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) |
82591e6e NP |
24 | * inode->i_alloc_sem (vmtruncate_range) |
25 | * mm->mmap_sem | |
26 | * page->flags PG_locked (lock_page) | |
27 | * mapping->i_mmap_lock | |
28 | * anon_vma->lock | |
29 | * mm->page_table_lock or pte_lock | |
30 | * zone->lru_lock (in mark_page_accessed, isolate_lru_page) | |
31 | * swap_lock (in swap_duplicate, swap_info_get) | |
32 | * mmlist_lock (in mmput, drain_mmlist and others) | |
33 | * mapping->private_lock (in __set_page_dirty_buffers) | |
34 | * inode_lock (in set_page_dirty's __mark_inode_dirty) | |
35 | * sb_lock (within inode_lock in fs/fs-writeback.c) | |
36 | * mapping->tree_lock (widely used, in set_page_dirty, | |
37 | * in arch-dependent flush_dcache_mmap_lock, | |
38 | * within inode_lock in __sync_single_inode) | |
1da177e4 LT |
39 | */ |
40 | ||
41 | #include <linux/mm.h> | |
42 | #include <linux/pagemap.h> | |
43 | #include <linux/swap.h> | |
44 | #include <linux/swapops.h> | |
45 | #include <linux/slab.h> | |
46 | #include <linux/init.h> | |
47 | #include <linux/rmap.h> | |
48 | #include <linux/rcupdate.h> | |
a48d07af | 49 | #include <linux/module.h> |
7de6b805 | 50 | #include <linux/kallsyms.h> |
1da177e4 LT |
51 | |
52 | #include <asm/tlbflush.h> | |
53 | ||
fcc234f8 | 54 | struct kmem_cache *anon_vma_cachep; |
1da177e4 LT |
55 | |
56 | static inline void validate_anon_vma(struct vm_area_struct *find_vma) | |
57 | { | |
b7ab795b | 58 | #ifdef CONFIG_DEBUG_VM |
1da177e4 LT |
59 | struct anon_vma *anon_vma = find_vma->anon_vma; |
60 | struct vm_area_struct *vma; | |
61 | unsigned int mapcount = 0; | |
62 | int found = 0; | |
63 | ||
64 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
65 | mapcount++; | |
66 | BUG_ON(mapcount > 100000); | |
67 | if (vma == find_vma) | |
68 | found = 1; | |
69 | } | |
70 | BUG_ON(!found); | |
71 | #endif | |
72 | } | |
73 | ||
74 | /* This must be called under the mmap_sem. */ | |
75 | int anon_vma_prepare(struct vm_area_struct *vma) | |
76 | { | |
77 | struct anon_vma *anon_vma = vma->anon_vma; | |
78 | ||
79 | might_sleep(); | |
80 | if (unlikely(!anon_vma)) { | |
81 | struct mm_struct *mm = vma->vm_mm; | |
82 | struct anon_vma *allocated, *locked; | |
83 | ||
84 | anon_vma = find_mergeable_anon_vma(vma); | |
85 | if (anon_vma) { | |
86 | allocated = NULL; | |
87 | locked = anon_vma; | |
88 | spin_lock(&locked->lock); | |
89 | } else { | |
90 | anon_vma = anon_vma_alloc(); | |
91 | if (unlikely(!anon_vma)) | |
92 | return -ENOMEM; | |
93 | allocated = anon_vma; | |
94 | locked = NULL; | |
95 | } | |
96 | ||
97 | /* page_table_lock to protect against threads */ | |
98 | spin_lock(&mm->page_table_lock); | |
99 | if (likely(!vma->anon_vma)) { | |
100 | vma->anon_vma = anon_vma; | |
0697212a | 101 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
102 | allocated = NULL; |
103 | } | |
104 | spin_unlock(&mm->page_table_lock); | |
105 | ||
106 | if (locked) | |
107 | spin_unlock(&locked->lock); | |
108 | if (unlikely(allocated)) | |
109 | anon_vma_free(allocated); | |
110 | } | |
111 | return 0; | |
112 | } | |
113 | ||
114 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) | |
115 | { | |
116 | BUG_ON(vma->anon_vma != next->anon_vma); | |
117 | list_del(&next->anon_vma_node); | |
118 | } | |
119 | ||
120 | void __anon_vma_link(struct vm_area_struct *vma) | |
121 | { | |
122 | struct anon_vma *anon_vma = vma->anon_vma; | |
123 | ||
124 | if (anon_vma) { | |
0697212a | 125 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
126 | validate_anon_vma(vma); |
127 | } | |
128 | } | |
129 | ||
130 | void anon_vma_link(struct vm_area_struct *vma) | |
131 | { | |
132 | struct anon_vma *anon_vma = vma->anon_vma; | |
133 | ||
134 | if (anon_vma) { | |
135 | spin_lock(&anon_vma->lock); | |
0697212a | 136 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
137 | validate_anon_vma(vma); |
138 | spin_unlock(&anon_vma->lock); | |
139 | } | |
140 | } | |
141 | ||
142 | void anon_vma_unlink(struct vm_area_struct *vma) | |
143 | { | |
144 | struct anon_vma *anon_vma = vma->anon_vma; | |
145 | int empty; | |
146 | ||
147 | if (!anon_vma) | |
148 | return; | |
149 | ||
150 | spin_lock(&anon_vma->lock); | |
151 | validate_anon_vma(vma); | |
152 | list_del(&vma->anon_vma_node); | |
153 | ||
154 | /* We must garbage collect the anon_vma if it's empty */ | |
155 | empty = list_empty(&anon_vma->head); | |
156 | spin_unlock(&anon_vma->lock); | |
157 | ||
158 | if (empty) | |
159 | anon_vma_free(anon_vma); | |
160 | } | |
161 | ||
fcc234f8 PE |
162 | static void anon_vma_ctor(void *data, struct kmem_cache *cachep, |
163 | unsigned long flags) | |
1da177e4 LT |
164 | { |
165 | if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == | |
166 | SLAB_CTOR_CONSTRUCTOR) { | |
167 | struct anon_vma *anon_vma = data; | |
168 | ||
169 | spin_lock_init(&anon_vma->lock); | |
170 | INIT_LIST_HEAD(&anon_vma->head); | |
171 | } | |
172 | } | |
173 | ||
174 | void __init anon_vma_init(void) | |
175 | { | |
176 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
177 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL); | |
178 | } | |
179 | ||
180 | /* | |
181 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
182 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
183 | */ | |
184 | static struct anon_vma *page_lock_anon_vma(struct page *page) | |
185 | { | |
186 | struct anon_vma *anon_vma = NULL; | |
187 | unsigned long anon_mapping; | |
188 | ||
189 | rcu_read_lock(); | |
190 | anon_mapping = (unsigned long) page->mapping; | |
191 | if (!(anon_mapping & PAGE_MAPPING_ANON)) | |
192 | goto out; | |
193 | if (!page_mapped(page)) | |
194 | goto out; | |
195 | ||
196 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
197 | spin_lock(&anon_vma->lock); | |
198 | out: | |
199 | rcu_read_unlock(); | |
200 | return anon_vma; | |
201 | } | |
202 | ||
203 | /* | |
204 | * At what user virtual address is page expected in vma? | |
205 | */ | |
206 | static inline unsigned long | |
207 | vma_address(struct page *page, struct vm_area_struct *vma) | |
208 | { | |
209 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
210 | unsigned long address; | |
211 | ||
212 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
213 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
214 | /* page should be within any vma from prio_tree_next */ | |
215 | BUG_ON(!PageAnon(page)); | |
216 | return -EFAULT; | |
217 | } | |
218 | return address; | |
219 | } | |
220 | ||
221 | /* | |
222 | * At what user virtual address is page expected in vma? checking that the | |
ee498ed7 | 223 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
1da177e4 LT |
224 | */ |
225 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
226 | { | |
227 | if (PageAnon(page)) { | |
228 | if ((void *)vma->anon_vma != | |
229 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
230 | return -EFAULT; | |
231 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
232 | if (!vma->vm_file || |
233 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
234 | return -EFAULT; |
235 | } else | |
236 | return -EFAULT; | |
237 | return vma_address(page, vma); | |
238 | } | |
239 | ||
81b4082d ND |
240 | /* |
241 | * Check that @page is mapped at @address into @mm. | |
242 | * | |
b8072f09 | 243 | * On success returns with pte mapped and locked. |
81b4082d | 244 | */ |
ceffc078 | 245 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
c0718806 | 246 | unsigned long address, spinlock_t **ptlp) |
81b4082d ND |
247 | { |
248 | pgd_t *pgd; | |
249 | pud_t *pud; | |
250 | pmd_t *pmd; | |
251 | pte_t *pte; | |
c0718806 | 252 | spinlock_t *ptl; |
81b4082d | 253 | |
81b4082d | 254 | pgd = pgd_offset(mm, address); |
c0718806 HD |
255 | if (!pgd_present(*pgd)) |
256 | return NULL; | |
257 | ||
258 | pud = pud_offset(pgd, address); | |
259 | if (!pud_present(*pud)) | |
260 | return NULL; | |
261 | ||
262 | pmd = pmd_offset(pud, address); | |
263 | if (!pmd_present(*pmd)) | |
264 | return NULL; | |
265 | ||
266 | pte = pte_offset_map(pmd, address); | |
267 | /* Make a quick check before getting the lock */ | |
268 | if (!pte_present(*pte)) { | |
269 | pte_unmap(pte); | |
270 | return NULL; | |
271 | } | |
272 | ||
4c21e2f2 | 273 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
274 | spin_lock(ptl); |
275 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
276 | *ptlp = ptl; | |
277 | return pte; | |
81b4082d | 278 | } |
c0718806 HD |
279 | pte_unmap_unlock(pte, ptl); |
280 | return NULL; | |
81b4082d ND |
281 | } |
282 | ||
1da177e4 LT |
283 | /* |
284 | * Subfunctions of page_referenced: page_referenced_one called | |
285 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
286 | */ | |
287 | static int page_referenced_one(struct page *page, | |
f7b7fd8f | 288 | struct vm_area_struct *vma, unsigned int *mapcount) |
1da177e4 LT |
289 | { |
290 | struct mm_struct *mm = vma->vm_mm; | |
291 | unsigned long address; | |
1da177e4 | 292 | pte_t *pte; |
c0718806 | 293 | spinlock_t *ptl; |
1da177e4 LT |
294 | int referenced = 0; |
295 | ||
1da177e4 LT |
296 | address = vma_address(page, vma); |
297 | if (address == -EFAULT) | |
298 | goto out; | |
299 | ||
c0718806 HD |
300 | pte = page_check_address(page, mm, address, &ptl); |
301 | if (!pte) | |
302 | goto out; | |
1da177e4 | 303 | |
c0718806 HD |
304 | if (ptep_clear_flush_young(vma, address, pte)) |
305 | referenced++; | |
1da177e4 | 306 | |
c0718806 HD |
307 | /* Pretend the page is referenced if the task has the |
308 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 309 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
310 | rwsem_is_locked(&mm->mmap_sem)) |
311 | referenced++; | |
312 | ||
313 | (*mapcount)--; | |
314 | pte_unmap_unlock(pte, ptl); | |
1da177e4 LT |
315 | out: |
316 | return referenced; | |
317 | } | |
318 | ||
f7b7fd8f | 319 | static int page_referenced_anon(struct page *page) |
1da177e4 LT |
320 | { |
321 | unsigned int mapcount; | |
322 | struct anon_vma *anon_vma; | |
323 | struct vm_area_struct *vma; | |
324 | int referenced = 0; | |
325 | ||
326 | anon_vma = page_lock_anon_vma(page); | |
327 | if (!anon_vma) | |
328 | return referenced; | |
329 | ||
330 | mapcount = page_mapcount(page); | |
331 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
f7b7fd8f | 332 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
333 | if (!mapcount) |
334 | break; | |
335 | } | |
336 | spin_unlock(&anon_vma->lock); | |
337 | return referenced; | |
338 | } | |
339 | ||
340 | /** | |
341 | * page_referenced_file - referenced check for object-based rmap | |
342 | * @page: the page we're checking references on. | |
343 | * | |
344 | * For an object-based mapped page, find all the places it is mapped and | |
345 | * check/clear the referenced flag. This is done by following the page->mapping | |
346 | * pointer, then walking the chain of vmas it holds. It returns the number | |
347 | * of references it found. | |
348 | * | |
349 | * This function is only called from page_referenced for object-based pages. | |
350 | */ | |
f7b7fd8f | 351 | static int page_referenced_file(struct page *page) |
1da177e4 LT |
352 | { |
353 | unsigned int mapcount; | |
354 | struct address_space *mapping = page->mapping; | |
355 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
356 | struct vm_area_struct *vma; | |
357 | struct prio_tree_iter iter; | |
358 | int referenced = 0; | |
359 | ||
360 | /* | |
361 | * The caller's checks on page->mapping and !PageAnon have made | |
362 | * sure that this is a file page: the check for page->mapping | |
363 | * excludes the case just before it gets set on an anon page. | |
364 | */ | |
365 | BUG_ON(PageAnon(page)); | |
366 | ||
367 | /* | |
368 | * The page lock not only makes sure that page->mapping cannot | |
369 | * suddenly be NULLified by truncation, it makes sure that the | |
370 | * structure at mapping cannot be freed and reused yet, | |
371 | * so we can safely take mapping->i_mmap_lock. | |
372 | */ | |
373 | BUG_ON(!PageLocked(page)); | |
374 | ||
375 | spin_lock(&mapping->i_mmap_lock); | |
376 | ||
377 | /* | |
378 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
379 | * is more likely to be accurate if we note it after spinning. | |
380 | */ | |
381 | mapcount = page_mapcount(page); | |
382 | ||
383 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
384 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
385 | == (VM_LOCKED|VM_MAYSHARE)) { | |
386 | referenced++; | |
387 | break; | |
388 | } | |
f7b7fd8f | 389 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
390 | if (!mapcount) |
391 | break; | |
392 | } | |
393 | ||
394 | spin_unlock(&mapping->i_mmap_lock); | |
395 | return referenced; | |
396 | } | |
397 | ||
398 | /** | |
399 | * page_referenced - test if the page was referenced | |
400 | * @page: the page to test | |
401 | * @is_locked: caller holds lock on the page | |
402 | * | |
403 | * Quick test_and_clear_referenced for all mappings to a page, | |
404 | * returns the number of ptes which referenced the page. | |
405 | */ | |
f7b7fd8f | 406 | int page_referenced(struct page *page, int is_locked) |
1da177e4 LT |
407 | { |
408 | int referenced = 0; | |
409 | ||
1da177e4 LT |
410 | if (page_test_and_clear_young(page)) |
411 | referenced++; | |
412 | ||
413 | if (TestClearPageReferenced(page)) | |
414 | referenced++; | |
415 | ||
416 | if (page_mapped(page) && page->mapping) { | |
417 | if (PageAnon(page)) | |
f7b7fd8f | 418 | referenced += page_referenced_anon(page); |
1da177e4 | 419 | else if (is_locked) |
f7b7fd8f | 420 | referenced += page_referenced_file(page); |
1da177e4 LT |
421 | else if (TestSetPageLocked(page)) |
422 | referenced++; | |
423 | else { | |
424 | if (page->mapping) | |
f7b7fd8f | 425 | referenced += page_referenced_file(page); |
1da177e4 LT |
426 | unlock_page(page); |
427 | } | |
428 | } | |
429 | return referenced; | |
430 | } | |
431 | ||
d08b3851 PZ |
432 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma) |
433 | { | |
434 | struct mm_struct *mm = vma->vm_mm; | |
435 | unsigned long address; | |
436 | pte_t *pte, entry; | |
437 | spinlock_t *ptl; | |
438 | int ret = 0; | |
439 | ||
440 | address = vma_address(page, vma); | |
441 | if (address == -EFAULT) | |
442 | goto out; | |
443 | ||
444 | pte = page_check_address(page, mm, address, &ptl); | |
445 | if (!pte) | |
446 | goto out; | |
447 | ||
448 | if (!pte_dirty(*pte) && !pte_write(*pte)) | |
449 | goto unlock; | |
450 | ||
451 | entry = ptep_get_and_clear(mm, address, pte); | |
452 | entry = pte_mkclean(entry); | |
453 | entry = pte_wrprotect(entry); | |
454 | ptep_establish(vma, address, pte, entry); | |
455 | lazy_mmu_prot_update(entry); | |
456 | ret = 1; | |
457 | ||
458 | unlock: | |
459 | pte_unmap_unlock(pte, ptl); | |
460 | out: | |
461 | return ret; | |
462 | } | |
463 | ||
464 | static int page_mkclean_file(struct address_space *mapping, struct page *page) | |
465 | { | |
466 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
467 | struct vm_area_struct *vma; | |
468 | struct prio_tree_iter iter; | |
469 | int ret = 0; | |
470 | ||
471 | BUG_ON(PageAnon(page)); | |
472 | ||
473 | spin_lock(&mapping->i_mmap_lock); | |
474 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
475 | if (vma->vm_flags & VM_SHARED) | |
476 | ret += page_mkclean_one(page, vma); | |
477 | } | |
478 | spin_unlock(&mapping->i_mmap_lock); | |
479 | return ret; | |
480 | } | |
481 | ||
482 | int page_mkclean(struct page *page) | |
483 | { | |
484 | int ret = 0; | |
485 | ||
486 | BUG_ON(!PageLocked(page)); | |
487 | ||
488 | if (page_mapped(page)) { | |
489 | struct address_space *mapping = page_mapping(page); | |
490 | if (mapping) | |
491 | ret = page_mkclean_file(mapping, page); | |
492 | } | |
493 | ||
494 | return ret; | |
495 | } | |
496 | ||
9617d95e NP |
497 | /** |
498 | * page_set_anon_rmap - setup new anonymous rmap | |
499 | * @page: the page to add the mapping to | |
500 | * @vma: the vm area in which the mapping is added | |
501 | * @address: the user virtual address mapped | |
502 | */ | |
503 | static void __page_set_anon_rmap(struct page *page, | |
504 | struct vm_area_struct *vma, unsigned long address) | |
505 | { | |
506 | struct anon_vma *anon_vma = vma->anon_vma; | |
507 | ||
508 | BUG_ON(!anon_vma); | |
509 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
510 | page->mapping = (struct address_space *) anon_vma; | |
511 | ||
512 | page->index = linear_page_index(vma, address); | |
513 | ||
a74609fa NP |
514 | /* |
515 | * nr_mapped state can be updated without turning off | |
516 | * interrupts because it is not modified via interrupt. | |
517 | */ | |
f3dbd344 | 518 | __inc_zone_page_state(page, NR_ANON_PAGES); |
9617d95e NP |
519 | } |
520 | ||
1da177e4 LT |
521 | /** |
522 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
523 | * @page: the page to add the mapping to | |
524 | * @vma: the vm area in which the mapping is added | |
525 | * @address: the user virtual address mapped | |
526 | * | |
b8072f09 | 527 | * The caller needs to hold the pte lock. |
1da177e4 LT |
528 | */ |
529 | void page_add_anon_rmap(struct page *page, | |
530 | struct vm_area_struct *vma, unsigned long address) | |
531 | { | |
9617d95e NP |
532 | if (atomic_inc_and_test(&page->_mapcount)) |
533 | __page_set_anon_rmap(page, vma, address); | |
1da177e4 LT |
534 | /* else checking page index and mapping is racy */ |
535 | } | |
536 | ||
9617d95e NP |
537 | /* |
538 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
539 | * @page: the page to add the mapping to | |
540 | * @vma: the vm area in which the mapping is added | |
541 | * @address: the user virtual address mapped | |
542 | * | |
543 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
544 | * This means the inc-and-test can be bypassed. | |
545 | */ | |
546 | void page_add_new_anon_rmap(struct page *page, | |
547 | struct vm_area_struct *vma, unsigned long address) | |
548 | { | |
549 | atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ | |
550 | __page_set_anon_rmap(page, vma, address); | |
551 | } | |
552 | ||
1da177e4 LT |
553 | /** |
554 | * page_add_file_rmap - add pte mapping to a file page | |
555 | * @page: the page to add the mapping to | |
556 | * | |
b8072f09 | 557 | * The caller needs to hold the pte lock. |
1da177e4 LT |
558 | */ |
559 | void page_add_file_rmap(struct page *page) | |
560 | { | |
1da177e4 | 561 | if (atomic_inc_and_test(&page->_mapcount)) |
65ba55f5 | 562 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
1da177e4 LT |
563 | } |
564 | ||
565 | /** | |
566 | * page_remove_rmap - take down pte mapping from a page | |
567 | * @page: page to remove mapping from | |
568 | * | |
b8072f09 | 569 | * The caller needs to hold the pte lock. |
1da177e4 | 570 | */ |
7de6b805 | 571 | void page_remove_rmap(struct page *page, struct vm_area_struct *vma) |
1da177e4 | 572 | { |
1da177e4 | 573 | if (atomic_add_negative(-1, &page->_mapcount)) { |
b7ab795b | 574 | if (unlikely(page_mapcount(page) < 0)) { |
ef2bf0dc | 575 | printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page)); |
7de6b805 | 576 | printk (KERN_EMERG " page pfn = %lx\n", page_to_pfn(page)); |
ef2bf0dc DJ |
577 | printk (KERN_EMERG " page->flags = %lx\n", page->flags); |
578 | printk (KERN_EMERG " page->count = %x\n", page_count(page)); | |
579 | printk (KERN_EMERG " page->mapping = %p\n", page->mapping); | |
7de6b805 NP |
580 | print_symbol (KERN_EMERG " vma->vm_ops = %s\n", (unsigned long)vma->vm_ops); |
581 | if (vma->vm_ops) | |
582 | print_symbol (KERN_EMERG " vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage); | |
583 | if (vma->vm_file && vma->vm_file->f_op) | |
584 | print_symbol (KERN_EMERG " vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap); | |
b16bc64d | 585 | BUG(); |
ef2bf0dc | 586 | } |
b16bc64d | 587 | |
1da177e4 LT |
588 | /* |
589 | * It would be tidy to reset the PageAnon mapping here, | |
590 | * but that might overwrite a racing page_add_anon_rmap | |
591 | * which increments mapcount after us but sets mapping | |
592 | * before us: so leave the reset to free_hot_cold_page, | |
593 | * and remember that it's only reliable while mapped. | |
594 | * Leaving it set also helps swapoff to reinstate ptes | |
595 | * faster for those pages still in swapcache. | |
596 | */ | |
597 | if (page_test_and_clear_dirty(page)) | |
598 | set_page_dirty(page); | |
f3dbd344 CL |
599 | __dec_zone_page_state(page, |
600 | PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED); | |
1da177e4 LT |
601 | } |
602 | } | |
603 | ||
604 | /* | |
605 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
606 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
607 | */ | |
a48d07af | 608 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
7352349a | 609 | int migration) |
1da177e4 LT |
610 | { |
611 | struct mm_struct *mm = vma->vm_mm; | |
612 | unsigned long address; | |
1da177e4 LT |
613 | pte_t *pte; |
614 | pte_t pteval; | |
c0718806 | 615 | spinlock_t *ptl; |
1da177e4 LT |
616 | int ret = SWAP_AGAIN; |
617 | ||
1da177e4 LT |
618 | address = vma_address(page, vma); |
619 | if (address == -EFAULT) | |
620 | goto out; | |
621 | ||
c0718806 HD |
622 | pte = page_check_address(page, mm, address, &ptl); |
623 | if (!pte) | |
81b4082d | 624 | goto out; |
1da177e4 LT |
625 | |
626 | /* | |
627 | * If the page is mlock()d, we cannot swap it out. | |
628 | * If it's recently referenced (perhaps page_referenced | |
629 | * skipped over this mm) then we should reactivate it. | |
630 | */ | |
e6a1530d CL |
631 | if (!migration && ((vma->vm_flags & VM_LOCKED) || |
632 | (ptep_clear_flush_young(vma, address, pte)))) { | |
1da177e4 LT |
633 | ret = SWAP_FAIL; |
634 | goto out_unmap; | |
635 | } | |
636 | ||
1da177e4 LT |
637 | /* Nuke the page table entry. */ |
638 | flush_cache_page(vma, address, page_to_pfn(page)); | |
639 | pteval = ptep_clear_flush(vma, address, pte); | |
640 | ||
641 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
642 | if (pte_dirty(pteval)) | |
643 | set_page_dirty(page); | |
644 | ||
365e9c87 HD |
645 | /* Update high watermark before we lower rss */ |
646 | update_hiwater_rss(mm); | |
647 | ||
1da177e4 | 648 | if (PageAnon(page)) { |
4c21e2f2 | 649 | swp_entry_t entry = { .val = page_private(page) }; |
0697212a CL |
650 | |
651 | if (PageSwapCache(page)) { | |
652 | /* | |
653 | * Store the swap location in the pte. | |
654 | * See handle_pte_fault() ... | |
655 | */ | |
656 | swap_duplicate(entry); | |
657 | if (list_empty(&mm->mmlist)) { | |
658 | spin_lock(&mmlist_lock); | |
659 | if (list_empty(&mm->mmlist)) | |
660 | list_add(&mm->mmlist, &init_mm.mmlist); | |
661 | spin_unlock(&mmlist_lock); | |
662 | } | |
442c9137 | 663 | dec_mm_counter(mm, anon_rss); |
04e62a29 | 664 | #ifdef CONFIG_MIGRATION |
0697212a CL |
665 | } else { |
666 | /* | |
667 | * Store the pfn of the page in a special migration | |
668 | * pte. do_swap_page() will wait until the migration | |
669 | * pte is removed and then restart fault handling. | |
670 | */ | |
671 | BUG_ON(!migration); | |
672 | entry = make_migration_entry(page, pte_write(pteval)); | |
04e62a29 | 673 | #endif |
1da177e4 LT |
674 | } |
675 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
676 | BUG_ON(pte_file(*pte)); | |
4294621f | 677 | } else |
04e62a29 CL |
678 | #ifdef CONFIG_MIGRATION |
679 | if (migration) { | |
680 | /* Establish migration entry for a file page */ | |
681 | swp_entry_t entry; | |
682 | entry = make_migration_entry(page, pte_write(pteval)); | |
683 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
684 | } else | |
685 | #endif | |
4294621f | 686 | dec_mm_counter(mm, file_rss); |
1da177e4 | 687 | |
04e62a29 | 688 | |
7de6b805 | 689 | page_remove_rmap(page, vma); |
1da177e4 LT |
690 | page_cache_release(page); |
691 | ||
692 | out_unmap: | |
c0718806 | 693 | pte_unmap_unlock(pte, ptl); |
1da177e4 LT |
694 | out: |
695 | return ret; | |
696 | } | |
697 | ||
698 | /* | |
699 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
700 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
701 | * Consequently, given a particular page and its ->index, we cannot locate the | |
702 | * ptes which are mapping that page without an exhaustive linear search. | |
703 | * | |
704 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
705 | * maps the file to which the target page belongs. The ->vm_private_data field | |
706 | * holds the current cursor into that scan. Successive searches will circulate | |
707 | * around the vma's virtual address space. | |
708 | * | |
709 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
710 | * more scanning pressure is placed against them as well. Eventually pages | |
711 | * will become fully unmapped and are eligible for eviction. | |
712 | * | |
713 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
714 | * there there won't be many ptes located within the scan cluster. In this case | |
715 | * maybe we could scan further - to the end of the pte page, perhaps. | |
716 | */ | |
717 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
718 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
719 | ||
720 | static void try_to_unmap_cluster(unsigned long cursor, | |
721 | unsigned int *mapcount, struct vm_area_struct *vma) | |
722 | { | |
723 | struct mm_struct *mm = vma->vm_mm; | |
724 | pgd_t *pgd; | |
725 | pud_t *pud; | |
726 | pmd_t *pmd; | |
c0718806 | 727 | pte_t *pte; |
1da177e4 | 728 | pte_t pteval; |
c0718806 | 729 | spinlock_t *ptl; |
1da177e4 LT |
730 | struct page *page; |
731 | unsigned long address; | |
732 | unsigned long end; | |
1da177e4 | 733 | |
1da177e4 LT |
734 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
735 | end = address + CLUSTER_SIZE; | |
736 | if (address < vma->vm_start) | |
737 | address = vma->vm_start; | |
738 | if (end > vma->vm_end) | |
739 | end = vma->vm_end; | |
740 | ||
741 | pgd = pgd_offset(mm, address); | |
742 | if (!pgd_present(*pgd)) | |
c0718806 | 743 | return; |
1da177e4 LT |
744 | |
745 | pud = pud_offset(pgd, address); | |
746 | if (!pud_present(*pud)) | |
c0718806 | 747 | return; |
1da177e4 LT |
748 | |
749 | pmd = pmd_offset(pud, address); | |
750 | if (!pmd_present(*pmd)) | |
c0718806 HD |
751 | return; |
752 | ||
753 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 754 | |
365e9c87 HD |
755 | /* Update high watermark before we lower rss */ |
756 | update_hiwater_rss(mm); | |
757 | ||
c0718806 | 758 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
759 | if (!pte_present(*pte)) |
760 | continue; | |
6aab341e LT |
761 | page = vm_normal_page(vma, address, *pte); |
762 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 LT |
763 | |
764 | if (ptep_clear_flush_young(vma, address, pte)) | |
765 | continue; | |
766 | ||
767 | /* Nuke the page table entry. */ | |
eca35133 | 768 | flush_cache_page(vma, address, pte_pfn(*pte)); |
1da177e4 LT |
769 | pteval = ptep_clear_flush(vma, address, pte); |
770 | ||
771 | /* If nonlinear, store the file page offset in the pte. */ | |
772 | if (page->index != linear_page_index(vma, address)) | |
773 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
774 | ||
775 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
776 | if (pte_dirty(pteval)) | |
777 | set_page_dirty(page); | |
778 | ||
7de6b805 | 779 | page_remove_rmap(page, vma); |
1da177e4 | 780 | page_cache_release(page); |
4294621f | 781 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
782 | (*mapcount)--; |
783 | } | |
c0718806 | 784 | pte_unmap_unlock(pte - 1, ptl); |
1da177e4 LT |
785 | } |
786 | ||
7352349a | 787 | static int try_to_unmap_anon(struct page *page, int migration) |
1da177e4 LT |
788 | { |
789 | struct anon_vma *anon_vma; | |
790 | struct vm_area_struct *vma; | |
791 | int ret = SWAP_AGAIN; | |
792 | ||
793 | anon_vma = page_lock_anon_vma(page); | |
794 | if (!anon_vma) | |
795 | return ret; | |
796 | ||
797 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
7352349a | 798 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
799 | if (ret == SWAP_FAIL || !page_mapped(page)) |
800 | break; | |
801 | } | |
802 | spin_unlock(&anon_vma->lock); | |
803 | return ret; | |
804 | } | |
805 | ||
806 | /** | |
807 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
808 | * @page: the page to unmap | |
809 | * | |
810 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
811 | * contained in the address_space struct it points to. | |
812 | * | |
813 | * This function is only called from try_to_unmap for object-based pages. | |
814 | */ | |
7352349a | 815 | static int try_to_unmap_file(struct page *page, int migration) |
1da177e4 LT |
816 | { |
817 | struct address_space *mapping = page->mapping; | |
818 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
819 | struct vm_area_struct *vma; | |
820 | struct prio_tree_iter iter; | |
821 | int ret = SWAP_AGAIN; | |
822 | unsigned long cursor; | |
823 | unsigned long max_nl_cursor = 0; | |
824 | unsigned long max_nl_size = 0; | |
825 | unsigned int mapcount; | |
826 | ||
827 | spin_lock(&mapping->i_mmap_lock); | |
828 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
7352349a | 829 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
830 | if (ret == SWAP_FAIL || !page_mapped(page)) |
831 | goto out; | |
832 | } | |
833 | ||
834 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
835 | goto out; | |
836 | ||
837 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
838 | shared.vm_set.list) { | |
e6a1530d | 839 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
840 | continue; |
841 | cursor = (unsigned long) vma->vm_private_data; | |
842 | if (cursor > max_nl_cursor) | |
843 | max_nl_cursor = cursor; | |
844 | cursor = vma->vm_end - vma->vm_start; | |
845 | if (cursor > max_nl_size) | |
846 | max_nl_size = cursor; | |
847 | } | |
848 | ||
849 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
850 | ret = SWAP_FAIL; | |
851 | goto out; | |
852 | } | |
853 | ||
854 | /* | |
855 | * We don't try to search for this page in the nonlinear vmas, | |
856 | * and page_referenced wouldn't have found it anyway. Instead | |
857 | * just walk the nonlinear vmas trying to age and unmap some. | |
858 | * The mapcount of the page we came in with is irrelevant, | |
859 | * but even so use it as a guide to how hard we should try? | |
860 | */ | |
861 | mapcount = page_mapcount(page); | |
862 | if (!mapcount) | |
863 | goto out; | |
864 | cond_resched_lock(&mapping->i_mmap_lock); | |
865 | ||
866 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
867 | if (max_nl_cursor == 0) | |
868 | max_nl_cursor = CLUSTER_SIZE; | |
869 | ||
870 | do { | |
871 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
872 | shared.vm_set.list) { | |
e6a1530d | 873 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
874 | continue; |
875 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 876 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
877 | cursor < vma->vm_end - vma->vm_start) { |
878 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
879 | cursor += CLUSTER_SIZE; | |
880 | vma->vm_private_data = (void *) cursor; | |
881 | if ((int)mapcount <= 0) | |
882 | goto out; | |
883 | } | |
884 | vma->vm_private_data = (void *) max_nl_cursor; | |
885 | } | |
886 | cond_resched_lock(&mapping->i_mmap_lock); | |
887 | max_nl_cursor += CLUSTER_SIZE; | |
888 | } while (max_nl_cursor <= max_nl_size); | |
889 | ||
890 | /* | |
891 | * Don't loop forever (perhaps all the remaining pages are | |
892 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
893 | * vmas, now forgetting on which ones it had fallen behind. | |
894 | */ | |
101d2be7 HD |
895 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
896 | vma->vm_private_data = NULL; | |
1da177e4 LT |
897 | out: |
898 | spin_unlock(&mapping->i_mmap_lock); | |
899 | return ret; | |
900 | } | |
901 | ||
902 | /** | |
903 | * try_to_unmap - try to remove all page table mappings to a page | |
904 | * @page: the page to get unmapped | |
905 | * | |
906 | * Tries to remove all the page table entries which are mapping this | |
907 | * page, used in the pageout path. Caller must hold the page lock. | |
908 | * Return values are: | |
909 | * | |
910 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
911 | * SWAP_AGAIN - we missed a mapping, try again later | |
912 | * SWAP_FAIL - the page is unswappable | |
913 | */ | |
7352349a | 914 | int try_to_unmap(struct page *page, int migration) |
1da177e4 LT |
915 | { |
916 | int ret; | |
917 | ||
1da177e4 LT |
918 | BUG_ON(!PageLocked(page)); |
919 | ||
920 | if (PageAnon(page)) | |
7352349a | 921 | ret = try_to_unmap_anon(page, migration); |
1da177e4 | 922 | else |
7352349a | 923 | ret = try_to_unmap_file(page, migration); |
1da177e4 LT |
924 | |
925 | if (!page_mapped(page)) | |
926 | ret = SWAP_SUCCESS; | |
927 | return ret; | |
928 | } | |
81b4082d | 929 |