5115814cb74503bd4d34502817e9713a6f90541d
[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         nilfs_mapping_init_once(btnc);
41 }
42
43 static const struct address_space_operations def_btnode_aops = {
44         .sync_page              = block_sync_page,
45 };
46
47 void nilfs_btnode_cache_init(struct address_space *btnc,
48                              struct backing_dev_info *bdi)
49 {
50         nilfs_mapping_init(btnc, bdi, &def_btnode_aops);
51 }
52
53 void nilfs_btnode_cache_clear(struct address_space *btnc)
54 {
55         invalidate_mapping_pages(btnc, 0, -1);
56         truncate_inode_pages(btnc, 0);
57 }
58
59 struct buffer_head *
60 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
61 {
62         struct inode *inode = NILFS_BTNC_I(btnc);
63         struct buffer_head *bh;
64
65         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
66         if (unlikely(!bh))
67                 return NULL;
68
69         if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
70                      buffer_dirty(bh))) {
71                 brelse(bh);
72                 BUG();
73         }
74         memset(bh->b_data, 0, 1 << inode->i_blkbits);
75         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
76         bh->b_blocknr = blocknr;
77         set_buffer_mapped(bh);
78         set_buffer_uptodate(bh);
79
80         unlock_page(bh->b_page);
81         page_cache_release(bh->b_page);
82         return bh;
83 }
84
85 int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
86                               sector_t pblocknr, int mode,
87                               struct buffer_head **pbh, sector_t *submit_ptr)
88 {
89         struct buffer_head *bh;
90         struct inode *inode = NILFS_BTNC_I(btnc);
91         struct page *page;
92         int err;
93
94         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
95         if (unlikely(!bh))
96                 return -ENOMEM;
97
98         err = -EEXIST; /* internal code */
99         page = bh->b_page;
100
101         if (buffer_uptodate(bh) || buffer_dirty(bh))
102                 goto found;
103
104         if (pblocknr == 0) {
105                 pblocknr = blocknr;
106                 if (inode->i_ino != NILFS_DAT_INO) {
107                         struct inode *dat =
108                                 nilfs_dat_inode(NILFS_I_NILFS(inode));
109
110                         /* blocknr is a virtual block number */
111                         err = nilfs_dat_translate(dat, blocknr, &pblocknr);
112                         if (unlikely(err)) {
113                                 brelse(bh);
114                                 goto out_locked;
115                         }
116                 }
117         }
118
119         if (mode == READA) {
120                 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
121                         err = -EBUSY; /* internal code */
122                         brelse(bh);
123                         goto out_locked;
124                 }
125         } else { /* mode == READ */
126                 lock_buffer(bh);
127         }
128         if (buffer_uptodate(bh)) {
129                 unlock_buffer(bh);
130                 err = -EEXIST; /* internal code */
131                 goto found;
132         }
133         set_buffer_mapped(bh);
134         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
135         bh->b_blocknr = pblocknr; /* set block address for read */
136         bh->b_end_io = end_buffer_read_sync;
137         get_bh(bh);
138         submit_bh(mode, bh);
139         bh->b_blocknr = blocknr; /* set back to the given block address */
140         *submit_ptr = pblocknr;
141         err = 0;
142 found:
143         *pbh = bh;
144
145 out_locked:
146         unlock_page(page);
147         page_cache_release(page);
148         return err;
149 }
150
151 /**
152  * nilfs_btnode_delete - delete B-tree node buffer
153  * @bh: buffer to be deleted
154  *
155  * nilfs_btnode_delete() invalidates the specified buffer and delete the page
156  * including the buffer if the page gets unbusy.
157  */
158 void nilfs_btnode_delete(struct buffer_head *bh)
159 {
160         struct address_space *mapping;
161         struct page *page = bh->b_page;
162         pgoff_t index = page_index(page);
163         int still_dirty;
164
165         page_cache_get(page);
166         lock_page(page);
167         wait_on_page_writeback(page);
168
169         nilfs_forget_buffer(bh);
170         still_dirty = PageDirty(page);
171         mapping = page->mapping;
172         unlock_page(page);
173         page_cache_release(page);
174
175         if (!still_dirty && mapping)
176                 invalidate_inode_pages2_range(mapping, index, index);
177 }
178
179 /**
180  * nilfs_btnode_prepare_change_key
181  *  prepare to move contents of the block for old key to one of new key.
182  *  the old buffer will not be removed, but might be reused for new buffer.
183  *  it might return -ENOMEM because of memory allocation errors,
184  *  and might return -EIO because of disk read errors.
185  */
186 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
187                                     struct nilfs_btnode_chkey_ctxt *ctxt)
188 {
189         struct buffer_head *obh, *nbh;
190         struct inode *inode = NILFS_BTNC_I(btnc);
191         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
192         int err;
193
194         if (oldkey == newkey)
195                 return 0;
196
197         obh = ctxt->bh;
198         ctxt->newbh = NULL;
199
200         if (inode->i_blkbits == PAGE_CACHE_SHIFT) {
201                 lock_page(obh->b_page);
202                 /*
203                  * We cannot call radix_tree_preload for the kernels older
204                  * than 2.6.23, because it is not exported for modules.
205                  */
206 retry:
207                 err = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
208                 if (err)
209                         goto failed_unlock;
210                 /* BUG_ON(oldkey != obh->b_page->index); */
211                 if (unlikely(oldkey != obh->b_page->index))
212                         NILFS_PAGE_BUG(obh->b_page,
213                                        "invalid oldkey %lld (newkey=%lld)",
214                                        (unsigned long long)oldkey,
215                                        (unsigned long long)newkey);
216
217                 spin_lock_irq(&btnc->tree_lock);
218                 err = radix_tree_insert(&btnc->page_tree, newkey, obh->b_page);
219                 spin_unlock_irq(&btnc->tree_lock);
220                 /*
221                  * Note: page->index will not change to newkey until
222                  * nilfs_btnode_commit_change_key() will be called.
223                  * To protect the page in intermediate state, the page lock
224                  * is held.
225                  */
226                 radix_tree_preload_end();
227                 if (!err)
228                         return 0;
229                 else if (err != -EEXIST)
230                         goto failed_unlock;
231
232                 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
233                 if (!err)
234                         goto retry;
235                 /* fallback to copy mode */
236                 unlock_page(obh->b_page);
237         }
238
239         nbh = nilfs_btnode_create_block(btnc, newkey);
240         if (!nbh)
241                 return -ENOMEM;
242
243         BUG_ON(nbh == obh);
244         ctxt->newbh = nbh;
245         return 0;
246
247  failed_unlock:
248         unlock_page(obh->b_page);
249         return err;
250 }
251
252 /**
253  * nilfs_btnode_commit_change_key
254  *  commit the change_key operation prepared by prepare_change_key().
255  */
256 void nilfs_btnode_commit_change_key(struct address_space *btnc,
257                                     struct nilfs_btnode_chkey_ctxt *ctxt)
258 {
259         struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
260         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
261         struct page *opage;
262
263         if (oldkey == newkey)
264                 return;
265
266         if (nbh == NULL) {      /* blocksize == pagesize */
267                 opage = obh->b_page;
268                 if (unlikely(oldkey != opage->index))
269                         NILFS_PAGE_BUG(opage,
270                                        "invalid oldkey %lld (newkey=%lld)",
271                                        (unsigned long long)oldkey,
272                                        (unsigned long long)newkey);
273                 nilfs_btnode_mark_dirty(obh);
274
275                 spin_lock_irq(&btnc->tree_lock);
276                 radix_tree_delete(&btnc->page_tree, oldkey);
277                 radix_tree_tag_set(&btnc->page_tree, newkey,
278                                    PAGECACHE_TAG_DIRTY);
279                 spin_unlock_irq(&btnc->tree_lock);
280
281                 opage->index = obh->b_blocknr = newkey;
282                 unlock_page(opage);
283         } else {
284                 nilfs_copy_buffer(nbh, obh);
285                 nilfs_btnode_mark_dirty(nbh);
286
287                 nbh->b_blocknr = newkey;
288                 ctxt->bh = nbh;
289                 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
290         }
291 }
292
293 /**
294  * nilfs_btnode_abort_change_key
295  *  abort the change_key operation prepared by prepare_change_key().
296  */
297 void nilfs_btnode_abort_change_key(struct address_space *btnc,
298                                    struct nilfs_btnode_chkey_ctxt *ctxt)
299 {
300         struct buffer_head *nbh = ctxt->newbh;
301         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
302
303         if (oldkey == newkey)
304                 return;
305
306         if (nbh == NULL) {      /* blocksize == pagesize */
307                 spin_lock_irq(&btnc->tree_lock);
308                 radix_tree_delete(&btnc->page_tree, newkey);
309                 spin_unlock_irq(&btnc->tree_lock);
310                 unlock_page(ctxt->bh->b_page);
311         } else
312                 brelse(nbh);
313 }