GNU Linux-libre 4.19.211-gnu1
[releases.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 #include <cluster/masklog.h>
20 #include "ocfs2.h"
21 #include "inode.h"
22 #include "alloc.h"
23 #include "suballoc.h"
24 #include "journal.h"
25 #include "uptodate.h"
26 #include "super.h"
27 #include "buffer_head_io.h"
28 #include "blockcheck.h"
29 #include "refcounttree.h"
30 #include "sysfile.h"
31 #include "dlmglue.h"
32 #include "extent_map.h"
33 #include "aops.h"
34 #include "xattr.h"
35 #include "namei.h"
36 #include "ocfs2_trace.h"
37 #include "file.h"
38
39 #include <linux/bio.h>
40 #include <linux/blkdev.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 #include <linux/posix_acl.h>
51
52 struct ocfs2_cow_context {
53         struct inode *inode;
54         u32 cow_start;
55         u32 cow_len;
56         struct ocfs2_extent_tree data_et;
57         struct ocfs2_refcount_tree *ref_tree;
58         struct buffer_head *ref_root_bh;
59         struct ocfs2_alloc_context *meta_ac;
60         struct ocfs2_alloc_context *data_ac;
61         struct ocfs2_cached_dealloc_ctxt dealloc;
62         void *cow_object;
63         struct ocfs2_post_refcount *post_refcount;
64         int extra_credits;
65         int (*get_clusters)(struct ocfs2_cow_context *context,
66                             u32 v_cluster, u32 *p_cluster,
67                             u32 *num_clusters,
68                             unsigned int *extent_flags);
69         int (*cow_duplicate_clusters)(handle_t *handle,
70                                       struct inode *inode,
71                                       u32 cpos, u32 old_cluster,
72                                       u32 new_cluster, u32 new_len);
73 };
74
75 static inline struct ocfs2_refcount_tree *
76 cache_info_to_refcount(struct ocfs2_caching_info *ci)
77 {
78         return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
79 }
80
81 static int ocfs2_validate_refcount_block(struct super_block *sb,
82                                          struct buffer_head *bh)
83 {
84         int rc;
85         struct ocfs2_refcount_block *rb =
86                 (struct ocfs2_refcount_block *)bh->b_data;
87
88         trace_ocfs2_validate_refcount_block((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                 rc = ocfs2_error(sb,
107                                  "Refcount block #%llu has bad signature %.*s\n",
108                                  (unsigned long long)bh->b_blocknr, 7,
109                                  rb->rf_signature);
110                 goto out;
111         }
112
113         if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
114                 rc = ocfs2_error(sb,
115                                  "Refcount block #%llu has an invalid rf_blkno of %llu\n",
116                                  (unsigned long long)bh->b_blocknr,
117                                  (unsigned long long)le64_to_cpu(rb->rf_blkno));
118                 goto out;
119         }
120
121         if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
122                 rc = ocfs2_error(sb,
123                                  "Refcount block #%llu has an invalid rf_fs_generation of #%u\n",
124                                  (unsigned long long)bh->b_blocknr,
125                                  le32_to_cpu(rb->rf_fs_generation));
126                 goto out;
127         }
128 out:
129         return rc;
130 }
131
132 static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
133                                      u64 rb_blkno,
134                                      struct buffer_head **bh)
135 {
136         int rc;
137         struct buffer_head *tmp = *bh;
138
139         rc = ocfs2_read_block(ci, rb_blkno, &tmp,
140                               ocfs2_validate_refcount_block);
141
142         /* If ocfs2_read_block() got us a new bh, pass it up. */
143         if (!rc && !*bh)
144                 *bh = tmp;
145
146         return rc;
147 }
148
149 static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
150 {
151         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
152
153         return rf->rf_blkno;
154 }
155
156 static struct super_block *
157 ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
158 {
159         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
160
161         return rf->rf_sb;
162 }
163
164 static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
165 {
166         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
167
168         spin_lock(&rf->rf_lock);
169 }
170
171 static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
172 {
173         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
174
175         spin_unlock(&rf->rf_lock);
176 }
177
178 static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
179 {
180         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
181
182         mutex_lock(&rf->rf_io_mutex);
183 }
184
185 static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
186 {
187         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
188
189         mutex_unlock(&rf->rf_io_mutex);
190 }
191
192 static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
193         .co_owner               = ocfs2_refcount_cache_owner,
194         .co_get_super           = ocfs2_refcount_cache_get_super,
195         .co_cache_lock          = ocfs2_refcount_cache_lock,
196         .co_cache_unlock        = ocfs2_refcount_cache_unlock,
197         .co_io_lock             = ocfs2_refcount_cache_io_lock,
198         .co_io_unlock           = ocfs2_refcount_cache_io_unlock,
199 };
200
201 static struct ocfs2_refcount_tree *
202 ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
203 {
204         struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
205         struct ocfs2_refcount_tree *tree = NULL;
206
207         while (n) {
208                 tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
209
210                 if (blkno < tree->rf_blkno)
211                         n = n->rb_left;
212                 else if (blkno > tree->rf_blkno)
213                         n = n->rb_right;
214                 else
215                         return tree;
216         }
217
218         return NULL;
219 }
220
221 /* osb_lock is already locked. */
222 static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
223                                        struct ocfs2_refcount_tree *new)
224 {
225         u64 rf_blkno = new->rf_blkno;
226         struct rb_node *parent = NULL;
227         struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
228         struct ocfs2_refcount_tree *tmp;
229
230         while (*p) {
231                 parent = *p;
232
233                 tmp = rb_entry(parent, struct ocfs2_refcount_tree,
234                                rf_node);
235
236                 if (rf_blkno < tmp->rf_blkno)
237                         p = &(*p)->rb_left;
238                 else if (rf_blkno > tmp->rf_blkno)
239                         p = &(*p)->rb_right;
240                 else {
241                         /* This should never happen! */
242                         mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
243                              (unsigned long long)rf_blkno);
244                         BUG();
245                 }
246         }
247
248         rb_link_node(&new->rf_node, parent, p);
249         rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
250 }
251
252 static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
253 {
254         ocfs2_metadata_cache_exit(&tree->rf_ci);
255         ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
256         ocfs2_lock_res_free(&tree->rf_lockres);
257         kfree(tree);
258 }
259
260 static inline void
261 ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
262                                         struct ocfs2_refcount_tree *tree)
263 {
264         rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
265         if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
266                 osb->osb_ref_tree_lru = NULL;
267 }
268
269 static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
270                                         struct ocfs2_refcount_tree *tree)
271 {
272         spin_lock(&osb->osb_lock);
273         ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
274         spin_unlock(&osb->osb_lock);
275 }
276
277 static void ocfs2_kref_remove_refcount_tree(struct kref *kref)
278 {
279         struct ocfs2_refcount_tree *tree =
280                 container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
281
282         ocfs2_free_refcount_tree(tree);
283 }
284
285 static inline void
286 ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
287 {
288         kref_get(&tree->rf_getcnt);
289 }
290
291 static inline void
292 ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
293 {
294         kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
295 }
296
297 static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
298                                                struct super_block *sb)
299 {
300         ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
301         mutex_init(&new->rf_io_mutex);
302         new->rf_sb = sb;
303         spin_lock_init(&new->rf_lock);
304 }
305
306 static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
307                                         struct ocfs2_refcount_tree *new,
308                                         u64 rf_blkno, u32 generation)
309 {
310         init_rwsem(&new->rf_sem);
311         ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
312                                      rf_blkno, generation);
313 }
314
315 static struct ocfs2_refcount_tree*
316 ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
317 {
318         struct ocfs2_refcount_tree *new;
319
320         new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
321         if (!new)
322                 return NULL;
323
324         new->rf_blkno = rf_blkno;
325         kref_init(&new->rf_getcnt);
326         ocfs2_init_refcount_tree_ci(new, osb->sb);
327
328         return new;
329 }
330
331 static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
332                                    struct ocfs2_refcount_tree **ret_tree)
333 {
334         int ret = 0;
335         struct ocfs2_refcount_tree *tree, *new = NULL;
336         struct buffer_head *ref_root_bh = NULL;
337         struct ocfs2_refcount_block *ref_rb;
338
339         spin_lock(&osb->osb_lock);
340         if (osb->osb_ref_tree_lru &&
341             osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
342                 tree = osb->osb_ref_tree_lru;
343         else
344                 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
345         if (tree)
346                 goto out;
347
348         spin_unlock(&osb->osb_lock);
349
350         new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
351         if (!new) {
352                 ret = -ENOMEM;
353                 mlog_errno(ret);
354                 return ret;
355         }
356         /*
357          * We need the generation to create the refcount tree lock and since
358          * it isn't changed during the tree modification, we are safe here to
359          * read without protection.
360          * We also have to purge the cache after we create the lock since the
361          * refcount block may have the stale data. It can only be trusted when
362          * we hold the refcount lock.
363          */
364         ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
365         if (ret) {
366                 mlog_errno(ret);
367                 ocfs2_metadata_cache_exit(&new->rf_ci);
368                 kfree(new);
369                 return ret;
370         }
371
372         ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
373         new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
374         ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
375                                       new->rf_generation);
376         ocfs2_metadata_cache_purge(&new->rf_ci);
377
378         spin_lock(&osb->osb_lock);
379         tree = ocfs2_find_refcount_tree(osb, rf_blkno);
380         if (tree)
381                 goto out;
382
383         ocfs2_insert_refcount_tree(osb, new);
384
385         tree = new;
386         new = NULL;
387
388 out:
389         *ret_tree = tree;
390
391         osb->osb_ref_tree_lru = tree;
392
393         spin_unlock(&osb->osb_lock);
394
395         if (new)
396                 ocfs2_free_refcount_tree(new);
397
398         brelse(ref_root_bh);
399         return ret;
400 }
401
402 static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
403 {
404         int ret;
405         struct buffer_head *di_bh = NULL;
406         struct ocfs2_dinode *di;
407
408         ret = ocfs2_read_inode_block(inode, &di_bh);
409         if (ret) {
410                 mlog_errno(ret);
411                 goto out;
412         }
413
414         BUG_ON(!ocfs2_is_refcount_inode(inode));
415
416         di = (struct ocfs2_dinode *)di_bh->b_data;
417         *ref_blkno = le64_to_cpu(di->i_refcount_loc);
418         brelse(di_bh);
419 out:
420         return ret;
421 }
422
423 static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
424                                       struct ocfs2_refcount_tree *tree, int rw)
425 {
426         int ret;
427
428         ret = ocfs2_refcount_lock(tree, rw);
429         if (ret) {
430                 mlog_errno(ret);
431                 goto out;
432         }
433
434         if (rw)
435                 down_write(&tree->rf_sem);
436         else
437                 down_read(&tree->rf_sem);
438
439 out:
440         return ret;
441 }
442
443 /*
444  * Lock the refcount tree pointed by ref_blkno and return the tree.
445  * In most case, we lock the tree and read the refcount block.
446  * So read it here if the caller really needs it.
447  *
448  * If the tree has been re-created by other node, it will free the
449  * old one and re-create it.
450  */
451 int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
452                              u64 ref_blkno, int rw,
453                              struct ocfs2_refcount_tree **ret_tree,
454                              struct buffer_head **ref_bh)
455 {
456         int ret, delete_tree = 0;
457         struct ocfs2_refcount_tree *tree = NULL;
458         struct buffer_head *ref_root_bh = NULL;
459         struct ocfs2_refcount_block *rb;
460
461 again:
462         ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
463         if (ret) {
464                 mlog_errno(ret);
465                 return ret;
466         }
467
468         ocfs2_refcount_tree_get(tree);
469
470         ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
471         if (ret) {
472                 mlog_errno(ret);
473                 ocfs2_refcount_tree_put(tree);
474                 goto out;
475         }
476
477         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
478                                         &ref_root_bh);
479         if (ret) {
480                 mlog_errno(ret);
481                 ocfs2_unlock_refcount_tree(osb, tree, rw);
482                 goto out;
483         }
484
485         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
486         /*
487          * If the refcount block has been freed and re-created, we may need
488          * to recreate the refcount tree also.
489          *
490          * Here we just remove the tree from the rb-tree, and the last
491          * kref holder will unlock and delete this refcount_tree.
492          * Then we goto "again" and ocfs2_get_refcount_tree will create
493          * the new refcount tree for us.
494          */
495         if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
496                 if (!tree->rf_removed) {
497                         ocfs2_erase_refcount_tree_from_list(osb, tree);
498                         tree->rf_removed = 1;
499                         delete_tree = 1;
500                 }
501
502                 ocfs2_unlock_refcount_tree(osb, tree, rw);
503                 /*
504                  * We get an extra reference when we create the refcount
505                  * tree, so another put will destroy it.
506                  */
507                 if (delete_tree)
508                         ocfs2_refcount_tree_put(tree);
509                 brelse(ref_root_bh);
510                 ref_root_bh = NULL;
511                 goto again;
512         }
513
514         *ret_tree = tree;
515         if (ref_bh) {
516                 *ref_bh = ref_root_bh;
517                 ref_root_bh = NULL;
518         }
519 out:
520         brelse(ref_root_bh);
521         return ret;
522 }
523
524 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
525                                 struct ocfs2_refcount_tree *tree, int rw)
526 {
527         if (rw)
528                 up_write(&tree->rf_sem);
529         else
530                 up_read(&tree->rf_sem);
531
532         ocfs2_refcount_unlock(tree, rw);
533         ocfs2_refcount_tree_put(tree);
534 }
535
536 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
537 {
538         struct rb_node *node;
539         struct ocfs2_refcount_tree *tree;
540         struct rb_root *root = &osb->osb_rf_lock_tree;
541
542         while ((node = rb_last(root)) != NULL) {
543                 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
544
545                 trace_ocfs2_purge_refcount_trees(
546                                 (unsigned long long) tree->rf_blkno);
547
548                 rb_erase(&tree->rf_node, root);
549                 ocfs2_free_refcount_tree(tree);
550         }
551 }
552
553 /*
554  * Create a refcount tree for an inode.
555  * We take for granted that the inode is already locked.
556  */
557 static int ocfs2_create_refcount_tree(struct inode *inode,
558                                       struct buffer_head *di_bh)
559 {
560         int ret;
561         handle_t *handle = NULL;
562         struct ocfs2_alloc_context *meta_ac = NULL;
563         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
564         struct ocfs2_inode_info *oi = OCFS2_I(inode);
565         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
566         struct buffer_head *new_bh = NULL;
567         struct ocfs2_refcount_block *rb;
568         struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
569         u16 suballoc_bit_start;
570         u32 num_got;
571         u64 suballoc_loc, first_blkno;
572
573         BUG_ON(ocfs2_is_refcount_inode(inode));
574
575         trace_ocfs2_create_refcount_tree(
576                 (unsigned long long)oi->ip_blkno);
577
578         ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
579         if (ret) {
580                 mlog_errno(ret);
581                 goto out;
582         }
583
584         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
585         if (IS_ERR(handle)) {
586                 ret = PTR_ERR(handle);
587                 mlog_errno(ret);
588                 goto out;
589         }
590
591         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
592                                       OCFS2_JOURNAL_ACCESS_WRITE);
593         if (ret) {
594                 mlog_errno(ret);
595                 goto out_commit;
596         }
597
598         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
599                                    &suballoc_bit_start, &num_got,
600                                    &first_blkno);
601         if (ret) {
602                 mlog_errno(ret);
603                 goto out_commit;
604         }
605
606         new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
607         if (!new_tree) {
608                 ret = -ENOMEM;
609                 mlog_errno(ret);
610                 goto out_commit;
611         }
612
613         new_bh = sb_getblk(inode->i_sb, first_blkno);
614         if (!new_bh) {
615                 ret = -ENOMEM;
616                 mlog_errno(ret);
617                 goto out_commit;
618         }
619         ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
620
621         ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
622                                       OCFS2_JOURNAL_ACCESS_CREATE);
623         if (ret) {
624                 mlog_errno(ret);
625                 goto out_commit;
626         }
627
628         /* Initialize ocfs2_refcount_block. */
629         rb = (struct ocfs2_refcount_block *)new_bh->b_data;
630         memset(rb, 0, inode->i_sb->s_blocksize);
631         strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
632         rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
633         rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
634         rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
635         rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
636         rb->rf_blkno = cpu_to_le64(first_blkno);
637         rb->rf_count = cpu_to_le32(1);
638         rb->rf_records.rl_count =
639                         cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
640         spin_lock(&osb->osb_lock);
641         rb->rf_generation = osb->s_next_generation++;
642         spin_unlock(&osb->osb_lock);
643
644         ocfs2_journal_dirty(handle, new_bh);
645
646         spin_lock(&oi->ip_lock);
647         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
648         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
649         di->i_refcount_loc = cpu_to_le64(first_blkno);
650         spin_unlock(&oi->ip_lock);
651
652         trace_ocfs2_create_refcount_tree_blkno((unsigned long long)first_blkno);
653
654         ocfs2_journal_dirty(handle, di_bh);
655
656         /*
657          * We have to init the tree lock here since it will use
658          * the generation number to create it.
659          */
660         new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
661         ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
662                                       new_tree->rf_generation);
663
664         spin_lock(&osb->osb_lock);
665         tree = ocfs2_find_refcount_tree(osb, first_blkno);
666
667         /*
668          * We've just created a new refcount tree in this block.  If
669          * we found a refcount tree on the ocfs2_super, it must be
670          * one we just deleted.  We free the old tree before
671          * inserting the new tree.
672          */
673         BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
674         if (tree)
675                 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
676         ocfs2_insert_refcount_tree(osb, new_tree);
677         spin_unlock(&osb->osb_lock);
678         new_tree = NULL;
679         if (tree)
680                 ocfs2_refcount_tree_put(tree);
681
682 out_commit:
683         ocfs2_commit_trans(osb, handle);
684
685 out:
686         if (new_tree) {
687                 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
688                 kfree(new_tree);
689         }
690
691         brelse(new_bh);
692         if (meta_ac)
693                 ocfs2_free_alloc_context(meta_ac);
694
695         return ret;
696 }
697
698 static int ocfs2_set_refcount_tree(struct inode *inode,
699                                    struct buffer_head *di_bh,
700                                    u64 refcount_loc)
701 {
702         int ret;
703         handle_t *handle = NULL;
704         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
705         struct ocfs2_inode_info *oi = OCFS2_I(inode);
706         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
707         struct buffer_head *ref_root_bh = NULL;
708         struct ocfs2_refcount_block *rb;
709         struct ocfs2_refcount_tree *ref_tree;
710
711         BUG_ON(ocfs2_is_refcount_inode(inode));
712
713         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
714                                        &ref_tree, &ref_root_bh);
715         if (ret) {
716                 mlog_errno(ret);
717                 return ret;
718         }
719
720         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
721         if (IS_ERR(handle)) {
722                 ret = PTR_ERR(handle);
723                 mlog_errno(ret);
724                 goto out;
725         }
726
727         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
728                                       OCFS2_JOURNAL_ACCESS_WRITE);
729         if (ret) {
730                 mlog_errno(ret);
731                 goto out_commit;
732         }
733
734         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
735                                       OCFS2_JOURNAL_ACCESS_WRITE);
736         if (ret) {
737                 mlog_errno(ret);
738                 goto out_commit;
739         }
740
741         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
742         le32_add_cpu(&rb->rf_count, 1);
743
744         ocfs2_journal_dirty(handle, ref_root_bh);
745
746         spin_lock(&oi->ip_lock);
747         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
748         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
749         di->i_refcount_loc = cpu_to_le64(refcount_loc);
750         spin_unlock(&oi->ip_lock);
751         ocfs2_journal_dirty(handle, di_bh);
752
753 out_commit:
754         ocfs2_commit_trans(osb, handle);
755 out:
756         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
757         brelse(ref_root_bh);
758
759         return ret;
760 }
761
762 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
763 {
764         int ret, delete_tree = 0;
765         handle_t *handle = NULL;
766         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
767         struct ocfs2_inode_info *oi = OCFS2_I(inode);
768         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
769         struct ocfs2_refcount_block *rb;
770         struct inode *alloc_inode = NULL;
771         struct buffer_head *alloc_bh = NULL;
772         struct buffer_head *blk_bh = NULL;
773         struct ocfs2_refcount_tree *ref_tree;
774         int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
775         u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
776         u16 bit = 0;
777
778         if (!ocfs2_is_refcount_inode(inode))
779                 return 0;
780
781         BUG_ON(!ref_blkno);
782         ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
783         if (ret) {
784                 mlog_errno(ret);
785                 return ret;
786         }
787
788         rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
789
790         /*
791          * If we are the last user, we need to free the block.
792          * So lock the allocator ahead.
793          */
794         if (le32_to_cpu(rb->rf_count) == 1) {
795                 blk = le64_to_cpu(rb->rf_blkno);
796                 bit = le16_to_cpu(rb->rf_suballoc_bit);
797                 if (rb->rf_suballoc_loc)
798                         bg_blkno = le64_to_cpu(rb->rf_suballoc_loc);
799                 else
800                         bg_blkno = ocfs2_which_suballoc_group(blk, bit);
801
802                 alloc_inode = ocfs2_get_system_file_inode(osb,
803                                         EXTENT_ALLOC_SYSTEM_INODE,
804                                         le16_to_cpu(rb->rf_suballoc_slot));
805                 if (!alloc_inode) {
806                         ret = -ENOMEM;
807                         mlog_errno(ret);
808                         goto out;
809                 }
810                 inode_lock(alloc_inode);
811
812                 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
813                 if (ret) {
814                         mlog_errno(ret);
815                         goto out_mutex;
816                 }
817
818                 credits += OCFS2_SUBALLOC_FREE;
819         }
820
821         handle = ocfs2_start_trans(osb, credits);
822         if (IS_ERR(handle)) {
823                 ret = PTR_ERR(handle);
824                 mlog_errno(ret);
825                 goto out_unlock;
826         }
827
828         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
829                                       OCFS2_JOURNAL_ACCESS_WRITE);
830         if (ret) {
831                 mlog_errno(ret);
832                 goto out_commit;
833         }
834
835         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
836                                       OCFS2_JOURNAL_ACCESS_WRITE);
837         if (ret) {
838                 mlog_errno(ret);
839                 goto out_commit;
840         }
841
842         spin_lock(&oi->ip_lock);
843         oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
844         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
845         di->i_refcount_loc = 0;
846         spin_unlock(&oi->ip_lock);
847         ocfs2_journal_dirty(handle, di_bh);
848
849         le32_add_cpu(&rb->rf_count , -1);
850         ocfs2_journal_dirty(handle, blk_bh);
851
852         if (!rb->rf_count) {
853                 delete_tree = 1;
854                 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
855                 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
856                                                alloc_bh, bit, bg_blkno, 1);
857                 if (ret)
858                         mlog_errno(ret);
859         }
860
861 out_commit:
862         ocfs2_commit_trans(osb, handle);
863 out_unlock:
864         if (alloc_inode) {
865                 ocfs2_inode_unlock(alloc_inode, 1);
866                 brelse(alloc_bh);
867         }
868 out_mutex:
869         if (alloc_inode) {
870                 inode_unlock(alloc_inode);
871                 iput(alloc_inode);
872         }
873 out:
874         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
875         if (delete_tree)
876                 ocfs2_refcount_tree_put(ref_tree);
877         brelse(blk_bh);
878
879         return ret;
880 }
881
882 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
883                                           struct buffer_head *ref_leaf_bh,
884                                           u64 cpos, unsigned int len,
885                                           struct ocfs2_refcount_rec *ret_rec,
886                                           int *index)
887 {
888         int i = 0;
889         struct ocfs2_refcount_block *rb =
890                 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
891         struct ocfs2_refcount_rec *rec = NULL;
892
893         for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
894                 rec = &rb->rf_records.rl_recs[i];
895
896                 if (le64_to_cpu(rec->r_cpos) +
897                     le32_to_cpu(rec->r_clusters) <= cpos)
898                         continue;
899                 else if (le64_to_cpu(rec->r_cpos) > cpos)
900                         break;
901
902                 /* ok, cpos fail in this rec. Just return. */
903                 if (ret_rec)
904                         *ret_rec = *rec;
905                 goto out;
906         }
907
908         if (ret_rec) {
909                 /* We meet with a hole here, so fake the rec. */
910                 ret_rec->r_cpos = cpu_to_le64(cpos);
911                 ret_rec->r_refcount = 0;
912                 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
913                     le64_to_cpu(rec->r_cpos) < cpos + len)
914                         ret_rec->r_clusters =
915                                 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
916                 else
917                         ret_rec->r_clusters = cpu_to_le32(len);
918         }
919
920 out:
921         *index = i;
922 }
923
924 /*
925  * Try to remove refcount tree. The mechanism is:
926  * 1) Check whether i_clusters == 0, if no, exit.
927  * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
928  * 3) Check whether we have inline xattr stored outside, if yes, exit.
929  * 4) Remove the tree.
930  */
931 int ocfs2_try_remove_refcount_tree(struct inode *inode,
932                                    struct buffer_head *di_bh)
933 {
934         int ret;
935         struct ocfs2_inode_info *oi = OCFS2_I(inode);
936         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
937
938         down_write(&oi->ip_xattr_sem);
939         down_write(&oi->ip_alloc_sem);
940
941         if (oi->ip_clusters)
942                 goto out;
943
944         if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
945                 goto out;
946
947         if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
948             ocfs2_has_inline_xattr_value_outside(inode, di))
949                 goto out;
950
951         ret = ocfs2_remove_refcount_tree(inode, di_bh);
952         if (ret)
953                 mlog_errno(ret);
954 out:
955         up_write(&oi->ip_alloc_sem);
956         up_write(&oi->ip_xattr_sem);
957         return 0;
958 }
959
960 /*
961  * Find the end range for a leaf refcount block indicated by
962  * el->l_recs[index].e_blkno.
963  */
964 static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci,
965                                        struct buffer_head *ref_root_bh,
966                                        struct ocfs2_extent_block *eb,
967                                        struct ocfs2_extent_list *el,
968                                        int index,  u32 *cpos_end)
969 {
970         int ret, i, subtree_root;
971         u32 cpos;
972         u64 blkno;
973         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
974         struct ocfs2_path *left_path = NULL, *right_path = NULL;
975         struct ocfs2_extent_tree et;
976         struct ocfs2_extent_list *tmp_el;
977
978         if (index < le16_to_cpu(el->l_next_free_rec) - 1) {
979                 /*
980                  * We have a extent rec after index, so just use the e_cpos
981                  * of the next extent rec.
982                  */
983                 *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos);
984                 return 0;
985         }
986
987         if (!eb || (eb && !eb->h_next_leaf_blk)) {
988                 /*
989                  * We are the last extent rec, so any high cpos should
990                  * be stored in this leaf refcount block.
991                  */
992                 *cpos_end = UINT_MAX;
993                 return 0;
994         }
995
996         /*
997          * If the extent block isn't the last one, we have to find
998          * the subtree root between this extent block and the next
999          * leaf extent block and get the corresponding e_cpos from
1000          * the subroot. Otherwise we may corrupt the b-tree.
1001          */
1002         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1003
1004         left_path = ocfs2_new_path_from_et(&et);
1005         if (!left_path) {
1006                 ret = -ENOMEM;
1007                 mlog_errno(ret);
1008                 goto out;
1009         }
1010
1011         cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos);
1012         ret = ocfs2_find_path(ci, left_path, cpos);
1013         if (ret) {
1014                 mlog_errno(ret);
1015                 goto out;
1016         }
1017
1018         right_path = ocfs2_new_path_from_path(left_path);
1019         if (!right_path) {
1020                 ret = -ENOMEM;
1021                 mlog_errno(ret);
1022                 goto out;
1023         }
1024
1025         ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos);
1026         if (ret) {
1027                 mlog_errno(ret);
1028                 goto out;
1029         }
1030
1031         ret = ocfs2_find_path(ci, right_path, cpos);
1032         if (ret) {
1033                 mlog_errno(ret);
1034                 goto out;
1035         }
1036
1037         subtree_root = ocfs2_find_subtree_root(&et, left_path,
1038                                                right_path);
1039
1040         tmp_el = left_path->p_node[subtree_root].el;
1041         blkno = left_path->p_node[subtree_root+1].bh->b_blocknr;
1042         for (i = 0; i < le16_to_cpu(tmp_el->l_next_free_rec); i++) {
1043                 if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) {
1044                         *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos);
1045                         break;
1046                 }
1047         }
1048
1049         BUG_ON(i == le16_to_cpu(tmp_el->l_next_free_rec));
1050
1051 out:
1052         ocfs2_free_path(left_path);
1053         ocfs2_free_path(right_path);
1054         return ret;
1055 }
1056
1057 /*
1058  * Given a cpos and len, try to find the refcount record which contains cpos.
1059  * 1. If cpos can be found in one refcount record, return the record.
1060  * 2. If cpos can't be found, return a fake record which start from cpos
1061  *    and end at a small value between cpos+len and start of the next record.
1062  *    This fake record has r_refcount = 0.
1063  */
1064 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
1065                                   struct buffer_head *ref_root_bh,
1066                                   u64 cpos, unsigned int len,
1067                                   struct ocfs2_refcount_rec *ret_rec,
1068                                   int *index,
1069                                   struct buffer_head **ret_bh)
1070 {
1071         int ret = 0, i, found;
1072         u32 low_cpos, uninitialized_var(cpos_end);
1073         struct ocfs2_extent_list *el;
1074         struct ocfs2_extent_rec *rec = NULL;
1075         struct ocfs2_extent_block *eb = NULL;
1076         struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
1077         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1078         struct ocfs2_refcount_block *rb =
1079                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1080
1081         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
1082                 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
1083                                               ret_rec, index);
1084                 *ret_bh = ref_root_bh;
1085                 get_bh(ref_root_bh);
1086                 return 0;
1087         }
1088
1089         el = &rb->rf_list;
1090         low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1091
1092         if (el->l_tree_depth) {
1093                 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1094                 if (ret) {
1095                         mlog_errno(ret);
1096                         goto out;
1097                 }
1098
1099                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1100                 el = &eb->h_list;
1101
1102                 if (el->l_tree_depth) {
1103                         ret = ocfs2_error(sb,
1104                                           "refcount tree %llu has non zero tree depth in leaf btree tree block %llu\n",
1105                                           (unsigned long long)ocfs2_metadata_cache_owner(ci),
1106                                           (unsigned long long)eb_bh->b_blocknr);
1107                         goto out;
1108                 }
1109         }
1110
1111         found = 0;
1112         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1113                 rec = &el->l_recs[i];
1114
1115                 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1116                         found = 1;
1117                         break;
1118                 }
1119         }
1120
1121         if (found) {
1122                 ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh,
1123                                                   eb, el, i, &cpos_end);
1124                 if (ret) {
1125                         mlog_errno(ret);
1126                         goto out;
1127                 }
1128
1129                 if (cpos_end < low_cpos + len)
1130                         len = cpos_end - low_cpos;
1131         }
1132
1133         ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1134                                         &ref_leaf_bh);
1135         if (ret) {
1136                 mlog_errno(ret);
1137                 goto out;
1138         }
1139
1140         ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1141                                       ret_rec, index);
1142         *ret_bh = ref_leaf_bh;
1143 out:
1144         brelse(eb_bh);
1145         return ret;
1146 }
1147
1148 enum ocfs2_ref_rec_contig {
1149         REF_CONTIG_NONE = 0,
1150         REF_CONTIG_LEFT,
1151         REF_CONTIG_RIGHT,
1152         REF_CONTIG_LEFTRIGHT,
1153 };
1154
1155 static enum ocfs2_ref_rec_contig
1156         ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1157                                     int index)
1158 {
1159         if ((rb->rf_records.rl_recs[index].r_refcount ==
1160             rb->rf_records.rl_recs[index + 1].r_refcount) &&
1161             (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1162             le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1163             le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1164                 return REF_CONTIG_RIGHT;
1165
1166         return REF_CONTIG_NONE;
1167 }
1168
1169 static enum ocfs2_ref_rec_contig
1170         ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1171                                   int index)
1172 {
1173         enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1174
1175         if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1176                 ret = ocfs2_refcount_rec_adjacent(rb, index);
1177
1178         if (index > 0) {
1179                 enum ocfs2_ref_rec_contig tmp;
1180
1181                 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1182
1183                 if (tmp == REF_CONTIG_RIGHT) {
1184                         if (ret == REF_CONTIG_RIGHT)
1185                                 ret = REF_CONTIG_LEFTRIGHT;
1186                         else
1187                                 ret = REF_CONTIG_LEFT;
1188                 }
1189         }
1190
1191         return ret;
1192 }
1193
1194 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1195                                            int index)
1196 {
1197         BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1198                rb->rf_records.rl_recs[index+1].r_refcount);
1199
1200         le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1201                      le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1202
1203         if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1204                 memmove(&rb->rf_records.rl_recs[index + 1],
1205                         &rb->rf_records.rl_recs[index + 2],
1206                         sizeof(struct ocfs2_refcount_rec) *
1207                         (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1208
1209         memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1210                0, sizeof(struct ocfs2_refcount_rec));
1211         le16_add_cpu(&rb->rf_records.rl_used, -1);
1212 }
1213
1214 /*
1215  * Merge the refcount rec if we are contiguous with the adjacent recs.
1216  */
1217 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1218                                      int index)
1219 {
1220         enum ocfs2_ref_rec_contig contig =
1221                                 ocfs2_refcount_rec_contig(rb, index);
1222
1223         if (contig == REF_CONTIG_NONE)
1224                 return;
1225
1226         if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1227                 BUG_ON(index == 0);
1228                 index--;
1229         }
1230
1231         ocfs2_rotate_refcount_rec_left(rb, index);
1232
1233         if (contig == REF_CONTIG_LEFTRIGHT)
1234                 ocfs2_rotate_refcount_rec_left(rb, index);
1235 }
1236
1237 /*
1238  * Change the refcount indexed by "index" in ref_bh.
1239  * If refcount reaches 0, remove it.
1240  */
1241 static int ocfs2_change_refcount_rec(handle_t *handle,
1242                                      struct ocfs2_caching_info *ci,
1243                                      struct buffer_head *ref_leaf_bh,
1244                                      int index, int merge, int change)
1245 {
1246         int ret;
1247         struct ocfs2_refcount_block *rb =
1248                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1249         struct ocfs2_refcount_list *rl = &rb->rf_records;
1250         struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1251
1252         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1253                                       OCFS2_JOURNAL_ACCESS_WRITE);
1254         if (ret) {
1255                 mlog_errno(ret);
1256                 goto out;
1257         }
1258
1259         trace_ocfs2_change_refcount_rec(
1260                 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1261                 index, le32_to_cpu(rec->r_refcount), change);
1262         le32_add_cpu(&rec->r_refcount, change);
1263
1264         if (!rec->r_refcount) {
1265                 if (index != le16_to_cpu(rl->rl_used) - 1) {
1266                         memmove(rec, rec + 1,
1267                                 (le16_to_cpu(rl->rl_used) - index - 1) *
1268                                 sizeof(struct ocfs2_refcount_rec));
1269                         memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1270                                0, sizeof(struct ocfs2_refcount_rec));
1271                 }
1272
1273                 le16_add_cpu(&rl->rl_used, -1);
1274         } else if (merge)
1275                 ocfs2_refcount_rec_merge(rb, index);
1276
1277         ocfs2_journal_dirty(handle, ref_leaf_bh);
1278 out:
1279         return ret;
1280 }
1281
1282 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1283                                         struct ocfs2_caching_info *ci,
1284                                         struct buffer_head *ref_root_bh,
1285                                         struct buffer_head **ref_leaf_bh,
1286                                         struct ocfs2_alloc_context *meta_ac)
1287 {
1288         int ret;
1289         u16 suballoc_bit_start;
1290         u32 num_got;
1291         u64 suballoc_loc, blkno;
1292         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1293         struct buffer_head *new_bh = NULL;
1294         struct ocfs2_refcount_block *new_rb;
1295         struct ocfs2_refcount_block *root_rb =
1296                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1297
1298         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1299                                       OCFS2_JOURNAL_ACCESS_WRITE);
1300         if (ret) {
1301                 mlog_errno(ret);
1302                 goto out;
1303         }
1304
1305         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1306                                    &suballoc_bit_start, &num_got,
1307                                    &blkno);
1308         if (ret) {
1309                 mlog_errno(ret);
1310                 goto out;
1311         }
1312
1313         new_bh = sb_getblk(sb, blkno);
1314         if (new_bh == NULL) {
1315                 ret = -ENOMEM;
1316                 mlog_errno(ret);
1317                 goto out;
1318         }
1319         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1320
1321         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1322                                       OCFS2_JOURNAL_ACCESS_CREATE);
1323         if (ret) {
1324                 mlog_errno(ret);
1325                 goto out;
1326         }
1327
1328         /*
1329          * Initialize ocfs2_refcount_block.
1330          * It should contain the same information as the old root.
1331          * so just memcpy it and change the corresponding field.
1332          */
1333         memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1334
1335         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1336         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1337         new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1338         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1339         new_rb->rf_blkno = cpu_to_le64(blkno);
1340         new_rb->rf_cpos = cpu_to_le32(0);
1341         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1342         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1343         ocfs2_journal_dirty(handle, new_bh);
1344
1345         /* Now change the root. */
1346         memset(&root_rb->rf_list, 0, sb->s_blocksize -
1347                offsetof(struct ocfs2_refcount_block, rf_list));
1348         root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1349         root_rb->rf_clusters = cpu_to_le32(1);
1350         root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1351         root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1352         root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1353         root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1354
1355         ocfs2_journal_dirty(handle, ref_root_bh);
1356
1357         trace_ocfs2_expand_inline_ref_root((unsigned long long)blkno,
1358                 le16_to_cpu(new_rb->rf_records.rl_used));
1359
1360         *ref_leaf_bh = new_bh;
1361         new_bh = NULL;
1362 out:
1363         brelse(new_bh);
1364         return ret;
1365 }
1366
1367 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1368                                            struct ocfs2_refcount_rec *next)
1369 {
1370         if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1371                 ocfs2_get_ref_rec_low_cpos(next))
1372                 return 1;
1373
1374         return 0;
1375 }
1376
1377 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1378 {
1379         const struct ocfs2_refcount_rec *l = a, *r = b;
1380         u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1381         u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1382
1383         if (l_cpos > r_cpos)
1384                 return 1;
1385         if (l_cpos < r_cpos)
1386                 return -1;
1387         return 0;
1388 }
1389
1390 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1391 {
1392         const struct ocfs2_refcount_rec *l = a, *r = b;
1393         u64 l_cpos = le64_to_cpu(l->r_cpos);
1394         u64 r_cpos = le64_to_cpu(r->r_cpos);
1395
1396         if (l_cpos > r_cpos)
1397                 return 1;
1398         if (l_cpos < r_cpos)
1399                 return -1;
1400         return 0;
1401 }
1402
1403 static void swap_refcount_rec(void *a, void *b, int size)
1404 {
1405         struct ocfs2_refcount_rec *l = a, *r = b;
1406
1407         swap(*l, *r);
1408 }
1409
1410 /*
1411  * The refcount cpos are ordered by their 64bit cpos,
1412  * But we will use the low 32 bit to be the e_cpos in the b-tree.
1413  * So we need to make sure that this pos isn't intersected with others.
1414  *
1415  * Note: The refcount block is already sorted by their low 32 bit cpos,
1416  *       So just try the middle pos first, and we will exit when we find
1417  *       the good position.
1418  */
1419 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1420                                          u32 *split_pos, int *split_index)
1421 {
1422         int num_used = le16_to_cpu(rl->rl_used);
1423         int delta, middle = num_used / 2;
1424
1425         for (delta = 0; delta < middle; delta++) {
1426                 /* Let's check delta earlier than middle */
1427                 if (ocfs2_refcount_rec_no_intersect(
1428                                         &rl->rl_recs[middle - delta - 1],
1429                                         &rl->rl_recs[middle - delta])) {
1430                         *split_index = middle - delta;
1431                         break;
1432                 }
1433
1434                 /* For even counts, don't walk off the end */
1435                 if ((middle + delta + 1) == num_used)
1436                         continue;
1437
1438                 /* Now try delta past middle */
1439                 if (ocfs2_refcount_rec_no_intersect(
1440                                         &rl->rl_recs[middle + delta],
1441                                         &rl->rl_recs[middle + delta + 1])) {
1442                         *split_index = middle + delta + 1;
1443                         break;
1444                 }
1445         }
1446
1447         if (delta >= middle)
1448                 return -ENOSPC;
1449
1450         *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1451         return 0;
1452 }
1453
1454 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1455                                             struct buffer_head *new_bh,
1456                                             u32 *split_cpos)
1457 {
1458         int split_index = 0, num_moved, ret;
1459         u32 cpos = 0;
1460         struct ocfs2_refcount_block *rb =
1461                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1462         struct ocfs2_refcount_list *rl = &rb->rf_records;
1463         struct ocfs2_refcount_block *new_rb =
1464                         (struct ocfs2_refcount_block *)new_bh->b_data;
1465         struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1466
1467         trace_ocfs2_divide_leaf_refcount_block(
1468                 (unsigned long long)ref_leaf_bh->b_blocknr,
1469                 le16_to_cpu(rl->rl_count), le16_to_cpu(rl->rl_used));
1470
1471         /*
1472          * XXX: Improvement later.
1473          * If we know all the high 32 bit cpos is the same, no need to sort.
1474          *
1475          * In order to make the whole process safe, we do:
1476          * 1. sort the entries by their low 32 bit cpos first so that we can
1477          *    find the split cpos easily.
1478          * 2. call ocfs2_insert_extent to insert the new refcount block.
1479          * 3. move the refcount rec to the new block.
1480          * 4. sort the entries by their 64 bit cpos.
1481          * 5. dirty the new_rb and rb.
1482          */
1483         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1484              sizeof(struct ocfs2_refcount_rec),
1485              cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1486
1487         ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1488         if (ret) {
1489                 mlog_errno(ret);
1490                 return ret;
1491         }
1492
1493         new_rb->rf_cpos = cpu_to_le32(cpos);
1494
1495         /* move refcount records starting from split_index to the new block. */
1496         num_moved = le16_to_cpu(rl->rl_used) - split_index;
1497         memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1498                num_moved * sizeof(struct ocfs2_refcount_rec));
1499
1500         /*ok, remove the entries we just moved over to the other block. */
1501         memset(&rl->rl_recs[split_index], 0,
1502                num_moved * sizeof(struct ocfs2_refcount_rec));
1503
1504         /* change old and new rl_used accordingly. */
1505         le16_add_cpu(&rl->rl_used, -num_moved);
1506         new_rl->rl_used = cpu_to_le16(num_moved);
1507
1508         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1509              sizeof(struct ocfs2_refcount_rec),
1510              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1511
1512         sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1513              sizeof(struct ocfs2_refcount_rec),
1514              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1515
1516         *split_cpos = cpos;
1517         return 0;
1518 }
1519
1520 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1521                                          struct ocfs2_caching_info *ci,
1522                                          struct buffer_head *ref_root_bh,
1523                                          struct buffer_head *ref_leaf_bh,
1524                                          struct ocfs2_alloc_context *meta_ac)
1525 {
1526         int ret;
1527         u16 suballoc_bit_start;
1528         u32 num_got, new_cpos;
1529         u64 suballoc_loc, blkno;
1530         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1531         struct ocfs2_refcount_block *root_rb =
1532                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1533         struct buffer_head *new_bh = NULL;
1534         struct ocfs2_refcount_block *new_rb;
1535         struct ocfs2_extent_tree ref_et;
1536
1537         BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1538
1539         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1540                                       OCFS2_JOURNAL_ACCESS_WRITE);
1541         if (ret) {
1542                 mlog_errno(ret);
1543                 goto out;
1544         }
1545
1546         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1547                                       OCFS2_JOURNAL_ACCESS_WRITE);
1548         if (ret) {
1549                 mlog_errno(ret);
1550                 goto out;
1551         }
1552
1553         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1554                                    &suballoc_bit_start, &num_got,
1555                                    &blkno);
1556         if (ret) {
1557                 mlog_errno(ret);
1558                 goto out;
1559         }
1560
1561         new_bh = sb_getblk(sb, blkno);
1562         if (new_bh == NULL) {
1563                 ret = -ENOMEM;
1564                 mlog_errno(ret);
1565                 goto out;
1566         }
1567         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1568
1569         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1570                                       OCFS2_JOURNAL_ACCESS_CREATE);
1571         if (ret) {
1572                 mlog_errno(ret);
1573                 goto out;
1574         }
1575
1576         /* Initialize ocfs2_refcount_block. */
1577         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1578         memset(new_rb, 0, sb->s_blocksize);
1579         strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1580         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1581         new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1582         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1583         new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1584         new_rb->rf_blkno = cpu_to_le64(blkno);
1585         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1586         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1587         new_rb->rf_records.rl_count =
1588                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1589         new_rb->rf_generation = root_rb->rf_generation;
1590
1591         ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1592         if (ret) {
1593                 mlog_errno(ret);
1594                 goto out;
1595         }
1596
1597         ocfs2_journal_dirty(handle, ref_leaf_bh);
1598         ocfs2_journal_dirty(handle, new_bh);
1599
1600         ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1601
1602         trace_ocfs2_new_leaf_refcount_block(
1603                         (unsigned long long)new_bh->b_blocknr, new_cpos);
1604
1605         /* Insert the new leaf block with the specific offset cpos. */
1606         ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1607                                   1, 0, meta_ac);
1608         if (ret)
1609                 mlog_errno(ret);
1610
1611 out:
1612         brelse(new_bh);
1613         return ret;
1614 }
1615
1616 static int ocfs2_expand_refcount_tree(handle_t *handle,
1617                                       struct ocfs2_caching_info *ci,
1618                                       struct buffer_head *ref_root_bh,
1619                                       struct buffer_head *ref_leaf_bh,
1620                                       struct ocfs2_alloc_context *meta_ac)
1621 {
1622         int ret;
1623         struct buffer_head *expand_bh = NULL;
1624
1625         if (ref_root_bh == ref_leaf_bh) {
1626                 /*
1627                  * the old root bh hasn't been expanded to a b-tree,
1628                  * so expand it first.
1629                  */
1630                 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1631                                                    &expand_bh, meta_ac);
1632                 if (ret) {
1633                         mlog_errno(ret);
1634                         goto out;
1635                 }
1636         } else {
1637                 expand_bh = ref_leaf_bh;
1638                 get_bh(expand_bh);
1639         }
1640
1641
1642         /* Now add a new refcount block into the tree.*/
1643         ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1644                                             expand_bh, meta_ac);
1645         if (ret)
1646                 mlog_errno(ret);
1647 out:
1648         brelse(expand_bh);
1649         return ret;
1650 }
1651
1652 /*
1653  * Adjust the extent rec in b-tree representing ref_leaf_bh.
1654  *
1655  * Only called when we have inserted a new refcount rec at index 0
1656  * which means ocfs2_extent_rec.e_cpos may need some change.
1657  */
1658 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1659                                      struct ocfs2_caching_info *ci,
1660                                      struct buffer_head *ref_root_bh,
1661                                      struct buffer_head *ref_leaf_bh,
1662                                      struct ocfs2_refcount_rec *rec)
1663 {
1664         int ret = 0, i;
1665         u32 new_cpos, old_cpos;
1666         struct ocfs2_path *path = NULL;
1667         struct ocfs2_extent_tree et;
1668         struct ocfs2_refcount_block *rb =
1669                 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1670         struct ocfs2_extent_list *el;
1671
1672         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1673                 goto out;
1674
1675         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1676         old_cpos = le32_to_cpu(rb->rf_cpos);
1677         new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1678         if (old_cpos <= new_cpos)
1679                 goto out;
1680
1681         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1682
1683         path = ocfs2_new_path_from_et(&et);
1684         if (!path) {
1685                 ret = -ENOMEM;
1686                 mlog_errno(ret);
1687                 goto out;
1688         }
1689
1690         ret = ocfs2_find_path(ci, path, old_cpos);
1691         if (ret) {
1692                 mlog_errno(ret);
1693                 goto out;
1694         }
1695
1696         /*
1697          * 2 more credits, one for the leaf refcount block, one for
1698          * the extent block contains the extent rec.
1699          */
1700         ret = ocfs2_extend_trans(handle, 2);
1701         if (ret < 0) {
1702                 mlog_errno(ret);
1703                 goto out;
1704         }
1705
1706         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1707                                       OCFS2_JOURNAL_ACCESS_WRITE);
1708         if (ret < 0) {
1709                 mlog_errno(ret);
1710                 goto out;
1711         }
1712
1713         ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1714                                       OCFS2_JOURNAL_ACCESS_WRITE);
1715         if (ret < 0) {
1716                 mlog_errno(ret);
1717                 goto out;
1718         }
1719
1720         /* change the leaf extent block first. */
1721         el = path_leaf_el(path);
1722
1723         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1724                 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1725                         break;
1726
1727         BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1728
1729         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1730
1731         /* change the r_cpos in the leaf block. */
1732         rb->rf_cpos = cpu_to_le32(new_cpos);
1733
1734         ocfs2_journal_dirty(handle, path_leaf_bh(path));
1735         ocfs2_journal_dirty(handle, ref_leaf_bh);
1736
1737 out:
1738         ocfs2_free_path(path);
1739         return ret;
1740 }
1741
1742 static int ocfs2_insert_refcount_rec(handle_t *handle,
1743                                      struct ocfs2_caching_info *ci,
1744                                      struct buffer_head *ref_root_bh,
1745                                      struct buffer_head *ref_leaf_bh,
1746                                      struct ocfs2_refcount_rec *rec,
1747                                      int index, int merge,
1748                                      struct ocfs2_alloc_context *meta_ac)
1749 {
1750         int ret;
1751         struct ocfs2_refcount_block *rb =
1752                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1753         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1754         struct buffer_head *new_bh = NULL;
1755
1756         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1757
1758         if (rf_list->rl_used == rf_list->rl_count) {
1759                 u64 cpos = le64_to_cpu(rec->r_cpos);
1760                 u32 len = le32_to_cpu(rec->r_clusters);
1761
1762                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1763                                                  ref_leaf_bh, meta_ac);
1764                 if (ret) {
1765                         mlog_errno(ret);
1766                         goto out;
1767                 }
1768
1769                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1770                                              cpos, len, NULL, &index,
1771                                              &new_bh);
1772                 if (ret) {
1773                         mlog_errno(ret);
1774                         goto out;
1775                 }
1776
1777                 ref_leaf_bh = new_bh;
1778                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1779                 rf_list = &rb->rf_records;
1780         }
1781
1782         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1783                                       OCFS2_JOURNAL_ACCESS_WRITE);
1784         if (ret) {
1785                 mlog_errno(ret);
1786                 goto out;
1787         }
1788
1789         if (index < le16_to_cpu(rf_list->rl_used))
1790                 memmove(&rf_list->rl_recs[index + 1],
1791                         &rf_list->rl_recs[index],
1792                         (le16_to_cpu(rf_list->rl_used) - index) *
1793                          sizeof(struct ocfs2_refcount_rec));
1794
1795         trace_ocfs2_insert_refcount_rec(
1796                 (unsigned long long)ref_leaf_bh->b_blocknr, index,
1797                 (unsigned long long)le64_to_cpu(rec->r_cpos),
1798                 le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount));
1799
1800         rf_list->rl_recs[index] = *rec;
1801
1802         le16_add_cpu(&rf_list->rl_used, 1);
1803
1804         if (merge)
1805                 ocfs2_refcount_rec_merge(rb, index);
1806
1807         ocfs2_journal_dirty(handle, ref_leaf_bh);
1808
1809         if (index == 0) {
1810                 ret = ocfs2_adjust_refcount_rec(handle, ci,
1811                                                 ref_root_bh,
1812                                                 ref_leaf_bh, rec);
1813                 if (ret)
1814                         mlog_errno(ret);
1815         }
1816 out:
1817         brelse(new_bh);
1818         return ret;
1819 }
1820
1821 /*
1822  * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1823  * This is much simple than our b-tree code.
1824  * split_rec is the new refcount rec we want to insert.
1825  * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1826  * increase refcount or decrease a refcount to non-zero).
1827  * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1828  * rec( in case we decrease a refcount to zero).
1829  */
1830 static int ocfs2_split_refcount_rec(handle_t *handle,
1831                                     struct ocfs2_caching_info *ci,
1832                                     struct buffer_head *ref_root_bh,
1833                                     struct buffer_head *ref_leaf_bh,
1834                                     struct ocfs2_refcount_rec *split_rec,
1835                                     int index, int merge,
1836                                     struct ocfs2_alloc_context *meta_ac,
1837                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
1838 {
1839         int ret, recs_need;
1840         u32 len;
1841         struct ocfs2_refcount_block *rb =
1842                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1843         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1844         struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1845         struct ocfs2_refcount_rec *tail_rec = NULL;
1846         struct buffer_head *new_bh = NULL;
1847
1848         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1849
1850         trace_ocfs2_split_refcount_rec(le64_to_cpu(orig_rec->r_cpos),
1851                 le32_to_cpu(orig_rec->r_clusters),
1852                 le32_to_cpu(orig_rec->r_refcount),
1853                 le64_to_cpu(split_rec->r_cpos),
1854                 le32_to_cpu(split_rec->r_clusters),
1855                 le32_to_cpu(split_rec->r_refcount));
1856
1857         /*
1858          * If we just need to split the header or tail clusters,
1859          * no more recs are needed, just split is OK.
1860          * Otherwise we at least need one new recs.
1861          */
1862         if (!split_rec->r_refcount &&
1863             (split_rec->r_cpos == orig_rec->r_cpos ||
1864              le64_to_cpu(split_rec->r_cpos) +
1865              le32_to_cpu(split_rec->r_clusters) ==
1866              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1867                 recs_need = 0;
1868         else
1869                 recs_need = 1;
1870
1871         /*
1872          * We need one more rec if we split in the middle and the new rec have
1873          * some refcount in it.
1874          */
1875         if (split_rec->r_refcount &&
1876             (split_rec->r_cpos != orig_rec->r_cpos &&
1877              le64_to_cpu(split_rec->r_cpos) +
1878              le32_to_cpu(split_rec->r_clusters) !=
1879              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1880                 recs_need++;
1881
1882         /* If the leaf block don't have enough record, expand it. */
1883         if (le16_to_cpu(rf_list->rl_used) + recs_need >
1884                                          le16_to_cpu(rf_list->rl_count)) {
1885                 struct ocfs2_refcount_rec tmp_rec;
1886                 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1887                 len = le32_to_cpu(orig_rec->r_clusters);
1888                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1889                                                  ref_leaf_bh, meta_ac);
1890                 if (ret) {
1891                         mlog_errno(ret);
1892                         goto out;
1893                 }
1894
1895                 /*
1896                  * We have to re-get it since now cpos may be moved to
1897                  * another leaf block.
1898                  */
1899                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1900                                              cpos, len, &tmp_rec, &index,
1901                                              &new_bh);
1902                 if (ret) {
1903                         mlog_errno(ret);
1904                         goto out;
1905                 }
1906
1907                 ref_leaf_bh = new_bh;
1908                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1909                 rf_list = &rb->rf_records;
1910                 orig_rec = &rf_list->rl_recs[index];
1911         }
1912
1913         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1914                                       OCFS2_JOURNAL_ACCESS_WRITE);
1915         if (ret) {
1916                 mlog_errno(ret);
1917                 goto out;
1918         }
1919
1920         /*
1921          * We have calculated out how many new records we need and store
1922          * in recs_need, so spare enough space first by moving the records
1923          * after "index" to the end.
1924          */
1925         if (index != le16_to_cpu(rf_list->rl_used) - 1)
1926                 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1927                         &rf_list->rl_recs[index + 1],
1928                         (le16_to_cpu(rf_list->rl_used) - index - 1) *
1929                          sizeof(struct ocfs2_refcount_rec));
1930
1931         len = (le64_to_cpu(orig_rec->r_cpos) +
1932               le32_to_cpu(orig_rec->r_clusters)) -
1933               (le64_to_cpu(split_rec->r_cpos) +
1934               le32_to_cpu(split_rec->r_clusters));
1935
1936         /*
1937          * If we have "len", the we will split in the tail and move it
1938          * to the end of the space we have just spared.
1939          */
1940         if (len) {
1941                 tail_rec = &rf_list->rl_recs[index + recs_need];
1942
1943                 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1944                 le64_add_cpu(&tail_rec->r_cpos,
1945                              le32_to_cpu(tail_rec->r_clusters) - len);
1946                 tail_rec->r_clusters = cpu_to_le32(len);
1947         }
1948
1949         /*
1950          * If the split pos isn't the same as the original one, we need to
1951          * split in the head.
1952          *
1953          * Note: We have the chance that split_rec.r_refcount = 0,
1954          * recs_need = 0 and len > 0, which means we just cut the head from
1955          * the orig_rec and in that case we have done some modification in
1956          * orig_rec above, so the check for r_cpos is faked.
1957          */
1958         if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1959                 len = le64_to_cpu(split_rec->r_cpos) -
1960                       le64_to_cpu(orig_rec->r_cpos);
1961                 orig_rec->r_clusters = cpu_to_le32(len);
1962                 index++;
1963         }
1964
1965         le16_add_cpu(&rf_list->rl_used, recs_need);
1966
1967         if (split_rec->r_refcount) {
1968                 rf_list->rl_recs[index] = *split_rec;
1969                 trace_ocfs2_split_refcount_rec_insert(
1970                         (unsigned long long)ref_leaf_bh->b_blocknr, index,
1971                         (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1972                         le32_to_cpu(split_rec->r_clusters),
1973                         le32_to_cpu(split_rec->r_refcount));
1974
1975                 if (merge)
1976                         ocfs2_refcount_rec_merge(rb, index);
1977         }
1978
1979         ocfs2_journal_dirty(handle, ref_leaf_bh);
1980
1981 out:
1982         brelse(new_bh);
1983         return ret;
1984 }
1985
1986 static int __ocfs2_increase_refcount(handle_t *handle,
1987                                      struct ocfs2_caching_info *ci,
1988                                      struct buffer_head *ref_root_bh,
1989                                      u64 cpos, u32 len, int merge,
1990                                      struct ocfs2_alloc_context *meta_ac,
1991                                      struct ocfs2_cached_dealloc_ctxt *dealloc)
1992 {
1993         int ret = 0, index;
1994         struct buffer_head *ref_leaf_bh = NULL;
1995         struct ocfs2_refcount_rec rec;
1996         unsigned int set_len = 0;
1997
1998         trace_ocfs2_increase_refcount_begin(
1999              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2000              (unsigned long long)cpos, len);
2001
2002         while (len) {
2003                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2004                                              cpos, len, &rec, &index,
2005                                              &ref_leaf_bh);
2006                 if (ret) {
2007                         mlog_errno(ret);
2008                         goto out;
2009                 }
2010
2011                 set_len = le32_to_cpu(rec.r_clusters);
2012
2013                 /*
2014                  * Here we may meet with 3 situations:
2015                  *
2016                  * 1. If we find an already existing record, and the length
2017                  *    is the same, cool, we just need to increase the r_refcount
2018                  *    and it is OK.
2019                  * 2. If we find a hole, just insert it with r_refcount = 1.
2020                  * 3. If we are in the middle of one extent record, split
2021                  *    it.
2022                  */
2023                 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
2024                     set_len <= len) {
2025                         trace_ocfs2_increase_refcount_change(
2026                                 (unsigned long long)cpos, set_len,
2027                                 le32_to_cpu(rec.r_refcount));
2028                         ret = ocfs2_change_refcount_rec(handle, ci,
2029                                                         ref_leaf_bh, index,
2030                                                         merge, 1);
2031                         if (ret) {
2032                                 mlog_errno(ret);
2033                                 goto out;
2034                         }
2035                 } else if (!rec.r_refcount) {
2036                         rec.r_refcount = cpu_to_le32(1);
2037
2038                         trace_ocfs2_increase_refcount_insert(
2039                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2040                              set_len);
2041                         ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
2042                                                         ref_leaf_bh,
2043                                                         &rec, index,
2044                                                         merge, meta_ac);
2045                         if (ret) {
2046                                 mlog_errno(ret);
2047                                 goto out;
2048                         }
2049                 } else  {
2050                         set_len = min((u64)(cpos + len),
2051                                       le64_to_cpu(rec.r_cpos) + set_len) - cpos;
2052                         rec.r_cpos = cpu_to_le64(cpos);
2053                         rec.r_clusters = cpu_to_le32(set_len);
2054                         le32_add_cpu(&rec.r_refcount, 1);
2055
2056                         trace_ocfs2_increase_refcount_split(
2057                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2058                              set_len, le32_to_cpu(rec.r_refcount));
2059                         ret = ocfs2_split_refcount_rec(handle, ci,
2060                                                        ref_root_bh, ref_leaf_bh,
2061                                                        &rec, index, merge,
2062                                                        meta_ac, dealloc);
2063                         if (ret) {
2064                                 mlog_errno(ret);
2065                                 goto out;
2066                         }
2067                 }
2068
2069                 cpos += set_len;
2070                 len -= set_len;
2071                 brelse(ref_leaf_bh);
2072                 ref_leaf_bh = NULL;
2073         }
2074
2075 out:
2076         brelse(ref_leaf_bh);
2077         return ret;
2078 }
2079
2080 static int ocfs2_remove_refcount_extent(handle_t *handle,
2081                                 struct ocfs2_caching_info *ci,
2082                                 struct buffer_head *ref_root_bh,
2083                                 struct buffer_head *ref_leaf_bh,
2084                                 struct ocfs2_alloc_context *meta_ac,
2085                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2086 {
2087         int ret;
2088         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2089         struct ocfs2_refcount_block *rb =
2090                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2091         struct ocfs2_extent_tree et;
2092
2093         BUG_ON(rb->rf_records.rl_used);
2094
2095         trace_ocfs2_remove_refcount_extent(
2096                 (unsigned long long)ocfs2_metadata_cache_owner(ci),
2097                 (unsigned long long)ref_leaf_bh->b_blocknr,
2098                 le32_to_cpu(rb->rf_cpos));
2099
2100         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2101         ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
2102                                   1, meta_ac, dealloc);
2103         if (ret) {
2104                 mlog_errno(ret);
2105                 goto out;
2106         }
2107
2108         ocfs2_remove_from_cache(ci, ref_leaf_bh);
2109
2110         /*
2111          * add the freed block to the dealloc so that it will be freed
2112          * when we run dealloc.
2113          */
2114         ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
2115                                         le16_to_cpu(rb->rf_suballoc_slot),
2116                                         le64_to_cpu(rb->rf_suballoc_loc),
2117                                         le64_to_cpu(rb->rf_blkno),
2118                                         le16_to_cpu(rb->rf_suballoc_bit));
2119         if (ret) {
2120                 mlog_errno(ret);
2121                 goto out;
2122         }
2123
2124         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
2125                                       OCFS2_JOURNAL_ACCESS_WRITE);
2126         if (ret) {
2127                 mlog_errno(ret);
2128                 goto out;
2129         }
2130
2131         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2132
2133         le32_add_cpu(&rb->rf_clusters, -1);
2134
2135         /*
2136          * check whether we need to restore the root refcount block if
2137          * there is no leaf extent block at atll.
2138          */
2139         if (!rb->rf_list.l_next_free_rec) {
2140                 BUG_ON(rb->rf_clusters);
2141
2142                 trace_ocfs2_restore_refcount_block(
2143                      (unsigned long long)ref_root_bh->b_blocknr);
2144
2145                 rb->rf_flags = 0;
2146                 rb->rf_parent = 0;
2147                 rb->rf_cpos = 0;
2148                 memset(&rb->rf_records, 0, sb->s_blocksize -
2149                        offsetof(struct ocfs2_refcount_block, rf_records));
2150                 rb->rf_records.rl_count =
2151                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
2152         }
2153
2154         ocfs2_journal_dirty(handle, ref_root_bh);
2155
2156 out:
2157         return ret;
2158 }
2159
2160 int ocfs2_increase_refcount(handle_t *handle,
2161                             struct ocfs2_caching_info *ci,
2162                             struct buffer_head *ref_root_bh,
2163                             u64 cpos, u32 len,
2164                             struct ocfs2_alloc_context *meta_ac,
2165                             struct ocfs2_cached_dealloc_ctxt *dealloc)
2166 {
2167         return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
2168                                          cpos, len, 1,
2169                                          meta_ac, dealloc);
2170 }
2171
2172 static int ocfs2_decrease_refcount_rec(handle_t *handle,
2173                                 struct ocfs2_caching_info *ci,
2174                                 struct buffer_head *ref_root_bh,
2175                                 struct buffer_head *ref_leaf_bh,
2176                                 int index, u64 cpos, unsigned int len,
2177                                 struct ocfs2_alloc_context *meta_ac,
2178                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2179 {
2180         int ret;
2181         struct ocfs2_refcount_block *rb =
2182                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2183         struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2184
2185         BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2186         BUG_ON(cpos + len >
2187                le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2188
2189         trace_ocfs2_decrease_refcount_rec(
2190                 (unsigned long long)ocfs2_metadata_cache_owner(ci),
2191                 (unsigned long long)cpos, len);
2192
2193         if (cpos == le64_to_cpu(rec->r_cpos) &&
2194             len == le32_to_cpu(rec->r_clusters))
2195                 ret = ocfs2_change_refcount_rec(handle, ci,
2196                                                 ref_leaf_bh, index, 1, -1);
2197         else {
2198                 struct ocfs2_refcount_rec split = *rec;
2199                 split.r_cpos = cpu_to_le64(cpos);
2200                 split.r_clusters = cpu_to_le32(len);
2201
2202                 le32_add_cpu(&split.r_refcount, -1);
2203
2204                 ret = ocfs2_split_refcount_rec(handle, ci,
2205                                                ref_root_bh, ref_leaf_bh,
2206                                                &split, index, 1,
2207                                                meta_ac, dealloc);
2208         }
2209
2210         if (ret) {
2211                 mlog_errno(ret);
2212                 goto out;
2213         }
2214
2215         /* Remove the leaf refcount block if it contains no refcount record. */
2216         if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2217                 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2218                                                    ref_leaf_bh, meta_ac,
2219                                                    dealloc);
2220                 if (ret)
2221                         mlog_errno(ret);
2222         }
2223
2224 out:
2225         return ret;
2226 }
2227
2228 static int __ocfs2_decrease_refcount(handle_t *handle,
2229                                      struct ocfs2_caching_info *ci,
2230                                      struct buffer_head *ref_root_bh,
2231                                      u64 cpos, u32 len,
2232                                      struct ocfs2_alloc_context *meta_ac,
2233                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2234                                      int delete)
2235 {
2236         int ret = 0, index = 0;
2237         struct ocfs2_refcount_rec rec;
2238         unsigned int r_count = 0, r_len;
2239         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2240         struct buffer_head *ref_leaf_bh = NULL;
2241
2242         trace_ocfs2_decrease_refcount(
2243                 (unsigned long long)ocfs2_metadata_cache_owner(ci),
2244                 (unsigned long long)cpos, len, delete);
2245
2246         while (len) {
2247                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2248                                              cpos, len, &rec, &index,
2249                                              &ref_leaf_bh);
2250                 if (ret) {
2251                         mlog_errno(ret);
2252                         goto out;
2253                 }
2254
2255                 r_count = le32_to_cpu(rec.r_refcount);
2256                 BUG_ON(r_count == 0);
2257                 if (!delete)
2258                         BUG_ON(r_count > 1);
2259
2260                 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2261                               le32_to_cpu(rec.r_clusters)) - cpos;
2262
2263                 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2264                                                   ref_leaf_bh, index,
2265                                                   cpos, r_len,
2266                                                   meta_ac, dealloc);
2267                 if (ret) {
2268                         mlog_errno(ret);
2269                         goto out;
2270                 }
2271
2272                 if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
2273                         ret = ocfs2_cache_cluster_dealloc(dealloc,
2274                                           ocfs2_clusters_to_blocks(sb, cpos),
2275                                                           r_len);
2276                         if (ret) {
2277                                 mlog_errno(ret);
2278                                 goto out;
2279                         }
2280                 }
2281
2282                 cpos += r_len;
2283                 len -= r_len;
2284                 brelse(ref_leaf_bh);
2285                 ref_leaf_bh = NULL;
2286         }
2287
2288 out:
2289         brelse(ref_leaf_bh);
2290         return ret;
2291 }
2292
2293 /* Caller must hold refcount tree lock. */
2294 int ocfs2_decrease_refcount(struct inode *inode,
2295                             handle_t *handle, u32 cpos, u32 len,
2296                             struct ocfs2_alloc_context *meta_ac,
2297                             struct ocfs2_cached_dealloc_ctxt *dealloc,
2298                             int delete)
2299 {
2300         int ret;
2301         u64 ref_blkno;
2302         struct buffer_head *ref_root_bh = NULL;
2303         struct ocfs2_refcount_tree *tree;
2304
2305         BUG_ON(!ocfs2_is_refcount_inode(inode));
2306
2307         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2308         if (ret) {
2309                 mlog_errno(ret);
2310                 goto out;
2311         }
2312
2313         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2314         if (ret) {
2315                 mlog_errno(ret);
2316                 goto out;
2317         }
2318
2319         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2320                                         &ref_root_bh);
2321         if (ret) {
2322                 mlog_errno(ret);
2323                 goto out;
2324         }
2325
2326         ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2327                                         cpos, len, meta_ac, dealloc, delete);
2328         if (ret)
2329                 mlog_errno(ret);
2330 out:
2331         brelse(ref_root_bh);
2332         return ret;
2333 }
2334
2335 /*
2336  * Mark the already-existing extent at cpos as refcounted for len clusters.
2337  * This adds the refcount extent flag.
2338  *
2339  * If the existing extent is larger than the request, initiate a
2340  * split. An attempt will be made at merging with adjacent extents.
2341  *
2342  * The caller is responsible for passing down meta_ac if we'll need it.
2343  */
2344 static int ocfs2_mark_extent_refcounted(struct inode *inode,
2345                                 struct ocfs2_extent_tree *et,
2346                                 handle_t *handle, u32 cpos,
2347                                 u32 len, u32 phys,
2348                                 struct ocfs2_alloc_context *meta_ac,
2349                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2350 {
2351         int ret;
2352
2353         trace_ocfs2_mark_extent_refcounted(OCFS2_I(inode)->ip_blkno,
2354                                            cpos, len, phys);
2355
2356         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2357                 ret = ocfs2_error(inode->i_sb, "Inode %lu want to use refcount tree, but the feature bit is not set in the super block\n",
2358                                   inode->i_ino);
2359                 goto out;
2360         }
2361
2362         ret = ocfs2_change_extent_flag(handle, et, cpos,
2363                                        len, phys, meta_ac, dealloc,
2364                                        OCFS2_EXT_REFCOUNTED, 0);
2365         if (ret)
2366                 mlog_errno(ret);
2367
2368 out:
2369         return ret;
2370 }
2371
2372 /*
2373  * Given some contiguous physical clusters, calculate what we need
2374  * for modifying their refcount.
2375  */
2376 static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2377                                             struct ocfs2_caching_info *ci,
2378                                             struct buffer_head *ref_root_bh,
2379                                             u64 start_cpos,
2380                                             u32 clusters,
2381                                             int *meta_add,
2382                                             int *credits)
2383 {
2384         int ret = 0, index, ref_blocks = 0, recs_add = 0;
2385         u64 cpos = start_cpos;
2386         struct ocfs2_refcount_block *rb;
2387         struct ocfs2_refcount_rec rec;
2388         struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2389         u32 len;
2390
2391         while (clusters) {
2392                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2393                                              cpos, clusters, &rec,
2394                                              &index, &ref_leaf_bh);
2395                 if (ret) {
2396                         mlog_errno(ret);
2397                         goto out;
2398                 }
2399
2400                 if (ref_leaf_bh != prev_bh) {
2401                         /*
2402                          * Now we encounter a new leaf block, so calculate
2403                          * whether we need to extend the old leaf.
2404                          */
2405                         if (prev_bh) {
2406                                 rb = (struct ocfs2_refcount_block *)
2407                                                         prev_bh->b_data;
2408
2409                                 if (le16_to_cpu(rb->rf_records.rl_used) +
2410                                     recs_add >
2411                                     le16_to_cpu(rb->rf_records.rl_count))
2412                                         ref_blocks++;
2413                         }
2414
2415                         recs_add = 0;
2416                         *credits += 1;
2417                         brelse(prev_bh);
2418                         prev_bh = ref_leaf_bh;
2419                         get_bh(prev_bh);
2420                 }
2421
2422                 trace_ocfs2_calc_refcount_meta_credits_iterate(
2423                                 recs_add, (unsigned long long)cpos, clusters,
2424                                 (unsigned long long)le64_to_cpu(rec.r_cpos),
2425                                 le32_to_cpu(rec.r_clusters),
2426                                 le32_to_cpu(rec.r_refcount), index);
2427
2428                 len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2429                           le32_to_cpu(rec.r_clusters)) - cpos;
2430                 /*
2431                  * We record all the records which will be inserted to the
2432                  * same refcount block, so that we can tell exactly whether
2433                  * we need a new refcount block or not.
2434                  *
2435                  * If we will insert a new one, this is easy and only happens
2436                  * during adding refcounted flag to the extent, so we don't
2437                  * have a chance of spliting. We just need one record.
2438                  *
2439                  * If the refcount rec already exists, that would be a little
2440                  * complicated. we may have to:
2441                  * 1) split at the beginning if the start pos isn't aligned.
2442                  *    we need 1 more record in this case.
2443                  * 2) split int the end if the end pos isn't aligned.
2444                  *    we need 1 more record in this case.
2445                  * 3) split in the middle because of file system fragmentation.
2446                  *    we need 2 more records in this case(we can't detect this
2447                  *    beforehand, so always think of the worst case).
2448                  */
2449                 if (rec.r_refcount) {
2450                         recs_add += 2;
2451                         /* Check whether we need a split at the beginning. */
2452                         if (cpos == start_cpos &&
2453                             cpos != le64_to_cpu(rec.r_cpos))
2454                                 recs_add++;
2455
2456                         /* Check whether we need a split in the end. */
2457                         if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2458                             le32_to_cpu(rec.r_clusters))
2459                                 recs_add++;
2460                 } else
2461                         recs_add++;
2462
2463                 brelse(ref_leaf_bh);
2464                 ref_leaf_bh = NULL;
2465                 clusters -= len;
2466                 cpos += len;
2467         }
2468
2469         if (prev_bh) {
2470                 rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2471
2472                 if (le16_to_cpu(rb->rf_records.rl_used) + recs_add >
2473                     le16_to_cpu(rb->rf_records.rl_count))
2474                         ref_blocks++;
2475
2476                 *credits += 1;
2477         }
2478
2479         if (!ref_blocks)
2480                 goto out;
2481
2482         *meta_add += ref_blocks;
2483         *credits += ref_blocks;
2484
2485         /*
2486          * So we may need ref_blocks to insert into the tree.
2487          * That also means we need to change the b-tree and add that number
2488          * of records since we never merge them.
2489          * We need one more block for expansion since the new created leaf
2490          * block is also full and needs split.
2491          */
2492         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2493         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2494                 struct ocfs2_extent_tree et;
2495
2496                 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2497                 *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2498                 *credits += ocfs2_calc_extend_credits(sb,
2499                                                       et.et_root_el);
2500         } else {
2501                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2502                 *meta_add += 1;
2503         }
2504
2505 out:
2506
2507         trace_ocfs2_calc_refcount_meta_credits(
2508                 (unsigned long long)start_cpos, clusters,
2509                 *meta_add, *credits);
2510         brelse(ref_leaf_bh);
2511         brelse(prev_bh);
2512         return ret;
2513 }
2514
2515 /*
2516  * For refcount tree, we will decrease some contiguous clusters
2517  * refcount count, so just go through it to see how many blocks
2518  * we gonna touch and whether we need to create new blocks.
2519  *
2520  * Normally the refcount blocks store these refcount should be
2521  * contiguous also, so that we can get the number easily.
2522  * We will at most add split 2 refcount records and 2 more
2523  * refcount blocks, so just check it in a rough way.
2524  *
2525  * Caller must hold refcount tree lock.
2526  */
2527 int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
2528                                           u64 refcount_loc,
2529                                           u64 phys_blkno,
2530                                           u32 clusters,
2531                                           int *credits,
2532                                           int *ref_blocks)
2533 {
2534         int ret;
2535         struct buffer_head *ref_root_bh = NULL;
2536         struct ocfs2_refcount_tree *tree;
2537         u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2538
2539         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2540                 ret = ocfs2_error(inode->i_sb, "Inode %lu want to use refcount tree, but the feature bit is not set in the super block\n",
2541                                   inode->i_ino);
2542                 goto out;
2543         }
2544
2545         BUG_ON(!ocfs2_is_refcount_inode(inode));
2546
2547         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
2548                                       refcount_loc, &tree);
2549         if (ret) {
2550                 mlog_errno(ret);
2551                 goto out;
2552         }
2553
2554         ret = ocfs2_read_refcount_block(&tree->rf_ci, refcount_loc,
2555                                         &ref_root_bh);
2556         if (ret) {
2557                 mlog_errno(ret);
2558                 goto out;
2559         }
2560
2561         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2562                                                &tree->rf_ci,
2563                                                ref_root_bh,
2564                                                start_cpos, clusters,
2565                                                ref_blocks, credits);
2566         if (ret) {
2567                 mlog_errno(ret);
2568                 goto out;
2569         }
2570
2571         trace_ocfs2_prepare_refcount_change_for_del(*ref_blocks, *credits);
2572
2573 out:
2574         brelse(ref_root_bh);
2575         return ret;
2576 }
2577
2578 #define MAX_CONTIG_BYTES        1048576
2579
2580 static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb)
2581 {
2582         return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES);
2583 }
2584
2585 static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb)
2586 {
2587         return ~(ocfs2_cow_contig_clusters(sb) - 1);
2588 }
2589
2590 /*
2591  * Given an extent that starts at 'start' and an I/O that starts at 'cpos',
2592  * find an offset (start + (n * contig_clusters)) that is closest to cpos
2593  * while still being less than or equal to it.
2594  *
2595  * The goal is to break the extent at a multiple of contig_clusters.
2596  */
2597 static inline unsigned int ocfs2_cow_align_start(struct super_block *sb,
2598                                                  unsigned int start,
2599                                                  unsigned int cpos)
2600 {
2601         BUG_ON(start > cpos);
2602
2603         return start + ((cpos - start) & ocfs2_cow_contig_mask(sb));
2604 }
2605
2606 /*
2607  * Given a cluster count of len, pad it out so that it is a multiple
2608  * of contig_clusters.
2609  */
2610 static inline unsigned int ocfs2_cow_align_length(struct super_block *sb,
2611                                                   unsigned int len)
2612 {
2613         unsigned int padded =
2614                 (len + (ocfs2_cow_contig_clusters(sb) - 1)) &
2615                 ocfs2_cow_contig_mask(sb);
2616
2617         /* Did we wrap? */
2618         if (padded < len)
2619                 padded = UINT_MAX;
2620
2621         return padded;
2622 }
2623
2624 /*
2625  * Calculate out the start and number of virtual clusters we need to to CoW.
2626  *
2627  * cpos is vitual start cluster position we want to do CoW in a
2628  * file and write_len is the cluster length.
2629  * max_cpos is the place where we want to stop CoW intentionally.
2630  *
2631  * Normal we will start CoW from the beginning of extent record cotaining cpos.
2632  * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we
2633  * get good I/O from the resulting extent tree.
2634  */
2635 static int ocfs2_refcount_cal_cow_clusters(struct inode *inode,
2636                                            struct ocfs2_extent_list *el,
2637                                            u32 cpos,
2638                                            u32 write_len,
2639                                            u32 max_cpos,
2640                                            u32 *cow_start,
2641                                            u32 *cow_len)
2642 {
2643         int ret = 0;
2644         int tree_height = le16_to_cpu(el->l_tree_depth), i;
2645         struct buffer_head *eb_bh = NULL;
2646         struct ocfs2_extent_block *eb = NULL;
2647         struct ocfs2_extent_rec *rec;
2648         unsigned int want_clusters, rec_end = 0;
2649         int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb);
2650         int leaf_clusters;
2651
2652         BUG_ON(cpos + write_len > max_cpos);
2653
2654         if (tree_height > 0) {
2655                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh);
2656                 if (ret) {
2657                         mlog_errno(ret);
2658                         goto out;
2659                 }
2660
2661                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2662                 el = &eb->h_list;
2663
2664                 if (el->l_tree_depth) {
2665                         ret = ocfs2_error(inode->i_sb,
2666                                           "Inode %lu has non zero tree depth in leaf block %llu\n",
2667                                           inode->i_ino,
2668                                           (unsigned long long)eb_bh->b_blocknr);
2669                         goto out;
2670                 }
2671         }
2672
2673         *cow_len = 0;
2674         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
2675                 rec = &el->l_recs[i];
2676
2677                 if (ocfs2_is_empty_extent(rec)) {
2678                         mlog_bug_on_msg(i != 0, "Inode %lu has empty record in "
2679                                         "index %d\n", inode->i_ino, i);
2680                         continue;
2681                 }
2682
2683                 if (le32_to_cpu(rec->e_cpos) +
2684                     le16_to_cpu(rec->e_leaf_clusters) <= cpos)
2685                         continue;
2686
2687                 if (*cow_len == 0) {
2688                         /*
2689                          * We should find a refcounted record in the
2690                          * first pass.
2691                          */
2692                         BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED));
2693                         *cow_start = le32_to_cpu(rec->e_cpos);
2694                 }
2695
2696                 /*
2697                  * If we encounter a hole, a non-refcounted record or
2698                  * pass the max_cpos, stop the search.
2699                  */
2700                 if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) ||
2701                     (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) ||
2702                     (max_cpos <= le32_to_cpu(rec->e_cpos)))
2703                         break;
2704
2705                 leaf_clusters = le16_to_cpu(rec->e_leaf_clusters);
2706                 rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters;
2707                 if (rec_end > max_cpos) {
2708                         rec_end = max_cpos;
2709                         leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos);
2710                 }
2711
2712                 /*
2713                  * How many clusters do we actually need from
2714                  * this extent?  First we see how many we actually
2715                  * need to complete the write.  If that's smaller
2716                  * than contig_clusters, we try for contig_clusters.
2717                  */
2718                 if (!*cow_len)
2719                         want_clusters = write_len;
2720                 else
2721                         want_clusters = (cpos + write_len) -
2722                                 (*cow_start + *cow_len);
2723                 if (want_clusters < contig_clusters)
2724                         want_clusters = contig_clusters;
2725
2726                 /*
2727                  * If the write does not cover the whole extent, we
2728                  * need to calculate how we're going to split the extent.
2729                  * We try to do it on contig_clusters boundaries.
2730                  *
2731                  * Any extent smaller than contig_clusters will be
2732                  * CoWed in its entirety.
2733                  */
2734                 if (leaf_clusters <= contig_clusters)
2735                         *cow_len += leaf_clusters;
2736                 else if (*cow_len || (*cow_start == cpos)) {
2737                         /*
2738                          * This extent needs to be CoW'd from its
2739                          * beginning, so all we have to do is compute
2740                          * how many clusters to grab.  We align
2741                          * want_clusters to the edge of contig_clusters
2742                          * to get better I/O.
2743                          */
2744                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2745                                                                want_clusters);
2746
2747                         if (leaf_clusters < want_clusters)
2748                                 *cow_len += leaf_clusters;
2749                         else
2750                                 *cow_len += want_clusters;
2751                 } else if ((*cow_start + contig_clusters) >=
2752                            (cpos + write_len)) {
2753                         /*
2754                          * Breaking off contig_clusters at the front
2755                          * of the extent will cover our write.  That's
2756                          * easy.
2757                          */
2758                         *cow_len = contig_clusters;
2759                 } else if ((rec_end - cpos) <= contig_clusters) {
2760                         /*
2761                          * Breaking off contig_clusters at the tail of
2762                          * this extent will cover cpos.
2763                          */
2764                         *cow_start = rec_end - contig_clusters;
2765                         *cow_len = contig_clusters;
2766                 } else if ((rec_end - cpos) <= want_clusters) {
2767                         /*
2768                          * While we can't fit the entire write in this
2769                          * extent, we know that the write goes from cpos
2770                          * to the end of the extent.  Break that off.
2771                          * We try to break it at some multiple of
2772                          * contig_clusters from the front of the extent.
2773                          * Failing that (ie, cpos is within
2774                          * contig_clusters of the front), we'll CoW the
2775                          * entire extent.
2776                          */
2777                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2778                                                            *cow_start, cpos);
2779                         *cow_len = rec_end - *cow_start;
2780                 } else {
2781                         /*
2782                          * Ok, the entire write lives in the middle of
2783                          * this extent.  Let's try to slice the extent up
2784                          * nicely.  Optimally, our CoW region starts at
2785                          * m*contig_clusters from the beginning of the
2786                          * extent and goes for n*contig_clusters,
2787                          * covering the entire write.
2788                          */
2789                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2790                                                            *cow_start, cpos);
2791
2792                         want_clusters = (cpos + write_len) - *cow_start;
2793                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2794                                                                want_clusters);
2795                         if (*cow_start + want_clusters <= rec_end)
2796                                 *cow_len = want_clusters;
2797                         else
2798                                 *cow_len = rec_end - *cow_start;
2799                 }
2800
2801                 /* Have we covered our entire write yet? */
2802                 if ((*cow_start + *cow_len) >= (cpos + write_len))
2803                         break;
2804
2805                 /*
2806                  * If we reach the end of the extent block and don't get enough
2807                  * clusters, continue with the next extent block if possible.
2808                  */
2809                 if (i + 1 == le16_to_cpu(el->l_next_free_rec) &&
2810                     eb && eb->h_next_leaf_blk) {
2811                         brelse(eb_bh);
2812                         eb_bh = NULL;
2813
2814                         ret = ocfs2_read_extent_block(INODE_CACHE(inode),
2815                                                le64_to_cpu(eb->h_next_leaf_blk),
2816                                                &eb_bh);
2817                         if (ret) {
2818                                 mlog_errno(ret);
2819                                 goto out;
2820                         }
2821
2822                         eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2823                         el = &eb->h_list;
2824                         i = -1;
2825                 }
2826         }
2827
2828 out:
2829         brelse(eb_bh);
2830         return ret;
2831 }
2832
2833 /*
2834  * Prepare meta_ac, data_ac and calculate credits when we want to add some
2835  * num_clusters in data_tree "et" and change the refcount for the old
2836  * clusters(starting form p_cluster) in the refcount tree.
2837  *
2838  * Note:
2839  * 1. since we may split the old tree, so we at most will need num_clusters + 2
2840  *    more new leaf records.
2841  * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so
2842  *    just give data_ac = NULL.
2843  */
2844 static int ocfs2_lock_refcount_allocators(struct super_block *sb,
2845                                         u32 p_cluster, u32 num_clusters,
2846                                         struct ocfs2_extent_tree *et,
2847                                         struct ocfs2_caching_info *ref_ci,
2848                                         struct buffer_head *ref_root_bh,
2849                                         struct ocfs2_alloc_context **meta_ac,
2850                                         struct ocfs2_alloc_context **data_ac,
2851                                         int *credits)
2852 {
2853         int ret = 0, meta_add = 0;
2854         int num_free_extents = ocfs2_num_free_extents(et);
2855
2856         if (num_free_extents < 0) {
2857                 ret = num_free_extents;
2858                 mlog_errno(ret);
2859                 goto out;
2860         }
2861
2862         if (num_free_extents < num_clusters + 2)
2863                 meta_add =
2864                         ocfs2_extend_meta_needed(et->et_root_el);
2865
2866         *credits += ocfs2_calc_extend_credits(sb, et->et_root_el);
2867
2868         ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh,
2869                                                p_cluster, num_clusters,
2870                                                &meta_add, credits);
2871         if (ret) {
2872                 mlog_errno(ret);
2873                 goto out;
2874         }
2875
2876         trace_ocfs2_lock_refcount_allocators(meta_add, *credits);
2877         ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add,
2878                                                 meta_ac);
2879         if (ret) {
2880                 mlog_errno(ret);
2881                 goto out;
2882         }
2883
2884         if (data_ac) {
2885                 ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters,
2886                                              data_ac);
2887                 if (ret)
2888                         mlog_errno(ret);
2889         }
2890
2891 out:
2892         if (ret) {
2893                 if (*meta_ac) {
2894                         ocfs2_free_alloc_context(*meta_ac);
2895                         *meta_ac = NULL;
2896                 }
2897         }
2898
2899         return ret;
2900 }
2901
2902 static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh)
2903 {
2904         BUG_ON(buffer_dirty(bh));
2905
2906         clear_buffer_mapped(bh);
2907
2908         return 0;
2909 }
2910
2911 int ocfs2_duplicate_clusters_by_page(handle_t *handle,
2912                                      struct inode *inode,
2913                                      u32 cpos, u32 old_cluster,
2914                                      u32 new_cluster, u32 new_len)
2915 {
2916         int ret = 0, partial;
2917         struct super_block *sb = inode->i_sb;
2918         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2919         struct page *page;
2920         pgoff_t page_index;
2921         unsigned int from, to;
2922         loff_t offset, end, map_end;
2923         struct address_space *mapping = inode->i_mapping;
2924
2925         trace_ocfs2_duplicate_clusters_by_page(cpos, old_cluster,
2926                                                new_cluster, new_len);
2927
2928         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
2929         end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2930         /*
2931          * We only duplicate pages until we reach the page contains i_size - 1.
2932          * So trim 'end' to i_size.
2933          */
2934         if (end > i_size_read(inode))
2935                 end = i_size_read(inode);
2936
2937         while (offset < end) {
2938                 page_index = offset >> PAGE_SHIFT;
2939                 map_end = ((loff_t)page_index + 1) << PAGE_SHIFT;
2940                 if (map_end > end)
2941                         map_end = end;
2942
2943                 /* from, to is the offset within the page. */
2944                 from = offset & (PAGE_SIZE - 1);
2945                 to = PAGE_SIZE;
2946                 if (map_end & (PAGE_SIZE - 1))
2947                         to = map_end & (PAGE_SIZE - 1);
2948
2949 retry:
2950                 page = find_or_create_page(mapping, page_index, GFP_NOFS);
2951                 if (!page) {
2952                         ret = -ENOMEM;
2953                         mlog_errno(ret);
2954                         break;
2955                 }
2956
2957                 /*
2958                  * In case PAGE_SIZE <= CLUSTER_SIZE, we do not expect a dirty
2959                  * page, so write it back.
2960                  */
2961                 if (PAGE_SIZE <= OCFS2_SB(sb)->s_clustersize) {
2962                         if (PageDirty(page)) {
2963                                 /*
2964                                  * write_on_page will unlock the page on return
2965                                  */
2966                                 ret = write_one_page(page);
2967                                 goto retry;
2968                         }
2969                 }
2970
2971                 if (!PageUptodate(page)) {
2972                         ret = block_read_full_page(page, ocfs2_get_block);
2973                         if (ret) {
2974                                 mlog_errno(ret);
2975                                 goto unlock;
2976                         }
2977                         lock_page(page);
2978                 }
2979
2980                 if (page_has_buffers(page)) {
2981                         ret = walk_page_buffers(handle, page_buffers(page),
2982                                                 from, to, &partial,
2983                                                 ocfs2_clear_cow_buffer);
2984                         if (ret) {
2985                                 mlog_errno(ret);
2986                                 goto unlock;
2987                         }
2988                 }
2989
2990                 ocfs2_map_and_dirty_page(inode,
2991                                          handle, from, to,
2992                                          page, 0, &new_block);
2993                 mark_page_accessed(page);
2994 unlock:
2995                 unlock_page(page);
2996                 put_page(page);
2997                 page = NULL;
2998                 offset = map_end;
2999                 if (ret)
3000                         break;
3001         }
3002
3003         return ret;
3004 }
3005
3006 int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
3007                                     struct inode *inode,
3008                                     u32 cpos, u32 old_cluster,
3009                                     u32 new_cluster, u32 new_len)
3010 {
3011         int ret = 0;
3012         struct super_block *sb = inode->i_sb;
3013         struct ocfs2_caching_info *ci = INODE_CACHE(inode);
3014         int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
3015         u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
3016         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
3017         struct ocfs2_super *osb = OCFS2_SB(sb);
3018         struct buffer_head *old_bh = NULL;
3019         struct buffer_head *new_bh = NULL;
3020
3021         trace_ocfs2_duplicate_clusters_by_page(cpos, old_cluster,
3022                                                new_cluster, new_len);
3023
3024         for (i = 0; i < blocks; i++, old_block++, new_block++) {
3025                 new_bh = sb_getblk(osb->sb, new_block);
3026                 if (new_bh == NULL) {
3027                         ret = -ENOMEM;
3028                         mlog_errno(ret);
3029                         break;
3030                 }
3031
3032                 ocfs2_set_new_buffer_uptodate(ci, new_bh);
3033
3034                 ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
3035                 if (ret) {
3036                         mlog_errno(ret);
3037                         break;
3038                 }
3039
3040                 ret = ocfs2_journal_access(handle, ci, new_bh,
3041                                            OCFS2_JOURNAL_ACCESS_CREATE);
3042                 if (ret) {
3043                         mlog_errno(ret);
3044                         break;
3045                 }
3046
3047                 memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
3048                 ocfs2_journal_dirty(handle, new_bh);
3049
3050                 brelse(new_bh);
3051                 brelse(old_bh);
3052                 new_bh = NULL;
3053                 old_bh = NULL;
3054         }
3055
3056         brelse(new_bh);
3057         brelse(old_bh);
3058         return ret;
3059 }
3060
3061 static int ocfs2_clear_ext_refcount(handle_t *handle,
3062                                     struct ocfs2_extent_tree *et,
3063                                     u32 cpos, u32 p_cluster, u32 len,
3064                                     unsigned int ext_flags,
3065                                     struct ocfs2_alloc_context *meta_ac,
3066                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
3067 {
3068         int ret, index;
3069         struct ocfs2_extent_rec replace_rec;
3070         struct ocfs2_path *path = NULL;
3071         struct ocfs2_extent_list *el;
3072         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
3073         u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
3074
3075         trace_ocfs2_clear_ext_refcount((unsigned long long)ino,
3076                                        cpos, len, p_cluster, ext_flags);
3077
3078         memset(&replace_rec, 0, sizeof(replace_rec));
3079         replace_rec.e_cpos = cpu_to_le32(cpos);
3080         replace_rec.e_leaf_clusters = cpu_to_le16(len);
3081         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
3082                                                                    p_cluster));
3083         replace_rec.e_flags = ext_flags;
3084         replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
3085
3086         path = ocfs2_new_path_from_et(et);
3087         if (!path) {
3088                 ret = -ENOMEM;
3089                 mlog_errno(ret);
3090                 goto out;
3091         }
3092
3093         ret = ocfs2_find_path(et->et_ci, path, cpos);
3094         if (ret) {
3095                 mlog_errno(ret);
3096                 goto out;
3097         }
3098
3099         el = path_leaf_el(path);
3100
3101         index = ocfs2_search_extent_list(el, cpos);
3102         if (index == -1) {
3103                 ret = ocfs2_error(sb,
3104                                   "Inode %llu has an extent at cpos %u which can no longer be found\n",
3105                                   (unsigned long long)ino, cpos);
3106                 goto out;
3107         }
3108
3109         ret = ocfs2_split_extent(handle, et, path, index,
3110                                  &replace_rec, meta_ac, dealloc);
3111         if (ret)
3112                 mlog_errno(ret);
3113
3114 out:
3115         ocfs2_free_path(path);
3116         return ret;
3117 }
3118
3119 static int ocfs2_replace_clusters(handle_t *handle,
3120                                   struct ocfs2_cow_context *context,
3121                                   u32 cpos, u32 old,
3122                                   u32 new, u32 len,
3123                                   unsigned int ext_flags)
3124 {
3125         int ret;
3126         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3127         u64 ino = ocfs2_metadata_cache_owner(ci);
3128
3129         trace_ocfs2_replace_clusters((unsigned long long)ino,
3130                                      cpos, old, new, len, ext_flags);
3131
3132         /*If the old clusters is unwritten, no need to duplicate. */
3133         if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3134                 ret = context->cow_duplicate_clusters(handle, context->inode,
3135                                                       cpos, old, new, len);
3136                 if (ret) {
3137                         mlog_errno(ret);
3138                         goto out;
3139                 }
3140         }
3141
3142         ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
3143                                        cpos, new, len, ext_flags,
3144                                        context->meta_ac, &context->dealloc);
3145         if (ret)
3146                 mlog_errno(ret);
3147 out:
3148         return ret;
3149 }
3150
3151 int ocfs2_cow_sync_writeback(struct super_block *sb,
3152                              struct inode *inode,
3153                              u32 cpos, u32 num_clusters)
3154 {
3155         int ret = 0;
3156         loff_t offset, end, map_end;
3157         pgoff_t page_index;
3158         struct page *page;
3159
3160         if (ocfs2_should_order_data(inode))
3161                 return 0;
3162
3163         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
3164         end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits);
3165
3166         ret = filemap_fdatawrite_range(inode->i_mapping,
3167                                        offset, end - 1);
3168         if (ret < 0) {
3169                 mlog_errno(ret);
3170                 return ret;
3171         }
3172
3173         while (offset < end) {
3174                 page_index = offset >> PAGE_SHIFT;
3175                 map_end = ((loff_t)page_index + 1) << PAGE_SHIFT;
3176                 if (map_end > end)
3177                         map_end = end;
3178
3179                 page = find_or_create_page(inode->i_mapping,
3180                                            page_index, GFP_NOFS);
3181                 BUG_ON(!page);
3182
3183                 wait_on_page_writeback(page);
3184                 if (PageError(page)) {
3185                         ret = -EIO;
3186                         mlog_errno(ret);
3187                 } else
3188                         mark_page_accessed(page);
3189
3190                 unlock_page(page);
3191                 put_page(page);
3192                 page = NULL;
3193                 offset = map_end;
3194                 if (ret)
3195                         break;
3196         }
3197
3198         return ret;
3199 }
3200
3201 static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3202                                  u32 v_cluster, u32 *p_cluster,
3203                                  u32 *num_clusters,
3204                                  unsigned int *extent_flags)
3205 {
3206         return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3207                                   num_clusters, extent_flags);
3208 }
3209
3210 static int ocfs2_make_clusters_writable(struct super_block *sb,
3211                                         struct ocfs2_cow_context *context,
3212                                         u32 cpos, u32 p_cluster,
3213                                         u32 num_clusters, unsigned int e_flags)
3214 {
3215         int ret, delete, index, credits =  0;
3216         u32 new_bit, new_len, orig_num_clusters;
3217         unsigned int set_len;
3218         struct ocfs2_super *osb = OCFS2_SB(sb);
3219         handle_t *handle;
3220         struct buffer_head *ref_leaf_bh = NULL;
3221         struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
3222         struct ocfs2_refcount_rec rec;
3223
3224         trace_ocfs2_make_clusters_writable(cpos, p_cluster,
3225                                            num_clusters, e_flags);
3226
3227         ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3228                                              &context->data_et,
3229                                              ref_ci,
3230                                              context->ref_root_bh,
3231                                              &context->meta_ac,
3232                                              &context->data_ac, &credits);
3233         if (ret) {
3234                 mlog_errno(ret);
3235                 return ret;
3236         }
3237
3238         if (context->post_refcount)
3239                 credits += context->post_refcount->credits;
3240
3241         credits += context->extra_credits;
3242         handle = ocfs2_start_trans(osb, credits);
3243         if (IS_ERR(handle)) {
3244                 ret = PTR_ERR(handle);
3245                 mlog_errno(ret);
3246                 goto out;
3247         }
3248
3249         orig_num_clusters = num_clusters;
3250
3251         while (num_clusters) {
3252                 ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
3253                                              p_cluster, num_clusters,
3254                                              &rec, &index, &ref_leaf_bh);
3255                 if (ret) {
3256                         mlog_errno(ret);
3257                         goto out_commit;
3258                 }
3259
3260                 BUG_ON(!rec.r_refcount);
3261                 set_len = min((u64)p_cluster + num_clusters,
3262                               le64_to_cpu(rec.r_cpos) +
3263                               le32_to_cpu(rec.r_clusters)) - p_cluster;
3264
3265                 /*
3266                  * There are many different situation here.
3267                  * 1. If refcount == 1, remove the flag and don't COW.
3268                  * 2. If refcount > 1, allocate clusters.
3269                  *    Here we may not allocate r_len once at a time, so continue
3270                  *    until we reach num_clusters.
3271                  */
3272                 if (le32_to_cpu(rec.r_refcount) == 1) {
3273                         delete = 0;
3274                         ret = ocfs2_clear_ext_refcount(handle,
3275                                                        &context->data_et,
3276                                                        cpos, p_cluster,
3277                                                        set_len, e_flags,
3278                                                        context->meta_ac,
3279                                                        &context->dealloc);
3280                         if (ret) {
3281                                 mlog_errno(ret);
3282                                 goto out_commit;
3283                         }
3284                 } else {
3285                         delete = 1;
3286
3287                         ret = __ocfs2_claim_clusters(handle,
3288                                                      context->data_ac,
3289                                                      1, set_len,
3290                                                      &new_bit, &new_len);
3291                         if (ret) {
3292                                 mlog_errno(ret);
3293                                 goto out_commit;
3294                         }
3295
3296                         ret = ocfs2_replace_clusters(handle, context,
3297                                                      cpos, p_cluster, new_bit,
3298                                                      new_len, e_flags);
3299                         if (ret) {
3300                                 mlog_errno(ret);
3301                                 goto out_commit;
3302                         }
3303                         set_len = new_len;
3304                 }
3305
3306                 ret = __ocfs2_decrease_refcount(handle, ref_ci,
3307                                                 context->ref_root_bh,
3308                                                 p_cluster, set_len,
3309                                                 context->meta_ac,
3310                                                 &context->dealloc, delete);
3311                 if (ret) {
3312                         mlog_errno(ret);
3313                         goto out_commit;
3314                 }
3315
3316                 cpos += set_len;
3317                 p_cluster += set_len;
3318                 num_clusters -= set_len;
3319                 brelse(ref_leaf_bh);
3320                 ref_leaf_bh = NULL;
3321         }
3322
3323         /* handle any post_cow action. */
3324         if (context->post_refcount && context->post_refcount->func) {
3325                 ret = context->post_refcount->func(context->inode, handle,
3326                                                 context->post_refcount->para);
3327                 if (ret) {
3328                         mlog_errno(ret);
3329                         goto out_commit;
3330                 }
3331         }
3332
3333         /*
3334          * Here we should write the new page out first if we are
3335          * in write-back mode.
3336          */
3337         if (context->get_clusters == ocfs2_di_get_clusters) {
3338                 ret = ocfs2_cow_sync_writeback(sb, context->inode, cpos,
3339                                                orig_num_clusters);
3340                 if (ret)
3341                         mlog_errno(ret);
3342         }
3343
3344 out_commit:
3345         ocfs2_commit_trans(osb, handle);
3346
3347 out:
3348         if (context->data_ac) {
3349                 ocfs2_free_alloc_context(context->data_ac);
3350                 context->data_ac = NULL;
3351         }
3352         if (context->meta_ac) {
3353                 ocfs2_free_alloc_context(context->meta_ac);
3354                 context->meta_ac = NULL;
3355         }
3356         brelse(ref_leaf_bh);
3357
3358         return ret;
3359 }
3360
3361 static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
3362 {
3363         int ret = 0;
3364         struct inode *inode = context->inode;
3365         u32 cow_start = context->cow_start, cow_len = context->cow_len;
3366         u32 p_cluster, num_clusters;
3367         unsigned int ext_flags;
3368         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3369
3370         if (!ocfs2_refcount_tree(osb)) {
3371                 return ocfs2_error(inode->i_sb, "Inode %lu want to use refcount tree, but the feature bit is not set in the super block\n",
3372                                    inode->i_ino);
3373         }
3374
3375         ocfs2_init_dealloc_ctxt(&context->dealloc);
3376
3377         while (cow_len) {
3378                 ret = context->get_clusters(context, cow_start, &p_cluster,
3379                                             &num_clusters, &ext_flags);
3380                 if (ret) {
3381                         mlog_errno(ret);
3382                         break;
3383                 }
3384
3385                 BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
3386
3387                 if (cow_len < num_clusters)
3388                         num_clusters = cow_len;
3389
3390                 ret = ocfs2_make_clusters_writable(inode->i_sb, context,
3391                                                    cow_start, p_cluster,
3392                                                    num_clusters, ext_flags);
3393                 if (ret) {
3394                         mlog_errno(ret);
3395                         break;
3396                 }
3397
3398                 cow_len -= num_clusters;
3399                 cow_start += num_clusters;
3400         }
3401
3402         if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
3403                 ocfs2_schedule_truncate_log_flush(osb, 1);
3404                 ocfs2_run_deallocs(osb, &context->dealloc);
3405         }
3406
3407         return ret;
3408 }
3409
3410 /*
3411  * Starting at cpos, try to CoW write_len clusters.  Don't CoW
3412  * past max_cpos.  This will stop when it runs into a hole or an
3413  * unrefcounted extent.
3414  */
3415 static int ocfs2_refcount_cow_hunk(struct inode *inode,
3416                                    struct buffer_head *di_bh,
3417                                    u32 cpos, u32 write_len, u32 max_cpos)
3418 {
3419         int ret;
3420         u32 cow_start = 0, cow_len = 0;
3421         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3422         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3423         struct buffer_head *ref_root_bh = NULL;
3424         struct ocfs2_refcount_tree *ref_tree;
3425         struct ocfs2_cow_context *context = NULL;
3426
3427         BUG_ON(!ocfs2_is_refcount_inode(inode));
3428
3429         ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
3430                                               cpos, write_len, max_cpos,
3431                                               &cow_start, &cow_len);
3432         if (ret) {
3433                 mlog_errno(ret);
3434                 goto out;
3435         }
3436
3437         trace_ocfs2_refcount_cow_hunk(OCFS2_I(inode)->ip_blkno,
3438                                       cpos, write_len, max_cpos,
3439                                       cow_start, cow_len);
3440
3441         BUG_ON(cow_len == 0);
3442
3443         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3444         if (!context) {
3445                 ret = -ENOMEM;
3446                 mlog_errno(ret);
3447                 goto out;
3448         }
3449
3450         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3451                                        1, &ref_tree, &ref_root_bh);
3452         if (ret) {
3453                 mlog_errno(ret);
3454                 goto out;
3455         }
3456
3457         context->inode = inode;
3458         context->cow_start = cow_start;
3459         context->cow_len = cow_len;
3460         context->ref_tree = ref_tree;
3461         context->ref_root_bh = ref_root_bh;
3462         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3463         context->get_clusters = ocfs2_di_get_clusters;
3464
3465         ocfs2_init_dinode_extent_tree(&context->data_et,
3466                                       INODE_CACHE(inode), di_bh);
3467
3468         ret = ocfs2_replace_cow(context);
3469         if (ret)
3470                 mlog_errno(ret);
3471
3472         /*
3473          * truncate the extent map here since no matter whether we meet with
3474          * any error during the action, we shouldn't trust cached extent map
3475          * any more.
3476          */
3477         ocfs2_extent_map_trunc(inode, cow_start);
3478
3479         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3480         brelse(ref_root_bh);
3481 out:
3482         kfree(context);
3483         return ret;
3484 }
3485
3486 /*
3487  * CoW any and all clusters between cpos and cpos+write_len.
3488  * Don't CoW past max_cpos.  If this returns successfully, all
3489  * clusters between cpos and cpos+write_len are safe to modify.
3490  */
3491 int ocfs2_refcount_cow(struct inode *inode,
3492                        struct buffer_head *di_bh,
3493                        u32 cpos, u32 write_len, u32 max_cpos)
3494 {
3495         int ret = 0;
3496         u32 p_cluster, num_clusters;
3497         unsigned int ext_flags;
3498
3499         while (write_len) {
3500                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3501                                          &num_clusters, &ext_flags);
3502                 if (ret) {
3503                         mlog_errno(ret);
3504                         break;
3505                 }
3506
3507                 if (write_len < num_clusters)
3508                         num_clusters = write_len;
3509
3510                 if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3511                         ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
3512                                                       num_clusters, max_cpos);
3513                         if (ret) {
3514                                 mlog_errno(ret);
3515                                 break;
3516                         }
3517                 }
3518
3519                 write_len -= num_clusters;
3520                 cpos += num_clusters;
3521         }
3522
3523         return ret;
3524 }
3525
3526 static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3527                                           u32 v_cluster, u32 *p_cluster,
3528                                           u32 *num_clusters,
3529                                           unsigned int *extent_flags)
3530 {
3531         struct inode *inode = context->inode;
3532         struct ocfs2_xattr_value_root *xv = context->cow_object;
3533
3534         return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3535                                         num_clusters, &xv->xr_list,
3536                                         extent_flags);
3537 }
3538
3539 /*
3540  * Given a xattr value root, calculate the most meta/credits we need for
3541  * refcount tree change if we truncate it to 0.
3542  */
3543 int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3544                                        struct ocfs2_caching_info *ref_ci,
3545                                        struct buffer_head *ref_root_bh,
3546                                        struct ocfs2_xattr_value_root *xv,
3547                                        int *meta_add, int *credits)
3548 {
3549         int ret = 0, index, ref_blocks = 0;
3550         u32 p_cluster, num_clusters;
3551         u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3552         struct ocfs2_refcount_block *rb;
3553         struct ocfs2_refcount_rec rec;
3554         struct buffer_head *ref_leaf_bh = NULL;
3555
3556         while (cpos < clusters) {
3557                 ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3558                                                &num_clusters, &xv->xr_list,
3559                                                NULL);
3560                 if (ret) {
3561                         mlog_errno(ret);
3562                         goto out;
3563                 }
3564
3565                 cpos += num_clusters;
3566
3567                 while (num_clusters) {
3568                         ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3569                                                      p_cluster, num_clusters,
3570                                                      &rec, &index,
3571                                                      &ref_leaf_bh);
3572                         if (ret) {
3573                                 mlog_errno(ret);
3574                                 goto out;
3575                         }
3576
3577                         BUG_ON(!rec.r_refcount);
3578
3579                         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3580
3581                         /*
3582                          * We really don't know whether the other clusters is in
3583                          * this refcount block or not, so just take the worst
3584                          * case that all the clusters are in this block and each
3585                          * one will split a refcount rec, so totally we need
3586                          * clusters * 2 new refcount rec.
3587                          */
3588                         if (le16_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3589                             le16_to_cpu(rb->rf_records.rl_count))
3590                                 ref_blocks++;
3591
3592                         *credits += 1;
3593                         brelse(ref_leaf_bh);
3594                         ref_leaf_bh = NULL;
3595
3596                         if (num_clusters <= le32_to_cpu(rec.r_clusters))
3597                                 break;
3598                         else
3599                                 num_clusters -= le32_to_cpu(rec.r_clusters);
3600                         p_cluster += num_clusters;
3601                 }
3602         }
3603
3604         *meta_add += ref_blocks;
3605         if (!ref_blocks)
3606                 goto out;
3607
3608         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3609         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3610                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3611         else {
3612                 struct ocfs2_extent_tree et;
3613
3614                 ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3615                 *credits += ocfs2_calc_extend_credits(inode->i_sb,
3616                                                       et.et_root_el);
3617         }
3618
3619 out:
3620         brelse(ref_leaf_bh);
3621         return ret;
3622 }
3623
3624 /*
3625  * Do CoW for xattr.
3626  */
3627 int ocfs2_refcount_cow_xattr(struct inode *inode,
3628                              struct ocfs2_dinode *di,
3629                              struct ocfs2_xattr_value_buf *vb,
3630                              struct ocfs2_refcount_tree *ref_tree,
3631                              struct buffer_head *ref_root_bh,
3632                              u32 cpos, u32 write_len,
3633                              struct ocfs2_post_refcount *post)
3634 {
3635         int ret;
3636         struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3637         struct ocfs2_cow_context *context = NULL;
3638         u32 cow_start, cow_len;
3639
3640         BUG_ON(!ocfs2_is_refcount_inode(inode));
3641
3642         ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3643                                               cpos, write_len, UINT_MAX,
3644                                               &cow_start, &cow_len);
3645         if (ret) {
3646                 mlog_errno(ret);
3647                 goto out;
3648         }
3649
3650         BUG_ON(cow_len == 0);
3651
3652         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3653         if (!context) {
3654                 ret = -ENOMEM;
3655                 mlog_errno(ret);
3656                 goto out;
3657         }
3658
3659         context->inode = inode;
3660         context->cow_start = cow_start;
3661         context->cow_len = cow_len;
3662         context->ref_tree = ref_tree;
3663         context->ref_root_bh = ref_root_bh;
3664         context->cow_object = xv;
3665
3666         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3667         /* We need the extra credits for duplicate_clusters by jbd. */
3668         context->extra_credits =
3669                 ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3670         context->get_clusters = ocfs2_xattr_value_get_clusters;
3671         context->post_refcount = post;
3672
3673         ocfs2_init_xattr_value_extent_tree(&context->data_et,
3674                                            INODE_CACHE(inode), vb);
3675
3676         ret = ocfs2_replace_cow(context);
3677         if (ret)
3678                 mlog_errno(ret);
3679
3680 out:
3681         kfree(context);
3682         return ret;
3683 }
3684
3685 /*
3686  * Insert a new extent into refcount tree and mark a extent rec
3687  * as refcounted in the dinode tree.
3688  */
3689 int ocfs2_add_refcount_flag(struct inode *inode,
3690                             struct ocfs2_extent_tree *data_et,
3691                             struct ocfs2_caching_info *ref_ci,
3692                             struct buffer_head *ref_root_bh,
3693                             u32 cpos, u32 p_cluster, u32 num_clusters,
3694                             struct ocfs2_cached_dealloc_ctxt *dealloc,
3695                             struct ocfs2_post_refcount *post)
3696 {
3697         int ret;
3698         handle_t *handle;
3699         int credits = 1, ref_blocks = 0;
3700         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3701         struct ocfs2_alloc_context *meta_ac = NULL;
3702
3703         /* We need to be able to handle at least an extent tree split. */
3704         ref_blocks = ocfs2_extend_meta_needed(data_et->et_root_el);
3705
3706         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3707                                                ref_ci, ref_root_bh,
3708                                                p_cluster, num_clusters,
3709                                                &ref_blocks, &credits);
3710         if (ret) {
3711                 mlog_errno(ret);
3712                 goto out;
3713         }
3714
3715         trace_ocfs2_add_refcount_flag(ref_blocks, credits);
3716
3717         if (ref_blocks) {
3718                 ret = ocfs2_reserve_new_metadata_blocks(osb,
3719                                                         ref_blocks, &meta_ac);
3720                 if (ret) {
3721                         mlog_errno(ret);
3722                         goto out;
3723                 }
3724         }
3725
3726         if (post)
3727                 credits += post->credits;
3728
3729         handle = ocfs2_start_trans(osb, credits);
3730         if (IS_ERR(handle)) {
3731                 ret = PTR_ERR(handle);
3732                 mlog_errno(ret);
3733                 goto out;
3734         }
3735
3736         ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3737                                            cpos, num_clusters, p_cluster,
3738                                            meta_ac, dealloc);
3739         if (ret) {
3740                 mlog_errno(ret);
3741                 goto out_commit;
3742         }
3743
3744         ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3745                                         p_cluster, num_clusters, 0,
3746                                         meta_ac, dealloc);
3747         if (ret) {
3748                 mlog_errno(ret);
3749                 goto out_commit;
3750         }
3751
3752         if (post && post->func) {
3753                 ret = post->func(inode, handle, post->para);
3754                 if (ret)
3755                         mlog_errno(ret);
3756         }
3757
3758 out_commit:
3759         ocfs2_commit_trans(osb, handle);
3760 out:
3761         if (meta_ac)
3762                 ocfs2_free_alloc_context(meta_ac);
3763         return ret;
3764 }
3765
3766 static int ocfs2_change_ctime(struct inode *inode,
3767                               struct buffer_head *di_bh)
3768 {
3769         int ret;
3770         handle_t *handle;
3771         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3772
3773         handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3774                                    OCFS2_INODE_UPDATE_CREDITS);
3775         if (IS_ERR(handle)) {
3776                 ret = PTR_ERR(handle);
3777                 mlog_errno(ret);
3778                 goto out;
3779         }
3780
3781         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3782                                       OCFS2_JOURNAL_ACCESS_WRITE);
3783         if (ret) {
3784                 mlog_errno(ret);
3785                 goto out_commit;
3786         }
3787
3788         inode->i_ctime = current_time(inode);
3789         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
3790         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
3791
3792         ocfs2_journal_dirty(handle, di_bh);
3793
3794 out_commit:
3795         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3796 out:
3797         return ret;
3798 }
3799
3800 static int ocfs2_attach_refcount_tree(struct inode *inode,
3801                                       struct buffer_head *di_bh)
3802 {
3803         int ret, data_changed = 0;
3804         struct buffer_head *ref_root_bh = NULL;
3805         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3806         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3807         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3808         struct ocfs2_refcount_tree *ref_tree;
3809         unsigned int ext_flags;
3810         loff_t size;
3811         u32 cpos, num_clusters, clusters, p_cluster;
3812         struct ocfs2_cached_dealloc_ctxt dealloc;
3813         struct ocfs2_extent_tree di_et;
3814
3815         ocfs2_init_dealloc_ctxt(&dealloc);
3816
3817         if (!ocfs2_is_refcount_inode(inode)) {
3818                 ret = ocfs2_create_refcount_tree(inode, di_bh);
3819                 if (ret) {
3820                         mlog_errno(ret);
3821                         goto out;
3822                 }
3823         }
3824
3825         BUG_ON(!di->i_refcount_loc);
3826         ret = ocfs2_lock_refcount_tree(osb,
3827                                        le64_to_cpu(di->i_refcount_loc), 1,
3828                                        &ref_tree, &ref_root_bh);
3829         if (ret) {
3830                 mlog_errno(ret);
3831                 goto out;
3832         }
3833
3834         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
3835                 goto attach_xattr;
3836
3837         ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3838
3839         size = i_size_read(inode);
3840         clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3841
3842         cpos = 0;
3843         while (cpos < clusters) {
3844                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3845                                          &num_clusters, &ext_flags);
3846                 if (ret) {
3847                         mlog_errno(ret);
3848                         goto unlock;
3849                 }
3850                 if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3851                         ret = ocfs2_add_refcount_flag(inode, &di_et,
3852                                                       &ref_tree->rf_ci,
3853                                                       ref_root_bh, cpos,
3854                                                       p_cluster, num_clusters,
3855                                                       &dealloc, NULL);
3856                         if (ret) {
3857                                 mlog_errno(ret);
3858                                 goto unlock;
3859                         }
3860
3861                         data_changed = 1;
3862                 }
3863                 cpos += num_clusters;
3864         }
3865
3866 attach_xattr:
3867         if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
3868                 ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
3869                                                        &ref_tree->rf_ci,
3870                                                        ref_root_bh,
3871                                                        &dealloc);
3872                 if (ret) {
3873                         mlog_errno(ret);
3874                         goto unlock;
3875                 }
3876         }
3877
3878         if (data_changed) {
3879                 ret = ocfs2_change_ctime(inode, di_bh);
3880                 if (ret)
3881                         mlog_errno(ret);
3882         }
3883
3884 unlock:
3885         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3886         brelse(ref_root_bh);
3887
3888         if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3889                 ocfs2_schedule_truncate_log_flush(osb, 1);
3890                 ocfs2_run_deallocs(osb, &dealloc);
3891         }
3892 out:
3893         /*
3894          * Empty the extent map so that we may get the right extent
3895          * record from the disk.
3896          */
3897         ocfs2_extent_map_trunc(inode, 0);
3898
3899         return ret;
3900 }
3901
3902 static int ocfs2_add_refcounted_extent(struct inode *inode,
3903                                    struct ocfs2_extent_tree *et,
3904                                    struct ocfs2_caching_info *ref_ci,
3905                                    struct buffer_head *ref_root_bh,
3906                                    u32 cpos, u32 p_cluster, u32 num_clusters,
3907                                    unsigned int ext_flags,
3908                                    struct ocfs2_cached_dealloc_ctxt *dealloc)
3909 {
3910         int ret;
3911         handle_t *handle;
3912         int credits = 0;
3913         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3914         struct ocfs2_alloc_context *meta_ac = NULL;
3915
3916         ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3917                                              p_cluster, num_clusters,
3918                                              et, ref_ci,
3919                                              ref_root_bh, &meta_ac,
3920                                              NULL, &credits);
3921         if (ret) {
3922                 mlog_errno(ret);
3923                 goto out;
3924         }
3925
3926         handle = ocfs2_start_trans(osb, credits);
3927         if (IS_ERR(handle)) {
3928                 ret = PTR_ERR(handle);
3929                 mlog_errno(ret);
3930                 goto out;
3931         }
3932
3933         ret = ocfs2_insert_extent(handle, et, cpos,
3934                         ocfs2_clusters_to_blocks(inode->i_sb, p_cluster),
3935                         num_clusters, ext_flags, meta_ac);
3936         if (ret) {
3937                 mlog_errno(ret);
3938                 goto out_commit;
3939         }
3940
3941         ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3942                                       p_cluster, num_clusters,
3943                                       meta_ac, dealloc);
3944         if (ret) {
3945                 mlog_errno(ret);
3946                 goto out_commit;
3947         }
3948
3949         ret = dquot_alloc_space_nodirty(inode,
3950                 ocfs2_clusters_to_bytes(osb->sb, num_clusters));
3951         if (ret)
3952                 mlog_errno(ret);
3953
3954 out_commit:
3955         ocfs2_commit_trans(osb, handle);
3956 out:
3957         if (meta_ac)
3958                 ocfs2_free_alloc_context(meta_ac);
3959         return ret;
3960 }
3961
3962 static int ocfs2_duplicate_inline_data(struct inode *s_inode,
3963                                        struct buffer_head *s_bh,
3964                                        struct inode *t_inode,
3965                                        struct buffer_head *t_bh)
3966 {
3967         int ret;
3968         handle_t *handle;
3969         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
3970         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
3971         struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data;
3972
3973         BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
3974
3975         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
3976         if (IS_ERR(handle)) {
3977                 ret = PTR_ERR(handle);
3978                 mlog_errno(ret);
3979                 goto out;
3980         }
3981
3982         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
3983                                       OCFS2_JOURNAL_ACCESS_WRITE);
3984         if (ret) {
3985                 mlog_errno(ret);
3986                 goto out_commit;
3987         }
3988
3989         t_di->id2.i_data.id_count = s_di->id2.i_data.id_count;
3990         memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data,
3991                le16_to_cpu(s_di->id2.i_data.id_count));
3992         spin_lock(&OCFS2_I(t_inode)->ip_lock);
3993         OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
3994         t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features);
3995         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
3996
3997         ocfs2_journal_dirty(handle, t_bh);
3998
3999 out_commit:
4000         ocfs2_commit_trans(osb, handle);
4001 out:
4002         return ret;
4003 }
4004
4005 static int ocfs2_duplicate_extent_list(struct inode *s_inode,
4006                                 struct inode *t_inode,
4007                                 struct buffer_head *t_bh,
4008                                 struct ocfs2_caching_info *ref_ci,
4009                                 struct buffer_head *ref_root_bh,
4010                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
4011 {
4012         int ret = 0;
4013         u32 p_cluster, num_clusters, clusters, cpos;
4014         loff_t size;
4015         unsigned int ext_flags;
4016         struct ocfs2_extent_tree et;
4017
4018         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
4019
4020         size = i_size_read(s_inode);
4021         clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
4022
4023         cpos = 0;
4024         while (cpos < clusters) {
4025                 ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
4026                                          &num_clusters, &ext_flags);
4027                 if (ret) {
4028                         mlog_errno(ret);
4029                         goto out;
4030                 }
4031                 if (p_cluster) {
4032                         ret = ocfs2_add_refcounted_extent(t_inode, &et,
4033                                                           ref_ci, ref_root_bh,
4034                                                           cpos, p_cluster,
4035                                                           num_clusters,
4036                                                           ext_flags,
4037                                                           dealloc);
4038                         if (ret) {
4039                                 mlog_errno(ret);
4040                                 goto out;
4041                         }
4042                 }
4043
4044                 cpos += num_clusters;
4045         }
4046
4047 out:
4048         return ret;
4049 }
4050
4051 /*
4052  * change the new file's attributes to the src.
4053  *
4054  * reflink creates a snapshot of a file, that means the attributes
4055  * must be identical except for three exceptions - nlink, ino, and ctime.
4056  */
4057 static int ocfs2_complete_reflink(struct inode *s_inode,
4058                                   struct buffer_head *s_bh,
4059                                   struct inode *t_inode,
4060                                   struct buffer_head *t_bh,
4061                                   bool preserve)
4062 {
4063         int ret;
4064         handle_t *handle;
4065         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
4066         struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
4067         loff_t size = i_size_read(s_inode);
4068
4069         handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
4070                                    OCFS2_INODE_UPDATE_CREDITS);
4071         if (IS_ERR(handle)) {
4072                 ret = PTR_ERR(handle);
4073                 mlog_errno(ret);
4074                 return ret;
4075         }
4076
4077         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4078                                       OCFS2_JOURNAL_ACCESS_WRITE);
4079         if (ret) {
4080                 mlog_errno(ret);
4081                 goto out_commit;
4082         }
4083
4084         spin_lock(&OCFS2_I(t_inode)->ip_lock);
4085         OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
4086         OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
4087         OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
4088         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4089         i_size_write(t_inode, size);
4090         t_inode->i_blocks = s_inode->i_blocks;
4091
4092         di->i_xattr_inline_size = s_di->i_xattr_inline_size;
4093         di->i_clusters = s_di->i_clusters;
4094         di->i_size = s_di->i_size;
4095         di->i_dyn_features = s_di->i_dyn_features;
4096         di->i_attr = s_di->i_attr;
4097
4098         if (preserve) {
4099                 t_inode->i_uid = s_inode->i_uid;
4100                 t_inode->i_gid = s_inode->i_gid;
4101                 t_inode->i_mode = s_inode->i_mode;
4102                 di->i_uid = s_di->i_uid;
4103                 di->i_gid = s_di->i_gid;
4104                 di->i_mode = s_di->i_mode;
4105
4106                 /*
4107                  * update time.
4108                  * we want mtime to appear identical to the source and
4109                  * update ctime.
4110                  */
4111                 t_inode->i_ctime = current_time(t_inode);
4112
4113                 di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec);
4114                 di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec);
4115
4116                 t_inode->i_mtime = s_inode->i_mtime;
4117                 di->i_mtime = s_di->i_mtime;
4118                 di->i_mtime_nsec = s_di->i_mtime_nsec;
4119         }
4120
4121         ocfs2_journal_dirty(handle, t_bh);
4122
4123 out_commit:
4124         ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
4125         return ret;
4126 }
4127
4128 static int ocfs2_create_reflink_node(struct inode *s_inode,
4129                                      struct buffer_head *s_bh,
4130                                      struct inode *t_inode,
4131                                      struct buffer_head *t_bh,
4132                                      bool preserve)
4133 {
4134         int ret;
4135         struct buffer_head *ref_root_bh = NULL;
4136         struct ocfs2_cached_dealloc_ctxt dealloc;
4137         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
4138         struct ocfs2_refcount_block *rb;
4139         struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
4140         struct ocfs2_refcount_tree *ref_tree;
4141
4142         ocfs2_init_dealloc_ctxt(&dealloc);
4143
4144         ret = ocfs2_set_refcount_tree(t_inode, t_bh,
4145                                       le64_to_cpu(di->i_refcount_loc));
4146         if (ret) {
4147                 mlog_errno(ret);
4148                 goto out;
4149         }
4150
4151         if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4152                 ret = ocfs2_duplicate_inline_data(s_inode, s_bh,
4153                                                   t_inode, t_bh);
4154                 if (ret)
4155                         mlog_errno(ret);
4156                 goto out;
4157         }
4158
4159         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
4160                                        1, &ref_tree, &ref_root_bh);
4161         if (ret) {
4162                 mlog_errno(ret);
4163                 goto out;
4164         }
4165         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
4166
4167         ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4168                                           &ref_tree->rf_ci, ref_root_bh,
4169                                           &dealloc);
4170         if (ret) {
4171                 mlog_errno(ret);
4172                 goto out_unlock_refcount;
4173         }
4174
4175 out_unlock_refcount:
4176         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4177         brelse(ref_root_bh);
4178 out:
4179         if (ocfs2_dealloc_has_cluster(&dealloc)) {
4180                 ocfs2_schedule_truncate_log_flush(osb, 1);
4181                 ocfs2_run_deallocs(osb, &dealloc);
4182         }
4183
4184         return ret;
4185 }
4186
4187 static int __ocfs2_reflink(struct dentry *old_dentry,
4188                            struct buffer_head *old_bh,
4189                            struct inode *new_inode,
4190                            bool preserve)
4191 {
4192         int ret;
4193         struct inode *inode = d_inode(old_dentry);
4194         struct buffer_head *new_bh = NULL;
4195
4196         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) {
4197                 ret = -EINVAL;
4198                 mlog_errno(ret);
4199                 goto out;
4200         }
4201
4202         ret = filemap_fdatawrite(inode->i_mapping);
4203         if (ret) {
4204                 mlog_errno(ret);
4205                 goto out;
4206         }
4207
4208         ret = ocfs2_attach_refcount_tree(inode, old_bh);
4209         if (ret) {
4210                 mlog_errno(ret);
4211                 goto out;
4212         }
4213
4214         inode_lock_nested(new_inode, I_MUTEX_CHILD);
4215         ret = ocfs2_inode_lock_nested(new_inode, &new_bh, 1,
4216                                       OI_LS_REFLINK_TARGET);
4217         if (ret) {
4218                 mlog_errno(ret);
4219                 goto out_unlock;
4220         }
4221
4222         ret = ocfs2_create_reflink_node(inode, old_bh,
4223                                         new_inode, new_bh, preserve);
4224         if (ret) {
4225                 mlog_errno(ret);
4226                 goto inode_unlock;
4227         }
4228
4229         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
4230                 ret = ocfs2_reflink_xattrs(inode, old_bh,
4231                                            new_inode, new_bh,
4232                                            preserve);
4233                 if (ret) {
4234                         mlog_errno(ret);
4235                         goto inode_unlock;
4236                 }
4237         }
4238
4239         ret = ocfs2_complete_reflink(inode, old_bh,
4240                                      new_inode, new_bh, preserve);
4241         if (ret)
4242                 mlog_errno(ret);
4243
4244 inode_unlock:
4245         ocfs2_inode_unlock(new_inode, 1);
4246         brelse(new_bh);
4247 out_unlock:
4248         inode_unlock(new_inode);
4249 out:
4250         if (!ret) {
4251                 ret = filemap_fdatawait(inode->i_mapping);
4252                 if (ret)
4253                         mlog_errno(ret);
4254         }
4255         return ret;
4256 }
4257
4258 static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
4259                          struct dentry *new_dentry, bool preserve)
4260 {
4261         int error, had_lock;
4262         struct inode *inode = d_inode(old_dentry);
4263         struct buffer_head *old_bh = NULL;
4264         struct inode *new_orphan_inode = NULL;
4265         struct ocfs2_lock_holder oh;
4266
4267         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4268                 return -EOPNOTSUPP;
4269
4270
4271         error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
4272                                              &new_orphan_inode);
4273         if (error) {
4274                 mlog_errno(error);
4275                 goto out;
4276         }
4277
4278         error = ocfs2_rw_lock(inode, 1);
4279         if (error) {
4280                 mlog_errno(error);
4281                 goto out;
4282         }
4283
4284         error = ocfs2_inode_lock(inode, &old_bh, 1);
4285         if (error) {
4286                 mlog_errno(error);
4287                 ocfs2_rw_unlock(inode, 1);
4288                 goto out;
4289         }
4290
4291         down_write(&OCFS2_I(inode)->ip_xattr_sem);
4292         down_write(&OCFS2_I(inode)->ip_alloc_sem);
4293         error = __ocfs2_reflink(old_dentry, old_bh,
4294                                 new_orphan_inode, preserve);
4295         up_write(&OCFS2_I(inode)->ip_alloc_sem);
4296         up_write(&OCFS2_I(inode)->ip_xattr_sem);
4297
4298         ocfs2_inode_unlock(inode, 1);
4299         ocfs2_rw_unlock(inode, 1);
4300         brelse(old_bh);
4301
4302         if (error) {
4303                 mlog_errno(error);
4304                 goto out;
4305         }
4306
4307         had_lock = ocfs2_inode_lock_tracker(new_orphan_inode, NULL, 1,
4308                                             &oh);
4309         if (had_lock < 0) {
4310                 error = had_lock;
4311                 mlog_errno(error);
4312                 goto out;
4313         }
4314
4315         /* If the security isn't preserved, we need to re-initialize them. */
4316         if (!preserve) {
4317                 error = ocfs2_init_security_and_acl(dir, new_orphan_inode,
4318                                                     &new_dentry->d_name);
4319                 if (error)
4320                         mlog_errno(error);
4321         }
4322         if (!error) {
4323                 error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
4324                                                        new_dentry);
4325                 if (error)
4326                         mlog_errno(error);
4327         }
4328         ocfs2_inode_unlock_tracker(new_orphan_inode, 1, &oh, had_lock);
4329
4330 out:
4331         if (new_orphan_inode) {
4332                 /*
4333                  * We need to open_unlock the inode no matter whether we
4334                  * succeed or not, so that other nodes can delete it later.
4335                  */
4336                 ocfs2_open_unlock(new_orphan_inode);
4337                 if (error)
4338                         iput(new_orphan_inode);
4339         }
4340
4341         return error;
4342 }
4343
4344 /*
4345  * Below here are the bits used by OCFS2_IOC_REFLINK() to fake
4346  * sys_reflink().  This will go away when vfs_reflink() exists in
4347  * fs/namei.c.
4348  */
4349
4350 /* copied from may_create in VFS. */
4351 static inline int ocfs2_may_create(struct inode *dir, struct dentry *child)
4352 {
4353         if (d_really_is_positive(child))
4354                 return -EEXIST;
4355         if (IS_DEADDIR(dir))
4356                 return -ENOENT;
4357         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
4358 }
4359
4360 /**
4361  * ocfs2_vfs_reflink - Create a reference-counted link
4362  *
4363  * @old_dentry:        source dentry + inode
4364  * @dir:       directory to create the target
4365  * @new_dentry:        target dentry
4366  * @preserve:  if true, preserve all file attributes
4367  */
4368 static int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir,
4369                              struct dentry *new_dentry, bool preserve)
4370 {
4371         struct inode *inode = d_inode(old_dentry);
4372         int error;
4373
4374         if (!inode)
4375                 return -ENOENT;
4376
4377         error = ocfs2_may_create(dir, new_dentry);
4378         if (error)
4379                 return error;
4380
4381         if (dir->i_sb != inode->i_sb)
4382                 return -EXDEV;
4383
4384         /*
4385          * A reflink to an append-only or immutable file cannot be created.
4386          */
4387         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4388                 return -EPERM;
4389
4390         /* Only regular files can be reflinked. */
4391         if (!S_ISREG(inode->i_mode))
4392                 return -EPERM;
4393
4394         /*
4395          * If the caller wants to preserve ownership, they require the
4396          * rights to do so.
4397          */
4398         if (preserve) {
4399                 if (!uid_eq(current_fsuid(), inode->i_uid) && !capable(CAP_CHOWN))
4400                         return -EPERM;
4401                 if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN))
4402                         return -EPERM;
4403         }
4404
4405         /*
4406          * If the caller is modifying any aspect of the attributes, they
4407          * are not creating a snapshot.  They need read permission on the
4408          * file.
4409          */
4410         if (!preserve) {
4411                 error = inode_permission(inode, MAY_READ);
4412                 if (error)
4413                         return error;
4414         }
4415
4416         inode_lock(inode);
4417         error = dquot_initialize(dir);
4418         if (!error)
4419                 error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve);
4420         inode_unlock(inode);
4421         if (!error)
4422                 fsnotify_create(dir, new_dentry);
4423         return error;
4424 }
4425 /*
4426  * Most codes are copied from sys_linkat.
4427  */
4428 int ocfs2_reflink_ioctl(struct inode *inode,
4429                         const char __user *oldname,
4430                         const char __user *newname,
4431                         bool preserve)
4432 {
4433         struct dentry *new_dentry;
4434         struct path old_path, new_path;
4435         int error;
4436
4437         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4438                 return -EOPNOTSUPP;
4439
4440         error = user_path_at(AT_FDCWD, oldname, 0, &old_path);
4441         if (error) {
4442                 mlog_errno(error);
4443                 return error;
4444         }
4445
4446         new_dentry = user_path_create(AT_FDCWD, newname, &new_path, 0);
4447         error = PTR_ERR(new_dentry);
4448         if (IS_ERR(new_dentry)) {
4449                 mlog_errno(error);
4450                 goto out;
4451         }
4452
4453         error = -EXDEV;
4454         if (old_path.mnt != new_path.mnt) {
4455                 mlog_errno(error);
4456                 goto out_dput;
4457         }
4458
4459         error = ocfs2_vfs_reflink(old_path.dentry,
4460                                   d_inode(new_path.dentry),
4461                                   new_dentry, preserve);
4462 out_dput:
4463         done_path_create(&new_path, new_dentry);
4464 out:
4465         path_put(&old_path);
4466
4467         return error;
4468 }
4469
4470 /* Update destination inode size, if necessary. */
4471 static int ocfs2_reflink_update_dest(struct inode *dest,
4472                                      struct buffer_head *d_bh,
4473                                      loff_t newlen)
4474 {
4475         handle_t *handle;
4476         int ret;
4477
4478         dest->i_blocks = ocfs2_inode_sector_count(dest);
4479
4480         if (newlen <= i_size_read(dest))
4481                 return 0;
4482
4483         handle = ocfs2_start_trans(OCFS2_SB(dest->i_sb),
4484                                    OCFS2_INODE_UPDATE_CREDITS);
4485         if (IS_ERR(handle)) {
4486                 ret = PTR_ERR(handle);
4487                 mlog_errno(ret);
4488                 return ret;
4489         }
4490
4491         /* Extend i_size if needed. */
4492         spin_lock(&OCFS2_I(dest)->ip_lock);
4493         if (newlen > i_size_read(dest))
4494                 i_size_write(dest, newlen);
4495         spin_unlock(&OCFS2_I(dest)->ip_lock);
4496         dest->i_ctime = dest->i_mtime = current_time(dest);
4497
4498         ret = ocfs2_mark_inode_dirty(handle, dest, d_bh);
4499         if (ret) {
4500                 mlog_errno(ret);
4501                 goto out_commit;
4502         }
4503
4504 out_commit:
4505         ocfs2_commit_trans(OCFS2_SB(dest->i_sb), handle);
4506         return ret;
4507 }
4508
4509 /* Remap the range pos_in:len in s_inode to pos_out:len in t_inode. */
4510 static int ocfs2_reflink_remap_extent(struct inode *s_inode,
4511                                       struct buffer_head *s_bh,
4512                                       loff_t pos_in,
4513                                       struct inode *t_inode,
4514                                       struct buffer_head *t_bh,
4515                                       loff_t pos_out,
4516                                       loff_t len,
4517                                       struct ocfs2_cached_dealloc_ctxt *dealloc)
4518 {
4519         struct ocfs2_extent_tree s_et;
4520         struct ocfs2_extent_tree t_et;
4521         struct ocfs2_dinode *dis;
4522         struct buffer_head *ref_root_bh = NULL;
4523         struct ocfs2_refcount_tree *ref_tree;
4524         struct ocfs2_super *osb;
4525         loff_t pstart, plen;
4526         u32 p_cluster, num_clusters, slast, spos, tpos;
4527         unsigned int ext_flags;
4528         int ret = 0;
4529
4530         osb = OCFS2_SB(s_inode->i_sb);
4531         dis = (struct ocfs2_dinode *)s_bh->b_data;
4532         ocfs2_init_dinode_extent_tree(&s_et, INODE_CACHE(s_inode), s_bh);
4533         ocfs2_init_dinode_extent_tree(&t_et, INODE_CACHE(t_inode), t_bh);
4534
4535         spos = ocfs2_bytes_to_clusters(s_inode->i_sb, pos_in);
4536         tpos = ocfs2_bytes_to_clusters(t_inode->i_sb, pos_out);
4537         slast = ocfs2_clusters_for_bytes(s_inode->i_sb, pos_in + len);
4538
4539         while (spos < slast) {
4540                 if (fatal_signal_pending(current)) {
4541                         ret = -EINTR;
4542                         goto out;
4543                 }
4544
4545                 /* Look up the extent. */
4546                 ret = ocfs2_get_clusters(s_inode, spos, &p_cluster,
4547                                          &num_clusters, &ext_flags);
4548                 if (ret) {
4549                         mlog_errno(ret);
4550                         goto out;
4551                 }
4552
4553                 num_clusters = min_t(u32, num_clusters, slast - spos);
4554
4555                 /* Punch out the dest range. */
4556                 pstart = ocfs2_clusters_to_bytes(t_inode->i_sb, tpos);
4557                 plen = ocfs2_clusters_to_bytes(t_inode->i_sb, num_clusters);
4558                 ret = ocfs2_remove_inode_range(t_inode, t_bh, pstart, plen);
4559                 if (ret) {
4560                         mlog_errno(ret);
4561                         goto out;
4562                 }
4563
4564                 if (p_cluster == 0)
4565                         goto next_loop;
4566
4567                 /* Lock the refcount btree... */
4568                 ret = ocfs2_lock_refcount_tree(osb,
4569                                                le64_to_cpu(dis->i_refcount_loc),
4570                                                1, &ref_tree, &ref_root_bh);
4571                 if (ret) {
4572                         mlog_errno(ret);
4573                         goto out;
4574                 }
4575
4576                 /* Mark s_inode's extent as refcounted. */
4577                 if (!(ext_flags & OCFS2_EXT_REFCOUNTED)) {
4578                         ret = ocfs2_add_refcount_flag(s_inode, &s_et,
4579                                                       &ref_tree->rf_ci,
4580                                                       ref_root_bh, spos,
4581                                                       p_cluster, num_clusters,
4582                                                       dealloc, NULL);
4583                         if (ret) {
4584                                 mlog_errno(ret);
4585                                 goto out_unlock_refcount;
4586                         }
4587                 }
4588
4589                 /* Map in the new extent. */
4590                 ext_flags |= OCFS2_EXT_REFCOUNTED;
4591                 ret = ocfs2_add_refcounted_extent(t_inode, &t_et,
4592                                                   &ref_tree->rf_ci,
4593                                                   ref_root_bh,
4594                                                   tpos, p_cluster,
4595                                                   num_clusters,
4596                                                   ext_flags,
4597                                                   dealloc);
4598                 if (ret) {
4599                         mlog_errno(ret);
4600                         goto out_unlock_refcount;
4601                 }
4602
4603                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4604                 brelse(ref_root_bh);
4605 next_loop:
4606                 spos += num_clusters;
4607                 tpos += num_clusters;
4608         }
4609
4610 out:
4611         return ret;
4612 out_unlock_refcount:
4613         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4614         brelse(ref_root_bh);
4615         return ret;
4616 }
4617
4618 /* Set up refcount tree and remap s_inode to t_inode. */
4619 static int ocfs2_reflink_remap_blocks(struct inode *s_inode,
4620                                       struct buffer_head *s_bh,
4621                                       loff_t pos_in,
4622                                       struct inode *t_inode,
4623                                       struct buffer_head *t_bh,
4624                                       loff_t pos_out,
4625                                       loff_t len)
4626 {
4627         struct ocfs2_cached_dealloc_ctxt dealloc;
4628         struct ocfs2_super *osb;
4629         struct ocfs2_dinode *dis;
4630         struct ocfs2_dinode *dit;
4631         int ret;
4632
4633         osb = OCFS2_SB(s_inode->i_sb);
4634         dis = (struct ocfs2_dinode *)s_bh->b_data;
4635         dit = (struct ocfs2_dinode *)t_bh->b_data;
4636         ocfs2_init_dealloc_ctxt(&dealloc);
4637
4638         /*
4639          * If we're reflinking the entire file and the source is inline
4640          * data, just copy the contents.
4641          */
4642         if (pos_in == pos_out && pos_in == 0 && len == i_size_read(s_inode) &&
4643             i_size_read(t_inode) <= len &&
4644             (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) {
4645                 ret = ocfs2_duplicate_inline_data(s_inode, s_bh, t_inode, t_bh);
4646                 if (ret)
4647                         mlog_errno(ret);
4648                 goto out;
4649         }
4650
4651         /*
4652          * If both inodes belong to two different refcount groups then
4653          * forget it because we don't know how (or want) to go merging
4654          * refcount trees.
4655          */
4656         ret = -EOPNOTSUPP;
4657         if (ocfs2_is_refcount_inode(s_inode) &&
4658             ocfs2_is_refcount_inode(t_inode) &&
4659             le64_to_cpu(dis->i_refcount_loc) !=
4660             le64_to_cpu(dit->i_refcount_loc))
4661                 goto out;
4662
4663         /* Neither inode has a refcount tree.  Add one to s_inode. */
4664         if (!ocfs2_is_refcount_inode(s_inode) &&
4665             !ocfs2_is_refcount_inode(t_inode)) {
4666                 ret = ocfs2_create_refcount_tree(s_inode, s_bh);
4667                 if (ret) {
4668                         mlog_errno(ret);
4669                         goto out;
4670                 }
4671         }
4672
4673         /* Ensure that both inodes end up with the same refcount tree. */
4674         if (!ocfs2_is_refcount_inode(s_inode)) {
4675                 ret = ocfs2_set_refcount_tree(s_inode, s_bh,
4676                                               le64_to_cpu(dit->i_refcount_loc));
4677                 if (ret) {
4678                         mlog_errno(ret);
4679                         goto out;
4680                 }
4681         }
4682         if (!ocfs2_is_refcount_inode(t_inode)) {
4683                 ret = ocfs2_set_refcount_tree(t_inode, t_bh,
4684                                               le64_to_cpu(dis->i_refcount_loc));
4685                 if (ret) {
4686                         mlog_errno(ret);
4687                         goto out;
4688                 }
4689         }
4690
4691         /* Turn off inline data in the dest file. */
4692         if (OCFS2_I(t_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4693                 ret = ocfs2_convert_inline_data_to_extents(t_inode, t_bh);
4694                 if (ret) {
4695                         mlog_errno(ret);
4696                         goto out;
4697                 }
4698         }
4699
4700         /* Actually remap extents now. */
4701         ret = ocfs2_reflink_remap_extent(s_inode, s_bh, pos_in, t_inode, t_bh,
4702                                          pos_out, len, &dealloc);
4703         if (ret) {
4704                 mlog_errno(ret);
4705                 goto out;
4706         }
4707
4708 out:
4709         if (ocfs2_dealloc_has_cluster(&dealloc)) {
4710                 ocfs2_schedule_truncate_log_flush(osb, 1);
4711                 ocfs2_run_deallocs(osb, &dealloc);
4712         }
4713
4714         return ret;
4715 }
4716
4717 /* Lock an inode and grab a bh pointing to the inode. */
4718 static int ocfs2_reflink_inodes_lock(struct inode *s_inode,
4719                                      struct buffer_head **bh_s,
4720                                      struct inode *t_inode,
4721                                      struct buffer_head **bh_t)
4722 {
4723         struct inode *inode1 = s_inode;
4724         struct inode *inode2 = t_inode;
4725         struct ocfs2_inode_info *oi1;
4726         struct ocfs2_inode_info *oi2;
4727         struct buffer_head *bh1 = NULL;
4728         struct buffer_head *bh2 = NULL;
4729         bool same_inode = (s_inode == t_inode);
4730         bool need_swap = (inode1->i_ino > inode2->i_ino);
4731         int status;
4732
4733         /* First grab the VFS and rw locks. */
4734         lock_two_nondirectories(s_inode, t_inode);
4735         if (need_swap)
4736                 swap(inode1, inode2);
4737
4738         status = ocfs2_rw_lock(inode1, 1);
4739         if (status) {
4740                 mlog_errno(status);
4741                 goto out_i1;
4742         }
4743         if (!same_inode) {
4744                 status = ocfs2_rw_lock(inode2, 1);
4745                 if (status) {
4746                         mlog_errno(status);
4747                         goto out_i2;
4748                 }
4749         }
4750
4751         /* Now go for the cluster locks */
4752         oi1 = OCFS2_I(inode1);
4753         oi2 = OCFS2_I(inode2);
4754
4755         trace_ocfs2_double_lock((unsigned long long)oi1->ip_blkno,
4756                                 (unsigned long long)oi2->ip_blkno);
4757
4758         /* We always want to lock the one with the lower lockid first. */
4759         if (oi1->ip_blkno > oi2->ip_blkno)
4760                 mlog_errno(-ENOLCK);
4761
4762         /* lock id1 */
4763         status = ocfs2_inode_lock_nested(inode1, &bh1, 1,
4764                                          OI_LS_REFLINK_TARGET);
4765         if (status < 0) {
4766                 if (status != -ENOENT)
4767                         mlog_errno(status);
4768                 goto out_rw2;
4769         }
4770
4771         /* lock id2 */
4772         if (!same_inode) {
4773                 status = ocfs2_inode_lock_nested(inode2, &bh2, 1,
4774                                                  OI_LS_REFLINK_TARGET);
4775                 if (status < 0) {
4776                         if (status != -ENOENT)
4777                                 mlog_errno(status);
4778                         goto out_cl1;
4779                 }
4780         } else {
4781                 bh2 = bh1;
4782         }
4783
4784         /*
4785          * If we swapped inode order above, we have to swap the buffer heads
4786          * before passing them back to the caller.
4787          */
4788         if (need_swap)
4789                 swap(bh1, bh2);
4790         *bh_s = bh1;
4791         *bh_t = bh2;
4792
4793         trace_ocfs2_double_lock_end(
4794                         (unsigned long long)oi1->ip_blkno,
4795                         (unsigned long long)oi2->ip_blkno);
4796
4797         return 0;
4798
4799 out_cl1:
4800         ocfs2_inode_unlock(inode1, 1);
4801         brelse(bh1);
4802 out_rw2:
4803         ocfs2_rw_unlock(inode2, 1);
4804 out_i2:
4805         ocfs2_rw_unlock(inode1, 1);
4806 out_i1:
4807         unlock_two_nondirectories(s_inode, t_inode);
4808         return status;
4809 }
4810
4811 /* Unlock both inodes and release buffers. */
4812 static void ocfs2_reflink_inodes_unlock(struct inode *s_inode,
4813                                         struct buffer_head *s_bh,
4814                                         struct inode *t_inode,
4815                                         struct buffer_head *t_bh)
4816 {
4817         ocfs2_inode_unlock(s_inode, 1);
4818         ocfs2_rw_unlock(s_inode, 1);
4819         brelse(s_bh);
4820         if (s_inode != t_inode) {
4821                 ocfs2_inode_unlock(t_inode, 1);
4822                 ocfs2_rw_unlock(t_inode, 1);
4823                 brelse(t_bh);
4824         }
4825         unlock_two_nondirectories(s_inode, t_inode);
4826 }
4827
4828 /* Link a range of blocks from one file to another. */
4829 int ocfs2_reflink_remap_range(struct file *file_in,
4830                               loff_t pos_in,
4831                               struct file *file_out,
4832                               loff_t pos_out,
4833                               u64 len,
4834                               bool is_dedupe)
4835 {
4836         struct inode *inode_in = file_inode(file_in);
4837         struct inode *inode_out = file_inode(file_out);
4838         struct ocfs2_super *osb = OCFS2_SB(inode_in->i_sb);
4839         struct buffer_head *in_bh = NULL, *out_bh = NULL;
4840         bool same_inode = (inode_in == inode_out);
4841         ssize_t ret;
4842
4843         if (!ocfs2_refcount_tree(osb))
4844                 return -EOPNOTSUPP;
4845         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
4846                 return -EROFS;
4847
4848         /* Lock both files against IO */
4849         ret = ocfs2_reflink_inodes_lock(inode_in, &in_bh, inode_out, &out_bh);
4850         if (ret)
4851                 return ret;
4852
4853         /* Check file eligibility and prepare for block sharing. */
4854         ret = -EINVAL;
4855         if ((OCFS2_I(inode_in)->ip_flags & OCFS2_INODE_SYSTEM_FILE) ||
4856             (OCFS2_I(inode_out)->ip_flags & OCFS2_INODE_SYSTEM_FILE))
4857                 goto out_unlock;
4858
4859         ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
4860                         &len, is_dedupe);
4861         if (ret <= 0)
4862                 goto out_unlock;
4863
4864         /* Lock out changes to the allocation maps and remap. */
4865         down_write(&OCFS2_I(inode_in)->ip_alloc_sem);
4866         if (!same_inode)
4867                 down_write_nested(&OCFS2_I(inode_out)->ip_alloc_sem,
4868                                   SINGLE_DEPTH_NESTING);
4869
4870         ret = ocfs2_reflink_remap_blocks(inode_in, in_bh, pos_in, inode_out,
4871                                          out_bh, pos_out, len);
4872
4873         /* Zap any page cache for the destination file's range. */
4874         if (!ret)
4875                 truncate_inode_pages_range(&inode_out->i_data, pos_out,
4876                                            PAGE_ALIGN(pos_out + len) - 1);
4877
4878         up_write(&OCFS2_I(inode_in)->ip_alloc_sem);
4879         if (!same_inode)
4880                 up_write(&OCFS2_I(inode_out)->ip_alloc_sem);
4881         if (ret) {
4882                 mlog_errno(ret);
4883                 goto out_unlock;
4884         }
4885
4886         /*
4887          * Empty the extent map so that we may get the right extent
4888          * record from the disk.
4889          */
4890         ocfs2_extent_map_trunc(inode_in, 0);
4891         ocfs2_extent_map_trunc(inode_out, 0);
4892
4893         ret = ocfs2_reflink_update_dest(inode_out, out_bh, pos_out + len);
4894         if (ret) {
4895                 mlog_errno(ret);
4896                 goto out_unlock;
4897         }
4898
4899         ocfs2_reflink_inodes_unlock(inode_in, in_bh, inode_out, out_bh);
4900         return 0;
4901
4902 out_unlock:
4903         ocfs2_reflink_inodes_unlock(inode_in, in_bh, inode_out, out_bh);
4904         return ret;
4905 }