Merge branch 'x86/pebs' into x86/unify-cpu-detect
[linux-2.6] / 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/backing-dev.h>
12 #include <linux/mm.h>
13 #include <linux/swap.h>
14 #include <linux/module.h>
15 #include <linux/pagemap.h>
16 #include <linux/highmem.h>
17 #include <linux/pagevec.h>
18 #include <linux/task_io_accounting_ops.h>
19 #include <linux/buffer_head.h>  /* grr. try_to_release_page,
20                                    do_invalidatepage */
21
22
23 /**
24  * do_invalidatepage - invalidate part or all of a page
25  * @page: the page which is affected
26  * @offset: the index of the truncation point
27  *
28  * do_invalidatepage() is called when all or part of the page has become
29  * invalidated by a truncate operation.
30  *
31  * do_invalidatepage() does not have to release all buffers, but it must
32  * ensure that no dirty buffer is left outside @offset and that no I/O
33  * is underway against any of the blocks which are outside the truncation
34  * point.  Because the caller is about to free (and possibly reuse) those
35  * blocks on-disk.
36  */
37 void do_invalidatepage(struct page *page, unsigned long offset)
38 {
39         void (*invalidatepage)(struct page *, unsigned long);
40         invalidatepage = page->mapping->a_ops->invalidatepage;
41 #ifdef CONFIG_BLOCK
42         if (!invalidatepage)
43                 invalidatepage = block_invalidatepage;
44 #endif
45         if (invalidatepage)
46                 (*invalidatepage)(page, offset);
47 }
48
49 static inline void truncate_partial_page(struct page *page, unsigned partial)
50 {
51         zero_user_segment(page, partial, PAGE_CACHE_SIZE);
52         if (PagePrivate(page))
53                 do_invalidatepage(page, partial);
54 }
55
56 /*
57  * This cancels just the dirty bit on the kernel page itself, it
58  * does NOT actually remove dirty bits on any mmap's that may be
59  * around. It also leaves the page tagged dirty, so any sync
60  * activity will still find it on the dirty lists, and in particular,
61  * clear_page_dirty_for_io() will still look at the dirty bits in
62  * the VM.
63  *
64  * Doing this should *normally* only ever be done when a page
65  * is truncated, and is not actually mapped anywhere at all. However,
66  * fs/buffer.c does this when it notices that somebody has cleaned
67  * out all the buffers on a page without actually doing it through
68  * the VM. Can you say "ext3 is horribly ugly"? Tought you could.
69  */
70 void cancel_dirty_page(struct page *page, unsigned int account_size)
71 {
72         if (TestClearPageDirty(page)) {
73                 struct address_space *mapping = page->mapping;
74                 if (mapping && mapping_cap_account_dirty(mapping)) {
75                         dec_zone_page_state(page, NR_FILE_DIRTY);
76                         dec_bdi_stat(mapping->backing_dev_info,
77                                         BDI_RECLAIMABLE);
78                         if (account_size)
79                                 task_io_account_cancelled_write(account_size);
80                 }
81         }
82 }
83 EXPORT_SYMBOL(cancel_dirty_page);
84
85 /*
86  * If truncate cannot remove the fs-private metadata from the page, the page
87  * becomes orphaned.  It will be left on the LRU and may even be mapped into
88  * user pagetables if we're racing with filemap_fault().
89  *
90  * We need to bale out if page->mapping is no longer equal to the original
91  * mapping.  This happens a) when the VM reclaimed the page while we waited on
92  * its lock, b) when a concurrent invalidate_mapping_pages got there first and
93  * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
94  */
95 static void
96 truncate_complete_page(struct address_space *mapping, struct page *page)
97 {
98         if (page->mapping != mapping)
99                 return;
100
101         if (PagePrivate(page))
102                 do_invalidatepage(page, 0);
103
104         cancel_dirty_page(page, PAGE_CACHE_SIZE);
105
106         remove_from_page_cache(page);
107         ClearPageMappedToDisk(page);
108         page_cache_release(page);       /* pagecache ref */
109 }
110
111 /*
112  * This is for invalidate_mapping_pages().  That function can be called at
113  * any time, and is not supposed to throw away dirty pages.  But pages can
114  * be marked dirty at any time too, so use remove_mapping which safely
115  * discards clean, unused pages.
116  *
117  * Returns non-zero if the page was successfully invalidated.
118  */
119 static int
120 invalidate_complete_page(struct address_space *mapping, struct page *page)
121 {
122         int ret;
123
124         if (page->mapping != mapping)
125                 return 0;
126
127         if (PagePrivate(page) && !try_to_release_page(page, 0))
128                 return 0;
129
130         ret = remove_mapping(mapping, page);
131
132         return ret;
133 }
134
135 /**
136  * truncate_inode_pages - truncate range of pages specified by start & end byte offsets
137  * @mapping: mapping to truncate
138  * @lstart: offset from which to truncate
139  * @lend: offset to which to truncate
140  *
141  * Truncate the page cache, removing the pages that are between
142  * specified offsets (and zeroing out partial page
143  * (if lstart is not page aligned)).
144  *
145  * Truncate takes two passes - the first pass is nonblocking.  It will not
146  * block on page locks and it will not block on writeback.  The second pass
147  * will wait.  This is to prevent as much IO as possible in the affected region.
148  * The first pass will remove most pages, so the search cost of the second pass
149  * is low.
150  *
151  * When looking at page->index outside the page lock we need to be careful to
152  * copy it into a local to avoid races (it could change at any time).
153  *
154  * We pass down the cache-hot hint to the page freeing code.  Even if the
155  * mapping is large, it is probably the case that the final pages are the most
156  * recently touched, and freeing happens in ascending file offset order.
157  */
158 void truncate_inode_pages_range(struct address_space *mapping,
159                                 loff_t lstart, loff_t lend)
160 {
161         const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
162         pgoff_t end;
163         const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
164         struct pagevec pvec;
165         pgoff_t next;
166         int i;
167
168         if (mapping->nrpages == 0)
169                 return;
170
171         BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
172         end = (lend >> PAGE_CACHE_SHIFT);
173
174         pagevec_init(&pvec, 0);
175         next = start;
176         while (next <= end &&
177                pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
178                 for (i = 0; i < pagevec_count(&pvec); i++) {
179                         struct page *page = pvec.pages[i];
180                         pgoff_t page_index = page->index;
181
182                         if (page_index > end) {
183                                 next = page_index;
184                                 break;
185                         }
186
187                         if (page_index > next)
188                                 next = page_index;
189                         next++;
190                         if (!trylock_page(page))
191                                 continue;
192                         if (PageWriteback(page)) {
193                                 unlock_page(page);
194                                 continue;
195                         }
196                         if (page_mapped(page)) {
197                                 unmap_mapping_range(mapping,
198                                   (loff_t)page_index<<PAGE_CACHE_SHIFT,
199                                   PAGE_CACHE_SIZE, 0);
200                         }
201                         truncate_complete_page(mapping, page);
202                         unlock_page(page);
203                 }
204                 pagevec_release(&pvec);
205                 cond_resched();
206         }
207
208         if (partial) {
209                 struct page *page = find_lock_page(mapping, start - 1);
210                 if (page) {
211                         wait_on_page_writeback(page);
212                         truncate_partial_page(page, partial);
213                         unlock_page(page);
214                         page_cache_release(page);
215                 }
216         }
217
218         next = start;
219         for ( ; ; ) {
220                 cond_resched();
221                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
222                         if (next == start)
223                                 break;
224                         next = start;
225                         continue;
226                 }
227                 if (pvec.pages[0]->index > end) {
228                         pagevec_release(&pvec);
229                         break;
230                 }
231                 for (i = 0; i < pagevec_count(&pvec); i++) {
232                         struct page *page = pvec.pages[i];
233
234                         if (page->index > end)
235                                 break;
236                         lock_page(page);
237                         wait_on_page_writeback(page);
238                         if (page_mapped(page)) {
239                                 unmap_mapping_range(mapping,
240                                   (loff_t)page->index<<PAGE_CACHE_SHIFT,
241                                   PAGE_CACHE_SIZE, 0);
242                         }
243                         if (page->index > next)
244                                 next = page->index;
245                         next++;
246                         truncate_complete_page(mapping, page);
247                         unlock_page(page);
248                 }
249                 pagevec_release(&pvec);
250         }
251 }
252 EXPORT_SYMBOL(truncate_inode_pages_range);
253
254 /**
255  * truncate_inode_pages - truncate *all* the pages from an offset
256  * @mapping: mapping to truncate
257  * @lstart: offset from which to truncate
258  *
259  * Called under (and serialised by) inode->i_mutex.
260  */
261 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
262 {
263         truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
264 }
265 EXPORT_SYMBOL(truncate_inode_pages);
266
267 unsigned long __invalidate_mapping_pages(struct address_space *mapping,
268                                 pgoff_t start, pgoff_t end, bool be_atomic)
269 {
270         struct pagevec pvec;
271         pgoff_t next = start;
272         unsigned long ret = 0;
273         int i;
274
275         pagevec_init(&pvec, 0);
276         while (next <= end &&
277                         pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
278                 for (i = 0; i < pagevec_count(&pvec); i++) {
279                         struct page *page = pvec.pages[i];
280                         pgoff_t index;
281                         int lock_failed;
282
283                         lock_failed = !trylock_page(page);
284
285                         /*
286                          * We really shouldn't be looking at the ->index of an
287                          * unlocked page.  But we're not allowed to lock these
288                          * pages.  So we rely upon nobody altering the ->index
289                          * of this (pinned-by-us) page.
290                          */
291                         index = page->index;
292                         if (index > next)
293                                 next = index;
294                         next++;
295                         if (lock_failed)
296                                 continue;
297
298                         if (PageDirty(page) || PageWriteback(page))
299                                 goto unlock;
300                         if (page_mapped(page))
301                                 goto unlock;
302                         ret += invalidate_complete_page(mapping, page);
303 unlock:
304                         unlock_page(page);
305                         if (next > end)
306                                 break;
307                 }
308                 pagevec_release(&pvec);
309                 if (likely(!be_atomic))
310                         cond_resched();
311         }
312         return ret;
313 }
314
315 /**
316  * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
317  * @mapping: the address_space which holds the pages to invalidate
318  * @start: the offset 'from' which to invalidate
319  * @end: the offset 'to' which to invalidate (inclusive)
320  *
321  * This function only removes the unlocked pages, if you want to
322  * remove all the pages of one inode, you must call truncate_inode_pages.
323  *
324  * invalidate_mapping_pages() will not block on IO activity. It will not
325  * invalidate pages which are dirty, locked, under writeback or mapped into
326  * pagetables.
327  */
328 unsigned long invalidate_mapping_pages(struct address_space *mapping,
329                                 pgoff_t start, pgoff_t end)
330 {
331         return __invalidate_mapping_pages(mapping, start, end, false);
332 }
333 EXPORT_SYMBOL(invalidate_mapping_pages);
334
335 /*
336  * This is like invalidate_complete_page(), except it ignores the page's
337  * refcount.  We do this because invalidate_inode_pages2() needs stronger
338  * invalidation guarantees, and cannot afford to leave pages behind because
339  * shrink_page_list() has a temp ref on them, or because they're transiently
340  * sitting in the lru_cache_add() pagevecs.
341  */
342 static int
343 invalidate_complete_page2(struct address_space *mapping, struct page *page)
344 {
345         if (page->mapping != mapping)
346                 return 0;
347
348         if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
349                 return 0;
350
351         spin_lock_irq(&mapping->tree_lock);
352         if (PageDirty(page))
353                 goto failed;
354
355         BUG_ON(PagePrivate(page));
356         __remove_from_page_cache(page);
357         spin_unlock_irq(&mapping->tree_lock);
358         page_cache_release(page);       /* pagecache ref */
359         return 1;
360 failed:
361         spin_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 -EBUSY 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 ret2 = 0;
393         int did_range_unmap = 0;
394         int wrapped = 0;
395
396         pagevec_init(&pvec, 0);
397         next = start;
398         while (next <= end && !wrapped &&
399                 pagevec_lookup(&pvec, mapping, next,
400                         min(end - next, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
401                 for (i = 0; i < pagevec_count(&pvec); i++) {
402                         struct page *page = pvec.pages[i];
403                         pgoff_t page_index;
404
405                         lock_page(page);
406                         if (page->mapping != mapping) {
407                                 unlock_page(page);
408                                 continue;
409                         }
410                         page_index = page->index;
411                         next = page_index + 1;
412                         if (next == 0)
413                                 wrapped = 1;
414                         if (page_index > end) {
415                                 unlock_page(page);
416                                 break;
417                         }
418                         wait_on_page_writeback(page);
419                         if (page_mapped(page)) {
420                                 if (!did_range_unmap) {
421                                         /*
422                                          * Zap the rest of the file in one hit.
423                                          */
424                                         unmap_mapping_range(mapping,
425                                            (loff_t)page_index<<PAGE_CACHE_SHIFT,
426                                            (loff_t)(end - page_index + 1)
427                                                         << PAGE_CACHE_SHIFT,
428                                             0);
429                                         did_range_unmap = 1;
430                                 } else {
431                                         /*
432                                          * Just zap this page
433                                          */
434                                         unmap_mapping_range(mapping,
435                                           (loff_t)page_index<<PAGE_CACHE_SHIFT,
436                                           PAGE_CACHE_SIZE, 0);
437                                 }
438                         }
439                         BUG_ON(page_mapped(page));
440                         ret2 = do_launder_page(mapping, page);
441                         if (ret2 == 0) {
442                                 if (!invalidate_complete_page2(mapping, page))
443                                         ret2 = -EBUSY;
444                         }
445                         if (ret2 < 0)
446                                 ret = ret2;
447                         unlock_page(page);
448                 }
449                 pagevec_release(&pvec);
450                 cond_resched();
451         }
452         return ret;
453 }
454 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
455
456 /**
457  * invalidate_inode_pages2 - remove all pages from an address_space
458  * @mapping: the address_space
459  *
460  * Any pages which are found to be mapped into pagetables are unmapped prior to
461  * invalidation.
462  *
463  * Returns -EIO if any pages could not be invalidated.
464  */
465 int invalidate_inode_pages2(struct address_space *mapping)
466 {
467         return invalidate_inode_pages2_range(mapping, 0, -1);
468 }
469 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);