of zero.
2267                                                  */
2268                                                 goto out;
2269                                         }
2270                                         /*
2271                                          * The rightmost record points to our
2272                                          * leaf - we need to travel up the
2273                                          * tree one level.
2274                                          */
2275                                         goto next_node;
2276                                 }
2277
2278                                 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2279                                 goto out;
2280                         }
2281                 }
2282
2283                 /*
2284                  * If we got here, we never found a valid node where
2285                  * the tree indicated one should be.
2286                  */
2287                 ocfs2_error(sb,
2288                             "Invalid extent tree at extent block %llu\n",
2289                             (unsigned long long)blkno);
2290                 ret = -EROFS;
2291                 goto out;
2292
2293 next_node:
2294                 blkno = path->p_node[i].bh->b_blocknr;
2295                 i--;
2296         }
2297
2298 out:
2299         return ret;
2300 }
2301
2302 static int ocfs2_rotate_rightmost_leaf_left(struct inode *inode,
2303                                             handle_t *handle,
2304                                             struct buffer_head *bh,
2305                                             struct ocfs2_extent_list *el)
2306 {
2307         int ret;
2308
2309         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2310                 return 0;
2311
2312         ret = ocfs2_journal_access(handle, inode, bh,
2313                                    OCFS2_JOURNAL_ACCESS_WRITE);
2314         if (ret) {
2315                 mlog_errno(ret);
2316                 goto out;
2317         }
2318
2319         ocfs2_remove_empty_extent(el);
2320
2321         ret = ocfs2_journal_dirty(handle, bh);
2322         if (ret)
2323                 mlog_errno(ret);
2324
2325 out:
2326         return ret;
2327 }
2328
2329 static int __ocfs2_rotate_tree_left(struct inode *inode,
2330                                     handle_t *handle, int orig_credits,
2331                                     struct ocfs2_path *path,
2332                                     struct ocfs2_cached_dealloc_ctxt *dealloc,
2333                                     struct ocfs2_path **empty_extent_path)
2334 {
2335         int ret, subtree_root, deleted;
2336         u32 right_cpos;
2337         struct ocfs2_path *left_path = NULL;
2338         struct ocfs2_path *right_path = NULL;
2339
2340         BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2341
2342         *empty_extent_path = NULL;
2343
2344         ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, path,
2345                                              &right_cpos);
2346         if (ret) {
2347                 mlog_errno(ret);
2348                 goto out;
2349         }
2350
2351         left_path = ocfs2_new_path(path_root_bh(path),
2352                                    path_root_el(path));
2353         if (!left_path) {
2354                 ret = -ENOMEM;
2355                 mlog_errno(ret);
2356                 goto out;
2357         }
2358
2359         ocfs2_cp_path(left_path, path);
2360
2361         right_path = ocfs2_new_path(path_root_bh(path),
2362                                     path_root_el(path));
2363         if (!right_path) {
2364                 ret = -ENOMEM;
2365                 mlog_errno(ret);
2366                 goto out;
2367         }
2368
2369         while (right_cpos) {
2370                 ret = ocfs2_find_path(inode, right_path, right_cpos);
2371                 if (ret) {
2372                         mlog_errno(ret);
2373                         goto out;
2374                 }
2375
2376                 subtree_root = ocfs2_find_subtree_root(inode, left_path,
2377                                                        right_path);
2378
2379                 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
2380                      subtree_root,
2381                      (unsigned long long)
2382                      right_path->p_node[subtree_root].bh->b_blocknr,
2383                      right_path->p_tree_depth);
2384
2385                 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2386                                                       orig_credits, left_path);
2387                 if (ret) {
2388                         mlog_errno(ret);
2389                         goto out;
2390                 }
2391
2392                 ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
2393                                                 right_path, subtree_root,
2394                                                 dealloc, &deleted);
2395                 if (ret == -EAGAIN) {
2396                         /*
2397                          * The rotation has to temporarily stop due to
2398                          * the right subtree having an empty
2399                          * extent. Pass it back to the caller for a
2400                          * fixup.
2401                          */
2402                         *empty_extent_path = right_path;
2403                         right_path = NULL;
2404                         goto out;
2405                 }
2406                 if (ret) {
2407                         mlog_errno(ret);
2408                         goto out;
2409                 }
2410
2411                 /*
2412                  * The subtree rotate might have removed records on
2413                  * the rightmost edge. If so, then rotation is
2414                  * complete.
2415                  */
2416                 if (deleted)
2417                         break;
2418
2419                 ocfs2_mv_path(left_path, right_path);
2420
2421                 ret = ocfs2_find_cpos_for_right_leaf(inode->i_sb, left_path,
2422                                                      &right_cpos);
2423                 if (ret) {
2424                         mlog_errno(ret);
2425                         goto out;
2426                 }
2427         }
2428
2429 out:
2430         ocfs2_free_path(right_path);
2431         ocfs2_free_path(left_path);
2432
2433         return ret;
2434 }
2435
2436 static int ocfs2_remove_rightmost_path(struct inode *inode, handle_t *handle,
2437                                        struct ocfs2_path *path,
2438                                        struct ocfs2_cached_dealloc_ctxt *dealloc)
2439 {
2440         int ret, subtree_index;
2441         u32 cpos;
2442         struct ocfs2_path *left_path = NULL;
2443         struct ocfs2_dinode *di;
2444         struct ocfs2_extent_block *eb;
2445         struct ocfs2_extent_list *el;
2446
2447         /*
2448          * XXX: This code assumes that the root is an inode, which is
2449          * true for now but may change as tree code gets generic.
2450          */
2451         di = (struct ocfs2_dinode *)path_root_bh(path)->b_data;
2452         if (!OCFS2_IS_VALID_DINODE(di)) {
2453                 ret = -EIO;
2454                 ocfs2_error(inode->i_sb,
2455                             "Inode %llu has invalid path root",
2456                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
2457                 goto out;
2458         }
2459
2460         /*
2461          * There's two ways we handle this depending on
2462          * whether path is the only existing one.
2463          */
2464         ret = ocfs2_extend_rotate_transaction(handle, 0,
2465                                               handle->h_buffer_credits,
2466                                               path);
2467         if (ret) {
2468                 mlog_errno(ret);
2469                 goto out;
2470         }
2471
2472         ret = ocfs2_journal_access_path(inode, handle, path);
2473         if (ret) {
2474                 mlog_errno(ret);
2475                 goto out;
2476         }
2477
2478         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
2479         if (ret) {
2480                 mlog_errno(ret);
2481                 goto out;
2482         }
2483
2484         if (cpos) {
2485                 /*
2486                  * We have a path to the left of this one - it needs
2487                  * an update too.
2488                  */
2489                 left_path = ocfs2_new_path(path_root_bh(path),
2490                                            path_root_el(path));
2491                 if (!left_path) {
2492                         ret = -ENOMEM;
2493                         mlog_errno(ret);
2494                         goto out;
2495                 }
2496
2497                 ret = ocfs2_find_path(inode, left_path, cpos);
2498                 if (ret) {
2499                         mlog_errno(ret);
2500                         goto out;
2501                 }
2502
2503                 ret = ocfs2_journal_access_path(inode, handle, left_path);
2504                 if (ret) {
2505                         mlog_errno(ret);
2506                         goto out;
2507                 }
2508
2509                 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
2510
2511                 ocfs2_unlink_subtree(inode, handle, left_path, path,
2512                                      subtree_index, dealloc);
2513                 ocfs2_update_edge_lengths(inode, handle, left_path);
2514
2515                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2516                 di->i_last_eb_blk = eb->h_blkno;
2517         } else {
2518                 /*
2519                  * 'path' is also the leftmost path which
2520                  * means it must be the only one. This gets
2521                  * handled differently because we want to
2522                  * revert the inode back to having extents
2523                  * in-line.
2524                  */
2525                 ocfs2_unlink_path(inode, handle, dealloc, path, 1);
2526
2527                 el = &di->id2.i_list;
2528                 el->l_tree_depth = 0;
2529                 el->l_next_free_rec = 0;
2530                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2531
2532                 di->i_last_eb_blk = 0;
2533         }
2534
2535         ocfs2_journal_dirty(handle, path_root_bh(path));
2536
2537 out:
2538         ocfs2_free_path(left_path);
2539         return ret;
2540 }
2541
2542 /*
2543  * Left rotation of btree records.
2544  *
2545  * In many ways, this is (unsurprisingly) the opposite of right
2546  * rotation. We start at some non-rightmost path containing an empty
2547  * extent in the leaf block. The code works its way to the rightmost
2548  * path by rotating records to the left in every subtree.
2549  *
2550  * This is used by any code which reduces the number of extent records
2551  * in a leaf. After removal, an empty record should be placed in the
2552  * leftmost list position.
2553  *
2554  * This won't handle a length update of the rightmost path records if
2555  * the rightmost tree leaf record is removed so the caller is
2556  * responsible for detecting and correcting that.
2557  */
2558 static int ocfs2_rotate_tree_left(struct inode *inode, handle_t *handle,
2559                                   struct ocfs2_path *path,
2560                                   struct ocfs2_cached_dealloc_ctxt *dealloc)
2561 {
2562         int ret, orig_credits = handle->h_buffer_credits;
2563         struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
2564         struct ocfs2_extent_block *eb;
2565         struct ocfs2_extent_list *el;
2566
2567         el = path_leaf_el(path);
2568         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2569                 return 0;
2570
2571         if (path->p_tree_depth == 0) {
2572 rightmost_no_delete:
2573                 /*
2574                  * In-inode extents. This is trivially handled, so do
2575                  * it up front.
2576                  */
2577                 ret = ocfs2_rotate_rightmost_leaf_left(inode, handle,
2578                                                        path_leaf_bh(path),
2579                                                        path_leaf_el(path));
2580                 if (ret)
2581                         mlog_errno(ret);
2582                 goto out;
2583         }
2584
2585         /*
2586          * Handle rightmost branch now. There's several cases:
2587          *  1) simple rotation leaving records in there. That's trivial.
2588          *  2) rotation requiring a branch delete - there's no more
2589          *     records left. Two cases of this:
2590          *     a) There are branches to the left.
2591          *     b) This is also the leftmost (the only) branch.
2592          *
2593          *  1) is handled via ocfs2_rotate_rightmost_leaf_left()
2594          *  2a) we need the left branch so that we can update it with the unlink
2595          *  2b) we need to bring the inode back to inline extents.
2596          */
2597
2598         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2599         el = &eb->h_list;
2600         if (eb->h_next_leaf_blk == 0) {
2601                 /*
2602                  * This gets a bit tricky if we're going to delete the
2603                  * rightmost path. Get the other cases out of the way
2604                  * 1st.
2605                  */
2606                 if (le16_to_cpu(el->l_next_free_rec) > 1)
2607                         goto rightmost_no_delete;
2608
2609                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
2610                         ret = -EIO;
2611                         ocfs2_error(inode->i_sb,
2612                                     "Inode %llu has empty extent block at %llu",
2613                                     (unsigned long long)OCFS2_I(inode)->ip_blkno,
2614                                     (unsigned long long)le64_to_cpu(eb->h_blkno));
2615                         goto out;
2616                 }
2617
2618                 /*
2619                  * XXX: The caller can not trust "path" any more after
2620                  * this as it will have been deleted. What do we do?
2621                  *
2622                  * In theory the rotate-for-merge code will never get
2623                  * here because it'll always ask for a rotate in a
2624                  * nonempty list.
2625                  */
2626
2627                 ret = ocfs2_remove_rightmost_path(inode, handle, path,
2628                                                   dealloc);
2629                 if (ret)
2630                         mlog_errno(ret);
2631                 goto out;
2632         }
2633
2634         /*
2635          * Now we can loop, remembering the path we get from -EAGAIN
2636          * and restarting from there.
2637          */
2638 try_rotate:
2639         ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits, path,
2640                                        dealloc, &restart_path);
2641         if (ret && ret != -EAGAIN) {
2642                 mlog_errno(ret);
2643                 goto out;
2644         }
2645
2646         while (ret == -EAGAIN) {
2647                 tmp_path = restart_path;
2648                 restart_path = NULL;
2649
2650                 ret = __ocfs2_rotate_tree_left(inode, handle, orig_credits,
2651                                                tmp_path, dealloc,
2652                                                &restart_path);
2653                 if (ret && ret != -EAGAIN) {
2654                         mlog_errno(ret);
2655                         goto out;
2656                 }
2657
2658                 ocfs2_free_path(tmp_path);
2659                 tmp_path = NULL;
2660
2661                 if (ret == 0)
2662                         goto try_rotate;
2663         }
2664
2665 out:
2666         ocfs2_free_path(tmp_path);
2667         ocfs2_free_path(restart_path);
2668         return ret;
2669 }
2670
2671 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
2672                                 int index)
2673 {
2674         struct ocfs2_extent_rec *rec = &el->l_recs[index];
2675         unsigned int size;
2676
2677         if (rec->e_leaf_clusters == 0) {
2678                 /*
2679                  * We consumed all of the merged-from record. An empty
2680                  * extent cannot exist anywhere but the 1st array
2681                  * position, so move things over if the merged-from
2682                  * record doesn't occupy that position.
2683                  *
2684                  * This creates a new empty extent so the caller
2685                  * should be smart enough to have removed any existing
2686                  * ones.
2687                  */
2688                 if (index > 0) {
2689                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
2690                         size = index * sizeof(struct ocfs2_extent_rec);
2691                         memmove(&el->l_recs[1], &el->l_recs[0], size);
2692                 }
2693
2694                 /*
2695                  * Always memset - the caller doesn't check whether it
2696                  * created an empty extent, so there could be junk in
2697                  * the other fields.
2698                  */
2699                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2700         }
2701 }
2702
2703 /*
2704  * Remove split_rec clusters from the record at index and merge them
2705  * onto the beginning of the record at index + 1.
2706  */
2707 static int ocfs2_merge_rec_right(struct inode *inode, struct buffer_head *bh,
2708                                 handle_t *handle,
2709                                 struct ocfs2_extent_rec *split_rec,
2710                                 struct ocfs2_extent_list *el, int index)
2711 {
2712         int ret;
2713         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
2714         struct ocfs2_extent_rec *left_rec;
2715         struct ocfs2_extent_rec *right_rec;
2716
2717         BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
2718
2719         left_rec = &el->l_recs[index];
2720         right_rec = &el->l_recs[index + 1];
2721
2722         ret = ocfs2_journal_access(handle, inode, bh,
2723                                    OCFS2_JOURNAL_ACCESS_WRITE);
2724         if (ret) {
2725                 mlog_errno(ret);
2726                 goto out;
2727         }
2728
2729         le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
2730
2731         le32_add_cpu(&right_rec->e_cpos, -split_clusters);
2732         le64_add_cpu(&right_rec->e_blkno,
2733                      -ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
2734         le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
2735
2736         ocfs2_cleanup_merge(el, index);
2737
2738         ret = ocfs2_journal_dirty(handle, bh);
2739         if (ret)
2740                 mlog_errno(ret);
2741
2742 out:
2743         return ret;
2744 }
2745
2746 /*
2747  * Remove split_rec clusters from the record at index and merge them
2748  * onto the tail of the record at index - 1.
2749  */
2750 static int ocfs2_merge_rec_left(struct inode *inode, struct buffer_head *bh,
2751                                 handle_t *handle,
2752                                 struct ocfs2_extent_rec *split_rec,
2753                                 struct ocfs2_extent_list *el, int index)
2754 {
2755         int ret, has_empty_extent = 0;
2756         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
2757         struct ocfs2_extent_rec *left_rec;
2758         struct ocfs2_extent_rec *right_rec;
2759
2760         BUG_ON(index <= 0);
2761
2762         left_rec = &el->l_recs[index - 1];
2763         right_rec = &el->l_recs[index];
2764         if (ocfs2_is_empty_extent(&el->l_recs[0]))
2765                 has_empty_extent = 1;
2766
2767         ret = ocfs2_journal_access(handle, inode, bh,
2768                                    OCFS2_JOURNAL_ACCESS_WRITE);
2769         if (ret) {
2770                 mlog_errno(ret);
2771                 goto out;
2772         }
2773
2774         if (has_empty_extent && index == 1) {
2775                 /*
2776                  * The easy case - we can just plop the record right in.
2777                  */
2778                 *left_rec = *split_rec;
2779
2780                 has_empty_extent = 0;
2781         } else {
2782                 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
2783         }
2784
2785         le32_add_cpu(&right_rec->e_cpos, split_clusters);
2786         le64_add_cpu(&right_rec->e_blkno,
2787                      ocfs2_clusters_to_blocks(inode->i_sb, split_clusters));
2788         le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
2789
2790         ocfs2_cleanup_merge(el, index);
2791
2792         ret = ocfs2_journal_dirty(handle, bh);
2793         if (ret)
2794                 mlog_errno(ret);
2795
2796 out:
2797         return ret;
2798 }
2799
2800 static int ocfs2_try_to_merge_extent(struct inode *inode,
2801                                      handle_t *handle,
2802                                      struct ocfs2_path *left_path,
2803                                      int split_index,
2804                                      struct ocfs2_extent_rec *split_rec,
2805                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2806                                      struct ocfs2_merge_ctxt *ctxt)
2807
2808 {
2809         int ret = 0;
2810         struct ocfs2_extent_list *el = path_leaf_el(left_path);
2811         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
2812
2813         BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
2814
2815         if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
2816                 /*
2817                  * The merge code will need to create an empty
2818                  * extent to take the place of the newly
2819                  * emptied slot. Remove any pre-existing empty
2820                  * extents - having more than one in a leaf is
2821                  * illegal.
2822                  */
2823                 ret = ocfs2_rotate_tree_left(inode, handle, left_path,
2824                                              dealloc);
2825                 if (ret) {
2826                         mlog_errno(ret);
2827                         goto out;
2828                 }
2829                 split_index--;
2830                 rec = &el->l_recs[split_index];
2831         }
2832
2833         if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
2834                 /*
2835                  * Left-right contig implies this.
2836                  */
2837                 BUG_ON(!ctxt->c_split_covers_rec);
2838                 BUG_ON(split_index == 0);
2839
2840                 /*
2841                  * Since the leftright insert always covers the entire
2842                  * extent, this call will delete the insert record
2843                  * entirely, resulting in an empty extent record added to
2844                  * the extent block.
2845                  *
2846                  * Since the adding of an empty extent shifts
2847                  * everything back to the right, there's no need to
2848                  * update split_index here.
2849                  */
2850                 ret = ocfs2_merge_rec_left(inode, path_leaf_bh(left_path),
2851                                            handle, split_rec, el, split_index);
2852                 if (ret) {
2853                         mlog_errno(ret);
2854                         goto out;
2855                 }
2856
2857                 /*
2858                  * We can only get this from logic error above.
2859                  */
2860                 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
2861
2862                 /*
2863                  * The left merge left us with an empty extent, remove
2864                  * it.
2865                  */
2866                 ret = ocfs2_rotate_tree_left(inode, handle, left_path, dealloc);
2867                 if (ret) {
2868                         mlog_errno(ret);
2869                         goto out;
2870                 }
2871                 split_index--;
2872                 rec = &el->l_recs[split_index];
2873
2874                 /*
2875                  * Note that we don't pass split_rec here on purpose -
2876                  * we've merged it into the left side.
2877                  */
2878                 ret = ocfs2_merge_rec_right(inode, path_leaf_bh(left_path),
2879                                             handle, rec, el, split_index);
2880                 if (ret) {
2881                         mlog_errno(ret);
2882                         goto out;
2883                 }
2884
2885                 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
2886
2887                 ret = ocfs2_rotate_tree_left(inode, handle, left_path,
2888                                              dealloc);
2889                 /*
2890                  * Error from this last rotate is not critical, so
2891                  * print but don't bubble it up.
2892                  */
2893                 if (ret)
2894                         mlog_errno(ret);
2895                 ret = 0;
2896         } else {
2897                 /*
2898                  * Merge a record to the left or right.
2899                  *
2900                  * 'contig_type' is relative to the existing record,
2901                  * so for example, if we're "right contig", it's to
2902                  * the record on the left (hence the left merge).
2903                  */
2904                 if (ctxt->c_contig_type == CONTIG_RIGHT) {
2905                         ret = ocfs2_merge_rec_left(inode,
2906                                                    path_leaf_bh(left_path),
2907                                                    handle, split_rec, el,
2908                                                    split_index);
2909                         if (ret) {
2910                                 mlog_errno(ret);
2911                                 goto out;
2912                         }
2913                 } else {
2914                         ret = ocfs2_merge_rec_right(inode,
2915                                                     path_leaf_bh(left_path),
2916                                                     handle, split_rec, el,
2917                                                     split_index);
2918                         if (ret) {
2919                                 mlog_errno(ret);
2920                                 goto out;
2921                         }
2922                 }
2923
2924                 if (ctxt->c_split_covers_rec) {
2925                         /*
2926                          * The merge may have left an empty extent in
2927                          * our leaf. Try to rotate it away.
2928                          */
2929                         ret = ocfs2_rotate_tree_left(inode, handle, left_path,
2930                                                      dealloc);
2931                         if (ret)
2932                                 mlog_errno(ret);
2933                         ret = 0;
2934                 }
2935         }
2936
2937 out:
2938         return ret;
2939 }
2940
2941 static void ocfs2_subtract_from_rec(struct super_block *sb,
2942                                     enum ocfs2_split_type split,
2943                                     struct ocfs2_extent_rec *rec,
2944                                     struct ocfs2_extent_rec *split_rec)
2945 {
2946         u64 len_blocks;
2947
2948         len_blocks = ocfs2_clusters_to_blocks(sb,
2949                                 le16_to_cpu(split_rec->e_leaf_clusters));
2950
2951         if (split == SPLIT_LEFT) {
2952                 /*
2953                  * Region is on the left edge of the existing
2954                  * record.
2955                  */
2956                 le32_add_cpu(&rec->e_cpos,
2957                              le16_to_cpu(split_rec->e_leaf_clusters));
2958                 le64_add_cpu(&rec->e_blkno, len_blocks);
2959                 le16_add_cpu(&rec->e_leaf_clusters,
2960                              -le16_to_cpu(split_rec->e_leaf_clusters));
2961         } else {
2962                 /*
2963                  * Region is on the right edge of the existing
2964                  * record.
2965                  */
2966                 le16_add_cpu(&rec->e_leaf_clusters,
2967                              -le16_to_cpu(split_rec->e_leaf_clusters));
2968         }
2969 }
2970
2971 /*
2972  * Do the final bits of extent record insertion at the target leaf
2973  * list. If this leaf is part of an allocation tree, it is assumed
2974  * that the tree above has been prepared.
2975  */
2976 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
2977                                  struct ocfs2_extent_list *el,
2978                                  struct ocfs2_insert_type *insert,
2979                                  struct inode *inode)
2980 {
2981         int i = insert->ins_contig_index;
2982         unsigned int range;
2983         struct ocfs2_extent_rec *rec;
2984
2985         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
2986
2987         if (insert->ins_split != SPLIT_NONE) {
2988                 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
2989                 BUG_ON(i == -1);
2990                 rec = &el->l_recs[i];
2991                 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec,
2992                                         insert_rec);
2993                 goto rotate;
2994         }
2995
2996         /*
2997          * Contiguous insert - either left or right.
2998          */
2999         if (insert->ins_contig != CONTIG_NONE) {
3000                 rec = &el->l_recs[i];
3001                 if (insert->ins_contig == CONTIG_LEFT) {
3002                         rec->e_blkno = insert_rec->e_blkno;
3003                         rec->e_cpos = insert_rec->e_cpos;
3004                 }
3005                 le16_add_cpu(&rec->e_leaf_clusters,
3006                              le16_to_cpu(insert_rec->e_leaf_clusters));
3007                 return;
3008         }
3009
3010         /*
3011          * Handle insert into an empty leaf.
3012          */
3013         if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3014             ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3015              ocfs2_is_empty_extent(&el->l_recs[0]))) {
3016                 el->l_recs[0] = *insert_rec;
3017                 el->l_next_free_rec = cpu_to_le16(1);
3018                 return;
3019         }
3020
3021         /*
3022          * Appending insert.
3023          */
3024         if (insert->ins_appending == APPEND_TAIL) {
3025                 i = le16_to_cpu(el->l_next_free_rec) - 1;
3026                 rec = &el->l_recs[i];
3027                 range = le32_to_cpu(rec->e_cpos)
3028                         + le16_to_cpu(rec->e_leaf_clusters);
3029                 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3030
3031                 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3032                                 le16_to_cpu(el->l_count),
3033                                 "inode %lu, depth %u, count %u, next free %u, "
3034                                 "rec.cpos %u, rec.clusters %u, "
3035                                 "insert.cpos %u, insert.clusters %u\n",
3036                                 inode->i_ino,
3037                                 le16_to_cpu(el->l_tree_depth),
3038                                 le16_to_cpu(el->l_count),
3039                                 le16_to_cpu(el->l_next_free_rec),
3040                                 le32_to_cpu(el->l_recs[i].e_cpos),
3041                                 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3042                                 le32_to_cpu(insert_rec->e_cpos),
3043                                 le16_to_cpu(insert_rec->e_leaf_clusters));
3044                 i++;
3045                 el->l_recs[i] = *insert_rec;
3046                 le16_add_cpu(&el->l_next_free_rec, 1);
3047                 return;
3048         }
3049
3050 rotate:
3051         /*
3052          * Ok, we have to rotate.
3053          *
3054          * At this point, it is safe to assume that inserting into an
3055          * empty leaf and appending to a leaf have both been handled
3056          * above.
3057          *
3058          * This leaf needs to have space, either by the empty 1st
3059          * extent record, or by virtue of an l_next_rec < l_count.
3060          */
3061         ocfs2_rotate_leaf(el, insert_rec);
3062 }
3063
3064 static inline void ocfs2_update_dinode_clusters(struct inode *inode,
3065                                                 struct ocfs2_dinode *di,
3066                                                 u32 clusters)
3067 {
3068         le32_add_cpu(&di->i_clusters, clusters);
3069         spin_lock(&OCFS2_I(inode)->ip_lock);
3070         OCFS2_I(inode)->ip_clusters = le32_to_cpu(di->i_clusters);
3071         spin_unlock(&OCFS2_I(inode)->ip_lock);
3072 }
3073
3074 static void ocfs2_adjust_rightmost_records(struct inode *inode,
3075                                            handle_t *handle,
3076                                            struct ocfs2_path *path,
3077                                            struct ocfs2_extent_rec *insert_rec)
3078 {
3079         int ret, i, next_free;
3080         struct buffer_head *bh;
3081         struct ocfs2_extent_list *el;
3082         struct ocfs2_extent_rec *rec;
3083
3084         /*
3085          * Update everything except the leaf block.
3086          */
3087         for (i = 0; i < path->p_tree_depth; i++) {
3088                 bh = path->p_node[i].bh;
3089                 el = path->p_node[i].el;
3090
3091                 next_free = le16_to_cpu(el->l_next_free_rec);
3092                 if (next_free == 0) {
3093                         ocfs2_error(inode->i_sb,
3094                                     "Dinode %llu has a bad extent list",
3095                                     (unsigned long long)OCFS2_I(inode)->ip_blkno);
3096                         ret = -EIO;
3097                         return;
3098                 }
3099
3100                 rec = &el->l_recs[next_free - 1];
3101
3102                 rec->e_int_clusters = insert_rec->e_cpos;
3103                 le32_add_cpu(&rec->e_int_clusters,
3104                              le16_to_cpu(insert_rec->e_leaf_clusters));
3105                 le32_add_cpu(&rec->e_int_clusters,
3106                              -le32_to_cpu(rec->e_cpos));
3107
3108                 ret = ocfs2_journal_dirty(handle, bh);
3109                 if (ret)
3110                         mlog_errno(ret);
3111
3112         }
3113 }
3114
3115 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
3116                                     struct ocfs2_extent_rec *insert_rec,
3117                                     struct ocfs2_path *right_path,
3118                                     struct ocfs2_path **ret_left_path)
3119 {
3120         int ret, next_free;
3121         struct ocfs2_extent_list *el;
3122         struct ocfs2_path *left_path = NULL;
3123
3124         *ret_left_path = NULL;
3125
3126         /*
3127          * This shouldn't happen for non-trees. The extent rec cluster
3128          * count manipulation below only works for interior nodes.
3129          */
3130         BUG_ON(right_path->p_tree_depth == 0);
3131
3132         /*
3133          * If our appending insert is at the leftmost edge of a leaf,
3134          * then we might need to update the rightmost records of the
3135          * neighboring path.
3136          */
3137         el = path_leaf_el(right_path);
3138         next_free = le16_to_cpu(el->l_next_free_rec);
3139         if (next_free == 0 ||
3140             (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3141                 u32 left_cpos;
3142
3143                 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
3144                                                     &left_cpos);
3145                 if (ret) {
3146                         mlog_errno(ret);
3147                         goto out;
3148                 }
3149
3150                 mlog(0, "Append may need a left path update. cpos: %u, "
3151                      "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
3152                      left_cpos);
3153
3154                 /*
3155                  * No need to worry if the append is already in the
3156                  * leftmost leaf.
3157                  */
3158                 if (left_cpos) {
3159                         left_path = ocfs2_new_path(path_root_bh(right_path),
3160                                                    path_root_el(right_path));
3161                         if (!left_path) {
3162                                 ret = -ENOMEM;
3163                                 mlog_errno(ret);
3164                                 goto out;
3165                         }
3166
3167                         ret = ocfs2_find_path(inode, left_path, left_cpos);
3168                         if (ret) {
3169                                 mlog_errno(ret);
3170                                 goto out;
3171                         }
3172
3173                         /*
3174                          * ocfs2_insert_path() will pass the left_path to the
3175                          * journal for us.
3176                          */
3177                 }
3178         }
3179
3180         ret = ocfs2_journal_access_path(inode, handle, right_path);
3181         if (ret) {
3182                 mlog_errno(ret);
3183                 goto out;
3184         }
3185
3186         ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec);
3187
3188         *ret_left_path = left_path;
3189         ret = 0;
3190 out:
3191         if (ret != 0)
3192                 ocfs2_free_path(left_path);
3193
3194         return ret;
3195 }
3196
3197 static void ocfs2_split_record(struct inode *inode,
3198                                struct ocfs2_path *left_path,
3199                                struct ocfs2_path *right_path,
3200                                struct ocfs2_extent_rec *split_rec,
3201                                enum ocfs2_split_type split)
3202 {
3203         int index;
3204         u32 cpos = le32_to_cpu(split_rec->e_cpos);
3205         struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
3206         struct ocfs2_extent_rec *rec, *tmprec;
3207
3208         right_el = path_leaf_el(right_path);;
3209         if (left_path)
3210                 left_el = path_leaf_el(left_path);
3211
3212         el = right_el;
3213         insert_el = right_el;
3214         index = ocfs2_search_extent_list(el, cpos);
3215         if (index != -1) {
3216                 if (index == 0 && left_path) {
3217                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3218
3219                         /*
3220                          * This typically means that the record
3221                          * started in the left path but moved to the
3222                          * right as a result of rotation. We either
3223                          * move the existing record to the left, or we
3224                          * do the later insert there.
3225                          *
3226                          * In this case, the left path should always
3227                          * exist as the rotate code will have passed
3228                          * it back for a post-insert update.
3229                          */
3230
3231                         if (split == SPLIT_LEFT) {
3232                                 /*
3233                                  * It's a left split. Since we know
3234                                  * that the rotate code gave us an
3235                                  * empty extent in the left path, we
3236                                  * can just do the insert there.
3237                                  */
3238                                 insert_el = left_el;
3239                         } else {
3240                                 /*
3241                                  * Right split - we have to move the
3242                                  * existing record over to the left
3243                                  * leaf. The insert will be into the
3244                                  * newly created empty extent in the
3245                                  * right leaf.
3246                                  */
3247                                 tmprec = &right_el->l_recs[index];
3248                                 ocfs2_rotate_leaf(left_el, tmprec);
3249                                 el = left_el;
3250
3251                                 memset(tmprec, 0, sizeof(*tmprec));
3252                                 index = ocfs2_search_extent_list(left_el, cpos);
3253                                 BUG_ON(index == -1);
3254                         }
3255                 }
3256         } else {
3257                 BUG_ON(!left_path);
3258                 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
3259                 /*
3260                  * Left path is easy - we can just allow the insert to
3261                  * happen.
3262                  */
3263                 el = left_el;
3264                 insert_el = left_el;
3265                 index = ocfs2_search_extent_list(el, cpos);
3266                 BUG_ON(index == -1);
3267         }
3268
3269         rec = &el->l_recs[index];
3270         ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec);
3271         ocfs2_rotate_leaf(insert_el, split_rec);
3272 }
3273
3274 /*
3275  * This function only does inserts on an allocation b-tree. For dinode
3276  * lists, ocfs2_insert_at_leaf() is called directly.
3277  *
3278  * right_path is the path we want to do the actual insert
3279  * in. left_path should only be passed in if we need to update that
3280  * portion of the tree after an edge insert.
3281  */
3282 static int ocfs2_insert_path(struct inode *inode,
3283                              handle_t *handle,
3284                              struct ocfs2_path *left_path,
3285                              struct ocfs2_path *right_path,
3286                              struct ocfs2_extent_rec *insert_rec,
3287                              struct ocfs2_insert_type *insert)
3288 {
3289         int ret, subtree_index;
3290         struct buffer_head *leaf_bh = path_leaf_bh(right_path);
3291
3292         /*
3293          * Pass both paths to the journal. The majority of inserts
3294          * will be touching all components anyway.
3295          */
3296         ret = ocfs2_journal_access_path(inode, handle, right_path);
3297         if (ret < 0) {
3298                 mlog_errno(ret);
3299                 goto out;
3300         }
3301
3302         if (left_path) {
3303                 int credits = handle->h_buffer_credits;
3304
3305                 /*
3306                  * There's a chance that left_path got passed back to
3307                  * us without being accounted for in the
3308                  * journal. Extend our transaction here to be sure we
3309                  * can change those blocks.
3310                  */
3311                 credits += left_path->p_tree_depth;
3312
3313                 ret = ocfs2_extend_trans(handle, credits);
3314                 if (ret < 0) {
3315                         mlog_errno(ret);
3316                         goto out;
3317                 }
3318
3319                 ret = ocfs2_journal_access_path(inode, handle, left_path);
3320                 if (ret < 0) {
3321                         mlog_errno(ret);
3322                         goto out;
3323                 }
3324         }
3325
3326         if (insert->ins_split != SPLIT_NONE) {
3327                 /*
3328                  * We could call ocfs2_insert_at_leaf() for some types
3329                  * of splits, but it's easier to just let one seperate
3330                  * function sort it all out.
3331                  */
3332                 ocfs2_split_record(inode, left_path, right_path,
3333                                    insert_rec, insert->ins_split);
3334         } else
3335                 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
3336                                      insert, inode);
3337
3338         ret = ocfs2_journal_dirty(handle, leaf_bh);
3339         if (ret)
3340                 mlog_errno(ret);
3341
3342         if (left_path) {
3343                 /*
3344                  * The rotate code has indicated that we need to fix
3345                  * up portions of the tree after the insert.
3346                  *
3347                  * XXX: Should we extend the transaction here?
3348                  */
3349                 subtree_index = ocfs2_find_subtree_root(inode, left_path,
3350                                                         right_path);
3351                 ocfs2_complete_edge_insert(inode, handle, left_path,
3352                                            right_path, subtree_index);
3353         }
3354
3355         ret = 0;
3356 out:
3357         return ret;
3358 }
3359
3360 static int ocfs2_do_insert_extent(struct inode *inode,
3361                                   handle_t *handle,
3362                                   struct buffer_head *di_bh,
3363                                   struct ocfs2_extent_rec *insert_rec,
3364                                   struct ocfs2_insert_type *type)
3365 {
3366         int ret, rotate = 0;
3367         u32 cpos;
3368         struct ocfs2_path *right_path = NULL;
3369         struct ocfs2_path *left_path = NULL;
3370         struct ocfs2_dinode *di;
3371         struct ocfs2_extent_list *el;
3372
3373         di = (struct ocfs2_dinode *) di_bh->b_data;
3374         el = &di->id2.i_list;
3375
3376         ret = ocfs2_journal_access(handle, inode, di_bh,
3377                                    OCFS2_JOURNAL_ACCESS_WRITE);
3378         if (ret) {
3379                 mlog_errno(ret);
3380                 goto out;
3381         }
3382
3383         if (le16_to_cpu(el->l_tree_depth) == 0) {
3384                 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
3385                 goto out_update_clusters;
3386         }
3387
3388         right_path = ocfs2_new_inode_path(di_bh);
3389         if (!right_path) {
3390                 ret = -ENOMEM;
3391                 mlog_errno(ret);
3392                 goto out;
3393         }
3394
3395         /*
3396          * Determine the path to start with. Rotations need the
3397          * rightmost path, everything else can go directly to the
3398          * target leaf.
3399          */
3400         cpos = le32_to_cpu(insert_rec->e_cpos);
3401         if (type->ins_appending == APPEND_NONE &&
3402             type->ins_contig == CONTIG_NONE) {
3403                 rotate = 1;
3404                 cpos = UINT_MAX;
3405         }
3406
3407         ret = ocfs2_find_path(inode, right_path, cpos);
3408         if (ret) {
3409                 mlog_errno(ret);
3410                 goto out;
3411         }
3412
3413         /*
3414          * Rotations and appends need special treatment - they modify
3415          * parts of the tree's above them.
3416          *
3417          * Both might pass back a path immediate to the left of the
3418          * one being inserted to. This will be cause
3419          * ocfs2_insert_path() to modify the rightmost records of
3420          * left_path to account for an edge insert.
3421          *
3422          * XXX: When modifying this code, keep in mind that an insert
3423          * can wind up skipping both of these two special cases...
3424          */
3425         if (rotate) {
3426                 ret = ocfs2_rotate_tree_right(inode, handle, type->ins_split,
3427                                               le32_to_cpu(insert_rec->e_cpos),
3428                                               right_path, &left_path);
3429                 if (ret) {
3430                         mlog_errno(ret);
3431                         goto out;
3432                 }
3433         } else if (type->ins_appending == APPEND_TAIL
3434                    && type->ins_contig != CONTIG_LEFT) {
3435                 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
3436                                                right_path, &left_path);
3437                 if (ret) {
3438                         mlog_errno(ret);
3439                         goto out;
3440                 }
3441         }
3442
3443         ret = ocfs2_insert_path(inode, handle, left_path, right_path,
3444                                 insert_rec, type);
3445         if (ret) {
3446                 mlog_errno(ret);
3447                 goto out;
3448         }
3449
3450 out_update_clusters:
3451         if (type->ins_split == SPLIT_NONE)
3452                 ocfs2_update_dinode_clusters(inode, di,
3453                                              le16_to_cpu(insert_rec->e_leaf_clusters));
3454
3455         ret = ocfs2_journal_dirty(handle, di_bh);
3456         if (ret)
3457                 mlog_errno(ret);
3458
3459 out:
3460         ocfs2_free_path(left_path);
3461         ocfs2_free_path(right_path);
3462
3463         return ret;
3464 }
3465
3466 static enum ocfs2_contig_type
3467 ocfs2_figure_merge_contig_type(struct inode *inode,
3468                                struct ocfs2_extent_list *el, int index,
3469                                struct ocfs2_extent_rec *split_rec)
3470 {
3471         struct ocfs2_extent_rec *rec;
3472         enum ocfs2_contig_type ret = CONTIG_NONE;
3473
3474         /*
3475          * We're careful to check for an empty extent record here -
3476          * the merge code will know what to do if it sees one.
3477          */
3478
3479         if (index > 0) {
3480                 rec = &el->l_recs[index - 1];
3481                 if (index == 1 && ocfs2_is_empty_extent(rec)) {
3482                         if (split_rec->e_cpos == el->l_recs[index].e_cpos)
3483                                 ret = CONTIG_RIGHT;
3484                 } else {
3485                         ret = ocfs2_extent_contig(inode, rec, split_rec);
3486                 }
3487         }
3488
3489         if (index < (le16_to_cpu(el->l_next_free_rec) - 1)) {
3490                 enum ocfs2_contig_type contig_type;
3491
3492                 rec = &el->l_recs[index + 1];
3493                 contig_type = ocfs2_extent_contig(inode, rec, split_rec);
3494
3495                 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
3496                         ret = CONTIG_LEFTRIGHT;
3497                 else if (ret == CONTIG_NONE)
3498                         ret = contig_type;
3499         }
3500
3501         return ret;
3502 }
3503
3504 static void ocfs2_figure_contig_type(struct inode *inode,
3505                                      struct ocfs2_insert_type *insert,
3506                                      struct ocfs2_extent_list *el,
3507                                      struct ocfs2_extent_rec *insert_rec)
3508 {
3509         int i;
3510         enum ocfs2_contig_type contig_type = CONTIG_NONE;
3511
3512         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3513
3514         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
3515                 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
3516                                                   insert_rec);
3517                 if (contig_type != CONTIG_NONE) {
3518                         insert->ins_contig_index = i;
3519                         break;
3520                 }
3521         }
3522         insert->ins_contig = contig_type;
3523 }
3524
3525 /*
3526  * This should only be called against the righmost leaf extent list.
3527  *
3528  * ocfs2_figure_appending_type() will figure out whether we'll have to
3529  * insert at the tail of the rightmost leaf.
3530  *
3531  * This should also work against the dinode list for tree's with 0
3532  * depth. If we consider the dinode list to be the rightmost leaf node
3533  * then the logic here makes sense.
3534  */
3535 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
3536                                         struct ocfs2_extent_list *el,
3537                                         struct ocfs2_extent_rec *insert_rec)
3538 {
3539         int i;
3540         u32 cpos = le32_to_cpu(insert_rec->e_cpos);
3541         struct ocfs2_extent_rec *rec;
3542
3543         insert->ins_appending = APPEND_NONE;
3544
3545         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3546
3547         if (!el->l_next_free_rec)
3548                 goto set_tail_append;
3549
3550         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
3551                 /* Were all records empty? */
3552                 if (le16_to_cpu(el->l_next_free_rec) == 1)
3553                         goto set_tail_append;
3554         }
3555
3556         i = le16_to_cpu(el->l_next_free_rec) - 1;
3557         rec = &el->l_recs[i];
3558
3559         if (cpos >=
3560             (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
3561                 goto set_tail_append;
3562
3563         return;
3564
3565 set_tail_append:
3566         insert->ins_appending = APPEND_TAIL;
3567 }
3568
3569 /*
3570  * Helper function called at the begining of an insert.
3571  *
3572  * This computes a few things that are commonly used in the process of
3573  * inserting into the btree:
3574  *   - Whether the new extent is contiguous with an existing one.
3575  *   - The current tree depth.
3576  *   - Whether the insert is an appending one.
3577  *   - The total # of free records in the tree.
3578  *
3579  * All of the information is stored on the ocfs2_insert_type
3580  * structure.
3581  */
3582 static int ocfs2_figure_insert_type(struct inode *inode,
3583                                     struct buffer_head *di_bh,
3584                                     struct buffer_head **last_eb_bh,
3585                                     struct ocfs2_extent_rec *insert_rec,
3586                                     int *free_records,
3587                                     struct ocfs2_insert_type *insert)
3588 {
3589         int ret;
3590         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3591         struct ocfs2_extent_block *eb;
3592         struct ocfs2_extent_list *el;
3593         struct ocfs2_path *path = NULL;
3594         struct buffer_head *bh = NULL;
3595
3596         insert->ins_split = SPLIT_NONE;
3597
3598         el = &di->id2.i_list;
3599         insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
3600
3601         if (el->l_tree_depth) {
3602                 /*
3603                  * If we have tree depth, we read in the
3604                  * rightmost extent block ahead of time as
3605                  * ocfs2_figure_insert_type() and ocfs2_add_branch()
3606                  * may want it later.
3607                  */
3608                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
3609                                        le64_to_cpu(di->i_last_eb_blk), &bh,
3610                                        OCFS2_BH_CACHED, inode);
3611                 if (ret) {
3612                         mlog_exit(ret);
3613                         goto out;
3614                 }
3615                 eb = (struct ocfs2_extent_block *) bh->b_data;
3616                 el = &eb->h_list;
3617         }
3618
3619         /*
3620          * Unless we have a contiguous insert, we'll need to know if
3621          * there is room left in our allocation tree for another
3622          * extent record.
3623          *
3624          * XXX: This test is simplistic, we can search for empty
3625          * extent records too.
3626          */
3627         *free_records = le16_to_cpu(el->l_count) -
3628                 le16_to_cpu(el->l_next_free_rec);
3629
3630         if (!insert->ins_tree_depth) {
3631                 ocfs2_figure_contig_type(inode, insert, el, insert_rec);
3632                 ocfs2_figure_appending_type(insert, el, insert_rec);
3633                 return 0;
3634         }
3635
3636         path = ocfs2_new_inode_path(di_bh);
3637         if (!path) {
3638                 ret = -ENOMEM;
3639                 mlog_errno(ret);
3640                 goto out;
3641         }
3642
3643         /*
3644          * In the case that we're inserting past what the tree
3645          * currently accounts for, ocfs2_find_path() will return for
3646          * us the rightmost tree path. This is accounted for below in
3647          * the appending code.
3648          */
3649         ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
3650         if (ret) {
3651                 mlog_errno(ret);
3652                 goto out;
3653         }
3654
3655         el = path_leaf_el(path);
3656
3657         /*
3658          * Now that we have the path, there's two things we want to determine:
3659          * 1) Contiguousness (also set contig_index if this is so)
3660          *
3661          * 2) Are we doing an append? We can trivially break this up
3662          *     into two types of appends: simple record append, or a
3663          *     rotate inside the tail leaf.
3664          */
3665         ocfs2_figure_contig_type(inode, insert, el, insert_rec);
3666
3667         /*
3668          * The insert code isn't quite ready to deal with all cases of
3669          * left contiguousness. Specifically, if it's an insert into
3670          * the 1st record in a leaf, it will require the adjustment of
3671          * cluster count on the last record of the path directly to it's
3672          * left. For now, just catch that case and fool the layers
3673          * above us. This works just fine for tree_depth == 0, which
3674          * is why we allow that above.
3675          */
3676         if (insert->ins_contig == CONTIG_LEFT &&
3677             insert->ins_contig_index == 0)
3678                 insert->ins_contig = CONTIG_NONE;
3679
3680         /*
3681          * Ok, so we can simply compare against last_eb to figure out
3682          * whether the path doesn't exist. This will only happen in
3683          * the case that we're doing a tail append, so maybe we can
3684          * take advantage of that information somehow.
3685          */
3686         if (le64_to_cpu(di->i_last_eb_blk) == path_leaf_bh(path)->b_blocknr) {
3687                 /*
3688                  * Ok, ocfs2_find_path() returned us the rightmost
3689                  * tree path. This might be an appending insert. There are
3690                  * two cases:
3691                  *    1) We're doing a true append at the tail:
3692                  *      -This might even be off the end of the leaf
3693                  *    2) We're "appending" by rotating in the tail
3694                  */
3695                 ocfs2_figure_appending_type(insert, el, insert_rec);
3696         }
3697
3698 out:
3699         ocfs2_free_path(path);
3700
3701         if (ret == 0)
3702                 *last_eb_bh = bh;
3703         else
3704                 brelse(bh);
3705         return ret;
3706 }
3707
3708 /*
3709  * Insert an extent into an inode btree.
3710  *
3711  * The caller needs to update fe->i_clusters
3712  */
3713 int ocfs2_insert_extent(struct ocfs2_super *osb,
3714                         handle_t *handle,
3715                         struct inode *inode,
3716                         struct buffer_head *fe_bh,
3717                         u32 cpos,
3718                         u64 start_blk,
3719                         u32 new_clusters,
3720                         u8 flags,
3721                         struct ocfs2_alloc_context *meta_ac)
3722 {
3723         int status;
3724         int uninitialized_var(free_records);
3725         struct buffer_head *last_eb_bh = NULL;
3726         struct ocfs2_insert_type insert = {0, };
3727         struct ocfs2_extent_rec rec;
3728
3729         BUG_ON(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL);
3730
3731         mlog(0, "add %u clusters at position %u to inode %llu\n",
3732              new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
3733
3734         mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
3735                         (OCFS2_I(inode)->ip_clusters != cpos),
3736                         "Device %s, asking for sparse allocation: inode %llu, "
3737                         "cpos %u, clusters %u\n",
3738                         osb->dev_str,
3739                         (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos,
3740                         OCFS2_I(inode)->ip_clusters);
3741
3742         memset(&rec, 0, sizeof(rec));
3743         rec.e_cpos = cpu_to_le32(cpos);
3744         rec.e_blkno = cpu_to_le64(start_blk);
3745         rec.e_leaf_clusters = cpu_to_le16(new_clusters);
3746         rec.e_flags = flags;
3747
3748         status = ocfs2_figure_insert_type(inode, fe_bh, &last_eb_bh, &rec,
3749                                           &free_records, &insert);
3750         if (status < 0) {
3751                 mlog_errno(status);
3752                 goto bail;
3753         }
3754
3755         mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
3756              "Insert.contig_index: %d, Insert.free_records: %d, "
3757              "Insert.tree_depth: %d\n",
3758              insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
3759              free_records, insert.ins_tree_depth);
3760
3761         if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
3762                 status = ocfs2_grow_tree(inode, handle, fe_bh,
3763                                          &insert.ins_tree_depth, &last_eb_bh,
3764                                          meta_ac);
3765                 if (status) {
3766                         mlog_errno(status);
3767                         goto bail;
3768                 }
3769         }
3770
3771         /* Finally, we can add clusters. This might rotate the tree for us. */
3772         status = ocfs2_do_insert_extent(inode, handle, fe_bh, &rec, &insert);
3773         if (status < 0)
3774                 mlog_errno(status);
3775         else
3776                 ocfs2_extent_map_insert_rec(inode, &rec);
3777
3778 bail:
3779         if (last_eb_bh)
3780                 brelse(last_eb_bh);
3781
3782         mlog_exit(status);
3783         return status;
3784 }
3785
3786 static void ocfs2_make_right_split_rec(struct super_block *sb,
3787                                        struct ocfs2_extent_rec *split_rec,
3788                                        u32 cpos,
3789                                        struct ocfs2_extent_rec *rec)
3790 {
3791         u32 rec_cpos = le32_to_cpu(rec->e_cpos);
3792         u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
3793
3794         memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
3795
3796         split_rec->e_cpos = cpu_to_le32(cpos);
3797         split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
3798
3799         split_rec->e_blkno = rec->e_blkno;
3800         le64_add_cpu(&split_rec->e_blkno,
3801                      ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
3802
3803         split_rec->e_flags = rec->e_flags;
3804 }
3805
3806 static int ocfs2_split_and_insert(struct inode *inode,
3807                                   handle_t *handle,
3808                                   struct ocfs2_path *path,
3809                                   struct buffer_head *di_bh,
3810                                   struct buffer_head **last_eb_bh,
3811                                   int split_index,
3812                                   struct ocfs2_extent_rec *orig_split_rec,
3813                                   struct ocfs2_alloc_context *meta_ac)
3814 {
3815         int ret = 0, depth;
3816         unsigned int insert_range, rec_range, do_leftright = 0;
3817         struct ocfs2_extent_rec tmprec;
3818         struct ocfs2_extent_list *rightmost_el;
3819         struct ocfs2_extent_rec rec;
3820         struct ocfs2_extent_rec split_rec = *orig_split_rec;
3821         struct ocfs2_insert_type insert;
3822         struct ocfs2_extent_block *eb;
3823         struct ocfs2_dinode *di;
3824
3825 leftright:
3826         /*
3827          * Store a copy of the record on the stack - it might move
3828          * around as the tree is manipulated below.
3829          */
3830         rec = path_leaf_el(path)->l_recs[split_index];
3831
3832         di = (struct ocfs2_dinode *)di_bh->b_data;
3833         rightmost_el = &di->id2.i_list;
3834
3835         depth = le16_to_cpu(rightmost_el->l_tree_depth);
3836         if (depth) {
3837                 BUG_ON(!(*last_eb_bh));
3838                 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
3839                 rightmost_el = &eb->h_list;
3840         }
3841
3842         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
3843             le16_to_cpu(rightmost_el->l_count)) {
3844                 ret = ocfs2_grow_tree(inode, handle, di_bh, &depth, last_eb_bh,
3845                                       meta_ac);
3846                 if (ret) {
3847                         mlog_errno(ret);
3848                         goto out;
3849                 }
3850         }
3851
3852         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
3853         insert.ins_appending = APPEND_NONE;
3854         insert.ins_contig = CONTIG_NONE;
3855         insert.ins_tree_depth = depth;
3856
3857         insert_range = le32_to_cpu(split_rec.e_cpos) +
3858                 le16_to_cpu(split_rec.e_leaf_clusters);
3859         rec_range = le32_to_cpu(rec.e_cpos) +
3860                 le16_to_cpu(rec.e_leaf_clusters);
3861
3862         if (split_rec.e_cpos == rec.e_cpos) {
3863                 insert.ins_split = SPLIT_LEFT;
3864         } else if (insert_range == rec_range) {
3865                 insert.ins_split = SPLIT_RIGHT;
3866         } else {
3867                 /*
3868                  * Left/right split. We fake this as a right split
3869                  * first and then make a second pass as a left split.
3870                  */
3871                 insert.ins_split = SPLIT_RIGHT;
3872
3873                 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range,
3874                                            &rec);
3875
3876                 split_rec = tmprec;
3877
3878                 BUG_ON(do_leftright);
3879                 do_leftright = 1;
3880         }
3881
3882         ret = ocfs2_do_insert_extent(inode, handle, di_bh, &split_rec,
3883                                      &insert);
3884         if (ret) {
3885                 mlog_errno(ret);
3886                 goto out;
3887         }
3888
3889         if (do_leftright == 1) {
3890                 u32 cpos;
3891                 struct ocfs2_extent_list *el;
3892
3893                 do_leftright++;
3894                 split_rec = *orig_split_rec;
3895
3896                 ocfs2_reinit_path(path, 1);
3897
3898                 cpos = le32_to_cpu(split_rec.e_cpos);
3899                 ret = ocfs2_find_path(inode, path, cpos);
3900                 if (ret) {
3901                         mlog_errno(ret);
3902                         goto out;
3903                 }
3904
3905                 el = path_leaf_el(path);
3906                 split_index = ocfs2_search_extent_list(el, cpos);
3907                 goto leftright;
3908         }
3909 out:
3910
3911         return ret;
3912 }
3913
3914 /*
3915  * Mark part or all of the extent record at split_index in the leaf
3916  * pointed to by path as written. This removes the unwritten
3917  * extent flag.
3918  *
3919  * Care is taken to handle contiguousness so as to not grow the tree.
3920  *
3921  * meta_ac is not strictly necessary - we only truly need it if growth
3922  * of the tree is required. All other cases will degrade into a less
3923  * optimal tree layout.
3924  *
3925  * last_eb_bh should be the rightmost leaf block for any inode with a
3926  * btree. Since a split may grow the tree or a merge might shrink it, the caller cannot trust the contents of that buffer after this call.
3927  *
3928  * This code is optimized for readability - several passes might be
3929  * made over certain portions of the tree. All of those blocks will
3930  * have been brought into cache (and pinned via the journal), so the
3931  * extra overhead is not expressed in terms of disk reads.
3932  */
3933 static int __ocfs2_mark_extent_written(struct inode *inode,
3934                                        struct buffer_head *di_bh,
3935                                        handle_t *handle,
3936                                        struct ocfs2_path *path,
3937                                        int split_index,
3938                                        struct ocfs2_extent_rec *split_rec,
3939                                        struct ocfs2_alloc_context *meta_ac,
3940                                        struct ocfs2_cached_dealloc_ctxt *dealloc)
3941 {
3942         int ret = 0;
3943         struct ocfs2_extent_list *el = path_leaf_el(path);
3944         struct buffer_head *eb_bh, *last_eb_bh = NULL;
3945         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3946         struct ocfs2_merge_ctxt ctxt;
3947         struct ocfs2_extent_list *rightmost_el;
3948
3949         if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) {
3950                 ret = -EIO;
3951                 mlog_errno(ret);
3952                 goto out;
3953         }
3954
3955         if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
3956             ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
3957              (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
3958                 ret = -EIO;
3959                 mlog_errno(ret);
3960                 goto out;
3961         }
3962
3963         eb_bh = path_leaf_bh(path);
3964         ret = ocfs2_journal_access(handle, inode, eb_bh,
3965                                    OCFS2_JOURNAL_ACCESS_WRITE);
3966         if (ret) {
3967                 mlog_errno(ret);
3968                 goto out;
3969         }
3970
3971         ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, el,
3972                                                             split_index,
3973                                                             split_rec);
3974
3975         /*
3976          * The core merge / split code wants to know how much room is
3977          * left in this inodes allocation tree, so we pass the
3978          * rightmost extent list.
3979          */
3980         if (path->p_tree_depth) {
3981                 struct ocfs2_extent_block *eb;
3982                 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3983
3984                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
3985                                        le64_to_cpu(di->i_last_eb_blk),
3986                                        &last_eb_bh, OCFS2_BH_CACHED, inode);
3987                 if (ret) {
3988                         mlog_exit(ret);
3989                         goto out;
3990                 }
3991
3992                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
3993                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
3994                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
3995                         ret = -EROFS;
3996                         goto out;
3997                 }
3998
3999                 rightmost_el = &eb->h_list;
4000         } else
4001                 rightmost_el = path_root_el(path);
4002
4003         if (rec->e_cpos == split_rec->e_cpos &&
4004             rec->e_leaf_clusters == split_rec->e_leaf_clusters)
4005                 ctxt.c_split_covers_rec = 1;
4006         else
4007                 ctxt.c_split_covers_rec = 0;
4008
4009         ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
4010
4011         mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n",
4012              split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent,
4013              ctxt.c_split_covers_rec);
4014
4015         if (ctxt.c_contig_type == CONTIG_NONE) {
4016                 if (ctxt.c_split_covers_rec)
4017                         el->l_recs[split_index] = *split_rec;
4018                 else
4019                         ret = ocfs2_split_and_insert(inode, handle, path, di_bh,
4020                                                      &last_eb_bh, split_index,
4021                                                      split_rec, meta_ac);
4022                 if (ret)
4023                         mlog_errno(ret);
4024         } else {
4025                 ret = ocfs2_try_to_merge_extent(inode, handle, path,
4026                                                 split_index, split_rec,
4027                                                 dealloc, &ctxt);
4028                 if (ret)
4029                         mlog_errno(ret);
4030         }
4031
4032         ocfs2_journal_dirty(handle, eb_bh);
4033
4034 out:
4035         brelse(last_eb_bh);
4036         return ret;
4037 }
4038
4039 /*
4040  * Mark the already-existing extent at cpos as written for len clusters.
4041  *
4042  * If the existing extent is larger than the request, initiate a
4043  * split. An attempt will be made at merging with adjacent extents.
4044  *
4045  * The caller is responsible for passing down meta_ac if we'll need it.
4046  */
4047 int ocfs2_mark_extent_written(struct inode *inode, struct buffer_head *di_bh,
4048                               handle_t *handle, u32 cpos, u32 len, u32 phys,
4049                               struct ocfs2_alloc_context *meta_ac,
4050                               struct ocfs2_cached_dealloc_ctxt *dealloc)
4051 {
4052         int ret, index;
4053         u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys);
4054         struct ocfs2_extent_rec split_rec;
4055         struct ocfs2_path *left_path = NULL;
4056         struct ocfs2_extent_list *el;
4057
4058         mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n",
4059              inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno);
4060
4061         if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
4062                 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
4063                             "that are being written to, but the feature bit "
4064                             "is not set in the super block.",
4065                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
4066                 ret = -EROFS;
4067                 goto out;
4068         }
4069
4070         /*
4071          * XXX: This should be fixed up so that we just re-insert the
4072          * next extent records.
4073          */
4074         ocfs2_extent_map_trunc(inode, 0);
4075
4076         left_path = ocfs2_new_inode_path(di_bh);
4077         if (!left_path) {
4078                 ret = -ENOMEM;
4079                 mlog_errno(ret);
4080                 goto out;
4081         }
4082
4083         ret = ocfs2_find_path(inode, left_path, cpos);
4084         if (ret) {
4085                 mlog_errno(ret);
4086                 goto out;
4087         }
4088         el = path_leaf_el(left_path);
4089
4090         index = ocfs2_search_extent_list(el, cpos);
4091         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4092                 ocfs2_error(inode->i_sb,
4093                             "Inode %llu has an extent at cpos %u which can no "
4094                             "longer be found.\n",
4095                             (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
4096                 ret = -EROFS;
4097                 goto out;
4098         }
4099
4100         memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
4101         split_rec.e_cpos = cpu_to_le32(cpos);
4102         split_rec.e_leaf_clusters = cpu_to_le16(len);
4103         split_rec.e_blkno = cpu_to_le64(start_blkno);
4104         split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags;
4105         split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN;
4106
4107         ret = __ocfs2_mark_extent_written(inode, di_bh, handle, left_path,
4108                                           index, &split_rec, meta_ac, dealloc);
4109         if (ret)
4110                 mlog_errno(ret);
4111
4112 out:
4113         ocfs2_free_path(left_path);
4114         return ret;
4115 }
4116
4117 static int ocfs2_split_tree(struct inode *inode, struct buffer_head *di_bh,
4118                             handle_t *handle, struct ocfs2_path *path,
4119                             int index, u32 new_range,
4120                             struct ocfs2_alloc_context *meta_ac)
4121 {
4122         int ret, depth, credits = handle->h_buffer_credits;
4123         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4124         struct buffer_head *last_eb_bh = NULL;
4125         struct ocfs2_extent_block *eb;
4126         struct ocfs2_extent_list *rightmost_el, *el;
4127         struct ocfs2_extent_rec split_rec;
4128         struct ocfs2_extent_rec *rec;
4129         struct ocfs2_insert_type insert;
4130
4131         /*
4132          * Setup the record to split before we grow the tree.
4133          */
4134         el = path_leaf_el(path);
4135         rec = &el->l_recs[index];
4136         ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec);
4137
4138         depth = path->p_tree_depth;
4139         if (depth > 0) {
4140                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
4141                                        le64_to_cpu(di->i_last_eb_blk),
4142                                        &last_eb_bh, OCFS2_BH_CACHED, inode);
4143                 if (ret < 0) {
4144                         mlog_errno(ret);
4145                         goto out;
4146                 }
4147
4148                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
4149                 rightmost_el = &eb->h_list;
4150         } else
4151                 rightmost_el = path_leaf_el(path);
4152
4153         credits += path->p_tree_depth + ocfs2_extend_meta_needed(di);
4154         ret = ocfs2_extend_trans(handle, credits);
4155         if (ret) {
4156                 mlog_errno(ret);
4157                 goto out;
4158         }
4159
4160         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4161             le16_to_cpu(rightmost_el->l_count)) {
4162                 ret = ocfs2_grow_tree(inode, handle, di_bh, &depth, &last_eb_bh,
4163                                       meta_ac);
4164                 if (ret) {
4165                         mlog_errno(ret);
4166                         goto out;
4167                 }
4168         }
4169
4170         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4171         insert.ins_appending = APPEND_NONE;
4172         insert.ins_contig = CONTIG_NONE;
4173         insert.ins_split = SPLIT_RIGHT;
4174         insert.ins_tree_depth = depth;
4175
4176         ret = ocfs2_do_insert_extent(inode, handle, di_bh, &split_rec, &insert);
4177         if (ret)
4178                 mlog_errno(ret);
4179
4180 out:
4181         brelse(last_eb_bh);
4182         return ret;
4183 }
4184
4185 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle,
4186                               struct ocfs2_path *path, int index,
4187                               struct ocfs2_cached_dealloc_ctxt *dealloc,
4188                               u32 cpos, u32 len)
4189 {
4190         int ret;
4191         u32 left_cpos, rec_range, trunc_range;
4192         int wants_rotate = 0, is_rightmost_tree_rec = 0;
4193         struct super_block *sb = inode->i_sb;
4194         struct ocfs2_path *left_path = NULL;
4195         struct ocfs2_extent_list *el = path_leaf_el(path);
4196         struct ocfs2_extent_rec *rec;
4197         struct ocfs2_extent_block *eb;
4198
4199         if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
4200                 ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc);
4201                 if (ret) {
4202                         mlog_errno(ret);
4203                         goto out;
4204                 }
4205
4206                 index--;
4207         }
4208
4209         if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
4210             path->p_tree_depth) {
4211                 /*
4212                  * Check whether this is the rightmost tree record. If
4213                  * we remove all of this record or part of its right
4214                  * edge then an update of the record lengths above it
4215                  * will be required.
4216                  */
4217                 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
4218                 if (eb->h_next_leaf_blk == 0)
4219                         is_rightmost_tree_rec = 1;
4220         }
4221
4222         rec = &el->l_recs[index];
4223         if (index == 0 && path->p_tree_depth &&
4224             le32_to_cpu(rec->e_cpos) == cpos) {
4225                 /*
4226                  * Changing the leftmost offset (via partial or whole
4227                  * record truncate) of an interior (or rightmost) path
4228                  * means we have to update the subtree that is formed
4229                  * by this leaf and the one to it's left.
4230                  *
4231                  * There are two cases we can skip:
4232                  *   1) Path is the leftmost one in our inode tree.
4233                  *   2) The leaf is rightmost and will be empty after
4234                  *      we remove the extent record - the rotate code
4235                  *      knows how to update the newly formed edge.
4236                  */
4237
4238                 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path,
4239                                                     &left_cpos);
4240                 if (ret) {
4241                         mlog_errno(ret);
4242                         goto out;
4243                 }
4244
4245                 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
4246                         left_path = ocfs2_new_path(path_root_bh(path),
4247                                                    path_root_el(path));
4248                         if (!left_path) {
4249                                 ret = -ENOMEM;
4250                                 mlog_errno(ret);
4251                                 goto out;
4252                         }
4253
4254                         ret = ocfs2_find_path(inode, left_path, left_cpos);
4255                         if (ret) {
4256                                 mlog_errno(ret);
4257                                 goto out;
4258                         }
4259                 }
4260         }
4261
4262         ret = ocfs2_extend_rotate_transaction(handle, 0,
4263                                               handle->h_buffer_credits,
4264                                               path);
4265         if (ret) {
4266                 mlog_errno(ret);
4267                 goto out;
4268         }
4269
4270         ret = ocfs2_journal_access_path(inode, handle, path);
4271         if (ret) {
4272                 mlog_errno(ret);
4273                 goto out;
4274         }
4275
4276         ret = ocfs2_journal_access_path(inode, handle, left_path);
4277         if (ret) {
4278                 mlog_errno(ret);
4279                 goto out;
4280         }
4281
4282         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
4283         trunc_range = cpos + len;
4284
4285         if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
4286                 int next_free;
4287
4288                 memset(rec, 0, sizeof(*rec));
4289                 ocfs2_cleanup_merge(el, index);
4290                 wants_rotate = 1;
4291
4292                 next_free = le16_to_cpu(el->l_next_free_rec);
4293                 if (is_rightmost_tree_rec && next_free > 1) {
4294                         /*
4295                          * We skip the edge update if this path will
4296                          * be deleted by the rotate code.
4297                          */
4298                         rec = &el->l_recs[next_free - 1];
4299                         ocfs2_adjust_rightmost_records(inode, handle, path,
4300                                                        rec);
4301                 }
4302         } else if (le32_to_cpu(rec->e_cpos) == cpos) {
4303                 /* Remove leftmost portion of the record. */
4304                 le32_add_cpu(&rec->e_cpos, len);
4305                 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
4306                 le16_add_cpu(&rec->e_leaf_clusters, -len);
4307         } else if (rec_range == trunc_range) {
4308                 /* Remove rightmost portion of the record */
4309                 le16_add_cpu(&rec->e_leaf_clusters, -len);
4310                 if (is_rightmost_tree_rec)
4311                         ocfs2_adjust_rightmost_records(inode, handle, path, rec);
4312         } else {
4313                 /* Caller should have trapped this. */
4314                 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) "
4315                      "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno,
4316                      le32_to_cpu(rec->e_cpos),
4317                      le16_to_cpu(rec->e_leaf_clusters), cpos, len);
4318                 BUG();
4319         }
4320
4321         if (left_path) {
4322                 int subtree_index;
4323
4324                 subtree_index = ocfs2_find_subtree_root(inode, left_path, path);
4325                 ocfs2_complete_edge_insert(inode, handle, left_path, path,
4326                                            subtree_index);
4327         }
4328
4329         ocfs2_journal_dirty(handle, path_leaf_bh(path));
4330
4331         ret = ocfs2_rotate_tree_left(inode, handle, path, dealloc);
4332         if (ret) {
4333                 mlog_errno(ret);
4334                 goto out;
4335         }
4336
4337 out:
4338         ocfs2_free_path(left_path);
4339         return ret;
4340 }
4341
4342 int ocfs2_remove_extent(struct inode *inode, struct buffer_head *di_bh,
4343                         u32 cpos, u32 len, handle_t *handle,
4344                         struct ocfs2_alloc_context *meta_ac,
4345                         struct ocfs2_cached_dealloc_ctxt *dealloc)
4346 {
4347         int ret, index;
4348         u32 rec_range, trunc_range;
4349         struct ocfs2_extent_rec *rec;
4350         struct ocfs2_extent_list *el;
4351         struct ocfs2_path *path;
4352
4353         ocfs2_extent_map_trunc(inode, 0);
4354
4355         path = ocfs2_new_inode_path(di_bh);
4356         if (!path) {
4357                 ret = -ENOMEM;
4358                 mlog_errno(ret);
4359                 goto out;
4360         }
4361
4362         ret = ocfs2_find_path(inode, path, cpos);
4363         if (ret) {
4364                 mlog_errno(ret);
4365                 goto out;
4366         }
4367
4368         el = path_leaf_el(path);
4369         index = ocfs2_search_extent_list(el, cpos);
4370         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4371                 ocfs2_error(inode->i_sb,
4372                             "Inode %llu has an extent at cpos %u which can no "
4373                             "longer be found.\n",
4374                             (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
4375                 ret = -EROFS;
4376                 goto out;
4377         }
4378
4379         /*
4380          * We have 3 cases of extent removal:
4381          *   1) Range covers the entire extent rec
4382          *   2) Range begins or ends on one edge of the extent rec
4383          *   3) Range is in the middle of the extent rec (no shared edges)
4384          *
4385          * For case 1 we remove the extent rec and left rotate to
4386          * fill the hole.
4387          *
4388          * For case 2 we just shrink the existing extent rec, with a
4389          * tree update if the shrinking edge is also the edge of an
4390          * extent block.
4391          *
4392          * For case 3 we do a right split to turn the extent rec into
4393          * something case 2 can handle.
4394          */
4395         rec = &el->l_recs[index];
4396         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
4397         trunc_range = cpos + len;
4398
4399         BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
4400
4401         mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d "
4402              "(cpos %u, len %u)\n",
4403              (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index,
4404              le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec));
4405
4406         if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
4407                 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
4408                                          cpos, len);
4409                 if (ret) {
4410                         mlog_errno(ret);
4411                         goto out;
4412                 }
4413         } else {
4414                 ret = ocfs2_split_tree(inode, di_bh, handle, path, index,
4415                                        trunc_range, meta_ac);
4416                 if (ret) {
4417                         mlog_errno(ret);
4418                         goto out;
4419                 }
4420
4421                 /*
4422                  * The split could have manipulated the tree enough to
4423                  * move the record location, so we have to look for it again.
4424                  */
4425                 ocfs2_reinit_path(path, 1);
4426
4427                 ret = ocfs2_find_path(inode, path, cpos);
4428                 if (ret) {
4429                         mlog_errno(ret);
4430                         goto out;
4431                 }
4432
4433                 el = path_leaf_el(path);
4434                 index = ocfs2_search_extent_list(el, cpos);
4435                 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
4436                         ocfs2_error(inode->i_sb,
4437                                     "Inode %llu: split at cpos %u lost record.",
4438                                     (unsigned long long)OCFS2_I(inode)->ip_blkno,
4439                                     cpos);
4440                         ret = -EROFS;
4441                         goto out;
4442                 }
4443
4444                 /*
4445                  * Double check our values here. If anything is fishy,
4446                  * it's easier to catch it at the top level.
4447                  */
4448                 rec = &el->l_recs[index];
4449                 rec_range = le32_to_cpu(rec->e_cpos) +
4450                         ocfs2_rec_clusters(el, rec);
4451                 if (rec_range != trunc_range) {
4452                         ocfs2_error(inode->i_sb,
4453                                     "Inode %llu: error after split at cpos %u"
4454                                     "trunc len %u, existing record is (%u,%u)",
4455                                     (unsigned long long)OCFS2_I(inode)->ip_blkno,
4456                                     cpos, len, le32_to_cpu(rec->e_cpos),
4457                                     ocfs2_rec_clusters(el, rec));
4458                         ret = -EROFS;
4459                         goto out;
4460                 }
4461
4462                 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc,
4463                                          cpos, len);
4464                 if (ret) {
4465                         mlog_errno(ret);
4466                         goto out;
4467                 }
4468         }
4469
4470 out:
4471         ocfs2_free_path(path);
4472         return ret;
4473 }
4474
4475 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
4476 {
4477         struct buffer_head *tl_bh = osb->osb_tl_bh;
4478         struct ocfs2_dinode *di;
4479         struct ocfs2_truncate_log *tl;
4480
4481         di = (struct ocfs2_dinode *) tl_bh->b_data;
4482         tl = &di->id2.i_dealloc;
4483
4484         mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
4485                         "slot %d, invalid truncate log parameters: used = "
4486                         "%u, count = %u\n", osb->slot_num,
4487                         le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
4488         return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
4489 }
4490
4491 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
4492                                            unsigned int new_start)
4493 {
4494         unsigned int tail_index;
4495         unsigned int current_tail;
4496
4497         /* No records, nothing to coalesce */
4498         if (!le16_to_cpu(tl->tl_used))
4499                 return 0;
4500
4501         tail_index = le16_to_cpu(tl->tl_used) - 1;
4502         current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
4503         current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
4504
4505         return current_tail == new_start;
4506 }
4507
4508 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
4509                               handle_t *handle,
4510                               u64 start_blk,
4511                               unsigned int num_clusters)
4512 {
4513         int status, index;
4514         unsigned int start_cluster, tl_count;
4515         struct inode *tl_inode = osb->osb_tl_inode;
4516         struct buffer_head *tl_bh = osb->osb_tl_bh;
4517         struct ocfs2_dinode *di;
4518         struct ocfs2_truncate_log *tl;
4519
4520         mlog_entry("start_blk = %llu, num_clusters = %u\n",
4521                    (unsigned long long)start_blk, num_clusters);
4522
4523         BUG_ON(mutex_trylock(&tl_inode->i_mutex));
4524
4525         start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
4526
4527         di = (struct ocfs2_dinode *) tl_bh->b_data;
4528         tl = &di->id2.i_dealloc;
4529         if (!OCFS2_IS_VALID_DINODE(di)) {
4530                 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
4531                 status = -EIO;
4532                 goto bail;
4533         }
4534
4535         tl_count = le16_to_cpu(tl->tl_count);
4536         mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
4537                         tl_count == 0,
4538                         "Truncate record count on #%llu invalid "
4539                         "wanted %u, actual %u\n",
4540                         (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
4541                         ocfs2_truncate_recs_per_inode(osb->sb),
4542                         le16_to_cpu(tl->tl_count));
4543
4544         /* Caller should have known to flush before calling us. */
4545         index = le16_to_cpu(tl->tl_used);
4546         if (index >= tl_count) {
4547                 status = -ENOSPC;
4548                 mlog_errno(status);
4549                 goto bail;
4550         }
4551
4552         status = ocfs2_journal_access(handle, tl_inode, tl_bh,
4553                                       OCFS2_JOURNAL_ACCESS_WRITE);
4554         if (status < 0) {
4555                 mlog_errno(status);
4556                 goto bail;
4557         }
4558
4559         mlog(0, "Log truncate of %u clusters starting at cluster %u to "
4560              "%llu (index = %d)\n", num_clusters, start_cluster,
4561              (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index);
4562
4563         if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
4564                 /*
4565                  * Move index back to the record we are coalescing with.
4566                  * ocfs2_truncate_log_can_coalesce() guarantees nonzero
4567                  */
4568                 index--;
4569
4570                 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
4571                 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n",
4572                      index, le32_to_cpu(tl->tl_recs[index].t_start),
4573                      num_clusters);
4574         } else {
4575                 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
4576                 tl->tl_used = cpu_to_le16(index + 1);
4577         }
4578         tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
4579
4580         status = ocfs2_journal_dirty(handle, tl_bh);
4581         if (status < 0) {
4582                 mlog_errno(status);
4583                 goto bail;
4584         }
4585
4586 bail:
4587         mlog_exit(status);
4588         return status;
4589 }
4590
4591 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
4592                                          handle_t *handle,
4593                                          struct inode *data_alloc_inode,
4594                                          struct buffer_head *data_alloc_bh)
4595 {
4596         int status = 0;
4597         int i;
4598         unsigned int num_clusters;
4599         u64 start_blk;
4600         struct ocfs2_truncate_rec rec;
4601         struct ocfs2_dinode *di;
4602         struct ocfs2_truncate_log *tl;
4603         struct inode *tl_inode = osb->osb_tl_inode;
4604         struct buffer_head *tl_bh = osb->osb_tl_bh;
4605
4606         mlog_entry_void();
4607
4608         di = (struct ocfs2_dinode *) tl_bh->b_data;
4609         tl = &di->id2.i_dealloc;
4610         i = le16_to_cpu(tl->tl_used) - 1;
4611         while (i >= 0) {
4612                 /* Caller has given us at least enough credits to
4613                  * update the truncate log dinode */
4614                 status = ocfs2_journal_access(handle, tl_inode, tl_bh,
4615                                               OCFS2_JOURNAL_ACCESS_WRITE);
4616                 if (status < 0) {
4617                         mlog_errno(status);
4618                         goto bail;
4619                 }
4620
4621                 tl->tl_used = cpu_to_le16(i);
4622
4623                 status = ocfs2_journal_dirty(handle, tl_bh);
4624                 if (status < 0) {
4625                         mlog_errno(status);
4626                         goto bail;
4627                 }
4628
4629                 /* TODO: Perhaps we can calculate the bulk of the
4630                  * credits up front rather than extending like
4631                  * this. */
4632                 status = ocfs2_extend_trans(handle,
4633                                             OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
4634                 if (status < 0) {
4635                         mlog_errno(status);
4636                         goto bail;
4637                 }
4638
4639                 rec = tl->tl_recs[i];
4640                 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
4641                                                     le32_to_cpu(rec.t_start));
4642                 num_clusters = le32_to_cpu(rec.t_clusters);
4643
4644                 /* if start_blk is not set, we ignore the record as
4645                  * invalid. */
4646                 if (start_blk) {
4647                         mlog(0, "free record %d, start = %u, clusters = %u\n",
4648                              i, le32_to_cpu(rec.t_start), num_clusters);
4649
4650                         status = ocfs2_free_clusters(handle, data_alloc_inode,
4651                                                      data_alloc_bh, start_blk,
4652                                                      num_clusters);
4653                         if (status < 0) {
4654                                 mlog_errno(status);
4655                                 goto bail;
4656                         }
4657                 }
4658                 i--;
4659         }
4660
4661 bail:
4662         mlog_exit(status);
4663         return status;
4664 }
4665
4666 /* Expects you to already be holding tl_inode->i_mutex */
4667 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
4668 {
4669         int status;
4670         unsigned int num_to_flush;
4671         handle_t *handle;
4672         struct inode *tl_inode = osb->osb_tl_inode;
4673         struct inode *data_alloc_inode = NULL;
4674         struct buffer_head *tl_bh = osb->osb_tl_bh;
4675         struct buffer_head *data_alloc_bh = NULL;
4676         struct ocfs2_dinode *di;
4677         struct ocfs2_truncate_log *tl;
4678
4679         mlog_entry_void();
4680
4681         BUG_ON(mutex_trylock(&tl_inode->i_mutex));
4682
4683         di = (struct ocfs2_dinode *) tl_bh->b_data;
4684         tl = &di->id2.i_dealloc;
4685         if (!OCFS2_IS_VALID_DINODE(di)) {
4686                 OCFS2_RO_ON_INVALID_DINODE(osb->sb, di);
4687                 status = -EIO;
4688                 goto out;
4689         }
4690
4691         num_to_flush = le16_to_cpu(tl->tl_used);
4692         mlog(0, "Flush %u records from truncate log #%llu\n",
4693              num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno);
4694         if (!num_to_flush) {
4695                 status = 0;
4696                 goto out;
4697         }
4698
4699         data_alloc_inode = ocfs2_get_system_file_inode(osb,
4700                                                        GLOBAL_BITMAP_SYSTEM_INODE,
4701                                                        OCFS2_INVALID_SLOT);
4702         if (!data_alloc_inode) {
4703                 status = -EINVAL;
4704                 mlog(ML_ERROR, "Could not get bitmap inode!\n");
4705                 goto out;
4706         }
4707
4708         mutex_lock(&data_alloc_inode->i_mutex);
4709
4710         status = ocfs2_meta_lock(data_alloc_inode, &data_alloc_bh, 1);
4711         if (status < 0) {
4712                 mlog_errno(status);
4713                 goto out_mutex;
4714         }
4715
4716         handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
4717         if (IS_ERR(handle)) {
4718                 status = PTR_ERR(handle);
4719                 mlog_errno(status);
4720                 goto out_unlock;
4721         }
4722
4723         status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
4724                                                data_alloc_bh);
4725         if (status < 0)
4726                 mlog_errno(status);
4727
4728         ocfs2_commit_trans(osb, handle);
4729
4730 out_unlock:
4731         brelse(data_alloc_bh);
4732         ocfs2_meta_unlock(data_alloc_inode, 1);
4733
4734 out_mutex:
4735         mutex_unlock(&data_alloc_inode->i_mutex);
4736         iput(data_alloc_inode);
4737
4738 out:
4739         mlog_exit(status);
4740         return status;
4741 }
4742
4743 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
4744 {
4745         int status;
4746         struct inode *tl_inode = osb->osb_tl_inode;
4747
4748         mutex_lock(&tl_inode->i_mutex);
4749         status = __ocfs2_flush_truncate_log(osb);
4750         mutex_unlock(&tl_inode->i_mutex);
4751
4752         return status;
4753 }
4754
4755 static void ocfs2_truncate_log_worker(struct work_struct *work)
4756 {
4757         int status;
4758         struct ocfs2_super *osb =
4759                 container_of(work, struct ocfs2_super,
4760                              osb_truncate_log_wq.work);
4761
4762         mlog_entry_void();
4763
4764         status = ocfs2_flush_truncate_log(osb);
4765         if (status < 0)
4766                 mlog_errno(status);
4767
4768         mlog_exit(status);
4769 }
4770
4771 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
4772 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
4773                                        int cancel)
4774 {
4775         if (osb->osb_tl_inode) {
4776                 /* We want to push off log flushes while truncates are
4777                  * still running. */
4778                 if (cancel)
4779                         cancel_delayed_work(&osb->osb_truncate_log_wq);
4780
4781                 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
4782                                    OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
4783         }
4784 }
4785
4786 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
4787                                        int slot_num,
4788                                        struct inode **tl_inode,
4789                                        struct buffer_head **tl_bh)
4790 {
4791         int status;
4792         struct inode *inode = NULL;
4793         struct buffer_head *bh = NULL;
4794
4795         inode = ocfs2_get_system_file_inode(osb,
4796                                            TRUNCATE_LOG_SYSTEM_INODE,
4797                                            slot_num);
4798         if (!inode) {
4799                 status = -EINVAL;
4800                 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
4801                 goto bail;
4802         }
4803
4804         status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
4805                                   OCFS2_BH_CACHED, inode);
4806         if (status < 0) {
4807                 iput(inode);
4808                 mlog_errno(status);
4809                 goto bail;
4810         }
4811
4812         *tl_inode = inode;
4813         *tl_bh    = bh;
4814 bail:
4815         mlog_exit(status);
4816         return status;
4817 }
4818
4819 /* called during the 1st stage of node recovery. we stamp a clean
4820  * truncate log and pass back a copy for processing later. if the
4821  * truncate log does not require processing, a *tl_copy is set to
4822  * NULL. */
4823 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
4824                                       int slot_num,
4825                                       struct ocfs2_dinode **tl_copy)
4826 {
4827         int status;
4828         struct inode *tl_inode = NULL;
4829         struct buffer_head *tl_bh = NULL;
4830         struct ocfs2_dinode *di;
4831         struct ocfs2_truncate_log *tl;
4832
4833         *tl_copy = NULL;
4834
4835         mlog(0, "recover truncate log from slot %d\n", slot_num);
4836
4837         status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
4838         if (status < 0) {
4839                 mlog_errno(status);
4840                 goto bail;
4841         }
4842
4843         di = (struct ocfs2_dinode *) tl_bh->b_data;
4844         tl = &di->id2.i_dealloc;
4845         if (!OCFS2_IS_VALID_DINODE(di)) {
4846                 OCFS2_RO_ON_INVALID_DINODE(tl_inode->i_sb, di);
4847                 status = -EIO;
4848                 goto bail;
4849         }
4850
4851         if (le16_to_cpu(tl->tl_used)) {
4852                 mlog(0, "We'll have %u logs to recover\n",
4853                      le16_to_cpu(tl->tl_used));
4854
4855                 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
4856                 if (!(*tl_copy)) {
4857                         status = -ENOMEM;
4858                         mlog_errno(status);
4859                         goto bail;
4860                 }
4861
4862                 /* Assuming the write-out below goes well, this copy
4863                  * will be passed back to recovery for processing. */
4864                 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
4865
4866                 /* All we need to do to clear the truncate log is set
4867                  * tl_used. */
4868                 tl->tl_used = 0;
4869
4870                 status = ocfs2_write_block(osb, tl_bh, tl_inode);
4871                 if (status < 0) {
4872                         mlog_errno(status);
4873                         goto bail;
4874                 }
4875         }
4876
4877 bail:
4878         if (tl_inode)
4879                 iput(tl_inode);
4880         if (tl_bh)
4881                 brelse(tl_bh);
4882
4883         if (status < 0 && (*tl_copy)) {
4884                 kfree(*tl_copy);
4885                 *tl_copy = NULL;
4886         }
4887
4888         mlog_exit(status);
4889         return status;
4890 }
4891
4892 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
4893                                          struct ocfs2_dinode *tl_copy)
4894 {
4895         int status = 0;
4896         int i;
4897         unsigned int clusters, num_recs, start_cluster;
4898         u64 start_blk;
4899         handle_t *handle;
4900         struct inode *tl_inode = osb->osb_tl_inode;
4901         struct ocfs2_truncate_log *tl;
4902
4903         mlog_entry_void();
4904
4905         if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
4906                 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
4907                 return -EINVAL;
4908         }
4909
4910         tl = &tl_copy->id2.i_dealloc;
4911         num_recs = le16_to_cpu(tl->tl_used);
4912         mlog(0, "cleanup %u records from %llu\n", num_recs,
4913              (unsigned long long)le64_to_cpu(tl_copy->i_blkno));
4914
4915         mutex_lock(&tl_inode->i_mutex);
4916         for(i = 0; i < num_recs; i++) {
4917                 if (ocfs2_truncate_log_needs_flush(osb)) {
4918                         status = __ocfs2_flush_truncate_log(osb);
4919                         if (status < 0) {
4920                                 mlog_errno(status);
4921                                 goto bail_up;
4922                         }
4923                 }
4924
4925                 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
4926                 if (IS_ERR(handle)) {
4927                         status = PTR_ERR(handle);
4928                         mlog_errno(status);
4929                         goto bail_up;
4930                 }
4931
4932                 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
4933                 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
4934                 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
4935
4936                 status = ocfs2_truncate_log_append(osb, handle,
4937                                                    start_blk, clusters);
4938                 ocfs2_commit_trans(osb, handle);
4939                 if (status < 0) {
4940                         mlog_errno(status);
4941                         goto bail_up;
4942                 }
4943         }
4944
4945 bail_up:
4946         mutex_unlock(&tl_inode->i_mutex);
4947
4948         mlog_exit(status);
4949         return status;
4950 }
4951
4952 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
4953 {
4954         int status;
4955         struct inode *tl_inode = osb->osb_tl_inode;
4956
4957         mlog_entry_void();
4958
4959         if (tl_inode) {
4960                 cancel_delayed_work(&osb->osb_truncate_log_wq);
4961                 flush_workqueue(ocfs2_wq);
4962
4963                 status = ocfs2_flush_truncate_log(osb);
4964                 if (status < 0)
4965                         mlog_errno(status);
4966
4967                 brelse(osb->osb_tl_bh);
4968                 iput(osb->osb_tl_inode);
4969         }
4970
4971         mlog_exit_void();
4972 }
4973
4974 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
4975 {
4976         int status;
4977         struct inode *tl_inode = NULL;
4978         struct buffer_head *tl_bh = NULL;
4979
4980         mlog_entry_void();
4981
4982         status = ocfs2_get_truncate_log_info(osb,
4983                                              osb->slot_num,
4984                                              &tl_inode,
4985                                              &tl_bh);
4986         if (status < 0)
4987                 mlog_errno(status);
4988
4989         /* ocfs2_truncate_log_shutdown keys on the existence of
4990          * osb->osb_tl_inode so we don't set any of the osb variables
4991          * until we're sure all is well. */
4992         INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
4993                           ocfs2_truncate_log_worker);
4994         osb->osb_tl_bh    = tl_bh;
4995         osb->osb_tl_inode = tl_inode;
4996
4997         mlog_exit(status);
4998         return status;
4999 }
5000
5001 /*
5002  * Delayed de-allocation of suballocator blocks.
5003  *
5004  * Some sets of block de-allocations might involve multiple suballocator inodes.
5005  *
5006  * The locking for this can get extremely complicated, especially when
5007  * the suballocator inodes to delete from aren't known until deep
5008  * within an unrelated codepath.
5009  *
5010  * ocfs2_extent_block structures are a good example of this - an inode
5011  * btree could have been grown by any number of nodes each allocating
5012  * out of their own suballoc inode.
5013  *
5014  * These structures allow the delay of block de-allocation until a
5015  * later time, when locking of multiple cluster inodes won't cause
5016  * deadlock.
5017  */
5018
5019 /*
5020  * Describes a single block free from a suballocator
5021  */
5022 struct ocfs2_cached_block_free {
5023         struct ocfs2_cached_block_free          *free_next;
5024         u64                                     free_blk;
5025         unsigned int                            free_bit;
5026 };
5027
5028 struct ocfs2_per_slot_free_list {
5029         struct ocfs2_per_slot_free_list         *f_next_suballocator;
5030         int                                     f_inode_type;
5031         int                                     f_slot;
5032         struct ocfs2_cached_block_free          *f_first;
5033 };
5034
5035 static int ocfs2_free_cached_items(struct ocfs2_super *osb,
5036                                    int sysfile_type,
5037                                    int slot,
5038                                    struct ocfs2_cached_block_free *head)
5039 {
5040         int ret;
5041         u64 bg_blkno;
5042         handle_t *handle;
5043         struct inode *inode;
5044         struct buffer_head *di_bh = NULL;
5045         struct ocfs2_cached_block_free *tmp;
5046
5047         inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
5048         if (!inode) {
5049                 ret = -EINVAL;
5050                 mlog_errno(ret);
5051                 goto out;
5052         }
5053
5054         mutex_lock(&inode->i_mutex);
5055
5056         ret = ocfs2_meta_lock(inode, &di_bh, 1);
5057         if (ret) {
5058                 mlog_errno(ret);
5059                 goto out_mutex;
5060         }
5061
5062         handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
5063         if (IS_ERR(handle)) {
5064                 ret = PTR_ERR(handle);
5065                 mlog_errno(ret);
5066                 goto out_unlock;
5067         }
5068
5069         while (head) {
5070                 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
5071                                                       head->free_bit);
5072                 mlog(0, "Free bit: (bit %u, blkno %llu)\n",
5073                      head->free_bit, (unsigned long long)head->free_blk);
5074
5075                 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
5076                                                head->free_bit, bg_blkno, 1);
5077                 if (ret) {
5078                         mlog_errno(ret);
5079                         goto out_journal;
5080                 }
5081
5082                 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
5083                 if (ret) {
5084                         mlog_errno(ret);
5085                         goto out_journal;
5086                 }
5087
5088                 tmp = head;
5089                 head = head->free_next;
5090                 kfree(tmp);
5091         }
5092
5093 out_journal:
5094         ocfs2_commit_trans(osb, handle);
5095
5096 out_unlock:
5097         ocfs2_meta_unlock(inode, 1);
5098         brelse(di_bh);
5099 out_mutex:
5100         mutex_unlock(&inode->i_mutex);
5101         iput(inode);
5102 out:
5103         while(head) {
5104                 /* Premature exit may have left some dangling items. */
5105                 tmp = head;
5106                 head = head->free_next;
5107                 kfree(tmp);
5108         }
5109
5110         return ret;
5111 }
5112
5113 int ocfs2_run_deallocs(struct ocfs2_super *osb,
5114                        struct ocfs2_cached_dealloc_ctxt *ctxt)
5115 {
5116         int ret = 0, ret2;
5117         struct ocfs2_per_slot_free_list *fl;
5118
5119         if (!ctxt)
5120                 return 0;
5121
5122         while (ctxt->c_first_suballocator) {
5123                 fl = ctxt->c_first_suballocator;
5124
5125                 if (fl->f_first) {
5126                         mlog(0, "Free items: (type %u, slot %d)\n",
5127                              fl->f_inode_type, fl->f_slot);
5128                         ret2 = ocfs2_free_cached_items(osb, fl->f_inode_type,
5129                                                        fl->f_slot, fl->f_first);
5130                         if (ret2)
5131                                 mlog_errno(ret2);
5132                         if (!ret)
5133                                 ret = ret2;
5134                 }
5135
5136                 ctxt->c_first_suballocator = fl->f_next_suballocator;
5137                 kfree(fl);
5138         }
5139
5140         return ret;
5141 }
5142
5143 static struct ocfs2_per_slot_free_list *
5144 ocfs2_find_per_slot_free_list(int type,
5145                               int slot,
5146                               struct ocfs2_cached_dealloc_ctxt *ctxt)
5147 {
5148         struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
5149
5150         while (fl) {
5151                 if (fl->f_inode_type == type && fl->f_slot == slot)
5152                         return fl;
5153
5154                 fl = fl->f_next_suballocator;
5155         }
5156
5157         fl = kmalloc(sizeof(*fl), GFP_NOFS);
5158         if (fl) {
5159                 fl->f_inode_type = type;
5160                 fl->f_slot = slot;
5161                 fl->f_first = NULL;
5162                 fl->f_next_suballocator = ctxt->c_first_suballocator;
5163
5164                 ctxt->c_first_suballocator = fl;
5165         }
5166         return fl;
5167 }
5168
5169 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
5170                                      int type, int slot, u64 blkno,
5171                                      unsigned int bit)
5172 {
5173         int ret;
5174         struct ocfs2_per_slot_free_list *fl;
5175         struct ocfs2_cached_block_free *item;
5176
5177         fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
5178         if (fl == NULL) {
5179                 ret = -ENOMEM;
5180                 mlog_errno(ret);
5181                 goto out;
5182         }
5183
5184         item = kmalloc(sizeof(*item), GFP_NOFS);
5185         if (item == NULL) {
5186                 ret = -ENOMEM;
5187                 mlog_errno(ret);
5188                 goto out;
5189         }
5190
5191         mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n",
5192              type, slot, bit, (unsigned long long)blkno);
5193
5194         item->free_blk = blkno;
5195         item->free_bit = bit;
5196         item->free_next = fl->f_first;
5197
5198         fl->f_first = item;
5199
5200         ret = 0;
5201 out:
5202         return ret;
5203 }
5204
5205 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
5206                                          struct ocfs2_extent_block *eb)
5207 {
5208         return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
5209                                          le16_to_cpu(eb->h_suballoc_slot),
5210                                          le64_to_cpu(eb->h_blkno),
5211                                          le16_to_cpu(eb->h_suballoc_bit));
5212 }
5213
5214 /* This function will figure out whether the currently last extent
5215  * block will be deleted, and if it will, what the new last extent
5216  * block will be so we can update his h_next_leaf_blk field, as well
5217  * as the dinodes i_last_eb_blk */
5218 static int ocfs2_find_new_last_ext_blk(struct inode *inode,
5219                                        unsigned int clusters_to_del,
5220                                        struct ocfs2_path *path,
5221                                        struct buffer_head **new_last_eb)
5222 {
5223         int next_free, ret = 0;
5224         u32 cpos;
5225         struct ocfs2_extent_rec *rec;
5226         struct ocfs2_extent_block *eb;
5227         struct ocfs2_extent_list *el;
5228         struct buffer_head *bh = NULL;
5229
5230         *new_last_eb = NULL;
5231
5232         /* we have no tree, so of course, no last_eb. */
5233         if (!path->p_tree_depth)
5234                 goto out;
5235
5236         /* trunc to zero special case - this makes tree_depth = 0
5237          * regardless of what it is.  */
5238         if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
5239                 goto out;
5240
5241         el = path_leaf_el(path);
5242         BUG_ON(!el->l_next_free_rec);
5243
5244         /*
5245          * Make sure that this extent list will actually be empty
5246          * after we clear away the data. We can shortcut out if
5247          * there's more than one non-empty extent in the
5248          * list. Otherwise, a check of the remaining extent is
5249          * necessary.
5250          */
5251         next_free = le16_to_cpu(el->l_next_free_rec);
5252         rec = NULL;
5253         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
5254                 if (next_free > 2)
5255                         goto out;
5256
5257                 /* We may have a valid extent in index 1, check it. */
5258                 if (next_free == 2)
5259                         rec = &el->l_recs[1];
5260
5261                 /*
5262                  * Fall through - no more nonempty extents, so we want
5263                  * to delete this leaf.
5264                  */
5265         } else {
5266                 if (next_free > 1)
5267                         goto out;
5268
5269                 rec = &el->l_recs[0];
5270         }
5271
5272         if (rec) {
5273                 /*
5274                  * Check it we'll only be trimming off the end of this
5275                  * cluster.
5276                  */
5277                 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
5278                         goto out;
5279         }
5280
5281         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
5282         if (ret) {
5283                 mlog_errno(ret);
5284                 goto out;
5285         }
5286
5287         ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
5288         if (ret) {
5289                 mlog_errno(ret);
5290                 goto out;
5291         }
5292
5293         eb = (struct ocfs2_extent_block *) bh->b_data;
5294         el = &eb->h_list;
5295         if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
5296                 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
5297                 ret = -EROFS;
5298                 goto out;
5299         }
5300
5301         *new_last_eb = bh;
5302         get_bh(*new_last_eb);
5303         mlog(0, "returning block %llu, (cpos: %u)\n",
5304              (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
5305 out:
5306         brelse(bh);
5307
5308         return ret;
5309 }
5310
5311 /*
5312  * Trim some clusters off the rightmost edge of a tree. Only called
5313  * during truncate.
5314  *
5315  * The caller needs to:
5316  *   - start journaling of each path component.
5317  *   - compute and fully set up any new last ext block
5318  */
5319 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
5320                            handle_t *handle, struct ocfs2_truncate_context *tc,
5321                            u32 clusters_to_del, u64 *delete_start)
5322 {
5323         int ret, i, index = path->p_tree_depth;
5324         u32 new_edge = 0;
5325         u64 deleted_eb = 0;
5326         struct buffer_head *bh;
5327         struct ocfs2_extent_list *el;
5328         struct ocfs2_extent_rec *rec;
5329
5330         *delete_start = 0;
5331
5332         while (index >= 0) {
5333                 bh = path->p_node[index].bh;
5334                 el = path->p_node[index].el;
5335
5336                 mlog(0, "traveling tree (index = %d, block = %llu)\n",
5337                      index,  (unsigned long long)bh->b_blocknr);
5338
5339                 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
5340
5341                 if (index !=
5342                     (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
5343                         ocfs2_error(inode->i_sb,
5344                                     "Inode %lu has invalid ext. block %llu",
5345                                     inode->i_ino,
5346                                     (unsigned long long)bh->b_blocknr);
5347                         ret = -EROFS;
5348                         goto out;
5349                 }
5350
5351 find_tail_record:
5352                 i = le16_to_cpu(el->l_next_free_rec) - 1;
5353                 rec = &el->l_recs[i];
5354
5355                 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
5356                      "next = %u\n", i, le32_to_cpu(rec->e_cpos),
5357                      ocfs2_rec_clusters(el, rec),
5358                      (unsigned long long)le64_to_cpu(rec->e_blkno),
5359                      le16_to_cpu(el->l_next_free_rec));
5360
5361                 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
5362
5363                 if (le16_to_cpu(el->l_tree_depth) == 0) {
5364                         /*
5365                          * If the leaf block contains a single empty
5366                          * extent and no records, we can just remove
5367                          * the block.
5368                          */
5369                         if (i == 0 && ocfs2_is_empty_extent(rec)) {
5370                                 memset(rec, 0,
5371                                        sizeof(struct ocfs2_extent_rec));
5372                                 el->l_next_free_rec = cpu_to_le16(0);
5373
5374                                 goto delete;
5375                         }
5376
5377                         /*
5378                          * Remove any empty extents by shifting things
5379                          * left. That should make life much easier on
5380                          * the code below. This condition is rare
5381                          * enough that we shouldn't see a performance
5382                          * hit.
5383                          */
5384                         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
5385                                 le16_add_cpu(&el->l_next_free_rec, -1);
5386
5387                                 for(i = 0;
5388                                     i < le16_to_cpu(el->l_next_free_rec); i++)
5389                                         el->l_recs[i] = el->l_recs[i + 1];
5390
5391                                 memset(&el->l_recs[i], 0,
5392                                        sizeof(struct ocfs2_extent_rec));
5393
5394                                 /*
5395                                  * We've modified our extent list. The
5396                                  * simplest way to handle this change
5397                                  * is to being the search from the
5398                                  * start again.
5399                                  */
5400                                 goto find_tail_record;
5401                         }
5402
5403                         le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
5404
5405                         /*
5406                          * We'll use "new_edge" on our way back up the
5407                          * tree to know what our rightmost cpos is.
5408                          */
5409                         new_edge = le16_to_cpu(rec->e_leaf_clusters);
5410                         new_edge += le32_to_cpu(rec->e_cpos);
5411
5412                         /*
5413                          * The caller will use this to delete data blocks.
5414                          */
5415                         *delete_start = le64_to_cpu(rec->e_blkno)
5416                                 + ocfs2_clusters_to_blocks(inode->i_sb,
5417                                         le16_to_cpu(rec->e_leaf_clusters));
5418
5419                         /*
5420                          * If it's now empty, remove this record.
5421                          */
5422                         if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
5423                                 memset(rec, 0,
5424                                        sizeof(struct ocfs2_extent_rec));
5425                                 le16_add_cpu(&el->l_next_free_rec, -1);
5426                         }
5427                 } else {
5428                         if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
5429                                 memset(rec, 0,
5430                                        sizeof(struct ocfs2_extent_rec));
5431                                 le16_add_cpu(&el->l_next_free_rec, -1);
5432
5433                                 goto delete;
5434                         }
5435
5436                         /* Can this actually happen? */
5437                         if (le16_to_cpu(el->l_next_free_rec) == 0)
5438                                 goto delete;
5439
5440                         /*
5441                          * We never actually deleted any clusters
5442                          * because our leaf was empty. There's no
5443                          * reason to adjust the rightmost edge then.
5444                          */
5445                         if (new_edge == 0)
5446                                 goto delete;
5447
5448                         rec->e_int_clusters = cpu_to_le32(new_edge);
5449                         le32_add_cpu(&rec->e_int_clusters,
5450                                      -le32_to_cpu(rec->e_cpos));
5451
5452                          /*
5453                           * A deleted child record should have been
5454                           * caught above.
5455                           */
5456                          BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
5457                 }
5458
5459 delete:
5460                 ret = ocfs2_journal_dirty(handle, bh);
5461                 if (ret) {
5462                         mlog_errno(ret);
5463                         goto out;
5464                 }
5465
5466                 mlog(0, "extent list container %llu, after: record %d: "
5467                      "(%u, %u, %llu), next = %u.\n",
5468                      (unsigned long long)bh->b_blocknr, i,
5469                      le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
5470                      (unsigned long long)le64_to_cpu(rec->e_blkno),
5471                      le16_to_cpu(el->l_next_free_rec));
5472
5473                 /*
5474                  * We must be careful to only attempt delete of an
5475                  * extent block (and not the root inode block).
5476                  */
5477                 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
5478                         struct ocfs2_extent_block *eb =
5479                                 (struct ocfs2_extent_block *)bh->b_data;
5480
5481                         /*
5482                          * Save this for use when processing the
5483                          * parent block.
5484                          */
5485                         deleted_eb = le64_to_cpu(eb->h_blkno);
5486
5487                         mlog(0, "deleting this extent block.\n");
5488
5489                         ocfs2_remove_from_cache(inode, bh);
5490
5491                         BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
5492                         BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
5493                         BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
5494
5495                         ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb);
5496                         /* An error here is not fatal. */
5497                         if (ret < 0)
5498                                 mlog_errno(ret);
5499                 } else {
5500                         deleted_eb = 0;
5501                 }
5502
5503                 index--;
5504         }
5505
5506         ret = 0;
5507 out:
5508         return ret;
5509 }
5510
5511 static int ocfs2_do_truncate(struct ocfs2_super *osb,
5512                              unsigned int clusters_to_del,
5513                              struct inode *inode,
5514                              struct buffer_head *fe_bh,
5515                              handle_t *handle,
5516                              struct ocfs2_truncate_context *tc,
5517                              struct ocfs2_path *path)
5518 {
5519         int status;
5520         struct ocfs2_dinode *fe;
5521         struct ocfs2_extent_block *last_eb = NULL;
5522         struct ocfs2_extent_list *el;
5523         struct buffer_head *last_eb_bh = NULL;
5524         u64 delete_blk = 0;
5525
5526         fe = (struct ocfs2_dinode *) fe_bh->b_data;
5527
5528         status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
5529                                              path, &last_eb_bh);
5530         if (status < 0) {
5531                 mlog_errno(status);
5532                 goto bail;
5533         }
5534
5535         /*
5536          * Each component will be touched, so we might as well journal
5537          * here to avoid having to handle errors later.
5538          */
5539         status = ocfs2_journal_access_path(inode, handle, path);
5540         if (status < 0) {
5541                 mlog_errno(status);
5542                 goto bail;
5543         }
5544
5545         if (last_eb_bh) {
5546                 status = ocfs2_journal_access(handle, inode, last_eb_bh,
5547                                               OCFS2_JOURNAL_ACCESS_WRITE);
5548                 if (status < 0) {
5549                         mlog_errno(status);
5550                         goto bail;
5551                 }
5552
5553                 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5554         }
5555
5556         el = &(fe->id2.i_list);
5557
5558         /*
5559          * Lower levels depend on this never happening, but it's best
5560          * to check it up here before changing the tree.
5561          */
5562         if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
5563                 ocfs2_error(inode->i_sb,
5564                             "Inode %lu has an empty extent record, depth %u\n",
5565                             inode->i_ino, le16_to_cpu(el->l_tree_depth));
5566                 status = -EROFS;
5567                 goto bail;
5568         }
5569
5570         spin_lock(&OCFS2_I(inode)->ip_lock);
5571         OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
5572                                       clusters_to_del;
5573         spin_unlock(&OCFS2_I(inode)->ip_lock);
5574         le32_add_cpu(&fe->i_clusters, -clusters_to_del);
5575         inode->i_blocks = ocfs2_inode_sector_count(inode);
5576
5577         status = ocfs2_trim_tree(inode, path, handle, tc,
5578                                  clusters_to_del, &delete_blk);
5579         if (status) {
5580                 mlog_errno(status);
5581                 goto bail;
5582         }
5583
5584         if (le32_to_cpu(fe->i_clusters) == 0) {
5585                 /* trunc to zero is a special case. */
5586                 el->l_tree_depth = 0;
5587                 fe->i_last_eb_blk = 0;
5588         } else if (last_eb)
5589                 fe->i_last_eb_blk = last_eb->h_blkno;
5590
5591         status = ocfs2_journal_dirty(handle, fe_bh);
5592         if (status < 0) {
5593                 mlog_errno(status);
5594                 goto bail;
5595         }
5596
5597         if (last_eb) {
5598                 /* If there will be a new last extent block, then by
5599                  * definition, there cannot be any leaves to the right of
5600                  * him. */
5601                 last_eb->h_next_leaf_blk = 0;
5602                 status = ocfs2_journal_dirty(handle, last_eb_bh);
5603                 if (status < 0) {
5604                         mlog_errno(status);
5605                         goto bail;
5606                 }
5607         }
5608
5609         if (delete_blk) {
5610                 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
5611                                                    clusters_to_del);
5612                 if (status < 0) {
5613                         mlog_errno(status);
5614                         goto bail;
5615                 }
5616         }
5617         status = 0;
5618 bail:
5619
5620         mlog_exit(status);
5621         return status;
5622 }
5623
5624 static int ocfs2_writeback_zero_func(handle_t *handle, struct buffer_head *bh)
5625 {
5626         set_buffer_uptodate(bh);
5627         mark_buffer_dirty(bh);
5628         return 0;
5629 }
5630
5631 static int ocfs2_ordered_zero_func(handle_t *handle, struct buffer_head *bh)
5632 {
5633         set_buffer_uptodate(bh);
5634         mark_buffer_dirty(bh);
5635         return ocfs2_journal_dirty_data(handle, bh);
5636 }
5637
5638 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
5639                                      unsigned int from, unsigned int to,
5640                                      struct page *page, int zero, u64 *phys)
5641 {
5642         int ret, partial = 0;
5643
5644         ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
5645         if (ret)
5646                 mlog_errno(ret);
5647
5648         if (zero)
5649                 zero_user_page(page, from, to - from, KM_USER0);
5650
5651         /*
5652          * Need to set the buffers we zero'd into uptodate
5653          * here if they aren't - ocfs2_map_page_blocks()
5654          * might've skipped some
5655          */
5656         if (ocfs2_should_order_data(inode)) {
5657                 ret = walk_page_buffers(handle,
5658                                         page_buffers(page),
5659                                         from, to, &partial,
5660                                         ocfs2_ordered_zero_func);
5661                 if (ret < 0)
5662                         mlog_errno(ret);
5663         } else {
5664                 ret = walk_page_buffers(handle, page_buffers(page),
5665                                         from, to, &partial,
5666                                         ocfs2_writeback_zero_func);
5667                 if (ret < 0)
5668                         mlog_errno(ret);
5669         }
5670
5671         if (!partial)
5672                 SetPageUptodate(page);
5673
5674         flush_dcache_page(page);
5675 }
5676
5677 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
5678                                      loff_t end, struct page **pages,
5679                                      int numpages, u64 phys, handle_t *handle)
5680 {
5681         int i;
5682         struct page *page;
5683         unsigned int from, to = PAGE_CACHE_SIZE;
5684         struct super_block *sb = inode->i_sb;
5685
5686         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
5687
5688         if (numpages == 0)
5689                 goto out;
5690
5691         to = PAGE_CACHE_SIZE;
5692         for(i = 0; i < numpages; i++) {
5693                 page = pages[i];
5694
5695                 from = start & (PAGE_CACHE_SIZE - 1);
5696                 if ((end >> PAGE_CACHE_SHIFT) == page->index)
5697                         to = end & (PAGE_CACHE_SIZE - 1);
5698
5699                 BUG_ON(from > PAGE_CACHE_SIZE);
5700                 BUG_ON(to > PAGE_CACHE_SIZE);
5701
5702                 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
5703                                          &phys);
5704
5705                 start = (page->index + 1) << PAGE_CACHE_SHIFT;
5706         }
5707 out:
5708         if (pages)
5709                 ocfs2_unlock_and_free_pages(pages, numpages);
5710 }
5711
5712 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
5713                                 struct page **pages, int *num)
5714 {
5715         int numpages, ret = 0;
5716         struct super_block *sb = inode->i_sb;
5717         struct address_space *mapping = inode->i_mapping;
5718         unsigned long index;
5719         loff_t last_page_bytes;
5720
5721         BUG_ON(start > end);
5722
5723         BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
5724                (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
5725
5726         numpages = 0;
5727         last_page_bytes = PAGE_ALIGN(end);
5728         index = start >> PAGE_CACHE_SHIFT;
5729         do {
5730                 pages[numpages] = grab_cache_page(mapping, index);
5731                 if (!pages[numpages]) {
5732                         ret = -ENOMEM;
5733                         mlog_errno(ret);
5734                         goto out;
5735                 }
5736
5737                 numpages++;
5738                 index++;
5739         } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
5740
5741 out:
5742         if (ret != 0) {
5743                 if (pages)
5744                         ocfs2_unlock_and_free_pages(pages, numpages);
5745                 numpages = 0;
5746         }
5747
5748         *num = numpages;
5749
5750         return ret;
5751 }
5752
5753 /*
5754  * Zero the area past i_size but still within an allocated
5755  * cluster. This avoids exposing nonzero data on subsequent file
5756  * extends.
5757  *
5758  * We need to call this before i_size is updated on the inode because
5759  * otherwise block_write_full_page() will skip writeout of pages past
5760  * i_size. The new_i_size parameter is passed for this reason.
5761  */
5762 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
5763                                   u64 range_start, u64 range_end)
5764 {
5765         int ret = 0, numpages;
5766         struct page **pages = NULL;
5767         u64 phys;
5768         unsigned int ext_flags;
5769         struct super_block *sb = inode->i_sb;
5770
5771         /*
5772          * File systems which don't support sparse files zero on every
5773          * extend.
5774          */
5775         if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
5776                 return 0;
5777
5778         pages = kcalloc(ocfs2_pages_per_cluster(sb),
5779                         sizeof(struct page *), GFP_NOFS);
5780         if (pages == NULL) {
5781                 ret = -ENOMEM;
5782                 mlog_errno(ret);
5783                 goto out;
5784         }
5785
5786         if (range_start == range_end)
5787                 goto out;
5788
5789         ret = ocfs2_extent_map_get_blocks(inode,
5790                                           range_start >> sb->s_blocksize_bits,
5791                                           &phys, NULL, &ext_flags);
5792         if (ret) {
5793                 mlog_errno(ret);
5794                 goto out;
5795         }
5796
5797         /*
5798          * Tail is a hole, or is marked unwritten. In either case, we
5799          * can count on read and write to return/push zero's.
5800          */
5801         if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
5802                 goto out;
5803
5804         ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
5805                                    &numpages);
5806         if (ret) {
5807                 mlog_errno(ret);
5808                 goto out;
5809         }
5810
5811         ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
5812                                  numpages, phys, handle);
5813
5814         /*
5815          * Initiate writeout of the pages we zero'd here. We don't
5816          * wait on them - the truncate_inode_pages() call later will
5817          * do that for us.
5818          */
5819         ret = do_sync_mapping_range(inode->i_mapping, range_start,
5820                                     range_end - 1, SYNC_FILE_RANGE_WRITE);
5821         if (ret)
5822                 mlog_errno(ret);
5823
5824 out:
5825         if (pages)
5826                 kfree(pages);
5827
5828         return ret;
5829 }
5830
5831 static void ocfs2_zero_dinode_id2(struct inode *inode, struct ocfs2_dinode *di)
5832 {
5833         unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
5834
5835         memset(&di->id2, 0, blocksize - offsetof(struct ocfs2_dinode, id2));
5836 }
5837
5838 void ocfs2_dinode_new_extent_list(struct inode *inode,
5839                                   struct ocfs2_dinode *di)
5840 {
5841         ocfs2_zero_dinode_id2(inode, di);
5842         di->id2.i_list.l_tree_depth = 0;
5843         di->id2.i_list.l_next_free_rec = 0;
5844         di->id2.i_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_inode(inode->i_sb));
5845 }
5846
5847 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
5848 {
5849         struct ocfs2_inode_info *oi = OCFS2_I(inode);
5850         struct ocfs2_inline_data *idata = &di->id2.i_data;
5851
5852         spin_lock(&oi->ip_lock);
5853         oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
5854         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
5855         spin_unlock(&oi->ip_lock);
5856
5857         /*
5858          * We clear the entire i_data structure here so that all
5859          * fields can be properly initialized.
5860          */
5861         ocfs2_zero_dinode_id2(inode, di);
5862
5863         idata->id_count = cpu_to_le16(ocfs2_max_inline_data(inode->i_sb));
5864 }
5865
5866 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
5867                                          struct buffer_head *di_bh)
5868 {
5869         int ret, i, has_data, num_pages = 0;
5870         handle_t *handle;
5871         u64 uninitialized_var(block);
5872         struct ocfs2_inode_info *oi = OCFS2_I(inode);
5873         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5874         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
5875         struct ocfs2_alloc_context *data_ac = NULL;
5876         struct page **pages = NULL;
5877         loff_t end = osb->s_clustersize;
5878
5879         has_data = i_size_read(inode) ? 1 : 0;
5880
5881         if (has_data) {
5882                 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
5883                                 sizeof(struct page *), GFP_NOFS);
5884                 if (pages == NULL) {
5885                         ret = -ENOMEM;
5886                         mlog_errno(ret);
5887                         goto out;
5888                 }
5889
5890                 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
5891                 if (ret) {
5892                         mlog_errno(ret);
5893                         goto out;
5894                 }
5895         }
5896
5897         handle = ocfs2_start_trans(osb, OCFS2_INLINE_TO_EXTENTS_CREDITS);
5898         if (IS_ERR(handle)) {
5899                 ret = PTR_ERR(handle);
5900                 mlog_errno(ret);
5901                 goto out_unlock;
5902         }
5903
5904         ret = ocfs2_journal_access(handle, inode, di_bh,
5905                                    OCFS2_JOURNAL_ACCESS_WRITE);
5906         if (ret) {
5907                 mlog_errno(ret);
5908                 goto out_commit;
5909         }
5910
5911         if (has_data) {
5912                 u32 bit_off, num;
5913                 unsigned int page_end;
5914                 u64 phys;
5915
5916                 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
5917                                            &num);
5918                 if (ret) {
5919                         mlog_errno(ret);
5920                         goto out_commit;
5921                 }
5922
5923                 /*
5924                  * Save two copies, one for insert, and one that can
5925                  * be changed by ocfs2_map_and_dirty_page() below.
5926                  */
5927                 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
5928
5929                 /*
5930                  * Non sparse file systems zero on extend, so no need
5931                  * to do that now.
5932                  */
5933                 if (!ocfs2_sparse_alloc(osb) &&
5934                     PAGE_CACHE_SIZE < osb->s_clustersize)
5935                         end = PAGE_CACHE_SIZE;
5936
5937                 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
5938                 if (ret) {
5939                         mlog_errno(ret);
5940                         goto out_commit;
5941                 }
5942
5943                 /*
5944                  * This should populate the 1st page for us and mark
5945                  * it up to date.
5946                  */
5947                 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
5948                 if (ret) {
5949                         mlog_errno(ret);
5950                         goto out_commit;
5951                 }
5952
5953                 page_end = PAGE_CACHE_SIZE;
5954                 if (PAGE_CACHE_SIZE > osb->s_clustersize)
5955                         page_end = osb->s_clustersize;
5956
5957                 for (i = 0; i < num_pages; i++)
5958                         ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
5959                                                  pages[i], i > 0, &phys);
5960         }
5961
5962         spin_lock(&oi->ip_lock);
5963         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
5964         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
5965         spin_unlock(&oi->ip_lock);
5966
5967         ocfs2_dinode_new_extent_list(inode, di);
5968
5969         ocfs2_journal_dirty(handle, di_bh);
5970
5971         if (has_data) {
5972                 /*
5973                  * An error at this point should be extremely rare. If
5974                  * this proves to be false, we could always re-build
5975                  * the in-inode data from our pages.
5976                  */
5977                 ret = ocfs2_insert_extent(osb, handle, inode, di_bh,
5978                                           0, block, 1, 0, NULL);
5979                 if (ret) {
5980                         mlog_errno(ret);
5981                         goto out_commit;
5982                 }
5983
5984                 inode->i_blocks = ocfs2_inode_sector_count(inode);
5985         }
5986
5987 out_commit:
5988         ocfs2_commit_trans(osb, handle);
5989
5990 out_unlock:
5991         if (data_ac)
5992                 ocfs2_free_alloc_context(data_ac);
5993
5994 out:
5995         if (pages) {
5996                 ocfs2_unlock_and_free_pages(pages, num_pages);
5997                 kfree(pages);
5998         }
5999
6000         return ret;
6001 }
6002
6003 /*
6004  * It is expected, that by the time you call this function,
6005  * inode->i_size and fe->i_size have been adjusted.
6006  *
6007  * WARNING: This will kfree the truncate context
6008  */
6009 int ocfs2_commit_truncate(struct ocfs2_super *osb,
6010                           struct inode *inode,
6011                           struct buffer_head *fe_bh,
6012                           struct ocfs2_truncate_context *tc)
6013 {
6014         int status, i, credits, tl_sem = 0;
6015         u32 clusters_to_del, new_highest_cpos, range;
6016         struct ocfs2_extent_list *el;
6017         handle_t *handle = NULL;
6018         struct inode *tl_inode = osb->osb_tl_inode;
6019         struct ocfs2_path *path = NULL;
6020
6021         mlog_entry_void();
6022
6023         new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
6024                                                      i_size_read(inode));
6025
6026         path = ocfs2_new_inode_path(fe_bh);
6027         if (!path) {
6028                 status = -ENOMEM;
6029                 mlog_errno(status);
6030                 goto bail;
6031         }
6032
6033         ocfs2_extent_map_trunc(inode, new_highest_cpos);
6034
6035 start:
6036         /*
6037          * Check that we still have allocation to delete.
6038          */
6039         if (OCFS2_I(inode)->ip_clusters == 0) {
6040                 status = 0;
6041                 goto bail;
6042         }
6043
6044         /*
6045          * Truncate always works against the rightmost tree branch.
6046          */
6047         status = ocfs2_find_path(inode, path, UINT_MAX);
6048         if (status) {
6049                 mlog_errno(status);
6050                 goto bail;
6051         }
6052
6053         mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
6054              OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
6055
6056         /*
6057          * By now, el will point to the extent list on the bottom most
6058          * portion of this tree. Only the tail record is considered in
6059          * each pass.
6060          *
6061          * We handle the following cases, in order:
6062          * - empty extent: delete the remaining branch
6063          * - remove the entire record
6064          * - remove a partial record
6065          * - no record needs to be removed (truncate has completed)
6066          */
6067         el = path_leaf_el(path);
6068         if (le16_to_cpu(el->l_next_free_rec) == 0) {
6069                 ocfs2_error(inode->i_sb,
6070                             "Inode %llu has empty extent block at %llu\n",
6071                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
6072                             (unsigned long long)path_leaf_bh(path)->b_blocknr);
6073                 status = -EROFS;
6074                 goto bail;
6075         }
6076
6077         i = le16_to_cpu(el->l_next_free_rec) - 1;
6078         range = le32_to_cpu(el->l_recs[i].e_cpos) +
6079                 ocfs2_rec_clusters(el, &el->l_recs[i]);
6080         if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
6081                 clusters_to_del = 0;
6082         } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
6083                 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
6084         } else if (range > new_highest_cpos) {
6085                 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
6086                                    le32_to_cpu(el->l_recs[i].e_cpos)) -
6087                                   new_highest_cpos;
6088         } else {
6089                 status = 0;
6090                 goto bail;
6091         }
6092
6093         mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
6094              clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
6095
6096         BUG_ON(clusters_to_del == 0);
6097
6098         mutex_lock(&tl_inode->i_mutex);
6099         tl_sem = 1;
6100         /* ocfs2_truncate_log_needs_flush guarantees us at least one
6101          * record is free for use. If there isn't any, we flush to get
6102          * an empty truncate log.  */
6103         if (ocfs2_truncate_log_needs_flush(osb)) {
6104                 status = __ocfs2_flush_truncate_log(osb);
6105                 if (status < 0) {
6106                         mlog_errno(status);
6107                         goto bail;
6108                 }
6109         }
6110
6111         credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
6112                                                 (struct ocfs2_dinode *)fe_bh->b_data,
6113                                                 el);
6114         handle = ocfs2_start_trans(osb, credits);
6115         if (IS_ERR(handle)) {
6116                 status = PTR_ERR(handle);
6117                 handle = NULL;
6118                 mlog_errno(status);
6119                 goto bail;
6120         }
6121
6122         status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
6123                                    tc, path);
6124         if (status < 0) {
6125                 mlog_errno(status);
6126                 goto bail;
6127         }
6128
6129         mutex_unlock(&tl_inode->i_mutex);
6130         tl_sem = 0;
6131
6132         ocfs2_commit_trans(osb, handle);
6133         handle = NULL;
6134
6135         ocfs2_reinit_path(path, 1);
6136
6137         /*
6138          * The check above will catch the case where we've truncated
6139          * away all allocation.
6140          */
6141         goto start;
6142
6143 bail:
6144
6145         ocfs2_schedule_truncate_log_flush(osb, 1);
6146
6147         if (tl_sem)
6148                 mutex_unlock(&tl_inode->i_mutex);
6149
6150         if (handle)
6151                 ocfs2_commit_trans(osb, handle);
6152
6153         ocfs2_run_deallocs(osb, &tc->tc_dealloc);
6154
6155         ocfs2_free_path(path);
6156
6157         /* This will drop the ext_alloc cluster lock for us */
6158         ocfs2_free_truncate_context(tc);
6159
6160         mlog_exit(status);
6161         return status;
6162 }
6163
6164 /*
6165  * Expects the inode to already be locked.
6166  */
6167 int ocfs2_prepare_truncate(struct ocfs2_super *osb,
6168                            struct inode *inode,
6169                            struct buffer_head *fe_bh,
6170                            struct ocfs2_truncate_context **tc)
6171 {
6172         int status;
6173         unsigned int new_i_clusters;
6174         struct ocfs2_dinode *fe;
6175         struct ocfs2_extent_block *eb;
6176         struct buffer_head *last_eb_bh = NULL;
6177
6178         mlog_entry_void();
6179
6180         *tc = NULL;
6181
6182         new_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
6183                                                   i_size_read(inode));
6184         fe = (struct ocfs2_dinode *) fe_bh->b_data;
6185
6186         mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size ="
6187              "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters,
6188              (unsigned long long)le64_to_cpu(fe->i_size));
6189
6190         *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
6191         if (!(*tc)) {
6192                 status = -ENOMEM;
6193                 mlog_errno(status);
6194                 goto bail;
6195         }
6196         ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc);
6197
6198         if (fe->id2.i_list.l_tree_depth) {
6199                 status = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
6200                                           &last_eb_bh, OCFS2_BH_CACHED, inode);
6201                 if (status < 0) {
6202                         mlog_errno(status);
6203                         goto bail;
6204                 }
6205                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
6206                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
6207                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
6208
6209                         brelse(last_eb_bh);
6210                         status = -EIO;
6211                         goto bail;
6212                 }
6213         }
6214
6215         (*tc)->tc_last_eb_bh = last_eb_bh;
6216
6217         status = 0;
6218 bail:
6219         if (status < 0) {
6220                 if (*tc)
6221                         ocfs2_free_truncate_context(*tc);
6222                 *tc = NULL;
6223         }
6224         mlog_exit_void();
6225         return status;
6226 }
6227
6228 /*
6229  * 'start' is inclusive, 'end' is not.
6230  */
6231 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
6232                           unsigned int start, unsigned int end, int trunc)
6233 {
6234         int ret;
6235         unsigned int numbytes;
6236         handle_t *handle;
6237         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6238         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6239         struct ocfs2_inline_data *idata = &di->id2.i_data;
6240
6241         if (end > i_size_read(inode))
6242                 end = i_size_read(inode);
6243
6244         BUG_ON(start >= end);
6245
6246         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
6247             !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
6248             !ocfs2_supports_inline_data(osb)) {
6249                 ocfs2_error(inode->i_sb,
6250                             "Inline data flags for inode %llu don't agree! "
6251                             "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
6252                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
6253                             le16_to_cpu(di->i_dyn_features),
6254                             OCFS2_I(inode)->ip_dyn_features,
6255                             osb->s_feature_incompat);
6256                 ret = -EROFS;
6257                 goto out;
6258         }
6259
6260         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
6261         if (IS_ERR(handle)) {
6262                 ret = PTR_ERR(handle);
6263                 mlog_errno(ret);
6264                 goto out;
6265         }
6266
6267         ret = ocfs2_journal_access(handle, inode, di_bh,
6268                                    OCFS2_JOURNAL_ACCESS_WRITE);
6269         if (ret) {
6270                 mlog_errno(ret);
6271                 goto out_commit;
6272         }
6273
6274         numbytes = end - start;
6275         memset(idata->id_data + start, 0, numbytes);
6276
6277         /*
6278          * No need to worry about the data page here - it's been
6279          * truncated already and inline data doesn't need it for
6280          * pushing zero's to disk, so we'll let readpage pick it up
6281          * later.
6282          */
6283         if (trunc) {
6284                 i_size_write(inode, start);
6285                 di->i_size = cpu_to_le64(start);
6286         }
6287
6288         inode->i_blocks = ocfs2_inode_sector_count(inode);
6289         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
6290
6291         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
6292         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
6293
6294         ocfs2_journal_dirty(handle, di_bh);
6295
6296 out_commit:
6297         ocfs2_commit_trans(osb, handle);
6298
6299 out:
6300         return ret;
6301 }
6302
6303 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc)
6304 {
6305         /*
6306          * The caller is responsible for completing deallocation
6307          * before freeing the context.
6308          */
6309         if (tc->tc_dealloc.c_first_suballocator != NULL)
6310                 mlog(ML_NOTICE,
6311                      "Truncate completion has non-empty dealloc context\n");
6312
6313         if (tc->tc_last_eb_bh)
6314                 brelse(tc->tc_last_eb_bh);
6315
6316         kfree(tc);
6317 }