Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/kaber/nf-2.6
[linux-2.6] / fs / nilfs2 / page.c
1 /*
2  * page.c - buffer/page management specific to NILFS
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  * Written by Ryusuke Konishi <ryusuke@osrg.net>,
21  *            Seiji Kihara <kihara@osrg.net>.
22  */
23
24 #include <linux/pagemap.h>
25 #include <linux/writeback.h>
26 #include <linux/swap.h>
27 #include <linux/bitops.h>
28 #include <linux/page-flags.h>
29 #include <linux/list.h>
30 #include <linux/highmem.h>
31 #include <linux/pagevec.h>
32 #include "nilfs.h"
33 #include "page.h"
34 #include "mdt.h"
35
36
37 #define NILFS_BUFFER_INHERENT_BITS  \
38         ((1UL << BH_Uptodate) | (1UL << BH_Mapped) | (1UL << BH_NILFS_Node) | \
39          (1UL << BH_NILFS_Volatile) | (1UL << BH_NILFS_Allocated))
40
41 static struct buffer_head *
42 __nilfs_get_page_block(struct page *page, unsigned long block, pgoff_t index,
43                        int blkbits, unsigned long b_state)
44
45 {
46         unsigned long first_block;
47         struct buffer_head *bh;
48
49         if (!page_has_buffers(page))
50                 create_empty_buffers(page, 1 << blkbits, b_state);
51
52         first_block = (unsigned long)index << (PAGE_CACHE_SHIFT - blkbits);
53         bh = nilfs_page_get_nth_block(page, block - first_block);
54
55         touch_buffer(bh);
56         wait_on_buffer(bh);
57         return bh;
58 }
59
60 /*
61  * Since the page cache of B-tree node pages or data page cache of pseudo
62  * inodes does not have a valid mapping->host pointer, calling
63  * mark_buffer_dirty() for their buffers causes a NULL pointer dereference;
64  * it calls __mark_inode_dirty(NULL) through __set_page_dirty().
65  * To avoid this problem, the old style mark_buffer_dirty() is used instead.
66  */
67 void nilfs_mark_buffer_dirty(struct buffer_head *bh)
68 {
69         if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
70                 __set_page_dirty_nobuffers(bh->b_page);
71 }
72
73 struct buffer_head *nilfs_grab_buffer(struct inode *inode,
74                                       struct address_space *mapping,
75                                       unsigned long blkoff,
76                                       unsigned long b_state)
77 {
78         int blkbits = inode->i_blkbits;
79         pgoff_t index = blkoff >> (PAGE_CACHE_SHIFT - blkbits);
80         struct page *page, *opage;
81         struct buffer_head *bh, *obh;
82
83         page = grab_cache_page(mapping, index);
84         if (unlikely(!page))
85                 return NULL;
86
87         bh = __nilfs_get_page_block(page, blkoff, index, blkbits, b_state);
88         if (unlikely(!bh)) {
89                 unlock_page(page);
90                 page_cache_release(page);
91                 return NULL;
92         }
93         if (!buffer_uptodate(bh) && mapping->assoc_mapping != NULL) {
94                 /*
95                  * Shadow page cache uses assoc_mapping to point its original
96                  * page cache.  The following code tries the original cache
97                  * if the given cache is a shadow and it didn't hit.
98                  */
99                 opage = find_lock_page(mapping->assoc_mapping, index);
100                 if (!opage)
101                         return bh;
102
103                 obh = __nilfs_get_page_block(opage, blkoff, index, blkbits,
104                                              b_state);
105                 if (buffer_uptodate(obh)) {
106                         nilfs_copy_buffer(bh, obh);
107                         if (buffer_dirty(obh)) {
108                                 nilfs_mark_buffer_dirty(bh);
109                                 if (!buffer_nilfs_node(bh) && NILFS_MDT(inode))
110                                         nilfs_mdt_mark_dirty(inode);
111                         }
112                 }
113                 brelse(obh);
114                 unlock_page(opage);
115                 page_cache_release(opage);
116         }
117         return bh;
118 }
119
120 /**
121  * nilfs_forget_buffer - discard dirty state
122  * @inode: owner inode of the buffer
123  * @bh: buffer head of the buffer to be discarded
124  */
125 void nilfs_forget_buffer(struct buffer_head *bh)
126 {
127         struct page *page = bh->b_page;
128
129         lock_buffer(bh);
130         clear_buffer_nilfs_volatile(bh);
131         if (test_clear_buffer_dirty(bh) && nilfs_page_buffers_clean(page))
132                 __nilfs_clear_page_dirty(page);
133
134         clear_buffer_uptodate(bh);
135         clear_buffer_mapped(bh);
136         bh->b_blocknr = -1;
137         ClearPageUptodate(page);
138         ClearPageMappedToDisk(page);
139         unlock_buffer(bh);
140         brelse(bh);
141 }
142
143 /**
144  * nilfs_copy_buffer -- copy buffer data and flags
145  * @dbh: destination buffer
146  * @sbh: source buffer
147  */
148 void nilfs_copy_buffer(struct buffer_head *dbh, struct buffer_head *sbh)
149 {
150         void *kaddr0, *kaddr1;
151         unsigned long bits;
152         struct page *spage = sbh->b_page, *dpage = dbh->b_page;
153         struct buffer_head *bh;
154
155         kaddr0 = kmap_atomic(spage, KM_USER0);
156         kaddr1 = kmap_atomic(dpage, KM_USER1);
157         memcpy(kaddr1 + bh_offset(dbh), kaddr0 + bh_offset(sbh), sbh->b_size);
158         kunmap_atomic(kaddr1, KM_USER1);
159         kunmap_atomic(kaddr0, KM_USER0);
160
161         dbh->b_state = sbh->b_state & NILFS_BUFFER_INHERENT_BITS;
162         dbh->b_blocknr = sbh->b_blocknr;
163         dbh->b_bdev = sbh->b_bdev;
164
165         bh = dbh;
166         bits = sbh->b_state & ((1UL << BH_Uptodate) | (1UL << BH_Mapped));
167         while ((bh = bh->b_this_page) != dbh) {
168                 lock_buffer(bh);
169                 bits &= bh->b_state;
170                 unlock_buffer(bh);
171         }
172         if (bits & (1UL << BH_Uptodate))
173                 SetPageUptodate(dpage);
174         else
175                 ClearPageUptodate(dpage);
176         if (bits & (1UL << BH_Mapped))
177                 SetPageMappedToDisk(dpage);
178         else
179                 ClearPageMappedToDisk(dpage);
180 }
181
182 /**
183  * nilfs_page_buffers_clean - check if a page has dirty buffers or not.
184  * @page: page to be checked
185  *
186  * nilfs_page_buffers_clean() returns zero if the page has dirty buffers.
187  * Otherwise, it returns non-zero value.
188  */
189 int nilfs_page_buffers_clean(struct page *page)
190 {
191         struct buffer_head *bh, *head;
192
193         bh = head = page_buffers(page);
194         do {
195                 if (buffer_dirty(bh))
196                         return 0;
197                 bh = bh->b_this_page;
198         } while (bh != head);
199         return 1;
200 }
201
202 void nilfs_page_bug(struct page *page)
203 {
204         struct address_space *m;
205         unsigned long ino = 0;
206
207         if (unlikely(!page)) {
208                 printk(KERN_CRIT "NILFS_PAGE_BUG(NULL)\n");
209                 return;
210         }
211
212         m = page->mapping;
213         if (m) {
214                 struct inode *inode = NILFS_AS_I(m);
215                 if (inode != NULL)
216                         ino = inode->i_ino;
217         }
218         printk(KERN_CRIT "NILFS_PAGE_BUG(%p): cnt=%d index#=%llu flags=0x%lx "
219                "mapping=%p ino=%lu\n",
220                page, atomic_read(&page->_count),
221                (unsigned long long)page->index, page->flags, m, ino);
222
223         if (page_has_buffers(page)) {
224                 struct buffer_head *bh, *head;
225                 int i = 0;
226
227                 bh = head = page_buffers(page);
228                 do {
229                         printk(KERN_CRIT
230                                " BH[%d] %p: cnt=%d block#=%llu state=0x%lx\n",
231                                i++, bh, atomic_read(&bh->b_count),
232                                (unsigned long long)bh->b_blocknr, bh->b_state);
233                         bh = bh->b_this_page;
234                 } while (bh != head);
235         }
236 }
237
238 /**
239  * nilfs_alloc_private_page - allocate a private page with buffer heads
240  *
241  * Return Value: On success, a pointer to the allocated page is returned.
242  * On error, NULL is returned.
243  */
244 struct page *nilfs_alloc_private_page(struct block_device *bdev, int size,
245                                       unsigned long state)
246 {
247         struct buffer_head *bh, *head, *tail;
248         struct page *page;
249
250         page = alloc_page(GFP_NOFS); /* page_count of the returned page is 1 */
251         if (unlikely(!page))
252                 return NULL;
253
254         lock_page(page);
255         head = alloc_page_buffers(page, size, 0);
256         if (unlikely(!head)) {
257                 unlock_page(page);
258                 __free_page(page);
259                 return NULL;
260         }
261
262         bh = head;
263         do {
264                 bh->b_state = (1UL << BH_NILFS_Allocated) | state;
265                 tail = bh;
266                 bh->b_bdev = bdev;
267                 bh = bh->b_this_page;
268         } while (bh);
269
270         tail->b_this_page = head;
271         attach_page_buffers(page, head);
272
273         return page;
274 }
275
276 void nilfs_free_private_page(struct page *page)
277 {
278         BUG_ON(!PageLocked(page));
279         BUG_ON(page->mapping);
280
281         if (page_has_buffers(page) && !try_to_free_buffers(page))
282                 NILFS_PAGE_BUG(page, "failed to free page");
283
284         unlock_page(page);
285         __free_page(page);
286 }
287
288 /**
289  * nilfs_copy_page -- copy the page with buffers
290  * @dst: destination page
291  * @src: source page
292  * @copy_dirty: flag whether to copy dirty states on the page's buffer heads.
293  *
294  * This fuction is for both data pages and btnode pages.  The dirty flag
295  * should be treated by caller.  The page must not be under i/o.
296  * Both src and dst page must be locked
297  */
298 static void nilfs_copy_page(struct page *dst, struct page *src, int copy_dirty)
299 {
300         struct buffer_head *dbh, *dbufs, *sbh, *sbufs;
301         unsigned long mask = NILFS_BUFFER_INHERENT_BITS;
302
303         BUG_ON(PageWriteback(dst));
304
305         sbh = sbufs = page_buffers(src);
306         if (!page_has_buffers(dst))
307                 create_empty_buffers(dst, sbh->b_size, 0);
308
309         if (copy_dirty)
310                 mask |= (1UL << BH_Dirty);
311
312         dbh = dbufs = page_buffers(dst);
313         do {
314                 lock_buffer(sbh);
315                 lock_buffer(dbh);
316                 dbh->b_state = sbh->b_state & mask;
317                 dbh->b_blocknr = sbh->b_blocknr;
318                 dbh->b_bdev = sbh->b_bdev;
319                 sbh = sbh->b_this_page;
320                 dbh = dbh->b_this_page;
321         } while (dbh != dbufs);
322
323         copy_highpage(dst, src);
324
325         if (PageUptodate(src) && !PageUptodate(dst))
326                 SetPageUptodate(dst);
327         else if (!PageUptodate(src) && PageUptodate(dst))
328                 ClearPageUptodate(dst);
329         if (PageMappedToDisk(src) && !PageMappedToDisk(dst))
330                 SetPageMappedToDisk(dst);
331         else if (!PageMappedToDisk(src) && PageMappedToDisk(dst))
332                 ClearPageMappedToDisk(dst);
333
334         do {
335                 unlock_buffer(sbh);
336                 unlock_buffer(dbh);
337                 sbh = sbh->b_this_page;
338                 dbh = dbh->b_this_page;
339         } while (dbh != dbufs);
340 }
341
342 int nilfs_copy_dirty_pages(struct address_space *dmap,
343                            struct address_space *smap)
344 {
345         struct pagevec pvec;
346         unsigned int i;
347         pgoff_t index = 0;
348         int err = 0;
349
350         pagevec_init(&pvec, 0);
351 repeat:
352         if (!pagevec_lookup_tag(&pvec, smap, &index, PAGECACHE_TAG_DIRTY,
353                                 PAGEVEC_SIZE))
354                 return 0;
355
356         for (i = 0; i < pagevec_count(&pvec); i++) {
357                 struct page *page = pvec.pages[i], *dpage;
358
359                 lock_page(page);
360                 if (unlikely(!PageDirty(page)))
361                         NILFS_PAGE_BUG(page, "inconsistent dirty state");
362
363                 dpage = grab_cache_page(dmap, page->index);
364                 if (unlikely(!dpage)) {
365                         /* No empty page is added to the page cache */
366                         err = -ENOMEM;
367                         unlock_page(page);
368                         break;
369                 }
370                 if (unlikely(!page_has_buffers(page)))
371                         NILFS_PAGE_BUG(page,
372                                        "found empty page in dat page cache");
373
374                 nilfs_copy_page(dpage, page, 1);
375                 __set_page_dirty_nobuffers(dpage);
376
377                 unlock_page(dpage);
378                 page_cache_release(dpage);
379                 unlock_page(page);
380         }
381         pagevec_release(&pvec);
382         cond_resched();
383
384         if (likely(!err))
385                 goto repeat;
386         return err;
387 }
388
389 /**
390  * nilfs_copy_back_pages -- copy back pages to orignal cache from shadow cache
391  * @dmap: destination page cache
392  * @smap: source page cache
393  *
394  * No pages must no be added to the cache during this process.
395  * This must be ensured by the caller.
396  */
397 void nilfs_copy_back_pages(struct address_space *dmap,
398                            struct address_space *smap)
399 {
400         struct pagevec pvec;
401         unsigned int i, n;
402         pgoff_t index = 0;
403         int err;
404
405         pagevec_init(&pvec, 0);
406 repeat:
407         n = pagevec_lookup(&pvec, smap, index, PAGEVEC_SIZE);
408         if (!n)
409                 return;
410         index = pvec.pages[n - 1]->index + 1;
411
412         for (i = 0; i < pagevec_count(&pvec); i++) {
413                 struct page *page = pvec.pages[i], *dpage;
414                 pgoff_t offset = page->index;
415
416                 lock_page(page);
417                 dpage = find_lock_page(dmap, offset);
418                 if (dpage) {
419                         /* override existing page on the destination cache */
420                         WARN_ON(PageDirty(dpage));
421                         nilfs_copy_page(dpage, page, 0);
422                         unlock_page(dpage);
423                         page_cache_release(dpage);
424                 } else {
425                         struct page *page2;
426
427                         /* move the page to the destination cache */
428                         spin_lock_irq(&smap->tree_lock);
429                         page2 = radix_tree_delete(&smap->page_tree, offset);
430                         WARN_ON(page2 != page);
431
432                         smap->nrpages--;
433                         spin_unlock_irq(&smap->tree_lock);
434
435                         spin_lock_irq(&dmap->tree_lock);
436                         err = radix_tree_insert(&dmap->page_tree, offset, page);
437                         if (unlikely(err < 0)) {
438                                 WARN_ON(err == -EEXIST);
439                                 page->mapping = NULL;
440                                 page_cache_release(page); /* for cache */
441                         } else {
442                                 page->mapping = dmap;
443                                 dmap->nrpages++;
444                                 if (PageDirty(page))
445                                         radix_tree_tag_set(&dmap->page_tree,
446                                                            offset,
447                                                            PAGECACHE_TAG_DIRTY);
448                         }
449                         spin_unlock_irq(&dmap->tree_lock);
450                 }
451                 unlock_page(page);
452         }
453         pagevec_release(&pvec);
454         cond_resched();
455
456         goto repeat;
457 }
458
459 void nilfs_clear_dirty_pages(struct address_space *mapping)
460 {
461         struct pagevec pvec;
462         unsigned int i;
463         pgoff_t index = 0;
464
465         pagevec_init(&pvec, 0);
466
467         while (pagevec_lookup_tag(&pvec, mapping, &index, PAGECACHE_TAG_DIRTY,
468                                   PAGEVEC_SIZE)) {
469                 for (i = 0; i < pagevec_count(&pvec); i++) {
470                         struct page *page = pvec.pages[i];
471                         struct buffer_head *bh, *head;
472
473                         lock_page(page);
474                         ClearPageUptodate(page);
475                         ClearPageMappedToDisk(page);
476                         bh = head = page_buffers(page);
477                         do {
478                                 lock_buffer(bh);
479                                 clear_buffer_dirty(bh);
480                                 clear_buffer_nilfs_volatile(bh);
481                                 clear_buffer_uptodate(bh);
482                                 clear_buffer_mapped(bh);
483                                 unlock_buffer(bh);
484                                 bh = bh->b_this_page;
485                         } while (bh != head);
486
487                         __nilfs_clear_page_dirty(page);
488                         unlock_page(page);
489                 }
490                 pagevec_release(&pvec);
491                 cond_resched();
492         }
493 }
494
495 unsigned nilfs_page_count_clean_buffers(struct page *page,
496                                         unsigned from, unsigned to)
497 {
498         unsigned block_start, block_end;
499         struct buffer_head *bh, *head;
500         unsigned nc = 0;
501
502         for (bh = head = page_buffers(page), block_start = 0;
503              bh != head || !block_start;
504              block_start = block_end, bh = bh->b_this_page) {
505                 block_end = block_start + bh->b_size;
506                 if (block_end > from && block_start < to && !buffer_dirty(bh))
507                         nc++;
508         }
509         return nc;
510 }
511
512 /*
513  * NILFS2 needs clear_page_dirty() in the following two cases:
514  *
515  * 1) For B-tree node pages and data pages of the dat/gcdat, NILFS2 clears
516  *    page dirty flags when it copies back pages from the shadow cache
517  *    (gcdat->{i_mapping,i_btnode_cache}) to its original cache
518  *    (dat->{i_mapping,i_btnode_cache}).
519  *
520  * 2) Some B-tree operations like insertion or deletion may dispose buffers
521  *    in dirty state, and this needs to cancel the dirty state of their pages.
522  */
523 int __nilfs_clear_page_dirty(struct page *page)
524 {
525         struct address_space *mapping = page->mapping;
526
527         if (mapping) {
528                 spin_lock_irq(&mapping->tree_lock);
529                 if (test_bit(PG_dirty, &page->flags)) {
530                         radix_tree_tag_clear(&mapping->page_tree,
531                                              page_index(page),
532                                              PAGECACHE_TAG_DIRTY);
533                         spin_unlock_irq(&mapping->tree_lock);
534                         return clear_page_dirty_for_io(page);
535                 }
536                 spin_unlock_irq(&mapping->tree_lock);
537                 return 0;
538         }
539         return TestClearPageDirty(page);
540 }