[PATCH] mutex subsystem, semaphore to mutex: VFS, ->i_sem
[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/module.h>
13 #include <linux/pagemap.h>
14 #include <linux/pagevec.h>
15 #include <linux/buffer_head.h>  /* grr. try_to_release_page,
16                                    do_invalidatepage */
17
18
19 static inline void truncate_partial_page(struct page *page, unsigned partial)
20 {
21         memclear_highpage_flush(page, partial, PAGE_CACHE_SIZE-partial);
22         if (PagePrivate(page))
23                 do_invalidatepage(page, partial);
24 }
25
26 /*
27  * If truncate cannot remove the fs-private metadata from the page, the page
28  * becomes anonymous.  It will be left on the LRU and may even be mapped into
29  * user pagetables if we're racing with filemap_nopage().
30  *
31  * We need to bale out if page->mapping is no longer equal to the original
32  * mapping.  This happens a) when the VM reclaimed the page while we waited on
33  * its lock, b) when a concurrent invalidate_inode_pages got there first and
34  * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
35  */
36 static void
37 truncate_complete_page(struct address_space *mapping, struct page *page)
38 {
39         if (page->mapping != mapping)
40                 return;
41
42         if (PagePrivate(page))
43                 do_invalidatepage(page, 0);
44
45         clear_page_dirty(page);
46         ClearPageUptodate(page);
47         ClearPageMappedToDisk(page);
48         remove_from_page_cache(page);
49         page_cache_release(page);       /* pagecache ref */
50 }
51
52 /*
53  * This is for invalidate_inode_pages().  That function can be called at
54  * any time, and is not supposed to throw away dirty pages.  But pages can
55  * be marked dirty at any time too.  So we re-check the dirtiness inside
56  * ->tree_lock.  That provides exclusion against the __set_page_dirty
57  * functions.
58  *
59  * Returns non-zero if the page was successfully invalidated.
60  */
61 static int
62 invalidate_complete_page(struct address_space *mapping, struct page *page)
63 {
64         if (page->mapping != mapping)
65                 return 0;
66
67         if (PagePrivate(page) && !try_to_release_page(page, 0))
68                 return 0;
69
70         write_lock_irq(&mapping->tree_lock);
71         if (PageDirty(page)) {
72                 write_unlock_irq(&mapping->tree_lock);
73                 return 0;
74         }
75
76         BUG_ON(PagePrivate(page));
77         __remove_from_page_cache(page);
78         write_unlock_irq(&mapping->tree_lock);
79         ClearPageUptodate(page);
80         page_cache_release(page);       /* pagecache ref */
81         return 1;
82 }
83
84 /**
85  * truncate_inode_pages - truncate range of pages specified by start and
86  * end byte offsets
87  * @mapping: mapping to truncate
88  * @lstart: offset from which to truncate
89  * @lend: offset to which to truncate
90  *
91  * Truncate the page cache, removing the pages that are between
92  * specified offsets (and zeroing out partial page
93  * (if lstart is not page aligned)).
94  *
95  * Truncate takes two passes - the first pass is nonblocking.  It will not
96  * block on page locks and it will not block on writeback.  The second pass
97  * will wait.  This is to prevent as much IO as possible in the affected region.
98  * The first pass will remove most pages, so the search cost of the second pass
99  * is low.
100  *
101  * When looking at page->index outside the page lock we need to be careful to
102  * copy it into a local to avoid races (it could change at any time).
103  *
104  * We pass down the cache-hot hint to the page freeing code.  Even if the
105  * mapping is large, it is probably the case that the final pages are the most
106  * recently touched, and freeing happens in ascending file offset order.
107  */
108 void truncate_inode_pages_range(struct address_space *mapping,
109                                 loff_t lstart, loff_t lend)
110 {
111         const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
112         pgoff_t end;
113         const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
114         struct pagevec pvec;
115         pgoff_t next;
116         int i;
117
118         if (mapping->nrpages == 0)
119                 return;
120
121         BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
122         end = (lend >> PAGE_CACHE_SHIFT);
123
124         pagevec_init(&pvec, 0);
125         next = start;
126         while (next <= end &&
127                pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
128                 for (i = 0; i < pagevec_count(&pvec); i++) {
129                         struct page *page = pvec.pages[i];
130                         pgoff_t page_index = page->index;
131
132                         if (page_index > end) {
133                                 next = page_index;
134                                 break;
135                         }
136
137                         if (page_index > next)
138                                 next = page_index;
139                         next++;
140                         if (TestSetPageLocked(page))
141                                 continue;
142                         if (PageWriteback(page)) {
143                                 unlock_page(page);
144                                 continue;
145                         }
146                         truncate_complete_page(mapping, page);
147                         unlock_page(page);
148                 }
149                 pagevec_release(&pvec);
150                 cond_resched();
151         }
152
153         if (partial) {
154                 struct page *page = find_lock_page(mapping, start - 1);
155                 if (page) {
156                         wait_on_page_writeback(page);
157                         truncate_partial_page(page, partial);
158                         unlock_page(page);
159                         page_cache_release(page);
160                 }
161         }
162
163         next = start;
164         for ( ; ; ) {
165                 cond_resched();
166                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
167                         if (next == start)
168                                 break;
169                         next = start;
170                         continue;
171                 }
172                 if (pvec.pages[0]->index > end) {
173                         pagevec_release(&pvec);
174                         break;
175                 }
176                 for (i = 0; i < pagevec_count(&pvec); i++) {
177                         struct page *page = pvec.pages[i];
178
179                         if (page->index > end)
180                                 break;
181                         lock_page(page);
182                         wait_on_page_writeback(page);
183                         if (page->index > next)
184                                 next = page->index;
185                         next++;
186                         truncate_complete_page(mapping, page);
187                         unlock_page(page);
188                 }
189                 pagevec_release(&pvec);
190         }
191 }
192 EXPORT_SYMBOL(truncate_inode_pages_range);
193
194 /**
195  * truncate_inode_pages - truncate *all* the pages from an offset
196  * @mapping: mapping to truncate
197  * @lstart: offset from which to truncate
198  *
199  * Called under (and serialised by) inode->i_mutex.
200  */
201 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
202 {
203         truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
204 }
205 EXPORT_SYMBOL(truncate_inode_pages);
206
207 /**
208  * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
209  * @mapping: the address_space which holds the pages to invalidate
210  * @start: the offset 'from' which to invalidate
211  * @end: the offset 'to' which to invalidate (inclusive)
212  *
213  * This function only removes the unlocked pages, if you want to
214  * remove all the pages of one inode, you must call truncate_inode_pages.
215  *
216  * invalidate_mapping_pages() will not block on IO activity. It will not
217  * invalidate pages which are dirty, locked, under writeback or mapped into
218  * pagetables.
219  */
220 unsigned long invalidate_mapping_pages(struct address_space *mapping,
221                                 pgoff_t start, pgoff_t end)
222 {
223         struct pagevec pvec;
224         pgoff_t next = start;
225         unsigned long ret = 0;
226         int i;
227
228         pagevec_init(&pvec, 0);
229         while (next <= end &&
230                         pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
231                 for (i = 0; i < pagevec_count(&pvec); i++) {
232                         struct page *page = pvec.pages[i];
233
234                         if (TestSetPageLocked(page)) {
235                                 next++;
236                                 continue;
237                         }
238                         if (page->index > next)
239                                 next = page->index;
240                         next++;
241                         if (PageDirty(page) || PageWriteback(page))
242                                 goto unlock;
243                         if (page_mapped(page))
244                                 goto unlock;
245                         ret += invalidate_complete_page(mapping, page);
246 unlock:
247                         unlock_page(page);
248                         if (next > end)
249                                 break;
250                 }
251                 pagevec_release(&pvec);
252         }
253         return ret;
254 }
255
256 unsigned long invalidate_inode_pages(struct address_space *mapping)
257 {
258         return invalidate_mapping_pages(mapping, 0, ~0UL);
259 }
260
261 EXPORT_SYMBOL(invalidate_inode_pages);
262
263 /**
264  * invalidate_inode_pages2_range - remove range of pages from an address_space
265  * @mapping: the address_space
266  * @start: the page offset 'from' which to invalidate
267  * @end: the page offset 'to' which to invalidate (inclusive)
268  *
269  * Any pages which are found to be mapped into pagetables are unmapped prior to
270  * invalidation.
271  *
272  * Returns -EIO if any pages could not be invalidated.
273  */
274 int invalidate_inode_pages2_range(struct address_space *mapping,
275                                   pgoff_t start, pgoff_t end)
276 {
277         struct pagevec pvec;
278         pgoff_t next;
279         int i;
280         int ret = 0;
281         int did_range_unmap = 0;
282         int wrapped = 0;
283
284         pagevec_init(&pvec, 0);
285         next = start;
286         while (next <= end && !ret && !wrapped &&
287                 pagevec_lookup(&pvec, mapping, next,
288                         min(end - next, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
289                 for (i = 0; !ret && i < pagevec_count(&pvec); i++) {
290                         struct page *page = pvec.pages[i];
291                         pgoff_t page_index;
292                         int was_dirty;
293
294                         lock_page(page);
295                         if (page->mapping != mapping) {
296                                 unlock_page(page);
297                                 continue;
298                         }
299                         page_index = page->index;
300                         next = page_index + 1;
301                         if (next == 0)
302                                 wrapped = 1;
303                         if (page_index > end) {
304                                 unlock_page(page);
305                                 break;
306                         }
307                         wait_on_page_writeback(page);
308                         while (page_mapped(page)) {
309                                 if (!did_range_unmap) {
310                                         /*
311                                          * Zap the rest of the file in one hit.
312                                          */
313                                         unmap_mapping_range(mapping,
314                                            (loff_t)page_index<<PAGE_CACHE_SHIFT,
315                                            (loff_t)(end - page_index + 1)
316                                                         << PAGE_CACHE_SHIFT,
317                                             0);
318                                         did_range_unmap = 1;
319                                 } else {
320                                         /*
321                                          * Just zap this page
322                                          */
323                                         unmap_mapping_range(mapping,
324                                           (loff_t)page_index<<PAGE_CACHE_SHIFT,
325                                           PAGE_CACHE_SIZE, 0);
326                                 }
327                         }
328                         was_dirty = test_clear_page_dirty(page);
329                         if (!invalidate_complete_page(mapping, page)) {
330                                 if (was_dirty)
331                                         set_page_dirty(page);
332                                 ret = -EIO;
333                         }
334                         unlock_page(page);
335                 }
336                 pagevec_release(&pvec);
337                 cond_resched();
338         }
339         return ret;
340 }
341 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
342
343 /**
344  * invalidate_inode_pages2 - remove all pages from an address_space
345  * @mapping: the address_space
346  *
347  * Any pages which are found to be mapped into pagetables are unmapped prior to
348  * invalidation.
349  *
350  * Returns -EIO if any pages could not be invalidated.
351  */
352 int invalidate_inode_pages2(struct address_space *mapping)
353 {
354         return invalidate_inode_pages2_range(mapping, 0, -1);
355 }
356 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);