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