Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * mm/rmap.c - physical to virtual reverse mappings | |
3 | * | |
4 | * Copyright 2001, Rik van Riel <riel@conectiva.com.br> | |
5 | * Released under the General Public License (GPL). | |
6 | * | |
7 | * Simple, low overhead reverse mapping scheme. | |
8 | * Please try to keep this thing as modular as possible. | |
9 | * | |
10 | * Provides methods for unmapping each kind of mapped page: | |
11 | * the anon methods track anonymous pages, and | |
12 | * the file methods track pages belonging to an inode. | |
13 | * | |
14 | * Original design by Rik van Riel <riel@conectiva.com.br> 2001 | |
15 | * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 | |
16 | * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 | |
17 | * Contributions by Hugh Dickins <hugh@veritas.com> 2003, 2004 | |
18 | */ | |
19 | ||
20 | /* | |
21 | * Lock ordering in mm: | |
22 | * | |
1b1dcc1b | 23 | * inode->i_mutex (while writing or truncating, not reading or faulting) |
1da177e4 LT |
24 | * inode->i_alloc_sem |
25 | * | |
26 | * When a page fault occurs in writing from user to file, down_read | |
1b1dcc1b JS |
27 | * of mmap_sem nests within i_mutex; in sys_msync, i_mutex nests within |
28 | * down_read of mmap_sem; i_mutex and down_write of mmap_sem are never | |
29 | * taken together; in truncation, i_mutex is taken outermost. | |
1da177e4 LT |
30 | * |
31 | * mm->mmap_sem | |
32 | * page->flags PG_locked (lock_page) | |
33 | * mapping->i_mmap_lock | |
34 | * anon_vma->lock | |
b8072f09 | 35 | * mm->page_table_lock or pte_lock |
053837fc | 36 | * zone->lru_lock (in mark_page_accessed, isolate_lru_page) |
5d337b91 | 37 | * swap_lock (in swap_duplicate, swap_info_get) |
1da177e4 | 38 | * mmlist_lock (in mmput, drain_mmlist and others) |
1da177e4 LT |
39 | * mapping->private_lock (in __set_page_dirty_buffers) |
40 | * inode_lock (in set_page_dirty's __mark_inode_dirty) | |
41 | * sb_lock (within inode_lock in fs/fs-writeback.c) | |
42 | * mapping->tree_lock (widely used, in set_page_dirty, | |
43 | * in arch-dependent flush_dcache_mmap_lock, | |
44 | * within inode_lock in __sync_single_inode) | |
45 | */ | |
46 | ||
47 | #include <linux/mm.h> | |
48 | #include <linux/pagemap.h> | |
49 | #include <linux/swap.h> | |
50 | #include <linux/swapops.h> | |
51 | #include <linux/slab.h> | |
52 | #include <linux/init.h> | |
53 | #include <linux/rmap.h> | |
54 | #include <linux/rcupdate.h> | |
a48d07af | 55 | #include <linux/module.h> |
1da177e4 LT |
56 | |
57 | #include <asm/tlbflush.h> | |
58 | ||
fcc234f8 | 59 | struct kmem_cache *anon_vma_cachep; |
1da177e4 LT |
60 | |
61 | static inline void validate_anon_vma(struct vm_area_struct *find_vma) | |
62 | { | |
b7ab795b | 63 | #ifdef CONFIG_DEBUG_VM |
1da177e4 LT |
64 | struct anon_vma *anon_vma = find_vma->anon_vma; |
65 | struct vm_area_struct *vma; | |
66 | unsigned int mapcount = 0; | |
67 | int found = 0; | |
68 | ||
69 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
70 | mapcount++; | |
71 | BUG_ON(mapcount > 100000); | |
72 | if (vma == find_vma) | |
73 | found = 1; | |
74 | } | |
75 | BUG_ON(!found); | |
76 | #endif | |
77 | } | |
78 | ||
79 | /* This must be called under the mmap_sem. */ | |
80 | int anon_vma_prepare(struct vm_area_struct *vma) | |
81 | { | |
82 | struct anon_vma *anon_vma = vma->anon_vma; | |
83 | ||
84 | might_sleep(); | |
85 | if (unlikely(!anon_vma)) { | |
86 | struct mm_struct *mm = vma->vm_mm; | |
87 | struct anon_vma *allocated, *locked; | |
88 | ||
89 | anon_vma = find_mergeable_anon_vma(vma); | |
90 | if (anon_vma) { | |
91 | allocated = NULL; | |
92 | locked = anon_vma; | |
93 | spin_lock(&locked->lock); | |
94 | } else { | |
95 | anon_vma = anon_vma_alloc(); | |
96 | if (unlikely(!anon_vma)) | |
97 | return -ENOMEM; | |
98 | allocated = anon_vma; | |
99 | locked = NULL; | |
100 | } | |
101 | ||
102 | /* page_table_lock to protect against threads */ | |
103 | spin_lock(&mm->page_table_lock); | |
104 | if (likely(!vma->anon_vma)) { | |
105 | vma->anon_vma = anon_vma; | |
106 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
107 | allocated = NULL; | |
108 | } | |
109 | spin_unlock(&mm->page_table_lock); | |
110 | ||
111 | if (locked) | |
112 | spin_unlock(&locked->lock); | |
113 | if (unlikely(allocated)) | |
114 | anon_vma_free(allocated); | |
115 | } | |
116 | return 0; | |
117 | } | |
118 | ||
119 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) | |
120 | { | |
121 | BUG_ON(vma->anon_vma != next->anon_vma); | |
122 | list_del(&next->anon_vma_node); | |
123 | } | |
124 | ||
125 | void __anon_vma_link(struct vm_area_struct *vma) | |
126 | { | |
127 | struct anon_vma *anon_vma = vma->anon_vma; | |
128 | ||
129 | if (anon_vma) { | |
130 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
131 | validate_anon_vma(vma); | |
132 | } | |
133 | } | |
134 | ||
135 | void anon_vma_link(struct vm_area_struct *vma) | |
136 | { | |
137 | struct anon_vma *anon_vma = vma->anon_vma; | |
138 | ||
139 | if (anon_vma) { | |
140 | spin_lock(&anon_vma->lock); | |
141 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
142 | validate_anon_vma(vma); | |
143 | spin_unlock(&anon_vma->lock); | |
144 | } | |
145 | } | |
146 | ||
147 | void anon_vma_unlink(struct vm_area_struct *vma) | |
148 | { | |
149 | struct anon_vma *anon_vma = vma->anon_vma; | |
150 | int empty; | |
151 | ||
152 | if (!anon_vma) | |
153 | return; | |
154 | ||
155 | spin_lock(&anon_vma->lock); | |
156 | validate_anon_vma(vma); | |
157 | list_del(&vma->anon_vma_node); | |
158 | ||
159 | /* We must garbage collect the anon_vma if it's empty */ | |
160 | empty = list_empty(&anon_vma->head); | |
161 | spin_unlock(&anon_vma->lock); | |
162 | ||
163 | if (empty) | |
164 | anon_vma_free(anon_vma); | |
165 | } | |
166 | ||
fcc234f8 PE |
167 | static void anon_vma_ctor(void *data, struct kmem_cache *cachep, |
168 | unsigned long flags) | |
1da177e4 LT |
169 | { |
170 | if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == | |
171 | SLAB_CTOR_CONSTRUCTOR) { | |
172 | struct anon_vma *anon_vma = data; | |
173 | ||
174 | spin_lock_init(&anon_vma->lock); | |
175 | INIT_LIST_HEAD(&anon_vma->head); | |
176 | } | |
177 | } | |
178 | ||
179 | void __init anon_vma_init(void) | |
180 | { | |
181 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
182 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL); | |
183 | } | |
184 | ||
185 | /* | |
186 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
187 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
188 | */ | |
189 | static struct anon_vma *page_lock_anon_vma(struct page *page) | |
190 | { | |
191 | struct anon_vma *anon_vma = NULL; | |
192 | unsigned long anon_mapping; | |
193 | ||
194 | rcu_read_lock(); | |
195 | anon_mapping = (unsigned long) page->mapping; | |
196 | if (!(anon_mapping & PAGE_MAPPING_ANON)) | |
197 | goto out; | |
198 | if (!page_mapped(page)) | |
199 | goto out; | |
200 | ||
201 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
202 | spin_lock(&anon_vma->lock); | |
203 | out: | |
204 | rcu_read_unlock(); | |
205 | return anon_vma; | |
206 | } | |
207 | ||
a3351e52 CL |
208 | #ifdef CONFIG_MIGRATION |
209 | /* | |
210 | * Remove an anonymous page from swap replacing the swap pte's | |
211 | * through real pte's pointing to valid pages and then releasing | |
212 | * the page from the swap cache. | |
213 | * | |
e8788c0c CL |
214 | * Must hold page lock on page and mmap_sem of one vma that contains |
215 | * the page. | |
a3351e52 CL |
216 | */ |
217 | void remove_from_swap(struct page *page) | |
218 | { | |
219 | struct anon_vma *anon_vma; | |
220 | struct vm_area_struct *vma; | |
e8788c0c | 221 | unsigned long mapping; |
a3351e52 | 222 | |
e8788c0c | 223 | if (!PageSwapCache(page)) |
a3351e52 CL |
224 | return; |
225 | ||
e8788c0c CL |
226 | mapping = (unsigned long)page->mapping; |
227 | ||
228 | if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) | |
a3351e52 CL |
229 | return; |
230 | ||
e8788c0c CL |
231 | /* |
232 | * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. | |
233 | */ | |
234 | anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); | |
235 | spin_lock(&anon_vma->lock); | |
236 | ||
a3351e52 CL |
237 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) |
238 | remove_vma_swap(vma, page); | |
239 | ||
240 | spin_unlock(&anon_vma->lock); | |
a3351e52 CL |
241 | delete_from_swap_cache(page); |
242 | } | |
e965f963 | 243 | EXPORT_SYMBOL(remove_from_swap); |
a3351e52 CL |
244 | #endif |
245 | ||
1da177e4 LT |
246 | /* |
247 | * At what user virtual address is page expected in vma? | |
248 | */ | |
249 | static inline unsigned long | |
250 | vma_address(struct page *page, struct vm_area_struct *vma) | |
251 | { | |
252 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
253 | unsigned long address; | |
254 | ||
255 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
256 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
257 | /* page should be within any vma from prio_tree_next */ | |
258 | BUG_ON(!PageAnon(page)); | |
259 | return -EFAULT; | |
260 | } | |
261 | return address; | |
262 | } | |
263 | ||
264 | /* | |
265 | * At what user virtual address is page expected in vma? checking that the | |
ee498ed7 | 266 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
1da177e4 LT |
267 | */ |
268 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
269 | { | |
270 | if (PageAnon(page)) { | |
271 | if ((void *)vma->anon_vma != | |
272 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
273 | return -EFAULT; | |
274 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
275 | if (!vma->vm_file || |
276 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
277 | return -EFAULT; |
278 | } else | |
279 | return -EFAULT; | |
280 | return vma_address(page, vma); | |
281 | } | |
282 | ||
81b4082d ND |
283 | /* |
284 | * Check that @page is mapped at @address into @mm. | |
285 | * | |
b8072f09 | 286 | * On success returns with pte mapped and locked. |
81b4082d | 287 | */ |
ceffc078 | 288 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
c0718806 | 289 | unsigned long address, spinlock_t **ptlp) |
81b4082d ND |
290 | { |
291 | pgd_t *pgd; | |
292 | pud_t *pud; | |
293 | pmd_t *pmd; | |
294 | pte_t *pte; | |
c0718806 | 295 | spinlock_t *ptl; |
81b4082d | 296 | |
81b4082d | 297 | pgd = pgd_offset(mm, address); |
c0718806 HD |
298 | if (!pgd_present(*pgd)) |
299 | return NULL; | |
300 | ||
301 | pud = pud_offset(pgd, address); | |
302 | if (!pud_present(*pud)) | |
303 | return NULL; | |
304 | ||
305 | pmd = pmd_offset(pud, address); | |
306 | if (!pmd_present(*pmd)) | |
307 | return NULL; | |
308 | ||
309 | pte = pte_offset_map(pmd, address); | |
310 | /* Make a quick check before getting the lock */ | |
311 | if (!pte_present(*pte)) { | |
312 | pte_unmap(pte); | |
313 | return NULL; | |
314 | } | |
315 | ||
4c21e2f2 | 316 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
317 | spin_lock(ptl); |
318 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
319 | *ptlp = ptl; | |
320 | return pte; | |
81b4082d | 321 | } |
c0718806 HD |
322 | pte_unmap_unlock(pte, ptl); |
323 | return NULL; | |
81b4082d ND |
324 | } |
325 | ||
1da177e4 LT |
326 | /* |
327 | * Subfunctions of page_referenced: page_referenced_one called | |
328 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
329 | */ | |
330 | static int page_referenced_one(struct page *page, | |
f7b7fd8f | 331 | struct vm_area_struct *vma, unsigned int *mapcount) |
1da177e4 LT |
332 | { |
333 | struct mm_struct *mm = vma->vm_mm; | |
334 | unsigned long address; | |
1da177e4 | 335 | pte_t *pte; |
c0718806 | 336 | spinlock_t *ptl; |
1da177e4 LT |
337 | int referenced = 0; |
338 | ||
1da177e4 LT |
339 | address = vma_address(page, vma); |
340 | if (address == -EFAULT) | |
341 | goto out; | |
342 | ||
c0718806 HD |
343 | pte = page_check_address(page, mm, address, &ptl); |
344 | if (!pte) | |
345 | goto out; | |
1da177e4 | 346 | |
c0718806 HD |
347 | if (ptep_clear_flush_young(vma, address, pte)) |
348 | referenced++; | |
1da177e4 | 349 | |
c0718806 HD |
350 | /* Pretend the page is referenced if the task has the |
351 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 352 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
353 | rwsem_is_locked(&mm->mmap_sem)) |
354 | referenced++; | |
355 | ||
356 | (*mapcount)--; | |
357 | pte_unmap_unlock(pte, ptl); | |
1da177e4 LT |
358 | out: |
359 | return referenced; | |
360 | } | |
361 | ||
f7b7fd8f | 362 | static int page_referenced_anon(struct page *page) |
1da177e4 LT |
363 | { |
364 | unsigned int mapcount; | |
365 | struct anon_vma *anon_vma; | |
366 | struct vm_area_struct *vma; | |
367 | int referenced = 0; | |
368 | ||
369 | anon_vma = page_lock_anon_vma(page); | |
370 | if (!anon_vma) | |
371 | return referenced; | |
372 | ||
373 | mapcount = page_mapcount(page); | |
374 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
f7b7fd8f | 375 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
376 | if (!mapcount) |
377 | break; | |
378 | } | |
379 | spin_unlock(&anon_vma->lock); | |
380 | return referenced; | |
381 | } | |
382 | ||
383 | /** | |
384 | * page_referenced_file - referenced check for object-based rmap | |
385 | * @page: the page we're checking references on. | |
386 | * | |
387 | * For an object-based mapped page, find all the places it is mapped and | |
388 | * check/clear the referenced flag. This is done by following the page->mapping | |
389 | * pointer, then walking the chain of vmas it holds. It returns the number | |
390 | * of references it found. | |
391 | * | |
392 | * This function is only called from page_referenced for object-based pages. | |
393 | */ | |
f7b7fd8f | 394 | static int page_referenced_file(struct page *page) |
1da177e4 LT |
395 | { |
396 | unsigned int mapcount; | |
397 | struct address_space *mapping = page->mapping; | |
398 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
399 | struct vm_area_struct *vma; | |
400 | struct prio_tree_iter iter; | |
401 | int referenced = 0; | |
402 | ||
403 | /* | |
404 | * The caller's checks on page->mapping and !PageAnon have made | |
405 | * sure that this is a file page: the check for page->mapping | |
406 | * excludes the case just before it gets set on an anon page. | |
407 | */ | |
408 | BUG_ON(PageAnon(page)); | |
409 | ||
410 | /* | |
411 | * The page lock not only makes sure that page->mapping cannot | |
412 | * suddenly be NULLified by truncation, it makes sure that the | |
413 | * structure at mapping cannot be freed and reused yet, | |
414 | * so we can safely take mapping->i_mmap_lock. | |
415 | */ | |
416 | BUG_ON(!PageLocked(page)); | |
417 | ||
418 | spin_lock(&mapping->i_mmap_lock); | |
419 | ||
420 | /* | |
421 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
422 | * is more likely to be accurate if we note it after spinning. | |
423 | */ | |
424 | mapcount = page_mapcount(page); | |
425 | ||
426 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
427 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
428 | == (VM_LOCKED|VM_MAYSHARE)) { | |
429 | referenced++; | |
430 | break; | |
431 | } | |
f7b7fd8f | 432 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
433 | if (!mapcount) |
434 | break; | |
435 | } | |
436 | ||
437 | spin_unlock(&mapping->i_mmap_lock); | |
438 | return referenced; | |
439 | } | |
440 | ||
441 | /** | |
442 | * page_referenced - test if the page was referenced | |
443 | * @page: the page to test | |
444 | * @is_locked: caller holds lock on the page | |
445 | * | |
446 | * Quick test_and_clear_referenced for all mappings to a page, | |
447 | * returns the number of ptes which referenced the page. | |
448 | */ | |
f7b7fd8f | 449 | int page_referenced(struct page *page, int is_locked) |
1da177e4 LT |
450 | { |
451 | int referenced = 0; | |
452 | ||
1da177e4 LT |
453 | if (page_test_and_clear_young(page)) |
454 | referenced++; | |
455 | ||
456 | if (TestClearPageReferenced(page)) | |
457 | referenced++; | |
458 | ||
459 | if (page_mapped(page) && page->mapping) { | |
460 | if (PageAnon(page)) | |
f7b7fd8f | 461 | referenced += page_referenced_anon(page); |
1da177e4 | 462 | else if (is_locked) |
f7b7fd8f | 463 | referenced += page_referenced_file(page); |
1da177e4 LT |
464 | else if (TestSetPageLocked(page)) |
465 | referenced++; | |
466 | else { | |
467 | if (page->mapping) | |
f7b7fd8f | 468 | referenced += page_referenced_file(page); |
1da177e4 LT |
469 | unlock_page(page); |
470 | } | |
471 | } | |
472 | return referenced; | |
473 | } | |
474 | ||
9617d95e NP |
475 | /** |
476 | * page_set_anon_rmap - setup new anonymous rmap | |
477 | * @page: the page to add the mapping to | |
478 | * @vma: the vm area in which the mapping is added | |
479 | * @address: the user virtual address mapped | |
480 | */ | |
481 | static void __page_set_anon_rmap(struct page *page, | |
482 | struct vm_area_struct *vma, unsigned long address) | |
483 | { | |
484 | struct anon_vma *anon_vma = vma->anon_vma; | |
485 | ||
486 | BUG_ON(!anon_vma); | |
487 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
488 | page->mapping = (struct address_space *) anon_vma; | |
489 | ||
490 | page->index = linear_page_index(vma, address); | |
491 | ||
a74609fa NP |
492 | /* |
493 | * nr_mapped state can be updated without turning off | |
494 | * interrupts because it is not modified via interrupt. | |
495 | */ | |
496 | __inc_page_state(nr_mapped); | |
9617d95e NP |
497 | } |
498 | ||
1da177e4 LT |
499 | /** |
500 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
501 | * @page: the page to add the mapping to | |
502 | * @vma: the vm area in which the mapping is added | |
503 | * @address: the user virtual address mapped | |
504 | * | |
b8072f09 | 505 | * The caller needs to hold the pte lock. |
1da177e4 LT |
506 | */ |
507 | void page_add_anon_rmap(struct page *page, | |
508 | struct vm_area_struct *vma, unsigned long address) | |
509 | { | |
9617d95e NP |
510 | if (atomic_inc_and_test(&page->_mapcount)) |
511 | __page_set_anon_rmap(page, vma, address); | |
1da177e4 LT |
512 | /* else checking page index and mapping is racy */ |
513 | } | |
514 | ||
9617d95e NP |
515 | /* |
516 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
517 | * @page: the page to add the mapping to | |
518 | * @vma: the vm area in which the mapping is added | |
519 | * @address: the user virtual address mapped | |
520 | * | |
521 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
522 | * This means the inc-and-test can be bypassed. | |
523 | */ | |
524 | void page_add_new_anon_rmap(struct page *page, | |
525 | struct vm_area_struct *vma, unsigned long address) | |
526 | { | |
527 | atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ | |
528 | __page_set_anon_rmap(page, vma, address); | |
529 | } | |
530 | ||
1da177e4 LT |
531 | /** |
532 | * page_add_file_rmap - add pte mapping to a file page | |
533 | * @page: the page to add the mapping to | |
534 | * | |
b8072f09 | 535 | * The caller needs to hold the pte lock. |
1da177e4 LT |
536 | */ |
537 | void page_add_file_rmap(struct page *page) | |
538 | { | |
1da177e4 | 539 | if (atomic_inc_and_test(&page->_mapcount)) |
a74609fa | 540 | __inc_page_state(nr_mapped); |
1da177e4 LT |
541 | } |
542 | ||
543 | /** | |
544 | * page_remove_rmap - take down pte mapping from a page | |
545 | * @page: page to remove mapping from | |
546 | * | |
b8072f09 | 547 | * The caller needs to hold the pte lock. |
1da177e4 LT |
548 | */ |
549 | void page_remove_rmap(struct page *page) | |
550 | { | |
1da177e4 | 551 | if (atomic_add_negative(-1, &page->_mapcount)) { |
b7ab795b NP |
552 | #ifdef CONFIG_DEBUG_VM |
553 | if (unlikely(page_mapcount(page) < 0)) { | |
ef2bf0dc DJ |
554 | printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page)); |
555 | printk (KERN_EMERG " page->flags = %lx\n", page->flags); | |
556 | printk (KERN_EMERG " page->count = %x\n", page_count(page)); | |
557 | printk (KERN_EMERG " page->mapping = %p\n", page->mapping); | |
558 | } | |
b7ab795b | 559 | #endif |
1da177e4 LT |
560 | BUG_ON(page_mapcount(page) < 0); |
561 | /* | |
562 | * It would be tidy to reset the PageAnon mapping here, | |
563 | * but that might overwrite a racing page_add_anon_rmap | |
564 | * which increments mapcount after us but sets mapping | |
565 | * before us: so leave the reset to free_hot_cold_page, | |
566 | * and remember that it's only reliable while mapped. | |
567 | * Leaving it set also helps swapoff to reinstate ptes | |
568 | * faster for those pages still in swapcache. | |
569 | */ | |
570 | if (page_test_and_clear_dirty(page)) | |
571 | set_page_dirty(page); | |
a74609fa | 572 | __dec_page_state(nr_mapped); |
1da177e4 LT |
573 | } |
574 | } | |
575 | ||
576 | /* | |
577 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
578 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
579 | */ | |
a48d07af CL |
580 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
581 | int ignore_refs) | |
1da177e4 LT |
582 | { |
583 | struct mm_struct *mm = vma->vm_mm; | |
584 | unsigned long address; | |
1da177e4 LT |
585 | pte_t *pte; |
586 | pte_t pteval; | |
c0718806 | 587 | spinlock_t *ptl; |
1da177e4 LT |
588 | int ret = SWAP_AGAIN; |
589 | ||
1da177e4 LT |
590 | address = vma_address(page, vma); |
591 | if (address == -EFAULT) | |
592 | goto out; | |
593 | ||
c0718806 HD |
594 | pte = page_check_address(page, mm, address, &ptl); |
595 | if (!pte) | |
81b4082d | 596 | goto out; |
1da177e4 LT |
597 | |
598 | /* | |
599 | * If the page is mlock()d, we cannot swap it out. | |
600 | * If it's recently referenced (perhaps page_referenced | |
601 | * skipped over this mm) then we should reactivate it. | |
602 | */ | |
101d2be7 | 603 | if ((vma->vm_flags & VM_LOCKED) || |
a48d07af CL |
604 | (ptep_clear_flush_young(vma, address, pte) |
605 | && !ignore_refs)) { | |
1da177e4 LT |
606 | ret = SWAP_FAIL; |
607 | goto out_unmap; | |
608 | } | |
609 | ||
1da177e4 LT |
610 | /* Nuke the page table entry. */ |
611 | flush_cache_page(vma, address, page_to_pfn(page)); | |
612 | pteval = ptep_clear_flush(vma, address, pte); | |
613 | ||
614 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
615 | if (pte_dirty(pteval)) | |
616 | set_page_dirty(page); | |
617 | ||
365e9c87 HD |
618 | /* Update high watermark before we lower rss */ |
619 | update_hiwater_rss(mm); | |
620 | ||
1da177e4 | 621 | if (PageAnon(page)) { |
4c21e2f2 | 622 | swp_entry_t entry = { .val = page_private(page) }; |
1da177e4 LT |
623 | /* |
624 | * Store the swap location in the pte. | |
625 | * See handle_pte_fault() ... | |
626 | */ | |
627 | BUG_ON(!PageSwapCache(page)); | |
628 | swap_duplicate(entry); | |
629 | if (list_empty(&mm->mmlist)) { | |
630 | spin_lock(&mmlist_lock); | |
f412ac08 HD |
631 | if (list_empty(&mm->mmlist)) |
632 | list_add(&mm->mmlist, &init_mm.mmlist); | |
1da177e4 LT |
633 | spin_unlock(&mmlist_lock); |
634 | } | |
635 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
636 | BUG_ON(pte_file(*pte)); | |
637 | dec_mm_counter(mm, anon_rss); | |
4294621f HD |
638 | } else |
639 | dec_mm_counter(mm, file_rss); | |
1da177e4 | 640 | |
1da177e4 LT |
641 | page_remove_rmap(page); |
642 | page_cache_release(page); | |
643 | ||
644 | out_unmap: | |
c0718806 | 645 | pte_unmap_unlock(pte, ptl); |
1da177e4 LT |
646 | out: |
647 | return ret; | |
648 | } | |
649 | ||
650 | /* | |
651 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
652 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
653 | * Consequently, given a particular page and its ->index, we cannot locate the | |
654 | * ptes which are mapping that page without an exhaustive linear search. | |
655 | * | |
656 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
657 | * maps the file to which the target page belongs. The ->vm_private_data field | |
658 | * holds the current cursor into that scan. Successive searches will circulate | |
659 | * around the vma's virtual address space. | |
660 | * | |
661 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
662 | * more scanning pressure is placed against them as well. Eventually pages | |
663 | * will become fully unmapped and are eligible for eviction. | |
664 | * | |
665 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
666 | * there there won't be many ptes located within the scan cluster. In this case | |
667 | * maybe we could scan further - to the end of the pte page, perhaps. | |
668 | */ | |
669 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
670 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
671 | ||
672 | static void try_to_unmap_cluster(unsigned long cursor, | |
673 | unsigned int *mapcount, struct vm_area_struct *vma) | |
674 | { | |
675 | struct mm_struct *mm = vma->vm_mm; | |
676 | pgd_t *pgd; | |
677 | pud_t *pud; | |
678 | pmd_t *pmd; | |
c0718806 | 679 | pte_t *pte; |
1da177e4 | 680 | pte_t pteval; |
c0718806 | 681 | spinlock_t *ptl; |
1da177e4 LT |
682 | struct page *page; |
683 | unsigned long address; | |
684 | unsigned long end; | |
1da177e4 | 685 | |
1da177e4 LT |
686 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
687 | end = address + CLUSTER_SIZE; | |
688 | if (address < vma->vm_start) | |
689 | address = vma->vm_start; | |
690 | if (end > vma->vm_end) | |
691 | end = vma->vm_end; | |
692 | ||
693 | pgd = pgd_offset(mm, address); | |
694 | if (!pgd_present(*pgd)) | |
c0718806 | 695 | return; |
1da177e4 LT |
696 | |
697 | pud = pud_offset(pgd, address); | |
698 | if (!pud_present(*pud)) | |
c0718806 | 699 | return; |
1da177e4 LT |
700 | |
701 | pmd = pmd_offset(pud, address); | |
702 | if (!pmd_present(*pmd)) | |
c0718806 HD |
703 | return; |
704 | ||
705 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 706 | |
365e9c87 HD |
707 | /* Update high watermark before we lower rss */ |
708 | update_hiwater_rss(mm); | |
709 | ||
c0718806 | 710 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
711 | if (!pte_present(*pte)) |
712 | continue; | |
6aab341e LT |
713 | page = vm_normal_page(vma, address, *pte); |
714 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 LT |
715 | |
716 | if (ptep_clear_flush_young(vma, address, pte)) | |
717 | continue; | |
718 | ||
719 | /* Nuke the page table entry. */ | |
eca35133 | 720 | flush_cache_page(vma, address, pte_pfn(*pte)); |
1da177e4 LT |
721 | pteval = ptep_clear_flush(vma, address, pte); |
722 | ||
723 | /* If nonlinear, store the file page offset in the pte. */ | |
724 | if (page->index != linear_page_index(vma, address)) | |
725 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
726 | ||
727 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
728 | if (pte_dirty(pteval)) | |
729 | set_page_dirty(page); | |
730 | ||
731 | page_remove_rmap(page); | |
732 | page_cache_release(page); | |
4294621f | 733 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
734 | (*mapcount)--; |
735 | } | |
c0718806 | 736 | pte_unmap_unlock(pte - 1, ptl); |
1da177e4 LT |
737 | } |
738 | ||
a48d07af | 739 | static int try_to_unmap_anon(struct page *page, int ignore_refs) |
1da177e4 LT |
740 | { |
741 | struct anon_vma *anon_vma; | |
742 | struct vm_area_struct *vma; | |
743 | int ret = SWAP_AGAIN; | |
744 | ||
745 | anon_vma = page_lock_anon_vma(page); | |
746 | if (!anon_vma) | |
747 | return ret; | |
748 | ||
749 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
a48d07af | 750 | ret = try_to_unmap_one(page, vma, ignore_refs); |
1da177e4 LT |
751 | if (ret == SWAP_FAIL || !page_mapped(page)) |
752 | break; | |
753 | } | |
754 | spin_unlock(&anon_vma->lock); | |
755 | return ret; | |
756 | } | |
757 | ||
758 | /** | |
759 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
760 | * @page: the page to unmap | |
761 | * | |
762 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
763 | * contained in the address_space struct it points to. | |
764 | * | |
765 | * This function is only called from try_to_unmap for object-based pages. | |
766 | */ | |
a48d07af | 767 | static int try_to_unmap_file(struct page *page, int ignore_refs) |
1da177e4 LT |
768 | { |
769 | struct address_space *mapping = page->mapping; | |
770 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
771 | struct vm_area_struct *vma; | |
772 | struct prio_tree_iter iter; | |
773 | int ret = SWAP_AGAIN; | |
774 | unsigned long cursor; | |
775 | unsigned long max_nl_cursor = 0; | |
776 | unsigned long max_nl_size = 0; | |
777 | unsigned int mapcount; | |
778 | ||
779 | spin_lock(&mapping->i_mmap_lock); | |
780 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
a48d07af | 781 | ret = try_to_unmap_one(page, vma, ignore_refs); |
1da177e4 LT |
782 | if (ret == SWAP_FAIL || !page_mapped(page)) |
783 | goto out; | |
784 | } | |
785 | ||
786 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
787 | goto out; | |
788 | ||
789 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
790 | shared.vm_set.list) { | |
101d2be7 | 791 | if (vma->vm_flags & VM_LOCKED) |
1da177e4 LT |
792 | continue; |
793 | cursor = (unsigned long) vma->vm_private_data; | |
794 | if (cursor > max_nl_cursor) | |
795 | max_nl_cursor = cursor; | |
796 | cursor = vma->vm_end - vma->vm_start; | |
797 | if (cursor > max_nl_size) | |
798 | max_nl_size = cursor; | |
799 | } | |
800 | ||
801 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
802 | ret = SWAP_FAIL; | |
803 | goto out; | |
804 | } | |
805 | ||
806 | /* | |
807 | * We don't try to search for this page in the nonlinear vmas, | |
808 | * and page_referenced wouldn't have found it anyway. Instead | |
809 | * just walk the nonlinear vmas trying to age and unmap some. | |
810 | * The mapcount of the page we came in with is irrelevant, | |
811 | * but even so use it as a guide to how hard we should try? | |
812 | */ | |
813 | mapcount = page_mapcount(page); | |
814 | if (!mapcount) | |
815 | goto out; | |
816 | cond_resched_lock(&mapping->i_mmap_lock); | |
817 | ||
818 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
819 | if (max_nl_cursor == 0) | |
820 | max_nl_cursor = CLUSTER_SIZE; | |
821 | ||
822 | do { | |
823 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
824 | shared.vm_set.list) { | |
101d2be7 | 825 | if (vma->vm_flags & VM_LOCKED) |
1da177e4 LT |
826 | continue; |
827 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 828 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
829 | cursor < vma->vm_end - vma->vm_start) { |
830 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
831 | cursor += CLUSTER_SIZE; | |
832 | vma->vm_private_data = (void *) cursor; | |
833 | if ((int)mapcount <= 0) | |
834 | goto out; | |
835 | } | |
836 | vma->vm_private_data = (void *) max_nl_cursor; | |
837 | } | |
838 | cond_resched_lock(&mapping->i_mmap_lock); | |
839 | max_nl_cursor += CLUSTER_SIZE; | |
840 | } while (max_nl_cursor <= max_nl_size); | |
841 | ||
842 | /* | |
843 | * Don't loop forever (perhaps all the remaining pages are | |
844 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
845 | * vmas, now forgetting on which ones it had fallen behind. | |
846 | */ | |
101d2be7 HD |
847 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
848 | vma->vm_private_data = NULL; | |
1da177e4 LT |
849 | out: |
850 | spin_unlock(&mapping->i_mmap_lock); | |
851 | return ret; | |
852 | } | |
853 | ||
854 | /** | |
855 | * try_to_unmap - try to remove all page table mappings to a page | |
856 | * @page: the page to get unmapped | |
857 | * | |
858 | * Tries to remove all the page table entries which are mapping this | |
859 | * page, used in the pageout path. Caller must hold the page lock. | |
860 | * Return values are: | |
861 | * | |
862 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
863 | * SWAP_AGAIN - we missed a mapping, try again later | |
864 | * SWAP_FAIL - the page is unswappable | |
865 | */ | |
a48d07af | 866 | int try_to_unmap(struct page *page, int ignore_refs) |
1da177e4 LT |
867 | { |
868 | int ret; | |
869 | ||
1da177e4 LT |
870 | BUG_ON(!PageLocked(page)); |
871 | ||
872 | if (PageAnon(page)) | |
a48d07af | 873 | ret = try_to_unmap_anon(page, ignore_refs); |
1da177e4 | 874 | else |
a48d07af | 875 | ret = try_to_unmap_file(page, ignore_refs); |
1da177e4 LT |
876 | |
877 | if (!page_mapped(page)) | |
878 | ret = SWAP_SUCCESS; | |
879 | return ret; | |
880 | } | |
81b4082d | 881 |