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