]>
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 | ||
59 | //#define RMAP_DEBUG /* can be enabled only for debugging */ | |
60 | ||
fcc234f8 | 61 | struct kmem_cache *anon_vma_cachep; |
1da177e4 LT |
62 | |
63 | static inline void validate_anon_vma(struct vm_area_struct *find_vma) | |
64 | { | |
65 | #ifdef RMAP_DEBUG | |
66 | struct anon_vma *anon_vma = find_vma->anon_vma; | |
67 | struct vm_area_struct *vma; | |
68 | unsigned int mapcount = 0; | |
69 | int found = 0; | |
70 | ||
71 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
72 | mapcount++; | |
73 | BUG_ON(mapcount > 100000); | |
74 | if (vma == find_vma) | |
75 | found = 1; | |
76 | } | |
77 | BUG_ON(!found); | |
78 | #endif | |
79 | } | |
80 | ||
81 | /* This must be called under the mmap_sem. */ | |
82 | int anon_vma_prepare(struct vm_area_struct *vma) | |
83 | { | |
84 | struct anon_vma *anon_vma = vma->anon_vma; | |
85 | ||
86 | might_sleep(); | |
87 | if (unlikely(!anon_vma)) { | |
88 | struct mm_struct *mm = vma->vm_mm; | |
89 | struct anon_vma *allocated, *locked; | |
90 | ||
91 | anon_vma = find_mergeable_anon_vma(vma); | |
92 | if (anon_vma) { | |
93 | allocated = NULL; | |
94 | locked = anon_vma; | |
95 | spin_lock(&locked->lock); | |
96 | } else { | |
97 | anon_vma = anon_vma_alloc(); | |
98 | if (unlikely(!anon_vma)) | |
99 | return -ENOMEM; | |
100 | allocated = anon_vma; | |
101 | locked = NULL; | |
102 | } | |
103 | ||
104 | /* page_table_lock to protect against threads */ | |
105 | spin_lock(&mm->page_table_lock); | |
106 | if (likely(!vma->anon_vma)) { | |
107 | vma->anon_vma = anon_vma; | |
108 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
109 | allocated = NULL; | |
110 | } | |
111 | spin_unlock(&mm->page_table_lock); | |
112 | ||
113 | if (locked) | |
114 | spin_unlock(&locked->lock); | |
115 | if (unlikely(allocated)) | |
116 | anon_vma_free(allocated); | |
117 | } | |
118 | return 0; | |
119 | } | |
120 | ||
121 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) | |
122 | { | |
123 | BUG_ON(vma->anon_vma != next->anon_vma); | |
124 | list_del(&next->anon_vma_node); | |
125 | } | |
126 | ||
127 | void __anon_vma_link(struct vm_area_struct *vma) | |
128 | { | |
129 | struct anon_vma *anon_vma = vma->anon_vma; | |
130 | ||
131 | if (anon_vma) { | |
132 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
133 | validate_anon_vma(vma); | |
134 | } | |
135 | } | |
136 | ||
137 | void anon_vma_link(struct vm_area_struct *vma) | |
138 | { | |
139 | struct anon_vma *anon_vma = vma->anon_vma; | |
140 | ||
141 | if (anon_vma) { | |
142 | spin_lock(&anon_vma->lock); | |
143 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
144 | validate_anon_vma(vma); | |
145 | spin_unlock(&anon_vma->lock); | |
146 | } | |
147 | } | |
148 | ||
149 | void anon_vma_unlink(struct vm_area_struct *vma) | |
150 | { | |
151 | struct anon_vma *anon_vma = vma->anon_vma; | |
152 | int empty; | |
153 | ||
154 | if (!anon_vma) | |
155 | return; | |
156 | ||
157 | spin_lock(&anon_vma->lock); | |
158 | validate_anon_vma(vma); | |
159 | list_del(&vma->anon_vma_node); | |
160 | ||
161 | /* We must garbage collect the anon_vma if it's empty */ | |
162 | empty = list_empty(&anon_vma->head); | |
163 | spin_unlock(&anon_vma->lock); | |
164 | ||
165 | if (empty) | |
166 | anon_vma_free(anon_vma); | |
167 | } | |
168 | ||
fcc234f8 PE |
169 | static void anon_vma_ctor(void *data, struct kmem_cache *cachep, |
170 | unsigned long flags) | |
1da177e4 LT |
171 | { |
172 | if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == | |
173 | SLAB_CTOR_CONSTRUCTOR) { | |
174 | struct anon_vma *anon_vma = data; | |
175 | ||
176 | spin_lock_init(&anon_vma->lock); | |
177 | INIT_LIST_HEAD(&anon_vma->head); | |
178 | } | |
179 | } | |
180 | ||
181 | void __init anon_vma_init(void) | |
182 | { | |
183 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
184 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL); | |
185 | } | |
186 | ||
187 | /* | |
188 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
189 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
190 | */ | |
191 | static struct anon_vma *page_lock_anon_vma(struct page *page) | |
192 | { | |
193 | struct anon_vma *anon_vma = NULL; | |
194 | unsigned long anon_mapping; | |
195 | ||
196 | rcu_read_lock(); | |
197 | anon_mapping = (unsigned long) page->mapping; | |
198 | if (!(anon_mapping & PAGE_MAPPING_ANON)) | |
199 | goto out; | |
200 | if (!page_mapped(page)) | |
201 | goto out; | |
202 | ||
203 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
204 | spin_lock(&anon_vma->lock); | |
205 | out: | |
206 | rcu_read_unlock(); | |
207 | return anon_vma; | |
208 | } | |
209 | ||
a3351e52 CL |
210 | #ifdef CONFIG_MIGRATION |
211 | /* | |
212 | * Remove an anonymous page from swap replacing the swap pte's | |
213 | * through real pte's pointing to valid pages and then releasing | |
214 | * the page from the swap cache. | |
215 | * | |
e8788c0c CL |
216 | * Must hold page lock on page and mmap_sem of one vma that contains |
217 | * the page. | |
a3351e52 CL |
218 | */ |
219 | void remove_from_swap(struct page *page) | |
220 | { | |
221 | struct anon_vma *anon_vma; | |
222 | struct vm_area_struct *vma; | |
e8788c0c | 223 | unsigned long mapping; |
a3351e52 | 224 | |
e8788c0c | 225 | if (!PageSwapCache(page)) |
a3351e52 CL |
226 | return; |
227 | ||
e8788c0c CL |
228 | mapping = (unsigned long)page->mapping; |
229 | ||
230 | if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) | |
a3351e52 CL |
231 | return; |
232 | ||
e8788c0c CL |
233 | /* |
234 | * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. | |
235 | */ | |
236 | anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); | |
237 | spin_lock(&anon_vma->lock); | |
238 | ||
a3351e52 CL |
239 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) |
240 | remove_vma_swap(vma, page); | |
241 | ||
242 | spin_unlock(&anon_vma->lock); | |
a3351e52 CL |
243 | delete_from_swap_cache(page); |
244 | } | |
e965f963 | 245 | EXPORT_SYMBOL(remove_from_swap); |
a3351e52 CL |
246 | #endif |
247 | ||
1da177e4 LT |
248 | /* |
249 | * At what user virtual address is page expected in vma? | |
250 | */ | |
251 | static inline unsigned long | |
252 | vma_address(struct page *page, struct vm_area_struct *vma) | |
253 | { | |
254 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
255 | unsigned long address; | |
256 | ||
257 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
258 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
259 | /* page should be within any vma from prio_tree_next */ | |
260 | BUG_ON(!PageAnon(page)); | |
261 | return -EFAULT; | |
262 | } | |
263 | return address; | |
264 | } | |
265 | ||
266 | /* | |
267 | * At what user virtual address is page expected in vma? checking that the | |
ee498ed7 | 268 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
1da177e4 LT |
269 | */ |
270 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
271 | { | |
272 | if (PageAnon(page)) { | |
273 | if ((void *)vma->anon_vma != | |
274 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
275 | return -EFAULT; | |
276 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
277 | if (!vma->vm_file || |
278 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
279 | return -EFAULT; |
280 | } else | |
281 | return -EFAULT; | |
282 | return vma_address(page, vma); | |
283 | } | |
284 | ||
81b4082d ND |
285 | /* |
286 | * Check that @page is mapped at @address into @mm. | |
287 | * | |
b8072f09 | 288 | * On success returns with pte mapped and locked. |
81b4082d | 289 | */ |
ceffc078 | 290 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
c0718806 | 291 | unsigned long address, spinlock_t **ptlp) |
81b4082d ND |
292 | { |
293 | pgd_t *pgd; | |
294 | pud_t *pud; | |
295 | pmd_t *pmd; | |
296 | pte_t *pte; | |
c0718806 | 297 | spinlock_t *ptl; |
81b4082d | 298 | |
81b4082d | 299 | pgd = pgd_offset(mm, address); |
c0718806 HD |
300 | if (!pgd_present(*pgd)) |
301 | return NULL; | |
302 | ||
303 | pud = pud_offset(pgd, address); | |
304 | if (!pud_present(*pud)) | |
305 | return NULL; | |
306 | ||
307 | pmd = pmd_offset(pud, address); | |
308 | if (!pmd_present(*pmd)) | |
309 | return NULL; | |
310 | ||
311 | pte = pte_offset_map(pmd, address); | |
312 | /* Make a quick check before getting the lock */ | |
313 | if (!pte_present(*pte)) { | |
314 | pte_unmap(pte); | |
315 | return NULL; | |
316 | } | |
317 | ||
4c21e2f2 | 318 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
319 | spin_lock(ptl); |
320 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
321 | *ptlp = ptl; | |
322 | return pte; | |
81b4082d | 323 | } |
c0718806 HD |
324 | pte_unmap_unlock(pte, ptl); |
325 | return NULL; | |
81b4082d ND |
326 | } |
327 | ||
1da177e4 LT |
328 | /* |
329 | * Subfunctions of page_referenced: page_referenced_one called | |
330 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
331 | */ | |
332 | static int page_referenced_one(struct page *page, | |
f7b7fd8f | 333 | struct vm_area_struct *vma, unsigned int *mapcount) |
1da177e4 LT |
334 | { |
335 | struct mm_struct *mm = vma->vm_mm; | |
336 | unsigned long address; | |
1da177e4 | 337 | pte_t *pte; |
c0718806 | 338 | spinlock_t *ptl; |
1da177e4 LT |
339 | int referenced = 0; |
340 | ||
1da177e4 LT |
341 | address = vma_address(page, vma); |
342 | if (address == -EFAULT) | |
343 | goto out; | |
344 | ||
c0718806 HD |
345 | pte = page_check_address(page, mm, address, &ptl); |
346 | if (!pte) | |
347 | goto out; | |
1da177e4 | 348 | |
c0718806 HD |
349 | if (ptep_clear_flush_young(vma, address, pte)) |
350 | referenced++; | |
1da177e4 | 351 | |
c0718806 HD |
352 | /* Pretend the page is referenced if the task has the |
353 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 354 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
355 | rwsem_is_locked(&mm->mmap_sem)) |
356 | referenced++; | |
357 | ||
358 | (*mapcount)--; | |
359 | pte_unmap_unlock(pte, ptl); | |
1da177e4 LT |
360 | out: |
361 | return referenced; | |
362 | } | |
363 | ||
f7b7fd8f | 364 | static int page_referenced_anon(struct page *page) |
1da177e4 LT |
365 | { |
366 | unsigned int mapcount; | |
367 | struct anon_vma *anon_vma; | |
368 | struct vm_area_struct *vma; | |
369 | int referenced = 0; | |
370 | ||
371 | anon_vma = page_lock_anon_vma(page); | |
372 | if (!anon_vma) | |
373 | return referenced; | |
374 | ||
375 | mapcount = page_mapcount(page); | |
376 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
f7b7fd8f | 377 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
378 | if (!mapcount) |
379 | break; | |
380 | } | |
381 | spin_unlock(&anon_vma->lock); | |
382 | return referenced; | |
383 | } | |
384 | ||
385 | /** | |
386 | * page_referenced_file - referenced check for object-based rmap | |
387 | * @page: the page we're checking references on. | |
388 | * | |
389 | * For an object-based mapped page, find all the places it is mapped and | |
390 | * check/clear the referenced flag. This is done by following the page->mapping | |
391 | * pointer, then walking the chain of vmas it holds. It returns the number | |
392 | * of references it found. | |
393 | * | |
394 | * This function is only called from page_referenced for object-based pages. | |
395 | */ | |
f7b7fd8f | 396 | static int page_referenced_file(struct page *page) |
1da177e4 LT |
397 | { |
398 | unsigned int mapcount; | |
399 | struct address_space *mapping = page->mapping; | |
400 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
401 | struct vm_area_struct *vma; | |
402 | struct prio_tree_iter iter; | |
403 | int referenced = 0; | |
404 | ||
405 | /* | |
406 | * The caller's checks on page->mapping and !PageAnon have made | |
407 | * sure that this is a file page: the check for page->mapping | |
408 | * excludes the case just before it gets set on an anon page. | |
409 | */ | |
410 | BUG_ON(PageAnon(page)); | |
411 | ||
412 | /* | |
413 | * The page lock not only makes sure that page->mapping cannot | |
414 | * suddenly be NULLified by truncation, it makes sure that the | |
415 | * structure at mapping cannot be freed and reused yet, | |
416 | * so we can safely take mapping->i_mmap_lock. | |
417 | */ | |
418 | BUG_ON(!PageLocked(page)); | |
419 | ||
420 | spin_lock(&mapping->i_mmap_lock); | |
421 | ||
422 | /* | |
423 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
424 | * is more likely to be accurate if we note it after spinning. | |
425 | */ | |
426 | mapcount = page_mapcount(page); | |
427 | ||
428 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
429 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
430 | == (VM_LOCKED|VM_MAYSHARE)) { | |
431 | referenced++; | |
432 | break; | |
433 | } | |
f7b7fd8f | 434 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
435 | if (!mapcount) |
436 | break; | |
437 | } | |
438 | ||
439 | spin_unlock(&mapping->i_mmap_lock); | |
440 | return referenced; | |
441 | } | |
442 | ||
443 | /** | |
444 | * page_referenced - test if the page was referenced | |
445 | * @page: the page to test | |
446 | * @is_locked: caller holds lock on the page | |
447 | * | |
448 | * Quick test_and_clear_referenced for all mappings to a page, | |
449 | * returns the number of ptes which referenced the page. | |
450 | */ | |
f7b7fd8f | 451 | int page_referenced(struct page *page, int is_locked) |
1da177e4 LT |
452 | { |
453 | int referenced = 0; | |
454 | ||
1da177e4 LT |
455 | if (page_test_and_clear_young(page)) |
456 | referenced++; | |
457 | ||
458 | if (TestClearPageReferenced(page)) | |
459 | referenced++; | |
460 | ||
461 | if (page_mapped(page) && page->mapping) { | |
462 | if (PageAnon(page)) | |
f7b7fd8f | 463 | referenced += page_referenced_anon(page); |
1da177e4 | 464 | else if (is_locked) |
f7b7fd8f | 465 | referenced += page_referenced_file(page); |
1da177e4 LT |
466 | else if (TestSetPageLocked(page)) |
467 | referenced++; | |
468 | else { | |
469 | if (page->mapping) | |
f7b7fd8f | 470 | referenced += page_referenced_file(page); |
1da177e4 LT |
471 | unlock_page(page); |
472 | } | |
473 | } | |
474 | return referenced; | |
475 | } | |
476 | ||
9617d95e NP |
477 | /** |
478 | * page_set_anon_rmap - setup new anonymous rmap | |
479 | * @page: the page to add the mapping to | |
480 | * @vma: the vm area in which the mapping is added | |
481 | * @address: the user virtual address mapped | |
482 | */ | |
483 | static void __page_set_anon_rmap(struct page *page, | |
484 | struct vm_area_struct *vma, unsigned long address) | |
485 | { | |
486 | struct anon_vma *anon_vma = vma->anon_vma; | |
487 | ||
488 | BUG_ON(!anon_vma); | |
489 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
490 | page->mapping = (struct address_space *) anon_vma; | |
491 | ||
492 | page->index = linear_page_index(vma, address); | |
493 | ||
a74609fa NP |
494 | /* |
495 | * nr_mapped state can be updated without turning off | |
496 | * interrupts because it is not modified via interrupt. | |
497 | */ | |
498 | __inc_page_state(nr_mapped); | |
9617d95e NP |
499 | } |
500 | ||
1da177e4 LT |
501 | /** |
502 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
503 | * @page: the page to add the mapping to | |
504 | * @vma: the vm area in which the mapping is added | |
505 | * @address: the user virtual address mapped | |
506 | * | |
b8072f09 | 507 | * The caller needs to hold the pte lock. |
1da177e4 LT |
508 | */ |
509 | void page_add_anon_rmap(struct page *page, | |
510 | struct vm_area_struct *vma, unsigned long address) | |
511 | { | |
9617d95e NP |
512 | if (atomic_inc_and_test(&page->_mapcount)) |
513 | __page_set_anon_rmap(page, vma, address); | |
1da177e4 LT |
514 | /* else checking page index and mapping is racy */ |
515 | } | |
516 | ||
9617d95e NP |
517 | /* |
518 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
519 | * @page: the page to add the mapping to | |
520 | * @vma: the vm area in which the mapping is added | |
521 | * @address: the user virtual address mapped | |
522 | * | |
523 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
524 | * This means the inc-and-test can be bypassed. | |
525 | */ | |
526 | void page_add_new_anon_rmap(struct page *page, | |
527 | struct vm_area_struct *vma, unsigned long address) | |
528 | { | |
529 | atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ | |
530 | __page_set_anon_rmap(page, vma, address); | |
531 | } | |
532 | ||
1da177e4 LT |
533 | /** |
534 | * page_add_file_rmap - add pte mapping to a file page | |
535 | * @page: the page to add the mapping to | |
536 | * | |
b8072f09 | 537 | * The caller needs to hold the pte lock. |
1da177e4 LT |
538 | */ |
539 | void page_add_file_rmap(struct page *page) | |
540 | { | |
1da177e4 | 541 | if (atomic_inc_and_test(&page->_mapcount)) |
a74609fa | 542 | __inc_page_state(nr_mapped); |
1da177e4 LT |
543 | } |
544 | ||
545 | /** | |
546 | * page_remove_rmap - take down pte mapping from a page | |
547 | * @page: page to remove mapping from | |
548 | * | |
b8072f09 | 549 | * The caller needs to hold the pte lock. |
1da177e4 LT |
550 | */ |
551 | void page_remove_rmap(struct page *page) | |
552 | { | |
1da177e4 | 553 | if (atomic_add_negative(-1, &page->_mapcount)) { |
ef2bf0dc DJ |
554 | if (page_mapcount(page) < 0) { |
555 | printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page)); | |
556 | printk (KERN_EMERG " page->flags = %lx\n", page->flags); | |
557 | printk (KERN_EMERG " page->count = %x\n", page_count(page)); | |
558 | printk (KERN_EMERG " page->mapping = %p\n", page->mapping); | |
559 | } | |
560 | ||
1da177e4 LT |
561 | BUG_ON(page_mapcount(page) < 0); |
562 | /* | |
563 | * It would be tidy to reset the PageAnon mapping here, | |
564 | * but that might overwrite a racing page_add_anon_rmap | |
565 | * which increments mapcount after us but sets mapping | |
566 | * before us: so leave the reset to free_hot_cold_page, | |
567 | * and remember that it's only reliable while mapped. | |
568 | * Leaving it set also helps swapoff to reinstate ptes | |
569 | * faster for those pages still in swapcache. | |
570 | */ | |
571 | if (page_test_and_clear_dirty(page)) | |
572 | set_page_dirty(page); | |
a74609fa | 573 | __dec_page_state(nr_mapped); |
1da177e4 LT |
574 | } |
575 | } | |
576 | ||
577 | /* | |
578 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
579 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
580 | */ | |
a48d07af CL |
581 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
582 | int ignore_refs) | |
1da177e4 LT |
583 | { |
584 | struct mm_struct *mm = vma->vm_mm; | |
585 | unsigned long address; | |
1da177e4 LT |
586 | pte_t *pte; |
587 | pte_t pteval; | |
c0718806 | 588 | spinlock_t *ptl; |
1da177e4 LT |
589 | int ret = SWAP_AGAIN; |
590 | ||
1da177e4 LT |
591 | address = vma_address(page, vma); |
592 | if (address == -EFAULT) | |
593 | goto out; | |
594 | ||
c0718806 HD |
595 | pte = page_check_address(page, mm, address, &ptl); |
596 | if (!pte) | |
81b4082d | 597 | goto out; |
1da177e4 LT |
598 | |
599 | /* | |
600 | * If the page is mlock()d, we cannot swap it out. | |
601 | * If it's recently referenced (perhaps page_referenced | |
602 | * skipped over this mm) then we should reactivate it. | |
603 | */ | |
101d2be7 | 604 | if ((vma->vm_flags & VM_LOCKED) || |
a48d07af CL |
605 | (ptep_clear_flush_young(vma, address, pte) |
606 | && !ignore_refs)) { | |
1da177e4 LT |
607 | ret = SWAP_FAIL; |
608 | goto out_unmap; | |
609 | } | |
610 | ||
1da177e4 LT |
611 | /* Nuke the page table entry. */ |
612 | flush_cache_page(vma, address, page_to_pfn(page)); | |
613 | pteval = ptep_clear_flush(vma, address, pte); | |
614 | ||
615 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
616 | if (pte_dirty(pteval)) | |
617 | set_page_dirty(page); | |
618 | ||
365e9c87 HD |
619 | /* Update high watermark before we lower rss */ |
620 | update_hiwater_rss(mm); | |
621 | ||
1da177e4 | 622 | if (PageAnon(page)) { |
4c21e2f2 | 623 | swp_entry_t entry = { .val = page_private(page) }; |
1da177e4 LT |
624 | /* |
625 | * Store the swap location in the pte. | |
626 | * See handle_pte_fault() ... | |
627 | */ | |
628 | BUG_ON(!PageSwapCache(page)); | |
629 | swap_duplicate(entry); | |
630 | if (list_empty(&mm->mmlist)) { | |
631 | spin_lock(&mmlist_lock); | |
f412ac08 HD |
632 | if (list_empty(&mm->mmlist)) |
633 | list_add(&mm->mmlist, &init_mm.mmlist); | |
1da177e4 LT |
634 | spin_unlock(&mmlist_lock); |
635 | } | |
636 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
637 | BUG_ON(pte_file(*pte)); | |
638 | dec_mm_counter(mm, anon_rss); | |
4294621f HD |
639 | } else |
640 | dec_mm_counter(mm, file_rss); | |
1da177e4 | 641 | |
1da177e4 LT |
642 | page_remove_rmap(page); |
643 | page_cache_release(page); | |
644 | ||
645 | out_unmap: | |
c0718806 | 646 | pte_unmap_unlock(pte, ptl); |
1da177e4 LT |
647 | out: |
648 | return ret; | |
649 | } | |
650 | ||
651 | /* | |
652 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
653 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
654 | * Consequently, given a particular page and its ->index, we cannot locate the | |
655 | * ptes which are mapping that page without an exhaustive linear search. | |
656 | * | |
657 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
658 | * maps the file to which the target page belongs. The ->vm_private_data field | |
659 | * holds the current cursor into that scan. Successive searches will circulate | |
660 | * around the vma's virtual address space. | |
661 | * | |
662 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
663 | * more scanning pressure is placed against them as well. Eventually pages | |
664 | * will become fully unmapped and are eligible for eviction. | |
665 | * | |
666 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
667 | * there there won't be many ptes located within the scan cluster. In this case | |
668 | * maybe we could scan further - to the end of the pte page, perhaps. | |
669 | */ | |
670 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
671 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
672 | ||
673 | static void try_to_unmap_cluster(unsigned long cursor, | |
674 | unsigned int *mapcount, struct vm_area_struct *vma) | |
675 | { | |
676 | struct mm_struct *mm = vma->vm_mm; | |
677 | pgd_t *pgd; | |
678 | pud_t *pud; | |
679 | pmd_t *pmd; | |
c0718806 | 680 | pte_t *pte; |
1da177e4 | 681 | pte_t pteval; |
c0718806 | 682 | spinlock_t *ptl; |
1da177e4 LT |
683 | struct page *page; |
684 | unsigned long address; | |
685 | unsigned long end; | |
1da177e4 | 686 | |
1da177e4 LT |
687 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
688 | end = address + CLUSTER_SIZE; | |
689 | if (address < vma->vm_start) | |
690 | address = vma->vm_start; | |
691 | if (end > vma->vm_end) | |
692 | end = vma->vm_end; | |
693 | ||
694 | pgd = pgd_offset(mm, address); | |
695 | if (!pgd_present(*pgd)) | |
c0718806 | 696 | return; |
1da177e4 LT |
697 | |
698 | pud = pud_offset(pgd, address); | |
699 | if (!pud_present(*pud)) | |
c0718806 | 700 | return; |
1da177e4 LT |
701 | |
702 | pmd = pmd_offset(pud, address); | |
703 | if (!pmd_present(*pmd)) | |
c0718806 HD |
704 | return; |
705 | ||
706 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 707 | |
365e9c87 HD |
708 | /* Update high watermark before we lower rss */ |
709 | update_hiwater_rss(mm); | |
710 | ||
c0718806 | 711 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
712 | if (!pte_present(*pte)) |
713 | continue; | |
6aab341e LT |
714 | page = vm_normal_page(vma, address, *pte); |
715 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 LT |
716 | |
717 | if (ptep_clear_flush_young(vma, address, pte)) | |
718 | continue; | |
719 | ||
720 | /* Nuke the page table entry. */ | |
eca35133 | 721 | flush_cache_page(vma, address, pte_pfn(*pte)); |
1da177e4 LT |
722 | pteval = ptep_clear_flush(vma, address, pte); |
723 | ||
724 | /* If nonlinear, store the file page offset in the pte. */ | |
725 | if (page->index != linear_page_index(vma, address)) | |
726 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
727 | ||
728 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
729 | if (pte_dirty(pteval)) | |
730 | set_page_dirty(page); | |
731 | ||
732 | page_remove_rmap(page); | |
733 | page_cache_release(page); | |
4294621f | 734 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
735 | (*mapcount)--; |
736 | } | |
c0718806 | 737 | pte_unmap_unlock(pte - 1, ptl); |
1da177e4 LT |
738 | } |
739 | ||
a48d07af | 740 | static int try_to_unmap_anon(struct page *page, int ignore_refs) |
1da177e4 LT |
741 | { |
742 | struct anon_vma *anon_vma; | |
743 | struct vm_area_struct *vma; | |
744 | int ret = SWAP_AGAIN; | |
745 | ||
746 | anon_vma = page_lock_anon_vma(page); | |
747 | if (!anon_vma) | |
748 | return ret; | |
749 | ||
750 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
a48d07af | 751 | ret = try_to_unmap_one(page, vma, ignore_refs); |
1da177e4 LT |
752 | if (ret == SWAP_FAIL || !page_mapped(page)) |
753 | break; | |
754 | } | |
755 | spin_unlock(&anon_vma->lock); | |
756 | return ret; | |
757 | } | |
758 | ||
759 | /** | |
760 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
761 | * @page: the page to unmap | |
762 | * | |
763 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
764 | * contained in the address_space struct it points to. | |
765 | * | |
766 | * This function is only called from try_to_unmap for object-based pages. | |
767 | */ | |
a48d07af | 768 | static int try_to_unmap_file(struct page *page, int ignore_refs) |
1da177e4 LT |
769 | { |
770 | struct address_space *mapping = page->mapping; | |
771 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
772 | struct vm_area_struct *vma; | |
773 | struct prio_tree_iter iter; | |
774 | int ret = SWAP_AGAIN; | |
775 | unsigned long cursor; | |
776 | unsigned long max_nl_cursor = 0; | |
777 | unsigned long max_nl_size = 0; | |
778 | unsigned int mapcount; | |
779 | ||
780 | spin_lock(&mapping->i_mmap_lock); | |
781 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
a48d07af | 782 | ret = try_to_unmap_one(page, vma, ignore_refs); |
1da177e4 LT |
783 | if (ret == SWAP_FAIL || !page_mapped(page)) |
784 | goto out; | |
785 | } | |
786 | ||
787 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
788 | goto out; | |
789 | ||
790 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
791 | shared.vm_set.list) { | |
101d2be7 | 792 | if (vma->vm_flags & VM_LOCKED) |
1da177e4 LT |
793 | continue; |
794 | cursor = (unsigned long) vma->vm_private_data; | |
795 | if (cursor > max_nl_cursor) | |
796 | max_nl_cursor = cursor; | |
797 | cursor = vma->vm_end - vma->vm_start; | |
798 | if (cursor > max_nl_size) | |
799 | max_nl_size = cursor; | |
800 | } | |
801 | ||
802 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
803 | ret = SWAP_FAIL; | |
804 | goto out; | |
805 | } | |
806 | ||
807 | /* | |
808 | * We don't try to search for this page in the nonlinear vmas, | |
809 | * and page_referenced wouldn't have found it anyway. Instead | |
810 | * just walk the nonlinear vmas trying to age and unmap some. | |
811 | * The mapcount of the page we came in with is irrelevant, | |
812 | * but even so use it as a guide to how hard we should try? | |
813 | */ | |
814 | mapcount = page_mapcount(page); | |
815 | if (!mapcount) | |
816 | goto out; | |
817 | cond_resched_lock(&mapping->i_mmap_lock); | |
818 | ||
819 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
820 | if (max_nl_cursor == 0) | |
821 | max_nl_cursor = CLUSTER_SIZE; | |
822 | ||
823 | do { | |
824 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
825 | shared.vm_set.list) { | |
101d2be7 | 826 | if (vma->vm_flags & VM_LOCKED) |
1da177e4 LT |
827 | continue; |
828 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 829 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
830 | cursor < vma->vm_end - vma->vm_start) { |
831 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
832 | cursor += CLUSTER_SIZE; | |
833 | vma->vm_private_data = (void *) cursor; | |
834 | if ((int)mapcount <= 0) | |
835 | goto out; | |
836 | } | |
837 | vma->vm_private_data = (void *) max_nl_cursor; | |
838 | } | |
839 | cond_resched_lock(&mapping->i_mmap_lock); | |
840 | max_nl_cursor += CLUSTER_SIZE; | |
841 | } while (max_nl_cursor <= max_nl_size); | |
842 | ||
843 | /* | |
844 | * Don't loop forever (perhaps all the remaining pages are | |
845 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
846 | * vmas, now forgetting on which ones it had fallen behind. | |
847 | */ | |
101d2be7 HD |
848 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
849 | vma->vm_private_data = NULL; | |
1da177e4 LT |
850 | out: |
851 | spin_unlock(&mapping->i_mmap_lock); | |
852 | return ret; | |
853 | } | |
854 | ||
855 | /** | |
856 | * try_to_unmap - try to remove all page table mappings to a page | |
857 | * @page: the page to get unmapped | |
858 | * | |
859 | * Tries to remove all the page table entries which are mapping this | |
860 | * page, used in the pageout path. Caller must hold the page lock. | |
861 | * Return values are: | |
862 | * | |
863 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
864 | * SWAP_AGAIN - we missed a mapping, try again later | |
865 | * SWAP_FAIL - the page is unswappable | |
866 | */ | |
a48d07af | 867 | int try_to_unmap(struct page *page, int ignore_refs) |
1da177e4 LT |
868 | { |
869 | int ret; | |
870 | ||
1da177e4 LT |
871 | BUG_ON(!PageLocked(page)); |
872 | ||
873 | if (PageAnon(page)) | |
a48d07af | 874 | ret = try_to_unmap_anon(page, ignore_refs); |
1da177e4 | 875 | else |
a48d07af | 876 | ret = try_to_unmap_file(page, ignore_refs); |
1da177e4 LT |
877 | |
878 | if (!page_mapped(page)) | |
879 | ret = SWAP_SUCCESS; | |
880 | return ret; | |
881 | } | |
81b4082d | 882 |