tree-wide: fix assorted typos all over the place
[linux-2.6.git] / fs / ocfs2 / refcounttree.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * refcounttree.c
5  *
6  * Copyright (C) 2009 Oracle.  All rights reserved.
7  *
8  * This program is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU General Public
10  * License version 2 as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License for more details.
16  */
17
18 #include <linux/sort.h>
19 #define MLOG_MASK_PREFIX ML_REFCOUNT
20 #include <cluster/masklog.h>
21 #include "ocfs2.h"
22 #include "inode.h"
23 #include "alloc.h"
24 #include "suballoc.h"
25 #include "journal.h"
26 #include "uptodate.h"
27 #include "super.h"
28 #include "buffer_head_io.h"
29 #include "blockcheck.h"
30 #include "refcounttree.h"
31 #include "sysfile.h"
32 #include "dlmglue.h"
33 #include "extent_map.h"
34 #include "aops.h"
35 #include "xattr.h"
36 #include "namei.h"
37
38 #include <linux/bio.h>
39 #include <linux/blkdev.h>
40 #include <linux/gfp.h>
41 #include <linux/slab.h>
42 #include <linux/writeback.h>
43 #include <linux/pagevec.h>
44 #include <linux/swap.h>
45 #include <linux/security.h>
46 #include <linux/fsnotify.h>
47 #include <linux/quotaops.h>
48 #include <linux/namei.h>
49 #include <linux/mount.h>
50
51 struct ocfs2_cow_context {
52         struct inode *inode;
53         u32 cow_start;
54         u32 cow_len;
55         struct ocfs2_extent_tree data_et;
56         struct ocfs2_refcount_tree *ref_tree;
57         struct buffer_head *ref_root_bh;
58         struct ocfs2_alloc_context *meta_ac;
59         struct ocfs2_alloc_context *data_ac;
60         struct ocfs2_cached_dealloc_ctxt dealloc;
61         void *cow_object;
62         struct ocfs2_post_refcount *post_refcount;
63         int extra_credits;
64         int (*get_clusters)(struct ocfs2_cow_context *context,
65                             u32 v_cluster, u32 *p_cluster,
66                             u32 *num_clusters,
67                             unsigned int *extent_flags);
68         int (*cow_duplicate_clusters)(handle_t *handle,
69                                       struct ocfs2_cow_context *context,
70                                       u32 cpos, u32 old_cluster,
71                                       u32 new_cluster, u32 new_len);
72 };
73
74 static inline struct ocfs2_refcount_tree *
75 cache_info_to_refcount(struct ocfs2_caching_info *ci)
76 {
77         return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
78 }
79
80 static int ocfs2_validate_refcount_block(struct super_block *sb,
81                                          struct buffer_head *bh)
82 {
83         int rc;
84         struct ocfs2_refcount_block *rb =
85                 (struct ocfs2_refcount_block *)bh->b_data;
86
87         mlog(0, "Validating refcount block %llu\n",
88              (unsigned long long)bh->b_blocknr);
89
90         BUG_ON(!buffer_uptodate(bh));
91
92         /*
93          * If the ecc fails, we return the error but otherwise
94          * leave the filesystem running.  We know any error is
95          * local to this block.
96          */
97         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check);
98         if (rc) {
99                 mlog(ML_ERROR, "Checksum failed for refcount block %llu\n",
100                      (unsigned long long)bh->b_blocknr);
101                 return rc;
102         }
103
104
105         if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) {
106                 ocfs2_error(sb,
107                             "Refcount block #%llu has bad signature %.*s",
108                             (unsigned long long)bh->b_blocknr, 7,
109                             rb->rf_signature);
110                 return -EINVAL;
111         }
112
113         if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
114                 ocfs2_error(sb,
115                             "Refcount block #%llu has an invalid rf_blkno "
116                             "of %llu",
117                             (unsigned long long)bh->b_blocknr,
118                             (unsigned long long)le64_to_cpu(rb->rf_blkno));
119                 return -EINVAL;
120         }
121
122         if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
123                 ocfs2_error(sb,
124                             "Refcount block #%llu has an invalid "
125                             "rf_fs_generation of #%u",
126                             (unsigned long long)bh->b_blocknr,
127                             le32_to_cpu(rb->rf_fs_generation));
128                 return -EINVAL;
129         }
130
131         return 0;
132 }
133
134 static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
135                                      u64 rb_blkno,
136                                      struct buffer_head **bh)
137 {
138         int rc;
139         struct buffer_head *tmp = *bh;
140
141         rc = ocfs2_read_block(ci, rb_blkno, &tmp,
142                               ocfs2_validate_refcount_block);
143
144         /* If ocfs2_read_block() got us a new bh, pass it up. */
145         if (!rc && !*bh)
146                 *bh = tmp;
147
148         return rc;
149 }
150
151 static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
152 {
153         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
154
155         return rf->rf_blkno;
156 }
157
158 static struct super_block *
159 ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
160 {
161         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
162
163         return rf->rf_sb;
164 }
165
166 static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
167 {
168         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
169
170         spin_lock(&rf->rf_lock);
171 }
172
173 static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
174 {
175         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
176
177         spin_unlock(&rf->rf_lock);
178 }
179
180 static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
181 {
182         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
183
184         mutex_lock(&rf->rf_io_mutex);
185 }
186
187 static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
188 {
189         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
190
191         mutex_unlock(&rf->rf_io_mutex);
192 }
193
194 static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
195         .co_owner               = ocfs2_refcount_cache_owner,
196         .co_get_super           = ocfs2_refcount_cache_get_super,
197         .co_cache_lock          = ocfs2_refcount_cache_lock,
198         .co_cache_unlock        = ocfs2_refcount_cache_unlock,
199         .co_io_lock             = ocfs2_refcount_cache_io_lock,
200         .co_io_unlock           = ocfs2_refcount_cache_io_unlock,
201 };
202
203 static struct ocfs2_refcount_tree *
204 ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
205 {
206         struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
207         struct ocfs2_refcount_tree *tree = NULL;
208
209         while (n) {
210                 tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
211
212                 if (blkno < tree->rf_blkno)
213                         n = n->rb_left;
214                 else if (blkno > tree->rf_blkno)
215                         n = n->rb_right;
216                 else
217                         return tree;
218         }
219
220         return NULL;
221 }
222
223 /* osb_lock is already locked. */
224 static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
225                                        struct ocfs2_refcount_tree *new)
226 {
227         u64 rf_blkno = new->rf_blkno;
228         struct rb_node *parent = NULL;
229         struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
230         struct ocfs2_refcount_tree *tmp;
231
232         while (*p) {
233                 parent = *p;
234
235                 tmp = rb_entry(parent, struct ocfs2_refcount_tree,
236                                rf_node);
237
238                 if (rf_blkno < tmp->rf_blkno)
239                         p = &(*p)->rb_left;
240                 else if (rf_blkno > tmp->rf_blkno)
241                         p = &(*p)->rb_right;
242                 else {
243                         /* This should never happen! */
244                         mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
245                              (unsigned long long)rf_blkno);
246                         BUG();
247                 }
248         }
249
250         rb_link_node(&new->rf_node, parent, p);
251         rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
252 }
253
254 static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
255 {
256         ocfs2_metadata_cache_exit(&tree->rf_ci);
257         ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
258         ocfs2_lock_res_free(&tree->rf_lockres);
259         kfree(tree);
260 }
261
262 static inline void
263 ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
264                                         struct ocfs2_refcount_tree *tree)
265 {
266         rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
267         if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
268                 osb->osb_ref_tree_lru = NULL;
269 }
270
271 static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
272                                         struct ocfs2_refcount_tree *tree)
273 {
274         spin_lock(&osb->osb_lock);
275         ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
276         spin_unlock(&osb->osb_lock);
277 }
278
279 void ocfs2_kref_remove_refcount_tree(struct kref *kref)
280 {
281         struct ocfs2_refcount_tree *tree =
282                 container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
283
284         ocfs2_free_refcount_tree(tree);
285 }
286
287 static inline void
288 ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
289 {
290         kref_get(&tree->rf_getcnt);
291 }
292
293 static inline void
294 ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
295 {
296         kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
297 }
298
299 static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
300                                                struct super_block *sb)
301 {
302         ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
303         mutex_init(&new->rf_io_mutex);
304         new->rf_sb = sb;
305         spin_lock_init(&new->rf_lock);
306 }
307
308 static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
309                                         struct ocfs2_refcount_tree *new,
310                                         u64 rf_blkno, u32 generation)
311 {
312         init_rwsem(&new->rf_sem);
313         ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
314                                      rf_blkno, generation);
315 }
316
317 static struct ocfs2_refcount_tree*
318 ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
319 {
320         struct ocfs2_refcount_tree *new;
321
322         new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
323         if (!new)
324                 return NULL;
325
326         new->rf_blkno = rf_blkno;
327         kref_init(&new->rf_getcnt);
328         ocfs2_init_refcount_tree_ci(new, osb->sb);
329
330         return new;
331 }
332
333 static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
334                                    struct ocfs2_refcount_tree **ret_tree)
335 {
336         int ret = 0;
337         struct ocfs2_refcount_tree *tree, *new = NULL;
338         struct buffer_head *ref_root_bh = NULL;
339         struct ocfs2_refcount_block *ref_rb;
340
341         spin_lock(&osb->osb_lock);
342         if (osb->osb_ref_tree_lru &&
343             osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
344                 tree = osb->osb_ref_tree_lru;
345         else
346                 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
347         if (tree)
348                 goto out;
349
350         spin_unlock(&osb->osb_lock);
351
352         new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
353         if (!new) {
354                 ret = -ENOMEM;
355                 mlog_errno(ret);
356                 return ret;
357         }
358         /*
359          * We need the generation to create the refcount tree lock and since
360          * it isn't changed during the tree modification, we are safe here to
361          * read without protection.
362          * We also have to purge the cache after we create the lock since the
363          * refcount block may have the stale data. It can only be trusted when
364          * we hold the refcount lock.
365          */
366         ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
367         if (ret) {
368                 mlog_errno(ret);
369                 ocfs2_metadata_cache_exit(&new->rf_ci);
370                 kfree(new);
371                 return ret;
372         }
373
374         ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
375         new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
376         ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
377                                       new->rf_generation);
378         ocfs2_metadata_cache_purge(&new->rf_ci);
379
380         spin_lock(&osb->osb_lock);
381         tree = ocfs2_find_refcount_tree(osb, rf_blkno);
382         if (tree)
383                 goto out;
384
385         ocfs2_insert_refcount_tree(osb, new);
386
387         tree = new;
388         new = NULL;
389
390 out:
391         *ret_tree = tree;
392
393         osb->osb_ref_tree_lru = tree;
394
395         spin_unlock(&osb->osb_lock);
396
397         if (new)
398                 ocfs2_free_refcount_tree(new);
399
400         brelse(ref_root_bh);
401         return ret;
402 }
403
404 static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
405 {
406         int ret;
407         struct buffer_head *di_bh = NULL;
408         struct ocfs2_dinode *di;
409
410         ret = ocfs2_read_inode_block(inode, &di_bh);
411         if (ret) {
412                 mlog_errno(ret);
413                 goto out;
414         }
415
416         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
417
418         di = (struct ocfs2_dinode *)di_bh->b_data;
419         *ref_blkno = le64_to_cpu(di->i_refcount_loc);
420         brelse(di_bh);
421 out:
422         return ret;
423 }
424
425 static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
426                                       struct ocfs2_refcount_tree *tree, int rw)
427 {
428         int ret;
429
430         ret = ocfs2_refcount_lock(tree, rw);
431         if (ret) {
432                 mlog_errno(ret);
433                 goto out;
434         }
435
436         if (rw)
437                 down_write(&tree->rf_sem);
438         else
439                 down_read(&tree->rf_sem);
440
441 out:
442         return ret;
443 }
444
445 /*
446  * Lock the refcount tree pointed by ref_blkno and return the tree.
447  * In most case, we lock the tree and read the refcount block.
448  * So read it here if the caller really needs it.
449  *
450  * If the tree has been re-created by other node, it will free the
451  * old one and re-create it.
452  */
453 int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
454                              u64 ref_blkno, int rw,
455                              struct ocfs2_refcount_tree **ret_tree,
456                              struct buffer_head **ref_bh)
457 {
458         int ret, delete_tree = 0;
459         struct ocfs2_refcount_tree *tree = NULL;
460         struct buffer_head *ref_root_bh = NULL;
461         struct ocfs2_refcount_block *rb;
462
463 again:
464         ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
465         if (ret) {
466                 mlog_errno(ret);
467                 return ret;
468         }
469
470         ocfs2_refcount_tree_get(tree);
471
472         ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
473         if (ret) {
474                 mlog_errno(ret);
475                 ocfs2_refcount_tree_put(tree);
476                 goto out;
477         }
478
479         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
480                                         &ref_root_bh);
481         if (ret) {
482                 mlog_errno(ret);
483                 ocfs2_unlock_refcount_tree(osb, tree, rw);
484                 ocfs2_refcount_tree_put(tree);
485                 goto out;
486         }
487
488         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
489         /*
490          * If the refcount block has been freed and re-created, we may need
491          * to recreate the refcount tree also.
492          *
493          * Here we just remove the tree from the rb-tree, and the last
494          * kref holder will unlock and delete this refcount_tree.
495          * Then we goto "again" and ocfs2_get_refcount_tree will create
496          * the new refcount tree for us.
497          */
498         if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
499                 if (!tree->rf_removed) {
500                         ocfs2_erase_refcount_tree_from_list(osb, tree);
501                         tree->rf_removed = 1;
502                         delete_tree = 1;
503                 }
504
505                 ocfs2_unlock_refcount_tree(osb, tree, rw);
506                 /*
507                  * We get an extra reference when we create the refcount
508                  * tree, so another put will destroy it.
509                  */
510                 if (delete_tree)
511                         ocfs2_refcount_tree_put(tree);
512                 brelse(ref_root_bh);
513                 ref_root_bh = NULL;
514                 goto again;
515         }
516
517         *ret_tree = tree;
518         if (ref_bh) {
519                 *ref_bh = ref_root_bh;
520                 ref_root_bh = NULL;
521         }
522 out:
523         brelse(ref_root_bh);
524         return ret;
525 }
526
527 int ocfs2_lock_refcount_tree_by_inode(struct inode *inode, int rw,
528                                       struct ocfs2_refcount_tree **ret_tree,
529                                       struct buffer_head **ref_bh)
530 {
531         int ret;
532         u64 ref_blkno;
533
534         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
535         if (ret) {
536                 mlog_errno(ret);
537                 return ret;
538         }
539
540         return ocfs2_lock_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno,
541                                         rw, ret_tree, ref_bh);
542 }
543
544 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
545                                 struct ocfs2_refcount_tree *tree, int rw)
546 {
547         if (rw)
548                 up_write(&tree->rf_sem);
549         else
550                 up_read(&tree->rf_sem);
551
552         ocfs2_refcount_unlock(tree, rw);
553         ocfs2_refcount_tree_put(tree);
554 }
555
556 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
557 {
558         struct rb_node *node;
559         struct ocfs2_refcount_tree *tree;
560         struct rb_root *root = &osb->osb_rf_lock_tree;
561
562         while ((node = rb_last(root)) != NULL) {
563                 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
564
565                 mlog(0, "Purge tree %llu\n",
566                      (unsigned long long) tree->rf_blkno);
567
568                 rb_erase(&tree->rf_node, root);
569                 ocfs2_free_refcount_tree(tree);
570         }
571 }
572
573 /*
574  * Create a refcount tree for an inode.
575  * We take for granted that the inode is already locked.
576  */
577 static int ocfs2_create_refcount_tree(struct inode *inode,
578                                       struct buffer_head *di_bh)
579 {
580         int ret;
581         handle_t *handle = NULL;
582         struct ocfs2_alloc_context *meta_ac = NULL;
583         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
584         struct ocfs2_inode_info *oi = OCFS2_I(inode);
585         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
586         struct buffer_head *new_bh = NULL;
587         struct ocfs2_refcount_block *rb;
588         struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
589         u16 suballoc_bit_start;
590         u32 num_got;
591         u64 first_blkno;
592
593         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
594
595         mlog(0, "create tree for inode %lu\n", inode->i_ino);
596
597         ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
598         if (ret) {
599                 mlog_errno(ret);
600                 goto out;
601         }
602
603         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
604         if (IS_ERR(handle)) {
605                 ret = PTR_ERR(handle);
606                 mlog_errno(ret);
607                 goto out;
608         }
609
610         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
611                                       OCFS2_JOURNAL_ACCESS_WRITE);
612         if (ret) {
613                 mlog_errno(ret);
614                 goto out_commit;
615         }
616
617         ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1,
618                                    &suballoc_bit_start, &num_got,
619                                    &first_blkno);
620         if (ret) {
621                 mlog_errno(ret);
622                 goto out_commit;
623         }
624
625         new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
626         if (!new_tree) {
627                 ret = -ENOMEM;
628                 mlog_errno(ret);
629                 goto out_commit;
630         }
631
632         new_bh = sb_getblk(inode->i_sb, first_blkno);
633         ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
634
635         ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
636                                       OCFS2_JOURNAL_ACCESS_CREATE);
637         if (ret) {
638                 mlog_errno(ret);
639                 goto out_commit;
640         }
641
642         /* Initialize ocfs2_refcount_block. */
643         rb = (struct ocfs2_refcount_block *)new_bh->b_data;
644         memset(rb, 0, inode->i_sb->s_blocksize);
645         strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
646         rb->rf_suballoc_slot = cpu_to_le16(osb->slot_num);
647         rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
648         rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
649         rb->rf_blkno = cpu_to_le64(first_blkno);
650         rb->rf_count = cpu_to_le32(1);
651         rb->rf_records.rl_count =
652                         cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
653         spin_lock(&osb->osb_lock);
654         rb->rf_generation = osb->s_next_generation++;
655         spin_unlock(&osb->osb_lock);
656
657         ocfs2_journal_dirty(handle, new_bh);
658
659         spin_lock(&oi->ip_lock);
660         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
661         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
662         di->i_refcount_loc = cpu_to_le64(first_blkno);
663         spin_unlock(&oi->ip_lock);
664
665         mlog(0, "created tree for inode %lu, refblock %llu\n",
666              inode->i_ino, (unsigned long long)first_blkno);
667
668         ocfs2_journal_dirty(handle, di_bh);
669
670         /*
671          * We have to init the tree lock here since it will use
672          * the generation number to create it.
673          */
674         new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
675         ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
676                                       new_tree->rf_generation);
677
678         spin_lock(&osb->osb_lock);
679         tree = ocfs2_find_refcount_tree(osb, first_blkno);
680
681         /*
682          * We've just created a new refcount tree in this block.  If
683          * we found a refcount tree on the ocfs2_super, it must be
684          * one we just deleted.  We free the old tree before
685          * inserting the new tree.
686          */
687         BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
688         if (tree)
689                 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
690         ocfs2_insert_refcount_tree(osb, new_tree);
691         spin_unlock(&osb->osb_lock);
692         new_tree = NULL;
693         if (tree)
694                 ocfs2_refcount_tree_put(tree);
695
696 out_commit:
697         ocfs2_commit_trans(osb, handle);
698
699 out:
700         if (new_tree) {
701                 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
702                 kfree(new_tree);
703         }
704
705         brelse(new_bh);
706         if (meta_ac)
707                 ocfs2_free_alloc_context(meta_ac);
708
709         return ret;
710 }
711
712 static int ocfs2_set_refcount_tree(struct inode *inode,
713                                    struct buffer_head *di_bh,
714                                    u64 refcount_loc)
715 {
716         int ret;
717         handle_t *handle = NULL;
718         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
719         struct ocfs2_inode_info *oi = OCFS2_I(inode);
720         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
721         struct buffer_head *ref_root_bh = NULL;
722         struct ocfs2_refcount_block *rb;
723         struct ocfs2_refcount_tree *ref_tree;
724
725         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
726
727         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
728                                        &ref_tree, &ref_root_bh);
729         if (ret) {
730                 mlog_errno(ret);
731                 return ret;
732         }
733
734         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
735         if (IS_ERR(handle)) {
736                 ret = PTR_ERR(handle);
737                 mlog_errno(ret);
738                 goto out;
739         }
740
741         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
742                                       OCFS2_JOURNAL_ACCESS_WRITE);
743         if (ret) {
744                 mlog_errno(ret);
745                 goto out_commit;
746         }
747
748         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
749                                       OCFS2_JOURNAL_ACCESS_WRITE);
750         if (ret) {
751                 mlog_errno(ret);
752                 goto out_commit;
753         }
754
755         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
756         le32_add_cpu(&rb->rf_count, 1);
757
758         ocfs2_journal_dirty(handle, ref_root_bh);
759
760         spin_lock(&oi->ip_lock);
761         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
762         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
763         di->i_refcount_loc = cpu_to_le64(refcount_loc);
764         spin_unlock(&oi->ip_lock);
765         ocfs2_journal_dirty(handle, di_bh);
766
767 out_commit:
768         ocfs2_commit_trans(osb, handle);
769 out:
770         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
771         brelse(ref_root_bh);
772
773         return ret;
774 }
775
776 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
777 {
778         int ret, delete_tree = 0;
779         handle_t *handle = NULL;
780         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
781         struct ocfs2_inode_info *oi = OCFS2_I(inode);
782         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
783         struct ocfs2_refcount_block *rb;
784         struct inode *alloc_inode = NULL;
785         struct buffer_head *alloc_bh = NULL;
786         struct buffer_head *blk_bh = NULL;
787         struct ocfs2_refcount_tree *ref_tree;
788         int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
789         u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
790         u16 bit = 0;
791
792         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
793                 return 0;
794
795         BUG_ON(!ref_blkno);
796         ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
797         if (ret) {
798                 mlog_errno(ret);
799                 return ret;
800         }
801
802         rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
803
804         /*
805          * If we are the last user, we need to free the block.
806          * So lock the allocator ahead.
807          */
808         if (le32_to_cpu(rb->rf_count) == 1) {
809                 blk = le64_to_cpu(rb->rf_blkno);
810                 bit = le16_to_cpu(rb->rf_suballoc_bit);
811                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
812
813                 alloc_inode = ocfs2_get_system_file_inode(osb,
814                                         EXTENT_ALLOC_SYSTEM_INODE,
815                                         le16_to_cpu(rb->rf_suballoc_slot));
816                 if (!alloc_inode) {
817                         ret = -ENOMEM;
818                         mlog_errno(ret);
819                         goto out;
820                 }
821                 mutex_lock(&alloc_inode->i_mutex);
822
823                 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
824                 if (ret) {
825                         mlog_errno(ret);
826                         goto out_mutex;
827                 }
828
829                 credits += OCFS2_SUBALLOC_FREE;
830         }
831
832         handle = ocfs2_start_trans(osb, credits);
833         if (IS_ERR(handle)) {
834                 ret = PTR_ERR(handle);
835                 mlog_errno(ret);
836                 goto out_unlock;
837         }
838
839         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
840                                       OCFS2_JOURNAL_ACCESS_WRITE);
841         if (ret) {
842                 mlog_errno(ret);
843                 goto out_commit;
844         }
845
846         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
847                                       OCFS2_JOURNAL_ACCESS_WRITE);
848         if (ret) {
849                 mlog_errno(ret);
850                 goto out_commit;
851         }
852
853         spin_lock(&oi->ip_lock);
854         oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
855         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
856         di->i_refcount_loc = 0;
857         spin_unlock(&oi->ip_lock);
858         ocfs2_journal_dirty(handle, di_bh);
859
860         le32_add_cpu(&rb->rf_count , -1);
861         ocfs2_journal_dirty(handle, blk_bh);
862
863         if (!rb->rf_count) {
864                 delete_tree = 1;
865                 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
866                 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
867                                                alloc_bh, bit, bg_blkno, 1);
868                 if (ret)
869                         mlog_errno(ret);
870         }
871
872 out_commit:
873         ocfs2_commit_trans(osb, handle);
874 out_unlock:
875         if (alloc_inode) {
876                 ocfs2_inode_unlock(alloc_inode, 1);
877                 brelse(alloc_bh);
878         }
879 out_mutex:
880         if (alloc_inode) {
881                 mutex_unlock(&alloc_inode->i_mutex);
882                 iput(alloc_inode);
883         }
884 out:
885         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
886         if (delete_tree)
887                 ocfs2_refcount_tree_put(ref_tree);
888         brelse(blk_bh);
889
890         return ret;
891 }
892
893 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
894                                           struct buffer_head *ref_leaf_bh,
895                                           u64 cpos, unsigned int len,
896                                           struct ocfs2_refcount_rec *ret_rec,
897                                           int *index)
898 {
899         int i = 0;
900         struct ocfs2_refcount_block *rb =
901                 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
902         struct ocfs2_refcount_rec *rec = NULL;
903
904         for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
905                 rec = &rb->rf_records.rl_recs[i];
906
907                 if (le64_to_cpu(rec->r_cpos) +
908                     le32_to_cpu(rec->r_clusters) <= cpos)
909                         continue;
910                 else if (le64_to_cpu(rec->r_cpos) > cpos)
911                         break;
912
913                 /* ok, cpos fail in this rec. Just return. */
914                 if (ret_rec)
915                         *ret_rec = *rec;
916                 goto out;
917         }
918
919         if (ret_rec) {
920                 /* We meet with a hole here, so fake the rec. */
921                 ret_rec->r_cpos = cpu_to_le64(cpos);
922                 ret_rec->r_refcount = 0;
923                 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
924                     le64_to_cpu(rec->r_cpos) < cpos + len)
925                         ret_rec->r_clusters =
926                                 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
927                 else
928                         ret_rec->r_clusters = cpu_to_le32(len);
929         }
930
931 out:
932         *index = i;
933 }
934
935 /*
936  * Try to remove refcount tree. The mechanism is:
937  * 1) Check whether i_clusters == 0, if no, exit.
938  * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
939  * 3) Check whether we have inline xattr stored outside, if yes, exit.
940  * 4) Remove the tree.
941  */
942 int ocfs2_try_remove_refcount_tree(struct inode *inode,
943                                    struct buffer_head *di_bh)
944 {
945         int ret;
946         struct ocfs2_inode_info *oi = OCFS2_I(inode);
947         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
948
949         down_write(&oi->ip_xattr_sem);
950         down_write(&oi->ip_alloc_sem);
951
952         if (oi->ip_clusters)
953                 goto out;
954
955         if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
956                 goto out;
957
958         if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
959             ocfs2_has_inline_xattr_value_outside(inode, di))
960                 goto out;
961
962         ret = ocfs2_remove_refcount_tree(inode, di_bh);
963         if (ret)
964                 mlog_errno(ret);
965 out:
966         up_write(&oi->ip_alloc_sem);
967         up_write(&oi->ip_xattr_sem);
968         return 0;
969 }
970
971 /*
972  * Given a cpos and len, try to find the refcount record which contains cpos.
973  * 1. If cpos can be found in one refcount record, return the record.
974  * 2. If cpos can't be found, return a fake record which start from cpos
975  *    and end at a small value between cpos+len and start of the next record.
976  *    This fake record has r_refcount = 0.
977  */
978 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
979                                   struct buffer_head *ref_root_bh,
980                                   u64 cpos, unsigned int len,
981                                   struct ocfs2_refcount_rec *ret_rec,
982                                   int *index,
983                                   struct buffer_head **ret_bh)
984 {
985         int ret = 0, i, found;
986         u32 low_cpos;
987         struct ocfs2_extent_list *el;
988         struct ocfs2_extent_rec *tmp, *rec = NULL;
989         struct ocfs2_extent_block *eb;
990         struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
991         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
992         struct ocfs2_refcount_block *rb =
993                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
994
995         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
996                 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
997                                               ret_rec, index);
998                 *ret_bh = ref_root_bh;
999                 get_bh(ref_root_bh);
1000                 return 0;
1001         }
1002
1003         el = &rb->rf_list;
1004         low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1005
1006         if (el->l_tree_depth) {
1007                 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1008                 if (ret) {
1009                         mlog_errno(ret);
1010                         goto out;
1011                 }
1012
1013                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1014                 el = &eb->h_list;
1015
1016                 if (el->l_tree_depth) {
1017                         ocfs2_error(sb,
1018                         "refcount tree %llu has non zero tree "
1019                         "depth in leaf btree tree block %llu\n",
1020                         (unsigned long long)ocfs2_metadata_cache_owner(ci),
1021                         (unsigned long long)eb_bh->b_blocknr);
1022                         ret = -EROFS;
1023                         goto out;
1024                 }
1025         }
1026
1027         found = 0;
1028         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1029                 rec = &el->l_recs[i];
1030
1031                 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1032                         found = 1;
1033                         break;
1034                 }
1035         }
1036
1037         /* adjust len when we have ocfs2_extent_rec after it. */
1038         if (found && i < le16_to_cpu(el->l_next_free_rec) - 1) {
1039                 tmp = &el->l_recs[i+1];
1040
1041                 if (le32_to_cpu(tmp->e_cpos) < cpos + len)
1042                         len = le32_to_cpu(tmp->e_cpos) - cpos;
1043         }
1044
1045         ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1046                                         &ref_leaf_bh);
1047         if (ret) {
1048                 mlog_errno(ret);
1049                 goto out;
1050         }
1051
1052         ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1053                                       ret_rec, index);
1054         *ret_bh = ref_leaf_bh;
1055 out:
1056         brelse(eb_bh);
1057         return ret;
1058 }
1059
1060 enum ocfs2_ref_rec_contig {
1061         REF_CONTIG_NONE = 0,
1062         REF_CONTIG_LEFT,
1063         REF_CONTIG_RIGHT,
1064         REF_CONTIG_LEFTRIGHT,
1065 };
1066
1067 static enum ocfs2_ref_rec_contig
1068         ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1069                                     int index)
1070 {
1071         if ((rb->rf_records.rl_recs[index].r_refcount ==
1072             rb->rf_records.rl_recs[index + 1].r_refcount) &&
1073             (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1074             le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1075             le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1076                 return REF_CONTIG_RIGHT;
1077
1078         return REF_CONTIG_NONE;
1079 }
1080
1081 static enum ocfs2_ref_rec_contig
1082         ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1083                                   int index)
1084 {
1085         enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1086
1087         if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1088                 ret = ocfs2_refcount_rec_adjacent(rb, index);
1089
1090         if (index > 0) {
1091                 enum ocfs2_ref_rec_contig tmp;
1092
1093                 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1094
1095                 if (tmp == REF_CONTIG_RIGHT) {
1096                         if (ret == REF_CONTIG_RIGHT)
1097                                 ret = REF_CONTIG_LEFTRIGHT;
1098                         else
1099                                 ret = REF_CONTIG_LEFT;
1100                 }
1101         }
1102
1103         return ret;
1104 }
1105
1106 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1107                                            int index)
1108 {
1109         BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1110                rb->rf_records.rl_recs[index+1].r_refcount);
1111
1112         le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1113                      le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1114
1115         if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1116                 memmove(&rb->rf_records.rl_recs[index + 1],
1117                         &rb->rf_records.rl_recs[index + 2],
1118                         sizeof(struct ocfs2_refcount_rec) *
1119                         (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1120
1121         memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1122                0, sizeof(struct ocfs2_refcount_rec));
1123         le16_add_cpu(&rb->rf_records.rl_used, -1);
1124 }
1125
1126 /*
1127  * Merge the refcount rec if we are contiguous with the adjacent recs.
1128  */
1129 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1130                                      int index)
1131 {
1132         enum ocfs2_ref_rec_contig contig =
1133                                 ocfs2_refcount_rec_contig(rb, index);
1134
1135         if (contig == REF_CONTIG_NONE)
1136                 return;
1137
1138         if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1139                 BUG_ON(index == 0);
1140                 index--;
1141         }
1142
1143         ocfs2_rotate_refcount_rec_left(rb, index);
1144
1145         if (contig == REF_CONTIG_LEFTRIGHT)
1146                 ocfs2_rotate_refcount_rec_left(rb, index);
1147 }
1148
1149 /*
1150  * Change the refcount indexed by "index" in ref_bh.
1151  * If refcount reaches 0, remove it.
1152  */
1153 static int ocfs2_change_refcount_rec(handle_t *handle,
1154                                      struct ocfs2_caching_info *ci,
1155                                      struct buffer_head *ref_leaf_bh,
1156                                      int index, int merge, int change)
1157 {
1158         int ret;
1159         struct ocfs2_refcount_block *rb =
1160                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1161         struct ocfs2_refcount_list *rl = &rb->rf_records;
1162         struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1163
1164         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1165                                       OCFS2_JOURNAL_ACCESS_WRITE);
1166         if (ret) {
1167                 mlog_errno(ret);
1168                 goto out;
1169         }
1170
1171         mlog(0, "change index %d, old count %u, change %d\n", index,
1172              le32_to_cpu(rec->r_refcount), change);
1173         le32_add_cpu(&rec->r_refcount, change);
1174
1175         if (!rec->r_refcount) {
1176                 if (index != le16_to_cpu(rl->rl_used) - 1) {
1177                         memmove(rec, rec + 1,
1178                                 (le16_to_cpu(rl->rl_used) - index - 1) *
1179                                 sizeof(struct ocfs2_refcount_rec));
1180                         memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1181                                0, sizeof(struct ocfs2_refcount_rec));
1182                 }
1183
1184                 le16_add_cpu(&rl->rl_used, -1);
1185         } else if (merge)
1186                 ocfs2_refcount_rec_merge(rb, index);
1187
1188         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1189         if (ret)
1190                 mlog_errno(ret);
1191 out:
1192         return ret;
1193 }
1194
1195 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1196                                         struct ocfs2_caching_info *ci,
1197                                         struct buffer_head *ref_root_bh,
1198                                         struct buffer_head **ref_leaf_bh,
1199                                         struct ocfs2_alloc_context *meta_ac)
1200 {
1201         int ret;
1202         u16 suballoc_bit_start;
1203         u32 num_got;
1204         u64 blkno;
1205         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1206         struct buffer_head *new_bh = NULL;
1207         struct ocfs2_refcount_block *new_rb;
1208         struct ocfs2_refcount_block *root_rb =
1209                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1210
1211         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1212                                       OCFS2_JOURNAL_ACCESS_WRITE);
1213         if (ret) {
1214                 mlog_errno(ret);
1215                 goto out;
1216         }
1217
1218         ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1219                                    &suballoc_bit_start, &num_got,
1220                                    &blkno);
1221         if (ret) {
1222                 mlog_errno(ret);
1223                 goto out;
1224         }
1225
1226         new_bh = sb_getblk(sb, blkno);
1227         if (new_bh == NULL) {
1228                 ret = -EIO;
1229                 mlog_errno(ret);
1230                 goto out;
1231         }
1232         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1233
1234         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1235                                       OCFS2_JOURNAL_ACCESS_CREATE);
1236         if (ret) {
1237                 mlog_errno(ret);
1238                 goto out;
1239         }
1240
1241         /*
1242          * Initialize ocfs2_refcount_block.
1243          * It should contain the same information as the old root.
1244          * so just memcpy it and change the corresponding field.
1245          */
1246         memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1247
1248         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1249         new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1250         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1251         new_rb->rf_blkno = cpu_to_le64(blkno);
1252         new_rb->rf_cpos = cpu_to_le32(0);
1253         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1254         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1255         ocfs2_journal_dirty(handle, new_bh);
1256
1257         /* Now change the root. */
1258         memset(&root_rb->rf_list, 0, sb->s_blocksize -
1259                offsetof(struct ocfs2_refcount_block, rf_list));
1260         root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1261         root_rb->rf_clusters = cpu_to_le32(1);
1262         root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1263         root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1264         root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1265         root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1266
1267         ocfs2_journal_dirty(handle, ref_root_bh);
1268
1269         mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno,
1270              le16_to_cpu(new_rb->rf_records.rl_used));
1271
1272         *ref_leaf_bh = new_bh;
1273         new_bh = NULL;
1274 out:
1275         brelse(new_bh);
1276         return ret;
1277 }
1278
1279 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1280                                            struct ocfs2_refcount_rec *next)
1281 {
1282         if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1283                 ocfs2_get_ref_rec_low_cpos(next))
1284                 return 1;
1285
1286         return 0;
1287 }
1288
1289 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1290 {
1291         const struct ocfs2_refcount_rec *l = a, *r = b;
1292         u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1293         u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1294
1295         if (l_cpos > r_cpos)
1296                 return 1;
1297         if (l_cpos < r_cpos)
1298                 return -1;
1299         return 0;
1300 }
1301
1302 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1303 {
1304         const struct ocfs2_refcount_rec *l = a, *r = b;
1305         u64 l_cpos = le64_to_cpu(l->r_cpos);
1306         u64 r_cpos = le64_to_cpu(r->r_cpos);
1307
1308         if (l_cpos > r_cpos)
1309                 return 1;
1310         if (l_cpos < r_cpos)
1311                 return -1;
1312         return 0;
1313 }
1314
1315 static void swap_refcount_rec(void *a, void *b, int size)
1316 {
1317         struct ocfs2_refcount_rec *l = a, *r = b, tmp;
1318
1319         tmp = *(struct ocfs2_refcount_rec *)l;
1320         *(struct ocfs2_refcount_rec *)l =
1321                         *(struct ocfs2_refcount_rec *)r;
1322         *(struct ocfs2_refcount_rec *)r = tmp;
1323 }
1324
1325 /*
1326  * The refcount cpos are ordered by their 64bit cpos,
1327  * But we will use the low 32 bit to be the e_cpos in the b-tree.
1328  * So we need to make sure that this pos isn't intersected with others.
1329  *
1330  * Note: The refcount block is already sorted by their low 32 bit cpos,
1331  *       So just try the middle pos first, and we will exit when we find
1332  *       the good position.
1333  */
1334 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1335                                          u32 *split_pos, int *split_index)
1336 {
1337         int num_used = le16_to_cpu(rl->rl_used);
1338         int delta, middle = num_used / 2;
1339
1340         for (delta = 0; delta < middle; delta++) {
1341                 /* Let's check delta earlier than middle */
1342                 if (ocfs2_refcount_rec_no_intersect(
1343                                         &rl->rl_recs[middle - delta - 1],
1344                                         &rl->rl_recs[middle - delta])) {
1345                         *split_index = middle - delta;
1346                         break;
1347                 }
1348
1349                 /* For even counts, don't walk off the end */
1350                 if ((middle + delta + 1) == num_used)
1351                         continue;
1352
1353                 /* Now try delta past middle */
1354                 if (ocfs2_refcount_rec_no_intersect(
1355                                         &rl->rl_recs[middle + delta],
1356                                         &rl->rl_recs[middle + delta + 1])) {
1357                         *split_index = middle + delta + 1;
1358                         break;
1359                 }
1360         }
1361
1362         if (delta >= middle)
1363                 return -ENOSPC;
1364
1365         *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1366         return 0;
1367 }
1368
1369 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1370                                             struct buffer_head *new_bh,
1371                                             u32 *split_cpos)
1372 {
1373         int split_index = 0, num_moved, ret;
1374         u32 cpos = 0;
1375         struct ocfs2_refcount_block *rb =
1376                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1377         struct ocfs2_refcount_list *rl = &rb->rf_records;
1378         struct ocfs2_refcount_block *new_rb =
1379                         (struct ocfs2_refcount_block *)new_bh->b_data;
1380         struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1381
1382         mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n",
1383              (unsigned long long)ref_leaf_bh->b_blocknr,
1384              le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used));
1385
1386         /*
1387          * XXX: Improvement later.
1388          * If we know all the high 32 bit cpos is the same, no need to sort.
1389          *
1390          * In order to make the whole process safe, we do:
1391          * 1. sort the entries by their low 32 bit cpos first so that we can
1392          *    find the split cpos easily.
1393          * 2. call ocfs2_insert_extent to insert the new refcount block.
1394          * 3. move the refcount rec to the new block.
1395          * 4. sort the entries by their 64 bit cpos.
1396          * 5. dirty the new_rb and rb.
1397          */
1398         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1399              sizeof(struct ocfs2_refcount_rec),
1400              cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1401
1402         ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1403         if (ret) {
1404                 mlog_errno(ret);
1405                 return ret;
1406         }
1407
1408         new_rb->rf_cpos = cpu_to_le32(cpos);
1409
1410         /* move refcount records starting from split_index to the new block. */
1411         num_moved = le16_to_cpu(rl->rl_used) - split_index;
1412         memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1413                num_moved * sizeof(struct ocfs2_refcount_rec));
1414
1415         /*ok, remove the entries we just moved over to the other block. */
1416         memset(&rl->rl_recs[split_index], 0,
1417                num_moved * sizeof(struct ocfs2_refcount_rec));
1418
1419         /* change old and new rl_used accordingly. */
1420         le16_add_cpu(&rl->rl_used, -num_moved);
1421         new_rl->rl_used = cpu_to_le32(num_moved);
1422
1423         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1424              sizeof(struct ocfs2_refcount_rec),
1425              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1426
1427         sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1428              sizeof(struct ocfs2_refcount_rec),
1429              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1430
1431         *split_cpos = cpos;
1432         return 0;
1433 }
1434
1435 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1436                                          struct ocfs2_caching_info *ci,
1437                                          struct buffer_head *ref_root_bh,
1438                                          struct buffer_head *ref_leaf_bh,
1439                                          struct ocfs2_alloc_context *meta_ac)
1440 {
1441         int ret;
1442         u16 suballoc_bit_start;
1443         u32 num_got, new_cpos;
1444         u64 blkno;
1445         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1446         struct ocfs2_refcount_block *root_rb =
1447                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1448         struct buffer_head *new_bh = NULL;
1449         struct ocfs2_refcount_block *new_rb;
1450         struct ocfs2_extent_tree ref_et;
1451
1452         BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1453
1454         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1455                                       OCFS2_JOURNAL_ACCESS_WRITE);
1456         if (ret) {
1457                 mlog_errno(ret);
1458                 goto out;
1459         }
1460
1461         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1462                                       OCFS2_JOURNAL_ACCESS_WRITE);
1463         if (ret) {
1464                 mlog_errno(ret);
1465                 goto out;
1466         }
1467
1468         ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1469                                    &suballoc_bit_start, &num_got,
1470                                    &blkno);
1471         if (ret) {
1472                 mlog_errno(ret);
1473                 goto out;
1474         }
1475
1476         new_bh = sb_getblk(sb, blkno);
1477         if (new_bh == NULL) {
1478                 ret = -EIO;
1479                 mlog_errno(ret);
1480                 goto out;
1481         }
1482         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1483
1484         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1485                                       OCFS2_JOURNAL_ACCESS_CREATE);
1486         if (ret) {
1487                 mlog_errno(ret);
1488                 goto out;
1489         }
1490
1491         /* Initialize ocfs2_refcount_block. */
1492         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1493         memset(new_rb, 0, sb->s_blocksize);
1494         strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1495         new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1496         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1497         new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1498         new_rb->rf_blkno = cpu_to_le64(blkno);
1499         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1500         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1501         new_rb->rf_records.rl_count =
1502                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1503         new_rb->rf_generation = root_rb->rf_generation;
1504
1505         ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1506         if (ret) {
1507                 mlog_errno(ret);
1508                 goto out;
1509         }
1510
1511         ocfs2_journal_dirty(handle, ref_leaf_bh);
1512         ocfs2_journal_dirty(handle, new_bh);
1513
1514         ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1515
1516         mlog(0, "insert new leaf block %llu at %u\n",
1517              (unsigned long long)new_bh->b_blocknr, new_cpos);
1518
1519         /* Insert the new leaf block with the specific offset cpos. */
1520         ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1521                                   1, 0, meta_ac);
1522         if (ret)
1523                 mlog_errno(ret);
1524
1525 out:
1526         brelse(new_bh);
1527         return ret;
1528 }
1529
1530 static int ocfs2_expand_refcount_tree(handle_t *handle,
1531                                       struct ocfs2_caching_info *ci,
1532                                       struct buffer_head *ref_root_bh,
1533                                       struct buffer_head *ref_leaf_bh,
1534                                       struct ocfs2_alloc_context *meta_ac)
1535 {
1536         int ret;
1537         struct buffer_head *expand_bh = NULL;
1538
1539         if (ref_root_bh == ref_leaf_bh) {
1540                 /*
1541                  * the old root bh hasn't been expanded to a b-tree,
1542                  * so expand it first.
1543                  */
1544                 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1545                                                    &expand_bh, meta_ac);
1546                 if (ret) {
1547                         mlog_errno(ret);
1548                         goto out;
1549                 }
1550         } else {
1551                 expand_bh = ref_leaf_bh;
1552                 get_bh(expand_bh);
1553         }
1554
1555
1556         /* Now add a new refcount block into the tree.*/
1557         ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1558                                             expand_bh, meta_ac);
1559         if (ret)
1560                 mlog_errno(ret);
1561 out:
1562         brelse(expand_bh);
1563         return ret;
1564 }
1565
1566 /*
1567  * Adjust the extent rec in b-tree representing ref_leaf_bh.
1568  *
1569  * Only called when we have inserted a new refcount rec at index 0
1570  * which means ocfs2_extent_rec.e_cpos may need some change.
1571  */
1572 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1573                                      struct ocfs2_caching_info *ci,
1574                                      struct buffer_head *ref_root_bh,
1575                                      struct buffer_head *ref_leaf_bh,
1576                                      struct ocfs2_refcount_rec *rec)
1577 {
1578         int ret = 0, i;
1579         u32 new_cpos, old_cpos;
1580         struct ocfs2_path *path = NULL;
1581         struct ocfs2_extent_tree et;
1582         struct ocfs2_refcount_block *rb =
1583                 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1584         struct ocfs2_extent_list *el;
1585
1586         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1587                 goto out;
1588
1589         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1590         old_cpos = le32_to_cpu(rb->rf_cpos);
1591         new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1592         if (old_cpos <= new_cpos)
1593                 goto out;
1594
1595         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1596
1597         path = ocfs2_new_path_from_et(&et);
1598         if (!path) {
1599                 ret = -ENOMEM;
1600                 mlog_errno(ret);
1601                 goto out;
1602         }
1603
1604         ret = ocfs2_find_path(ci, path, old_cpos);
1605         if (ret) {
1606                 mlog_errno(ret);
1607                 goto out;
1608         }
1609
1610         /*
1611          * 2 more credits, one for the leaf refcount block, one for
1612          * the extent block contains the extent rec.
1613          */
1614         ret = ocfs2_extend_trans(handle, handle->h_buffer_credits + 2);
1615         if (ret < 0) {
1616                 mlog_errno(ret);
1617                 goto out;
1618         }
1619
1620         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1621                                       OCFS2_JOURNAL_ACCESS_WRITE);
1622         if (ret < 0) {
1623                 mlog_errno(ret);
1624                 goto out;
1625         }
1626
1627         ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1628                                       OCFS2_JOURNAL_ACCESS_WRITE);
1629         if (ret < 0) {
1630                 mlog_errno(ret);
1631                 goto out;
1632         }
1633
1634         /* change the leaf extent block first. */
1635         el = path_leaf_el(path);
1636
1637         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1638                 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1639                         break;
1640
1641         BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1642
1643         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1644
1645         /* change the r_cpos in the leaf block. */
1646         rb->rf_cpos = cpu_to_le32(new_cpos);
1647
1648         ocfs2_journal_dirty(handle, path_leaf_bh(path));
1649         ocfs2_journal_dirty(handle, ref_leaf_bh);
1650
1651 out:
1652         ocfs2_free_path(path);
1653         return ret;
1654 }
1655
1656 static int ocfs2_insert_refcount_rec(handle_t *handle,
1657                                      struct ocfs2_caching_info *ci,
1658                                      struct buffer_head *ref_root_bh,
1659                                      struct buffer_head *ref_leaf_bh,
1660                                      struct ocfs2_refcount_rec *rec,
1661                                      int index, int merge,
1662                                      struct ocfs2_alloc_context *meta_ac)
1663 {
1664         int ret;
1665         struct ocfs2_refcount_block *rb =
1666                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1667         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1668         struct buffer_head *new_bh = NULL;
1669
1670         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1671
1672         if (rf_list->rl_used == rf_list->rl_count) {
1673                 u64 cpos = le64_to_cpu(rec->r_cpos);
1674                 u32 len = le32_to_cpu(rec->r_clusters);
1675
1676                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1677                                                  ref_leaf_bh, meta_ac);
1678                 if (ret) {
1679                         mlog_errno(ret);
1680                         goto out;
1681                 }
1682
1683                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1684                                              cpos, len, NULL, &index,
1685                                              &new_bh);
1686                 if (ret) {
1687                         mlog_errno(ret);
1688                         goto out;
1689                 }
1690
1691                 ref_leaf_bh = new_bh;
1692                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1693                 rf_list = &rb->rf_records;
1694         }
1695
1696         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1697                                       OCFS2_JOURNAL_ACCESS_WRITE);
1698         if (ret) {
1699                 mlog_errno(ret);
1700                 goto out;
1701         }
1702
1703         if (index < le16_to_cpu(rf_list->rl_used))
1704                 memmove(&rf_list->rl_recs[index + 1],
1705                         &rf_list->rl_recs[index],
1706                         (le16_to_cpu(rf_list->rl_used) - index) *
1707                          sizeof(struct ocfs2_refcount_rec));
1708
1709         mlog(0, "insert refcount record start %llu, len %u, count %u "
1710              "to leaf block %llu at index %d\n",
1711              (unsigned long long)le64_to_cpu(rec->r_cpos),
1712              le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount),
1713              (unsigned long long)ref_leaf_bh->b_blocknr, index);
1714
1715         rf_list->rl_recs[index] = *rec;
1716
1717         le16_add_cpu(&rf_list->rl_used, 1);
1718
1719         if (merge)
1720                 ocfs2_refcount_rec_merge(rb, index);
1721
1722         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1723         if (ret) {
1724                 mlog_errno(ret);
1725                 goto out;
1726         }
1727
1728         if (index == 0) {
1729                 ret = ocfs2_adjust_refcount_rec(handle, ci,
1730                                                 ref_root_bh,
1731                                                 ref_leaf_bh, rec);
1732                 if (ret)
1733                         mlog_errno(ret);
1734         }
1735 out:
1736         brelse(new_bh);
1737         return ret;
1738 }
1739
1740 /*
1741  * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1742  * This is much simple than our b-tree code.
1743  * split_rec is the new refcount rec we want to insert.
1744  * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1745  * increase refcount or decrease a refcount to non-zero).
1746  * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1747  * rec( in case we decrease a refcount to zero).
1748  */
1749 static int ocfs2_split_refcount_rec(handle_t *handle,
1750                                     struct ocfs2_caching_info *ci,
1751                                     struct buffer_head *ref_root_bh,
1752                                     struct buffer_head *ref_leaf_bh,
1753                                     struct ocfs2_refcount_rec *split_rec,
1754                                     int index, int merge,
1755                                     struct ocfs2_alloc_context *meta_ac,
1756                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
1757 {
1758         int ret, recs_need;
1759         u32 len;
1760         struct ocfs2_refcount_block *rb =
1761                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1762         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1763         struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1764         struct ocfs2_refcount_rec *tail_rec = NULL;
1765         struct buffer_head *new_bh = NULL;
1766
1767         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1768
1769         mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n",
1770              le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters),
1771              le64_to_cpu(split_rec->r_cpos),
1772              le32_to_cpu(split_rec->r_clusters));
1773
1774         /*
1775          * If we just need to split the header or tail clusters,
1776          * no more recs are needed, just split is OK.
1777          * Otherwise we at least need one new recs.
1778          */
1779         if (!split_rec->r_refcount &&
1780             (split_rec->r_cpos == orig_rec->r_cpos ||
1781              le64_to_cpu(split_rec->r_cpos) +
1782              le32_to_cpu(split_rec->r_clusters) ==
1783              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1784                 recs_need = 0;
1785         else
1786                 recs_need = 1;
1787
1788         /*
1789          * We need one more rec if we split in the middle and the new rec have
1790          * some refcount in it.
1791          */
1792         if (split_rec->r_refcount &&
1793             (split_rec->r_cpos != orig_rec->r_cpos &&
1794              le64_to_cpu(split_rec->r_cpos) +
1795              le32_to_cpu(split_rec->r_clusters) !=
1796              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1797                 recs_need++;
1798
1799         /* If the leaf block don't have enough record, expand it. */
1800         if (le16_to_cpu(rf_list->rl_used) + recs_need > rf_list->rl_count) {
1801                 struct ocfs2_refcount_rec tmp_rec;
1802                 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1803                 len = le32_to_cpu(orig_rec->r_clusters);
1804                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1805                                                  ref_leaf_bh, meta_ac);
1806                 if (ret) {
1807                         mlog_errno(ret);
1808                         goto out;
1809                 }
1810
1811                 /*
1812                  * We have to re-get it since now cpos may be moved to
1813                  * another leaf block.
1814                  */
1815                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1816                                              cpos, len, &tmp_rec, &index,
1817                                              &new_bh);
1818                 if (ret) {
1819                         mlog_errno(ret);
1820                         goto out;
1821                 }
1822
1823                 ref_leaf_bh = new_bh;
1824                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1825                 rf_list = &rb->rf_records;
1826                 orig_rec = &rf_list->rl_recs[index];
1827         }
1828
1829         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1830                                       OCFS2_JOURNAL_ACCESS_WRITE);
1831         if (ret) {
1832                 mlog_errno(ret);
1833                 goto out;
1834         }
1835
1836         /*
1837          * We have calculated out how many new records we need and store
1838          * in recs_need, so spare enough space first by moving the records
1839          * after "index" to the end.
1840          */
1841         if (index != le16_to_cpu(rf_list->rl_used) - 1)
1842                 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1843                         &rf_list->rl_recs[index + 1],
1844                         (le16_to_cpu(rf_list->rl_used) - index - 1) *
1845                          sizeof(struct ocfs2_refcount_rec));
1846
1847         len = (le64_to_cpu(orig_rec->r_cpos) +
1848               le32_to_cpu(orig_rec->r_clusters)) -
1849               (le64_to_cpu(split_rec->r_cpos) +
1850               le32_to_cpu(split_rec->r_clusters));
1851
1852         /*
1853          * If we have "len", the we will split in the tail and move it
1854          * to the end of the space we have just spared.
1855          */
1856         if (len) {
1857                 tail_rec = &rf_list->rl_recs[index + recs_need];
1858
1859                 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1860                 le64_add_cpu(&tail_rec->r_cpos,
1861                              le32_to_cpu(tail_rec->r_clusters) - len);
1862                 tail_rec->r_clusters = le32_to_cpu(len);
1863         }
1864
1865         /*
1866          * If the split pos isn't the same as the original one, we need to
1867          * split in the head.
1868          *
1869          * Note: We have the chance that split_rec.r_refcount = 0,
1870          * recs_need = 0 and len > 0, which means we just cut the head from
1871          * the orig_rec and in that case we have done some modification in
1872          * orig_rec above, so the check for r_cpos is faked.
1873          */
1874         if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1875                 len = le64_to_cpu(split_rec->r_cpos) -
1876                       le64_to_cpu(orig_rec->r_cpos);
1877                 orig_rec->r_clusters = cpu_to_le32(len);
1878                 index++;
1879         }
1880
1881         le16_add_cpu(&rf_list->rl_used, recs_need);
1882
1883         if (split_rec->r_refcount) {
1884                 rf_list->rl_recs[index] = *split_rec;
1885                 mlog(0, "insert refcount record start %llu, len %u, count %u "
1886                      "to leaf block %llu at index %d\n",
1887                      (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1888                      le32_to_cpu(split_rec->r_clusters),
1889                      le32_to_cpu(split_rec->r_refcount),
1890                      (unsigned long long)ref_leaf_bh->b_blocknr, index);
1891
1892                 if (merge)
1893                         ocfs2_refcount_rec_merge(rb, index);
1894         }
1895
1896         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1897         if (ret)
1898                 mlog_errno(ret);
1899
1900 out:
1901         brelse(new_bh);
1902         return ret;
1903 }
1904
1905 static int __ocfs2_increase_refcount(handle_t *handle,
1906                                      struct ocfs2_caching_info *ci,
1907                                      struct buffer_head *ref_root_bh,
1908                                      u64 cpos, u32 len, int merge,
1909                                      struct ocfs2_alloc_context *meta_ac,
1910                                      struct ocfs2_cached_dealloc_ctxt *dealloc)
1911 {
1912         int ret = 0, index;
1913         struct buffer_head *ref_leaf_bh = NULL;
1914         struct ocfs2_refcount_rec rec;
1915         unsigned int set_len = 0;
1916
1917         mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n",
1918              (unsigned long long)ocfs2_metadata_cache_owner(ci),
1919              (unsigned long long)cpos, len);
1920
1921         while (len) {
1922                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1923                                              cpos, len, &rec, &index,
1924                                              &ref_leaf_bh);
1925                 if (ret) {
1926                         mlog_errno(ret);
1927                         goto out;
1928                 }
1929
1930                 set_len = le32_to_cpu(rec.r_clusters);
1931
1932                 /*
1933                  * Here we may meet with 3 situations:
1934                  *
1935                  * 1. If we find an already existing record, and the length
1936                  *    is the same, cool, we just need to increase the r_refcount
1937                  *    and it is OK.
1938                  * 2. If we find a hole, just insert it with r_refcount = 1.
1939                  * 3. If we are in the middle of one extent record, split
1940                  *    it.
1941                  */
1942                 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
1943                     set_len <= len) {
1944                         mlog(0, "increase refcount rec, start %llu, len %u, "
1945                              "count %u\n", (unsigned long long)cpos, set_len,
1946                              le32_to_cpu(rec.r_refcount));
1947                         ret = ocfs2_change_refcount_rec(handle, ci,
1948                                                         ref_leaf_bh, index,
1949                                                         merge, 1);
1950                         if (ret) {
1951                                 mlog_errno(ret);
1952                                 goto out;
1953                         }
1954                 } else if (!rec.r_refcount) {
1955                         rec.r_refcount = cpu_to_le32(1);
1956
1957                         mlog(0, "insert refcount rec, start %llu, len %u\n",
1958                              (unsigned long long)le64_to_cpu(rec.r_cpos),
1959                              set_len);
1960                         ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
1961                                                         ref_leaf_bh,
1962                                                         &rec, index,
1963                                                         merge, meta_ac);
1964                         if (ret) {
1965                                 mlog_errno(ret);
1966                                 goto out;
1967                         }
1968                 } else  {
1969                         set_len = min((u64)(cpos + len),
1970                                       le64_to_cpu(rec.r_cpos) + set_len) - cpos;
1971                         rec.r_cpos = cpu_to_le64(cpos);
1972                         rec.r_clusters = cpu_to_le32(set_len);
1973                         le32_add_cpu(&rec.r_refcount, 1);
1974
1975                         mlog(0, "split refcount rec, start %llu, "
1976                              "len %u, count %u\n",
1977                              (unsigned long long)le64_to_cpu(rec.r_cpos),
1978                              set_len, le32_to_cpu(rec.r_refcount));
1979                         ret = ocfs2_split_refcount_rec(handle, ci,
1980                                                        ref_root_bh, ref_leaf_bh,
1981                                                        &rec, index, merge,
1982                                                        meta_ac, dealloc);
1983                         if (ret) {
1984                                 mlog_errno(ret);
1985                                 goto out;
1986                         }
1987                 }
1988
1989                 cpos += set_len;
1990                 len -= set_len;
1991                 brelse(ref_leaf_bh);
1992                 ref_leaf_bh = NULL;
1993         }
1994
1995 out:
1996         brelse(ref_leaf_bh);
1997         return ret;
1998 }
1999
2000 static int ocfs2_remove_refcount_extent(handle_t *handle,
2001                                 struct ocfs2_caching_info *ci,
2002                                 struct buffer_head *ref_root_bh,
2003                                 struct buffer_head *ref_leaf_bh,
2004                                 struct ocfs2_alloc_context *meta_ac,
2005                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2006 {
2007         int ret;
2008         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2009         struct ocfs2_refcount_block *rb =
2010                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2011         struct ocfs2_extent_tree et;
2012
2013         BUG_ON(rb->rf_records.rl_used);
2014
2015         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2016         ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
2017                                   1, meta_ac, dealloc);
2018         if (ret) {
2019                 mlog_errno(ret);
2020                 goto out;
2021         }
2022
2023         ocfs2_remove_from_cache(ci, ref_leaf_bh);
2024
2025         /*
2026          * add the freed block to the dealloc so that it will be freed
2027          * when we run dealloc.
2028          */
2029         ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
2030                                         le16_to_cpu(rb->rf_suballoc_slot),
2031                                         le64_to_cpu(rb->rf_blkno),
2032                                         le16_to_cpu(rb->rf_suballoc_bit));
2033         if (ret) {
2034                 mlog_errno(ret);
2035                 goto out;
2036         }
2037
2038         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
2039                                       OCFS2_JOURNAL_ACCESS_WRITE);
2040         if (ret) {
2041                 mlog_errno(ret);
2042                 goto out;
2043         }
2044
2045         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2046
2047         le32_add_cpu(&rb->rf_clusters, -1);
2048
2049         /*
2050          * check whether we need to restore the root refcount block if
2051          * there is no leaf extent block at atll.
2052          */
2053         if (!rb->rf_list.l_next_free_rec) {
2054                 BUG_ON(rb->rf_clusters);
2055
2056                 mlog(0, "reset refcount tree root %llu to be a record block.\n",
2057                      (unsigned long long)ref_root_bh->b_blocknr);
2058
2059                 rb->rf_flags = 0;
2060                 rb->rf_parent = 0;
2061                 rb->rf_cpos = 0;
2062                 memset(&rb->rf_records, 0, sb->s_blocksize -
2063                        offsetof(struct ocfs2_refcount_block, rf_records));
2064                 rb->rf_records.rl_count =
2065                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
2066         }
2067
2068         ocfs2_journal_dirty(handle, ref_root_bh);
2069
2070 out:
2071         return ret;
2072 }
2073
2074 int ocfs2_increase_refcount(handle_t *handle,
2075                             struct ocfs2_caching_info *ci,
2076                             struct buffer_head *ref_root_bh,
2077                             u64 cpos, u32 len,
2078                             struct ocfs2_alloc_context *meta_ac,
2079                             struct ocfs2_cached_dealloc_ctxt *dealloc)
2080 {
2081         return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
2082                                          cpos, len, 1,
2083                                          meta_ac, dealloc);
2084 }
2085
2086 static int ocfs2_decrease_refcount_rec(handle_t *handle,
2087                                 struct ocfs2_caching_info *ci,
2088                                 struct buffer_head *ref_root_bh,
2089                                 struct buffer_head *ref_leaf_bh,
2090                                 int index, u64 cpos, unsigned int len,
2091                                 struct ocfs2_alloc_context *meta_ac,
2092                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2093 {
2094         int ret;
2095         struct ocfs2_refcount_block *rb =
2096                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2097         struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2098
2099         BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2100         BUG_ON(cpos + len >
2101                le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2102
2103         if (cpos == le64_to_cpu(rec->r_cpos) &&
2104             len == le32_to_cpu(rec->r_clusters))
2105                 ret = ocfs2_change_refcount_rec(handle, ci,
2106                                                 ref_leaf_bh, index, 1, -1);
2107         else {
2108                 struct ocfs2_refcount_rec split = *rec;
2109                 split.r_cpos = cpu_to_le64(cpos);
2110                 split.r_clusters = cpu_to_le32(len);
2111
2112                 le32_add_cpu(&split.r_refcount, -1);
2113
2114                 mlog(0, "split refcount rec, start %llu, "
2115                      "len %u, count %u, original start %llu, len %u\n",
2116                      (unsigned long long)le64_to_cpu(split.r_cpos),
2117                      len, le32_to_cpu(split.r_refcount),
2118                      (unsigned long long)le64_to_cpu(rec->r_cpos),
2119                      le32_to_cpu(rec->r_clusters));
2120                 ret = ocfs2_split_refcount_rec(handle, ci,
2121                                                ref_root_bh, ref_leaf_bh,
2122                                                &split, index, 1,
2123                                                meta_ac, dealloc);
2124         }
2125
2126         if (ret) {
2127                 mlog_errno(ret);
2128                 goto out;
2129         }
2130
2131         /* Remove the leaf refcount block if it contains no refcount record. */
2132         if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2133                 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2134                                                    ref_leaf_bh, meta_ac,
2135                                                    dealloc);
2136                 if (ret)
2137                         mlog_errno(ret);
2138         }
2139
2140 out:
2141         return ret;
2142 }
2143
2144 static int __ocfs2_decrease_refcount(handle_t *handle,
2145                                      struct ocfs2_caching_info *ci,
2146                                      struct buffer_head *ref_root_bh,
2147                                      u64 cpos, u32 len,
2148                                      struct ocfs2_alloc_context *meta_ac,
2149                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2150                                      int delete)
2151 {
2152         int ret = 0, index = 0;
2153         struct ocfs2_refcount_rec rec;
2154         unsigned int r_count = 0, r_len;
2155         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2156         struct buffer_head *ref_leaf_bh = NULL;
2157
2158         mlog(0, "Tree owner %llu, decrease refcount start %llu, "
2159              "len %u, delete %u\n",
2160              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2161              (unsigned long long)cpos, len, delete);
2162
2163         while (len) {
2164                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2165                                              cpos, len, &rec, &index,
2166                                              &ref_leaf_bh);
2167                 if (ret) {
2168                         mlog_errno(ret);
2169                         goto out;
2170                 }
2171
2172                 r_count = le32_to_cpu(rec.r_refcount);
2173                 BUG_ON(r_count == 0);
2174                 if (!delete)
2175                         BUG_ON(r_count > 1);
2176
2177                 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2178                               le32_to_cpu(rec.r_clusters)) - cpos;
2179
2180                 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2181                                                   ref_leaf_bh, index,
2182                                                   cpos, r_len,
2183                                                   meta_ac, dealloc);
2184                 if (ret) {
2185                         mlog_errno(ret);
2186                         goto out;
2187                 }
2188
2189                 if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
2190                         ret = ocfs2_cache_cluster_dealloc(dealloc,
2191                                           ocfs2_clusters_to_blocks(sb, cpos),
2192                                                           r_len);
2193                         if (ret) {
2194                                 mlog_errno(ret);
2195                                 goto out;
2196                         }
2197                 }
2198
2199                 cpos += r_len;
2200                 len -= r_len;
2201                 brelse(ref_leaf_bh);
2202                 ref_leaf_bh = NULL;
2203         }
2204
2205 out:
2206         brelse(ref_leaf_bh);
2207         return ret;
2208 }
2209
2210 /* Caller must hold refcount tree lock. */
2211 int ocfs2_decrease_refcount(struct inode *inode,
2212                             handle_t *handle, u32 cpos, u32 len,
2213                             struct ocfs2_alloc_context *meta_ac,
2214                             struct ocfs2_cached_dealloc_ctxt *dealloc,
2215                             int delete)
2216 {
2217         int ret;
2218         u64 ref_blkno;
2219         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2220         struct buffer_head *ref_root_bh = NULL;
2221         struct ocfs2_refcount_tree *tree;
2222
2223         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2224
2225         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2226         if (ret) {
2227                 mlog_errno(ret);
2228                 goto out;
2229         }
2230
2231         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2232         if (ret) {
2233                 mlog_errno(ret);
2234                 goto out;
2235         }
2236
2237         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2238                                         &ref_root_bh);
2239         if (ret) {
2240                 mlog_errno(ret);
2241                 goto out;
2242         }
2243
2244         ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2245                                         cpos, len, meta_ac, dealloc, delete);
2246         if (ret)
2247                 mlog_errno(ret);
2248 out:
2249         brelse(ref_root_bh);
2250         return ret;
2251 }
2252
2253 /*
2254  * Mark the already-existing extent at cpos as refcounted for len clusters.
2255  * This adds the refcount extent flag.
2256  *
2257  * If the existing extent is larger than the request, initiate a
2258  * split. An attempt will be made at merging with adjacent extents.
2259  *
2260  * The caller is responsible for passing down meta_ac if we'll need it.
2261  */
2262 static int ocfs2_mark_extent_refcounted(struct inode *inode,
2263                                 struct ocfs2_extent_tree *et,
2264                                 handle_t *handle, u32 cpos,
2265                                 u32 len, u32 phys,
2266                                 struct ocfs2_alloc_context *meta_ac,
2267                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2268 {
2269         int ret;
2270
2271         mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n",
2272              inode->i_ino, cpos, len, phys);
2273
2274         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2275                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2276                             "tree, but the feature bit is not set in the "
2277                             "super block.", inode->i_ino);
2278                 ret = -EROFS;
2279                 goto out;
2280         }
2281
2282         ret = ocfs2_change_extent_flag(handle, et, cpos,
2283                                        len, phys, meta_ac, dealloc,
2284                                        OCFS2_EXT_REFCOUNTED, 0);
2285         if (ret)
2286                 mlog_errno(ret);
2287
2288 out:
2289         return ret;
2290 }
2291
2292 /*
2293  * Given some contiguous physical clusters, calculate what we need
2294  * for modifying their refcount.
2295  */
2296 static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2297                                             struct ocfs2_caching_info *ci,
2298                                             struct buffer_head *ref_root_bh,
2299                                             u64 start_cpos,
2300                                             u32 clusters,
2301                                             int *meta_add,
2302                                             int *credits)
2303 {
2304         int ret = 0, index, ref_blocks = 0, recs_add = 0;
2305         u64 cpos = start_cpos;
2306         struct ocfs2_refcount_block *rb;
2307         struct ocfs2_refcount_rec rec;
2308         struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2309         u32 len;
2310
2311         mlog(0, "start_cpos %llu, clusters %u\n",
2312              (unsigned long long)start_cpos, clusters);
2313         while (clusters) {
2314                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2315                                              cpos, clusters, &rec,
2316                                              &index, &ref_leaf_bh);
2317                 if (ret) {
2318                         mlog_errno(ret);
2319                         goto out;
2320                 }
2321
2322                 if (ref_leaf_bh != prev_bh) {
2323                         /*
2324                          * Now we encounter a new leaf block, so calculate
2325                          * whether we need to extend the old leaf.
2326                          */
2327                         if (prev_bh) {
2328                                 rb = (struct ocfs2_refcount_block *)
2329                                                         prev_bh->b_data;
2330
2331                                 if (le64_to_cpu(rb->rf_records.rl_used) +
2332                                     recs_add >
2333                                     le16_to_cpu(rb->rf_records.rl_count))
2334                                         ref_blocks++;
2335                         }
2336
2337                         recs_add = 0;
2338                         *credits += 1;
2339                         brelse(prev_bh);
2340                         prev_bh = ref_leaf_bh;
2341                         get_bh(prev_bh);
2342                 }
2343
2344                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2345
2346                 mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu,"
2347                      "rec->r_clusters %u, rec->r_refcount %u, index %d\n",
2348                      recs_add, (unsigned long long)cpos, clusters,
2349                      (unsigned long long)le64_to_cpu(rec.r_cpos),
2350                      le32_to_cpu(rec.r_clusters),
2351                      le32_to_cpu(rec.r_refcount), index);
2352
2353                 len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2354                           le32_to_cpu(rec.r_clusters)) - cpos;
2355                 /*
2356                  * If the refcount rec already exist, cool. We just need
2357                  * to check whether there is a split. Otherwise we just need
2358                  * to increase the refcount.
2359                  * If we will insert one, increases recs_add.
2360                  *
2361                  * We record all the records which will be inserted to the
2362                  * same refcount block, so that we can tell exactly whether
2363                  * we need a new refcount block or not.
2364                  */
2365                 if (rec.r_refcount) {
2366                         /* Check whether we need a split at the beginning. */
2367                         if (cpos == start_cpos &&
2368                             cpos != le64_to_cpu(rec.r_cpos))
2369                                 recs_add++;
2370
2371                         /* Check whether we need a split in the end. */
2372                         if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2373                             le32_to_cpu(rec.r_clusters))
2374                                 recs_add++;
2375                 } else
2376                         recs_add++;
2377
2378                 brelse(ref_leaf_bh);
2379                 ref_leaf_bh = NULL;
2380                 clusters -= len;
2381                 cpos += len;
2382         }
2383
2384         if (prev_bh) {
2385                 rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2386
2387                 if (le64_to_cpu(rb->rf_records.rl_used) + recs_add >
2388                     le16_to_cpu(rb->rf_records.rl_count))
2389                         ref_blocks++;
2390
2391                 *credits += 1;
2392         }
2393
2394         if (!ref_blocks)
2395                 goto out;
2396
2397         mlog(0, "we need ref_blocks %d\n", ref_blocks);
2398         *meta_add += ref_blocks;
2399         *credits += ref_blocks;
2400
2401         /*
2402          * So we may need ref_blocks to insert into the tree.
2403          * That also means we need to change the b-tree and add that number
2404          * of records since we never merge them.
2405          * We need one more block for expansion since the new created leaf
2406          * block is also full and needs split.
2407          */
2408         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2409         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2410                 struct ocfs2_extent_tree et;
2411
2412                 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2413                 *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2414                 *credits += ocfs2_calc_extend_credits(sb,
2415                                                       et.et_root_el,
2416                                                       ref_blocks);
2417         } else {
2418                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2419                 *meta_add += 1;
2420         }
2421
2422 out:
2423         brelse(ref_leaf_bh);
2424         brelse(prev_bh);
2425         return ret;
2426 }
2427
2428 /*
2429  * For refcount tree, we will decrease some contiguous clusters
2430  * refcount count, so just go through it to see how many blocks
2431  * we gonna touch and whether we need to create new blocks.
2432  *
2433  * Normally the refcount blocks store these refcount should be
2434  * contiguous also, so that we can get the number easily.
2435  * As for meta_ac, we will at most add split 2 refcount record and
2436  * 2 more refcount block, so just check it in a rough way.
2437  *
2438  * Caller must hold refcount tree lock.
2439  */
2440 int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
2441                                           struct buffer_head *di_bh,
2442                                           u64 phys_blkno,
2443                                           u32 clusters,
2444                                           int *credits,
2445                                           struct ocfs2_alloc_context **meta_ac)
2446 {
2447         int ret, ref_blocks = 0;
2448         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2449         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2450         struct buffer_head *ref_root_bh = NULL;
2451         struct ocfs2_refcount_tree *tree;
2452         u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2453
2454         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2455                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2456                             "tree, but the feature bit is not set in the "
2457                             "super block.", inode->i_ino);
2458                 ret = -EROFS;
2459                 goto out;
2460         }
2461
2462         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2463
2464         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
2465                                       le64_to_cpu(di->i_refcount_loc), &tree);
2466         if (ret) {
2467                 mlog_errno(ret);
2468                 goto out;
2469         }
2470
2471         ret = ocfs2_read_refcount_block(&tree->rf_ci,
2472                                         le64_to_cpu(di->i_refcount_loc),
2473                                         &ref_root_bh);
2474         if (ret) {
2475                 mlog_errno(ret);
2476                 goto out;
2477         }
2478
2479         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2480                                                &tree->rf_ci,
2481                                                ref_root_bh,
2482                                                start_cpos, clusters,
2483                                                &ref_blocks, credits);
2484         if (ret) {
2485                 mlog_errno(ret);
2486                 goto out;
2487         }
2488
2489         mlog(0, "reserve new metadata %d, credits = %d\n",
2490              ref_blocks, *credits);
2491
2492         if (ref_blocks) {
2493                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
2494                                                         ref_blocks, meta_ac);
2495                 if (ret)
2496                         mlog_errno(ret);
2497         }
2498
2499 out:
2500         brelse(ref_root_bh);
2501         return ret;
2502 }
2503
2504 #define MAX_CONTIG_BYTES        1048576
2505
2506 static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb)
2507 {
2508         return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES);
2509 }
2510
2511 static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb)
2512 {
2513         return ~(ocfs2_cow_contig_clusters(sb) - 1);
2514 }
2515
2516 /*
2517  * Given an extent that starts at 'start' and an I/O that starts at 'cpos',
2518  * find an offset (start + (n * contig_clusters)) that is closest to cpos
2519  * while still being less than or equal to it.
2520  *
2521  * The goal is to break the extent at a multiple of contig_clusters.
2522  */
2523 static inline unsigned int ocfs2_cow_align_start(struct super_block *sb,
2524                                                  unsigned int start,
2525                                                  unsigned int cpos)
2526 {
2527         BUG_ON(start > cpos);
2528
2529         return start + ((cpos - start) & ocfs2_cow_contig_mask(sb));
2530 }
2531
2532 /*
2533  * Given a cluster count of len, pad it out so that it is a multiple
2534  * of contig_clusters.
2535  */
2536 static inline unsigned int ocfs2_cow_align_length(struct super_block *sb,
2537                                                   unsigned int len)
2538 {
2539         unsigned int padded =
2540                 (len + (ocfs2_cow_contig_clusters(sb) - 1)) &
2541                 ocfs2_cow_contig_mask(sb);
2542
2543         /* Did we wrap? */
2544         if (padded < len)
2545                 padded = UINT_MAX;
2546
2547         return padded;
2548 }
2549
2550 /*
2551  * Calculate out the start and number of virtual clusters we need to to CoW.
2552  *
2553  * cpos is vitual start cluster position we want to do CoW in a
2554  * file and write_len is the cluster length.
2555  * max_cpos is the place where we want to stop CoW intentionally.
2556  *
2557  * Normal we will start CoW from the beginning of extent record cotaining cpos.
2558  * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we
2559  * get good I/O from the resulting extent tree.
2560  */
2561 static int ocfs2_refcount_cal_cow_clusters(struct inode *inode,
2562                                            struct ocfs2_extent_list *el,
2563                                            u32 cpos,
2564                                            u32 write_len,
2565                                            u32 max_cpos,
2566                                            u32 *cow_start,
2567                                            u32 *cow_len)
2568 {
2569         int ret = 0;
2570         int tree_height = le16_to_cpu(el->l_tree_depth), i;
2571         struct buffer_head *eb_bh = NULL;
2572         struct ocfs2_extent_block *eb = NULL;
2573         struct ocfs2_extent_rec *rec;
2574         unsigned int want_clusters, rec_end = 0;
2575         int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb);
2576         int leaf_clusters;
2577
2578         BUG_ON(cpos + write_len > max_cpos);
2579
2580         if (tree_height > 0) {
2581                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh);
2582                 if (ret) {
2583                         mlog_errno(ret);
2584                         goto out;
2585                 }
2586
2587                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2588                 el = &eb->h_list;
2589
2590                 if (el->l_tree_depth) {
2591                         ocfs2_error(inode->i_sb,
2592                                     "Inode %lu has non zero tree depth in "
2593                                     "leaf block %llu\n", inode->i_ino,
2594                                     (unsigned long long)eb_bh->b_blocknr);
2595                         ret = -EROFS;
2596                         goto out;
2597                 }
2598         }
2599
2600         *cow_len = 0;
2601         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
2602                 rec = &el->l_recs[i];
2603
2604                 if (ocfs2_is_empty_extent(rec)) {
2605                         mlog_bug_on_msg(i != 0, "Inode %lu has empty record in "
2606                                         "index %d\n", inode->i_ino, i);
2607                         continue;
2608                 }
2609
2610                 if (le32_to_cpu(rec->e_cpos) +
2611                     le16_to_cpu(rec->e_leaf_clusters) <= cpos)
2612                         continue;
2613
2614                 if (*cow_len == 0) {
2615                         /*
2616                          * We should find a refcounted record in the
2617                          * first pass.
2618                          */
2619                         BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED));
2620                         *cow_start = le32_to_cpu(rec->e_cpos);
2621                 }
2622
2623                 /*
2624                  * If we encounter a hole, a non-refcounted record or
2625                  * pass the max_cpos, stop the search.
2626                  */
2627                 if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) ||
2628                     (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) ||
2629                     (max_cpos <= le32_to_cpu(rec->e_cpos)))
2630                         break;
2631
2632                 leaf_clusters = le16_to_cpu(rec->e_leaf_clusters);
2633                 rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters;
2634                 if (rec_end > max_cpos) {
2635                         rec_end = max_cpos;
2636                         leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos);
2637                 }
2638
2639                 /*
2640                  * How many clusters do we actually need from
2641                  * this extent?  First we see how many we actually
2642                  * need to complete the write.  If that's smaller
2643                  * than contig_clusters, we try for contig_clusters.
2644                  */
2645                 if (!*cow_len)
2646                         want_clusters = write_len;
2647                 else
2648                         want_clusters = (cpos + write_len) -
2649                                 (*cow_start + *cow_len);
2650                 if (want_clusters < contig_clusters)
2651                         want_clusters = contig_clusters;
2652
2653                 /*
2654                  * If the write does not cover the whole extent, we
2655                  * need to calculate how we're going to split the extent.
2656                  * We try to do it on contig_clusters boundaries.
2657                  *
2658                  * Any extent smaller than contig_clusters will be
2659                  * CoWed in its entirety.
2660                  */
2661                 if (leaf_clusters <= contig_clusters)
2662                         *cow_len += leaf_clusters;
2663                 else if (*cow_len || (*cow_start == cpos)) {
2664                         /*
2665                          * This extent needs to be CoW'd from its
2666                          * beginning, so all we have to do is compute
2667                          * how many clusters to grab.  We align
2668                          * want_clusters to the edge of contig_clusters
2669                          * to get better I/O.
2670                          */
2671                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2672                                                                want_clusters);
2673
2674                         if (leaf_clusters < want_clusters)
2675                                 *cow_len += leaf_clusters;
2676                         else
2677                                 *cow_len += want_clusters;
2678                 } else if ((*cow_start + contig_clusters) >=
2679                            (cpos + write_len)) {
2680                         /*
2681                          * Breaking off contig_clusters at the front
2682                          * of the extent will cover our write.  That's
2683                          * easy.
2684                          */
2685                         *cow_len = contig_clusters;
2686                 } else if ((rec_end - cpos) <= contig_clusters) {
2687                         /*
2688                          * Breaking off contig_clusters at the tail of
2689                          * this extent will cover cpos.
2690                          */
2691                         *cow_start = rec_end - contig_clusters;
2692                         *cow_len = contig_clusters;
2693                 } else if ((rec_end - cpos) <= want_clusters) {
2694                         /*
2695                          * While we can't fit the entire write in this
2696                          * extent, we know that the write goes from cpos
2697                          * to the end of the extent.  Break that off.
2698                          * We try to break it at some multiple of
2699                          * contig_clusters from the front of the extent.
2700                          * Failing that (ie, cpos is within
2701                          * contig_clusters of the front), we'll CoW the
2702                          * entire extent.
2703                          */
2704                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2705                                                            *cow_start, cpos);
2706                         *cow_len = rec_end - *cow_start;
2707                 } else {
2708                         /*
2709                          * Ok, the entire write lives in the middle of
2710                          * this extent.  Let's try to slice the extent up
2711                          * nicely.  Optimally, our CoW region starts at
2712                          * m*contig_clusters from the beginning of the
2713                          * extent and goes for n*contig_clusters,
2714                          * covering the entire write.
2715                          */
2716                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2717                                                            *cow_start, cpos);
2718
2719                         want_clusters = (cpos + write_len) - *cow_start;
2720                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2721                                                                want_clusters);
2722                         if (*cow_start + want_clusters <= rec_end)
2723                                 *cow_len = want_clusters;
2724                         else
2725                                 *cow_len = rec_end - *cow_start;
2726                 }
2727
2728                 /* Have we covered our entire write yet? */
2729                 if ((*cow_start + *cow_len) >= (cpos + write_len))
2730                         break;
2731
2732                 /*
2733                  * If we reach the end of the extent block and don't get enough
2734                  * clusters, continue with the next extent block if possible.
2735                  */
2736                 if (i + 1 == le16_to_cpu(el->l_next_free_rec) &&
2737                     eb && eb->h_next_leaf_blk) {
2738                         brelse(eb_bh);
2739                         eb_bh = NULL;
2740
2741                         ret = ocfs2_read_extent_block(INODE_CACHE(inode),
2742                                                le64_to_cpu(eb->h_next_leaf_blk),
2743                                                &eb_bh);
2744                         if (ret) {
2745                                 mlog_errno(ret);
2746                                 goto out;
2747                         }
2748
2749                         eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2750                         el = &eb->h_list;
2751                         i = -1;
2752                 }
2753         }
2754
2755 out:
2756         brelse(eb_bh);
2757         return ret;
2758 }
2759
2760 /*
2761  * Prepare meta_ac, data_ac and calculate credits when we want to add some
2762  * num_clusters in data_tree "et" and change the refcount for the old
2763  * clusters(starting form p_cluster) in the refcount tree.
2764  *
2765  * Note:
2766  * 1. since we may split the old tree, so we at most will need num_clusters + 2
2767  *    more new leaf records.
2768  * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so
2769  *    just give data_ac = NULL.
2770  */
2771 static int ocfs2_lock_refcount_allocators(struct super_block *sb,
2772                                         u32 p_cluster, u32 num_clusters,
2773                                         struct ocfs2_extent_tree *et,
2774                                         struct ocfs2_caching_info *ref_ci,
2775                                         struct buffer_head *ref_root_bh,
2776                                         struct ocfs2_alloc_context **meta_ac,
2777                                         struct ocfs2_alloc_context **data_ac,
2778                                         int *credits)
2779 {
2780         int ret = 0, meta_add = 0;
2781         int num_free_extents = ocfs2_num_free_extents(OCFS2_SB(sb), et);
2782
2783         if (num_free_extents < 0) {
2784                 ret = num_free_extents;
2785                 mlog_errno(ret);
2786                 goto out;
2787         }
2788
2789         if (num_free_extents < num_clusters + 2)
2790                 meta_add =
2791                         ocfs2_extend_meta_needed(et->et_root_el);
2792
2793         *credits += ocfs2_calc_extend_credits(sb, et->et_root_el,
2794                                               num_clusters + 2);
2795
2796         ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh,
2797                                                p_cluster, num_clusters,
2798                                                &meta_add, credits);
2799         if (ret) {
2800                 mlog_errno(ret);
2801                 goto out;
2802         }
2803
2804         mlog(0, "reserve new metadata %d, clusters %u, credits = %d\n",
2805              meta_add, num_clusters, *credits);
2806         ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add,
2807                                                 meta_ac);
2808         if (ret) {
2809                 mlog_errno(ret);
2810                 goto out;
2811         }
2812
2813         if (data_ac) {
2814                 ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters,
2815                                              data_ac);
2816                 if (ret)
2817                         mlog_errno(ret);
2818         }
2819
2820 out:
2821         if (ret) {
2822                 if (*meta_ac) {
2823                         ocfs2_free_alloc_context(*meta_ac);
2824                         *meta_ac = NULL;
2825                 }
2826         }
2827
2828         return ret;
2829 }
2830
2831 static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh)
2832 {
2833         BUG_ON(buffer_dirty(bh));
2834
2835         clear_buffer_mapped(bh);
2836
2837         return 0;
2838 }
2839
2840 static int ocfs2_duplicate_clusters_by_page(handle_t *handle,
2841                                             struct ocfs2_cow_context *context,
2842                                             u32 cpos, u32 old_cluster,
2843                                             u32 new_cluster, u32 new_len)
2844 {
2845         int ret = 0, partial;
2846         struct ocfs2_caching_info *ci = context->data_et.et_ci;
2847         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2848         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2849         struct page *page;
2850         pgoff_t page_index;
2851         unsigned int from, to;
2852         loff_t offset, end, map_end;
2853         struct address_space *mapping = context->inode->i_mapping;
2854
2855         mlog(0, "old_cluster %u, new %u, len %u at offset %u\n", old_cluster,
2856              new_cluster, new_len, cpos);
2857
2858         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
2859         end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2860
2861         while (offset < end) {
2862                 page_index = offset >> PAGE_CACHE_SHIFT;
2863                 map_end = (page_index + 1) << PAGE_CACHE_SHIFT;
2864                 if (map_end > end)
2865                         map_end = end;
2866
2867                 /* from, to is the offset within the page. */
2868                 from = offset & (PAGE_CACHE_SIZE - 1);
2869                 to = PAGE_CACHE_SIZE;
2870                 if (map_end & (PAGE_CACHE_SIZE - 1))
2871                         to = map_end & (PAGE_CACHE_SIZE - 1);
2872
2873                 page = grab_cache_page(mapping, page_index);
2874
2875                 /* This page can't be dirtied before we CoW it out. */
2876                 BUG_ON(PageDirty(page));
2877
2878                 if (!PageUptodate(page)) {
2879                         ret = block_read_full_page(page, ocfs2_get_block);
2880                         if (ret) {
2881                                 mlog_errno(ret);
2882                                 goto unlock;
2883                         }
2884                         lock_page(page);
2885                 }
2886
2887                 if (page_has_buffers(page)) {
2888                         ret = walk_page_buffers(handle, page_buffers(page),
2889                                                 from, to, &partial,
2890                                                 ocfs2_clear_cow_buffer);
2891                         if (ret) {
2892                                 mlog_errno(ret);
2893                                 goto unlock;
2894                         }
2895                 }
2896
2897                 ocfs2_map_and_dirty_page(context->inode,
2898                                          handle, from, to,
2899                                          page, 0, &new_block);
2900                 mark_page_accessed(page);
2901 unlock:
2902                 unlock_page(page);
2903                 page_cache_release(page);
2904                 page = NULL;
2905                 offset = map_end;
2906                 if (ret)
2907                         break;
2908         }
2909
2910         return ret;
2911 }
2912
2913 static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
2914                                            struct ocfs2_cow_context *context,
2915                                            u32 cpos, u32 old_cluster,
2916                                            u32 new_cluster, u32 new_len)
2917 {
2918         int ret = 0;
2919         struct super_block *sb = context->inode->i_sb;
2920         struct ocfs2_caching_info *ci = context->data_et.et_ci;
2921         int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
2922         u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
2923         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2924         struct ocfs2_super *osb = OCFS2_SB(sb);
2925         struct buffer_head *old_bh = NULL;
2926         struct buffer_head *new_bh = NULL;
2927
2928         mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster,
2929              new_cluster, new_len);
2930
2931         for (i = 0; i < blocks; i++, old_block++, new_block++) {
2932                 new_bh = sb_getblk(osb->sb, new_block);
2933                 if (new_bh == NULL) {
2934                         ret = -EIO;
2935                         mlog_errno(ret);
2936                         break;
2937                 }
2938
2939                 ocfs2_set_new_buffer_uptodate(ci, new_bh);
2940
2941                 ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
2942                 if (ret) {
2943                         mlog_errno(ret);
2944                         break;
2945                 }
2946
2947                 ret = ocfs2_journal_access(handle, ci, new_bh,
2948                                            OCFS2_JOURNAL_ACCESS_CREATE);
2949                 if (ret) {
2950                         mlog_errno(ret);
2951                         break;
2952                 }
2953
2954                 memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
2955                 ret = ocfs2_journal_dirty(handle, new_bh);
2956                 if (ret) {
2957                         mlog_errno(ret);
2958                         break;
2959                 }
2960
2961                 brelse(new_bh);
2962                 brelse(old_bh);
2963                 new_bh = NULL;
2964                 old_bh = NULL;
2965         }
2966
2967         brelse(new_bh);
2968         brelse(old_bh);
2969         return ret;
2970 }
2971
2972 static int ocfs2_clear_ext_refcount(handle_t *handle,
2973                                     struct ocfs2_extent_tree *et,
2974                                     u32 cpos, u32 p_cluster, u32 len,
2975                                     unsigned int ext_flags,
2976                                     struct ocfs2_alloc_context *meta_ac,
2977                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
2978 {
2979         int ret, index;
2980         struct ocfs2_extent_rec replace_rec;
2981         struct ocfs2_path *path = NULL;
2982         struct ocfs2_extent_list *el;
2983         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2984         u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
2985
2986         mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n",
2987              (unsigned long long)ino, cpos, len, p_cluster, ext_flags);
2988
2989         memset(&replace_rec, 0, sizeof(replace_rec));
2990         replace_rec.e_cpos = cpu_to_le32(cpos);
2991         replace_rec.e_leaf_clusters = cpu_to_le16(len);
2992         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
2993                                                                    p_cluster));
2994         replace_rec.e_flags = ext_flags;
2995         replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
2996
2997         path = ocfs2_new_path_from_et(et);
2998         if (!path) {
2999                 ret = -ENOMEM;
3000                 mlog_errno(ret);
3001                 goto out;
3002         }
3003
3004         ret = ocfs2_find_path(et->et_ci, path, cpos);
3005         if (ret) {
3006                 mlog_errno(ret);
3007                 goto out;
3008         }
3009
3010         el = path_leaf_el(path);
3011
3012         index = ocfs2_search_extent_list(el, cpos);
3013         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
3014                 ocfs2_error(sb,
3015                             "Inode %llu has an extent at cpos %u which can no "
3016                             "longer be found.\n",
3017                             (unsigned long long)ino, cpos);
3018                 ret = -EROFS;
3019                 goto out;
3020         }
3021
3022         ret = ocfs2_split_extent(handle, et, path, index,
3023                                  &replace_rec, meta_ac, dealloc);
3024         if (ret)
3025                 mlog_errno(ret);
3026
3027 out:
3028         ocfs2_free_path(path);
3029         return ret;
3030 }
3031
3032 static int ocfs2_replace_clusters(handle_t *handle,
3033                                   struct ocfs2_cow_context *context,
3034                                   u32 cpos, u32 old,
3035                                   u32 new, u32 len,
3036                                   unsigned int ext_flags)
3037 {
3038         int ret;
3039         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3040         u64 ino = ocfs2_metadata_cache_owner(ci);
3041
3042         mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n",
3043              (unsigned long long)ino, cpos, old, new, len, ext_flags);
3044
3045         /*If the old clusters is unwritten, no need to duplicate. */
3046         if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3047                 ret = context->cow_duplicate_clusters(handle, context, cpos,
3048                                                       old, new, len);
3049                 if (ret) {
3050                         mlog_errno(ret);
3051                         goto out;
3052                 }
3053         }
3054
3055         ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
3056                                        cpos, new, len, ext_flags,
3057                                        context->meta_ac, &context->dealloc);
3058         if (ret)
3059                 mlog_errno(ret);
3060 out:
3061         return ret;
3062 }
3063
3064 static int ocfs2_cow_sync_writeback(struct super_block *sb,
3065                                     struct ocfs2_cow_context *context,
3066                                     u32 cpos, u32 num_clusters)
3067 {
3068         int ret = 0;
3069         loff_t offset, end, map_end;
3070         pgoff_t page_index;
3071         struct page *page;
3072
3073         if (ocfs2_should_order_data(context->inode))
3074                 return 0;
3075
3076         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
3077         end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits);
3078
3079         ret = filemap_fdatawrite_range(context->inode->i_mapping,
3080                                        offset, end - 1);
3081         if (ret < 0) {
3082                 mlog_errno(ret);
3083                 return ret;
3084         }
3085
3086         while (offset < end) {
3087                 page_index = offset >> PAGE_CACHE_SHIFT;
3088                 map_end = (page_index + 1) << PAGE_CACHE_SHIFT;
3089                 if (map_end > end)
3090                         map_end = end;
3091
3092                 page = grab_cache_page(context->inode->i_mapping, page_index);
3093                 BUG_ON(!page);
3094
3095                 wait_on_page_writeback(page);
3096                 if (PageError(page)) {
3097                         ret = -EIO;
3098                         mlog_errno(ret);
3099                 } else
3100                         mark_page_accessed(page);
3101
3102                 unlock_page(page);
3103                 page_cache_release(page);
3104                 page = NULL;
3105                 offset = map_end;
3106                 if (ret)
3107                         break;
3108         }
3109
3110         return ret;
3111 }
3112
3113 static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3114                                  u32 v_cluster, u32 *p_cluster,
3115                                  u32 *num_clusters,
3116                                  unsigned int *extent_flags)
3117 {
3118         return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3119                                   num_clusters, extent_flags);
3120 }
3121
3122 static int ocfs2_make_clusters_writable(struct super_block *sb,
3123                                         struct ocfs2_cow_context *context,
3124                                         u32 cpos, u32 p_cluster,
3125                                         u32 num_clusters, unsigned int e_flags)
3126 {
3127         int ret, delete, index, credits =  0;
3128         u32 new_bit, new_len;
3129         unsigned int set_len;
3130         struct ocfs2_super *osb = OCFS2_SB(sb);
3131         handle_t *handle;
3132         struct buffer_head *ref_leaf_bh = NULL;
3133         struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
3134         struct ocfs2_refcount_rec rec;
3135
3136         mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n",
3137              cpos, p_cluster, num_clusters, e_flags);
3138
3139         ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3140                                              &context->data_et,
3141                                              ref_ci,
3142                                              context->ref_root_bh,
3143                                              &context->meta_ac,
3144                                              &context->data_ac, &credits);
3145         if (ret) {
3146                 mlog_errno(ret);
3147                 return ret;
3148         }
3149
3150         if (context->post_refcount)
3151                 credits += context->post_refcount->credits;
3152
3153         credits += context->extra_credits;
3154         handle = ocfs2_start_trans(osb, credits);
3155         if (IS_ERR(handle)) {
3156                 ret = PTR_ERR(handle);
3157                 mlog_errno(ret);
3158                 goto out;
3159         }
3160
3161         while (num_clusters) {
3162                 ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
3163                                              p_cluster, num_clusters,
3164                                              &rec, &index, &ref_leaf_bh);
3165                 if (ret) {
3166                         mlog_errno(ret);
3167                         goto out_commit;
3168                 }
3169
3170                 BUG_ON(!rec.r_refcount);
3171                 set_len = min((u64)p_cluster + num_clusters,
3172                               le64_to_cpu(rec.r_cpos) +
3173                               le32_to_cpu(rec.r_clusters)) - p_cluster;
3174
3175                 /*
3176                  * There are many different situation here.
3177                  * 1. If refcount == 1, remove the flag and don't COW.
3178                  * 2. If refcount > 1, allocate clusters.
3179                  *    Here we may not allocate r_len once at a time, so continue
3180                  *    until we reach num_clusters.
3181                  */
3182                 if (le32_to_cpu(rec.r_refcount) == 1) {
3183                         delete = 0;
3184                         ret = ocfs2_clear_ext_refcount(handle,
3185                                                        &context->data_et,
3186                                                        cpos, p_cluster,
3187                                                        set_len, e_flags,
3188                                                        context->meta_ac,
3189                                                        &context->dealloc);
3190                         if (ret) {
3191                                 mlog_errno(ret);
3192                                 goto out_commit;
3193                         }
3194                 } else {
3195                         delete = 1;
3196
3197                         ret = __ocfs2_claim_clusters(osb, handle,
3198                                                      context->data_ac,
3199                                                      1, set_len,
3200                                                      &new_bit, &new_len);
3201                         if (ret) {
3202                                 mlog_errno(ret);
3203                                 goto out_commit;
3204                         }
3205
3206                         ret = ocfs2_replace_clusters(handle, context,
3207                                                      cpos, p_cluster, new_bit,
3208                                                      new_len, e_flags);
3209                         if (ret) {
3210                                 mlog_errno(ret);
3211                                 goto out_commit;
3212                         }
3213                         set_len = new_len;
3214                 }
3215
3216                 ret = __ocfs2_decrease_refcount(handle, ref_ci,
3217                                                 context->ref_root_bh,
3218                                                 p_cluster, set_len,
3219                                                 context->meta_ac,
3220                                                 &context->dealloc, delete);
3221                 if (ret) {
3222                         mlog_errno(ret);
3223                         goto out_commit;
3224                 }
3225
3226                 cpos += set_len;
3227                 p_cluster += set_len;
3228                 num_clusters -= set_len;
3229                 brelse(ref_leaf_bh);
3230                 ref_leaf_bh = NULL;
3231         }
3232
3233         /* handle any post_cow action. */
3234         if (context->post_refcount && context->post_refcount->func) {
3235                 ret = context->post_refcount->func(context->inode, handle,
3236                                                 context->post_refcount->para);
3237                 if (ret) {
3238                         mlog_errno(ret);
3239                         goto out_commit;
3240                 }
3241         }
3242
3243         /*
3244          * Here we should write the new page out first if we are
3245          * in write-back mode.
3246          */
3247         if (context->get_clusters == ocfs2_di_get_clusters) {
3248                 ret = ocfs2_cow_sync_writeback(sb, context, cpos, num_clusters);
3249                 if (ret)
3250                         mlog_errno(ret);
3251         }
3252
3253 out_commit:
3254         ocfs2_commit_trans(osb, handle);
3255
3256 out:
3257         if (context->data_ac) {
3258                 ocfs2_free_alloc_context(context->data_ac);
3259                 context->data_ac = NULL;
3260         }
3261         if (context->meta_ac) {
3262                 ocfs2_free_alloc_context(context->meta_ac);
3263                 context->meta_ac = NULL;
3264         }
3265         brelse(ref_leaf_bh);
3266
3267         return ret;
3268 }
3269
3270 static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
3271 {
3272         int ret = 0;
3273         struct inode *inode = context->inode;
3274         u32 cow_start = context->cow_start, cow_len = context->cow_len;
3275         u32 p_cluster, num_clusters;
3276         unsigned int ext_flags;
3277         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3278
3279         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
3280                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
3281                             "tree, but the feature bit is not set in the "
3282                             "super block.", inode->i_ino);
3283                 return -EROFS;
3284         }
3285
3286         ocfs2_init_dealloc_ctxt(&context->dealloc);
3287
3288         while (cow_len) {
3289                 ret = context->get_clusters(context, cow_start, &p_cluster,
3290                                             &num_clusters, &ext_flags);
3291                 if (ret) {
3292                         mlog_errno(ret);
3293                         break;
3294                 }
3295
3296                 BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
3297
3298                 if (cow_len < num_clusters)
3299                         num_clusters = cow_len;
3300
3301                 ret = ocfs2_make_clusters_writable(inode->i_sb, context,
3302                                                    cow_start, p_cluster,
3303                                                    num_clusters, ext_flags);
3304                 if (ret) {
3305                         mlog_errno(ret);
3306                         break;
3307                 }
3308
3309                 cow_len -= num_clusters;
3310                 cow_start += num_clusters;
3311         }
3312
3313         if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
3314                 ocfs2_schedule_truncate_log_flush(osb, 1);
3315                 ocfs2_run_deallocs(osb, &context->dealloc);
3316         }
3317
3318         return ret;
3319 }
3320
3321 /*
3322  * Starting at cpos, try to CoW write_len clusters.  Don't CoW
3323  * past max_cpos.  This will stop when it runs into a hole or an
3324  * unrefcounted extent.
3325  */
3326 static int ocfs2_refcount_cow_hunk(struct inode *inode,
3327                                    struct buffer_head *di_bh,
3328                                    u32 cpos, u32 write_len, u32 max_cpos)
3329 {
3330         int ret;
3331         u32 cow_start = 0, cow_len = 0;
3332         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3333         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3334         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3335         struct buffer_head *ref_root_bh = NULL;
3336         struct ocfs2_refcount_tree *ref_tree;
3337         struct ocfs2_cow_context *context = NULL;
3338
3339         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3340
3341         ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
3342                                               cpos, write_len, max_cpos,
3343                                               &cow_start, &cow_len);
3344         if (ret) {
3345                 mlog_errno(ret);
3346                 goto out;
3347         }
3348
3349         mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, "
3350              "cow_len %u\n", inode->i_ino,
3351              cpos, write_len, cow_start, cow_len);
3352
3353         BUG_ON(cow_len == 0);
3354
3355         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3356         if (!context) {
3357                 ret = -ENOMEM;
3358                 mlog_errno(ret);
3359                 goto out;
3360         }
3361
3362         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3363                                        1, &ref_tree, &ref_root_bh);
3364         if (ret) {
3365                 mlog_errno(ret);
3366                 goto out;
3367         }
3368
3369         context->inode = inode;
3370         context->cow_start = cow_start;
3371         context->cow_len = cow_len;
3372         context->ref_tree = ref_tree;
3373         context->ref_root_bh = ref_root_bh;
3374         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3375         context->get_clusters = ocfs2_di_get_clusters;
3376
3377         ocfs2_init_dinode_extent_tree(&context->data_et,
3378                                       INODE_CACHE(inode), di_bh);
3379
3380         ret = ocfs2_replace_cow(context);
3381         if (ret)
3382                 mlog_errno(ret);
3383
3384         /*
3385          * truncate the extent map here since no matter whether we meet with
3386          * any error during the action, we shouldn't trust cached extent map
3387          * any more.
3388          */
3389         ocfs2_extent_map_trunc(inode, cow_start);
3390
3391         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3392         brelse(ref_root_bh);
3393 out:
3394         kfree(context);
3395         return ret;
3396 }
3397
3398 /*
3399  * CoW any and all clusters between cpos and cpos+write_len.
3400  * Don't CoW past max_cpos.  If this returns successfully, all
3401  * clusters between cpos and cpos+write_len are safe to modify.
3402  */
3403 int ocfs2_refcount_cow(struct inode *inode,
3404                        struct buffer_head *di_bh,
3405                        u32 cpos, u32 write_len, u32 max_cpos)
3406 {
3407         int ret = 0;
3408         u32 p_cluster, num_clusters;
3409         unsigned int ext_flags;
3410
3411         while (write_len) {
3412                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3413                                          &num_clusters, &ext_flags);
3414                 if (ret) {
3415                         mlog_errno(ret);
3416                         break;
3417                 }
3418
3419                 if (write_len < num_clusters)
3420                         num_clusters = write_len;
3421
3422                 if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3423                         ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
3424                                                       num_clusters, max_cpos);
3425                         if (ret) {
3426                                 mlog_errno(ret);
3427                                 break;
3428                         }
3429                 }
3430
3431                 write_len -= num_clusters;
3432                 cpos += num_clusters;
3433         }
3434
3435         return ret;
3436 }
3437
3438 static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3439                                           u32 v_cluster, u32 *p_cluster,
3440                                           u32 *num_clusters,
3441                                           unsigned int *extent_flags)
3442 {
3443         struct inode *inode = context->inode;
3444         struct ocfs2_xattr_value_root *xv = context->cow_object;
3445
3446         return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3447                                         num_clusters, &xv->xr_list,
3448                                         extent_flags);
3449 }
3450
3451 /*
3452  * Given a xattr value root, calculate the most meta/credits we need for
3453  * refcount tree change if we truncate it to 0.
3454  */
3455 int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3456                                        struct ocfs2_caching_info *ref_ci,
3457                                        struct buffer_head *ref_root_bh,
3458                                        struct ocfs2_xattr_value_root *xv,
3459                                        int *meta_add, int *credits)
3460 {
3461         int ret = 0, index, ref_blocks = 0;
3462         u32 p_cluster, num_clusters;
3463         u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3464         struct ocfs2_refcount_block *rb;
3465         struct ocfs2_refcount_rec rec;
3466         struct buffer_head *ref_leaf_bh = NULL;
3467
3468         while (cpos < clusters) {
3469                 ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3470                                                &num_clusters, &xv->xr_list,
3471                                                NULL);
3472                 if (ret) {
3473                         mlog_errno(ret);
3474                         goto out;
3475                 }
3476
3477                 cpos += num_clusters;
3478
3479                 while (num_clusters) {
3480                         ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3481                                                      p_cluster, num_clusters,
3482                                                      &rec, &index,
3483                                                      &ref_leaf_bh);
3484                         if (ret) {
3485                                 mlog_errno(ret);
3486                                 goto out;
3487                         }
3488
3489                         BUG_ON(!rec.r_refcount);
3490
3491                         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3492
3493                         /*
3494                          * We really don't know whether the other clusters is in
3495                          * this refcount block or not, so just take the worst
3496                          * case that all the clusters are in this block and each
3497                          * one will split a refcount rec, so totally we need
3498                          * clusters * 2 new refcount rec.
3499                          */
3500                         if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3501                             le16_to_cpu(rb->rf_records.rl_count))
3502                                 ref_blocks++;
3503
3504                         *credits += 1;
3505                         brelse(ref_leaf_bh);
3506                         ref_leaf_bh = NULL;
3507
3508                         if (num_clusters <= le32_to_cpu(rec.r_clusters))
3509                                 break;
3510                         else
3511                                 num_clusters -= le32_to_cpu(rec.r_clusters);
3512                         p_cluster += num_clusters;
3513                 }
3514         }
3515
3516         *meta_add += ref_blocks;
3517         if (!ref_blocks)
3518                 goto out;
3519
3520         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3521         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3522                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3523         else {
3524                 struct ocfs2_extent_tree et;
3525
3526                 ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3527                 *credits += ocfs2_calc_extend_credits(inode->i_sb,
3528                                                       et.et_root_el,
3529                                                       ref_blocks);
3530         }
3531
3532 out:
3533         brelse(ref_leaf_bh);
3534         return ret;
3535 }
3536
3537 /*
3538  * Do CoW for xattr.
3539  */
3540 int ocfs2_refcount_cow_xattr(struct inode *inode,
3541                              struct ocfs2_dinode *di,
3542                              struct ocfs2_xattr_value_buf *vb,
3543                              struct ocfs2_refcount_tree *ref_tree,
3544                              struct buffer_head *ref_root_bh,
3545                              u32 cpos, u32 write_len,
3546                              struct ocfs2_post_refcount *post)
3547 {
3548         int ret;
3549         struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3550         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3551         struct ocfs2_cow_context *context = NULL;
3552         u32 cow_start, cow_len;
3553
3554         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3555
3556         ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3557                                               cpos, write_len, UINT_MAX,
3558                                               &cow_start, &cow_len);
3559         if (ret) {
3560                 mlog_errno(ret);
3561                 goto out;
3562         }
3563
3564         BUG_ON(cow_len == 0);
3565
3566         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3567         if (!context) {
3568                 ret = -ENOMEM;
3569                 mlog_errno(ret);
3570                 goto out;
3571         }
3572
3573         context->inode = inode;
3574         context->cow_start = cow_start;
3575         context->cow_len = cow_len;
3576         context->ref_tree = ref_tree;
3577         context->ref_root_bh = ref_root_bh;;
3578         context->cow_object = xv;
3579
3580         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3581         /* We need the extra credits for duplicate_clusters by jbd. */
3582         context->extra_credits =
3583                 ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3584         context->get_clusters = ocfs2_xattr_value_get_clusters;
3585         context->post_refcount = post;
3586
3587         ocfs2_init_xattr_value_extent_tree(&context->data_et,
3588                                            INODE_CACHE(inode), vb);
3589
3590         ret = ocfs2_replace_cow(context);
3591         if (ret)
3592                 mlog_errno(ret);
3593
3594 out:
3595         kfree(context);
3596         return ret;
3597 }
3598
3599 /*
3600  * Insert a new extent into refcount tree and mark a extent rec
3601  * as refcounted in the dinode tree.
3602  */
3603 int ocfs2_add_refcount_flag(struct inode *inode,
3604                             struct ocfs2_extent_tree *data_et,
3605                             struct ocfs2_caching_info *ref_ci,
3606                             struct buffer_head *ref_root_bh,
3607                             u32 cpos, u32 p_cluster, u32 num_clusters,
3608                             struct ocfs2_cached_dealloc_ctxt *dealloc,
3609                             struct ocfs2_post_refcount *post)
3610 {
3611         int ret;
3612         handle_t *handle;
3613         int credits = 1, ref_blocks = 0;
3614         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3615         struct ocfs2_alloc_context *meta_ac = NULL;
3616
3617         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3618                                                ref_ci, ref_root_bh,
3619                                                p_cluster, num_clusters,
3620                                                &ref_blocks, &credits);
3621         if (ret) {
3622                 mlog_errno(ret);
3623                 goto out;
3624         }
3625
3626         mlog(0, "reserve new metadata %d, credits = %d\n",
3627              ref_blocks, credits);
3628
3629         if (ref_blocks) {
3630                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
3631                                                         ref_blocks, &meta_ac);
3632                 if (ret) {
3633                         mlog_errno(ret);
3634                         goto out;
3635                 }
3636         }
3637
3638         if (post)
3639                 credits += post->credits;
3640
3641         handle = ocfs2_start_trans(osb, credits);
3642         if (IS_ERR(handle)) {
3643                 ret = PTR_ERR(handle);
3644                 mlog_errno(ret);
3645                 goto out;
3646         }
3647
3648         ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3649                                            cpos, num_clusters, p_cluster,
3650                                            meta_ac, dealloc);
3651         if (ret) {
3652                 mlog_errno(ret);
3653                 goto out_commit;
3654         }
3655
3656         ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3657                                         p_cluster, num_clusters, 0,
3658                                         meta_ac, dealloc);
3659         if (ret) {
3660                 mlog_errno(ret);
3661                 goto out_commit;
3662         }
3663
3664         if (post && post->func) {
3665                 ret = post->func(inode, handle, post->para);
3666                 if (ret)
3667                         mlog_errno(ret);
3668         }
3669
3670 out_commit:
3671         ocfs2_commit_trans(osb, handle);
3672 out:
3673         if (meta_ac)
3674                 ocfs2_free_alloc_context(meta_ac);
3675         return ret;
3676 }
3677
3678 static int ocfs2_change_ctime(struct inode *inode,
3679                               struct buffer_head *di_bh)
3680 {
3681         int ret;
3682         handle_t *handle;
3683         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3684
3685         handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3686                                    OCFS2_INODE_UPDATE_CREDITS);
3687         if (IS_ERR(handle)) {
3688                 ret = PTR_ERR(handle);
3689                 mlog_errno(ret);
3690                 goto out;
3691         }
3692
3693         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3694                                       OCFS2_JOURNAL_ACCESS_WRITE);
3695         if (ret) {
3696                 mlog_errno(ret);
3697                 goto out_commit;
3698         }
3699
3700         inode->i_ctime = CURRENT_TIME;
3701         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
3702         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
3703
3704         ocfs2_journal_dirty(handle, di_bh);
3705
3706 out_commit:
3707         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3708 out:
3709         return ret;
3710 }
3711
3712 static int ocfs2_attach_refcount_tree(struct inode *inode,
3713                                       struct buffer_head *di_bh)
3714 {
3715         int ret, data_changed = 0;
3716         struct buffer_head *ref_root_bh = NULL;
3717         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3718         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3719         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3720         struct ocfs2_refcount_tree *ref_tree;
3721         unsigned int ext_flags;
3722         loff_t size;
3723         u32 cpos, num_clusters, clusters, p_cluster;
3724         struct ocfs2_cached_dealloc_ctxt dealloc;
3725         struct ocfs2_extent_tree di_et;
3726
3727         ocfs2_init_dealloc_ctxt(&dealloc);
3728
3729         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) {
3730                 ret = ocfs2_create_refcount_tree(inode, di_bh);
3731                 if (ret) {
3732                         mlog_errno(ret);
3733                         goto out;
3734                 }
3735         }
3736
3737         BUG_ON(!di->i_refcount_loc);
3738         ret = ocfs2_lock_refcount_tree(osb,
3739                                        le64_to_cpu(di->i_refcount_loc), 1,
3740                                        &ref_tree, &ref_root_bh);
3741         if (ret) {
3742                 mlog_errno(ret);
3743                 goto out;
3744         }
3745
3746         ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3747
3748         size = i_size_read(inode);
3749         clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3750
3751         cpos = 0;
3752         while (cpos < clusters) {
3753                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3754                                          &num_clusters, &ext_flags);
3755
3756                 if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3757                         ret = ocfs2_add_refcount_flag(inode, &di_et,
3758                                                       &ref_tree->rf_ci,
3759                                                       ref_root_bh, cpos,
3760                                                       p_cluster, num_clusters,
3761                                                       &dealloc, NULL);
3762                         if (ret) {
3763                                 mlog_errno(ret);
3764                                 goto unlock;
3765                         }
3766
3767                         data_changed = 1;
3768                 }
3769                 cpos += num_clusters;
3770         }
3771
3772         if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
3773                 ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
3774                                                        &ref_tree->rf_ci,
3775                                                        ref_root_bh,
3776                                                        &dealloc);
3777                 if (ret) {
3778                         mlog_errno(ret);
3779                         goto unlock;
3780                 }
3781         }
3782
3783         if (data_changed) {
3784                 ret = ocfs2_change_ctime(inode, di_bh);
3785                 if (ret)
3786                         mlog_errno(ret);
3787         }
3788
3789 unlock:
3790         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3791         brelse(ref_root_bh);
3792
3793         if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3794                 ocfs2_schedule_truncate_log_flush(osb, 1);
3795                 ocfs2_run_deallocs(osb, &dealloc);
3796         }
3797 out:
3798         /*
3799          * Empty the extent map so that we may get the right extent
3800          * record from the disk.
3801          */
3802         ocfs2_extent_map_trunc(inode, 0);
3803
3804         return ret;
3805 }
3806
3807 static int ocfs2_add_refcounted_extent(struct inode *inode,
3808                                    struct ocfs2_extent_tree *et,
3809                                    struct ocfs2_caching_info *ref_ci,
3810                                    struct buffer_head *ref_root_bh,
3811                                    u32 cpos, u32 p_cluster, u32 num_clusters,
3812                                    unsigned int ext_flags,
3813                                    struct ocfs2_cached_dealloc_ctxt *dealloc)
3814 {
3815         int ret;
3816         handle_t *handle;
3817         int credits = 0;
3818         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3819         struct ocfs2_alloc_context *meta_ac = NULL;
3820
3821         ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3822                                              p_cluster, num_clusters,
3823                                              et, ref_ci,
3824                                              ref_root_bh, &meta_ac,
3825                                              NULL, &credits);
3826         if (ret) {
3827                 mlog_errno(ret);
3828                 goto out;
3829         }
3830
3831         handle = ocfs2_start_trans(osb, credits);
3832         if (IS_ERR(handle)) {
3833                 ret = PTR_ERR(handle);
3834                 mlog_errno(ret);
3835                 goto out;
3836         }
3837
3838         ret = ocfs2_insert_extent(handle, et, cpos,
3839                         cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
3840                                                              p_cluster)),
3841                         num_clusters, ext_flags, meta_ac);
3842         if (ret) {
3843                 mlog_errno(ret);
3844                 goto out_commit;
3845         }
3846
3847         ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3848                                       p_cluster, num_clusters,
3849                                       meta_ac, dealloc);
3850         if (ret)
3851                 mlog_errno(ret);
3852
3853 out_commit:
3854         ocfs2_commit_trans(osb, handle);
3855 out:
3856         if (meta_ac)
3857                 ocfs2_free_alloc_context(meta_ac);
3858         return ret;
3859 }
3860
3861 static int ocfs2_duplicate_extent_list(struct inode *s_inode,
3862                                 struct inode *t_inode,
3863                                 struct buffer_head *t_bh,
3864                                 struct ocfs2_caching_info *ref_ci,
3865                                 struct buffer_head *ref_root_bh,
3866                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
3867 {
3868         int ret = 0;
3869         u32 p_cluster, num_clusters, clusters, cpos;
3870         loff_t size;
3871         unsigned int ext_flags;
3872         struct ocfs2_extent_tree et;
3873
3874         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
3875
3876         size = i_size_read(s_inode);
3877         clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
3878
3879         cpos = 0;
3880         while (cpos < clusters) {
3881                 ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
3882                                          &num_clusters, &ext_flags);
3883
3884                 if (p_cluster) {
3885                         ret = ocfs2_add_refcounted_extent(t_inode, &et,
3886                                                           ref_ci, ref_root_bh,
3887                                                           cpos, p_cluster,
3888                                                           num_clusters,
3889                                                           ext_flags,
3890                                                           dealloc);
3891                         if (ret) {
3892                                 mlog_errno(ret);
3893                                 goto out;
3894                         }
3895                 }
3896
3897                 cpos += num_clusters;
3898         }
3899
3900 out:
3901         return ret;
3902 }
3903
3904 /*
3905  * change the new file's attributes to the src.
3906  *
3907  * reflink creates a snapshot of a file, that means the attributes
3908  * must be identical except for three exceptions - nlink, ino, and ctime.
3909  */
3910 static int ocfs2_complete_reflink(struct inode *s_inode,
3911                                   struct buffer_head *s_bh,
3912                                   struct inode *t_inode,
3913                                   struct buffer_head *t_bh,
3914                                   bool preserve)
3915 {
3916         int ret;
3917         handle_t *handle;
3918         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
3919         struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
3920         loff_t size = i_size_read(s_inode);
3921
3922         handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
3923                                    OCFS2_INODE_UPDATE_CREDITS);
3924         if (IS_ERR(handle)) {
3925                 ret = PTR_ERR(handle);
3926                 mlog_errno(ret);
3927                 return ret;
3928         }
3929
3930         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
3931                                       OCFS2_JOURNAL_ACCESS_WRITE);
3932         if (ret) {
3933                 mlog_errno(ret);
3934                 goto out_commit;
3935         }
3936
3937         spin_lock(&OCFS2_I(t_inode)->ip_lock);
3938         OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
3939         OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
3940         OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
3941         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
3942         i_size_write(t_inode, size);
3943
3944         di->i_xattr_inline_size = s_di->i_xattr_inline_size;
3945         di->i_clusters = s_di->i_clusters;
3946         di->i_size = s_di->i_size;
3947         di->i_dyn_features = s_di->i_dyn_features;
3948         di->i_attr = s_di->i_attr;
3949
3950         if (preserve) {
3951                 di->i_uid = s_di->i_uid;
3952                 di->i_gid = s_di->i_gid;
3953                 di->i_mode = s_di->i_mode;
3954
3955                 /*
3956                  * update time.
3957                  * we want mtime to appear identical to the source and
3958                  * update ctime.
3959                  */
3960                 t_inode->i_ctime = CURRENT_TIME;
3961
3962                 di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec);
3963                 di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec);
3964
3965                 t_inode->i_mtime = s_inode->i_mtime;
3966                 di->i_mtime = s_di->i_mtime;
3967                 di->i_mtime_nsec = s_di->i_mtime_nsec;
3968         }
3969
3970         ocfs2_journal_dirty(handle, t_bh);
3971
3972 out_commit:
3973         ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
3974         return ret;
3975 }
3976
3977 static int ocfs2_create_reflink_node(struct inode *s_inode,
3978                                      struct buffer_head *s_bh,
3979                                      struct inode *t_inode,
3980                                      struct buffer_head *t_bh,
3981                                      bool preserve)
3982 {
3983         int ret;
3984         struct buffer_head *ref_root_bh = NULL;
3985         struct ocfs2_cached_dealloc_ctxt dealloc;
3986         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
3987         struct ocfs2_refcount_block *rb;
3988         struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
3989         struct ocfs2_refcount_tree *ref_tree;
3990
3991         ocfs2_init_dealloc_ctxt(&dealloc);
3992
3993         ret = ocfs2_set_refcount_tree(t_inode, t_bh,
3994                                       le64_to_cpu(di->i_refcount_loc));
3995         if (ret) {
3996                 mlog_errno(ret);
3997                 goto out;
3998         }
3999
4000         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
4001                                        1, &ref_tree, &ref_root_bh);
4002         if (ret) {
4003                 mlog_errno(ret);
4004                 goto out;
4005         }
4006         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
4007
4008         ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4009                                           &ref_tree->rf_ci, ref_root_bh,
4010                                           &dealloc);
4011         if (ret) {
4012                 mlog_errno(ret);
4013                 goto out_unlock_refcount;
4014         }
4015
4016         ret = ocfs2_complete_reflink(s_inode, s_bh, t_inode, t_bh, preserve);
4017         if (ret)
4018                 mlog_errno(ret);
4019
4020 out_unlock_refcount:
4021         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4022         brelse(ref_root_bh);
4023 out:
4024         if (ocfs2_dealloc_has_cluster(&dealloc)) {
4025                 ocfs2_schedule_truncate_log_flush(osb, 1);
4026                 ocfs2_run_deallocs(osb, &dealloc);
4027         }
4028
4029         return ret;
4030 }
4031
4032 static int __ocfs2_reflink(struct dentry *old_dentry,
4033                            struct buffer_head *old_bh,
4034                            struct inode *new_inode,
4035                            bool preserve)
4036 {
4037         int ret;
4038         struct inode *inode = old_dentry->d_inode;
4039         struct buffer_head *new_bh = NULL;
4040
4041         ret = filemap_fdatawrite(inode->i_mapping);
4042         if (ret) {
4043                 mlog_errno(ret);
4044                 goto out;
4045         }
4046
4047         ret = ocfs2_attach_refcount_tree(inode, old_bh);
4048         if (ret) {
4049                 mlog_errno(ret);
4050                 goto out;
4051         }
4052
4053         mutex_lock(&new_inode->i_mutex);
4054         ret = ocfs2_inode_lock(new_inode, &new_bh, 1);
4055         if (ret) {
4056                 mlog_errno(ret);
4057                 goto out_unlock;
4058         }
4059
4060         ret = ocfs2_create_reflink_node(inode, old_bh,
4061                                         new_inode, new_bh, preserve);
4062         if (ret) {
4063                 mlog_errno(ret);
4064                 goto inode_unlock;
4065         }
4066
4067         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
4068                 ret = ocfs2_reflink_xattrs(inode, old_bh,
4069                                            new_inode, new_bh,
4070                                            preserve);
4071                 if (ret)
4072                         mlog_errno(ret);
4073         }
4074 inode_unlock:
4075         ocfs2_inode_unlock(new_inode, 1);
4076         brelse(new_bh);
4077 out_unlock:
4078         mutex_unlock(&new_inode->i_mutex);
4079 out:
4080         if (!ret) {
4081                 ret = filemap_fdatawait(inode->i_mapping);
4082                 if (ret)
4083                         mlog_errno(ret);
4084         }
4085         return ret;
4086 }
4087
4088 static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
4089                          struct dentry *new_dentry, bool preserve)
4090 {
4091         int error;
4092         struct inode *inode = old_dentry->d_inode;
4093         struct buffer_head *old_bh = NULL;
4094         struct inode *new_orphan_inode = NULL;
4095
4096         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4097                 return -EOPNOTSUPP;
4098
4099         error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
4100                                              &new_orphan_inode);
4101         if (error) {
4102                 mlog_errno(error);
4103                 goto out;
4104         }
4105
4106         error = ocfs2_inode_lock(inode, &old_bh, 1);
4107         if (error) {
4108                 mlog_errno(error);
4109                 goto out;
4110         }
4111
4112         down_write(&OCFS2_I(inode)->ip_xattr_sem);
4113         down_write(&OCFS2_I(inode)->ip_alloc_sem);
4114         error = __ocfs2_reflink(old_dentry, old_bh,
4115                                 new_orphan_inode, preserve);
4116         up_write(&OCFS2_I(inode)->ip_alloc_sem);
4117         up_write(&OCFS2_I(inode)->ip_xattr_sem);
4118
4119         ocfs2_inode_unlock(inode, 1);
4120         brelse(old_bh);
4121
4122         if (error) {
4123                 mlog_errno(error);
4124                 goto out;
4125         }
4126
4127         /* If the security isn't preserved, we need to re-initialize them. */
4128         if (!preserve) {
4129                 error = ocfs2_init_security_and_acl(dir, new_orphan_inode);
4130                 if (error)
4131                         mlog_errno(error);
4132         }
4133 out:
4134         if (!error) {
4135                 error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
4136                                                        new_dentry);
4137                 if (error)
4138                         mlog_errno(error);
4139         }
4140
4141         if (new_orphan_inode) {
4142                 /*
4143                  * We need to open_unlock the inode no matter whether we
4144                  * succeed or not, so that other nodes can delete it later.
4145                  */
4146                 ocfs2_open_unlock(new_orphan_inode);
4147                 if (error)
4148                         iput(new_orphan_inode);
4149         }
4150
4151         return error;
4152 }
4153
4154 /*
4155  * Below here are the bits used by OCFS2_IOC_REFLINK() to fake
4156  * sys_reflink().  This will go away when vfs_reflink() exists in
4157  * fs/namei.c.
4158  */
4159
4160 /* copied from may_create in VFS. */
4161 static inline int ocfs2_may_create(struct inode *dir, struct dentry *child)
4162 {
4163         if (child->d_inode)
4164                 return -EEXIST;
4165         if (IS_DEADDIR(dir))
4166                 return -ENOENT;
4167         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
4168 }
4169
4170 /* copied from user_path_parent. */
4171 static int ocfs2_user_path_parent(const char __user *path,
4172                                   struct nameidata *nd, char **name)
4173 {
4174         char *s = getname(path);
4175         int error;
4176
4177         if (IS_ERR(s))
4178                 return PTR_ERR(s);
4179
4180         error = path_lookup(s, LOOKUP_PARENT, nd);
4181         if (error)
4182                 putname(s);
4183         else
4184                 *name = s;
4185
4186         return error;
4187 }
4188
4189 /**
4190  * ocfs2_vfs_reflink - Create a reference-counted link
4191  *
4192  * @old_dentry:        source dentry + inode
4193  * @dir:       directory to create the target
4194  * @new_dentry:        target dentry
4195  * @preserve:  if true, preserve all file attributes
4196  */
4197 int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir,
4198                       struct dentry *new_dentry, bool preserve)
4199 {
4200         struct inode *inode = old_dentry->d_inode;
4201         int error;
4202
4203         if (!inode)
4204                 return -ENOENT;
4205
4206         error = ocfs2_may_create(dir, new_dentry);
4207         if (error)
4208                 return error;
4209
4210         if (dir->i_sb != inode->i_sb)
4211                 return -EXDEV;
4212
4213         /*
4214          * A reflink to an append-only or immutable file cannot be created.
4215          */
4216         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4217                 return -EPERM;
4218
4219         /* Only regular files can be reflinked. */
4220         if (!S_ISREG(inode->i_mode))
4221                 return -EPERM;
4222
4223         /*
4224          * If the caller wants to preserve ownership, they require the
4225          * rights to do so.
4226          */
4227         if (preserve) {
4228                 if ((current_fsuid() != inode->i_uid) && !capable(CAP_CHOWN))
4229                         return -EPERM;
4230                 if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN))
4231                         return -EPERM;
4232         }
4233
4234         /*
4235          * If the caller is modifying any aspect of the attributes, they
4236          * are not creating a snapshot.  They need read permission on the
4237          * file.
4238          */
4239         if (!preserve) {
4240                 error = inode_permission(inode, MAY_READ);
4241                 if (error)
4242                         return error;
4243         }
4244
4245         mutex_lock(&inode->i_mutex);
4246         vfs_dq_init(dir);
4247         error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve);
4248         mutex_unlock(&inode->i_mutex);
4249         if (!error)
4250                 fsnotify_create(dir, new_dentry);
4251         return error;
4252 }
4253 /*
4254  * Most codes are copied from sys_linkat.
4255  */
4256 int ocfs2_reflink_ioctl(struct inode *inode,
4257                         const char __user *oldname,
4258                         const char __user *newname,
4259                         bool preserve)
4260 {
4261         struct dentry *new_dentry;
4262         struct nameidata nd;
4263         struct path old_path;
4264         int error;
4265         char *to = NULL;
4266
4267         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4268                 return -EOPNOTSUPP;
4269
4270         error = user_path_at(AT_FDCWD, oldname, 0, &old_path);
4271         if (error) {
4272                 mlog_errno(error);
4273                 return error;
4274         }
4275
4276         error = ocfs2_user_path_parent(newname, &nd, &to);
4277         if (error) {
4278                 mlog_errno(error);
4279                 goto out;
4280         }
4281
4282         error = -EXDEV;
4283         if (old_path.mnt != nd.path.mnt)
4284                 goto out_release;
4285         new_dentry = lookup_create(&nd, 0);
4286         error = PTR_ERR(new_dentry);
4287         if (IS_ERR(new_dentry)) {
4288                 mlog_errno(error);
4289                 goto out_unlock;
4290         }
4291
4292         error = mnt_want_write(nd.path.mnt);
4293         if (error) {
4294                 mlog_errno(error);
4295                 goto out_dput;
4296         }
4297
4298         error = ocfs2_vfs_reflink(old_path.dentry,
4299                                   nd.path.dentry->d_inode,
4300                                   new_dentry, preserve);
4301         mnt_drop_write(nd.path.mnt);
4302 out_dput:
4303         dput(new_dentry);
4304 out_unlock:
4305         mutex_unlock(&nd.path.dentry->d_inode->i_mutex);
4306 out_release:
4307         path_put(&nd.path);
4308         putname(to);
4309 out:
4310         path_put(&old_path);
4311
4312         return error;
4313 }