11802d0beSThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only
2fa60ce2cSMasahiro Yamada /*
3f2c870e3STao Ma * refcounttree.c
4f2c870e3STao Ma *
5f2c870e3STao Ma * Copyright (C) 2009 Oracle. All rights reserved.
6f2c870e3STao Ma */
7f2c870e3STao Ma
8e73a819dSTao Ma #include <linux/sort.h>
9f2c870e3STao Ma #include <cluster/masklog.h>
10f2c870e3STao Ma #include "ocfs2.h"
11f2c870e3STao Ma #include "inode.h"
12f2c870e3STao Ma #include "alloc.h"
13f2c870e3STao Ma #include "suballoc.h"
14f2c870e3STao Ma #include "journal.h"
15f2c870e3STao Ma #include "uptodate.h"
16f2c870e3STao Ma #include "super.h"
17f2c870e3STao Ma #include "buffer_head_io.h"
18f2c870e3STao Ma #include "blockcheck.h"
19c732eb16STao Ma #include "refcounttree.h"
208bf396deSTao Ma #include "sysfile.h"
21374a263eSTao Ma #include "dlmglue.h"
22e73a819dSTao Ma #include "extent_map.h"
236f70fa51STao Ma #include "aops.h"
24492a8a33STao Ma #include "xattr.h"
2509bf27a0STao Ma #include "namei.h"
26198aac28STao Ma #include "ocfs2_trace.h"
2729ac8e85SDarrick J. Wong #include "file.h"
286f70fa51STao Ma
296f70fa51STao Ma #include <linux/bio.h>
306f70fa51STao Ma #include <linux/blkdev.h>
316f70fa51STao Ma #include <linux/slab.h>
326f70fa51STao Ma #include <linux/writeback.h>
336f70fa51STao Ma #include <linux/pagevec.h>
346f70fa51STao Ma #include <linux/swap.h>
35bd50873dSTao Ma #include <linux/security.h>
36bd50873dSTao Ma #include <linux/fsnotify.h>
37bd50873dSTao Ma #include <linux/quotaops.h>
38bd50873dSTao Ma #include <linux/namei.h>
39bd50873dSTao Ma #include <linux/mount.h>
40702e5bc6SChristoph Hellwig #include <linux/posix_acl.h>
416f70fa51STao Ma
426f70fa51STao Ma struct ocfs2_cow_context {
436f70fa51STao Ma struct inode *inode;
446f70fa51STao Ma u32 cow_start;
456f70fa51STao Ma u32 cow_len;
46913580b4STao Ma struct ocfs2_extent_tree data_et;
47913580b4STao Ma struct ocfs2_refcount_tree *ref_tree;
486f70fa51STao Ma struct buffer_head *ref_root_bh;
496f70fa51STao Ma struct ocfs2_alloc_context *meta_ac;
506f70fa51STao Ma struct ocfs2_alloc_context *data_ac;
516f70fa51STao Ma struct ocfs2_cached_dealloc_ctxt dealloc;
52492a8a33STao Ma void *cow_object;
53492a8a33STao Ma struct ocfs2_post_refcount *post_refcount;
54492a8a33STao Ma int extra_credits;
55913580b4STao Ma int (*get_clusters)(struct ocfs2_cow_context *context,
56913580b4STao Ma u32 v_cluster, u32 *p_cluster,
57913580b4STao Ma u32 *num_clusters,
58913580b4STao Ma unsigned int *extent_flags);
59913580b4STao Ma int (*cow_duplicate_clusters)(handle_t *handle,
60c7dd3392STiger Yang struct inode *inode,
61913580b4STao Ma u32 cpos, u32 old_cluster,
62913580b4STao Ma u32 new_cluster, u32 new_len);
636f70fa51STao Ma };
64c732eb16STao Ma
65c732eb16STao Ma static inline struct ocfs2_refcount_tree *
cache_info_to_refcount(struct ocfs2_caching_info * ci)66c732eb16STao Ma cache_info_to_refcount(struct ocfs2_caching_info *ci)
67c732eb16STao Ma {
68c732eb16STao Ma return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
69c732eb16STao Ma }
70f2c870e3STao Ma
ocfs2_validate_refcount_block(struct super_block * sb,struct buffer_head * bh)71f2c870e3STao Ma static int ocfs2_validate_refcount_block(struct super_block *sb,
72f2c870e3STao Ma struct buffer_head *bh)
73f2c870e3STao Ma {
74f2c870e3STao Ma int rc;
75f2c870e3STao Ma struct ocfs2_refcount_block *rb =
76f2c870e3STao Ma (struct ocfs2_refcount_block *)bh->b_data;
77f2c870e3STao Ma
78198aac28STao Ma trace_ocfs2_validate_refcount_block((unsigned long long)bh->b_blocknr);
79f2c870e3STao Ma
80f2c870e3STao Ma BUG_ON(!buffer_uptodate(bh));
81f2c870e3STao Ma
82f2c870e3STao Ma /*
83f2c870e3STao Ma * If the ecc fails, we return the error but otherwise
84f2c870e3STao Ma * leave the filesystem running. We know any error is
85f2c870e3STao Ma * local to this block.
86f2c870e3STao Ma */
87f2c870e3STao Ma rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check);
88f2c870e3STao Ma if (rc) {
89f2c870e3STao Ma mlog(ML_ERROR, "Checksum failed for refcount block %llu\n",
90f2c870e3STao Ma (unsigned long long)bh->b_blocknr);
91f2c870e3STao Ma return rc;
92f2c870e3STao Ma }
93f2c870e3STao Ma
94f2c870e3STao Ma
95f2c870e3STao Ma if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) {
9617a5b9abSGoldwyn Rodrigues rc = ocfs2_error(sb,
977ecef14aSJoe Perches "Refcount block #%llu has bad signature %.*s\n",
98f2c870e3STao Ma (unsigned long long)bh->b_blocknr, 7,
99f2c870e3STao Ma rb->rf_signature);
10017a5b9abSGoldwyn Rodrigues goto out;
101f2c870e3STao Ma }
102f2c870e3STao Ma
103f2c870e3STao Ma if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
10417a5b9abSGoldwyn Rodrigues rc = ocfs2_error(sb,
1057ecef14aSJoe Perches "Refcount block #%llu has an invalid rf_blkno of %llu\n",
106f2c870e3STao Ma (unsigned long long)bh->b_blocknr,
107f2c870e3STao Ma (unsigned long long)le64_to_cpu(rb->rf_blkno));
10817a5b9abSGoldwyn Rodrigues goto out;
109f2c870e3STao Ma }
110f2c870e3STao Ma
111f2c870e3STao Ma if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
11217a5b9abSGoldwyn Rodrigues rc = ocfs2_error(sb,
1137ecef14aSJoe Perches "Refcount block #%llu has an invalid rf_fs_generation of #%u\n",
114f2c870e3STao Ma (unsigned long long)bh->b_blocknr,
115f2c870e3STao Ma le32_to_cpu(rb->rf_fs_generation));
11617a5b9abSGoldwyn Rodrigues goto out;
117f2c870e3STao Ma }
11817a5b9abSGoldwyn Rodrigues out:
11917a5b9abSGoldwyn Rodrigues return rc;
120f2c870e3STao Ma }
121f2c870e3STao Ma
ocfs2_read_refcount_block(struct ocfs2_caching_info * ci,u64 rb_blkno,struct buffer_head ** bh)122f2c870e3STao Ma static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
123f2c870e3STao Ma u64 rb_blkno,
124f2c870e3STao Ma struct buffer_head **bh)
125f2c870e3STao Ma {
126f2c870e3STao Ma int rc;
127f2c870e3STao Ma struct buffer_head *tmp = *bh;
128f2c870e3STao Ma
129f2c870e3STao Ma rc = ocfs2_read_block(ci, rb_blkno, &tmp,
130f2c870e3STao Ma ocfs2_validate_refcount_block);
131f2c870e3STao Ma
132f2c870e3STao Ma /* If ocfs2_read_block() got us a new bh, pass it up. */
133f2c870e3STao Ma if (!rc && !*bh)
134f2c870e3STao Ma *bh = tmp;
135f2c870e3STao Ma
136f2c870e3STao Ma return rc;
137f2c870e3STao Ma }
138c732eb16STao Ma
ocfs2_refcount_cache_owner(struct ocfs2_caching_info * ci)139c732eb16STao Ma static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
140c732eb16STao Ma {
141c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
142c732eb16STao Ma
143c732eb16STao Ma return rf->rf_blkno;
144c732eb16STao Ma }
145c732eb16STao Ma
146c732eb16STao Ma static struct super_block *
ocfs2_refcount_cache_get_super(struct ocfs2_caching_info * ci)147c732eb16STao Ma ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
148c732eb16STao Ma {
149c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
150c732eb16STao Ma
151c732eb16STao Ma return rf->rf_sb;
152c732eb16STao Ma }
153c732eb16STao Ma
ocfs2_refcount_cache_lock(struct ocfs2_caching_info * ci)154c732eb16STao Ma static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
155185a7321SJules Irenge __acquires(&rf->rf_lock)
156c732eb16STao Ma {
157c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
158c732eb16STao Ma
159c732eb16STao Ma spin_lock(&rf->rf_lock);
160c732eb16STao Ma }
161c732eb16STao Ma
ocfs2_refcount_cache_unlock(struct ocfs2_caching_info * ci)162c732eb16STao Ma static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
163185a7321SJules Irenge __releases(&rf->rf_lock)
164c732eb16STao Ma {
165c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
166c732eb16STao Ma
167c732eb16STao Ma spin_unlock(&rf->rf_lock);
168c732eb16STao Ma }
169c732eb16STao Ma
ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info * ci)170c732eb16STao Ma static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
171c732eb16STao Ma {
172c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
173c732eb16STao Ma
174c732eb16STao Ma mutex_lock(&rf->rf_io_mutex);
175c732eb16STao Ma }
176c732eb16STao Ma
ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info * ci)177c732eb16STao Ma static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
178c732eb16STao Ma {
179c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
180c732eb16STao Ma
181c732eb16STao Ma mutex_unlock(&rf->rf_io_mutex);
182c732eb16STao Ma }
183c732eb16STao Ma
184c732eb16STao Ma static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
185c732eb16STao Ma .co_owner = ocfs2_refcount_cache_owner,
186c732eb16STao Ma .co_get_super = ocfs2_refcount_cache_get_super,
187c732eb16STao Ma .co_cache_lock = ocfs2_refcount_cache_lock,
188c732eb16STao Ma .co_cache_unlock = ocfs2_refcount_cache_unlock,
189c732eb16STao Ma .co_io_lock = ocfs2_refcount_cache_io_lock,
190c732eb16STao Ma .co_io_unlock = ocfs2_refcount_cache_io_unlock,
191c732eb16STao Ma };
192374a263eSTao Ma
193374a263eSTao Ma static struct ocfs2_refcount_tree *
ocfs2_find_refcount_tree(struct ocfs2_super * osb,u64 blkno)194374a263eSTao Ma ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
195374a263eSTao Ma {
196374a263eSTao Ma struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
197374a263eSTao Ma struct ocfs2_refcount_tree *tree = NULL;
198374a263eSTao Ma
199374a263eSTao Ma while (n) {
200374a263eSTao Ma tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
201374a263eSTao Ma
202374a263eSTao Ma if (blkno < tree->rf_blkno)
203374a263eSTao Ma n = n->rb_left;
204374a263eSTao Ma else if (blkno > tree->rf_blkno)
205374a263eSTao Ma n = n->rb_right;
206374a263eSTao Ma else
207374a263eSTao Ma return tree;
208374a263eSTao Ma }
209374a263eSTao Ma
210374a263eSTao Ma return NULL;
211374a263eSTao Ma }
212374a263eSTao Ma
213374a263eSTao Ma /* osb_lock is already locked. */
ocfs2_insert_refcount_tree(struct ocfs2_super * osb,struct ocfs2_refcount_tree * new)214374a263eSTao Ma static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
215374a263eSTao Ma struct ocfs2_refcount_tree *new)
216374a263eSTao Ma {
217374a263eSTao Ma u64 rf_blkno = new->rf_blkno;
218374a263eSTao Ma struct rb_node *parent = NULL;
219374a263eSTao Ma struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
220374a263eSTao Ma struct ocfs2_refcount_tree *tmp;
221374a263eSTao Ma
222374a263eSTao Ma while (*p) {
223374a263eSTao Ma parent = *p;
224374a263eSTao Ma
225374a263eSTao Ma tmp = rb_entry(parent, struct ocfs2_refcount_tree,
226374a263eSTao Ma rf_node);
227374a263eSTao Ma
228374a263eSTao Ma if (rf_blkno < tmp->rf_blkno)
229374a263eSTao Ma p = &(*p)->rb_left;
230374a263eSTao Ma else if (rf_blkno > tmp->rf_blkno)
231374a263eSTao Ma p = &(*p)->rb_right;
232374a263eSTao Ma else {
233374a263eSTao Ma /* This should never happen! */
234374a263eSTao Ma mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
235374a263eSTao Ma (unsigned long long)rf_blkno);
236374a263eSTao Ma BUG();
237374a263eSTao Ma }
238374a263eSTao Ma }
239374a263eSTao Ma
240374a263eSTao Ma rb_link_node(&new->rf_node, parent, p);
241374a263eSTao Ma rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
242374a263eSTao Ma }
243374a263eSTao Ma
ocfs2_free_refcount_tree(struct ocfs2_refcount_tree * tree)244374a263eSTao Ma static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
245374a263eSTao Ma {
246374a263eSTao Ma ocfs2_metadata_cache_exit(&tree->rf_ci);
247374a263eSTao Ma ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
248374a263eSTao Ma ocfs2_lock_res_free(&tree->rf_lockres);
249374a263eSTao Ma kfree(tree);
250374a263eSTao Ma }
251374a263eSTao Ma
252374a263eSTao Ma static inline void
ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super * osb,struct ocfs2_refcount_tree * tree)253374a263eSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
254374a263eSTao Ma struct ocfs2_refcount_tree *tree)
255374a263eSTao Ma {
256374a263eSTao Ma rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
257374a263eSTao Ma if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
258374a263eSTao Ma osb->osb_ref_tree_lru = NULL;
259374a263eSTao Ma }
260374a263eSTao Ma
ocfs2_erase_refcount_tree_from_list(struct ocfs2_super * osb,struct ocfs2_refcount_tree * tree)261374a263eSTao Ma static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
262374a263eSTao Ma struct ocfs2_refcount_tree *tree)
263374a263eSTao Ma {
264374a263eSTao Ma spin_lock(&osb->osb_lock);
265374a263eSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
266374a263eSTao Ma spin_unlock(&osb->osb_lock);
267374a263eSTao Ma }
268374a263eSTao Ma
ocfs2_kref_remove_refcount_tree(struct kref * kref)26912d4cec9STao Ma static void ocfs2_kref_remove_refcount_tree(struct kref *kref)
270374a263eSTao Ma {
271374a263eSTao Ma struct ocfs2_refcount_tree *tree =
272374a263eSTao Ma container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
273374a263eSTao Ma
274374a263eSTao Ma ocfs2_free_refcount_tree(tree);
275374a263eSTao Ma }
276374a263eSTao Ma
277374a263eSTao Ma static inline void
ocfs2_refcount_tree_get(struct ocfs2_refcount_tree * tree)278374a263eSTao Ma ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
279374a263eSTao Ma {
280374a263eSTao Ma kref_get(&tree->rf_getcnt);
281374a263eSTao Ma }
282374a263eSTao Ma
283374a263eSTao Ma static inline void
ocfs2_refcount_tree_put(struct ocfs2_refcount_tree * tree)284374a263eSTao Ma ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
285374a263eSTao Ma {
286374a263eSTao Ma kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
287374a263eSTao Ma }
288374a263eSTao Ma
ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree * new,struct super_block * sb)289374a263eSTao Ma static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
290374a263eSTao Ma struct super_block *sb)
291374a263eSTao Ma {
292374a263eSTao Ma ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
293374a263eSTao Ma mutex_init(&new->rf_io_mutex);
294374a263eSTao Ma new->rf_sb = sb;
295374a263eSTao Ma spin_lock_init(&new->rf_lock);
296374a263eSTao Ma }
297374a263eSTao Ma
ocfs2_init_refcount_tree_lock(struct ocfs2_super * osb,struct ocfs2_refcount_tree * new,u64 rf_blkno,u32 generation)298374a263eSTao Ma static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
299374a263eSTao Ma struct ocfs2_refcount_tree *new,
300374a263eSTao Ma u64 rf_blkno, u32 generation)
301374a263eSTao Ma {
302374a263eSTao Ma init_rwsem(&new->rf_sem);
303374a263eSTao Ma ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
304374a263eSTao Ma rf_blkno, generation);
305374a263eSTao Ma }
306374a263eSTao Ma
3078bf396deSTao Ma static struct ocfs2_refcount_tree*
ocfs2_allocate_refcount_tree(struct ocfs2_super * osb,u64 rf_blkno)3088bf396deSTao Ma ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
3098bf396deSTao Ma {
3108bf396deSTao Ma struct ocfs2_refcount_tree *new;
3118bf396deSTao Ma
3128bf396deSTao Ma new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
3138bf396deSTao Ma if (!new)
3148bf396deSTao Ma return NULL;
3158bf396deSTao Ma
3168bf396deSTao Ma new->rf_blkno = rf_blkno;
3178bf396deSTao Ma kref_init(&new->rf_getcnt);
3188bf396deSTao Ma ocfs2_init_refcount_tree_ci(new, osb->sb);
3198bf396deSTao Ma
3208bf396deSTao Ma return new;
3218bf396deSTao Ma }
3228bf396deSTao Ma
ocfs2_get_refcount_tree(struct ocfs2_super * osb,u64 rf_blkno,struct ocfs2_refcount_tree ** ret_tree)323374a263eSTao Ma static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
324374a263eSTao Ma struct ocfs2_refcount_tree **ret_tree)
325374a263eSTao Ma {
326374a263eSTao Ma int ret = 0;
327374a263eSTao Ma struct ocfs2_refcount_tree *tree, *new = NULL;
328374a263eSTao Ma struct buffer_head *ref_root_bh = NULL;
329374a263eSTao Ma struct ocfs2_refcount_block *ref_rb;
330374a263eSTao Ma
331374a263eSTao Ma spin_lock(&osb->osb_lock);
332374a263eSTao Ma if (osb->osb_ref_tree_lru &&
333374a263eSTao Ma osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
334374a263eSTao Ma tree = osb->osb_ref_tree_lru;
335374a263eSTao Ma else
336374a263eSTao Ma tree = ocfs2_find_refcount_tree(osb, rf_blkno);
337374a263eSTao Ma if (tree)
338374a263eSTao Ma goto out;
339374a263eSTao Ma
340374a263eSTao Ma spin_unlock(&osb->osb_lock);
341374a263eSTao Ma
3428bf396deSTao Ma new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
343374a263eSTao Ma if (!new) {
344374a263eSTao Ma ret = -ENOMEM;
3458bf396deSTao Ma mlog_errno(ret);
346374a263eSTao Ma return ret;
347374a263eSTao Ma }
348374a263eSTao Ma /*
349374a263eSTao Ma * We need the generation to create the refcount tree lock and since
350374a263eSTao Ma * it isn't changed during the tree modification, we are safe here to
351374a263eSTao Ma * read without protection.
352374a263eSTao Ma * We also have to purge the cache after we create the lock since the
353374a263eSTao Ma * refcount block may have the stale data. It can only be trusted when
354374a263eSTao Ma * we hold the refcount lock.
355374a263eSTao Ma */
356374a263eSTao Ma ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
357374a263eSTao Ma if (ret) {
358374a263eSTao Ma mlog_errno(ret);
359374a263eSTao Ma ocfs2_metadata_cache_exit(&new->rf_ci);
360374a263eSTao Ma kfree(new);
361374a263eSTao Ma return ret;
362374a263eSTao Ma }
363374a263eSTao Ma
364374a263eSTao Ma ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
365374a263eSTao Ma new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
366374a263eSTao Ma ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
367374a263eSTao Ma new->rf_generation);
368374a263eSTao Ma ocfs2_metadata_cache_purge(&new->rf_ci);
369374a263eSTao Ma
370374a263eSTao Ma spin_lock(&osb->osb_lock);
371374a263eSTao Ma tree = ocfs2_find_refcount_tree(osb, rf_blkno);
372374a263eSTao Ma if (tree)
373374a263eSTao Ma goto out;
374374a263eSTao Ma
375374a263eSTao Ma ocfs2_insert_refcount_tree(osb, new);
376374a263eSTao Ma
377374a263eSTao Ma tree = new;
378374a263eSTao Ma new = NULL;
379374a263eSTao Ma
380374a263eSTao Ma out:
381374a263eSTao Ma *ret_tree = tree;
382374a263eSTao Ma
383374a263eSTao Ma osb->osb_ref_tree_lru = tree;
384374a263eSTao Ma
385374a263eSTao Ma spin_unlock(&osb->osb_lock);
386374a263eSTao Ma
387374a263eSTao Ma if (new)
388374a263eSTao Ma ocfs2_free_refcount_tree(new);
389374a263eSTao Ma
390374a263eSTao Ma brelse(ref_root_bh);
391374a263eSTao Ma return ret;
392374a263eSTao Ma }
393374a263eSTao Ma
ocfs2_get_refcount_block(struct inode * inode,u64 * ref_blkno)394374a263eSTao Ma static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
395374a263eSTao Ma {
396374a263eSTao Ma int ret;
397374a263eSTao Ma struct buffer_head *di_bh = NULL;
398374a263eSTao Ma struct ocfs2_dinode *di;
399374a263eSTao Ma
400374a263eSTao Ma ret = ocfs2_read_inode_block(inode, &di_bh);
401374a263eSTao Ma if (ret) {
402374a263eSTao Ma mlog_errno(ret);
403374a263eSTao Ma goto out;
404374a263eSTao Ma }
405374a263eSTao Ma
40684e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode));
407374a263eSTao Ma
408374a263eSTao Ma di = (struct ocfs2_dinode *)di_bh->b_data;
409374a263eSTao Ma *ref_blkno = le64_to_cpu(di->i_refcount_loc);
410374a263eSTao Ma brelse(di_bh);
411374a263eSTao Ma out:
412374a263eSTao Ma return ret;
413374a263eSTao Ma }
414374a263eSTao Ma
__ocfs2_lock_refcount_tree(struct ocfs2_super * osb,struct ocfs2_refcount_tree * tree,int rw)415374a263eSTao Ma static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
416374a263eSTao Ma struct ocfs2_refcount_tree *tree, int rw)
417374a263eSTao Ma {
418374a263eSTao Ma int ret;
419374a263eSTao Ma
420374a263eSTao Ma ret = ocfs2_refcount_lock(tree, rw);
421374a263eSTao Ma if (ret) {
422374a263eSTao Ma mlog_errno(ret);
423374a263eSTao Ma goto out;
424374a263eSTao Ma }
425374a263eSTao Ma
426374a263eSTao Ma if (rw)
427374a263eSTao Ma down_write(&tree->rf_sem);
428374a263eSTao Ma else
429374a263eSTao Ma down_read(&tree->rf_sem);
430374a263eSTao Ma
431374a263eSTao Ma out:
432374a263eSTao Ma return ret;
433374a263eSTao Ma }
434374a263eSTao Ma
435374a263eSTao Ma /*
436374a263eSTao Ma * Lock the refcount tree pointed by ref_blkno and return the tree.
437374a263eSTao Ma * In most case, we lock the tree and read the refcount block.
438374a263eSTao Ma * So read it here if the caller really needs it.
439374a263eSTao Ma *
440374a263eSTao Ma * If the tree has been re-created by other node, it will free the
441374a263eSTao Ma * old one and re-create it.
442374a263eSTao Ma */
ocfs2_lock_refcount_tree(struct ocfs2_super * osb,u64 ref_blkno,int rw,struct ocfs2_refcount_tree ** ret_tree,struct buffer_head ** ref_bh)443374a263eSTao Ma int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
444374a263eSTao Ma u64 ref_blkno, int rw,
445374a263eSTao Ma struct ocfs2_refcount_tree **ret_tree,
446374a263eSTao Ma struct buffer_head **ref_bh)
447374a263eSTao Ma {
448374a263eSTao Ma int ret, delete_tree = 0;
449374a263eSTao Ma struct ocfs2_refcount_tree *tree = NULL;
450374a263eSTao Ma struct buffer_head *ref_root_bh = NULL;
451374a263eSTao Ma struct ocfs2_refcount_block *rb;
452374a263eSTao Ma
453374a263eSTao Ma again:
454374a263eSTao Ma ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
455374a263eSTao Ma if (ret) {
456374a263eSTao Ma mlog_errno(ret);
457374a263eSTao Ma return ret;
458374a263eSTao Ma }
459374a263eSTao Ma
460374a263eSTao Ma ocfs2_refcount_tree_get(tree);
461374a263eSTao Ma
462374a263eSTao Ma ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
463374a263eSTao Ma if (ret) {
464374a263eSTao Ma mlog_errno(ret);
465374a263eSTao Ma ocfs2_refcount_tree_put(tree);
466374a263eSTao Ma goto out;
467374a263eSTao Ma }
468374a263eSTao Ma
469374a263eSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
470374a263eSTao Ma &ref_root_bh);
471374a263eSTao Ma if (ret) {
472374a263eSTao Ma mlog_errno(ret);
473374a263eSTao Ma ocfs2_unlock_refcount_tree(osb, tree, rw);
474374a263eSTao Ma goto out;
475374a263eSTao Ma }
476374a263eSTao Ma
477374a263eSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
478374a263eSTao Ma /*
479374a263eSTao Ma * If the refcount block has been freed and re-created, we may need
480374a263eSTao Ma * to recreate the refcount tree also.
481374a263eSTao Ma *
482374a263eSTao Ma * Here we just remove the tree from the rb-tree, and the last
483374a263eSTao Ma * kref holder will unlock and delete this refcount_tree.
484374a263eSTao Ma * Then we goto "again" and ocfs2_get_refcount_tree will create
485374a263eSTao Ma * the new refcount tree for us.
486374a263eSTao Ma */
487374a263eSTao Ma if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
488374a263eSTao Ma if (!tree->rf_removed) {
489374a263eSTao Ma ocfs2_erase_refcount_tree_from_list(osb, tree);
490374a263eSTao Ma tree->rf_removed = 1;
491374a263eSTao Ma delete_tree = 1;
492374a263eSTao Ma }
493374a263eSTao Ma
494374a263eSTao Ma ocfs2_unlock_refcount_tree(osb, tree, rw);
495374a263eSTao Ma /*
496374a263eSTao Ma * We get an extra reference when we create the refcount
497374a263eSTao Ma * tree, so another put will destroy it.
498374a263eSTao Ma */
499374a263eSTao Ma if (delete_tree)
500374a263eSTao Ma ocfs2_refcount_tree_put(tree);
501374a263eSTao Ma brelse(ref_root_bh);
502374a263eSTao Ma ref_root_bh = NULL;
503374a263eSTao Ma goto again;
504374a263eSTao Ma }
505374a263eSTao Ma
506374a263eSTao Ma *ret_tree = tree;
507374a263eSTao Ma if (ref_bh) {
508374a263eSTao Ma *ref_bh = ref_root_bh;
509374a263eSTao Ma ref_root_bh = NULL;
510374a263eSTao Ma }
511374a263eSTao Ma out:
512374a263eSTao Ma brelse(ref_root_bh);
513374a263eSTao Ma return ret;
514374a263eSTao Ma }
515374a263eSTao Ma
ocfs2_unlock_refcount_tree(struct ocfs2_super * osb,struct ocfs2_refcount_tree * tree,int rw)516374a263eSTao Ma void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
517374a263eSTao Ma struct ocfs2_refcount_tree *tree, int rw)
518374a263eSTao Ma {
519374a263eSTao Ma if (rw)
520374a263eSTao Ma up_write(&tree->rf_sem);
521374a263eSTao Ma else
522374a263eSTao Ma up_read(&tree->rf_sem);
523374a263eSTao Ma
524374a263eSTao Ma ocfs2_refcount_unlock(tree, rw);
525374a263eSTao Ma ocfs2_refcount_tree_put(tree);
526374a263eSTao Ma }
527374a263eSTao Ma
ocfs2_purge_refcount_trees(struct ocfs2_super * osb)528374a263eSTao Ma void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
529374a263eSTao Ma {
530374a263eSTao Ma struct rb_node *node;
531374a263eSTao Ma struct ocfs2_refcount_tree *tree;
532374a263eSTao Ma struct rb_root *root = &osb->osb_rf_lock_tree;
533374a263eSTao Ma
534374a263eSTao Ma while ((node = rb_last(root)) != NULL) {
535374a263eSTao Ma tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
536374a263eSTao Ma
537198aac28STao Ma trace_ocfs2_purge_refcount_trees(
538374a263eSTao Ma (unsigned long long) tree->rf_blkno);
539374a263eSTao Ma
540374a263eSTao Ma rb_erase(&tree->rf_node, root);
541374a263eSTao Ma ocfs2_free_refcount_tree(tree);
542374a263eSTao Ma }
543374a263eSTao Ma }
5448bf396deSTao Ma
5458bf396deSTao Ma /*
5468bf396deSTao Ma * Create a refcount tree for an inode.
5478bf396deSTao Ma * We take for granted that the inode is already locked.
5488bf396deSTao Ma */
ocfs2_create_refcount_tree(struct inode * inode,struct buffer_head * di_bh)5498bf396deSTao Ma static int ocfs2_create_refcount_tree(struct inode *inode,
5508bf396deSTao Ma struct buffer_head *di_bh)
5518bf396deSTao Ma {
5528bf396deSTao Ma int ret;
5538bf396deSTao Ma handle_t *handle = NULL;
5548bf396deSTao Ma struct ocfs2_alloc_context *meta_ac = NULL;
5558bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
5568bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode);
5578bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5588bf396deSTao Ma struct buffer_head *new_bh = NULL;
5598bf396deSTao Ma struct ocfs2_refcount_block *rb;
5608bf396deSTao Ma struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
5618bf396deSTao Ma u16 suballoc_bit_start;
5628bf396deSTao Ma u32 num_got;
5632b6cb576SJoel Becker u64 suballoc_loc, first_blkno;
5648bf396deSTao Ma
56584e40080SDarrick J. Wong BUG_ON(ocfs2_is_refcount_inode(inode));
5668bf396deSTao Ma
567198aac28STao Ma trace_ocfs2_create_refcount_tree(
568d324cd4cSpiaojun (unsigned long long)oi->ip_blkno);
5698bf396deSTao Ma
5708bf396deSTao Ma ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
5718bf396deSTao Ma if (ret) {
5728bf396deSTao Ma mlog_errno(ret);
5738bf396deSTao Ma goto out;
5748bf396deSTao Ma }
5758bf396deSTao Ma
5768bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
5778bf396deSTao Ma if (IS_ERR(handle)) {
5788bf396deSTao Ma ret = PTR_ERR(handle);
5798bf396deSTao Ma mlog_errno(ret);
5808bf396deSTao Ma goto out;
5818bf396deSTao Ma }
5828bf396deSTao Ma
5838bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
5848bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
5858bf396deSTao Ma if (ret) {
5868bf396deSTao Ma mlog_errno(ret);
5878bf396deSTao Ma goto out_commit;
5888bf396deSTao Ma }
5898bf396deSTao Ma
5902b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
5918bf396deSTao Ma &suballoc_bit_start, &num_got,
5928bf396deSTao Ma &first_blkno);
5938bf396deSTao Ma if (ret) {
5948bf396deSTao Ma mlog_errno(ret);
5958bf396deSTao Ma goto out_commit;
5968bf396deSTao Ma }
5978bf396deSTao Ma
5988bf396deSTao Ma new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
5998bf396deSTao Ma if (!new_tree) {
6008bf396deSTao Ma ret = -ENOMEM;
6018bf396deSTao Ma mlog_errno(ret);
6028bf396deSTao Ma goto out_commit;
6038bf396deSTao Ma }
6048bf396deSTao Ma
6058bf396deSTao Ma new_bh = sb_getblk(inode->i_sb, first_blkno);
60658796207SRui Xiang if (!new_bh) {
60758796207SRui Xiang ret = -ENOMEM;
60858796207SRui Xiang mlog_errno(ret);
60958796207SRui Xiang goto out_commit;
61058796207SRui Xiang }
6118bf396deSTao Ma ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
6128bf396deSTao Ma
6138bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
6148bf396deSTao Ma OCFS2_JOURNAL_ACCESS_CREATE);
6158bf396deSTao Ma if (ret) {
6168bf396deSTao Ma mlog_errno(ret);
6178bf396deSTao Ma goto out_commit;
6188bf396deSTao Ma }
6198bf396deSTao Ma
6208bf396deSTao Ma /* Initialize ocfs2_refcount_block. */
6218bf396deSTao Ma rb = (struct ocfs2_refcount_block *)new_bh->b_data;
6228bf396deSTao Ma memset(rb, 0, inode->i_sb->s_blocksize);
6238bf396deSTao Ma strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
624b89c5428STiger Yang rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
6252b6cb576SJoel Becker rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
6268bf396deSTao Ma rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
6278bf396deSTao Ma rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
6288bf396deSTao Ma rb->rf_blkno = cpu_to_le64(first_blkno);
6298bf396deSTao Ma rb->rf_count = cpu_to_le32(1);
6308bf396deSTao Ma rb->rf_records.rl_count =
6318bf396deSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
6328bf396deSTao Ma spin_lock(&osb->osb_lock);
6338bf396deSTao Ma rb->rf_generation = osb->s_next_generation++;
6348bf396deSTao Ma spin_unlock(&osb->osb_lock);
6358bf396deSTao Ma
6368bf396deSTao Ma ocfs2_journal_dirty(handle, new_bh);
6378bf396deSTao Ma
6388bf396deSTao Ma spin_lock(&oi->ip_lock);
6398bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
6408bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6418bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(first_blkno);
6428bf396deSTao Ma spin_unlock(&oi->ip_lock);
6438bf396deSTao Ma
644198aac28STao Ma trace_ocfs2_create_refcount_tree_blkno((unsigned long long)first_blkno);
6458bf396deSTao Ma
6468bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh);
6478bf396deSTao Ma
6488bf396deSTao Ma /*
6498bf396deSTao Ma * We have to init the tree lock here since it will use
6508bf396deSTao Ma * the generation number to create it.
6518bf396deSTao Ma */
6528bf396deSTao Ma new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
6538bf396deSTao Ma ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
6548bf396deSTao Ma new_tree->rf_generation);
6558bf396deSTao Ma
6568bf396deSTao Ma spin_lock(&osb->osb_lock);
6578bf396deSTao Ma tree = ocfs2_find_refcount_tree(osb, first_blkno);
6588bf396deSTao Ma
6598bf396deSTao Ma /*
6608bf396deSTao Ma * We've just created a new refcount tree in this block. If
6618bf396deSTao Ma * we found a refcount tree on the ocfs2_super, it must be
6628bf396deSTao Ma * one we just deleted. We free the old tree before
6638bf396deSTao Ma * inserting the new tree.
6648bf396deSTao Ma */
6658bf396deSTao Ma BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
6668bf396deSTao Ma if (tree)
6678bf396deSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
6688bf396deSTao Ma ocfs2_insert_refcount_tree(osb, new_tree);
6698bf396deSTao Ma spin_unlock(&osb->osb_lock);
6708bf396deSTao Ma new_tree = NULL;
6718bf396deSTao Ma if (tree)
6728bf396deSTao Ma ocfs2_refcount_tree_put(tree);
6738bf396deSTao Ma
6748bf396deSTao Ma out_commit:
6758bf396deSTao Ma ocfs2_commit_trans(osb, handle);
6768bf396deSTao Ma
6778bf396deSTao Ma out:
6788bf396deSTao Ma if (new_tree) {
6798bf396deSTao Ma ocfs2_metadata_cache_exit(&new_tree->rf_ci);
6808bf396deSTao Ma kfree(new_tree);
6818bf396deSTao Ma }
6828bf396deSTao Ma
6838bf396deSTao Ma brelse(new_bh);
6848bf396deSTao Ma if (meta_ac)
6858bf396deSTao Ma ocfs2_free_alloc_context(meta_ac);
6868bf396deSTao Ma
6878bf396deSTao Ma return ret;
6888bf396deSTao Ma }
6898bf396deSTao Ma
ocfs2_set_refcount_tree(struct inode * inode,struct buffer_head * di_bh,u64 refcount_loc)6908bf396deSTao Ma static int ocfs2_set_refcount_tree(struct inode *inode,
6918bf396deSTao Ma struct buffer_head *di_bh,
6928bf396deSTao Ma u64 refcount_loc)
6938bf396deSTao Ma {
6948bf396deSTao Ma int ret;
6958bf396deSTao Ma handle_t *handle = NULL;
6968bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6978bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode);
6988bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6998bf396deSTao Ma struct buffer_head *ref_root_bh = NULL;
7008bf396deSTao Ma struct ocfs2_refcount_block *rb;
7018bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree;
7028bf396deSTao Ma
70384e40080SDarrick J. Wong BUG_ON(ocfs2_is_refcount_inode(inode));
7048bf396deSTao Ma
7058bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7068bf396deSTao Ma &ref_tree, &ref_root_bh);
7078bf396deSTao Ma if (ret) {
7088bf396deSTao Ma mlog_errno(ret);
7098bf396deSTao Ma return ret;
7108bf396deSTao Ma }
7118bf396deSTao Ma
7128bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
7138bf396deSTao Ma if (IS_ERR(handle)) {
7148bf396deSTao Ma ret = PTR_ERR(handle);
7158bf396deSTao Ma mlog_errno(ret);
7168bf396deSTao Ma goto out;
7178bf396deSTao Ma }
7188bf396deSTao Ma
7198bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7208bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
7218bf396deSTao Ma if (ret) {
7228bf396deSTao Ma mlog_errno(ret);
7238bf396deSTao Ma goto out_commit;
7248bf396deSTao Ma }
7258bf396deSTao Ma
7268bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
7278bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
7288bf396deSTao Ma if (ret) {
7298bf396deSTao Ma mlog_errno(ret);
7308bf396deSTao Ma goto out_commit;
7318bf396deSTao Ma }
7328bf396deSTao Ma
7338bf396deSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
7348bf396deSTao Ma le32_add_cpu(&rb->rf_count, 1);
7358bf396deSTao Ma
7368bf396deSTao Ma ocfs2_journal_dirty(handle, ref_root_bh);
7378bf396deSTao Ma
7388bf396deSTao Ma spin_lock(&oi->ip_lock);
7398bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
7408bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7418bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(refcount_loc);
7428bf396deSTao Ma spin_unlock(&oi->ip_lock);
7438bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh);
7448bf396deSTao Ma
7458bf396deSTao Ma out_commit:
7468bf396deSTao Ma ocfs2_commit_trans(osb, handle);
7478bf396deSTao Ma out:
7488bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7498bf396deSTao Ma brelse(ref_root_bh);
7508bf396deSTao Ma
7518bf396deSTao Ma return ret;
7528bf396deSTao Ma }
7538bf396deSTao Ma
ocfs2_remove_refcount_tree(struct inode * inode,struct buffer_head * di_bh)7548bf396deSTao Ma int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
7558bf396deSTao Ma {
7568bf396deSTao Ma int ret, delete_tree = 0;
7578bf396deSTao Ma handle_t *handle = NULL;
7588bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7598bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode);
7608bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7618bf396deSTao Ma struct ocfs2_refcount_block *rb;
7628bf396deSTao Ma struct inode *alloc_inode = NULL;
7638bf396deSTao Ma struct buffer_head *alloc_bh = NULL;
7648bf396deSTao Ma struct buffer_head *blk_bh = NULL;
7658bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree;
7668bf396deSTao Ma int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
7678bf396deSTao Ma u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
7688bf396deSTao Ma u16 bit = 0;
7698bf396deSTao Ma
77084e40080SDarrick J. Wong if (!ocfs2_is_refcount_inode(inode))
7718bf396deSTao Ma return 0;
7728bf396deSTao Ma
7738bf396deSTao Ma BUG_ON(!ref_blkno);
7748bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
7758bf396deSTao Ma if (ret) {
7768bf396deSTao Ma mlog_errno(ret);
7778bf396deSTao Ma return ret;
7788bf396deSTao Ma }
7798bf396deSTao Ma
7808bf396deSTao Ma rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
7818bf396deSTao Ma
7828bf396deSTao Ma /*
7838bf396deSTao Ma * If we are the last user, we need to free the block.
7848bf396deSTao Ma * So lock the allocator ahead.
7858bf396deSTao Ma */
7868bf396deSTao Ma if (le32_to_cpu(rb->rf_count) == 1) {
7878bf396deSTao Ma blk = le64_to_cpu(rb->rf_blkno);
7888bf396deSTao Ma bit = le16_to_cpu(rb->rf_suballoc_bit);
78974380c47STao Ma if (rb->rf_suballoc_loc)
79074380c47STao Ma bg_blkno = le64_to_cpu(rb->rf_suballoc_loc);
79174380c47STao Ma else
7928bf396deSTao Ma bg_blkno = ocfs2_which_suballoc_group(blk, bit);
7938bf396deSTao Ma
7948bf396deSTao Ma alloc_inode = ocfs2_get_system_file_inode(osb,
7958bf396deSTao Ma EXTENT_ALLOC_SYSTEM_INODE,
7968bf396deSTao Ma le16_to_cpu(rb->rf_suballoc_slot));
7978bf396deSTao Ma if (!alloc_inode) {
7988bf396deSTao Ma ret = -ENOMEM;
7998bf396deSTao Ma mlog_errno(ret);
8008bf396deSTao Ma goto out;
8018bf396deSTao Ma }
8025955102cSAl Viro inode_lock(alloc_inode);
8038bf396deSTao Ma
8048bf396deSTao Ma ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
8058bf396deSTao Ma if (ret) {
8068bf396deSTao Ma mlog_errno(ret);
8078bf396deSTao Ma goto out_mutex;
8088bf396deSTao Ma }
8098bf396deSTao Ma
8108bf396deSTao Ma credits += OCFS2_SUBALLOC_FREE;
8118bf396deSTao Ma }
8128bf396deSTao Ma
8138bf396deSTao Ma handle = ocfs2_start_trans(osb, credits);
8148bf396deSTao Ma if (IS_ERR(handle)) {
8158bf396deSTao Ma ret = PTR_ERR(handle);
8168bf396deSTao Ma mlog_errno(ret);
8178bf396deSTao Ma goto out_unlock;
8188bf396deSTao Ma }
8198bf396deSTao Ma
8208bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
8218bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
8228bf396deSTao Ma if (ret) {
8238bf396deSTao Ma mlog_errno(ret);
8248bf396deSTao Ma goto out_commit;
8258bf396deSTao Ma }
8268bf396deSTao Ma
8278bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
8288bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
8298bf396deSTao Ma if (ret) {
8308bf396deSTao Ma mlog_errno(ret);
8318bf396deSTao Ma goto out_commit;
8328bf396deSTao Ma }
8338bf396deSTao Ma
8348bf396deSTao Ma spin_lock(&oi->ip_lock);
8358bf396deSTao Ma oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
8368bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
8378bf396deSTao Ma di->i_refcount_loc = 0;
8388bf396deSTao Ma spin_unlock(&oi->ip_lock);
8398bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh);
8408bf396deSTao Ma
8418bf396deSTao Ma le32_add_cpu(&rb->rf_count , -1);
8428bf396deSTao Ma ocfs2_journal_dirty(handle, blk_bh);
8438bf396deSTao Ma
8448bf396deSTao Ma if (!rb->rf_count) {
8458bf396deSTao Ma delete_tree = 1;
8468bf396deSTao Ma ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
8478bf396deSTao Ma ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
8488bf396deSTao Ma alloc_bh, bit, bg_blkno, 1);
8498bf396deSTao Ma if (ret)
8508bf396deSTao Ma mlog_errno(ret);
8518bf396deSTao Ma }
8528bf396deSTao Ma
8538bf396deSTao Ma out_commit:
8548bf396deSTao Ma ocfs2_commit_trans(osb, handle);
8558bf396deSTao Ma out_unlock:
8568bf396deSTao Ma if (alloc_inode) {
8578bf396deSTao Ma ocfs2_inode_unlock(alloc_inode, 1);
8588bf396deSTao Ma brelse(alloc_bh);
8598bf396deSTao Ma }
8608bf396deSTao Ma out_mutex:
8618bf396deSTao Ma if (alloc_inode) {
8625955102cSAl Viro inode_unlock(alloc_inode);
8638bf396deSTao Ma iput(alloc_inode);
8648bf396deSTao Ma }
8658bf396deSTao Ma out:
8668bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
8678bf396deSTao Ma if (delete_tree)
8688bf396deSTao Ma ocfs2_refcount_tree_put(ref_tree);
8698bf396deSTao Ma brelse(blk_bh);
8708bf396deSTao Ma
8718bf396deSTao Ma return ret;
8728bf396deSTao Ma }
873e73a819dSTao Ma
ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info * ci,struct buffer_head * ref_leaf_bh,u64 cpos,unsigned int len,struct ocfs2_refcount_rec * ret_rec,int * index)874e73a819dSTao Ma static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
875e73a819dSTao Ma struct buffer_head *ref_leaf_bh,
876e73a819dSTao Ma u64 cpos, unsigned int len,
877e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec,
878e73a819dSTao Ma int *index)
879e73a819dSTao Ma {
880e73a819dSTao Ma int i = 0;
881e73a819dSTao Ma struct ocfs2_refcount_block *rb =
882e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
883e73a819dSTao Ma struct ocfs2_refcount_rec *rec = NULL;
884e73a819dSTao Ma
885e73a819dSTao Ma for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
886e73a819dSTao Ma rec = &rb->rf_records.rl_recs[i];
887e73a819dSTao Ma
888e73a819dSTao Ma if (le64_to_cpu(rec->r_cpos) +
889e73a819dSTao Ma le32_to_cpu(rec->r_clusters) <= cpos)
890e73a819dSTao Ma continue;
891e73a819dSTao Ma else if (le64_to_cpu(rec->r_cpos) > cpos)
892e73a819dSTao Ma break;
893e73a819dSTao Ma
894e73a819dSTao Ma /* ok, cpos fail in this rec. Just return. */
895e73a819dSTao Ma if (ret_rec)
896e73a819dSTao Ma *ret_rec = *rec;
897e73a819dSTao Ma goto out;
898e73a819dSTao Ma }
899e73a819dSTao Ma
900e73a819dSTao Ma if (ret_rec) {
901e73a819dSTao Ma /* We meet with a hole here, so fake the rec. */
902e73a819dSTao Ma ret_rec->r_cpos = cpu_to_le64(cpos);
903e73a819dSTao Ma ret_rec->r_refcount = 0;
904e73a819dSTao Ma if (i < le16_to_cpu(rb->rf_records.rl_used) &&
905e73a819dSTao Ma le64_to_cpu(rec->r_cpos) < cpos + len)
906e73a819dSTao Ma ret_rec->r_clusters =
907e73a819dSTao Ma cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
908e73a819dSTao Ma else
909e73a819dSTao Ma ret_rec->r_clusters = cpu_to_le32(len);
910e73a819dSTao Ma }
911e73a819dSTao Ma
912e73a819dSTao Ma out:
913e73a819dSTao Ma *index = i;
914e73a819dSTao Ma }
915e73a819dSTao Ma
916e73a819dSTao Ma /*
9178b2c0dbaSTao Ma * Try to remove refcount tree. The mechanism is:
9188b2c0dbaSTao Ma * 1) Check whether i_clusters == 0, if no, exit.
9198b2c0dbaSTao Ma * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
9208b2c0dbaSTao Ma * 3) Check whether we have inline xattr stored outside, if yes, exit.
9218b2c0dbaSTao Ma * 4) Remove the tree.
9228b2c0dbaSTao Ma */
ocfs2_try_remove_refcount_tree(struct inode * inode,struct buffer_head * di_bh)9238b2c0dbaSTao Ma int ocfs2_try_remove_refcount_tree(struct inode *inode,
9248b2c0dbaSTao Ma struct buffer_head *di_bh)
9258b2c0dbaSTao Ma {
9268b2c0dbaSTao Ma int ret;
9278b2c0dbaSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode);
9288b2c0dbaSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
9298b2c0dbaSTao Ma
9308b2c0dbaSTao Ma down_write(&oi->ip_xattr_sem);
9318b2c0dbaSTao Ma down_write(&oi->ip_alloc_sem);
9328b2c0dbaSTao Ma
9338b2c0dbaSTao Ma if (oi->ip_clusters)
9348b2c0dbaSTao Ma goto out;
9358b2c0dbaSTao Ma
9368b2c0dbaSTao Ma if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
9378b2c0dbaSTao Ma goto out;
9388b2c0dbaSTao Ma
9398b2c0dbaSTao Ma if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
9408b2c0dbaSTao Ma ocfs2_has_inline_xattr_value_outside(inode, di))
9418b2c0dbaSTao Ma goto out;
9428b2c0dbaSTao Ma
9438b2c0dbaSTao Ma ret = ocfs2_remove_refcount_tree(inode, di_bh);
9448b2c0dbaSTao Ma if (ret)
9458b2c0dbaSTao Ma mlog_errno(ret);
9468b2c0dbaSTao Ma out:
9478b2c0dbaSTao Ma up_write(&oi->ip_alloc_sem);
9488b2c0dbaSTao Ma up_write(&oi->ip_xattr_sem);
9498b2c0dbaSTao Ma return 0;
9508b2c0dbaSTao Ma }
9518b2c0dbaSTao Ma
9528b2c0dbaSTao Ma /*
95338a04e43STao Ma * Find the end range for a leaf refcount block indicated by
95438a04e43STao Ma * el->l_recs[index].e_blkno.
95538a04e43STao Ma */
ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct ocfs2_extent_block * eb,struct ocfs2_extent_list * el,int index,u32 * cpos_end)95638a04e43STao Ma static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci,
95738a04e43STao Ma struct buffer_head *ref_root_bh,
95838a04e43STao Ma struct ocfs2_extent_block *eb,
95938a04e43STao Ma struct ocfs2_extent_list *el,
96038a04e43STao Ma int index, u32 *cpos_end)
96138a04e43STao Ma {
96238a04e43STao Ma int ret, i, subtree_root;
96338a04e43STao Ma u32 cpos;
96438a04e43STao Ma u64 blkno;
96538a04e43STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
96638a04e43STao Ma struct ocfs2_path *left_path = NULL, *right_path = NULL;
96738a04e43STao Ma struct ocfs2_extent_tree et;
96838a04e43STao Ma struct ocfs2_extent_list *tmp_el;
96938a04e43STao Ma
97038a04e43STao Ma if (index < le16_to_cpu(el->l_next_free_rec) - 1) {
97138a04e43STao Ma /*
97238a04e43STao Ma * We have a extent rec after index, so just use the e_cpos
97338a04e43STao Ma * of the next extent rec.
97438a04e43STao Ma */
97538a04e43STao Ma *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos);
97638a04e43STao Ma return 0;
97738a04e43STao Ma }
97838a04e43STao Ma
9797c908aecSJiapeng Chong if (!eb || !eb->h_next_leaf_blk) {
98038a04e43STao Ma /*
98138a04e43STao Ma * We are the last extent rec, so any high cpos should
98238a04e43STao Ma * be stored in this leaf refcount block.
98338a04e43STao Ma */
98438a04e43STao Ma *cpos_end = UINT_MAX;
98538a04e43STao Ma return 0;
98638a04e43STao Ma }
98738a04e43STao Ma
98838a04e43STao Ma /*
98938a04e43STao Ma * If the extent block isn't the last one, we have to find
99038a04e43STao Ma * the subtree root between this extent block and the next
99138a04e43STao Ma * leaf extent block and get the corresponding e_cpos from
99238a04e43STao Ma * the subroot. Otherwise we may corrupt the b-tree.
99338a04e43STao Ma */
99438a04e43STao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
99538a04e43STao Ma
99638a04e43STao Ma left_path = ocfs2_new_path_from_et(&et);
99738a04e43STao Ma if (!left_path) {
99838a04e43STao Ma ret = -ENOMEM;
99938a04e43STao Ma mlog_errno(ret);
100038a04e43STao Ma goto out;
100138a04e43STao Ma }
100238a04e43STao Ma
100338a04e43STao Ma cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos);
100438a04e43STao Ma ret = ocfs2_find_path(ci, left_path, cpos);
100538a04e43STao Ma if (ret) {
100638a04e43STao Ma mlog_errno(ret);
100738a04e43STao Ma goto out;
100838a04e43STao Ma }
100938a04e43STao Ma
101038a04e43STao Ma right_path = ocfs2_new_path_from_path(left_path);
101138a04e43STao Ma if (!right_path) {
101238a04e43STao Ma ret = -ENOMEM;
101338a04e43STao Ma mlog_errno(ret);
101438a04e43STao Ma goto out;
101538a04e43STao Ma }
101638a04e43STao Ma
101738a04e43STao Ma ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos);
101838a04e43STao Ma if (ret) {
101938a04e43STao Ma mlog_errno(ret);
102038a04e43STao Ma goto out;
102138a04e43STao Ma }
102238a04e43STao Ma
102338a04e43STao Ma ret = ocfs2_find_path(ci, right_path, cpos);
102438a04e43STao Ma if (ret) {
102538a04e43STao Ma mlog_errno(ret);
102638a04e43STao Ma goto out;
102738a04e43STao Ma }
102838a04e43STao Ma
102938a04e43STao Ma subtree_root = ocfs2_find_subtree_root(&et, left_path,
103038a04e43STao Ma right_path);
103138a04e43STao Ma
103238a04e43STao Ma tmp_el = left_path->p_node[subtree_root].el;
103338a04e43STao Ma blkno = left_path->p_node[subtree_root+1].bh->b_blocknr;
10343a251f04SAl Viro for (i = 0; i < le16_to_cpu(tmp_el->l_next_free_rec); i++) {
103538a04e43STao Ma if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) {
103638a04e43STao Ma *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos);
103738a04e43STao Ma break;
103838a04e43STao Ma }
103938a04e43STao Ma }
104038a04e43STao Ma
10413a251f04SAl Viro BUG_ON(i == le16_to_cpu(tmp_el->l_next_free_rec));
104238a04e43STao Ma
104338a04e43STao Ma out:
104438a04e43STao Ma ocfs2_free_path(left_path);
104538a04e43STao Ma ocfs2_free_path(right_path);
104638a04e43STao Ma return ret;
104738a04e43STao Ma }
104838a04e43STao Ma
104938a04e43STao Ma /*
1050e73a819dSTao Ma * Given a cpos and len, try to find the refcount record which contains cpos.
1051e73a819dSTao Ma * 1. If cpos can be found in one refcount record, return the record.
1052e73a819dSTao Ma * 2. If cpos can't be found, return a fake record which start from cpos
1053e73a819dSTao Ma * and end at a small value between cpos+len and start of the next record.
1054e73a819dSTao Ma * This fake record has r_refcount = 0.
1055e73a819dSTao Ma */
ocfs2_get_refcount_rec(struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,u64 cpos,unsigned int len,struct ocfs2_refcount_rec * ret_rec,int * index,struct buffer_head ** ret_bh)1056e73a819dSTao Ma static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
1057e73a819dSTao Ma struct buffer_head *ref_root_bh,
1058e73a819dSTao Ma u64 cpos, unsigned int len,
1059e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec,
1060e73a819dSTao Ma int *index,
1061e73a819dSTao Ma struct buffer_head **ret_bh)
1062e73a819dSTao Ma {
1063e73a819dSTao Ma int ret = 0, i, found;
10643f649ab7SKees Cook u32 low_cpos, cpos_end;
1065e73a819dSTao Ma struct ocfs2_extent_list *el;
106638a04e43STao Ma struct ocfs2_extent_rec *rec = NULL;
106738a04e43STao Ma struct ocfs2_extent_block *eb = NULL;
1068e73a819dSTao Ma struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
1069e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1070e73a819dSTao Ma struct ocfs2_refcount_block *rb =
1071e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1072e73a819dSTao Ma
1073e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
1074e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
1075e73a819dSTao Ma ret_rec, index);
1076e73a819dSTao Ma *ret_bh = ref_root_bh;
1077e73a819dSTao Ma get_bh(ref_root_bh);
1078e73a819dSTao Ma return 0;
1079e73a819dSTao Ma }
1080e73a819dSTao Ma
1081e73a819dSTao Ma el = &rb->rf_list;
1082e73a819dSTao Ma low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1083e73a819dSTao Ma
1084e73a819dSTao Ma if (el->l_tree_depth) {
1085e73a819dSTao Ma ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1086e73a819dSTao Ma if (ret) {
1087e73a819dSTao Ma mlog_errno(ret);
1088e73a819dSTao Ma goto out;
1089e73a819dSTao Ma }
1090e73a819dSTao Ma
1091e73a819dSTao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1092e73a819dSTao Ma el = &eb->h_list;
1093e73a819dSTao Ma
1094e73a819dSTao Ma if (el->l_tree_depth) {
109517a5b9abSGoldwyn Rodrigues ret = ocfs2_error(sb,
10967ecef14aSJoe Perches "refcount tree %llu has non zero tree depth in leaf btree tree block %llu\n",
1097e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci),
1098e73a819dSTao Ma (unsigned long long)eb_bh->b_blocknr);
1099e73a819dSTao Ma goto out;
1100e73a819dSTao Ma }
1101e73a819dSTao Ma }
1102e73a819dSTao Ma
1103e73a819dSTao Ma found = 0;
1104e73a819dSTao Ma for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1105e73a819dSTao Ma rec = &el->l_recs[i];
1106e73a819dSTao Ma
1107e73a819dSTao Ma if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1108e73a819dSTao Ma found = 1;
1109e73a819dSTao Ma break;
1110e73a819dSTao Ma }
1111e73a819dSTao Ma }
1112e73a819dSTao Ma
111338a04e43STao Ma if (found) {
111438a04e43STao Ma ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh,
111538a04e43STao Ma eb, el, i, &cpos_end);
111638a04e43STao Ma if (ret) {
111738a04e43STao Ma mlog_errno(ret);
111838a04e43STao Ma goto out;
111938a04e43STao Ma }
1120e73a819dSTao Ma
112138a04e43STao Ma if (cpos_end < low_cpos + len)
112238a04e43STao Ma len = cpos_end - low_cpos;
1123e73a819dSTao Ma }
1124e73a819dSTao Ma
1125e73a819dSTao Ma ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1126e73a819dSTao Ma &ref_leaf_bh);
1127e73a819dSTao Ma if (ret) {
1128e73a819dSTao Ma mlog_errno(ret);
1129e73a819dSTao Ma goto out;
1130e73a819dSTao Ma }
1131e73a819dSTao Ma
1132e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1133e73a819dSTao Ma ret_rec, index);
1134e73a819dSTao Ma *ret_bh = ref_leaf_bh;
1135e73a819dSTao Ma out:
1136e73a819dSTao Ma brelse(eb_bh);
1137e73a819dSTao Ma return ret;
1138e73a819dSTao Ma }
1139e73a819dSTao Ma
1140e73a819dSTao Ma enum ocfs2_ref_rec_contig {
1141e73a819dSTao Ma REF_CONTIG_NONE = 0,
1142e73a819dSTao Ma REF_CONTIG_LEFT,
1143e73a819dSTao Ma REF_CONTIG_RIGHT,
1144e73a819dSTao Ma REF_CONTIG_LEFTRIGHT,
1145e73a819dSTao Ma };
1146e73a819dSTao Ma
1147e73a819dSTao Ma static enum ocfs2_ref_rec_contig
ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block * rb,int index)1148e73a819dSTao Ma ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1149e73a819dSTao Ma int index)
1150e73a819dSTao Ma {
1151e73a819dSTao Ma if ((rb->rf_records.rl_recs[index].r_refcount ==
1152e73a819dSTao Ma rb->rf_records.rl_recs[index + 1].r_refcount) &&
1153e73a819dSTao Ma (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1154e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1155e73a819dSTao Ma le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1156e73a819dSTao Ma return REF_CONTIG_RIGHT;
1157e73a819dSTao Ma
1158e73a819dSTao Ma return REF_CONTIG_NONE;
1159e73a819dSTao Ma }
1160e73a819dSTao Ma
1161e73a819dSTao Ma static enum ocfs2_ref_rec_contig
ocfs2_refcount_rec_contig(struct ocfs2_refcount_block * rb,int index)1162e73a819dSTao Ma ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1163e73a819dSTao Ma int index)
1164e73a819dSTao Ma {
1165e73a819dSTao Ma enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1166e73a819dSTao Ma
1167e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1168e73a819dSTao Ma ret = ocfs2_refcount_rec_adjacent(rb, index);
1169e73a819dSTao Ma
1170e73a819dSTao Ma if (index > 0) {
1171e73a819dSTao Ma enum ocfs2_ref_rec_contig tmp;
1172e73a819dSTao Ma
1173e73a819dSTao Ma tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1174e73a819dSTao Ma
1175e73a819dSTao Ma if (tmp == REF_CONTIG_RIGHT) {
1176e73a819dSTao Ma if (ret == REF_CONTIG_RIGHT)
1177e73a819dSTao Ma ret = REF_CONTIG_LEFTRIGHT;
1178e73a819dSTao Ma else
1179e73a819dSTao Ma ret = REF_CONTIG_LEFT;
1180e73a819dSTao Ma }
1181e73a819dSTao Ma }
1182e73a819dSTao Ma
1183e73a819dSTao Ma return ret;
1184e73a819dSTao Ma }
1185e73a819dSTao Ma
ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block * rb,int index)1186e73a819dSTao Ma static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1187e73a819dSTao Ma int index)
1188e73a819dSTao Ma {
1189e73a819dSTao Ma BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1190e73a819dSTao Ma rb->rf_records.rl_recs[index+1].r_refcount);
1191e73a819dSTao Ma
1192e73a819dSTao Ma le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1193e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1194e73a819dSTao Ma
1195e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1196e73a819dSTao Ma memmove(&rb->rf_records.rl_recs[index + 1],
1197e73a819dSTao Ma &rb->rf_records.rl_recs[index + 2],
1198e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec) *
1199e73a819dSTao Ma (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1200e73a819dSTao Ma
1201e73a819dSTao Ma memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1202e73a819dSTao Ma 0, sizeof(struct ocfs2_refcount_rec));
1203e73a819dSTao Ma le16_add_cpu(&rb->rf_records.rl_used, -1);
1204e73a819dSTao Ma }
1205e73a819dSTao Ma
1206e73a819dSTao Ma /*
1207e73a819dSTao Ma * Merge the refcount rec if we are contiguous with the adjacent recs.
1208e73a819dSTao Ma */
ocfs2_refcount_rec_merge(struct ocfs2_refcount_block * rb,int index)1209e73a819dSTao Ma static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1210e73a819dSTao Ma int index)
1211e73a819dSTao Ma {
1212e73a819dSTao Ma enum ocfs2_ref_rec_contig contig =
1213e73a819dSTao Ma ocfs2_refcount_rec_contig(rb, index);
1214e73a819dSTao Ma
1215e73a819dSTao Ma if (contig == REF_CONTIG_NONE)
1216e73a819dSTao Ma return;
1217e73a819dSTao Ma
1218e73a819dSTao Ma if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1219e73a819dSTao Ma BUG_ON(index == 0);
1220e73a819dSTao Ma index--;
1221e73a819dSTao Ma }
1222e73a819dSTao Ma
1223e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index);
1224e73a819dSTao Ma
1225e73a819dSTao Ma if (contig == REF_CONTIG_LEFTRIGHT)
1226e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index);
1227e73a819dSTao Ma }
1228e73a819dSTao Ma
12291823cb0bSTao Ma /*
12301823cb0bSTao Ma * Change the refcount indexed by "index" in ref_bh.
12311823cb0bSTao Ma * If refcount reaches 0, remove it.
12321823cb0bSTao Ma */
ocfs2_change_refcount_rec(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_leaf_bh,int index,int merge,int change)1233e73a819dSTao Ma static int ocfs2_change_refcount_rec(handle_t *handle,
1234e73a819dSTao Ma struct ocfs2_caching_info *ci,
1235e73a819dSTao Ma struct buffer_head *ref_leaf_bh,
12367540c1a7STao Ma int index, int merge, int change)
1237e73a819dSTao Ma {
1238e73a819dSTao Ma int ret;
1239e73a819dSTao Ma struct ocfs2_refcount_block *rb =
1240e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
12411823cb0bSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records;
12421823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1243e73a819dSTao Ma
1244e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1245e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1246e73a819dSTao Ma if (ret) {
1247e73a819dSTao Ma mlog_errno(ret);
1248e73a819dSTao Ma goto out;
1249e73a819dSTao Ma }
1250e73a819dSTao Ma
1251198aac28STao Ma trace_ocfs2_change_refcount_rec(
1252198aac28STao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci),
1253198aac28STao Ma index, le32_to_cpu(rec->r_refcount), change);
1254e73a819dSTao Ma le32_add_cpu(&rec->r_refcount, change);
1255e73a819dSTao Ma
12561823cb0bSTao Ma if (!rec->r_refcount) {
12571823cb0bSTao Ma if (index != le16_to_cpu(rl->rl_used) - 1) {
12581823cb0bSTao Ma memmove(rec, rec + 1,
12591823cb0bSTao Ma (le16_to_cpu(rl->rl_used) - index - 1) *
12601823cb0bSTao Ma sizeof(struct ocfs2_refcount_rec));
12611823cb0bSTao Ma memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
12621823cb0bSTao Ma 0, sizeof(struct ocfs2_refcount_rec));
12631823cb0bSTao Ma }
12641823cb0bSTao Ma
12651823cb0bSTao Ma le16_add_cpu(&rl->rl_used, -1);
12667540c1a7STao Ma } else if (merge)
1267e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index);
1268e73a819dSTao Ma
1269ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh);
1270e73a819dSTao Ma out:
1271e73a819dSTao Ma return ret;
1272e73a819dSTao Ma }
1273e73a819dSTao Ma
ocfs2_expand_inline_ref_root(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head ** ref_leaf_bh,struct ocfs2_alloc_context * meta_ac)1274e73a819dSTao Ma static int ocfs2_expand_inline_ref_root(handle_t *handle,
1275e73a819dSTao Ma struct ocfs2_caching_info *ci,
1276e73a819dSTao Ma struct buffer_head *ref_root_bh,
1277e73a819dSTao Ma struct buffer_head **ref_leaf_bh,
1278e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac)
1279e73a819dSTao Ma {
1280e73a819dSTao Ma int ret;
1281e73a819dSTao Ma u16 suballoc_bit_start;
1282e73a819dSTao Ma u32 num_got;
12832b6cb576SJoel Becker u64 suballoc_loc, blkno;
1284e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1285e73a819dSTao Ma struct buffer_head *new_bh = NULL;
1286e73a819dSTao Ma struct ocfs2_refcount_block *new_rb;
1287e73a819dSTao Ma struct ocfs2_refcount_block *root_rb =
1288e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1289e73a819dSTao Ma
1290e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1291e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1292e73a819dSTao Ma if (ret) {
1293e73a819dSTao Ma mlog_errno(ret);
1294e73a819dSTao Ma goto out;
1295e73a819dSTao Ma }
1296e73a819dSTao Ma
12972b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1298e73a819dSTao Ma &suballoc_bit_start, &num_got,
1299e73a819dSTao Ma &blkno);
1300e73a819dSTao Ma if (ret) {
1301e73a819dSTao Ma mlog_errno(ret);
1302e73a819dSTao Ma goto out;
1303e73a819dSTao Ma }
1304e73a819dSTao Ma
1305e73a819dSTao Ma new_bh = sb_getblk(sb, blkno);
1306e73a819dSTao Ma if (new_bh == NULL) {
13077391a294SRui Xiang ret = -ENOMEM;
1308e73a819dSTao Ma mlog_errno(ret);
1309e73a819dSTao Ma goto out;
1310e73a819dSTao Ma }
1311e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh);
1312e73a819dSTao Ma
1313e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1314e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE);
1315e73a819dSTao Ma if (ret) {
1316e73a819dSTao Ma mlog_errno(ret);
1317e73a819dSTao Ma goto out;
1318e73a819dSTao Ma }
1319e73a819dSTao Ma
1320e73a819dSTao Ma /*
1321e73a819dSTao Ma * Initialize ocfs2_refcount_block.
1322e73a819dSTao Ma * It should contain the same information as the old root.
1323e73a819dSTao Ma * so just memcpy it and change the corresponding field.
1324e73a819dSTao Ma */
1325e73a819dSTao Ma memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1326e73a819dSTao Ma
1327e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1328b89c5428STiger Yang new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
13292b6cb576SJoel Becker new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1330e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1331e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno);
1332e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(0);
1333e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1334e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1335e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh);
1336e73a819dSTao Ma
1337e73a819dSTao Ma /* Now change the root. */
1338e73a819dSTao Ma memset(&root_rb->rf_list, 0, sb->s_blocksize -
1339e73a819dSTao Ma offsetof(struct ocfs2_refcount_block, rf_list));
1340e73a819dSTao Ma root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1341e73a819dSTao Ma root_rb->rf_clusters = cpu_to_le32(1);
1342e73a819dSTao Ma root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1343e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1344e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1345e73a819dSTao Ma root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1346e73a819dSTao Ma
1347e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_root_bh);
1348e73a819dSTao Ma
1349198aac28STao Ma trace_ocfs2_expand_inline_ref_root((unsigned long long)blkno,
1350e73a819dSTao Ma le16_to_cpu(new_rb->rf_records.rl_used));
1351e73a819dSTao Ma
1352e73a819dSTao Ma *ref_leaf_bh = new_bh;
1353e73a819dSTao Ma new_bh = NULL;
1354e73a819dSTao Ma out:
1355e73a819dSTao Ma brelse(new_bh);
1356e73a819dSTao Ma return ret;
1357e73a819dSTao Ma }
1358e73a819dSTao Ma
ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec * prev,struct ocfs2_refcount_rec * next)1359e73a819dSTao Ma static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1360e73a819dSTao Ma struct ocfs2_refcount_rec *next)
1361e73a819dSTao Ma {
1362e73a819dSTao Ma if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1363e73a819dSTao Ma ocfs2_get_ref_rec_low_cpos(next))
1364e73a819dSTao Ma return 1;
1365e73a819dSTao Ma
1366e73a819dSTao Ma return 0;
1367e73a819dSTao Ma }
1368e73a819dSTao Ma
cmp_refcount_rec_by_low_cpos(const void * a,const void * b)1369e73a819dSTao Ma static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1370e73a819dSTao Ma {
1371e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b;
1372e73a819dSTao Ma u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1373e73a819dSTao Ma u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1374e73a819dSTao Ma
1375e73a819dSTao Ma if (l_cpos > r_cpos)
1376e73a819dSTao Ma return 1;
1377e73a819dSTao Ma if (l_cpos < r_cpos)
1378e73a819dSTao Ma return -1;
1379e73a819dSTao Ma return 0;
1380e73a819dSTao Ma }
1381e73a819dSTao Ma
cmp_refcount_rec_by_cpos(const void * a,const void * b)1382e73a819dSTao Ma static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1383e73a819dSTao Ma {
1384e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b;
1385e73a819dSTao Ma u64 l_cpos = le64_to_cpu(l->r_cpos);
1386e73a819dSTao Ma u64 r_cpos = le64_to_cpu(r->r_cpos);
1387e73a819dSTao Ma
1388e73a819dSTao Ma if (l_cpos > r_cpos)
1389e73a819dSTao Ma return 1;
1390e73a819dSTao Ma if (l_cpos < r_cpos)
1391e73a819dSTao Ma return -1;
1392e73a819dSTao Ma return 0;
1393e73a819dSTao Ma }
1394e73a819dSTao Ma
swap_refcount_rec(void * a,void * b,int size)1395e73a819dSTao Ma static void swap_refcount_rec(void *a, void *b, int size)
1396e73a819dSTao Ma {
1397a612543fSFabian Frederick struct ocfs2_refcount_rec *l = a, *r = b;
1398e73a819dSTao Ma
1399a612543fSFabian Frederick swap(*l, *r);
1400e73a819dSTao Ma }
1401e73a819dSTao Ma
1402e73a819dSTao Ma /*
1403e73a819dSTao Ma * The refcount cpos are ordered by their 64bit cpos,
1404e73a819dSTao Ma * But we will use the low 32 bit to be the e_cpos in the b-tree.
1405e73a819dSTao Ma * So we need to make sure that this pos isn't intersected with others.
1406e73a819dSTao Ma *
1407e73a819dSTao Ma * Note: The refcount block is already sorted by their low 32 bit cpos,
1408e73a819dSTao Ma * So just try the middle pos first, and we will exit when we find
1409e73a819dSTao Ma * the good position.
1410e73a819dSTao Ma */
ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list * rl,u32 * split_pos,int * split_index)1411e73a819dSTao Ma static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1412e73a819dSTao Ma u32 *split_pos, int *split_index)
1413e73a819dSTao Ma {
1414e73a819dSTao Ma int num_used = le16_to_cpu(rl->rl_used);
1415e73a819dSTao Ma int delta, middle = num_used / 2;
1416e73a819dSTao Ma
1417e73a819dSTao Ma for (delta = 0; delta < middle; delta++) {
1418e73a819dSTao Ma /* Let's check delta earlier than middle */
1419e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect(
1420e73a819dSTao Ma &rl->rl_recs[middle - delta - 1],
1421e73a819dSTao Ma &rl->rl_recs[middle - delta])) {
1422e73a819dSTao Ma *split_index = middle - delta;
1423e73a819dSTao Ma break;
1424e73a819dSTao Ma }
1425e73a819dSTao Ma
1426e73a819dSTao Ma /* For even counts, don't walk off the end */
1427e73a819dSTao Ma if ((middle + delta + 1) == num_used)
1428e73a819dSTao Ma continue;
1429e73a819dSTao Ma
1430e73a819dSTao Ma /* Now try delta past middle */
1431e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect(
1432e73a819dSTao Ma &rl->rl_recs[middle + delta],
1433e73a819dSTao Ma &rl->rl_recs[middle + delta + 1])) {
1434e73a819dSTao Ma *split_index = middle + delta + 1;
1435e73a819dSTao Ma break;
1436e73a819dSTao Ma }
1437e73a819dSTao Ma }
1438e73a819dSTao Ma
1439e73a819dSTao Ma if (delta >= middle)
1440e73a819dSTao Ma return -ENOSPC;
1441e73a819dSTao Ma
1442e73a819dSTao Ma *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1443e73a819dSTao Ma return 0;
1444e73a819dSTao Ma }
1445e73a819dSTao Ma
ocfs2_divide_leaf_refcount_block(struct buffer_head * ref_leaf_bh,struct buffer_head * new_bh,u32 * split_cpos)1446e73a819dSTao Ma static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1447e73a819dSTao Ma struct buffer_head *new_bh,
1448e73a819dSTao Ma u32 *split_cpos)
1449e73a819dSTao Ma {
1450e73a819dSTao Ma int split_index = 0, num_moved, ret;
1451e73a819dSTao Ma u32 cpos = 0;
1452e73a819dSTao Ma struct ocfs2_refcount_block *rb =
1453e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1454e73a819dSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records;
1455e73a819dSTao Ma struct ocfs2_refcount_block *new_rb =
1456e73a819dSTao Ma (struct ocfs2_refcount_block *)new_bh->b_data;
1457e73a819dSTao Ma struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1458e73a819dSTao Ma
1459198aac28STao Ma trace_ocfs2_divide_leaf_refcount_block(
1460e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr,
146128748b32SAl Viro le16_to_cpu(rl->rl_count), le16_to_cpu(rl->rl_used));
1462e73a819dSTao Ma
1463e73a819dSTao Ma /*
1464e73a819dSTao Ma * XXX: Improvement later.
1465e73a819dSTao Ma * If we know all the high 32 bit cpos is the same, no need to sort.
1466e73a819dSTao Ma *
1467e73a819dSTao Ma * In order to make the whole process safe, we do:
1468e73a819dSTao Ma * 1. sort the entries by their low 32 bit cpos first so that we can
1469e73a819dSTao Ma * find the split cpos easily.
1470e73a819dSTao Ma * 2. call ocfs2_insert_extent to insert the new refcount block.
1471e73a819dSTao Ma * 3. move the refcount rec to the new block.
1472e73a819dSTao Ma * 4. sort the entries by their 64 bit cpos.
1473e73a819dSTao Ma * 5. dirty the new_rb and rb.
1474e73a819dSTao Ma */
1475e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1476e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec),
1477e73a819dSTao Ma cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1478e73a819dSTao Ma
1479e73a819dSTao Ma ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1480e73a819dSTao Ma if (ret) {
1481e73a819dSTao Ma mlog_errno(ret);
1482e73a819dSTao Ma return ret;
1483e73a819dSTao Ma }
1484e73a819dSTao Ma
1485e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(cpos);
1486e73a819dSTao Ma
1487e73a819dSTao Ma /* move refcount records starting from split_index to the new block. */
1488e73a819dSTao Ma num_moved = le16_to_cpu(rl->rl_used) - split_index;
1489e73a819dSTao Ma memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1490e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec));
1491e73a819dSTao Ma
1492e73a819dSTao Ma /*ok, remove the entries we just moved over to the other block. */
1493e73a819dSTao Ma memset(&rl->rl_recs[split_index], 0,
1494e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec));
1495e73a819dSTao Ma
1496e73a819dSTao Ma /* change old and new rl_used accordingly. */
1497e73a819dSTao Ma le16_add_cpu(&rl->rl_used, -num_moved);
149812d4cec9STao Ma new_rl->rl_used = cpu_to_le16(num_moved);
1499e73a819dSTao Ma
1500e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1501e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec),
1502e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec);
1503e73a819dSTao Ma
1504e73a819dSTao Ma sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1505e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec),
1506e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec);
1507e73a819dSTao Ma
1508e73a819dSTao Ma *split_cpos = cpos;
1509e73a819dSTao Ma return 0;
1510e73a819dSTao Ma }
1511e73a819dSTao Ma
ocfs2_new_leaf_refcount_block(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head * ref_leaf_bh,struct ocfs2_alloc_context * meta_ac)1512e73a819dSTao Ma static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1513e73a819dSTao Ma struct ocfs2_caching_info *ci,
1514e73a819dSTao Ma struct buffer_head *ref_root_bh,
1515e73a819dSTao Ma struct buffer_head *ref_leaf_bh,
1516e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac)
1517e73a819dSTao Ma {
1518e73a819dSTao Ma int ret;
1519e73a819dSTao Ma u16 suballoc_bit_start;
1520e73a819dSTao Ma u32 num_got, new_cpos;
15212b6cb576SJoel Becker u64 suballoc_loc, blkno;
1522e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1523e73a819dSTao Ma struct ocfs2_refcount_block *root_rb =
1524e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1525e73a819dSTao Ma struct buffer_head *new_bh = NULL;
1526e73a819dSTao Ma struct ocfs2_refcount_block *new_rb;
1527e73a819dSTao Ma struct ocfs2_extent_tree ref_et;
1528e73a819dSTao Ma
1529e73a819dSTao Ma BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1530e73a819dSTao Ma
1531e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1532e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1533e73a819dSTao Ma if (ret) {
1534e73a819dSTao Ma mlog_errno(ret);
1535e73a819dSTao Ma goto out;
1536e73a819dSTao Ma }
1537e73a819dSTao Ma
1538e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1539e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1540e73a819dSTao Ma if (ret) {
1541e73a819dSTao Ma mlog_errno(ret);
1542e73a819dSTao Ma goto out;
1543e73a819dSTao Ma }
1544e73a819dSTao Ma
15452b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1546e73a819dSTao Ma &suballoc_bit_start, &num_got,
1547e73a819dSTao Ma &blkno);
1548e73a819dSTao Ma if (ret) {
1549e73a819dSTao Ma mlog_errno(ret);
1550e73a819dSTao Ma goto out;
1551e73a819dSTao Ma }
1552e73a819dSTao Ma
1553e73a819dSTao Ma new_bh = sb_getblk(sb, blkno);
1554e73a819dSTao Ma if (new_bh == NULL) {
15557391a294SRui Xiang ret = -ENOMEM;
1556e73a819dSTao Ma mlog_errno(ret);
1557e73a819dSTao Ma goto out;
1558e73a819dSTao Ma }
1559e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh);
1560e73a819dSTao Ma
1561e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1562e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE);
1563e73a819dSTao Ma if (ret) {
1564e73a819dSTao Ma mlog_errno(ret);
1565e73a819dSTao Ma goto out;
1566e73a819dSTao Ma }
1567e73a819dSTao Ma
1568e73a819dSTao Ma /* Initialize ocfs2_refcount_block. */
1569e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1570e73a819dSTao Ma memset(new_rb, 0, sb->s_blocksize);
1571e73a819dSTao Ma strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1572b89c5428STiger Yang new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
15732b6cb576SJoel Becker new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1574e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1575e73a819dSTao Ma new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1576e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno);
1577e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1578e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1579e73a819dSTao Ma new_rb->rf_records.rl_count =
1580e73a819dSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1581e73a819dSTao Ma new_rb->rf_generation = root_rb->rf_generation;
1582e73a819dSTao Ma
1583e73a819dSTao Ma ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1584e73a819dSTao Ma if (ret) {
1585e73a819dSTao Ma mlog_errno(ret);
1586e73a819dSTao Ma goto out;
1587e73a819dSTao Ma }
1588e73a819dSTao Ma
1589e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh);
1590e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh);
1591e73a819dSTao Ma
1592e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1593e73a819dSTao Ma
1594198aac28STao Ma trace_ocfs2_new_leaf_refcount_block(
1595e73a819dSTao Ma (unsigned long long)new_bh->b_blocknr, new_cpos);
1596e73a819dSTao Ma
1597e73a819dSTao Ma /* Insert the new leaf block with the specific offset cpos. */
1598e73a819dSTao Ma ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1599e73a819dSTao Ma 1, 0, meta_ac);
1600e73a819dSTao Ma if (ret)
1601e73a819dSTao Ma mlog_errno(ret);
1602e73a819dSTao Ma
1603e73a819dSTao Ma out:
1604e73a819dSTao Ma brelse(new_bh);
1605e73a819dSTao Ma return ret;
1606e73a819dSTao Ma }
1607e73a819dSTao Ma
ocfs2_expand_refcount_tree(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head * ref_leaf_bh,struct ocfs2_alloc_context * meta_ac)1608e73a819dSTao Ma static int ocfs2_expand_refcount_tree(handle_t *handle,
1609e73a819dSTao Ma struct ocfs2_caching_info *ci,
1610e73a819dSTao Ma struct buffer_head *ref_root_bh,
1611e73a819dSTao Ma struct buffer_head *ref_leaf_bh,
1612e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac)
1613e73a819dSTao Ma {
1614e73a819dSTao Ma int ret;
1615e73a819dSTao Ma struct buffer_head *expand_bh = NULL;
1616e73a819dSTao Ma
1617e73a819dSTao Ma if (ref_root_bh == ref_leaf_bh) {
1618e73a819dSTao Ma /*
1619e73a819dSTao Ma * the old root bh hasn't been expanded to a b-tree,
1620e73a819dSTao Ma * so expand it first.
1621e73a819dSTao Ma */
1622e73a819dSTao Ma ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1623e73a819dSTao Ma &expand_bh, meta_ac);
1624e73a819dSTao Ma if (ret) {
1625e73a819dSTao Ma mlog_errno(ret);
1626e73a819dSTao Ma goto out;
1627e73a819dSTao Ma }
1628e73a819dSTao Ma } else {
1629e73a819dSTao Ma expand_bh = ref_leaf_bh;
1630e73a819dSTao Ma get_bh(expand_bh);
1631e73a819dSTao Ma }
1632e73a819dSTao Ma
1633e73a819dSTao Ma
1634e73a819dSTao Ma /* Now add a new refcount block into the tree.*/
1635e73a819dSTao Ma ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1636e73a819dSTao Ma expand_bh, meta_ac);
1637e73a819dSTao Ma if (ret)
1638e73a819dSTao Ma mlog_errno(ret);
1639e73a819dSTao Ma out:
1640e73a819dSTao Ma brelse(expand_bh);
1641e73a819dSTao Ma return ret;
1642e73a819dSTao Ma }
1643e73a819dSTao Ma
1644e73a819dSTao Ma /*
1645e73a819dSTao Ma * Adjust the extent rec in b-tree representing ref_leaf_bh.
1646e73a819dSTao Ma *
1647e73a819dSTao Ma * Only called when we have inserted a new refcount rec at index 0
1648e73a819dSTao Ma * which means ocfs2_extent_rec.e_cpos may need some change.
1649e73a819dSTao Ma */
ocfs2_adjust_refcount_rec(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head * ref_leaf_bh,struct ocfs2_refcount_rec * rec)1650e73a819dSTao Ma static int ocfs2_adjust_refcount_rec(handle_t *handle,
1651e73a819dSTao Ma struct ocfs2_caching_info *ci,
1652e73a819dSTao Ma struct buffer_head *ref_root_bh,
1653e73a819dSTao Ma struct buffer_head *ref_leaf_bh,
1654e73a819dSTao Ma struct ocfs2_refcount_rec *rec)
1655e73a819dSTao Ma {
1656e73a819dSTao Ma int ret = 0, i;
1657e73a819dSTao Ma u32 new_cpos, old_cpos;
1658e73a819dSTao Ma struct ocfs2_path *path = NULL;
1659e73a819dSTao Ma struct ocfs2_extent_tree et;
1660e73a819dSTao Ma struct ocfs2_refcount_block *rb =
1661e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1662e73a819dSTao Ma struct ocfs2_extent_list *el;
1663e73a819dSTao Ma
1664e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1665e73a819dSTao Ma goto out;
1666e73a819dSTao Ma
1667e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1668e73a819dSTao Ma old_cpos = le32_to_cpu(rb->rf_cpos);
1669e73a819dSTao Ma new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1670e73a819dSTao Ma if (old_cpos <= new_cpos)
1671e73a819dSTao Ma goto out;
1672e73a819dSTao Ma
1673e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1674e73a819dSTao Ma
1675e73a819dSTao Ma path = ocfs2_new_path_from_et(&et);
1676e73a819dSTao Ma if (!path) {
1677e73a819dSTao Ma ret = -ENOMEM;
1678e73a819dSTao Ma mlog_errno(ret);
1679e73a819dSTao Ma goto out;
1680e73a819dSTao Ma }
1681e73a819dSTao Ma
1682e73a819dSTao Ma ret = ocfs2_find_path(ci, path, old_cpos);
1683e73a819dSTao Ma if (ret) {
1684e73a819dSTao Ma mlog_errno(ret);
1685e73a819dSTao Ma goto out;
1686e73a819dSTao Ma }
1687e73a819dSTao Ma
1688e73a819dSTao Ma /*
1689e73a819dSTao Ma * 2 more credits, one for the leaf refcount block, one for
1690e73a819dSTao Ma * the extent block contains the extent rec.
1691e73a819dSTao Ma */
1692c901fb00STao Ma ret = ocfs2_extend_trans(handle, 2);
1693e73a819dSTao Ma if (ret < 0) {
1694e73a819dSTao Ma mlog_errno(ret);
1695e73a819dSTao Ma goto out;
1696e73a819dSTao Ma }
1697e73a819dSTao Ma
1698e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1699e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1700e73a819dSTao Ma if (ret < 0) {
1701e73a819dSTao Ma mlog_errno(ret);
1702e73a819dSTao Ma goto out;
1703e73a819dSTao Ma }
1704e73a819dSTao Ma
1705e73a819dSTao Ma ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1706e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1707e73a819dSTao Ma if (ret < 0) {
1708e73a819dSTao Ma mlog_errno(ret);
1709e73a819dSTao Ma goto out;
1710e73a819dSTao Ma }
1711e73a819dSTao Ma
1712e73a819dSTao Ma /* change the leaf extent block first. */
1713e73a819dSTao Ma el = path_leaf_el(path);
1714e73a819dSTao Ma
1715e73a819dSTao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1716e73a819dSTao Ma if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1717e73a819dSTao Ma break;
1718e73a819dSTao Ma
1719e73a819dSTao Ma BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1720e73a819dSTao Ma
1721e73a819dSTao Ma el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1722e73a819dSTao Ma
1723e73a819dSTao Ma /* change the r_cpos in the leaf block. */
1724e73a819dSTao Ma rb->rf_cpos = cpu_to_le32(new_cpos);
1725e73a819dSTao Ma
1726e73a819dSTao Ma ocfs2_journal_dirty(handle, path_leaf_bh(path));
1727e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh);
1728e73a819dSTao Ma
1729e73a819dSTao Ma out:
1730e73a819dSTao Ma ocfs2_free_path(path);
1731e73a819dSTao Ma return ret;
1732e73a819dSTao Ma }
1733e73a819dSTao Ma
ocfs2_insert_refcount_rec(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head * ref_leaf_bh,struct ocfs2_refcount_rec * rec,int index,int merge,struct ocfs2_alloc_context * meta_ac)1734e73a819dSTao Ma static int ocfs2_insert_refcount_rec(handle_t *handle,
1735e73a819dSTao Ma struct ocfs2_caching_info *ci,
1736e73a819dSTao Ma struct buffer_head *ref_root_bh,
1737e73a819dSTao Ma struct buffer_head *ref_leaf_bh,
1738e73a819dSTao Ma struct ocfs2_refcount_rec *rec,
17397540c1a7STao Ma int index, int merge,
1740e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac)
1741e73a819dSTao Ma {
1742e73a819dSTao Ma int ret;
1743e73a819dSTao Ma struct ocfs2_refcount_block *rb =
1744e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1745e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1746e73a819dSTao Ma struct buffer_head *new_bh = NULL;
1747e73a819dSTao Ma
1748e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1749e73a819dSTao Ma
1750e73a819dSTao Ma if (rf_list->rl_used == rf_list->rl_count) {
1751e73a819dSTao Ma u64 cpos = le64_to_cpu(rec->r_cpos);
1752e73a819dSTao Ma u32 len = le32_to_cpu(rec->r_clusters);
1753e73a819dSTao Ma
1754e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1755e73a819dSTao Ma ref_leaf_bh, meta_ac);
1756e73a819dSTao Ma if (ret) {
1757e73a819dSTao Ma mlog_errno(ret);
1758e73a819dSTao Ma goto out;
1759e73a819dSTao Ma }
1760e73a819dSTao Ma
1761e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1762e73a819dSTao Ma cpos, len, NULL, &index,
1763e73a819dSTao Ma &new_bh);
1764e73a819dSTao Ma if (ret) {
1765e73a819dSTao Ma mlog_errno(ret);
1766e73a819dSTao Ma goto out;
1767e73a819dSTao Ma }
1768e73a819dSTao Ma
1769e73a819dSTao Ma ref_leaf_bh = new_bh;
1770e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1771e73a819dSTao Ma rf_list = &rb->rf_records;
1772e73a819dSTao Ma }
1773e73a819dSTao Ma
1774e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1775e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1776e73a819dSTao Ma if (ret) {
1777e73a819dSTao Ma mlog_errno(ret);
1778e73a819dSTao Ma goto out;
1779e73a819dSTao Ma }
1780e73a819dSTao Ma
1781e73a819dSTao Ma if (index < le16_to_cpu(rf_list->rl_used))
1782e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1],
1783e73a819dSTao Ma &rf_list->rl_recs[index],
1784e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index) *
1785e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec));
1786e73a819dSTao Ma
1787198aac28STao Ma trace_ocfs2_insert_refcount_rec(
1788198aac28STao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index,
1789e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec->r_cpos),
1790198aac28STao Ma le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount));
1791e73a819dSTao Ma
1792e73a819dSTao Ma rf_list->rl_recs[index] = *rec;
1793e73a819dSTao Ma
1794e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, 1);
1795e73a819dSTao Ma
17967540c1a7STao Ma if (merge)
1797e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index);
1798e73a819dSTao Ma
1799ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh);
1800e73a819dSTao Ma
1801e73a819dSTao Ma if (index == 0) {
1802e73a819dSTao Ma ret = ocfs2_adjust_refcount_rec(handle, ci,
1803e73a819dSTao Ma ref_root_bh,
1804e73a819dSTao Ma ref_leaf_bh, rec);
1805e73a819dSTao Ma if (ret)
1806e73a819dSTao Ma mlog_errno(ret);
1807e73a819dSTao Ma }
1808e73a819dSTao Ma out:
1809e73a819dSTao Ma brelse(new_bh);
1810e73a819dSTao Ma return ret;
1811e73a819dSTao Ma }
1812e73a819dSTao Ma
1813e73a819dSTao Ma /*
1814e73a819dSTao Ma * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1815e73a819dSTao Ma * This is much simple than our b-tree code.
1816e73a819dSTao Ma * split_rec is the new refcount rec we want to insert.
1817e73a819dSTao Ma * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1818e73a819dSTao Ma * increase refcount or decrease a refcount to non-zero).
1819e73a819dSTao Ma * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1820e73a819dSTao Ma * rec( in case we decrease a refcount to zero).
1821e73a819dSTao Ma */
ocfs2_split_refcount_rec(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head * ref_leaf_bh,struct ocfs2_refcount_rec * split_rec,int index,int merge,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc)1822e73a819dSTao Ma static int ocfs2_split_refcount_rec(handle_t *handle,
1823e73a819dSTao Ma struct ocfs2_caching_info *ci,
1824e73a819dSTao Ma struct buffer_head *ref_root_bh,
1825e73a819dSTao Ma struct buffer_head *ref_leaf_bh,
1826e73a819dSTao Ma struct ocfs2_refcount_rec *split_rec,
18277540c1a7STao Ma int index, int merge,
1828e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac,
1829e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
1830e73a819dSTao Ma {
1831e73a819dSTao Ma int ret, recs_need;
1832e73a819dSTao Ma u32 len;
1833e73a819dSTao Ma struct ocfs2_refcount_block *rb =
1834e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1835e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1836e73a819dSTao Ma struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1837e73a819dSTao Ma struct ocfs2_refcount_rec *tail_rec = NULL;
1838e73a819dSTao Ma struct buffer_head *new_bh = NULL;
1839e73a819dSTao Ma
1840e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1841e73a819dSTao Ma
1842198aac28STao Ma trace_ocfs2_split_refcount_rec(le64_to_cpu(orig_rec->r_cpos),
1843198aac28STao Ma le32_to_cpu(orig_rec->r_clusters),
1844198aac28STao Ma le32_to_cpu(orig_rec->r_refcount),
1845e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos),
1846198aac28STao Ma le32_to_cpu(split_rec->r_clusters),
1847198aac28STao Ma le32_to_cpu(split_rec->r_refcount));
1848e73a819dSTao Ma
1849e73a819dSTao Ma /*
1850e73a819dSTao Ma * If we just need to split the header or tail clusters,
1851e73a819dSTao Ma * no more recs are needed, just split is OK.
1852e73a819dSTao Ma * Otherwise we at least need one new recs.
1853e73a819dSTao Ma */
1854e73a819dSTao Ma if (!split_rec->r_refcount &&
1855e73a819dSTao Ma (split_rec->r_cpos == orig_rec->r_cpos ||
1856e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) +
1857e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) ==
1858e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1859e73a819dSTao Ma recs_need = 0;
1860e73a819dSTao Ma else
1861e73a819dSTao Ma recs_need = 1;
1862e73a819dSTao Ma
1863e73a819dSTao Ma /*
1864e73a819dSTao Ma * We need one more rec if we split in the middle and the new rec have
1865e73a819dSTao Ma * some refcount in it.
1866e73a819dSTao Ma */
1867e73a819dSTao Ma if (split_rec->r_refcount &&
1868e73a819dSTao Ma (split_rec->r_cpos != orig_rec->r_cpos &&
1869e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) +
1870e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) !=
1871e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1872e73a819dSTao Ma recs_need++;
1873e73a819dSTao Ma
1874e73a819dSTao Ma /* If the leaf block don't have enough record, expand it. */
187512d4cec9STao Ma if (le16_to_cpu(rf_list->rl_used) + recs_need >
187612d4cec9STao Ma le16_to_cpu(rf_list->rl_count)) {
1877e73a819dSTao Ma struct ocfs2_refcount_rec tmp_rec;
1878e73a819dSTao Ma u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1879e73a819dSTao Ma len = le32_to_cpu(orig_rec->r_clusters);
1880e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1881e73a819dSTao Ma ref_leaf_bh, meta_ac);
1882e73a819dSTao Ma if (ret) {
1883e73a819dSTao Ma mlog_errno(ret);
1884e73a819dSTao Ma goto out;
1885e73a819dSTao Ma }
1886e73a819dSTao Ma
1887e73a819dSTao Ma /*
1888e73a819dSTao Ma * We have to re-get it since now cpos may be moved to
1889e73a819dSTao Ma * another leaf block.
1890e73a819dSTao Ma */
1891e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1892e73a819dSTao Ma cpos, len, &tmp_rec, &index,
1893e73a819dSTao Ma &new_bh);
1894e73a819dSTao Ma if (ret) {
1895e73a819dSTao Ma mlog_errno(ret);
1896e73a819dSTao Ma goto out;
1897e73a819dSTao Ma }
1898e73a819dSTao Ma
1899e73a819dSTao Ma ref_leaf_bh = new_bh;
1900e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1901e73a819dSTao Ma rf_list = &rb->rf_records;
1902e73a819dSTao Ma orig_rec = &rf_list->rl_recs[index];
1903e73a819dSTao Ma }
1904e73a819dSTao Ma
1905e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1906e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
1907e73a819dSTao Ma if (ret) {
1908e73a819dSTao Ma mlog_errno(ret);
1909e73a819dSTao Ma goto out;
1910e73a819dSTao Ma }
1911e73a819dSTao Ma
1912e73a819dSTao Ma /*
1913e73a819dSTao Ma * We have calculated out how many new records we need and store
1914e73a819dSTao Ma * in recs_need, so spare enough space first by moving the records
1915e73a819dSTao Ma * after "index" to the end.
1916e73a819dSTao Ma */
1917e73a819dSTao Ma if (index != le16_to_cpu(rf_list->rl_used) - 1)
1918e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1 + recs_need],
1919e73a819dSTao Ma &rf_list->rl_recs[index + 1],
1920e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index - 1) *
1921e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec));
1922e73a819dSTao Ma
1923e73a819dSTao Ma len = (le64_to_cpu(orig_rec->r_cpos) +
1924e73a819dSTao Ma le32_to_cpu(orig_rec->r_clusters)) -
1925e73a819dSTao Ma (le64_to_cpu(split_rec->r_cpos) +
1926e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters));
1927e73a819dSTao Ma
1928e73a819dSTao Ma /*
1929e73a819dSTao Ma * If we have "len", the we will split in the tail and move it
1930e73a819dSTao Ma * to the end of the space we have just spared.
1931e73a819dSTao Ma */
1932e73a819dSTao Ma if (len) {
1933e73a819dSTao Ma tail_rec = &rf_list->rl_recs[index + recs_need];
1934e73a819dSTao Ma
1935e73a819dSTao Ma memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1936e73a819dSTao Ma le64_add_cpu(&tail_rec->r_cpos,
1937e73a819dSTao Ma le32_to_cpu(tail_rec->r_clusters) - len);
193812d4cec9STao Ma tail_rec->r_clusters = cpu_to_le32(len);
1939e73a819dSTao Ma }
1940e73a819dSTao Ma
1941e73a819dSTao Ma /*
1942e73a819dSTao Ma * If the split pos isn't the same as the original one, we need to
1943e73a819dSTao Ma * split in the head.
1944e73a819dSTao Ma *
1945e73a819dSTao Ma * Note: We have the chance that split_rec.r_refcount = 0,
1946e73a819dSTao Ma * recs_need = 0 and len > 0, which means we just cut the head from
1947e73a819dSTao Ma * the orig_rec and in that case we have done some modification in
1948e73a819dSTao Ma * orig_rec above, so the check for r_cpos is faked.
1949e73a819dSTao Ma */
1950e73a819dSTao Ma if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1951e73a819dSTao Ma len = le64_to_cpu(split_rec->r_cpos) -
1952e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos);
1953e73a819dSTao Ma orig_rec->r_clusters = cpu_to_le32(len);
1954e73a819dSTao Ma index++;
1955e73a819dSTao Ma }
1956e73a819dSTao Ma
1957e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, recs_need);
1958e73a819dSTao Ma
1959e73a819dSTao Ma if (split_rec->r_refcount) {
1960e73a819dSTao Ma rf_list->rl_recs[index] = *split_rec;
1961198aac28STao Ma trace_ocfs2_split_refcount_rec_insert(
1962198aac28STao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index,
1963e73a819dSTao Ma (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1964e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters),
1965198aac28STao Ma le32_to_cpu(split_rec->r_refcount));
1966e73a819dSTao Ma
19677540c1a7STao Ma if (merge)
1968e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index);
1969e73a819dSTao Ma }
1970e73a819dSTao Ma
1971ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh);
1972e73a819dSTao Ma
1973e73a819dSTao Ma out:
1974e73a819dSTao Ma brelse(new_bh);
1975e73a819dSTao Ma return ret;
1976e73a819dSTao Ma }
1977e73a819dSTao Ma
__ocfs2_increase_refcount(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,u64 cpos,u32 len,int merge,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc)19787540c1a7STao Ma static int __ocfs2_increase_refcount(handle_t *handle,
1979e73a819dSTao Ma struct ocfs2_caching_info *ci,
1980e73a819dSTao Ma struct buffer_head *ref_root_bh,
19817540c1a7STao Ma u64 cpos, u32 len, int merge,
1982e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac,
1983e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
1984e73a819dSTao Ma {
1985e73a819dSTao Ma int ret = 0, index;
1986e73a819dSTao Ma struct buffer_head *ref_leaf_bh = NULL;
1987e73a819dSTao Ma struct ocfs2_refcount_rec rec;
1988e73a819dSTao Ma unsigned int set_len = 0;
1989e73a819dSTao Ma
1990198aac28STao Ma trace_ocfs2_increase_refcount_begin(
1991e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci),
1992e73a819dSTao Ma (unsigned long long)cpos, len);
1993e73a819dSTao Ma
1994e73a819dSTao Ma while (len) {
1995e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1996e73a819dSTao Ma cpos, len, &rec, &index,
1997e73a819dSTao Ma &ref_leaf_bh);
1998e73a819dSTao Ma if (ret) {
1999e73a819dSTao Ma mlog_errno(ret);
2000e73a819dSTao Ma goto out;
2001e73a819dSTao Ma }
2002e73a819dSTao Ma
2003e73a819dSTao Ma set_len = le32_to_cpu(rec.r_clusters);
2004e73a819dSTao Ma
2005e73a819dSTao Ma /*
2006e73a819dSTao Ma * Here we may meet with 3 situations:
2007e73a819dSTao Ma *
2008e73a819dSTao Ma * 1. If we find an already existing record, and the length
2009e73a819dSTao Ma * is the same, cool, we just need to increase the r_refcount
2010e73a819dSTao Ma * and it is OK.
2011e73a819dSTao Ma * 2. If we find a hole, just insert it with r_refcount = 1.
2012e73a819dSTao Ma * 3. If we are in the middle of one extent record, split
2013e73a819dSTao Ma * it.
2014e73a819dSTao Ma */
2015e73a819dSTao Ma if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
2016e73a819dSTao Ma set_len <= len) {
2017198aac28STao Ma trace_ocfs2_increase_refcount_change(
2018198aac28STao Ma (unsigned long long)cpos, set_len,
2019e73a819dSTao Ma le32_to_cpu(rec.r_refcount));
2020e73a819dSTao Ma ret = ocfs2_change_refcount_rec(handle, ci,
20217540c1a7STao Ma ref_leaf_bh, index,
20227540c1a7STao Ma merge, 1);
2023e73a819dSTao Ma if (ret) {
2024e73a819dSTao Ma mlog_errno(ret);
2025e73a819dSTao Ma goto out;
2026e73a819dSTao Ma }
2027e73a819dSTao Ma } else if (!rec.r_refcount) {
2028e73a819dSTao Ma rec.r_refcount = cpu_to_le32(1);
2029e73a819dSTao Ma
2030198aac28STao Ma trace_ocfs2_increase_refcount_insert(
2031e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos),
2032e73a819dSTao Ma set_len);
2033e73a819dSTao Ma ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
2034e73a819dSTao Ma ref_leaf_bh,
20357540c1a7STao Ma &rec, index,
20367540c1a7STao Ma merge, meta_ac);
2037e73a819dSTao Ma if (ret) {
2038e73a819dSTao Ma mlog_errno(ret);
2039e73a819dSTao Ma goto out;
2040e73a819dSTao Ma }
2041e73a819dSTao Ma } else {
2042e73a819dSTao Ma set_len = min((u64)(cpos + len),
2043e73a819dSTao Ma le64_to_cpu(rec.r_cpos) + set_len) - cpos;
2044e73a819dSTao Ma rec.r_cpos = cpu_to_le64(cpos);
2045e73a819dSTao Ma rec.r_clusters = cpu_to_le32(set_len);
2046e73a819dSTao Ma le32_add_cpu(&rec.r_refcount, 1);
2047e73a819dSTao Ma
2048198aac28STao Ma trace_ocfs2_increase_refcount_split(
2049e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos),
2050e73a819dSTao Ma set_len, le32_to_cpu(rec.r_refcount));
2051e73a819dSTao Ma ret = ocfs2_split_refcount_rec(handle, ci,
2052e73a819dSTao Ma ref_root_bh, ref_leaf_bh,
20537540c1a7STao Ma &rec, index, merge,
2054e73a819dSTao Ma meta_ac, dealloc);
2055e73a819dSTao Ma if (ret) {
2056e73a819dSTao Ma mlog_errno(ret);
2057e73a819dSTao Ma goto out;
2058e73a819dSTao Ma }
2059e73a819dSTao Ma }
2060e73a819dSTao Ma
2061e73a819dSTao Ma cpos += set_len;
2062e73a819dSTao Ma len -= set_len;
2063e73a819dSTao Ma brelse(ref_leaf_bh);
2064e73a819dSTao Ma ref_leaf_bh = NULL;
2065e73a819dSTao Ma }
2066e73a819dSTao Ma
2067e73a819dSTao Ma out:
2068e73a819dSTao Ma brelse(ref_leaf_bh);
2069e73a819dSTao Ma return ret;
2070e73a819dSTao Ma }
20711823cb0bSTao Ma
ocfs2_remove_refcount_extent(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head * ref_leaf_bh,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc)20721823cb0bSTao Ma static int ocfs2_remove_refcount_extent(handle_t *handle,
20731823cb0bSTao Ma struct ocfs2_caching_info *ci,
20741823cb0bSTao Ma struct buffer_head *ref_root_bh,
20751823cb0bSTao Ma struct buffer_head *ref_leaf_bh,
20761823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac,
20771823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
20781823cb0bSTao Ma {
20791823cb0bSTao Ma int ret;
20801823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
20811823cb0bSTao Ma struct ocfs2_refcount_block *rb =
20821823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
20831823cb0bSTao Ma struct ocfs2_extent_tree et;
20841823cb0bSTao Ma
20851823cb0bSTao Ma BUG_ON(rb->rf_records.rl_used);
20861823cb0bSTao Ma
2087198aac28STao Ma trace_ocfs2_remove_refcount_extent(
2088198aac28STao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci),
2089198aac28STao Ma (unsigned long long)ref_leaf_bh->b_blocknr,
2090198aac28STao Ma le32_to_cpu(rb->rf_cpos));
2091198aac28STao Ma
20921823cb0bSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
20931823cb0bSTao Ma ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
20941823cb0bSTao Ma 1, meta_ac, dealloc);
20951823cb0bSTao Ma if (ret) {
20961823cb0bSTao Ma mlog_errno(ret);
20971823cb0bSTao Ma goto out;
20981823cb0bSTao Ma }
20991823cb0bSTao Ma
21001823cb0bSTao Ma ocfs2_remove_from_cache(ci, ref_leaf_bh);
21011823cb0bSTao Ma
21021823cb0bSTao Ma /*
21031823cb0bSTao Ma * add the freed block to the dealloc so that it will be freed
21041823cb0bSTao Ma * when we run dealloc.
21051823cb0bSTao Ma */
21061823cb0bSTao Ma ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
21071823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_slot),
210874380c47STao Ma le64_to_cpu(rb->rf_suballoc_loc),
21091823cb0bSTao Ma le64_to_cpu(rb->rf_blkno),
21101823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_bit));
21111823cb0bSTao Ma if (ret) {
21121823cb0bSTao Ma mlog_errno(ret);
21131823cb0bSTao Ma goto out;
21141823cb0bSTao Ma }
21151823cb0bSTao Ma
21161823cb0bSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
21171823cb0bSTao Ma OCFS2_JOURNAL_ACCESS_WRITE);
21181823cb0bSTao Ma if (ret) {
21191823cb0bSTao Ma mlog_errno(ret);
21201823cb0bSTao Ma goto out;
21211823cb0bSTao Ma }
21221823cb0bSTao Ma
21231823cb0bSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
21241823cb0bSTao Ma
21251823cb0bSTao Ma le32_add_cpu(&rb->rf_clusters, -1);
21261823cb0bSTao Ma
21271823cb0bSTao Ma /*
21281823cb0bSTao Ma * check whether we need to restore the root refcount block if
21291823cb0bSTao Ma * there is no leaf extent block at atll.
21301823cb0bSTao Ma */
21311823cb0bSTao Ma if (!rb->rf_list.l_next_free_rec) {
21321823cb0bSTao Ma BUG_ON(rb->rf_clusters);
21331823cb0bSTao Ma
2134198aac28STao Ma trace_ocfs2_restore_refcount_block(
21351823cb0bSTao Ma (unsigned long long)ref_root_bh->b_blocknr);
21361823cb0bSTao Ma
21371823cb0bSTao Ma rb->rf_flags = 0;
21381823cb0bSTao Ma rb->rf_parent = 0;
21391823cb0bSTao Ma rb->rf_cpos = 0;
21401823cb0bSTao Ma memset(&rb->rf_records, 0, sb->s_blocksize -
21411823cb0bSTao Ma offsetof(struct ocfs2_refcount_block, rf_records));
21421823cb0bSTao Ma rb->rf_records.rl_count =
21431823cb0bSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
21441823cb0bSTao Ma }
21451823cb0bSTao Ma
21461823cb0bSTao Ma ocfs2_journal_dirty(handle, ref_root_bh);
21471823cb0bSTao Ma
21481823cb0bSTao Ma out:
21491823cb0bSTao Ma return ret;
21501823cb0bSTao Ma }
21511823cb0bSTao Ma
ocfs2_increase_refcount(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,u64 cpos,u32 len,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc)21527540c1a7STao Ma int ocfs2_increase_refcount(handle_t *handle,
21537540c1a7STao Ma struct ocfs2_caching_info *ci,
21547540c1a7STao Ma struct buffer_head *ref_root_bh,
21557540c1a7STao Ma u64 cpos, u32 len,
21567540c1a7STao Ma struct ocfs2_alloc_context *meta_ac,
21577540c1a7STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
21587540c1a7STao Ma {
21597540c1a7STao Ma return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
21607540c1a7STao Ma cpos, len, 1,
21617540c1a7STao Ma meta_ac, dealloc);
21627540c1a7STao Ma }
21637540c1a7STao Ma
ocfs2_decrease_refcount_rec(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,struct buffer_head * ref_leaf_bh,int index,u64 cpos,unsigned int len,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc)21641823cb0bSTao Ma static int ocfs2_decrease_refcount_rec(handle_t *handle,
21651823cb0bSTao Ma struct ocfs2_caching_info *ci,
21661823cb0bSTao Ma struct buffer_head *ref_root_bh,
21671823cb0bSTao Ma struct buffer_head *ref_leaf_bh,
21681823cb0bSTao Ma int index, u64 cpos, unsigned int len,
21691823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac,
21701823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
21711823cb0bSTao Ma {
21721823cb0bSTao Ma int ret;
21731823cb0bSTao Ma struct ocfs2_refcount_block *rb =
21741823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
21751823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
21761823cb0bSTao Ma
21771823cb0bSTao Ma BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
21781823cb0bSTao Ma BUG_ON(cpos + len >
21791823cb0bSTao Ma le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
21801823cb0bSTao Ma
2181198aac28STao Ma trace_ocfs2_decrease_refcount_rec(
2182198aac28STao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci),
2183198aac28STao Ma (unsigned long long)cpos, len);
2184198aac28STao Ma
21851823cb0bSTao Ma if (cpos == le64_to_cpu(rec->r_cpos) &&
21861823cb0bSTao Ma len == le32_to_cpu(rec->r_clusters))
21871823cb0bSTao Ma ret = ocfs2_change_refcount_rec(handle, ci,
21887540c1a7STao Ma ref_leaf_bh, index, 1, -1);
21891823cb0bSTao Ma else {
21901823cb0bSTao Ma struct ocfs2_refcount_rec split = *rec;
21911823cb0bSTao Ma split.r_cpos = cpu_to_le64(cpos);
21921823cb0bSTao Ma split.r_clusters = cpu_to_le32(len);
21931823cb0bSTao Ma
21941823cb0bSTao Ma le32_add_cpu(&split.r_refcount, -1);
21951823cb0bSTao Ma
21961823cb0bSTao Ma ret = ocfs2_split_refcount_rec(handle, ci,
21971823cb0bSTao Ma ref_root_bh, ref_leaf_bh,
21987540c1a7STao Ma &split, index, 1,
21991823cb0bSTao Ma meta_ac, dealloc);
22001823cb0bSTao Ma }
22011823cb0bSTao Ma
22021823cb0bSTao Ma if (ret) {
22031823cb0bSTao Ma mlog_errno(ret);
22041823cb0bSTao Ma goto out;
22051823cb0bSTao Ma }
22061823cb0bSTao Ma
22071823cb0bSTao Ma /* Remove the leaf refcount block if it contains no refcount record. */
22081823cb0bSTao Ma if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
22091823cb0bSTao Ma ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
22101823cb0bSTao Ma ref_leaf_bh, meta_ac,
22111823cb0bSTao Ma dealloc);
22121823cb0bSTao Ma if (ret)
22131823cb0bSTao Ma mlog_errno(ret);
22141823cb0bSTao Ma }
22151823cb0bSTao Ma
22161823cb0bSTao Ma out:
22171823cb0bSTao Ma return ret;
22181823cb0bSTao Ma }
22191823cb0bSTao Ma
__ocfs2_decrease_refcount(handle_t * handle,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,u64 cpos,u32 len,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc,int delete)22201823cb0bSTao Ma static int __ocfs2_decrease_refcount(handle_t *handle,
22211823cb0bSTao Ma struct ocfs2_caching_info *ci,
22221823cb0bSTao Ma struct buffer_head *ref_root_bh,
22231823cb0bSTao Ma u64 cpos, u32 len,
22241823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac,
22256ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc,
22266ae23c55STao Ma int delete)
22271823cb0bSTao Ma {
22281823cb0bSTao Ma int ret = 0, index = 0;
22291823cb0bSTao Ma struct ocfs2_refcount_rec rec;
22301823cb0bSTao Ma unsigned int r_count = 0, r_len;
22311823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
22321823cb0bSTao Ma struct buffer_head *ref_leaf_bh = NULL;
22331823cb0bSTao Ma
2234198aac28STao Ma trace_ocfs2_decrease_refcount(
22351823cb0bSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci),
22366ae23c55STao Ma (unsigned long long)cpos, len, delete);
22371823cb0bSTao Ma
22381823cb0bSTao Ma while (len) {
22391823cb0bSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
22401823cb0bSTao Ma cpos, len, &rec, &index,
22411823cb0bSTao Ma &ref_leaf_bh);
22421823cb0bSTao Ma if (ret) {
22431823cb0bSTao Ma mlog_errno(ret);
22441823cb0bSTao Ma goto out;
22451823cb0bSTao Ma }
22461823cb0bSTao Ma
22471823cb0bSTao Ma r_count = le32_to_cpu(rec.r_refcount);
22481823cb0bSTao Ma BUG_ON(r_count == 0);
22496ae23c55STao Ma if (!delete)
22506ae23c55STao Ma BUG_ON(r_count > 1);
22511823cb0bSTao Ma
22521823cb0bSTao Ma r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
22531823cb0bSTao Ma le32_to_cpu(rec.r_clusters)) - cpos;
22541823cb0bSTao Ma
22551823cb0bSTao Ma ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
22561823cb0bSTao Ma ref_leaf_bh, index,
22571823cb0bSTao Ma cpos, r_len,
22581823cb0bSTao Ma meta_ac, dealloc);
22591823cb0bSTao Ma if (ret) {
22601823cb0bSTao Ma mlog_errno(ret);
22611823cb0bSTao Ma goto out;
22621823cb0bSTao Ma }
22631823cb0bSTao Ma
22646ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
22651823cb0bSTao Ma ret = ocfs2_cache_cluster_dealloc(dealloc,
22661823cb0bSTao Ma ocfs2_clusters_to_blocks(sb, cpos),
22671823cb0bSTao Ma r_len);
22681823cb0bSTao Ma if (ret) {
22691823cb0bSTao Ma mlog_errno(ret);
22701823cb0bSTao Ma goto out;
22711823cb0bSTao Ma }
22721823cb0bSTao Ma }
22731823cb0bSTao Ma
22741823cb0bSTao Ma cpos += r_len;
22751823cb0bSTao Ma len -= r_len;
22761823cb0bSTao Ma brelse(ref_leaf_bh);
22771823cb0bSTao Ma ref_leaf_bh = NULL;
22781823cb0bSTao Ma }
22791823cb0bSTao Ma
22801823cb0bSTao Ma out:
22811823cb0bSTao Ma brelse(ref_leaf_bh);
22821823cb0bSTao Ma return ret;
22831823cb0bSTao Ma }
22841823cb0bSTao Ma
22851823cb0bSTao Ma /* Caller must hold refcount tree lock. */
ocfs2_decrease_refcount(struct inode * inode,handle_t * handle,u32 cpos,u32 len,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc,int delete)22861823cb0bSTao Ma int ocfs2_decrease_refcount(struct inode *inode,
22871823cb0bSTao Ma handle_t *handle, u32 cpos, u32 len,
22881823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac,
22896ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc,
22906ae23c55STao Ma int delete)
22911823cb0bSTao Ma {
22921823cb0bSTao Ma int ret;
22931823cb0bSTao Ma u64 ref_blkno;
22941823cb0bSTao Ma struct buffer_head *ref_root_bh = NULL;
22951823cb0bSTao Ma struct ocfs2_refcount_tree *tree;
22961823cb0bSTao Ma
229784e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode));
22981823cb0bSTao Ma
22991823cb0bSTao Ma ret = ocfs2_get_refcount_block(inode, &ref_blkno);
23001823cb0bSTao Ma if (ret) {
23011823cb0bSTao Ma mlog_errno(ret);
23021823cb0bSTao Ma goto out;
23031823cb0bSTao Ma }
23041823cb0bSTao Ma
23051823cb0bSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
23061823cb0bSTao Ma if (ret) {
23071823cb0bSTao Ma mlog_errno(ret);
23081823cb0bSTao Ma goto out;
23091823cb0bSTao Ma }
23101823cb0bSTao Ma
23111823cb0bSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
23121823cb0bSTao Ma &ref_root_bh);
23131823cb0bSTao Ma if (ret) {
23141823cb0bSTao Ma mlog_errno(ret);
23151823cb0bSTao Ma goto out;
23161823cb0bSTao Ma }
23171823cb0bSTao Ma
23181823cb0bSTao Ma ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
23196ae23c55STao Ma cpos, len, meta_ac, dealloc, delete);
23201823cb0bSTao Ma if (ret)
23211823cb0bSTao Ma mlog_errno(ret);
23221823cb0bSTao Ma out:
23231823cb0bSTao Ma brelse(ref_root_bh);
23241823cb0bSTao Ma return ret;
23251823cb0bSTao Ma }
23261aa75feaSTao Ma
23271aa75feaSTao Ma /*
23281aa75feaSTao Ma * Mark the already-existing extent at cpos as refcounted for len clusters.
23291aa75feaSTao Ma * This adds the refcount extent flag.
23301aa75feaSTao Ma *
23311aa75feaSTao Ma * If the existing extent is larger than the request, initiate a
23321aa75feaSTao Ma * split. An attempt will be made at merging with adjacent extents.
23331aa75feaSTao Ma *
23341aa75feaSTao Ma * The caller is responsible for passing down meta_ac if we'll need it.
23351aa75feaSTao Ma */
ocfs2_mark_extent_refcounted(struct inode * inode,struct ocfs2_extent_tree * et,handle_t * handle,u32 cpos,u32 len,u32 phys,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc)23361aa75feaSTao Ma static int ocfs2_mark_extent_refcounted(struct inode *inode,
23371aa75feaSTao Ma struct ocfs2_extent_tree *et,
23381aa75feaSTao Ma handle_t *handle, u32 cpos,
23391aa75feaSTao Ma u32 len, u32 phys,
23401aa75feaSTao Ma struct ocfs2_alloc_context *meta_ac,
23411aa75feaSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
23421aa75feaSTao Ma {
23431aa75feaSTao Ma int ret;
23441aa75feaSTao Ma
2345198aac28STao Ma trace_ocfs2_mark_extent_refcounted(OCFS2_I(inode)->ip_blkno,
2346198aac28STao Ma cpos, len, phys);
23471aa75feaSTao Ma
23481aa75feaSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
23497ecef14aSJoe Perches 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",
23507ecef14aSJoe Perches inode->i_ino);
23511aa75feaSTao Ma goto out;
23521aa75feaSTao Ma }
23531aa75feaSTao Ma
23541aa75feaSTao Ma ret = ocfs2_change_extent_flag(handle, et, cpos,
23551aa75feaSTao Ma len, phys, meta_ac, dealloc,
23561aa75feaSTao Ma OCFS2_EXT_REFCOUNTED, 0);
23571aa75feaSTao Ma if (ret)
23581aa75feaSTao Ma mlog_errno(ret);
23591aa75feaSTao Ma
23601aa75feaSTao Ma out:
23611aa75feaSTao Ma return ret;
23621aa75feaSTao Ma }
2363bcbbb24aSTao Ma
2364bcbbb24aSTao Ma /*
2365bcbbb24aSTao Ma * Given some contiguous physical clusters, calculate what we need
2366bcbbb24aSTao Ma * for modifying their refcount.
2367bcbbb24aSTao Ma */
ocfs2_calc_refcount_meta_credits(struct super_block * sb,struct ocfs2_caching_info * ci,struct buffer_head * ref_root_bh,u64 start_cpos,u32 clusters,int * meta_add,int * credits)2368bcbbb24aSTao Ma static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2369bcbbb24aSTao Ma struct ocfs2_caching_info *ci,
2370bcbbb24aSTao Ma struct buffer_head *ref_root_bh,
2371bcbbb24aSTao Ma u64 start_cpos,
2372bcbbb24aSTao Ma u32 clusters,
2373bcbbb24aSTao Ma int *meta_add,
2374bcbbb24aSTao Ma int *credits)
2375bcbbb24aSTao Ma {
2376bcbbb24aSTao Ma int ret = 0, index, ref_blocks = 0, recs_add = 0;
2377bcbbb24aSTao Ma u64 cpos = start_cpos;
2378bcbbb24aSTao Ma struct ocfs2_refcount_block *rb;
2379bcbbb24aSTao Ma struct ocfs2_refcount_rec rec;
2380bcbbb24aSTao Ma struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2381bcbbb24aSTao Ma u32 len;
2382bcbbb24aSTao Ma
2383bcbbb24aSTao Ma while (clusters) {
2384bcbbb24aSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2385bcbbb24aSTao Ma cpos, clusters, &rec,
2386bcbbb24aSTao Ma &index, &ref_leaf_bh);
2387bcbbb24aSTao Ma if (ret) {
2388bcbbb24aSTao Ma mlog_errno(ret);
2389bcbbb24aSTao Ma goto out;
2390bcbbb24aSTao Ma }
2391bcbbb24aSTao Ma
2392bcbbb24aSTao Ma if (ref_leaf_bh != prev_bh) {
2393bcbbb24aSTao Ma /*
2394bcbbb24aSTao Ma * Now we encounter a new leaf block, so calculate
2395bcbbb24aSTao Ma * whether we need to extend the old leaf.
2396bcbbb24aSTao Ma */
2397bcbbb24aSTao Ma if (prev_bh) {
2398bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)
2399bcbbb24aSTao Ma prev_bh->b_data;
2400bcbbb24aSTao Ma
2401e1bf4cc6SAl Viro if (le16_to_cpu(rb->rf_records.rl_used) +
2402bcbbb24aSTao Ma recs_add >
2403bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count))
2404bcbbb24aSTao Ma ref_blocks++;
2405bcbbb24aSTao Ma }
2406bcbbb24aSTao Ma
2407bcbbb24aSTao Ma recs_add = 0;
2408bcbbb24aSTao Ma *credits += 1;
2409bcbbb24aSTao Ma brelse(prev_bh);
2410bcbbb24aSTao Ma prev_bh = ref_leaf_bh;
2411bcbbb24aSTao Ma get_bh(prev_bh);
2412bcbbb24aSTao Ma }
2413bcbbb24aSTao Ma
2414198aac28STao Ma trace_ocfs2_calc_refcount_meta_credits_iterate(
2415bcbbb24aSTao Ma recs_add, (unsigned long long)cpos, clusters,
2416bcbbb24aSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos),
2417bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters),
2418bcbbb24aSTao Ma le32_to_cpu(rec.r_refcount), index);
2419bcbbb24aSTao Ma
2420bcbbb24aSTao Ma len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2421bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters)) - cpos;
2422bcbbb24aSTao Ma /*
2423bcbbb24aSTao Ma * We record all the records which will be inserted to the
2424bcbbb24aSTao Ma * same refcount block, so that we can tell exactly whether
2425bcbbb24aSTao Ma * we need a new refcount block or not.
24268a2e70c4STao Ma *
24278a2e70c4STao Ma * If we will insert a new one, this is easy and only happens
24288a2e70c4STao Ma * during adding refcounted flag to the extent, so we don't
24298a2e70c4STao Ma * have a chance of spliting. We just need one record.
24308a2e70c4STao Ma *
24318a2e70c4STao Ma * If the refcount rec already exists, that would be a little
24328a2e70c4STao Ma * complicated. we may have to:
24338a2e70c4STao Ma * 1) split at the beginning if the start pos isn't aligned.
24348a2e70c4STao Ma * we need 1 more record in this case.
24358a2e70c4STao Ma * 2) split int the end if the end pos isn't aligned.
24368a2e70c4STao Ma * we need 1 more record in this case.
24378a2e70c4STao Ma * 3) split in the middle because of file system fragmentation.
24388a2e70c4STao Ma * we need 2 more records in this case(we can't detect this
24398a2e70c4STao Ma * beforehand, so always think of the worst case).
2440bcbbb24aSTao Ma */
2441bcbbb24aSTao Ma if (rec.r_refcount) {
24428a2e70c4STao Ma recs_add += 2;
2443bcbbb24aSTao Ma /* Check whether we need a split at the beginning. */
2444bcbbb24aSTao Ma if (cpos == start_cpos &&
2445bcbbb24aSTao Ma cpos != le64_to_cpu(rec.r_cpos))
2446bcbbb24aSTao Ma recs_add++;
2447bcbbb24aSTao Ma
2448bcbbb24aSTao Ma /* Check whether we need a split in the end. */
2449bcbbb24aSTao Ma if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2450bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters))
2451bcbbb24aSTao Ma recs_add++;
2452bcbbb24aSTao Ma } else
2453bcbbb24aSTao Ma recs_add++;
2454bcbbb24aSTao Ma
2455bcbbb24aSTao Ma brelse(ref_leaf_bh);
2456bcbbb24aSTao Ma ref_leaf_bh = NULL;
2457bcbbb24aSTao Ma clusters -= len;
2458bcbbb24aSTao Ma cpos += len;
2459bcbbb24aSTao Ma }
2460bcbbb24aSTao Ma
2461bcbbb24aSTao Ma if (prev_bh) {
2462bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2463bcbbb24aSTao Ma
2464e1bf4cc6SAl Viro if (le16_to_cpu(rb->rf_records.rl_used) + recs_add >
2465bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count))
2466bcbbb24aSTao Ma ref_blocks++;
2467bcbbb24aSTao Ma
2468bcbbb24aSTao Ma *credits += 1;
2469bcbbb24aSTao Ma }
2470bcbbb24aSTao Ma
2471bcbbb24aSTao Ma if (!ref_blocks)
2472bcbbb24aSTao Ma goto out;
2473bcbbb24aSTao Ma
2474bcbbb24aSTao Ma *meta_add += ref_blocks;
2475bcbbb24aSTao Ma *credits += ref_blocks;
2476bcbbb24aSTao Ma
2477bcbbb24aSTao Ma /*
2478bcbbb24aSTao Ma * So we may need ref_blocks to insert into the tree.
2479bcbbb24aSTao Ma * That also means we need to change the b-tree and add that number
2480bcbbb24aSTao Ma * of records since we never merge them.
2481bcbbb24aSTao Ma * We need one more block for expansion since the new created leaf
2482bcbbb24aSTao Ma * block is also full and needs split.
2483bcbbb24aSTao Ma */
2484bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2485bcbbb24aSTao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2486bcbbb24aSTao Ma struct ocfs2_extent_tree et;
2487bcbbb24aSTao Ma
2488bcbbb24aSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2489bcbbb24aSTao Ma *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2490bcbbb24aSTao Ma *credits += ocfs2_calc_extend_credits(sb,
249106f9da6eSGoldwyn Rodrigues et.et_root_el);
2492bcbbb24aSTao Ma } else {
2493bcbbb24aSTao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2494bcbbb24aSTao Ma *meta_add += 1;
2495bcbbb24aSTao Ma }
2496bcbbb24aSTao Ma
2497bcbbb24aSTao Ma out:
2498198aac28STao Ma
2499198aac28STao Ma trace_ocfs2_calc_refcount_meta_credits(
2500198aac28STao Ma (unsigned long long)start_cpos, clusters,
2501198aac28STao Ma *meta_add, *credits);
2502bcbbb24aSTao Ma brelse(ref_leaf_bh);
2503bcbbb24aSTao Ma brelse(prev_bh);
2504bcbbb24aSTao Ma return ret;
2505bcbbb24aSTao Ma }
2506bcbbb24aSTao Ma
2507bcbbb24aSTao Ma /*
2508bcbbb24aSTao Ma * For refcount tree, we will decrease some contiguous clusters
2509bcbbb24aSTao Ma * refcount count, so just go through it to see how many blocks
2510bcbbb24aSTao Ma * we gonna touch and whether we need to create new blocks.
2511bcbbb24aSTao Ma *
2512bcbbb24aSTao Ma * Normally the refcount blocks store these refcount should be
2513af901ca1SAndré Goddard Rosa * contiguous also, so that we can get the number easily.
251478f94673STristan Ye * We will at most add split 2 refcount records and 2 more
251578f94673STristan Ye * refcount blocks, so just check it in a rough way.
2516bcbbb24aSTao Ma *
2517bcbbb24aSTao Ma * Caller must hold refcount tree lock.
2518bcbbb24aSTao Ma */
ocfs2_prepare_refcount_change_for_del(struct inode * inode,u64 refcount_loc,u64 phys_blkno,u32 clusters,int * credits,int * ref_blocks)2519bcbbb24aSTao Ma int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
252078f94673STristan Ye u64 refcount_loc,
2521bcbbb24aSTao Ma u64 phys_blkno,
2522bcbbb24aSTao Ma u32 clusters,
2523bcbbb24aSTao Ma int *credits,
252478f94673STristan Ye int *ref_blocks)
2525bcbbb24aSTao Ma {
252678f94673STristan Ye int ret;
2527bcbbb24aSTao Ma struct buffer_head *ref_root_bh = NULL;
2528bcbbb24aSTao Ma struct ocfs2_refcount_tree *tree;
2529bcbbb24aSTao Ma u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2530bcbbb24aSTao Ma
2531bcbbb24aSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
25327ecef14aSJoe Perches 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",
25337ecef14aSJoe Perches inode->i_ino);
2534bcbbb24aSTao Ma goto out;
2535bcbbb24aSTao Ma }
2536bcbbb24aSTao Ma
253784e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode));
2538bcbbb24aSTao Ma
2539bcbbb24aSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
254078f94673STristan Ye refcount_loc, &tree);
2541bcbbb24aSTao Ma if (ret) {
2542bcbbb24aSTao Ma mlog_errno(ret);
2543bcbbb24aSTao Ma goto out;
2544bcbbb24aSTao Ma }
2545bcbbb24aSTao Ma
254678f94673STristan Ye ret = ocfs2_read_refcount_block(&tree->rf_ci, refcount_loc,
2547bcbbb24aSTao Ma &ref_root_bh);
2548bcbbb24aSTao Ma if (ret) {
2549bcbbb24aSTao Ma mlog_errno(ret);
2550bcbbb24aSTao Ma goto out;
2551bcbbb24aSTao Ma }
2552bcbbb24aSTao Ma
2553bcbbb24aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2554bcbbb24aSTao Ma &tree->rf_ci,
2555bcbbb24aSTao Ma ref_root_bh,
2556bcbbb24aSTao Ma start_cpos, clusters,
255778f94673STristan Ye ref_blocks, credits);
2558bcbbb24aSTao Ma if (ret) {
2559bcbbb24aSTao Ma mlog_errno(ret);
2560bcbbb24aSTao Ma goto out;
2561bcbbb24aSTao Ma }
2562bcbbb24aSTao Ma
2563198aac28STao Ma trace_ocfs2_prepare_refcount_change_for_del(*ref_blocks, *credits);
2564bcbbb24aSTao Ma
2565bcbbb24aSTao Ma out:
2566bcbbb24aSTao Ma brelse(ref_root_bh);
2567bcbbb24aSTao Ma return ret;
2568bcbbb24aSTao Ma }
25696f70fa51STao Ma
25706f70fa51STao Ma #define MAX_CONTIG_BYTES 1048576
25716f70fa51STao Ma
ocfs2_cow_contig_clusters(struct super_block * sb)25726f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb)
25736f70fa51STao Ma {
25746f70fa51STao Ma return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES);
25756f70fa51STao Ma }
25766f70fa51STao Ma
ocfs2_cow_contig_mask(struct super_block * sb)25776f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb)
25786f70fa51STao Ma {
25796f70fa51STao Ma return ~(ocfs2_cow_contig_clusters(sb) - 1);
25806f70fa51STao Ma }
25816f70fa51STao Ma
25826f70fa51STao Ma /*
25836f70fa51STao Ma * Given an extent that starts at 'start' and an I/O that starts at 'cpos',
25846f70fa51STao Ma * find an offset (start + (n * contig_clusters)) that is closest to cpos
25856f70fa51STao Ma * while still being less than or equal to it.
25866f70fa51STao Ma *
25876f70fa51STao Ma * The goal is to break the extent at a multiple of contig_clusters.
25886f70fa51STao Ma */
ocfs2_cow_align_start(struct super_block * sb,unsigned int start,unsigned int cpos)25896f70fa51STao Ma static inline unsigned int ocfs2_cow_align_start(struct super_block *sb,
25906f70fa51STao Ma unsigned int start,
25916f70fa51STao Ma unsigned int cpos)
25926f70fa51STao Ma {
25936f70fa51STao Ma BUG_ON(start > cpos);
25946f70fa51STao Ma
25956f70fa51STao Ma return start + ((cpos - start) & ocfs2_cow_contig_mask(sb));
25966f70fa51STao Ma }
25976f70fa51STao Ma
25986f70fa51STao Ma /*
25996f70fa51STao Ma * Given a cluster count of len, pad it out so that it is a multiple
26006f70fa51STao Ma * of contig_clusters.
26016f70fa51STao Ma */
ocfs2_cow_align_length(struct super_block * sb,unsigned int len)26026f70fa51STao Ma static inline unsigned int ocfs2_cow_align_length(struct super_block *sb,
26036f70fa51STao Ma unsigned int len)
26046f70fa51STao Ma {
26056f70fa51STao Ma unsigned int padded =
26066f70fa51STao Ma (len + (ocfs2_cow_contig_clusters(sb) - 1)) &
26076f70fa51STao Ma ocfs2_cow_contig_mask(sb);
26086f70fa51STao Ma
26096f70fa51STao Ma /* Did we wrap? */
26106f70fa51STao Ma if (padded < len)
26116f70fa51STao Ma padded = UINT_MAX;
26126f70fa51STao Ma
26136f70fa51STao Ma return padded;
26146f70fa51STao Ma }
26156f70fa51STao Ma
26166f70fa51STao Ma /*
2617e77999c1Swangjianli * Calculate out the start and number of virtual clusters we need to CoW.
26186f70fa51STao Ma *
26196f70fa51STao Ma * cpos is vitual start cluster position we want to do CoW in a
26206f70fa51STao Ma * file and write_len is the cluster length.
262137f8a2bfSTao Ma * max_cpos is the place where we want to stop CoW intentionally.
26226f70fa51STao Ma *
26236f70fa51STao Ma * Normal we will start CoW from the beginning of extent record cotaining cpos.
26246f70fa51STao Ma * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we
26256f70fa51STao Ma * get good I/O from the resulting extent tree.
26266f70fa51STao Ma */
ocfs2_refcount_cal_cow_clusters(struct inode * inode,struct ocfs2_extent_list * el,u32 cpos,u32 write_len,u32 max_cpos,u32 * cow_start,u32 * cow_len)26276f70fa51STao Ma static int ocfs2_refcount_cal_cow_clusters(struct inode *inode,
2628913580b4STao Ma struct ocfs2_extent_list *el,
26296f70fa51STao Ma u32 cpos,
26306f70fa51STao Ma u32 write_len,
263137f8a2bfSTao Ma u32 max_cpos,
26326f70fa51STao Ma u32 *cow_start,
26336f70fa51STao Ma u32 *cow_len)
26346f70fa51STao Ma {
26356f70fa51STao Ma int ret = 0;
26366f70fa51STao Ma int tree_height = le16_to_cpu(el->l_tree_depth), i;
26376f70fa51STao Ma struct buffer_head *eb_bh = NULL;
26386f70fa51STao Ma struct ocfs2_extent_block *eb = NULL;
26396f70fa51STao Ma struct ocfs2_extent_rec *rec;
26406f70fa51STao Ma unsigned int want_clusters, rec_end = 0;
26416f70fa51STao Ma int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb);
26426f70fa51STao Ma int leaf_clusters;
26436f70fa51STao Ma
264437f8a2bfSTao Ma BUG_ON(cpos + write_len > max_cpos);
264537f8a2bfSTao Ma
26466f70fa51STao Ma if (tree_height > 0) {
26476f70fa51STao Ma ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh);
26486f70fa51STao Ma if (ret) {
26496f70fa51STao Ma mlog_errno(ret);
26506f70fa51STao Ma goto out;
26516f70fa51STao Ma }
26526f70fa51STao Ma
26536f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data;
26546f70fa51STao Ma el = &eb->h_list;
26556f70fa51STao Ma
26566f70fa51STao Ma if (el->l_tree_depth) {
265717a5b9abSGoldwyn Rodrigues ret = ocfs2_error(inode->i_sb,
26587ecef14aSJoe Perches "Inode %lu has non zero tree depth in leaf block %llu\n",
26597ecef14aSJoe Perches inode->i_ino,
26606f70fa51STao Ma (unsigned long long)eb_bh->b_blocknr);
26616f70fa51STao Ma goto out;
26626f70fa51STao Ma }
26636f70fa51STao Ma }
26646f70fa51STao Ma
26656f70fa51STao Ma *cow_len = 0;
26666f70fa51STao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
26676f70fa51STao Ma rec = &el->l_recs[i];
26686f70fa51STao Ma
26696f70fa51STao Ma if (ocfs2_is_empty_extent(rec)) {
26706f70fa51STao Ma mlog_bug_on_msg(i != 0, "Inode %lu has empty record in "
26716f70fa51STao Ma "index %d\n", inode->i_ino, i);
26726f70fa51STao Ma continue;
26736f70fa51STao Ma }
26746f70fa51STao Ma
26756f70fa51STao Ma if (le32_to_cpu(rec->e_cpos) +
26766f70fa51STao Ma le16_to_cpu(rec->e_leaf_clusters) <= cpos)
26776f70fa51STao Ma continue;
26786f70fa51STao Ma
26796f70fa51STao Ma if (*cow_len == 0) {
26806f70fa51STao Ma /*
26816f70fa51STao Ma * We should find a refcounted record in the
26826f70fa51STao Ma * first pass.
26836f70fa51STao Ma */
26846f70fa51STao Ma BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED));
26856f70fa51STao Ma *cow_start = le32_to_cpu(rec->e_cpos);
26866f70fa51STao Ma }
26876f70fa51STao Ma
26886f70fa51STao Ma /*
268937f8a2bfSTao Ma * If we encounter a hole, a non-refcounted record or
269037f8a2bfSTao Ma * pass the max_cpos, stop the search.
26916f70fa51STao Ma */
26926f70fa51STao Ma if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) ||
269337f8a2bfSTao Ma (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) ||
269437f8a2bfSTao Ma (max_cpos <= le32_to_cpu(rec->e_cpos)))
26956f70fa51STao Ma break;
26966f70fa51STao Ma
26976f70fa51STao Ma leaf_clusters = le16_to_cpu(rec->e_leaf_clusters);
26986f70fa51STao Ma rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters;
269937f8a2bfSTao Ma if (rec_end > max_cpos) {
270037f8a2bfSTao Ma rec_end = max_cpos;
270137f8a2bfSTao Ma leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos);
270237f8a2bfSTao Ma }
27036f70fa51STao Ma
27046f70fa51STao Ma /*
27056f70fa51STao Ma * How many clusters do we actually need from
27066f70fa51STao Ma * this extent? First we see how many we actually
27076f70fa51STao Ma * need to complete the write. If that's smaller
27086f70fa51STao Ma * than contig_clusters, we try for contig_clusters.
27096f70fa51STao Ma */
27106f70fa51STao Ma if (!*cow_len)
27116f70fa51STao Ma want_clusters = write_len;
27126f70fa51STao Ma else
27136f70fa51STao Ma want_clusters = (cpos + write_len) -
27146f70fa51STao Ma (*cow_start + *cow_len);
27156f70fa51STao Ma if (want_clusters < contig_clusters)
27166f70fa51STao Ma want_clusters = contig_clusters;
27176f70fa51STao Ma
27186f70fa51STao Ma /*
27196f70fa51STao Ma * If the write does not cover the whole extent, we
27206f70fa51STao Ma * need to calculate how we're going to split the extent.
27216f70fa51STao Ma * We try to do it on contig_clusters boundaries.
27226f70fa51STao Ma *
27236f70fa51STao Ma * Any extent smaller than contig_clusters will be
27246f70fa51STao Ma * CoWed in its entirety.
27256f70fa51STao Ma */
27266f70fa51STao Ma if (leaf_clusters <= contig_clusters)
27276f70fa51STao Ma *cow_len += leaf_clusters;
27286f70fa51STao Ma else if (*cow_len || (*cow_start == cpos)) {
27296f70fa51STao Ma /*
27306f70fa51STao Ma * This extent needs to be CoW'd from its
27316f70fa51STao Ma * beginning, so all we have to do is compute
27326f70fa51STao Ma * how many clusters to grab. We align
27336f70fa51STao Ma * want_clusters to the edge of contig_clusters
27346f70fa51STao Ma * to get better I/O.
27356f70fa51STao Ma */
27366f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb,
27376f70fa51STao Ma want_clusters);
27386f70fa51STao Ma
27396f70fa51STao Ma if (leaf_clusters < want_clusters)
27406f70fa51STao Ma *cow_len += leaf_clusters;
27416f70fa51STao Ma else
27426f70fa51STao Ma *cow_len += want_clusters;
27436f70fa51STao Ma } else if ((*cow_start + contig_clusters) >=
27446f70fa51STao Ma (cpos + write_len)) {
27456f70fa51STao Ma /*
27466f70fa51STao Ma * Breaking off contig_clusters at the front
27476f70fa51STao Ma * of the extent will cover our write. That's
27486f70fa51STao Ma * easy.
27496f70fa51STao Ma */
27506f70fa51STao Ma *cow_len = contig_clusters;
27516f70fa51STao Ma } else if ((rec_end - cpos) <= contig_clusters) {
27526f70fa51STao Ma /*
27536f70fa51STao Ma * Breaking off contig_clusters at the tail of
27546f70fa51STao Ma * this extent will cover cpos.
27556f70fa51STao Ma */
27566f70fa51STao Ma *cow_start = rec_end - contig_clusters;
27576f70fa51STao Ma *cow_len = contig_clusters;
27586f70fa51STao Ma } else if ((rec_end - cpos) <= want_clusters) {
27596f70fa51STao Ma /*
27606f70fa51STao Ma * While we can't fit the entire write in this
27616f70fa51STao Ma * extent, we know that the write goes from cpos
27626f70fa51STao Ma * to the end of the extent. Break that off.
27636f70fa51STao Ma * We try to break it at some multiple of
27646f70fa51STao Ma * contig_clusters from the front of the extent.
27656f70fa51STao Ma * Failing that (ie, cpos is within
27666f70fa51STao Ma * contig_clusters of the front), we'll CoW the
27676f70fa51STao Ma * entire extent.
27686f70fa51STao Ma */
27696f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb,
27706f70fa51STao Ma *cow_start, cpos);
27716f70fa51STao Ma *cow_len = rec_end - *cow_start;
27726f70fa51STao Ma } else {
27736f70fa51STao Ma /*
27746f70fa51STao Ma * Ok, the entire write lives in the middle of
27756f70fa51STao Ma * this extent. Let's try to slice the extent up
27766f70fa51STao Ma * nicely. Optimally, our CoW region starts at
27776f70fa51STao Ma * m*contig_clusters from the beginning of the
27786f70fa51STao Ma * extent and goes for n*contig_clusters,
27796f70fa51STao Ma * covering the entire write.
27806f70fa51STao Ma */
27816f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb,
27826f70fa51STao Ma *cow_start, cpos);
27836f70fa51STao Ma
27846f70fa51STao Ma want_clusters = (cpos + write_len) - *cow_start;
27856f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb,
27866f70fa51STao Ma want_clusters);
27876f70fa51STao Ma if (*cow_start + want_clusters <= rec_end)
27886f70fa51STao Ma *cow_len = want_clusters;
27896f70fa51STao Ma else
27906f70fa51STao Ma *cow_len = rec_end - *cow_start;
27916f70fa51STao Ma }
27926f70fa51STao Ma
27936f70fa51STao Ma /* Have we covered our entire write yet? */
27946f70fa51STao Ma if ((*cow_start + *cow_len) >= (cpos + write_len))
27956f70fa51STao Ma break;
27966f70fa51STao Ma
27976f70fa51STao Ma /*
27986f70fa51STao Ma * If we reach the end of the extent block and don't get enough
27996f70fa51STao Ma * clusters, continue with the next extent block if possible.
28006f70fa51STao Ma */
28016f70fa51STao Ma if (i + 1 == le16_to_cpu(el->l_next_free_rec) &&
28026f70fa51STao Ma eb && eb->h_next_leaf_blk) {
28036f70fa51STao Ma brelse(eb_bh);
28046f70fa51STao Ma eb_bh = NULL;
28056f70fa51STao Ma
28066f70fa51STao Ma ret = ocfs2_read_extent_block(INODE_CACHE(inode),
28076f70fa51STao Ma le64_to_cpu(eb->h_next_leaf_blk),
28086f70fa51STao Ma &eb_bh);
28096f70fa51STao Ma if (ret) {
28106f70fa51STao Ma mlog_errno(ret);
28116f70fa51STao Ma goto out;
28126f70fa51STao Ma }
28136f70fa51STao Ma
28146f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data;
28156f70fa51STao Ma el = &eb->h_list;
28166f70fa51STao Ma i = -1;
28176f70fa51STao Ma }
28186f70fa51STao Ma }
28196f70fa51STao Ma
28206f70fa51STao Ma out:
28216f70fa51STao Ma brelse(eb_bh);
28226f70fa51STao Ma return ret;
28236f70fa51STao Ma }
28246f70fa51STao Ma
28256f70fa51STao Ma /*
28266f70fa51STao Ma * Prepare meta_ac, data_ac and calculate credits when we want to add some
28276f70fa51STao Ma * num_clusters in data_tree "et" and change the refcount for the old
28286f70fa51STao Ma * clusters(starting form p_cluster) in the refcount tree.
28296f70fa51STao Ma *
28306f70fa51STao Ma * Note:
28316f70fa51STao Ma * 1. since we may split the old tree, so we at most will need num_clusters + 2
28326f70fa51STao Ma * more new leaf records.
28336f70fa51STao Ma * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so
28346f70fa51STao Ma * just give data_ac = NULL.
28356f70fa51STao Ma */
ocfs2_lock_refcount_allocators(struct super_block * sb,u32 p_cluster,u32 num_clusters,struct ocfs2_extent_tree * et,struct ocfs2_caching_info * ref_ci,struct buffer_head * ref_root_bh,struct ocfs2_alloc_context ** meta_ac,struct ocfs2_alloc_context ** data_ac,int * credits)28366f70fa51STao Ma static int ocfs2_lock_refcount_allocators(struct super_block *sb,
28376f70fa51STao Ma u32 p_cluster, u32 num_clusters,
28386f70fa51STao Ma struct ocfs2_extent_tree *et,
28396f70fa51STao Ma struct ocfs2_caching_info *ref_ci,
28406f70fa51STao Ma struct buffer_head *ref_root_bh,
28416f70fa51STao Ma struct ocfs2_alloc_context **meta_ac,
28426f70fa51STao Ma struct ocfs2_alloc_context **data_ac,
28436f70fa51STao Ma int *credits)
28446f70fa51STao Ma {
28456f70fa51STao Ma int ret = 0, meta_add = 0;
2846964f14a0SJun Piao int num_free_extents = ocfs2_num_free_extents(et);
28476f70fa51STao Ma
28486f70fa51STao Ma if (num_free_extents < 0) {
28496f70fa51STao Ma ret = num_free_extents;
28506f70fa51STao Ma mlog_errno(ret);
28516f70fa51STao Ma goto out;
28526f70fa51STao Ma }
28536f70fa51STao Ma
28546f70fa51STao Ma if (num_free_extents < num_clusters + 2)
28556f70fa51STao Ma meta_add =
28566f70fa51STao Ma ocfs2_extend_meta_needed(et->et_root_el);
28576f70fa51STao Ma
285806f9da6eSGoldwyn Rodrigues *credits += ocfs2_calc_extend_credits(sb, et->et_root_el);
28596f70fa51STao Ma
28606f70fa51STao Ma ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh,
28616f70fa51STao Ma p_cluster, num_clusters,
28626f70fa51STao Ma &meta_add, credits);
28636f70fa51STao Ma if (ret) {
28646f70fa51STao Ma mlog_errno(ret);
28656f70fa51STao Ma goto out;
28666f70fa51STao Ma }
28676f70fa51STao Ma
2868198aac28STao Ma trace_ocfs2_lock_refcount_allocators(meta_add, *credits);
28696f70fa51STao Ma ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add,
28706f70fa51STao Ma meta_ac);
28716f70fa51STao Ma if (ret) {
28726f70fa51STao Ma mlog_errno(ret);
28736f70fa51STao Ma goto out;
28746f70fa51STao Ma }
28756f70fa51STao Ma
28766f70fa51STao Ma if (data_ac) {
28776f70fa51STao Ma ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters,
28786f70fa51STao Ma data_ac);
28796f70fa51STao Ma if (ret)
28806f70fa51STao Ma mlog_errno(ret);
28816f70fa51STao Ma }
28826f70fa51STao Ma
28836f70fa51STao Ma out:
28846f70fa51STao Ma if (ret) {
28856f70fa51STao Ma if (*meta_ac) {
28866f70fa51STao Ma ocfs2_free_alloc_context(*meta_ac);
28876f70fa51STao Ma *meta_ac = NULL;
28886f70fa51STao Ma }
28896f70fa51STao Ma }
28906f70fa51STao Ma
28916f70fa51STao Ma return ret;
28926f70fa51STao Ma }
28936f70fa51STao Ma
ocfs2_clear_cow_buffer(handle_t * handle,struct buffer_head * bh)28946f70fa51STao Ma static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh)
28956f70fa51STao Ma {
28966f70fa51STao Ma BUG_ON(buffer_dirty(bh));
28976f70fa51STao Ma
28986f70fa51STao Ma clear_buffer_mapped(bh);
28996f70fa51STao Ma
29006f70fa51STao Ma return 0;
29016f70fa51STao Ma }
29026f70fa51STao Ma
ocfs2_duplicate_clusters_by_page(handle_t * handle,struct inode * inode,u32 cpos,u32 old_cluster,u32 new_cluster,u32 new_len)29033e19a25eSTristan Ye int ocfs2_duplicate_clusters_by_page(handle_t *handle,
2904c7dd3392STiger Yang struct inode *inode,
29056f70fa51STao Ma u32 cpos, u32 old_cluster,
29066f70fa51STao Ma u32 new_cluster, u32 new_len)
29076f70fa51STao Ma {
29086f70fa51STao Ma int ret = 0, partial;
2909c7dd3392STiger Yang struct super_block *sb = inode->i_sb;
29106f70fa51STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
29116f70fa51STao Ma struct page *page;
29126f70fa51STao Ma pgoff_t page_index;
2913262d8a87SJoseph Qi unsigned int from, to;
29146f70fa51STao Ma loff_t offset, end, map_end;
29153e19a25eSTristan Ye struct address_space *mapping = inode->i_mapping;
29166f70fa51STao Ma
2917198aac28STao Ma trace_ocfs2_duplicate_clusters_by_page(cpos, old_cluster,
2918198aac28STao Ma new_cluster, new_len);
29196f70fa51STao Ma
29206f70fa51STao Ma offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
29216f70fa51STao Ma end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2922f5e27b6dSTao Ma /*
2923f5e27b6dSTao Ma * We only duplicate pages until we reach the page contains i_size - 1.
2924f5e27b6dSTao Ma * So trim 'end' to i_size.
2925f5e27b6dSTao Ma */
29263e19a25eSTristan Ye if (end > i_size_read(inode))
29273e19a25eSTristan Ye end = i_size_read(inode);
29286f70fa51STao Ma
29296f70fa51STao Ma while (offset < end) {
293009cbfeafSKirill A. Shutemov page_index = offset >> PAGE_SHIFT;
293109cbfeafSKirill A. Shutemov map_end = ((loff_t)page_index + 1) << PAGE_SHIFT;
29326f70fa51STao Ma if (map_end > end)
29336f70fa51STao Ma map_end = end;
29346f70fa51STao Ma
29356f70fa51STao Ma /* from, to is the offset within the page. */
293609cbfeafSKirill A. Shutemov from = offset & (PAGE_SIZE - 1);
293709cbfeafSKirill A. Shutemov to = PAGE_SIZE;
293809cbfeafSKirill A. Shutemov if (map_end & (PAGE_SIZE - 1))
293909cbfeafSKirill A. Shutemov to = map_end & (PAGE_SIZE - 1);
29406f70fa51STao Ma
294169eb7765SLarry Chen retry:
29429b4c0ff3SJan Kara page = find_or_create_page(mapping, page_index, GFP_NOFS);
294362c61046SGu Zheng if (!page) {
294462c61046SGu Zheng ret = -ENOMEM;
294562c61046SGu Zheng mlog_errno(ret);
294662c61046SGu Zheng break;
294762c61046SGu Zheng }
29486f70fa51STao Ma
29490a1ea437STao Ma /*
295069eb7765SLarry Chen * In case PAGE_SIZE <= CLUSTER_SIZE, we do not expect a dirty
295169eb7765SLarry Chen * page, so write it back.
29520a1ea437STao Ma */
295369eb7765SLarry Chen if (PAGE_SIZE <= OCFS2_SB(sb)->s_clustersize) {
295469eb7765SLarry Chen if (PageDirty(page)) {
29558e4bfd13SChristoph Hellwig unlock_page(page);
29568e4bfd13SChristoph Hellwig put_page(page);
29578e4bfd13SChristoph Hellwig
29588e4bfd13SChristoph Hellwig ret = filemap_write_and_wait_range(mapping,
29598e4bfd13SChristoph Hellwig offset, map_end - 1);
296069eb7765SLarry Chen goto retry;
296169eb7765SLarry Chen }
296269eb7765SLarry Chen }
29636f70fa51STao Ma
29646f70fa51STao Ma if (!PageUptodate(page)) {
29652c69e205SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page);
29662c69e205SMatthew Wilcox (Oracle)
29672c69e205SMatthew Wilcox (Oracle) ret = block_read_full_folio(folio, ocfs2_get_block);
29686f70fa51STao Ma if (ret) {
29696f70fa51STao Ma mlog_errno(ret);
29706f70fa51STao Ma goto unlock;
29716f70fa51STao Ma }
29722c69e205SMatthew Wilcox (Oracle) folio_lock(folio);
29736f70fa51STao Ma }
29746f70fa51STao Ma
29756f70fa51STao Ma if (page_has_buffers(page)) {
29766f70fa51STao Ma ret = walk_page_buffers(handle, page_buffers(page),
29776f70fa51STao Ma from, to, &partial,
29786f70fa51STao Ma ocfs2_clear_cow_buffer);
29796f70fa51STao Ma if (ret) {
29806f70fa51STao Ma mlog_errno(ret);
29816f70fa51STao Ma goto unlock;
29826f70fa51STao Ma }
29836f70fa51STao Ma }
29846f70fa51STao Ma
2985c7dd3392STiger Yang ocfs2_map_and_dirty_page(inode,
2986c7dd3392STiger Yang handle, from, to,
29876f70fa51STao Ma page, 0, &new_block);
29886f70fa51STao Ma mark_page_accessed(page);
29896f70fa51STao Ma unlock:
29906f70fa51STao Ma unlock_page(page);
299109cbfeafSKirill A. Shutemov put_page(page);
29926f70fa51STao Ma page = NULL;
29936f70fa51STao Ma offset = map_end;
29946f70fa51STao Ma if (ret)
29956f70fa51STao Ma break;
29966f70fa51STao Ma }
29976f70fa51STao Ma
29986f70fa51STao Ma return ret;
29996f70fa51STao Ma }
30006f70fa51STao Ma
ocfs2_duplicate_clusters_by_jbd(handle_t * handle,struct inode * inode,u32 cpos,u32 old_cluster,u32 new_cluster,u32 new_len)30013e19a25eSTristan Ye int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
3002c7dd3392STiger Yang struct inode *inode,
3003492a8a33STao Ma u32 cpos, u32 old_cluster,
3004492a8a33STao Ma u32 new_cluster, u32 new_len)
3005492a8a33STao Ma {
3006492a8a33STao Ma int ret = 0;
30073e19a25eSTristan Ye struct super_block *sb = inode->i_sb;
30083e19a25eSTristan Ye struct ocfs2_caching_info *ci = INODE_CACHE(inode);
3009492a8a33STao Ma int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
3010492a8a33STao Ma u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
3011492a8a33STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
3012492a8a33STao Ma struct ocfs2_super *osb = OCFS2_SB(sb);
3013492a8a33STao Ma struct buffer_head *old_bh = NULL;
3014492a8a33STao Ma struct buffer_head *new_bh = NULL;
3015492a8a33STao Ma
3016198aac28STao Ma trace_ocfs2_duplicate_clusters_by_page(cpos, old_cluster,
3017492a8a33STao Ma new_cluster, new_len);
3018492a8a33STao Ma
3019492a8a33STao Ma for (i = 0; i < blocks; i++, old_block++, new_block++) {
3020492a8a33STao Ma new_bh = sb_getblk(osb->sb, new_block);
3021492a8a33STao Ma if (new_bh == NULL) {
30227391a294SRui Xiang ret = -ENOMEM;
3023492a8a33STao Ma mlog_errno(ret);
3024492a8a33STao Ma break;
3025492a8a33STao Ma }
3026492a8a33STao Ma
3027492a8a33STao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh);
3028492a8a33STao Ma
3029492a8a33STao Ma ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
3030492a8a33STao Ma if (ret) {
3031492a8a33STao Ma mlog_errno(ret);
3032492a8a33STao Ma break;
3033492a8a33STao Ma }
3034492a8a33STao Ma
3035492a8a33STao Ma ret = ocfs2_journal_access(handle, ci, new_bh,
3036492a8a33STao Ma OCFS2_JOURNAL_ACCESS_CREATE);
3037492a8a33STao Ma if (ret) {
3038492a8a33STao Ma mlog_errno(ret);
3039492a8a33STao Ma break;
3040492a8a33STao Ma }
3041492a8a33STao Ma
3042492a8a33STao Ma memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
3043ec20cec7SJoel Becker ocfs2_journal_dirty(handle, new_bh);
3044492a8a33STao Ma
3045492a8a33STao Ma brelse(new_bh);
3046492a8a33STao Ma brelse(old_bh);
3047492a8a33STao Ma new_bh = NULL;
3048492a8a33STao Ma old_bh = NULL;
3049492a8a33STao Ma }
3050492a8a33STao Ma
3051492a8a33STao Ma brelse(new_bh);
3052492a8a33STao Ma brelse(old_bh);
3053492a8a33STao Ma return ret;
3054492a8a33STao Ma }
3055492a8a33STao Ma
ocfs2_clear_ext_refcount(handle_t * handle,struct ocfs2_extent_tree * et,u32 cpos,u32 p_cluster,u32 len,unsigned int ext_flags,struct ocfs2_alloc_context * meta_ac,struct ocfs2_cached_dealloc_ctxt * dealloc)30566f70fa51STao Ma static int ocfs2_clear_ext_refcount(handle_t *handle,
30576f70fa51STao Ma struct ocfs2_extent_tree *et,
30586f70fa51STao Ma u32 cpos, u32 p_cluster, u32 len,
30596f70fa51STao Ma unsigned int ext_flags,
30606f70fa51STao Ma struct ocfs2_alloc_context *meta_ac,
30616f70fa51STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
30626f70fa51STao Ma {
30636f70fa51STao Ma int ret, index;
30646f70fa51STao Ma struct ocfs2_extent_rec replace_rec;
30656f70fa51STao Ma struct ocfs2_path *path = NULL;
30666f70fa51STao Ma struct ocfs2_extent_list *el;
30676f70fa51STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
30686f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
30696f70fa51STao Ma
3070198aac28STao Ma trace_ocfs2_clear_ext_refcount((unsigned long long)ino,
3071198aac28STao Ma cpos, len, p_cluster, ext_flags);
30726f70fa51STao Ma
30736f70fa51STao Ma memset(&replace_rec, 0, sizeof(replace_rec));
30746f70fa51STao Ma replace_rec.e_cpos = cpu_to_le32(cpos);
30756f70fa51STao Ma replace_rec.e_leaf_clusters = cpu_to_le16(len);
30766f70fa51STao Ma replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
30776f70fa51STao Ma p_cluster));
30786f70fa51STao Ma replace_rec.e_flags = ext_flags;
30796f70fa51STao Ma replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
30806f70fa51STao Ma
30816f70fa51STao Ma path = ocfs2_new_path_from_et(et);
30826f70fa51STao Ma if (!path) {
30836f70fa51STao Ma ret = -ENOMEM;
30846f70fa51STao Ma mlog_errno(ret);
30856f70fa51STao Ma goto out;
30866f70fa51STao Ma }
30876f70fa51STao Ma
30886f70fa51STao Ma ret = ocfs2_find_path(et->et_ci, path, cpos);
30896f70fa51STao Ma if (ret) {
30906f70fa51STao Ma mlog_errno(ret);
30916f70fa51STao Ma goto out;
30926f70fa51STao Ma }
30936f70fa51STao Ma
30946f70fa51STao Ma el = path_leaf_el(path);
30956f70fa51STao Ma
30966f70fa51STao Ma index = ocfs2_search_extent_list(el, cpos);
3097981035b4SYingtai Xie if (index == -1) {
309817a5b9abSGoldwyn Rodrigues ret = ocfs2_error(sb,
30997ecef14aSJoe Perches "Inode %llu has an extent at cpos %u which can no longer be found\n",
31006f70fa51STao Ma (unsigned long long)ino, cpos);
31016f70fa51STao Ma goto out;
31026f70fa51STao Ma }
31036f70fa51STao Ma
31046f70fa51STao Ma ret = ocfs2_split_extent(handle, et, path, index,
31056f70fa51STao Ma &replace_rec, meta_ac, dealloc);
31066f70fa51STao Ma if (ret)
31076f70fa51STao Ma mlog_errno(ret);
31086f70fa51STao Ma
31096f70fa51STao Ma out:
31106f70fa51STao Ma ocfs2_free_path(path);
31116f70fa51STao Ma return ret;
31126f70fa51STao Ma }
31136f70fa51STao Ma
ocfs2_replace_clusters(handle_t * handle,struct ocfs2_cow_context * context,u32 cpos,u32 old,u32 new,u32 len,unsigned int ext_flags)31146f70fa51STao Ma static int ocfs2_replace_clusters(handle_t *handle,
31156f70fa51STao Ma struct ocfs2_cow_context *context,
31166f70fa51STao Ma u32 cpos, u32 old,
31176f70fa51STao Ma u32 new, u32 len,
31186f70fa51STao Ma unsigned int ext_flags)
31196f70fa51STao Ma {
31206f70fa51STao Ma int ret;
3121913580b4STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci;
31226f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(ci);
31236f70fa51STao Ma
3124198aac28STao Ma trace_ocfs2_replace_clusters((unsigned long long)ino,
3125198aac28STao Ma cpos, old, new, len, ext_flags);
31266f70fa51STao Ma
31276f70fa51STao Ma /*If the old clusters is unwritten, no need to duplicate. */
31286f70fa51STao Ma if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3129c7dd3392STiger Yang ret = context->cow_duplicate_clusters(handle, context->inode,
31303e19a25eSTristan Ye cpos, old, new, len);
31316f70fa51STao Ma if (ret) {
31326f70fa51STao Ma mlog_errno(ret);
31336f70fa51STao Ma goto out;
31346f70fa51STao Ma }
31356f70fa51STao Ma }
31366f70fa51STao Ma
3137913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
31386f70fa51STao Ma cpos, new, len, ext_flags,
31396f70fa51STao Ma context->meta_ac, &context->dealloc);
31406f70fa51STao Ma if (ret)
31416f70fa51STao Ma mlog_errno(ret);
31426f70fa51STao Ma out:
31436f70fa51STao Ma return ret;
31446f70fa51STao Ma }
31456f70fa51STao Ma
ocfs2_cow_sync_writeback(struct super_block * sb,struct inode * inode,u32 cpos,u32 num_clusters)31463e19a25eSTristan Ye int ocfs2_cow_sync_writeback(struct super_block *sb,
31473e19a25eSTristan Ye struct inode *inode,
31486f70fa51STao Ma u32 cpos, u32 num_clusters)
31496f70fa51STao Ma {
315097a3a383SMatthew Wilcox (Oracle) int ret;
315197a3a383SMatthew Wilcox (Oracle) loff_t start, end;
31526f70fa51STao Ma
31533e19a25eSTristan Ye if (ocfs2_should_order_data(inode))
31546f70fa51STao Ma return 0;
31556f70fa51STao Ma
315697a3a383SMatthew Wilcox (Oracle) start = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
315797a3a383SMatthew Wilcox (Oracle) end = start + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits) - 1;
31586f70fa51STao Ma
315997a3a383SMatthew Wilcox (Oracle) ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
316097a3a383SMatthew Wilcox (Oracle) if (ret < 0)
31616f70fa51STao Ma mlog_errno(ret);
31626f70fa51STao Ma
31636f70fa51STao Ma return ret;
31646f70fa51STao Ma }
31656f70fa51STao Ma
ocfs2_di_get_clusters(struct ocfs2_cow_context * context,u32 v_cluster,u32 * p_cluster,u32 * num_clusters,unsigned int * extent_flags)3166913580b4STao Ma static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3167913580b4STao Ma u32 v_cluster, u32 *p_cluster,
3168913580b4STao Ma u32 *num_clusters,
3169913580b4STao Ma unsigned int *extent_flags)
3170913580b4STao Ma {
3171913580b4STao Ma return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3172913580b4STao Ma num_clusters, extent_flags);
3173913580b4STao Ma }
3174913580b4STao Ma
ocfs2_make_clusters_writable(struct super_block * sb,struct ocfs2_cow_context * context,u32 cpos,u32 p_cluster,u32 num_clusters,unsigned int e_flags)31756f70fa51STao Ma static int ocfs2_make_clusters_writable(struct super_block *sb,
31766f70fa51STao Ma struct ocfs2_cow_context *context,
31776f70fa51STao Ma u32 cpos, u32 p_cluster,
31786f70fa51STao Ma u32 num_clusters, unsigned int e_flags)
31796f70fa51STao Ma {
31806ae23c55STao Ma int ret, delete, index, credits = 0;
3181acf3bb00STristan Ye u32 new_bit, new_len, orig_num_clusters;
31826ae23c55STao Ma unsigned int set_len;
31836f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(sb);
31846f70fa51STao Ma handle_t *handle;
31856ae23c55STao Ma struct buffer_head *ref_leaf_bh = NULL;
3186913580b4STao Ma struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
31876ae23c55STao Ma struct ocfs2_refcount_rec rec;
31886ae23c55STao Ma
3189198aac28STao Ma trace_ocfs2_make_clusters_writable(cpos, p_cluster,
3190198aac28STao Ma num_clusters, e_flags);
31916f70fa51STao Ma
31926f70fa51STao Ma ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3193913580b4STao Ma &context->data_et,
3194913580b4STao Ma ref_ci,
31956f70fa51STao Ma context->ref_root_bh,
31966f70fa51STao Ma &context->meta_ac,
31976f70fa51STao Ma &context->data_ac, &credits);
31986f70fa51STao Ma if (ret) {
31996f70fa51STao Ma mlog_errno(ret);
32006f70fa51STao Ma return ret;
32016f70fa51STao Ma }
32026f70fa51STao Ma
3203492a8a33STao Ma if (context->post_refcount)
3204492a8a33STao Ma credits += context->post_refcount->credits;
3205492a8a33STao Ma
3206492a8a33STao Ma credits += context->extra_credits;
32076f70fa51STao Ma handle = ocfs2_start_trans(osb, credits);
32086f70fa51STao Ma if (IS_ERR(handle)) {
32096f70fa51STao Ma ret = PTR_ERR(handle);
32106f70fa51STao Ma mlog_errno(ret);
32116f70fa51STao Ma goto out;
32126f70fa51STao Ma }
32136f70fa51STao Ma
3214acf3bb00STristan Ye orig_num_clusters = num_clusters;
3215acf3bb00STristan Ye
32166f70fa51STao Ma while (num_clusters) {
3217913580b4STao Ma ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
32186ae23c55STao Ma p_cluster, num_clusters,
32196ae23c55STao Ma &rec, &index, &ref_leaf_bh);
32206ae23c55STao Ma if (ret) {
32216ae23c55STao Ma mlog_errno(ret);
32226ae23c55STao Ma goto out_commit;
32236ae23c55STao Ma }
32246ae23c55STao Ma
32256ae23c55STao Ma BUG_ON(!rec.r_refcount);
32266ae23c55STao Ma set_len = min((u64)p_cluster + num_clusters,
32276ae23c55STao Ma le64_to_cpu(rec.r_cpos) +
32286ae23c55STao Ma le32_to_cpu(rec.r_clusters)) - p_cluster;
32296ae23c55STao Ma
32306ae23c55STao Ma /*
32316ae23c55STao Ma * There are many different situation here.
32326ae23c55STao Ma * 1. If refcount == 1, remove the flag and don't COW.
32336ae23c55STao Ma * 2. If refcount > 1, allocate clusters.
32346ae23c55STao Ma * Here we may not allocate r_len once at a time, so continue
32356ae23c55STao Ma * until we reach num_clusters.
32366ae23c55STao Ma */
32376ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1) {
32386ae23c55STao Ma delete = 0;
3239913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle,
3240913580b4STao Ma &context->data_et,
32416ae23c55STao Ma cpos, p_cluster,
32426ae23c55STao Ma set_len, e_flags,
32436ae23c55STao Ma context->meta_ac,
32446ae23c55STao Ma &context->dealloc);
32456ae23c55STao Ma if (ret) {
32466ae23c55STao Ma mlog_errno(ret);
32476ae23c55STao Ma goto out_commit;
32486ae23c55STao Ma }
32496ae23c55STao Ma } else {
32506ae23c55STao Ma delete = 1;
32516ae23c55STao Ma
32521ed9b777SJoel Becker ret = __ocfs2_claim_clusters(handle,
32536ae23c55STao Ma context->data_ac,
32546ae23c55STao Ma 1, set_len,
32556f70fa51STao Ma &new_bit, &new_len);
32566f70fa51STao Ma if (ret) {
32576f70fa51STao Ma mlog_errno(ret);
32586f70fa51STao Ma goto out_commit;
32596f70fa51STao Ma }
32606f70fa51STao Ma
32616f70fa51STao Ma ret = ocfs2_replace_clusters(handle, context,
32626f70fa51STao Ma cpos, p_cluster, new_bit,
32636f70fa51STao Ma new_len, e_flags);
32646f70fa51STao Ma if (ret) {
32656f70fa51STao Ma mlog_errno(ret);
32666f70fa51STao Ma goto out_commit;
32676f70fa51STao Ma }
32686ae23c55STao Ma set_len = new_len;
32696f70fa51STao Ma }
32706f70fa51STao Ma
3271913580b4STao Ma ret = __ocfs2_decrease_refcount(handle, ref_ci,
32726f70fa51STao Ma context->ref_root_bh,
32736ae23c55STao Ma p_cluster, set_len,
32746f70fa51STao Ma context->meta_ac,
32756ae23c55STao Ma &context->dealloc, delete);
32766f70fa51STao Ma if (ret) {
32776f70fa51STao Ma mlog_errno(ret);
32786f70fa51STao Ma goto out_commit;
32796f70fa51STao Ma }
32806f70fa51STao Ma
32816ae23c55STao Ma cpos += set_len;
32826ae23c55STao Ma p_cluster += set_len;
32836ae23c55STao Ma num_clusters -= set_len;
32846ae23c55STao Ma brelse(ref_leaf_bh);
32856ae23c55STao Ma ref_leaf_bh = NULL;
32866ae23c55STao Ma }
32876ae23c55STao Ma
3288492a8a33STao Ma /* handle any post_cow action. */
3289492a8a33STao Ma if (context->post_refcount && context->post_refcount->func) {
3290492a8a33STao Ma ret = context->post_refcount->func(context->inode, handle,
3291492a8a33STao Ma context->post_refcount->para);
3292492a8a33STao Ma if (ret) {
3293492a8a33STao Ma mlog_errno(ret);
3294492a8a33STao Ma goto out_commit;
3295492a8a33STao Ma }
3296492a8a33STao Ma }
3297492a8a33STao Ma
32986f70fa51STao Ma /*
32996f70fa51STao Ma * Here we should write the new page out first if we are
33006f70fa51STao Ma * in write-back mode.
33016f70fa51STao Ma */
3302492a8a33STao Ma if (context->get_clusters == ocfs2_di_get_clusters) {
33033e19a25eSTristan Ye ret = ocfs2_cow_sync_writeback(sb, context->inode, cpos,
3304acf3bb00STristan Ye orig_num_clusters);
33056f70fa51STao Ma if (ret)
33066f70fa51STao Ma mlog_errno(ret);
3307492a8a33STao Ma }
33086f70fa51STao Ma
33096f70fa51STao Ma out_commit:
33106f70fa51STao Ma ocfs2_commit_trans(osb, handle);
33116f70fa51STao Ma
33126f70fa51STao Ma out:
33136f70fa51STao Ma if (context->data_ac) {
33146f70fa51STao Ma ocfs2_free_alloc_context(context->data_ac);
33156f70fa51STao Ma context->data_ac = NULL;
33166f70fa51STao Ma }
33176f70fa51STao Ma if (context->meta_ac) {
33186f70fa51STao Ma ocfs2_free_alloc_context(context->meta_ac);
33196f70fa51STao Ma context->meta_ac = NULL;
33206f70fa51STao Ma }
33216ae23c55STao Ma brelse(ref_leaf_bh);
33226f70fa51STao Ma
33236f70fa51STao Ma return ret;
33246f70fa51STao Ma }
33256f70fa51STao Ma
ocfs2_replace_cow(struct ocfs2_cow_context * context)3326913580b4STao Ma static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
33276f70fa51STao Ma {
33286f70fa51STao Ma int ret = 0;
3329913580b4STao Ma struct inode *inode = context->inode;
3330913580b4STao Ma u32 cow_start = context->cow_start, cow_len = context->cow_len;
3331913580b4STao Ma u32 p_cluster, num_clusters;
33326f70fa51STao Ma unsigned int ext_flags;
33336f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
33346f70fa51STao Ma
33351119d3c0Spiaojun if (!ocfs2_refcount_tree(osb)) {
33367ecef14aSJoe Perches 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",
33377ecef14aSJoe Perches inode->i_ino);
33386f70fa51STao Ma }
33396f70fa51STao Ma
33406f70fa51STao Ma ocfs2_init_dealloc_ctxt(&context->dealloc);
33416f70fa51STao Ma
33426f70fa51STao Ma while (cow_len) {
3343913580b4STao Ma ret = context->get_clusters(context, cow_start, &p_cluster,
33446f70fa51STao Ma &num_clusters, &ext_flags);
33456f70fa51STao Ma if (ret) {
33466f70fa51STao Ma mlog_errno(ret);
33476f70fa51STao Ma break;
33486f70fa51STao Ma }
33496f70fa51STao Ma
33506f70fa51STao Ma BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
33516f70fa51STao Ma
33526f70fa51STao Ma if (cow_len < num_clusters)
33536f70fa51STao Ma num_clusters = cow_len;
33546f70fa51STao Ma
33556f70fa51STao Ma ret = ocfs2_make_clusters_writable(inode->i_sb, context,
33566f70fa51STao Ma cow_start, p_cluster,
33576f70fa51STao Ma num_clusters, ext_flags);
33586f70fa51STao Ma if (ret) {
33596f70fa51STao Ma mlog_errno(ret);
33606f70fa51STao Ma break;
33616f70fa51STao Ma }
33626f70fa51STao Ma
33636f70fa51STao Ma cow_len -= num_clusters;
33646f70fa51STao Ma cow_start += num_clusters;
33656f70fa51STao Ma }
33666f70fa51STao Ma
33676f70fa51STao Ma if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
33686f70fa51STao Ma ocfs2_schedule_truncate_log_flush(osb, 1);
33696f70fa51STao Ma ocfs2_run_deallocs(osb, &context->dealloc);
33706f70fa51STao Ma }
33716f70fa51STao Ma
33726f70fa51STao Ma return ret;
33736f70fa51STao Ma }
33746f70fa51STao Ma
33756f70fa51STao Ma /*
337637f8a2bfSTao Ma * Starting at cpos, try to CoW write_len clusters. Don't CoW
337737f8a2bfSTao Ma * past max_cpos. This will stop when it runs into a hole or an
337837f8a2bfSTao Ma * unrefcounted extent.
33796f70fa51STao Ma */
ocfs2_refcount_cow_hunk(struct inode * inode,struct buffer_head * di_bh,u32 cpos,u32 write_len,u32 max_cpos)33806f70fa51STao Ma static int ocfs2_refcount_cow_hunk(struct inode *inode,
33816f70fa51STao Ma struct buffer_head *di_bh,
338237f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos)
33836f70fa51STao Ma {
33846f70fa51STao Ma int ret;
33856f70fa51STao Ma u32 cow_start = 0, cow_len = 0;
33866f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
33876f70fa51STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
33886f70fa51STao Ma struct buffer_head *ref_root_bh = NULL;
33896f70fa51STao Ma struct ocfs2_refcount_tree *ref_tree;
3390913580b4STao Ma struct ocfs2_cow_context *context = NULL;
33916f70fa51STao Ma
339284e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode));
33936f70fa51STao Ma
3394913580b4STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
339537f8a2bfSTao Ma cpos, write_len, max_cpos,
33966f70fa51STao Ma &cow_start, &cow_len);
33976f70fa51STao Ma if (ret) {
33986f70fa51STao Ma mlog_errno(ret);
33996f70fa51STao Ma goto out;
34006f70fa51STao Ma }
340137f8a2bfSTao Ma
3402198aac28STao Ma trace_ocfs2_refcount_cow_hunk(OCFS2_I(inode)->ip_blkno,
3403198aac28STao Ma cpos, write_len, max_cpos,
3404198aac28STao Ma cow_start, cow_len);
34056f70fa51STao Ma
34066f70fa51STao Ma BUG_ON(cow_len == 0);
34076f70fa51STao Ma
3408913580b4STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3409913580b4STao Ma if (!context) {
3410913580b4STao Ma ret = -ENOMEM;
3411913580b4STao Ma mlog_errno(ret);
3412913580b4STao Ma goto out;
3413913580b4STao Ma }
3414913580b4STao Ma
34156f70fa51STao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
34166f70fa51STao Ma 1, &ref_tree, &ref_root_bh);
34176f70fa51STao Ma if (ret) {
34186f70fa51STao Ma mlog_errno(ret);
34196f70fa51STao Ma goto out;
34206f70fa51STao Ma }
34216f70fa51STao Ma
3422913580b4STao Ma context->inode = inode;
3423913580b4STao Ma context->cow_start = cow_start;
3424913580b4STao Ma context->cow_len = cow_len;
3425913580b4STao Ma context->ref_tree = ref_tree;
3426913580b4STao Ma context->ref_root_bh = ref_root_bh;
3427913580b4STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3428913580b4STao Ma context->get_clusters = ocfs2_di_get_clusters;
3429913580b4STao Ma
3430913580b4STao Ma ocfs2_init_dinode_extent_tree(&context->data_et,
3431913580b4STao Ma INODE_CACHE(inode), di_bh);
3432913580b4STao Ma
3433913580b4STao Ma ret = ocfs2_replace_cow(context);
34346f70fa51STao Ma if (ret)
34356f70fa51STao Ma mlog_errno(ret);
34366f70fa51STao Ma
3437913580b4STao Ma /*
3438913580b4STao Ma * truncate the extent map here since no matter whether we meet with
3439913580b4STao Ma * any error during the action, we shouldn't trust cached extent map
3440913580b4STao Ma * any more.
3441913580b4STao Ma */
3442913580b4STao Ma ocfs2_extent_map_trunc(inode, cow_start);
3443913580b4STao Ma
34446f70fa51STao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
34456f70fa51STao Ma brelse(ref_root_bh);
34466f70fa51STao Ma out:
3447913580b4STao Ma kfree(context);
34486f70fa51STao Ma return ret;
34496f70fa51STao Ma }
34506f70fa51STao Ma
34516f70fa51STao Ma /*
34526f70fa51STao Ma * CoW any and all clusters between cpos and cpos+write_len.
345337f8a2bfSTao Ma * Don't CoW past max_cpos. If this returns successfully, all
345437f8a2bfSTao Ma * clusters between cpos and cpos+write_len are safe to modify.
34556f70fa51STao Ma */
ocfs2_refcount_cow(struct inode * inode,struct buffer_head * di_bh,u32 cpos,u32 write_len,u32 max_cpos)34566f70fa51STao Ma int ocfs2_refcount_cow(struct inode *inode,
34576f70fa51STao Ma struct buffer_head *di_bh,
345837f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos)
34596f70fa51STao Ma {
34606f70fa51STao Ma int ret = 0;
34616f70fa51STao Ma u32 p_cluster, num_clusters;
34626f70fa51STao Ma unsigned int ext_flags;
34636f70fa51STao Ma
34646f70fa51STao Ma while (write_len) {
34656f70fa51STao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
34666f70fa51STao Ma &num_clusters, &ext_flags);
34676f70fa51STao Ma if (ret) {
34686f70fa51STao Ma mlog_errno(ret);
34696f70fa51STao Ma break;
34706f70fa51STao Ma }
34716f70fa51STao Ma
34726f70fa51STao Ma if (write_len < num_clusters)
34736f70fa51STao Ma num_clusters = write_len;
34746f70fa51STao Ma
34756f70fa51STao Ma if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3476c7dd3392STiger Yang ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
347737f8a2bfSTao Ma num_clusters, max_cpos);
34786f70fa51STao Ma if (ret) {
34796f70fa51STao Ma mlog_errno(ret);
34806f70fa51STao Ma break;
34816f70fa51STao Ma }
34826f70fa51STao Ma }
34836f70fa51STao Ma
34846f70fa51STao Ma write_len -= num_clusters;
34856f70fa51STao Ma cpos += num_clusters;
34866f70fa51STao Ma }
34876f70fa51STao Ma
34886f70fa51STao Ma return ret;
34896f70fa51STao Ma }
3490110a045aSTao Ma
ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context * context,u32 v_cluster,u32 * p_cluster,u32 * num_clusters,unsigned int * extent_flags)3491492a8a33STao Ma static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3492492a8a33STao Ma u32 v_cluster, u32 *p_cluster,
3493492a8a33STao Ma u32 *num_clusters,
3494492a8a33STao Ma unsigned int *extent_flags)
3495492a8a33STao Ma {
3496492a8a33STao Ma struct inode *inode = context->inode;
3497492a8a33STao Ma struct ocfs2_xattr_value_root *xv = context->cow_object;
3498492a8a33STao Ma
3499492a8a33STao Ma return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3500492a8a33STao Ma num_clusters, &xv->xr_list,
3501492a8a33STao Ma extent_flags);
3502492a8a33STao Ma }
3503492a8a33STao Ma
3504492a8a33STao Ma /*
3505492a8a33STao Ma * Given a xattr value root, calculate the most meta/credits we need for
3506492a8a33STao Ma * refcount tree change if we truncate it to 0.
3507492a8a33STao Ma */
ocfs2_refcounted_xattr_delete_need(struct inode * inode,struct ocfs2_caching_info * ref_ci,struct buffer_head * ref_root_bh,struct ocfs2_xattr_value_root * xv,int * meta_add,int * credits)3508492a8a33STao Ma int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3509492a8a33STao Ma struct ocfs2_caching_info *ref_ci,
3510492a8a33STao Ma struct buffer_head *ref_root_bh,
3511492a8a33STao Ma struct ocfs2_xattr_value_root *xv,
3512492a8a33STao Ma int *meta_add, int *credits)
3513492a8a33STao Ma {
3514492a8a33STao Ma int ret = 0, index, ref_blocks = 0;
3515492a8a33STao Ma u32 p_cluster, num_clusters;
3516492a8a33STao Ma u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3517492a8a33STao Ma struct ocfs2_refcount_block *rb;
3518492a8a33STao Ma struct ocfs2_refcount_rec rec;
3519492a8a33STao Ma struct buffer_head *ref_leaf_bh = NULL;
3520492a8a33STao Ma
3521492a8a33STao Ma while (cpos < clusters) {
3522492a8a33STao Ma ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3523492a8a33STao Ma &num_clusters, &xv->xr_list,
3524492a8a33STao Ma NULL);
3525492a8a33STao Ma if (ret) {
3526492a8a33STao Ma mlog_errno(ret);
3527492a8a33STao Ma goto out;
3528492a8a33STao Ma }
3529492a8a33STao Ma
3530492a8a33STao Ma cpos += num_clusters;
3531492a8a33STao Ma
3532492a8a33STao Ma while (num_clusters) {
3533492a8a33STao Ma ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3534492a8a33STao Ma p_cluster, num_clusters,
3535492a8a33STao Ma &rec, &index,
3536492a8a33STao Ma &ref_leaf_bh);
3537492a8a33STao Ma if (ret) {
3538492a8a33STao Ma mlog_errno(ret);
3539492a8a33STao Ma goto out;
3540492a8a33STao Ma }
3541492a8a33STao Ma
3542492a8a33STao Ma BUG_ON(!rec.r_refcount);
3543492a8a33STao Ma
3544492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3545492a8a33STao Ma
3546492a8a33STao Ma /*
3547492a8a33STao Ma * We really don't know whether the other clusters is in
3548492a8a33STao Ma * this refcount block or not, so just take the worst
3549492a8a33STao Ma * case that all the clusters are in this block and each
3550492a8a33STao Ma * one will split a refcount rec, so totally we need
3551492a8a33STao Ma * clusters * 2 new refcount rec.
3552492a8a33STao Ma */
3553e1bf4cc6SAl Viro if (le16_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3554492a8a33STao Ma le16_to_cpu(rb->rf_records.rl_count))
3555492a8a33STao Ma ref_blocks++;
3556492a8a33STao Ma
3557492a8a33STao Ma *credits += 1;
3558492a8a33STao Ma brelse(ref_leaf_bh);
3559492a8a33STao Ma ref_leaf_bh = NULL;
3560492a8a33STao Ma
3561492a8a33STao Ma if (num_clusters <= le32_to_cpu(rec.r_clusters))
3562492a8a33STao Ma break;
3563492a8a33STao Ma else
3564492a8a33STao Ma num_clusters -= le32_to_cpu(rec.r_clusters);
3565492a8a33STao Ma p_cluster += num_clusters;
3566492a8a33STao Ma }
3567492a8a33STao Ma }
3568492a8a33STao Ma
3569492a8a33STao Ma *meta_add += ref_blocks;
3570492a8a33STao Ma if (!ref_blocks)
3571492a8a33STao Ma goto out;
3572492a8a33STao Ma
3573492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3574492a8a33STao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3575492a8a33STao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3576492a8a33STao Ma else {
3577492a8a33STao Ma struct ocfs2_extent_tree et;
3578492a8a33STao Ma
3579492a8a33STao Ma ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3580492a8a33STao Ma *credits += ocfs2_calc_extend_credits(inode->i_sb,
358106f9da6eSGoldwyn Rodrigues et.et_root_el);
3582492a8a33STao Ma }
3583492a8a33STao Ma
3584492a8a33STao Ma out:
3585492a8a33STao Ma brelse(ref_leaf_bh);
3586492a8a33STao Ma return ret;
3587492a8a33STao Ma }
3588492a8a33STao Ma
3589492a8a33STao Ma /*
3590492a8a33STao Ma * Do CoW for xattr.
3591492a8a33STao Ma */
ocfs2_refcount_cow_xattr(struct inode * inode,struct ocfs2_dinode * di,struct ocfs2_xattr_value_buf * vb,struct ocfs2_refcount_tree * ref_tree,struct buffer_head * ref_root_bh,u32 cpos,u32 write_len,struct ocfs2_post_refcount * post)3592492a8a33STao Ma int ocfs2_refcount_cow_xattr(struct inode *inode,
3593492a8a33STao Ma struct ocfs2_dinode *di,
3594492a8a33STao Ma struct ocfs2_xattr_value_buf *vb,
3595492a8a33STao Ma struct ocfs2_refcount_tree *ref_tree,
3596492a8a33STao Ma struct buffer_head *ref_root_bh,
3597492a8a33STao Ma u32 cpos, u32 write_len,
3598492a8a33STao Ma struct ocfs2_post_refcount *post)
3599492a8a33STao Ma {
3600492a8a33STao Ma int ret;
3601492a8a33STao Ma struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3602492a8a33STao Ma struct ocfs2_cow_context *context = NULL;
3603492a8a33STao Ma u32 cow_start, cow_len;
3604492a8a33STao Ma
360584e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode));
3606492a8a33STao Ma
3607492a8a33STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3608492a8a33STao Ma cpos, write_len, UINT_MAX,
3609492a8a33STao Ma &cow_start, &cow_len);
3610492a8a33STao Ma if (ret) {
3611492a8a33STao Ma mlog_errno(ret);
3612492a8a33STao Ma goto out;
3613492a8a33STao Ma }
3614492a8a33STao Ma
3615492a8a33STao Ma BUG_ON(cow_len == 0);
3616492a8a33STao Ma
3617492a8a33STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3618492a8a33STao Ma if (!context) {
3619492a8a33STao Ma ret = -ENOMEM;
3620492a8a33STao Ma mlog_errno(ret);
3621492a8a33STao Ma goto out;
3622492a8a33STao Ma }
3623492a8a33STao Ma
3624492a8a33STao Ma context->inode = inode;
3625492a8a33STao Ma context->cow_start = cow_start;
3626492a8a33STao Ma context->cow_len = cow_len;
3627492a8a33STao Ma context->ref_tree = ref_tree;
36286eab04a8SJustin P. Mattock context->ref_root_bh = ref_root_bh;
3629492a8a33STao Ma context->cow_object = xv;
3630492a8a33STao Ma
3631492a8a33STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3632492a8a33STao Ma /* We need the extra credits for duplicate_clusters by jbd. */
3633492a8a33STao Ma context->extra_credits =
3634492a8a33STao Ma ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3635492a8a33STao Ma context->get_clusters = ocfs2_xattr_value_get_clusters;
3636492a8a33STao Ma context->post_refcount = post;
3637492a8a33STao Ma
3638492a8a33STao Ma ocfs2_init_xattr_value_extent_tree(&context->data_et,
3639492a8a33STao Ma INODE_CACHE(inode), vb);
3640492a8a33STao Ma
3641492a8a33STao Ma ret = ocfs2_replace_cow(context);
3642492a8a33STao Ma if (ret)
3643492a8a33STao Ma mlog_errno(ret);
3644492a8a33STao Ma
3645492a8a33STao Ma out:
3646492a8a33STao Ma kfree(context);
3647492a8a33STao Ma return ret;
3648492a8a33STao Ma }
3649492a8a33STao Ma
3650110a045aSTao Ma /*
3651110a045aSTao Ma * Insert a new extent into refcount tree and mark a extent rec
3652110a045aSTao Ma * as refcounted in the dinode tree.
3653110a045aSTao Ma */
ocfs2_add_refcount_flag(struct inode * inode,struct ocfs2_extent_tree * data_et,struct ocfs2_caching_info * ref_ci,struct buffer_head * ref_root_bh,u32 cpos,u32 p_cluster,u32 num_clusters,struct ocfs2_cached_dealloc_ctxt * dealloc,struct ocfs2_post_refcount * post)3654110a045aSTao Ma int ocfs2_add_refcount_flag(struct inode *inode,
3655110a045aSTao Ma struct ocfs2_extent_tree *data_et,
3656110a045aSTao Ma struct ocfs2_caching_info *ref_ci,
3657110a045aSTao Ma struct buffer_head *ref_root_bh,
3658110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters,
36590129241eSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc,
36600129241eSTao Ma struct ocfs2_post_refcount *post)
3661110a045aSTao Ma {
3662110a045aSTao Ma int ret;
3663110a045aSTao Ma handle_t *handle;
3664110a045aSTao Ma int credits = 1, ref_blocks = 0;
3665110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3666110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL;
3667110a045aSTao Ma
36683e10b793SDarrick J. Wong /* We need to be able to handle at least an extent tree split. */
36693e10b793SDarrick J. Wong ref_blocks = ocfs2_extend_meta_needed(data_et->et_root_el);
36703e10b793SDarrick J. Wong
3671110a045aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3672110a045aSTao Ma ref_ci, ref_root_bh,
3673110a045aSTao Ma p_cluster, num_clusters,
3674110a045aSTao Ma &ref_blocks, &credits);
3675110a045aSTao Ma if (ret) {
3676110a045aSTao Ma mlog_errno(ret);
3677110a045aSTao Ma goto out;
3678110a045aSTao Ma }
3679110a045aSTao Ma
3680198aac28STao Ma trace_ocfs2_add_refcount_flag(ref_blocks, credits);
3681110a045aSTao Ma
3682110a045aSTao Ma if (ref_blocks) {
36831119d3c0Spiaojun ret = ocfs2_reserve_new_metadata_blocks(osb,
3684110a045aSTao Ma ref_blocks, &meta_ac);
3685110a045aSTao Ma if (ret) {
3686110a045aSTao Ma mlog_errno(ret);
3687110a045aSTao Ma goto out;
3688110a045aSTao Ma }
3689110a045aSTao Ma }
3690110a045aSTao Ma
36910129241eSTao Ma if (post)
36920129241eSTao Ma credits += post->credits;
36930129241eSTao Ma
3694110a045aSTao Ma handle = ocfs2_start_trans(osb, credits);
3695110a045aSTao Ma if (IS_ERR(handle)) {
3696110a045aSTao Ma ret = PTR_ERR(handle);
3697110a045aSTao Ma mlog_errno(ret);
3698110a045aSTao Ma goto out;
3699110a045aSTao Ma }
3700110a045aSTao Ma
3701110a045aSTao Ma ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3702110a045aSTao Ma cpos, num_clusters, p_cluster,
3703110a045aSTao Ma meta_ac, dealloc);
3704110a045aSTao Ma if (ret) {
3705110a045aSTao Ma mlog_errno(ret);
3706110a045aSTao Ma goto out_commit;
3707110a045aSTao Ma }
3708110a045aSTao Ma
37097540c1a7STao Ma ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
37107540c1a7STao Ma p_cluster, num_clusters, 0,
3711110a045aSTao Ma meta_ac, dealloc);
37120129241eSTao Ma if (ret) {
37130129241eSTao Ma mlog_errno(ret);
37140129241eSTao Ma goto out_commit;
37150129241eSTao Ma }
37160129241eSTao Ma
37170129241eSTao Ma if (post && post->func) {
37180129241eSTao Ma ret = post->func(inode, handle, post->para);
3719110a045aSTao Ma if (ret)
3720110a045aSTao Ma mlog_errno(ret);
37210129241eSTao Ma }
3722110a045aSTao Ma
3723110a045aSTao Ma out_commit:
3724110a045aSTao Ma ocfs2_commit_trans(osb, handle);
3725110a045aSTao Ma out:
3726110a045aSTao Ma if (meta_ac)
3727110a045aSTao Ma ocfs2_free_alloc_context(meta_ac);
3728110a045aSTao Ma return ret;
3729110a045aSTao Ma }
3730110a045aSTao Ma
ocfs2_change_ctime(struct inode * inode,struct buffer_head * di_bh)3731a9063ab9STao Ma static int ocfs2_change_ctime(struct inode *inode,
3732110a045aSTao Ma struct buffer_head *di_bh)
3733110a045aSTao Ma {
3734110a045aSTao Ma int ret;
3735a9063ab9STao Ma handle_t *handle;
3736a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3737a9063ab9STao Ma
3738a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3739a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS);
3740a9063ab9STao Ma if (IS_ERR(handle)) {
3741a9063ab9STao Ma ret = PTR_ERR(handle);
3742a9063ab9STao Ma mlog_errno(ret);
3743a9063ab9STao Ma goto out;
3744a9063ab9STao Ma }
3745a9063ab9STao Ma
3746a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3747a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE);
3748a9063ab9STao Ma if (ret) {
3749a9063ab9STao Ma mlog_errno(ret);
3750a9063ab9STao Ma goto out_commit;
3751a9063ab9STao Ma }
3752a9063ab9STao Ma
37536861de97SJeff Layton inode_set_ctime_current(inode);
3754*10fc3a18SJeff Layton di->i_ctime = cpu_to_le64(inode_get_ctime_sec(inode));
3755*10fc3a18SJeff Layton di->i_ctime_nsec = cpu_to_le32(inode_get_ctime_nsec(inode));
3756a9063ab9STao Ma
3757a9063ab9STao Ma ocfs2_journal_dirty(handle, di_bh);
3758a9063ab9STao Ma
3759a9063ab9STao Ma out_commit:
3760a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3761a9063ab9STao Ma out:
3762a9063ab9STao Ma return ret;
3763a9063ab9STao Ma }
3764a9063ab9STao Ma
ocfs2_attach_refcount_tree(struct inode * inode,struct buffer_head * di_bh)3765a9063ab9STao Ma static int ocfs2_attach_refcount_tree(struct inode *inode,
3766a9063ab9STao Ma struct buffer_head *di_bh)
3767a9063ab9STao Ma {
3768a9063ab9STao Ma int ret, data_changed = 0;
3769110a045aSTao Ma struct buffer_head *ref_root_bh = NULL;
3770110a045aSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode);
3771110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3772110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3773110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree;
3774110a045aSTao Ma unsigned int ext_flags;
3775110a045aSTao Ma loff_t size;
3776110a045aSTao Ma u32 cpos, num_clusters, clusters, p_cluster;
3777110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc;
3778110a045aSTao Ma struct ocfs2_extent_tree di_et;
3779110a045aSTao Ma
3780110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc);
3781110a045aSTao Ma
378284e40080SDarrick J. Wong if (!ocfs2_is_refcount_inode(inode)) {
3783110a045aSTao Ma ret = ocfs2_create_refcount_tree(inode, di_bh);
3784110a045aSTao Ma if (ret) {
3785110a045aSTao Ma mlog_errno(ret);
3786110a045aSTao Ma goto out;
3787110a045aSTao Ma }
3788110a045aSTao Ma }
3789110a045aSTao Ma
3790110a045aSTao Ma BUG_ON(!di->i_refcount_loc);
3791110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb,
3792110a045aSTao Ma le64_to_cpu(di->i_refcount_loc), 1,
3793110a045aSTao Ma &ref_tree, &ref_root_bh);
3794110a045aSTao Ma if (ret) {
3795110a045aSTao Ma mlog_errno(ret);
3796110a045aSTao Ma goto out;
3797110a045aSTao Ma }
3798110a045aSTao Ma
37992f48d593STao Ma if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
38002f48d593STao Ma goto attach_xattr;
38012f48d593STao Ma
3802110a045aSTao Ma ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3803110a045aSTao Ma
3804110a045aSTao Ma size = i_size_read(inode);
3805110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3806110a045aSTao Ma
3807110a045aSTao Ma cpos = 0;
3808110a045aSTao Ma while (cpos < clusters) {
3809110a045aSTao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3810110a045aSTao Ma &num_clusters, &ext_flags);
38112b0f6eaeSJoseph Qi if (ret) {
38122b0f6eaeSJoseph Qi mlog_errno(ret);
38132b0f6eaeSJoseph Qi goto unlock;
38142b0f6eaeSJoseph Qi }
3815110a045aSTao Ma if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3816110a045aSTao Ma ret = ocfs2_add_refcount_flag(inode, &di_et,
3817110a045aSTao Ma &ref_tree->rf_ci,
3818110a045aSTao Ma ref_root_bh, cpos,
3819110a045aSTao Ma p_cluster, num_clusters,
38200129241eSTao Ma &dealloc, NULL);
3821110a045aSTao Ma if (ret) {
3822110a045aSTao Ma mlog_errno(ret);
3823a9063ab9STao Ma goto unlock;
3824110a045aSTao Ma }
3825a9063ab9STao Ma
3826a9063ab9STao Ma data_changed = 1;
3827110a045aSTao Ma }
3828110a045aSTao Ma cpos += num_clusters;
3829110a045aSTao Ma }
3830110a045aSTao Ma
38312f48d593STao Ma attach_xattr:
38320129241eSTao Ma if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
38330129241eSTao Ma ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
38340129241eSTao Ma &ref_tree->rf_ci,
38350129241eSTao Ma ref_root_bh,
38360129241eSTao Ma &dealloc);
38370129241eSTao Ma if (ret) {
38380129241eSTao Ma mlog_errno(ret);
38390129241eSTao Ma goto unlock;
38400129241eSTao Ma }
38410129241eSTao Ma }
38420129241eSTao Ma
3843a9063ab9STao Ma if (data_changed) {
3844a9063ab9STao Ma ret = ocfs2_change_ctime(inode, di_bh);
3845a9063ab9STao Ma if (ret)
3846a9063ab9STao Ma mlog_errno(ret);
3847a9063ab9STao Ma }
3848a9063ab9STao Ma
3849a9063ab9STao Ma unlock:
3850110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3851110a045aSTao Ma brelse(ref_root_bh);
3852110a045aSTao Ma
3853110a045aSTao Ma if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3854110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1);
3855110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc);
3856110a045aSTao Ma }
3857110a045aSTao Ma out:
3858110a045aSTao Ma /*
3859110a045aSTao Ma * Empty the extent map so that we may get the right extent
3860110a045aSTao Ma * record from the disk.
3861110a045aSTao Ma */
3862110a045aSTao Ma ocfs2_extent_map_trunc(inode, 0);
3863110a045aSTao Ma
3864110a045aSTao Ma return ret;
3865110a045aSTao Ma }
3866110a045aSTao Ma
ocfs2_add_refcounted_extent(struct inode * inode,struct ocfs2_extent_tree * et,struct ocfs2_caching_info * ref_ci,struct buffer_head * ref_root_bh,u32 cpos,u32 p_cluster,u32 num_clusters,unsigned int ext_flags,struct ocfs2_cached_dealloc_ctxt * dealloc)3867110a045aSTao Ma static int ocfs2_add_refcounted_extent(struct inode *inode,
3868110a045aSTao Ma struct ocfs2_extent_tree *et,
3869110a045aSTao Ma struct ocfs2_caching_info *ref_ci,
3870110a045aSTao Ma struct buffer_head *ref_root_bh,
3871110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters,
3872110a045aSTao Ma unsigned int ext_flags,
3873110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
3874110a045aSTao Ma {
3875110a045aSTao Ma int ret;
3876110a045aSTao Ma handle_t *handle;
3877110a045aSTao Ma int credits = 0;
3878110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3879110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL;
3880110a045aSTao Ma
3881110a045aSTao Ma ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3882110a045aSTao Ma p_cluster, num_clusters,
3883110a045aSTao Ma et, ref_ci,
3884110a045aSTao Ma ref_root_bh, &meta_ac,
3885110a045aSTao Ma NULL, &credits);
3886110a045aSTao Ma if (ret) {
3887110a045aSTao Ma mlog_errno(ret);
3888110a045aSTao Ma goto out;
3889110a045aSTao Ma }
3890110a045aSTao Ma
3891110a045aSTao Ma handle = ocfs2_start_trans(osb, credits);
3892110a045aSTao Ma if (IS_ERR(handle)) {
3893110a045aSTao Ma ret = PTR_ERR(handle);
3894110a045aSTao Ma mlog_errno(ret);
3895110a045aSTao Ma goto out;
3896110a045aSTao Ma }
3897110a045aSTao Ma
3898110a045aSTao Ma ret = ocfs2_insert_extent(handle, et, cpos,
389912d4cec9STao Ma ocfs2_clusters_to_blocks(inode->i_sb, p_cluster),
3900110a045aSTao Ma num_clusters, ext_flags, meta_ac);
3901110a045aSTao Ma if (ret) {
3902110a045aSTao Ma mlog_errno(ret);
3903110a045aSTao Ma goto out_commit;
3904110a045aSTao Ma }
3905110a045aSTao Ma
39062999d12fSTao Ma ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3907110a045aSTao Ma p_cluster, num_clusters,
3908110a045aSTao Ma meta_ac, dealloc);
390986e59436SDarrick J. Wong if (ret) {
391086e59436SDarrick J. Wong mlog_errno(ret);
391186e59436SDarrick J. Wong goto out_commit;
391286e59436SDarrick J. Wong }
391386e59436SDarrick J. Wong
391486e59436SDarrick J. Wong ret = dquot_alloc_space_nodirty(inode,
391586e59436SDarrick J. Wong ocfs2_clusters_to_bytes(osb->sb, num_clusters));
3916110a045aSTao Ma if (ret)
3917110a045aSTao Ma mlog_errno(ret);
3918110a045aSTao Ma
3919110a045aSTao Ma out_commit:
3920110a045aSTao Ma ocfs2_commit_trans(osb, handle);
3921110a045aSTao Ma out:
3922110a045aSTao Ma if (meta_ac)
3923110a045aSTao Ma ocfs2_free_alloc_context(meta_ac);
3924110a045aSTao Ma return ret;
3925110a045aSTao Ma }
3926110a045aSTao Ma
ocfs2_duplicate_inline_data(struct inode * s_inode,struct buffer_head * s_bh,struct inode * t_inode,struct buffer_head * t_bh)39272f48d593STao Ma static int ocfs2_duplicate_inline_data(struct inode *s_inode,
39282f48d593STao Ma struct buffer_head *s_bh,
39292f48d593STao Ma struct inode *t_inode,
39302f48d593STao Ma struct buffer_head *t_bh)
39312f48d593STao Ma {
39322f48d593STao Ma int ret;
39332f48d593STao Ma handle_t *handle;
39342f48d593STao Ma struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
39352f48d593STao Ma struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
39362f48d593STao Ma struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data;
39372f48d593STao Ma
39382f48d593STao Ma BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
39392f48d593STao Ma
39402f48d593STao Ma handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
39412f48d593STao Ma if (IS_ERR(handle)) {
39422f48d593STao Ma ret = PTR_ERR(handle);
39432f48d593STao Ma mlog_errno(ret);
39442f48d593STao Ma goto out;
39452f48d593STao Ma }
39462f48d593STao Ma
39472f48d593STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
39482f48d593STao Ma OCFS2_JOURNAL_ACCESS_WRITE);
39492f48d593STao Ma if (ret) {
39502f48d593STao Ma mlog_errno(ret);
39512f48d593STao Ma goto out_commit;
39522f48d593STao Ma }
39532f48d593STao Ma
39542f48d593STao Ma t_di->id2.i_data.id_count = s_di->id2.i_data.id_count;
39552f48d593STao Ma memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data,
39562f48d593STao Ma le16_to_cpu(s_di->id2.i_data.id_count));
39572f48d593STao Ma spin_lock(&OCFS2_I(t_inode)->ip_lock);
39582f48d593STao Ma OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
39592f48d593STao Ma t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features);
39602f48d593STao Ma spin_unlock(&OCFS2_I(t_inode)->ip_lock);
39612f48d593STao Ma
39622f48d593STao Ma ocfs2_journal_dirty(handle, t_bh);
39632f48d593STao Ma
39642f48d593STao Ma out_commit:
39652f48d593STao Ma ocfs2_commit_trans(osb, handle);
39662f48d593STao Ma out:
39672f48d593STao Ma return ret;
39682f48d593STao Ma }
39692f48d593STao Ma
ocfs2_duplicate_extent_list(struct inode * s_inode,struct inode * t_inode,struct buffer_head * t_bh,struct ocfs2_caching_info * ref_ci,struct buffer_head * ref_root_bh,struct ocfs2_cached_dealloc_ctxt * dealloc)3970110a045aSTao Ma static int ocfs2_duplicate_extent_list(struct inode *s_inode,
3971110a045aSTao Ma struct inode *t_inode,
3972110a045aSTao Ma struct buffer_head *t_bh,
3973110a045aSTao Ma struct ocfs2_caching_info *ref_ci,
3974110a045aSTao Ma struct buffer_head *ref_root_bh,
3975110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc)
3976110a045aSTao Ma {
3977110a045aSTao Ma int ret = 0;
3978110a045aSTao Ma u32 p_cluster, num_clusters, clusters, cpos;
3979110a045aSTao Ma loff_t size;
3980110a045aSTao Ma unsigned int ext_flags;
3981110a045aSTao Ma struct ocfs2_extent_tree et;
3982110a045aSTao Ma
3983110a045aSTao Ma ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
3984110a045aSTao Ma
3985110a045aSTao Ma size = i_size_read(s_inode);
3986110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
3987110a045aSTao Ma
3988110a045aSTao Ma cpos = 0;
3989110a045aSTao Ma while (cpos < clusters) {
3990110a045aSTao Ma ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
3991110a045aSTao Ma &num_clusters, &ext_flags);
39922b0f6eaeSJoseph Qi if (ret) {
39932b0f6eaeSJoseph Qi mlog_errno(ret);
39942b0f6eaeSJoseph Qi goto out;
39952b0f6eaeSJoseph Qi }
3996110a045aSTao Ma if (p_cluster) {
3997110a045aSTao Ma ret = ocfs2_add_refcounted_extent(t_inode, &et,
3998110a045aSTao Ma ref_ci, ref_root_bh,
3999110a045aSTao Ma cpos, p_cluster,
4000110a045aSTao Ma num_clusters,
4001110a045aSTao Ma ext_flags,
4002110a045aSTao Ma dealloc);
4003110a045aSTao Ma if (ret) {
4004110a045aSTao Ma mlog_errno(ret);
4005110a045aSTao Ma goto out;
4006110a045aSTao Ma }
4007110a045aSTao Ma }
4008110a045aSTao Ma
4009110a045aSTao Ma cpos += num_clusters;
4010110a045aSTao Ma }
4011110a045aSTao Ma
4012110a045aSTao Ma out:
4013110a045aSTao Ma return ret;
4014110a045aSTao Ma }
4015110a045aSTao Ma
4016a9063ab9STao Ma /*
4017a9063ab9STao Ma * change the new file's attributes to the src.
4018a9063ab9STao Ma *
4019a9063ab9STao Ma * reflink creates a snapshot of a file, that means the attributes
4020a9063ab9STao Ma * must be identical except for three exceptions - nlink, ino, and ctime.
4021a9063ab9STao Ma */
ocfs2_complete_reflink(struct inode * s_inode,struct buffer_head * s_bh,struct inode * t_inode,struct buffer_head * t_bh,bool preserve)4022a9063ab9STao Ma static int ocfs2_complete_reflink(struct inode *s_inode,
4023a9063ab9STao Ma struct buffer_head *s_bh,
4024a9063ab9STao Ma struct inode *t_inode,
40250fe9b66cSTao Ma struct buffer_head *t_bh,
40260fe9b66cSTao Ma bool preserve)
4027a9063ab9STao Ma {
4028a9063ab9STao Ma int ret;
4029a9063ab9STao Ma handle_t *handle;
4030a9063ab9STao Ma struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
4031a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
4032a9063ab9STao Ma loff_t size = i_size_read(s_inode);
4033a9063ab9STao Ma
4034a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
4035a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS);
4036a9063ab9STao Ma if (IS_ERR(handle)) {
4037a9063ab9STao Ma ret = PTR_ERR(handle);
4038a9063ab9STao Ma mlog_errno(ret);
4039a9063ab9STao Ma return ret;
4040a9063ab9STao Ma }
4041a9063ab9STao Ma
4042a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4043a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE);
4044a9063ab9STao Ma if (ret) {
4045a9063ab9STao Ma mlog_errno(ret);
4046a9063ab9STao Ma goto out_commit;
4047a9063ab9STao Ma }
4048a9063ab9STao Ma
4049a9063ab9STao Ma spin_lock(&OCFS2_I(t_inode)->ip_lock);
4050a9063ab9STao Ma OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
4051a9063ab9STao Ma OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
4052a9063ab9STao Ma OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
4053a9063ab9STao Ma spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4054a9063ab9STao Ma i_size_write(t_inode, size);
40556527f8f8STao Ma t_inode->i_blocks = s_inode->i_blocks;
4056a9063ab9STao Ma
4057a9063ab9STao Ma di->i_xattr_inline_size = s_di->i_xattr_inline_size;
4058a9063ab9STao Ma di->i_clusters = s_di->i_clusters;
4059a9063ab9STao Ma di->i_size = s_di->i_size;
4060a9063ab9STao Ma di->i_dyn_features = s_di->i_dyn_features;
4061a9063ab9STao Ma di->i_attr = s_di->i_attr;
40620fe9b66cSTao Ma
40630fe9b66cSTao Ma if (preserve) {
4064c21a534eSTao Ma t_inode->i_uid = s_inode->i_uid;
4065c21a534eSTao Ma t_inode->i_gid = s_inode->i_gid;
4066c21a534eSTao Ma t_inode->i_mode = s_inode->i_mode;
4067a9063ab9STao Ma di->i_uid = s_di->i_uid;
4068a9063ab9STao Ma di->i_gid = s_di->i_gid;
4069a9063ab9STao Ma di->i_mode = s_di->i_mode;
4070a9063ab9STao Ma
4071a9063ab9STao Ma /*
4072a9063ab9STao Ma * update time.
40730fe9b66cSTao Ma * we want mtime to appear identical to the source and
40740fe9b66cSTao Ma * update ctime.
4075a9063ab9STao Ma */
40766861de97SJeff Layton inode_set_ctime_current(t_inode);
4077a9063ab9STao Ma
4078*10fc3a18SJeff Layton di->i_ctime = cpu_to_le64(inode_get_ctime_sec(t_inode));
4079*10fc3a18SJeff Layton di->i_ctime_nsec = cpu_to_le32(inode_get_ctime_nsec(t_inode));
4080a9063ab9STao Ma
4081*10fc3a18SJeff Layton inode_set_mtime_to_ts(t_inode, inode_get_mtime(s_inode));
4082a9063ab9STao Ma di->i_mtime = s_di->i_mtime;
4083a9063ab9STao Ma di->i_mtime_nsec = s_di->i_mtime_nsec;
40840fe9b66cSTao Ma }
4085a9063ab9STao Ma
4086a9063ab9STao Ma ocfs2_journal_dirty(handle, t_bh);
4087a9063ab9STao Ma
4088a9063ab9STao Ma out_commit:
4089a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
4090a9063ab9STao Ma return ret;
4091a9063ab9STao Ma }
4092a9063ab9STao Ma
ocfs2_create_reflink_node(struct inode * s_inode,struct buffer_head * s_bh,struct inode * t_inode,struct buffer_head * t_bh,bool preserve)4093110a045aSTao Ma static int ocfs2_create_reflink_node(struct inode *s_inode,
4094110a045aSTao Ma struct buffer_head *s_bh,
4095110a045aSTao Ma struct inode *t_inode,
40960fe9b66cSTao Ma struct buffer_head *t_bh,
40970fe9b66cSTao Ma bool preserve)
4098110a045aSTao Ma {
4099110a045aSTao Ma int ret;
4100110a045aSTao Ma struct buffer_head *ref_root_bh = NULL;
4101110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc;
4102110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
4103110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
4104110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree;
4105110a045aSTao Ma
4106110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc);
4107110a045aSTao Ma
4108110a045aSTao Ma ret = ocfs2_set_refcount_tree(t_inode, t_bh,
4109110a045aSTao Ma le64_to_cpu(di->i_refcount_loc));
4110110a045aSTao Ma if (ret) {
4111110a045aSTao Ma mlog_errno(ret);
4112110a045aSTao Ma goto out;
4113110a045aSTao Ma }
4114110a045aSTao Ma
41152f48d593STao Ma if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
41162f48d593STao Ma ret = ocfs2_duplicate_inline_data(s_inode, s_bh,
41172f48d593STao Ma t_inode, t_bh);
41182f48d593STao Ma if (ret)
41192f48d593STao Ma mlog_errno(ret);
41202f48d593STao Ma goto out;
41212f48d593STao Ma }
41222f48d593STao Ma
4123110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
4124110a045aSTao Ma 1, &ref_tree, &ref_root_bh);
4125110a045aSTao Ma if (ret) {
4126110a045aSTao Ma mlog_errno(ret);
4127110a045aSTao Ma goto out;
4128110a045aSTao Ma }
4129110a045aSTao Ma
4130110a045aSTao Ma ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4131110a045aSTao Ma &ref_tree->rf_ci, ref_root_bh,
4132110a045aSTao Ma &dealloc);
4133a9063ab9STao Ma if (ret) {
4134a9063ab9STao Ma mlog_errno(ret);
4135a9063ab9STao Ma goto out_unlock_refcount;
4136a9063ab9STao Ma }
4137a9063ab9STao Ma
4138a9063ab9STao Ma out_unlock_refcount:
4139110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4140110a045aSTao Ma brelse(ref_root_bh);
4141110a045aSTao Ma out:
4142110a045aSTao Ma if (ocfs2_dealloc_has_cluster(&dealloc)) {
4143110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1);
4144110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc);
4145110a045aSTao Ma }
4146110a045aSTao Ma
4147110a045aSTao Ma return ret;
4148110a045aSTao Ma }
414909bf27a0STao Ma
__ocfs2_reflink(struct dentry * old_dentry,struct buffer_head * old_bh,struct inode * new_inode,bool preserve)415009bf27a0STao Ma static int __ocfs2_reflink(struct dentry *old_dentry,
415109bf27a0STao Ma struct buffer_head *old_bh,
415209bf27a0STao Ma struct inode *new_inode,
415309bf27a0STao Ma bool preserve)
415409bf27a0STao Ma {
415509bf27a0STao Ma int ret;
41562b0143b5SDavid Howells struct inode *inode = d_inode(old_dentry);
415709bf27a0STao Ma struct buffer_head *new_bh = NULL;
415809bf27a0STao Ma
41595693486bSJoel Becker if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) {
41605693486bSJoel Becker ret = -EINVAL;
41615693486bSJoel Becker mlog_errno(ret);
41625693486bSJoel Becker goto out;
41635693486bSJoel Becker }
41645693486bSJoel Becker
416509bf27a0STao Ma ret = filemap_fdatawrite(inode->i_mapping);
416609bf27a0STao Ma if (ret) {
416709bf27a0STao Ma mlog_errno(ret);
416809bf27a0STao Ma goto out;
416909bf27a0STao Ma }
417009bf27a0STao Ma
417109bf27a0STao Ma ret = ocfs2_attach_refcount_tree(inode, old_bh);
417209bf27a0STao Ma if (ret) {
417309bf27a0STao Ma mlog_errno(ret);
417409bf27a0STao Ma goto out;
417509bf27a0STao Ma }
417609bf27a0STao Ma
41775955102cSAl Viro inode_lock_nested(new_inode, I_MUTEX_CHILD);
417807eaac94STao Ma ret = ocfs2_inode_lock_nested(new_inode, &new_bh, 1,
417907eaac94STao Ma OI_LS_REFLINK_TARGET);
418009bf27a0STao Ma if (ret) {
418109bf27a0STao Ma mlog_errno(ret);
418209bf27a0STao Ma goto out_unlock;
418309bf27a0STao Ma }
418409bf27a0STao Ma
418509bf27a0STao Ma ret = ocfs2_create_reflink_node(inode, old_bh,
418609bf27a0STao Ma new_inode, new_bh, preserve);
418709bf27a0STao Ma if (ret) {
418809bf27a0STao Ma mlog_errno(ret);
418909bf27a0STao Ma goto inode_unlock;
419009bf27a0STao Ma }
419109bf27a0STao Ma
419209bf27a0STao Ma if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
419309bf27a0STao Ma ret = ocfs2_reflink_xattrs(inode, old_bh,
419409bf27a0STao Ma new_inode, new_bh,
419509bf27a0STao Ma preserve);
419687f4b1bbSTao Ma if (ret) {
419787f4b1bbSTao Ma mlog_errno(ret);
419887f4b1bbSTao Ma goto inode_unlock;
419987f4b1bbSTao Ma }
420087f4b1bbSTao Ma }
420187f4b1bbSTao Ma
420287f4b1bbSTao Ma ret = ocfs2_complete_reflink(inode, old_bh,
420387f4b1bbSTao Ma new_inode, new_bh, preserve);
420409bf27a0STao Ma if (ret)
420509bf27a0STao Ma mlog_errno(ret);
420687f4b1bbSTao Ma
420709bf27a0STao Ma inode_unlock:
420809bf27a0STao Ma ocfs2_inode_unlock(new_inode, 1);
420909bf27a0STao Ma brelse(new_bh);
421009bf27a0STao Ma out_unlock:
42115955102cSAl Viro inode_unlock(new_inode);
421209bf27a0STao Ma out:
421309bf27a0STao Ma if (!ret) {
421409bf27a0STao Ma ret = filemap_fdatawait(inode->i_mapping);
421509bf27a0STao Ma if (ret)
421609bf27a0STao Ma mlog_errno(ret);
421709bf27a0STao Ma }
421809bf27a0STao Ma return ret;
421909bf27a0STao Ma }
422009bf27a0STao Ma
ocfs2_reflink(struct dentry * old_dentry,struct inode * dir,struct dentry * new_dentry,bool preserve)422109bf27a0STao Ma static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
422209bf27a0STao Ma struct dentry *new_dentry, bool preserve)
422309bf27a0STao Ma {
4224e4383029SAshish Samant int error, had_lock;
42252b0143b5SDavid Howells struct inode *inode = d_inode(old_dentry);
422609bf27a0STao Ma struct buffer_head *old_bh = NULL;
422709bf27a0STao Ma struct inode *new_orphan_inode = NULL;
4228e4383029SAshish Samant struct ocfs2_lock_holder oh;
422909bf27a0STao Ma
423009bf27a0STao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
423109bf27a0STao Ma return -EOPNOTSUPP;
423209bf27a0STao Ma
4233702e5bc6SChristoph Hellwig
4234c25a1e06SJunxiao Bi error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
423509bf27a0STao Ma &new_orphan_inode);
423609bf27a0STao Ma if (error) {
423709bf27a0STao Ma mlog_errno(error);
423809bf27a0STao Ma goto out;
423909bf27a0STao Ma }
424009bf27a0STao Ma
42418a8ad1c2SWengang Wang error = ocfs2_rw_lock(inode, 1);
42428a8ad1c2SWengang Wang if (error) {
42438a8ad1c2SWengang Wang mlog_errno(error);
42448a8ad1c2SWengang Wang goto out;
42458a8ad1c2SWengang Wang }
42468a8ad1c2SWengang Wang
424709bf27a0STao Ma error = ocfs2_inode_lock(inode, &old_bh, 1);
424809bf27a0STao Ma if (error) {
424909bf27a0STao Ma mlog_errno(error);
42508a8ad1c2SWengang Wang ocfs2_rw_unlock(inode, 1);
425109bf27a0STao Ma goto out;
425209bf27a0STao Ma }
425309bf27a0STao Ma
425409bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_xattr_sem);
425509bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_alloc_sem);
425609bf27a0STao Ma error = __ocfs2_reflink(old_dentry, old_bh,
425709bf27a0STao Ma new_orphan_inode, preserve);
425809bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_alloc_sem);
425909bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_xattr_sem);
426009bf27a0STao Ma
426109bf27a0STao Ma ocfs2_inode_unlock(inode, 1);
42628a8ad1c2SWengang Wang ocfs2_rw_unlock(inode, 1);
426309bf27a0STao Ma brelse(old_bh);
426409bf27a0STao Ma
426509bf27a0STao Ma if (error) {
426609bf27a0STao Ma mlog_errno(error);
426709bf27a0STao Ma goto out;
426809bf27a0STao Ma }
426909bf27a0STao Ma
4270e4383029SAshish Samant had_lock = ocfs2_inode_lock_tracker(new_orphan_inode, NULL, 1,
4271e4383029SAshish Samant &oh);
4272e4383029SAshish Samant if (had_lock < 0) {
4273e4383029SAshish Samant error = had_lock;
4274e4383029SAshish Samant mlog_errno(error);
4275e4383029SAshish Samant goto out;
4276e4383029SAshish Samant }
4277e4383029SAshish Samant
427809bf27a0STao Ma /* If the security isn't preserved, we need to re-initialize them. */
427909bf27a0STao Ma if (!preserve) {
42802a7dba39SEric Paris error = ocfs2_init_security_and_acl(dir, new_orphan_inode,
4281c25a1e06SJunxiao Bi &new_dentry->d_name);
428209bf27a0STao Ma if (error)
428309bf27a0STao Ma mlog_errno(error);
428409bf27a0STao Ma }
428509bf27a0STao Ma if (!error) {
428609bf27a0STao Ma error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
428709bf27a0STao Ma new_dentry);
428809bf27a0STao Ma if (error)
428909bf27a0STao Ma mlog_errno(error);
429009bf27a0STao Ma }
4291e4383029SAshish Samant ocfs2_inode_unlock_tracker(new_orphan_inode, 1, &oh, had_lock);
429209bf27a0STao Ma
4293e4383029SAshish Samant out:
429409bf27a0STao Ma if (new_orphan_inode) {
429509bf27a0STao Ma /*
429609bf27a0STao Ma * We need to open_unlock the inode no matter whether we
429709bf27a0STao Ma * succeed or not, so that other nodes can delete it later.
429809bf27a0STao Ma */
429909bf27a0STao Ma ocfs2_open_unlock(new_orphan_inode);
430009bf27a0STao Ma if (error)
430109bf27a0STao Ma iput(new_orphan_inode);
430209bf27a0STao Ma }
430309bf27a0STao Ma
430409bf27a0STao Ma return error;
430509bf27a0STao Ma }
4306bd50873dSTao Ma
4307bd50873dSTao Ma /*
4308bd50873dSTao Ma * Below here are the bits used by OCFS2_IOC_REFLINK() to fake
4309bd50873dSTao Ma * sys_reflink(). This will go away when vfs_reflink() exists in
4310bd50873dSTao Ma * fs/namei.c.
4311bd50873dSTao Ma */
4312bd50873dSTao Ma
4313bd50873dSTao Ma /* copied from may_create in VFS. */
ocfs2_may_create(struct inode * dir,struct dentry * child)4314bd50873dSTao Ma static inline int ocfs2_may_create(struct inode *dir, struct dentry *child)
4315bd50873dSTao Ma {
43162b0143b5SDavid Howells if (d_really_is_positive(child))
4317bd50873dSTao Ma return -EEXIST;
4318bd50873dSTao Ma if (IS_DEADDIR(dir))
4319bd50873dSTao Ma return -ENOENT;
43204609e1f1SChristian Brauner return inode_permission(&nop_mnt_idmap, dir, MAY_WRITE | MAY_EXEC);
4321bd50873dSTao Ma }
4322bd50873dSTao Ma
4323bd50873dSTao Ma /**
4324bd50873dSTao Ma * ocfs2_vfs_reflink - Create a reference-counted link
4325bd50873dSTao Ma *
4326bd50873dSTao Ma * @old_dentry: source dentry + inode
4327bd50873dSTao Ma * @dir: directory to create the target
4328bd50873dSTao Ma * @new_dentry: target dentry
4329bd50873dSTao Ma * @preserve: if true, preserve all file attributes
4330bd50873dSTao Ma */
ocfs2_vfs_reflink(struct dentry * old_dentry,struct inode * dir,struct dentry * new_dentry,bool preserve)433112d4cec9STao Ma static int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir,
4332bd50873dSTao Ma struct dentry *new_dentry, bool preserve)
4333bd50873dSTao Ma {
43342b0143b5SDavid Howells struct inode *inode = d_inode(old_dentry);
4335bd50873dSTao Ma int error;
4336bd50873dSTao Ma
4337bd50873dSTao Ma if (!inode)
4338bd50873dSTao Ma return -ENOENT;
4339bd50873dSTao Ma
4340bd50873dSTao Ma error = ocfs2_may_create(dir, new_dentry);
4341bd50873dSTao Ma if (error)
4342bd50873dSTao Ma return error;
4343bd50873dSTao Ma
4344bd50873dSTao Ma if (dir->i_sb != inode->i_sb)
4345bd50873dSTao Ma return -EXDEV;
4346bd50873dSTao Ma
4347bd50873dSTao Ma /*
4348bd50873dSTao Ma * A reflink to an append-only or immutable file cannot be created.
4349bd50873dSTao Ma */
4350bd50873dSTao Ma if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4351bd50873dSTao Ma return -EPERM;
4352bd50873dSTao Ma
4353bd50873dSTao Ma /* Only regular files can be reflinked. */
4354bd50873dSTao Ma if (!S_ISREG(inode->i_mode))
4355bd50873dSTao Ma return -EPERM;
4356bd50873dSTao Ma
4357bd50873dSTao Ma /*
4358bd50873dSTao Ma * If the caller wants to preserve ownership, they require the
4359bd50873dSTao Ma * rights to do so.
4360bd50873dSTao Ma */
4361bd50873dSTao Ma if (preserve) {
4362488c8ef0SEric W. Biederman if (!uid_eq(current_fsuid(), inode->i_uid) && !capable(CAP_CHOWN))
4363bd50873dSTao Ma return -EPERM;
4364bd50873dSTao Ma if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN))
4365bd50873dSTao Ma return -EPERM;
4366bd50873dSTao Ma }
4367bd50873dSTao Ma
4368bd50873dSTao Ma /*
4369bd50873dSTao Ma * If the caller is modifying any aspect of the attributes, they
4370bd50873dSTao Ma * are not creating a snapshot. They need read permission on the
4371bd50873dSTao Ma * file.
4372bd50873dSTao Ma */
4373bd50873dSTao Ma if (!preserve) {
43744609e1f1SChristian Brauner error = inode_permission(&nop_mnt_idmap, inode, MAY_READ);
4375bd50873dSTao Ma if (error)
4376bd50873dSTao Ma return error;
4377bd50873dSTao Ma }
4378bd50873dSTao Ma
43795955102cSAl Viro inode_lock(inode);
43809c89fe0aSJan Kara error = dquot_initialize(dir);
43819c89fe0aSJan Kara if (!error)
4382bd50873dSTao Ma error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve);
43835955102cSAl Viro inode_unlock(inode);
4384bd50873dSTao Ma if (!error)
4385bd50873dSTao Ma fsnotify_create(dir, new_dentry);
4386bd50873dSTao Ma return error;
4387bd50873dSTao Ma }
4388bd50873dSTao Ma /*
4389bd50873dSTao Ma * Most codes are copied from sys_linkat.
4390bd50873dSTao Ma */
ocfs2_reflink_ioctl(struct inode * inode,const char __user * oldname,const char __user * newname,bool preserve)4391bd50873dSTao Ma int ocfs2_reflink_ioctl(struct inode *inode,
4392bd50873dSTao Ma const char __user *oldname,
4393bd50873dSTao Ma const char __user *newname,
4394bd50873dSTao Ma bool preserve)
4395bd50873dSTao Ma {
4396bd50873dSTao Ma struct dentry *new_dentry;
4397dae6ad8fSAl Viro struct path old_path, new_path;
4398bd50873dSTao Ma int error;
4399bd50873dSTao Ma
4400bd50873dSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4401bd50873dSTao Ma return -EOPNOTSUPP;
4402bd50873dSTao Ma
4403bd50873dSTao Ma error = user_path_at(AT_FDCWD, oldname, 0, &old_path);
4404bd50873dSTao Ma if (error) {
4405bd50873dSTao Ma mlog_errno(error);
4406bd50873dSTao Ma return error;
4407bd50873dSTao Ma }
4408bd50873dSTao Ma
4409dae6ad8fSAl Viro new_dentry = user_path_create(AT_FDCWD, newname, &new_path, 0);
4410dae6ad8fSAl Viro error = PTR_ERR(new_dentry);
4411dae6ad8fSAl Viro if (IS_ERR(new_dentry)) {
4412bd50873dSTao Ma mlog_errno(error);
4413bd50873dSTao Ma goto out;
4414bd50873dSTao Ma }
4415bd50873dSTao Ma
4416bd50873dSTao Ma error = -EXDEV;
4417dae6ad8fSAl Viro if (old_path.mnt != new_path.mnt) {
4418bd50873dSTao Ma mlog_errno(error);
4419dae6ad8fSAl Viro goto out_dput;
4420bd50873dSTao Ma }
4421bd50873dSTao Ma
4422bd50873dSTao Ma error = ocfs2_vfs_reflink(old_path.dentry,
44232b0143b5SDavid Howells d_inode(new_path.dentry),
4424bd50873dSTao Ma new_dentry, preserve);
4425bd50873dSTao Ma out_dput:
4426921a1650SAl Viro done_path_create(&new_path, new_dentry);
4427bd50873dSTao Ma out:
4428bd50873dSTao Ma path_put(&old_path);
4429bd50873dSTao Ma
4430bd50873dSTao Ma return error;
4431bd50873dSTao Ma }
443229ac8e85SDarrick J. Wong
443329ac8e85SDarrick J. Wong /* Update destination inode size, if necessary. */
ocfs2_reflink_update_dest(struct inode * dest,struct buffer_head * d_bh,loff_t newlen)443465f098e9SDarrick J. Wong int ocfs2_reflink_update_dest(struct inode *dest,
443529ac8e85SDarrick J. Wong struct buffer_head *d_bh,
443629ac8e85SDarrick J. Wong loff_t newlen)
443729ac8e85SDarrick J. Wong {
443829ac8e85SDarrick J. Wong handle_t *handle;
443929ac8e85SDarrick J. Wong int ret;
444029ac8e85SDarrick J. Wong
444129ac8e85SDarrick J. Wong dest->i_blocks = ocfs2_inode_sector_count(dest);
444229ac8e85SDarrick J. Wong
444329ac8e85SDarrick J. Wong if (newlen <= i_size_read(dest))
444429ac8e85SDarrick J. Wong return 0;
444529ac8e85SDarrick J. Wong
444629ac8e85SDarrick J. Wong handle = ocfs2_start_trans(OCFS2_SB(dest->i_sb),
444729ac8e85SDarrick J. Wong OCFS2_INODE_UPDATE_CREDITS);
444829ac8e85SDarrick J. Wong if (IS_ERR(handle)) {
444929ac8e85SDarrick J. Wong ret = PTR_ERR(handle);
445029ac8e85SDarrick J. Wong mlog_errno(ret);
445129ac8e85SDarrick J. Wong return ret;
445229ac8e85SDarrick J. Wong }
445329ac8e85SDarrick J. Wong
445429ac8e85SDarrick J. Wong /* Extend i_size if needed. */
445529ac8e85SDarrick J. Wong spin_lock(&OCFS2_I(dest)->ip_lock);
445629ac8e85SDarrick J. Wong if (newlen > i_size_read(dest))
445729ac8e85SDarrick J. Wong i_size_write(dest, newlen);
445829ac8e85SDarrick J. Wong spin_unlock(&OCFS2_I(dest)->ip_lock);
4459*10fc3a18SJeff Layton inode_set_mtime_to_ts(dest, inode_set_ctime_current(dest));
446029ac8e85SDarrick J. Wong
446129ac8e85SDarrick J. Wong ret = ocfs2_mark_inode_dirty(handle, dest, d_bh);
446229ac8e85SDarrick J. Wong if (ret) {
446329ac8e85SDarrick J. Wong mlog_errno(ret);
446429ac8e85SDarrick J. Wong goto out_commit;
446529ac8e85SDarrick J. Wong }
446629ac8e85SDarrick J. Wong
446729ac8e85SDarrick J. Wong out_commit:
446829ac8e85SDarrick J. Wong ocfs2_commit_trans(OCFS2_SB(dest->i_sb), handle);
446929ac8e85SDarrick J. Wong return ret;
447029ac8e85SDarrick J. Wong }
447129ac8e85SDarrick J. Wong
447229ac8e85SDarrick J. Wong /* Remap the range pos_in:len in s_inode to pos_out:len in t_inode. */
ocfs2_reflink_remap_extent(struct inode * s_inode,struct buffer_head * s_bh,loff_t pos_in,struct inode * t_inode,struct buffer_head * t_bh,loff_t pos_out,loff_t len,struct ocfs2_cached_dealloc_ctxt * dealloc)4473900611a1SDarrick J. Wong static loff_t ocfs2_reflink_remap_extent(struct inode *s_inode,
447429ac8e85SDarrick J. Wong struct buffer_head *s_bh,
447529ac8e85SDarrick J. Wong loff_t pos_in,
447629ac8e85SDarrick J. Wong struct inode *t_inode,
447729ac8e85SDarrick J. Wong struct buffer_head *t_bh,
447829ac8e85SDarrick J. Wong loff_t pos_out,
447929ac8e85SDarrick J. Wong loff_t len,
448029ac8e85SDarrick J. Wong struct ocfs2_cached_dealloc_ctxt *dealloc)
448129ac8e85SDarrick J. Wong {
448229ac8e85SDarrick J. Wong struct ocfs2_extent_tree s_et;
448329ac8e85SDarrick J. Wong struct ocfs2_extent_tree t_et;
448429ac8e85SDarrick J. Wong struct ocfs2_dinode *dis;
448529ac8e85SDarrick J. Wong struct buffer_head *ref_root_bh = NULL;
448629ac8e85SDarrick J. Wong struct ocfs2_refcount_tree *ref_tree;
448729ac8e85SDarrick J. Wong struct ocfs2_super *osb;
4488900611a1SDarrick J. Wong loff_t remapped_bytes = 0;
448929ac8e85SDarrick J. Wong loff_t pstart, plen;
4490900611a1SDarrick J. Wong u32 p_cluster, num_clusters, slast, spos, tpos, remapped_clus = 0;
449129ac8e85SDarrick J. Wong unsigned int ext_flags;
449229ac8e85SDarrick J. Wong int ret = 0;
449329ac8e85SDarrick J. Wong
449429ac8e85SDarrick J. Wong osb = OCFS2_SB(s_inode->i_sb);
449529ac8e85SDarrick J. Wong dis = (struct ocfs2_dinode *)s_bh->b_data;
449629ac8e85SDarrick J. Wong ocfs2_init_dinode_extent_tree(&s_et, INODE_CACHE(s_inode), s_bh);
449729ac8e85SDarrick J. Wong ocfs2_init_dinode_extent_tree(&t_et, INODE_CACHE(t_inode), t_bh);
449829ac8e85SDarrick J. Wong
449929ac8e85SDarrick J. Wong spos = ocfs2_bytes_to_clusters(s_inode->i_sb, pos_in);
450029ac8e85SDarrick J. Wong tpos = ocfs2_bytes_to_clusters(t_inode->i_sb, pos_out);
450129ac8e85SDarrick J. Wong slast = ocfs2_clusters_for_bytes(s_inode->i_sb, pos_in + len);
450229ac8e85SDarrick J. Wong
450329ac8e85SDarrick J. Wong while (spos < slast) {
450429ac8e85SDarrick J. Wong if (fatal_signal_pending(current)) {
450529ac8e85SDarrick J. Wong ret = -EINTR;
450629ac8e85SDarrick J. Wong goto out;
450729ac8e85SDarrick J. Wong }
450829ac8e85SDarrick J. Wong
450929ac8e85SDarrick J. Wong /* Look up the extent. */
451029ac8e85SDarrick J. Wong ret = ocfs2_get_clusters(s_inode, spos, &p_cluster,
451129ac8e85SDarrick J. Wong &num_clusters, &ext_flags);
451229ac8e85SDarrick J. Wong if (ret) {
451329ac8e85SDarrick J. Wong mlog_errno(ret);
451429ac8e85SDarrick J. Wong goto out;
451529ac8e85SDarrick J. Wong }
451629ac8e85SDarrick J. Wong
451729ac8e85SDarrick J. Wong num_clusters = min_t(u32, num_clusters, slast - spos);
451829ac8e85SDarrick J. Wong
451929ac8e85SDarrick J. Wong /* Punch out the dest range. */
452029ac8e85SDarrick J. Wong pstart = ocfs2_clusters_to_bytes(t_inode->i_sb, tpos);
452129ac8e85SDarrick J. Wong plen = ocfs2_clusters_to_bytes(t_inode->i_sb, num_clusters);
452229ac8e85SDarrick J. Wong ret = ocfs2_remove_inode_range(t_inode, t_bh, pstart, plen);
452329ac8e85SDarrick J. Wong if (ret) {
452429ac8e85SDarrick J. Wong mlog_errno(ret);
452529ac8e85SDarrick J. Wong goto out;
452629ac8e85SDarrick J. Wong }
452729ac8e85SDarrick J. Wong
452829ac8e85SDarrick J. Wong if (p_cluster == 0)
452929ac8e85SDarrick J. Wong goto next_loop;
453029ac8e85SDarrick J. Wong
453129ac8e85SDarrick J. Wong /* Lock the refcount btree... */
453229ac8e85SDarrick J. Wong ret = ocfs2_lock_refcount_tree(osb,
453329ac8e85SDarrick J. Wong le64_to_cpu(dis->i_refcount_loc),
453429ac8e85SDarrick J. Wong 1, &ref_tree, &ref_root_bh);
453529ac8e85SDarrick J. Wong if (ret) {
453629ac8e85SDarrick J. Wong mlog_errno(ret);
453729ac8e85SDarrick J. Wong goto out;
453829ac8e85SDarrick J. Wong }
453929ac8e85SDarrick J. Wong
454029ac8e85SDarrick J. Wong /* Mark s_inode's extent as refcounted. */
454129ac8e85SDarrick J. Wong if (!(ext_flags & OCFS2_EXT_REFCOUNTED)) {
454229ac8e85SDarrick J. Wong ret = ocfs2_add_refcount_flag(s_inode, &s_et,
454329ac8e85SDarrick J. Wong &ref_tree->rf_ci,
454429ac8e85SDarrick J. Wong ref_root_bh, spos,
454529ac8e85SDarrick J. Wong p_cluster, num_clusters,
454629ac8e85SDarrick J. Wong dealloc, NULL);
454729ac8e85SDarrick J. Wong if (ret) {
454829ac8e85SDarrick J. Wong mlog_errno(ret);
454929ac8e85SDarrick J. Wong goto out_unlock_refcount;
455029ac8e85SDarrick J. Wong }
455129ac8e85SDarrick J. Wong }
455229ac8e85SDarrick J. Wong
455329ac8e85SDarrick J. Wong /* Map in the new extent. */
455429ac8e85SDarrick J. Wong ext_flags |= OCFS2_EXT_REFCOUNTED;
455529ac8e85SDarrick J. Wong ret = ocfs2_add_refcounted_extent(t_inode, &t_et,
455629ac8e85SDarrick J. Wong &ref_tree->rf_ci,
455729ac8e85SDarrick J. Wong ref_root_bh,
455829ac8e85SDarrick J. Wong tpos, p_cluster,
455929ac8e85SDarrick J. Wong num_clusters,
456029ac8e85SDarrick J. Wong ext_flags,
456129ac8e85SDarrick J. Wong dealloc);
456229ac8e85SDarrick J. Wong if (ret) {
456329ac8e85SDarrick J. Wong mlog_errno(ret);
456429ac8e85SDarrick J. Wong goto out_unlock_refcount;
456529ac8e85SDarrick J. Wong }
456629ac8e85SDarrick J. Wong
456729ac8e85SDarrick J. Wong ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
456829ac8e85SDarrick J. Wong brelse(ref_root_bh);
456929ac8e85SDarrick J. Wong next_loop:
457029ac8e85SDarrick J. Wong spos += num_clusters;
457129ac8e85SDarrick J. Wong tpos += num_clusters;
4572900611a1SDarrick J. Wong remapped_clus += num_clusters;
457329ac8e85SDarrick J. Wong }
457429ac8e85SDarrick J. Wong
4575900611a1SDarrick J. Wong goto out;
457629ac8e85SDarrick J. Wong out_unlock_refcount:
457729ac8e85SDarrick J. Wong ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
457829ac8e85SDarrick J. Wong brelse(ref_root_bh);
4579900611a1SDarrick J. Wong out:
4580900611a1SDarrick J. Wong remapped_bytes = ocfs2_clusters_to_bytes(t_inode->i_sb, remapped_clus);
4581900611a1SDarrick J. Wong remapped_bytes = min_t(loff_t, len, remapped_bytes);
4582900611a1SDarrick J. Wong
4583900611a1SDarrick J. Wong return remapped_bytes > 0 ? remapped_bytes : ret;
458429ac8e85SDarrick J. Wong }
458529ac8e85SDarrick J. Wong
458629ac8e85SDarrick J. Wong /* Set up refcount tree and remap s_inode to t_inode. */
ocfs2_reflink_remap_blocks(struct inode * s_inode,struct buffer_head * s_bh,loff_t pos_in,struct inode * t_inode,struct buffer_head * t_bh,loff_t pos_out,loff_t len)458765f098e9SDarrick J. Wong loff_t ocfs2_reflink_remap_blocks(struct inode *s_inode,
458829ac8e85SDarrick J. Wong struct buffer_head *s_bh,
458929ac8e85SDarrick J. Wong loff_t pos_in,
459029ac8e85SDarrick J. Wong struct inode *t_inode,
459129ac8e85SDarrick J. Wong struct buffer_head *t_bh,
459229ac8e85SDarrick J. Wong loff_t pos_out,
459329ac8e85SDarrick J. Wong loff_t len)
459429ac8e85SDarrick J. Wong {
459529ac8e85SDarrick J. Wong struct ocfs2_cached_dealloc_ctxt dealloc;
459629ac8e85SDarrick J. Wong struct ocfs2_super *osb;
459729ac8e85SDarrick J. Wong struct ocfs2_dinode *dis;
459829ac8e85SDarrick J. Wong struct ocfs2_dinode *dit;
4599900611a1SDarrick J. Wong loff_t ret;
460029ac8e85SDarrick J. Wong
460129ac8e85SDarrick J. Wong osb = OCFS2_SB(s_inode->i_sb);
460229ac8e85SDarrick J. Wong dis = (struct ocfs2_dinode *)s_bh->b_data;
460329ac8e85SDarrick J. Wong dit = (struct ocfs2_dinode *)t_bh->b_data;
460429ac8e85SDarrick J. Wong ocfs2_init_dealloc_ctxt(&dealloc);
460529ac8e85SDarrick J. Wong
460629ac8e85SDarrick J. Wong /*
460729ac8e85SDarrick J. Wong * If we're reflinking the entire file and the source is inline
460829ac8e85SDarrick J. Wong * data, just copy the contents.
460929ac8e85SDarrick J. Wong */
461029ac8e85SDarrick J. Wong if (pos_in == pos_out && pos_in == 0 && len == i_size_read(s_inode) &&
461129ac8e85SDarrick J. Wong i_size_read(t_inode) <= len &&
461229ac8e85SDarrick J. Wong (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) {
461329ac8e85SDarrick J. Wong ret = ocfs2_duplicate_inline_data(s_inode, s_bh, t_inode, t_bh);
461429ac8e85SDarrick J. Wong if (ret)
461529ac8e85SDarrick J. Wong mlog_errno(ret);
461629ac8e85SDarrick J. Wong goto out;
461729ac8e85SDarrick J. Wong }
461829ac8e85SDarrick J. Wong
461929ac8e85SDarrick J. Wong /*
462029ac8e85SDarrick J. Wong * If both inodes belong to two different refcount groups then
462129ac8e85SDarrick J. Wong * forget it because we don't know how (or want) to go merging
462229ac8e85SDarrick J. Wong * refcount trees.
462329ac8e85SDarrick J. Wong */
462429ac8e85SDarrick J. Wong ret = -EOPNOTSUPP;
462529ac8e85SDarrick J. Wong if (ocfs2_is_refcount_inode(s_inode) &&
462629ac8e85SDarrick J. Wong ocfs2_is_refcount_inode(t_inode) &&
462729ac8e85SDarrick J. Wong le64_to_cpu(dis->i_refcount_loc) !=
462829ac8e85SDarrick J. Wong le64_to_cpu(dit->i_refcount_loc))
462929ac8e85SDarrick J. Wong goto out;
463029ac8e85SDarrick J. Wong
463129ac8e85SDarrick J. Wong /* Neither inode has a refcount tree. Add one to s_inode. */
463229ac8e85SDarrick J. Wong if (!ocfs2_is_refcount_inode(s_inode) &&
463329ac8e85SDarrick J. Wong !ocfs2_is_refcount_inode(t_inode)) {
463429ac8e85SDarrick J. Wong ret = ocfs2_create_refcount_tree(s_inode, s_bh);
463529ac8e85SDarrick J. Wong if (ret) {
463629ac8e85SDarrick J. Wong mlog_errno(ret);
463729ac8e85SDarrick J. Wong goto out;
463829ac8e85SDarrick J. Wong }
463929ac8e85SDarrick J. Wong }
464029ac8e85SDarrick J. Wong
464129ac8e85SDarrick J. Wong /* Ensure that both inodes end up with the same refcount tree. */
464229ac8e85SDarrick J. Wong if (!ocfs2_is_refcount_inode(s_inode)) {
464329ac8e85SDarrick J. Wong ret = ocfs2_set_refcount_tree(s_inode, s_bh,
464429ac8e85SDarrick J. Wong le64_to_cpu(dit->i_refcount_loc));
464529ac8e85SDarrick J. Wong if (ret) {
464629ac8e85SDarrick J. Wong mlog_errno(ret);
464729ac8e85SDarrick J. Wong goto out;
464829ac8e85SDarrick J. Wong }
464929ac8e85SDarrick J. Wong }
465029ac8e85SDarrick J. Wong if (!ocfs2_is_refcount_inode(t_inode)) {
465129ac8e85SDarrick J. Wong ret = ocfs2_set_refcount_tree(t_inode, t_bh,
465229ac8e85SDarrick J. Wong le64_to_cpu(dis->i_refcount_loc));
465329ac8e85SDarrick J. Wong if (ret) {
465429ac8e85SDarrick J. Wong mlog_errno(ret);
465529ac8e85SDarrick J. Wong goto out;
465629ac8e85SDarrick J. Wong }
465729ac8e85SDarrick J. Wong }
465829ac8e85SDarrick J. Wong
465929ac8e85SDarrick J. Wong /* Turn off inline data in the dest file. */
466029ac8e85SDarrick J. Wong if (OCFS2_I(t_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
466129ac8e85SDarrick J. Wong ret = ocfs2_convert_inline_data_to_extents(t_inode, t_bh);
466229ac8e85SDarrick J. Wong if (ret) {
466329ac8e85SDarrick J. Wong mlog_errno(ret);
466429ac8e85SDarrick J. Wong goto out;
466529ac8e85SDarrick J. Wong }
466629ac8e85SDarrick J. Wong }
466729ac8e85SDarrick J. Wong
466829ac8e85SDarrick J. Wong /* Actually remap extents now. */
466929ac8e85SDarrick J. Wong ret = ocfs2_reflink_remap_extent(s_inode, s_bh, pos_in, t_inode, t_bh,
467029ac8e85SDarrick J. Wong pos_out, len, &dealloc);
4671900611a1SDarrick J. Wong if (ret < 0) {
467229ac8e85SDarrick J. Wong mlog_errno(ret);
467329ac8e85SDarrick J. Wong goto out;
467429ac8e85SDarrick J. Wong }
467529ac8e85SDarrick J. Wong
467629ac8e85SDarrick J. Wong out:
467729ac8e85SDarrick J. Wong if (ocfs2_dealloc_has_cluster(&dealloc)) {
467829ac8e85SDarrick J. Wong ocfs2_schedule_truncate_log_flush(osb, 1);
467929ac8e85SDarrick J. Wong ocfs2_run_deallocs(osb, &dealloc);
468029ac8e85SDarrick J. Wong }
468129ac8e85SDarrick J. Wong
468229ac8e85SDarrick J. Wong return ret;
468329ac8e85SDarrick J. Wong }
468429ac8e85SDarrick J. Wong
468529ac8e85SDarrick J. Wong /* Lock an inode and grab a bh pointing to the inode. */
ocfs2_reflink_inodes_lock(struct inode * s_inode,struct buffer_head ** bh_s,struct inode * t_inode,struct buffer_head ** bh_t)468665f098e9SDarrick J. Wong int ocfs2_reflink_inodes_lock(struct inode *s_inode,
4687e6a9467eSDarrick J. Wong struct buffer_head **bh_s,
468829ac8e85SDarrick J. Wong struct inode *t_inode,
4689e6a9467eSDarrick J. Wong struct buffer_head **bh_t)
469029ac8e85SDarrick J. Wong {
4691e6a9467eSDarrick J. Wong struct inode *inode1 = s_inode;
4692e6a9467eSDarrick J. Wong struct inode *inode2 = t_inode;
469329ac8e85SDarrick J. Wong struct ocfs2_inode_info *oi1;
469429ac8e85SDarrick J. Wong struct ocfs2_inode_info *oi2;
4695e6a9467eSDarrick J. Wong struct buffer_head *bh1 = NULL;
4696e6a9467eSDarrick J. Wong struct buffer_head *bh2 = NULL;
469729ac8e85SDarrick J. Wong bool same_inode = (s_inode == t_inode);
4698e6a9467eSDarrick J. Wong bool need_swap = (inode1->i_ino > inode2->i_ino);
469929ac8e85SDarrick J. Wong int status;
470029ac8e85SDarrick J. Wong
470129ac8e85SDarrick J. Wong /* First grab the VFS and rw locks. */
470229ac8e85SDarrick J. Wong lock_two_nondirectories(s_inode, t_inode);
4703e6a9467eSDarrick J. Wong if (need_swap)
470429ac8e85SDarrick J. Wong swap(inode1, inode2);
470529ac8e85SDarrick J. Wong
470629ac8e85SDarrick J. Wong status = ocfs2_rw_lock(inode1, 1);
470729ac8e85SDarrick J. Wong if (status) {
470829ac8e85SDarrick J. Wong mlog_errno(status);
470929ac8e85SDarrick J. Wong goto out_i1;
471029ac8e85SDarrick J. Wong }
471129ac8e85SDarrick J. Wong if (!same_inode) {
471229ac8e85SDarrick J. Wong status = ocfs2_rw_lock(inode2, 1);
471329ac8e85SDarrick J. Wong if (status) {
471429ac8e85SDarrick J. Wong mlog_errno(status);
471529ac8e85SDarrick J. Wong goto out_i2;
471629ac8e85SDarrick J. Wong }
471729ac8e85SDarrick J. Wong }
471829ac8e85SDarrick J. Wong
471929ac8e85SDarrick J. Wong /* Now go for the cluster locks */
472029ac8e85SDarrick J. Wong oi1 = OCFS2_I(inode1);
472129ac8e85SDarrick J. Wong oi2 = OCFS2_I(inode2);
472229ac8e85SDarrick J. Wong
472329ac8e85SDarrick J. Wong trace_ocfs2_double_lock((unsigned long long)oi1->ip_blkno,
472429ac8e85SDarrick J. Wong (unsigned long long)oi2->ip_blkno);
472529ac8e85SDarrick J. Wong
472629ac8e85SDarrick J. Wong /* We always want to lock the one with the lower lockid first. */
472729ac8e85SDarrick J. Wong if (oi1->ip_blkno > oi2->ip_blkno)
472829ac8e85SDarrick J. Wong mlog_errno(-ENOLCK);
472929ac8e85SDarrick J. Wong
473029ac8e85SDarrick J. Wong /* lock id1 */
4731e6a9467eSDarrick J. Wong status = ocfs2_inode_lock_nested(inode1, &bh1, 1,
4732e6a9467eSDarrick J. Wong OI_LS_REFLINK_TARGET);
473329ac8e85SDarrick J. Wong if (status < 0) {
473429ac8e85SDarrick J. Wong if (status != -ENOENT)
473529ac8e85SDarrick J. Wong mlog_errno(status);
473629ac8e85SDarrick J. Wong goto out_rw2;
473729ac8e85SDarrick J. Wong }
473829ac8e85SDarrick J. Wong
473929ac8e85SDarrick J. Wong /* lock id2 */
474029ac8e85SDarrick J. Wong if (!same_inode) {
4741e6a9467eSDarrick J. Wong status = ocfs2_inode_lock_nested(inode2, &bh2, 1,
474229ac8e85SDarrick J. Wong OI_LS_REFLINK_TARGET);
474329ac8e85SDarrick J. Wong if (status < 0) {
474429ac8e85SDarrick J. Wong if (status != -ENOENT)
474529ac8e85SDarrick J. Wong mlog_errno(status);
474629ac8e85SDarrick J. Wong goto out_cl1;
474729ac8e85SDarrick J. Wong }
4748e6a9467eSDarrick J. Wong } else {
4749e6a9467eSDarrick J. Wong bh2 = bh1;
4750e6a9467eSDarrick J. Wong }
4751e6a9467eSDarrick J. Wong
4752e6a9467eSDarrick J. Wong /*
4753e6a9467eSDarrick J. Wong * If we swapped inode order above, we have to swap the buffer heads
4754e6a9467eSDarrick J. Wong * before passing them back to the caller.
4755e6a9467eSDarrick J. Wong */
4756e6a9467eSDarrick J. Wong if (need_swap)
4757e6a9467eSDarrick J. Wong swap(bh1, bh2);
4758e6a9467eSDarrick J. Wong *bh_s = bh1;
4759e6a9467eSDarrick J. Wong *bh_t = bh2;
476029ac8e85SDarrick J. Wong
476129ac8e85SDarrick J. Wong trace_ocfs2_double_lock_end(
4762d324cd4cSpiaojun (unsigned long long)oi1->ip_blkno,
4763d324cd4cSpiaojun (unsigned long long)oi2->ip_blkno);
476429ac8e85SDarrick J. Wong
476529ac8e85SDarrick J. Wong return 0;
476629ac8e85SDarrick J. Wong
476729ac8e85SDarrick J. Wong out_cl1:
476829ac8e85SDarrick J. Wong ocfs2_inode_unlock(inode1, 1);
4769e6a9467eSDarrick J. Wong brelse(bh1);
477029ac8e85SDarrick J. Wong out_rw2:
477129ac8e85SDarrick J. Wong ocfs2_rw_unlock(inode2, 1);
477229ac8e85SDarrick J. Wong out_i2:
477329ac8e85SDarrick J. Wong ocfs2_rw_unlock(inode1, 1);
477429ac8e85SDarrick J. Wong out_i1:
477529ac8e85SDarrick J. Wong unlock_two_nondirectories(s_inode, t_inode);
477629ac8e85SDarrick J. Wong return status;
477729ac8e85SDarrick J. Wong }
477829ac8e85SDarrick J. Wong
477929ac8e85SDarrick J. Wong /* Unlock both inodes and release buffers. */
ocfs2_reflink_inodes_unlock(struct inode * s_inode,struct buffer_head * s_bh,struct inode * t_inode,struct buffer_head * t_bh)478065f098e9SDarrick J. Wong void ocfs2_reflink_inodes_unlock(struct inode *s_inode,
478129ac8e85SDarrick J. Wong struct buffer_head *s_bh,
478229ac8e85SDarrick J. Wong struct inode *t_inode,
478329ac8e85SDarrick J. Wong struct buffer_head *t_bh)
478429ac8e85SDarrick J. Wong {
478529ac8e85SDarrick J. Wong ocfs2_inode_unlock(s_inode, 1);
478629ac8e85SDarrick J. Wong ocfs2_rw_unlock(s_inode, 1);
478729ac8e85SDarrick J. Wong brelse(s_bh);
478829ac8e85SDarrick J. Wong if (s_inode != t_inode) {
478929ac8e85SDarrick J. Wong ocfs2_inode_unlock(t_inode, 1);
479029ac8e85SDarrick J. Wong ocfs2_rw_unlock(t_inode, 1);
479129ac8e85SDarrick J. Wong brelse(t_bh);
479229ac8e85SDarrick J. Wong }
479329ac8e85SDarrick J. Wong unlock_two_nondirectories(s_inode, t_inode);
479429ac8e85SDarrick J. Wong }
4795