nilfs2: fix buffer head leak in nilfs_btnode_submit_block
[linux-2.6.git] / fs / nilfs2 / btnode.c
1 /*
2  * btnode.c - NILFS B-tree node cache
3  *
4  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License as published by
8  * the Free Software Foundation; either version 2 of the License, or
9  * (at your option) any later version.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
19  *
20  * This file was originally written by Seiji Kihara <kihara@osrg.net>
21  * and fully revised by Ryusuke Konishi <ryusuke@osrg.net> for
22  * stabilization and simplification.
23  *
24  */
25
26 #include <linux/types.h>
27 #include <linux/buffer_head.h>
28 #include <linux/mm.h>
29 #include <linux/backing-dev.h>
30 #include <linux/gfp.h>
31 #include "nilfs.h"
32 #include "mdt.h"
33 #include "dat.h"
34 #include "page.h"
35 #include "btnode.h"
36
37
38 void nilfs_btnode_cache_init_once(struct address_space *btnc)
39 {
40         memset(btnc, 0, sizeof(*btnc));
41         INIT_RADIX_TREE(&btnc->page_tree, GFP_ATOMIC);
42         spin_lock_init(&btnc->tree_lock);
43         INIT_LIST_HEAD(&btnc->private_list);
44         spin_lock_init(&btnc->private_lock);
45
46         spin_lock_init(&btnc->i_mmap_lock);
47         INIT_RAW_PRIO_TREE_ROOT(&btnc->i_mmap);
48         INIT_LIST_HEAD(&btnc->i_mmap_nonlinear);
49 }
50
51 static const struct address_space_operations def_btnode_aops = {
52         .sync_page              = block_sync_page,
53 };
54
55 void nilfs_btnode_cache_init(struct address_space *btnc,
56                              struct backing_dev_info *bdi)
57 {
58         btnc->host = NULL;  /* can safely set to host inode ? */
59         btnc->flags = 0;
60         mapping_set_gfp_mask(btnc, GFP_NOFS);
61         btnc->assoc_mapping = NULL;
62         btnc->backing_dev_info = bdi;
63         btnc->a_ops = &def_btnode_aops;
64 }
65
66 void nilfs_btnode_cache_clear(struct address_space *btnc)
67 {
68         invalidate_mapping_pages(btnc, 0, -1);
69         truncate_inode_pages(btnc, 0);
70 }
71
72 struct buffer_head *
73 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
74 {
75         struct inode *inode = NILFS_BTNC_I(btnc);
76         struct buffer_head *bh;
77
78         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
79         if (unlikely(!bh))
80                 return NULL;
81
82         if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
83                      buffer_dirty(bh))) {
84                 brelse(bh);
85                 BUG();
86         }
87         memset(bh->b_data, 0, 1 << inode->i_blkbits);
88         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
89         bh->b_blocknr = blocknr;
90         set_buffer_mapped(bh);
91         set_buffer_uptodate(bh);
92
93         unlock_page(bh->b_page);
94         page_cache_release(bh->b_page);
95         return bh;
96 }
97
98 int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
99                               sector_t pblocknr, struct buffer_head **pbh)
100 {
101         struct buffer_head *bh;
102         struct inode *inode = NILFS_BTNC_I(btnc);
103         struct page *page;
104         int err;
105
106         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
107         if (unlikely(!bh))
108                 return -ENOMEM;
109
110         err = -EEXIST; /* internal code */
111         page = bh->b_page;
112
113         if (buffer_uptodate(bh) || buffer_dirty(bh))
114                 goto found;
115
116         if (pblocknr == 0) {
117                 pblocknr = blocknr;
118                 if (inode->i_ino != NILFS_DAT_INO) {
119                         struct inode *dat =
120                                 nilfs_dat_inode(NILFS_I_NILFS(inode));
121
122                         /* blocknr is a virtual block number */
123                         err = nilfs_dat_translate(dat, blocknr, &pblocknr);
124                         if (unlikely(err)) {
125                                 brelse(bh);
126                                 goto out_locked;
127                         }
128                 }
129         }
130         lock_buffer(bh);
131         if (buffer_uptodate(bh)) {
132                 unlock_buffer(bh);
133                 err = -EEXIST; /* internal code */
134                 goto found;
135         }
136         set_buffer_mapped(bh);
137         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
138         bh->b_blocknr = pblocknr; /* set block address for read */
139         bh->b_end_io = end_buffer_read_sync;
140         get_bh(bh);
141         submit_bh(READ, bh);
142         bh->b_blocknr = blocknr; /* set back to the given block address */
143         err = 0;
144 found:
145         *pbh = bh;
146
147 out_locked:
148         unlock_page(page);
149         page_cache_release(page);
150         return err;
151 }
152
153 /**
154  * nilfs_btnode_delete - delete B-tree node buffer
155  * @bh: buffer to be deleted
156  *
157  * nilfs_btnode_delete() invalidates the specified buffer and delete the page
158  * including the buffer if the page gets unbusy.
159  */
160 void nilfs_btnode_delete(struct buffer_head *bh)
161 {
162         struct address_space *mapping;
163         struct page *page = bh->b_page;
164         pgoff_t index = page_index(page);
165         int still_dirty;
166
167         page_cache_get(page);
168         lock_page(page);
169         wait_on_page_writeback(page);
170
171         nilfs_forget_buffer(bh);
172         still_dirty = PageDirty(page);
173         mapping = page->mapping;
174         unlock_page(page);
175         page_cache_release(page);
176
177         if (!still_dirty && mapping)
178                 invalidate_inode_pages2_range(mapping, index, index);
179 }
180
181 /**
182  * nilfs_btnode_prepare_change_key
183  *  prepare to move contents of the block for old key to one of new key.
184  *  the old buffer will not be removed, but might be reused for new buffer.
185  *  it might return -ENOMEM because of memory allocation errors,
186  *  and might return -EIO because of disk read errors.
187  */
188 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
189                                     struct nilfs_btnode_chkey_ctxt *ctxt)
190 {
191         struct buffer_head *obh, *nbh;
192         struct inode *inode = NILFS_BTNC_I(btnc);
193         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
194         int err;
195
196         if (oldkey == newkey)
197                 return 0;
198
199         obh = ctxt->bh;
200         ctxt->newbh = NULL;
201
202         if (inode->i_blkbits == PAGE_CACHE_SHIFT) {
203                 lock_page(obh->b_page);
204                 /*
205                  * We cannot call radix_tree_preload for the kernels older
206                  * than 2.6.23, because it is not exported for modules.
207                  */
208 retry:
209                 err = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
210                 if (err)
211                         goto failed_unlock;
212                 /* BUG_ON(oldkey != obh->b_page->index); */
213                 if (unlikely(oldkey != obh->b_page->index))
214                         NILFS_PAGE_BUG(obh->b_page,
215                                        "invalid oldkey %lld (newkey=%lld)",
216                                        (unsigned long long)oldkey,
217                                        (unsigned long long)newkey);
218
219                 spin_lock_irq(&btnc->tree_lock);
220                 err = radix_tree_insert(&btnc->page_tree, newkey, obh->b_page);
221                 spin_unlock_irq(&btnc->tree_lock);
222                 /*
223                  * Note: page->index will not change to newkey until
224                  * nilfs_btnode_commit_change_key() will be called.
225                  * To protect the page in intermediate state, the page lock
226                  * is held.
227                  */
228                 radix_tree_preload_end();
229                 if (!err)
230                         return 0;
231                 else if (err != -EEXIST)
232                         goto failed_unlock;
233
234                 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
235                 if (!err)
236                         goto retry;
237                 /* fallback to copy mode */
238                 unlock_page(obh->b_page);
239         }
240
241         nbh = nilfs_btnode_create_block(btnc, newkey);
242         if (!nbh)
243                 return -ENOMEM;
244
245         BUG_ON(nbh == obh);
246         ctxt->newbh = nbh;
247         return 0;
248
249  failed_unlock:
250         unlock_page(obh->b_page);
251         return err;
252 }
253
254 /**
255  * nilfs_btnode_commit_change_key
256  *  commit the change_key operation prepared by prepare_change_key().
257  */
258 void nilfs_btnode_commit_change_key(struct address_space *btnc,
259                                     struct nilfs_btnode_chkey_ctxt *ctxt)
260 {
261         struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
262         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
263         struct page *opage;
264
265         if (oldkey == newkey)
266                 return;
267
268         if (nbh == NULL) {      /* blocksize == pagesize */
269                 opage = obh->b_page;
270                 if (unlikely(oldkey != opage->index))
271                         NILFS_PAGE_BUG(opage,
272                                        "invalid oldkey %lld (newkey=%lld)",
273                                        (unsigned long long)oldkey,
274                                        (unsigned long long)newkey);
275                 nilfs_btnode_mark_dirty(obh);
276
277                 spin_lock_irq(&btnc->tree_lock);
278                 radix_tree_delete(&btnc->page_tree, oldkey);
279                 radix_tree_tag_set(&btnc->page_tree, newkey,
280                                    PAGECACHE_TAG_DIRTY);
281                 spin_unlock_irq(&btnc->tree_lock);
282
283                 opage->index = obh->b_blocknr = newkey;
284                 unlock_page(opage);
285         } else {
286                 nilfs_copy_buffer(nbh, obh);
287                 nilfs_btnode_mark_dirty(nbh);
288
289                 nbh->b_blocknr = newkey;
290                 ctxt->bh = nbh;
291                 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
292         }
293 }
294
295 /**
296  * nilfs_btnode_abort_change_key
297  *  abort the change_key operation prepared by prepare_change_key().
298  */
299 void nilfs_btnode_abort_change_key(struct address_space *btnc,
300                                    struct nilfs_btnode_chkey_ctxt *ctxt)
301 {
302         struct buffer_head *nbh = ctxt->newbh;
303         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
304
305         if (oldkey == newkey)
306                 return;
307
308         if (nbh == NULL) {      /* blocksize == pagesize */
309                 spin_lock_irq(&btnc->tree_lock);
310                 radix_tree_delete(&btnc->page_tree, newkey);
311                 spin_unlock_irq(&btnc->tree_lock);
312                 unlock_page(ctxt->bh->b_page);
313         } else
314                 brelse(nbh);
315 }