mm: merge populate and nopage into fault (fixes nonlinear)
[linux-2.6.git] / mm / truncate.c
1 /*
2  * mm/truncate.c - code for taking down pages from address_spaces
3  *
4  * Copyright (C) 2002, Linus Torvalds
5  *
6  * 10Sep2002    akpm@zip.com.au
7  *              Initial version.
8  */
9
10 #include <linux/kernel.h>
11 #include <linux/mm.h>
12 #include <linux/swap.h>
13 #include <linux/module.h>
14 #include <linux/pagemap.h>
15 #include <linux/highmem.h>
16 #include <linux/pagevec.h>
17 #include <linux/task_io_accounting_ops.h>
18 #include <linux/buffer_head.h>  /* grr. try_to_release_page,
19                                    do_invalidatepage */
20
21
22 /**
23  * do_invalidatepage - invalidate part of all of a page
24  * @page: the page which is affected
25  * @offset: the index of the truncation point
26  *
27  * do_invalidatepage() is called when all or part of the page has become
28  * invalidated by a truncate operation.
29  *
30  * do_invalidatepage() does not have to release all buffers, but it must
31  * ensure that no dirty buffer is left outside @offset and that no I/O
32  * is underway against any of the blocks which are outside the truncation
33  * point.  Because the caller is about to free (and possibly reuse) those
34  * blocks on-disk.
35  */
36 void do_invalidatepage(struct page *page, unsigned long offset)
37 {
38         void (*invalidatepage)(struct page *, unsigned long);
39         invalidatepage = page->mapping->a_ops->invalidatepage;
40 #ifdef CONFIG_BLOCK
41         if (!invalidatepage)
42                 invalidatepage = block_invalidatepage;
43 #endif
44         if (invalidatepage)
45                 (*invalidatepage)(page, offset);
46 }
47
48 static inline void truncate_partial_page(struct page *page, unsigned partial)
49 {
50         zero_user_page(page, partial, PAGE_CACHE_SIZE - partial, KM_USER0);
51         if (PagePrivate(page))
52                 do_invalidatepage(page, partial);
53 }
54
55 /*
56  * This cancels just the dirty bit on the kernel page itself, it
57  * does NOT actually remove dirty bits on any mmap's that may be
58  * around. It also leaves the page tagged dirty, so any sync
59  * activity will still find it on the dirty lists, and in particular,
60  * clear_page_dirty_for_io() will still look at the dirty bits in
61  * the VM.
62  *
63  * Doing this should *normally* only ever be done when a page
64  * is truncated, and is not actually mapped anywhere at all. However,
65  * fs/buffer.c does this when it notices that somebody has cleaned
66  * out all the buffers on a page without actually doing it through
67  * the VM. Can you say "ext3 is horribly ugly"? Tought you could.
68  */
69 void cancel_dirty_page(struct page *page, unsigned int account_size)
70 {
71         if (TestClearPageDirty(page)) {
72                 struct address_space *mapping = page->mapping;
73                 if (mapping && mapping_cap_account_dirty(mapping)) {
74                         dec_zone_page_state(page, NR_FILE_DIRTY);
75                         if (account_size)
76                                 task_io_account_cancelled_write(account_size);
77                 }
78         }
79 }
80 EXPORT_SYMBOL(cancel_dirty_page);
81
82 /*
83  * If truncate cannot remove the fs-private metadata from the page, the page
84  * becomes anonymous.  It will be left on the LRU and may even be mapped into
85  * user pagetables if we're racing with filemap_fault().
86  *
87  * We need to bale out if page->mapping is no longer equal to the original
88  * mapping.  This happens a) when the VM reclaimed the page while we waited on
89  * its lock, b) when a concurrent invalidate_mapping_pages got there first and
90  * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
91  */
92 static void
93 truncate_complete_page(struct address_space *mapping, struct page *page)
94 {
95         if (page->mapping != mapping)
96                 return;
97
98         cancel_dirty_page(page, PAGE_CACHE_SIZE);
99
100         if (PagePrivate(page))
101                 do_invalidatepage(page, 0);
102
103         remove_from_page_cache(page);
104         ClearPageUptodate(page);
105         ClearPageMappedToDisk(page);
106         page_cache_release(page);       /* pagecache ref */
107 }
108
109 /*
110  * This is for invalidate_mapping_pages().  That function can be called at
111  * any time, and is not supposed to throw away dirty pages.  But pages can
112  * be marked dirty at any time too, so use remove_mapping which safely
113  * discards clean, unused pages.
114  *
115  * Returns non-zero if the page was successfully invalidated.
116  */
117 static int
118 invalidate_complete_page(struct address_space *mapping, struct page *page)
119 {
120         int ret;
121
122         if (page->mapping != mapping)
123                 return 0;
124
125         if (PagePrivate(page) && !try_to_release_page(page, 0))
126                 return 0;
127
128         ret = remove_mapping(mapping, page);
129
130         return ret;
131 }
132
133 /**
134  * truncate_inode_pages - truncate range of pages specified by start and
135  * end byte offsets
136  * @mapping: mapping to truncate
137  * @lstart: offset from which to truncate
138  * @lend: offset to which to truncate
139  *
140  * Truncate the page cache, removing the pages that are between
141  * specified offsets (and zeroing out partial page
142  * (if lstart is not page aligned)).
143  *
144  * Truncate takes two passes - the first pass is nonblocking.  It will not
145  * block on page locks and it will not block on writeback.  The second pass
146  * will wait.  This is to prevent as much IO as possible in the affected region.
147  * The first pass will remove most pages, so the search cost of the second pass
148  * is low.
149  *
150  * When looking at page->index outside the page lock we need to be careful to
151  * copy it into a local to avoid races (it could change at any time).
152  *
153  * We pass down the cache-hot hint to the page freeing code.  Even if the
154  * mapping is large, it is probably the case that the final pages are the most
155  * recently touched, and freeing happens in ascending file offset order.
156  */
157 void truncate_inode_pages_range(struct address_space *mapping,
158                                 loff_t lstart, loff_t lend)
159 {
160         const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
161         pgoff_t end;
162         const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
163         struct pagevec pvec;
164         pgoff_t next;
165         int i;
166
167         if (mapping->nrpages == 0)
168                 return;
169
170         BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
171         end = (lend >> PAGE_CACHE_SHIFT);
172
173         pagevec_init(&pvec, 0);
174         next = start;
175         while (next <= end &&
176                pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
177                 for (i = 0; i < pagevec_count(&pvec); i++) {
178                         struct page *page = pvec.pages[i];
179                         pgoff_t page_index = page->index;
180
181                         if (page_index > end) {
182                                 next = page_index;
183                                 break;
184                         }
185
186                         if (page_index > next)
187                                 next = page_index;
188                         next++;
189                         if (TestSetPageLocked(page))
190                                 continue;
191                         if (PageWriteback(page)) {
192                                 unlock_page(page);
193                                 continue;
194                         }
195                         if (page_mapped(page)) {
196                                 unmap_mapping_range(mapping,
197                                   (loff_t)page_index<<PAGE_CACHE_SHIFT,
198                                   PAGE_CACHE_SIZE, 0);
199                         }
200                         truncate_complete_page(mapping, page);
201                         unlock_page(page);
202                 }
203                 pagevec_release(&pvec);
204                 cond_resched();
205         }
206
207         if (partial) {
208                 struct page *page = find_lock_page(mapping, start - 1);
209                 if (page) {
210                         wait_on_page_writeback(page);
211                         truncate_partial_page(page, partial);
212                         unlock_page(page);
213                         page_cache_release(page);
214                 }
215         }
216
217         next = start;
218         for ( ; ; ) {
219                 cond_resched();
220                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
221                         if (next == start)
222                                 break;
223                         next = start;
224                         continue;
225                 }
226                 if (pvec.pages[0]->index > end) {
227                         pagevec_release(&pvec);
228                         break;
229                 }
230                 for (i = 0; i < pagevec_count(&pvec); i++) {
231                         struct page *page = pvec.pages[i];
232
233                         if (page->index > end)
234                                 break;
235                         lock_page(page);
236                         wait_on_page_writeback(page);
237                         if (page_mapped(page)) {
238                                 unmap_mapping_range(mapping,
239                                   (loff_t)page->index<<PAGE_CACHE_SHIFT,
240                                   PAGE_CACHE_SIZE, 0);
241                         }
242                         if (page->index > next)
243                                 next = page->index;
244                         next++;
245                         truncate_complete_page(mapping, page);
246                         unlock_page(page);
247                 }
248                 pagevec_release(&pvec);
249         }
250 }
251 EXPORT_SYMBOL(truncate_inode_pages_range);
252
253 /**
254  * truncate_inode_pages - truncate *all* the pages from an offset
255  * @mapping: mapping to truncate
256  * @lstart: offset from which to truncate
257  *
258  * Called under (and serialised by) inode->i_mutex.
259  */
260 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
261 {
262         truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
263 }
264 EXPORT_SYMBOL(truncate_inode_pages);
265
266 unsigned long __invalidate_mapping_pages(struct address_space *mapping,
267                                 pgoff_t start, pgoff_t end, bool be_atomic)
268 {
269         struct pagevec pvec;
270         pgoff_t next = start;
271         unsigned long ret = 0;
272         int i;
273
274         pagevec_init(&pvec, 0);
275         while (next <= end &&
276                         pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
277                 for (i = 0; i < pagevec_count(&pvec); i++) {
278                         struct page *page = pvec.pages[i];
279                         pgoff_t index;
280                         int lock_failed;
281
282                         lock_failed = TestSetPageLocked(page);
283
284                         /*
285                          * We really shouldn't be looking at the ->index of an
286                          * unlocked page.  But we're not allowed to lock these
287                          * pages.  So we rely upon nobody altering the ->index
288                          * of this (pinned-by-us) page.
289                          */
290                         index = page->index;
291                         if (index > next)
292                                 next = index;
293                         next++;
294                         if (lock_failed)
295                                 continue;
296
297                         if (PageDirty(page) || PageWriteback(page))
298                                 goto unlock;
299                         if (page_mapped(page))
300                                 goto unlock;
301                         ret += invalidate_complete_page(mapping, page);
302 unlock:
303                         unlock_page(page);
304                         if (next > end)
305                                 break;
306                 }
307                 pagevec_release(&pvec);
308                 if (likely(!be_atomic))
309                         cond_resched();
310         }
311         return ret;
312 }
313
314 /**
315  * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
316  * @mapping: the address_space which holds the pages to invalidate
317  * @start: the offset 'from' which to invalidate
318  * @end: the offset 'to' which to invalidate (inclusive)
319  *
320  * This function only removes the unlocked pages, if you want to
321  * remove all the pages of one inode, you must call truncate_inode_pages.
322  *
323  * invalidate_mapping_pages() will not block on IO activity. It will not
324  * invalidate pages which are dirty, locked, under writeback or mapped into
325  * pagetables.
326  */
327 unsigned long invalidate_mapping_pages(struct address_space *mapping,
328                                 pgoff_t start, pgoff_t end)
329 {
330         return __invalidate_mapping_pages(mapping, start, end, false);
331 }
332 EXPORT_SYMBOL(invalidate_mapping_pages);
333
334 /*
335  * This is like invalidate_complete_page(), except it ignores the page's
336  * refcount.  We do this because invalidate_inode_pages2() needs stronger
337  * invalidation guarantees, and cannot afford to leave pages behind because
338  * shrink_page_list() has a temp ref on them, or because they're transiently
339  * sitting in the lru_cache_add() pagevecs.
340  */
341 static int
342 invalidate_complete_page2(struct address_space *mapping, struct page *page)
343 {
344         if (page->mapping != mapping)
345                 return 0;
346
347         if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
348                 return 0;
349
350         write_lock_irq(&mapping->tree_lock);
351         if (PageDirty(page))
352                 goto failed;
353
354         BUG_ON(PagePrivate(page));
355         __remove_from_page_cache(page);
356         write_unlock_irq(&mapping->tree_lock);
357         ClearPageUptodate(page);
358         page_cache_release(page);       /* pagecache ref */
359         return 1;
360 failed:
361         write_unlock_irq(&mapping->tree_lock);
362         return 0;
363 }
364
365 static int do_launder_page(struct address_space *mapping, struct page *page)
366 {
367         if (!PageDirty(page))
368                 return 0;
369         if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
370                 return 0;
371         return mapping->a_ops->launder_page(page);
372 }
373
374 /**
375  * invalidate_inode_pages2_range - remove range of pages from an address_space
376  * @mapping: the address_space
377  * @start: the page offset 'from' which to invalidate
378  * @end: the page offset 'to' which to invalidate (inclusive)
379  *
380  * Any pages which are found to be mapped into pagetables are unmapped prior to
381  * invalidation.
382  *
383  * Returns -EIO if any pages could not be invalidated.
384  */
385 int invalidate_inode_pages2_range(struct address_space *mapping,
386                                   pgoff_t start, pgoff_t end)
387 {
388         struct pagevec pvec;
389         pgoff_t next;
390         int i;
391         int ret = 0;
392         int did_range_unmap = 0;
393         int wrapped = 0;
394
395         pagevec_init(&pvec, 0);
396         next = start;
397         while (next <= end && !wrapped &&
398                 pagevec_lookup(&pvec, mapping, next,
399                         min(end - next, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
400                 for (i = 0; i < pagevec_count(&pvec); i++) {
401                         struct page *page = pvec.pages[i];
402                         pgoff_t page_index;
403
404                         lock_page(page);
405                         if (page->mapping != mapping) {
406                                 unlock_page(page);
407                                 continue;
408                         }
409                         page_index = page->index;
410                         next = page_index + 1;
411                         if (next == 0)
412                                 wrapped = 1;
413                         if (page_index > end) {
414                                 unlock_page(page);
415                                 break;
416                         }
417                         wait_on_page_writeback(page);
418                         if (page_mapped(page)) {
419                                 if (!did_range_unmap) {
420                                         /*
421                                          * Zap the rest of the file in one hit.
422                                          */
423                                         unmap_mapping_range(mapping,
424                                            (loff_t)page_index<<PAGE_CACHE_SHIFT,
425                                            (loff_t)(end - page_index + 1)
426                                                         << PAGE_CACHE_SHIFT,
427                                             0);
428                                         did_range_unmap = 1;
429                                 } else {
430                                         /*
431                                          * Just zap this page
432                                          */
433                                         unmap_mapping_range(mapping,
434                                           (loff_t)page_index<<PAGE_CACHE_SHIFT,
435                                           PAGE_CACHE_SIZE, 0);
436                                 }
437                         }
438                         BUG_ON(page_mapped(page));
439                         ret = do_launder_page(mapping, page);
440                         if (ret == 0 && !invalidate_complete_page2(mapping, page))
441                                 ret = -EIO;
442                         unlock_page(page);
443                 }
444                 pagevec_release(&pvec);
445                 cond_resched();
446         }
447         return ret;
448 }
449 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
450
451 /**
452  * invalidate_inode_pages2 - remove all pages from an address_space
453  * @mapping: the address_space
454  *
455  * Any pages which are found to be mapped into pagetables are unmapped prior to
456  * invalidation.
457  *
458  * Returns -EIO if any pages could not be invalidated.
459  */
460 int invalidate_inode_pages2(struct address_space *mapping)
461 {
462         return invalidate_inode_pages2_range(mapping, 0, -1);
463 }
464 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);