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 | * | |
23 | * inode->i_sem (while writing or truncating, not reading or faulting) | |
24 | * inode->i_alloc_sem | |
25 | * | |
26 | * When a page fault occurs in writing from user to file, down_read | |
27 | * of mmap_sem nests within i_sem; in sys_msync, i_sem nests within | |
28 | * down_read of mmap_sem; i_sem and down_write of mmap_sem are never | |
29 | * taken together; in truncation, i_sem is taken outermost. | |
30 | * | |
31 | * mm->mmap_sem | |
32 | * page->flags PG_locked (lock_page) | |
33 | * mapping->i_mmap_lock | |
34 | * anon_vma->lock | |
35 | * mm->page_table_lock | |
36 | * zone->lru_lock (in mark_page_accessed) | |
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> | |
55 | ||
56 | #include <asm/tlbflush.h> | |
57 | ||
58 | //#define RMAP_DEBUG /* can be enabled only for debugging */ | |
59 | ||
60 | kmem_cache_t *anon_vma_cachep; | |
61 | ||
62 | static inline void validate_anon_vma(struct vm_area_struct *find_vma) | |
63 | { | |
64 | #ifdef RMAP_DEBUG | |
65 | struct anon_vma *anon_vma = find_vma->anon_vma; | |
66 | struct vm_area_struct *vma; | |
67 | unsigned int mapcount = 0; | |
68 | int found = 0; | |
69 | ||
70 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
71 | mapcount++; | |
72 | BUG_ON(mapcount > 100000); | |
73 | if (vma == find_vma) | |
74 | found = 1; | |
75 | } | |
76 | BUG_ON(!found); | |
77 | #endif | |
78 | } | |
79 | ||
80 | /* This must be called under the mmap_sem. */ | |
81 | int anon_vma_prepare(struct vm_area_struct *vma) | |
82 | { | |
83 | struct anon_vma *anon_vma = vma->anon_vma; | |
84 | ||
85 | might_sleep(); | |
86 | if (unlikely(!anon_vma)) { | |
87 | struct mm_struct *mm = vma->vm_mm; | |
88 | struct anon_vma *allocated, *locked; | |
89 | ||
90 | anon_vma = find_mergeable_anon_vma(vma); | |
91 | if (anon_vma) { | |
92 | allocated = NULL; | |
93 | locked = anon_vma; | |
94 | spin_lock(&locked->lock); | |
95 | } else { | |
96 | anon_vma = anon_vma_alloc(); | |
97 | if (unlikely(!anon_vma)) | |
98 | return -ENOMEM; | |
99 | allocated = anon_vma; | |
100 | locked = NULL; | |
101 | } | |
102 | ||
103 | /* page_table_lock to protect against threads */ | |
104 | spin_lock(&mm->page_table_lock); | |
105 | if (likely(!vma->anon_vma)) { | |
106 | vma->anon_vma = anon_vma; | |
107 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
108 | allocated = NULL; | |
109 | } | |
110 | spin_unlock(&mm->page_table_lock); | |
111 | ||
112 | if (locked) | |
113 | spin_unlock(&locked->lock); | |
114 | if (unlikely(allocated)) | |
115 | anon_vma_free(allocated); | |
116 | } | |
117 | return 0; | |
118 | } | |
119 | ||
120 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) | |
121 | { | |
122 | BUG_ON(vma->anon_vma != next->anon_vma); | |
123 | list_del(&next->anon_vma_node); | |
124 | } | |
125 | ||
126 | void __anon_vma_link(struct vm_area_struct *vma) | |
127 | { | |
128 | struct anon_vma *anon_vma = vma->anon_vma; | |
129 | ||
130 | if (anon_vma) { | |
131 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
132 | validate_anon_vma(vma); | |
133 | } | |
134 | } | |
135 | ||
136 | void anon_vma_link(struct vm_area_struct *vma) | |
137 | { | |
138 | struct anon_vma *anon_vma = vma->anon_vma; | |
139 | ||
140 | if (anon_vma) { | |
141 | spin_lock(&anon_vma->lock); | |
142 | list_add(&vma->anon_vma_node, &anon_vma->head); | |
143 | validate_anon_vma(vma); | |
144 | spin_unlock(&anon_vma->lock); | |
145 | } | |
146 | } | |
147 | ||
148 | void anon_vma_unlink(struct vm_area_struct *vma) | |
149 | { | |
150 | struct anon_vma *anon_vma = vma->anon_vma; | |
151 | int empty; | |
152 | ||
153 | if (!anon_vma) | |
154 | return; | |
155 | ||
156 | spin_lock(&anon_vma->lock); | |
157 | validate_anon_vma(vma); | |
158 | list_del(&vma->anon_vma_node); | |
159 | ||
160 | /* We must garbage collect the anon_vma if it's empty */ | |
161 | empty = list_empty(&anon_vma->head); | |
162 | spin_unlock(&anon_vma->lock); | |
163 | ||
164 | if (empty) | |
165 | anon_vma_free(anon_vma); | |
166 | } | |
167 | ||
168 | static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags) | |
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 | ||
208 | /* | |
209 | * At what user virtual address is page expected in vma? | |
210 | */ | |
211 | static inline unsigned long | |
212 | vma_address(struct page *page, struct vm_area_struct *vma) | |
213 | { | |
214 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
215 | unsigned long address; | |
216 | ||
217 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
218 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
219 | /* page should be within any vma from prio_tree_next */ | |
220 | BUG_ON(!PageAnon(page)); | |
221 | return -EFAULT; | |
222 | } | |
223 | return address; | |
224 | } | |
225 | ||
226 | /* | |
227 | * At what user virtual address is page expected in vma? checking that the | |
228 | * page matches the vma: currently only used by unuse_process, on anon pages. | |
229 | */ | |
230 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
231 | { | |
232 | if (PageAnon(page)) { | |
233 | if ((void *)vma->anon_vma != | |
234 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
235 | return -EFAULT; | |
236 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
237 | if (vma->vm_file->f_mapping != page->mapping) | |
238 | return -EFAULT; | |
239 | } else | |
240 | return -EFAULT; | |
241 | return vma_address(page, vma); | |
242 | } | |
243 | ||
81b4082d ND |
244 | /* |
245 | * Check that @page is mapped at @address into @mm. | |
246 | * | |
247 | * On success returns with mapped pte and locked mm->page_table_lock. | |
248 | */ | |
ceffc078 CO |
249 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
250 | unsigned long address) | |
81b4082d ND |
251 | { |
252 | pgd_t *pgd; | |
253 | pud_t *pud; | |
254 | pmd_t *pmd; | |
255 | pte_t *pte; | |
256 | ||
257 | /* | |
258 | * We need the page_table_lock to protect us from page faults, | |
259 | * munmap, fork, etc... | |
260 | */ | |
261 | spin_lock(&mm->page_table_lock); | |
262 | pgd = pgd_offset(mm, address); | |
263 | if (likely(pgd_present(*pgd))) { | |
264 | pud = pud_offset(pgd, address); | |
265 | if (likely(pud_present(*pud))) { | |
266 | pmd = pmd_offset(pud, address); | |
267 | if (likely(pmd_present(*pmd))) { | |
268 | pte = pte_offset_map(pmd, address); | |
269 | if (likely(pte_present(*pte) && | |
270 | page_to_pfn(page) == pte_pfn(*pte))) | |
271 | return pte; | |
272 | pte_unmap(pte); | |
273 | } | |
274 | } | |
275 | } | |
276 | spin_unlock(&mm->page_table_lock); | |
277 | return ERR_PTR(-ENOENT); | |
278 | } | |
279 | ||
1da177e4 LT |
280 | /* |
281 | * Subfunctions of page_referenced: page_referenced_one called | |
282 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
283 | */ | |
284 | static int page_referenced_one(struct page *page, | |
285 | struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token) | |
286 | { | |
287 | struct mm_struct *mm = vma->vm_mm; | |
288 | unsigned long address; | |
1da177e4 LT |
289 | pte_t *pte; |
290 | int referenced = 0; | |
291 | ||
1da177e4 LT |
292 | address = vma_address(page, vma); |
293 | if (address == -EFAULT) | |
294 | goto out; | |
295 | ||
81b4082d ND |
296 | pte = page_check_address(page, mm, address); |
297 | if (!IS_ERR(pte)) { | |
298 | if (ptep_clear_flush_young(vma, address, pte)) | |
299 | referenced++; | |
1da177e4 | 300 | |
fcdae29a RVR |
301 | /* Pretend the page is referenced if the task has the |
302 | swap token and is in the middle of a page fault. */ | |
303 | if (mm != current->mm && !ignore_token && | |
304 | has_swap_token(mm) && | |
305 | rwsem_is_locked(&mm->mmap_sem)) | |
81b4082d | 306 | referenced++; |
1da177e4 | 307 | |
81b4082d ND |
308 | (*mapcount)--; |
309 | pte_unmap(pte); | |
310 | spin_unlock(&mm->page_table_lock); | |
311 | } | |
1da177e4 LT |
312 | out: |
313 | return referenced; | |
314 | } | |
315 | ||
316 | static int page_referenced_anon(struct page *page, int ignore_token) | |
317 | { | |
318 | unsigned int mapcount; | |
319 | struct anon_vma *anon_vma; | |
320 | struct vm_area_struct *vma; | |
321 | int referenced = 0; | |
322 | ||
323 | anon_vma = page_lock_anon_vma(page); | |
324 | if (!anon_vma) | |
325 | return referenced; | |
326 | ||
327 | mapcount = page_mapcount(page); | |
328 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
329 | referenced += page_referenced_one(page, vma, &mapcount, | |
330 | ignore_token); | |
331 | if (!mapcount) | |
332 | break; | |
333 | } | |
334 | spin_unlock(&anon_vma->lock); | |
335 | return referenced; | |
336 | } | |
337 | ||
338 | /** | |
339 | * page_referenced_file - referenced check for object-based rmap | |
340 | * @page: the page we're checking references on. | |
341 | * | |
342 | * For an object-based mapped page, find all the places it is mapped and | |
343 | * check/clear the referenced flag. This is done by following the page->mapping | |
344 | * pointer, then walking the chain of vmas it holds. It returns the number | |
345 | * of references it found. | |
346 | * | |
347 | * This function is only called from page_referenced for object-based pages. | |
348 | */ | |
349 | static int page_referenced_file(struct page *page, int ignore_token) | |
350 | { | |
351 | unsigned int mapcount; | |
352 | struct address_space *mapping = page->mapping; | |
353 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
354 | struct vm_area_struct *vma; | |
355 | struct prio_tree_iter iter; | |
356 | int referenced = 0; | |
357 | ||
358 | /* | |
359 | * The caller's checks on page->mapping and !PageAnon have made | |
360 | * sure that this is a file page: the check for page->mapping | |
361 | * excludes the case just before it gets set on an anon page. | |
362 | */ | |
363 | BUG_ON(PageAnon(page)); | |
364 | ||
365 | /* | |
366 | * The page lock not only makes sure that page->mapping cannot | |
367 | * suddenly be NULLified by truncation, it makes sure that the | |
368 | * structure at mapping cannot be freed and reused yet, | |
369 | * so we can safely take mapping->i_mmap_lock. | |
370 | */ | |
371 | BUG_ON(!PageLocked(page)); | |
372 | ||
373 | spin_lock(&mapping->i_mmap_lock); | |
374 | ||
375 | /* | |
376 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
377 | * is more likely to be accurate if we note it after spinning. | |
378 | */ | |
379 | mapcount = page_mapcount(page); | |
380 | ||
381 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
382 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
383 | == (VM_LOCKED|VM_MAYSHARE)) { | |
384 | referenced++; | |
385 | break; | |
386 | } | |
387 | referenced += page_referenced_one(page, vma, &mapcount, | |
388 | ignore_token); | |
389 | if (!mapcount) | |
390 | break; | |
391 | } | |
392 | ||
393 | spin_unlock(&mapping->i_mmap_lock); | |
394 | return referenced; | |
395 | } | |
396 | ||
397 | /** | |
398 | * page_referenced - test if the page was referenced | |
399 | * @page: the page to test | |
400 | * @is_locked: caller holds lock on the page | |
401 | * | |
402 | * Quick test_and_clear_referenced for all mappings to a page, | |
403 | * returns the number of ptes which referenced the page. | |
404 | */ | |
405 | int page_referenced(struct page *page, int is_locked, int ignore_token) | |
406 | { | |
407 | int referenced = 0; | |
408 | ||
409 | if (!swap_token_default_timeout) | |
410 | ignore_token = 1; | |
411 | ||
412 | if (page_test_and_clear_young(page)) | |
413 | referenced++; | |
414 | ||
415 | if (TestClearPageReferenced(page)) | |
416 | referenced++; | |
417 | ||
418 | if (page_mapped(page) && page->mapping) { | |
419 | if (PageAnon(page)) | |
420 | referenced += page_referenced_anon(page, ignore_token); | |
421 | else if (is_locked) | |
422 | referenced += page_referenced_file(page, ignore_token); | |
423 | else if (TestSetPageLocked(page)) | |
424 | referenced++; | |
425 | else { | |
426 | if (page->mapping) | |
427 | referenced += page_referenced_file(page, | |
428 | ignore_token); | |
429 | unlock_page(page); | |
430 | } | |
431 | } | |
432 | return referenced; | |
433 | } | |
434 | ||
435 | /** | |
436 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
437 | * @page: the page to add the mapping to | |
438 | * @vma: the vm area in which the mapping is added | |
439 | * @address: the user virtual address mapped | |
440 | * | |
441 | * The caller needs to hold the mm->page_table_lock. | |
442 | */ | |
443 | void page_add_anon_rmap(struct page *page, | |
444 | struct vm_area_struct *vma, unsigned long address) | |
445 | { | |
1da177e4 | 446 | if (atomic_inc_and_test(&page->_mapcount)) { |
2822c1aa | 447 | struct anon_vma *anon_vma = vma->anon_vma; |
2822c1aa NP |
448 | |
449 | BUG_ON(!anon_vma); | |
450 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
1da177e4 | 451 | page->mapping = (struct address_space *) anon_vma; |
2822c1aa | 452 | |
4d7670e0 | 453 | page->index = linear_page_index(vma, address); |
2822c1aa | 454 | |
1da177e4 LT |
455 | inc_page_state(nr_mapped); |
456 | } | |
457 | /* else checking page index and mapping is racy */ | |
458 | } | |
459 | ||
460 | /** | |
461 | * page_add_file_rmap - add pte mapping to a file page | |
462 | * @page: the page to add the mapping to | |
463 | * | |
464 | * The caller needs to hold the mm->page_table_lock. | |
465 | */ | |
466 | void page_add_file_rmap(struct page *page) | |
467 | { | |
468 | BUG_ON(PageAnon(page)); | |
b5810039 | 469 | BUG_ON(!pfn_valid(page_to_pfn(page))); |
1da177e4 LT |
470 | |
471 | if (atomic_inc_and_test(&page->_mapcount)) | |
472 | inc_page_state(nr_mapped); | |
473 | } | |
474 | ||
475 | /** | |
476 | * page_remove_rmap - take down pte mapping from a page | |
477 | * @page: page to remove mapping from | |
478 | * | |
479 | * Caller needs to hold the mm->page_table_lock. | |
480 | */ | |
481 | void page_remove_rmap(struct page *page) | |
482 | { | |
1da177e4 LT |
483 | if (atomic_add_negative(-1, &page->_mapcount)) { |
484 | BUG_ON(page_mapcount(page) < 0); | |
485 | /* | |
486 | * It would be tidy to reset the PageAnon mapping here, | |
487 | * but that might overwrite a racing page_add_anon_rmap | |
488 | * which increments mapcount after us but sets mapping | |
489 | * before us: so leave the reset to free_hot_cold_page, | |
490 | * and remember that it's only reliable while mapped. | |
491 | * Leaving it set also helps swapoff to reinstate ptes | |
492 | * faster for those pages still in swapcache. | |
493 | */ | |
494 | if (page_test_and_clear_dirty(page)) | |
495 | set_page_dirty(page); | |
496 | dec_page_state(nr_mapped); | |
497 | } | |
498 | } | |
499 | ||
500 | /* | |
501 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
502 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
503 | */ | |
504 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma) | |
505 | { | |
506 | struct mm_struct *mm = vma->vm_mm; | |
507 | unsigned long address; | |
1da177e4 LT |
508 | pte_t *pte; |
509 | pte_t pteval; | |
510 | int ret = SWAP_AGAIN; | |
511 | ||
1da177e4 LT |
512 | address = vma_address(page, vma); |
513 | if (address == -EFAULT) | |
514 | goto out; | |
515 | ||
81b4082d ND |
516 | pte = page_check_address(page, mm, address); |
517 | if (IS_ERR(pte)) | |
518 | goto out; | |
1da177e4 LT |
519 | |
520 | /* | |
521 | * If the page is mlock()d, we cannot swap it out. | |
522 | * If it's recently referenced (perhaps page_referenced | |
523 | * skipped over this mm) then we should reactivate it. | |
c3dce2d8 NP |
524 | * |
525 | * Pages belonging to VM_RESERVED regions should not happen here. | |
1da177e4 LT |
526 | */ |
527 | if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) || | |
528 | ptep_clear_flush_young(vma, address, pte)) { | |
529 | ret = SWAP_FAIL; | |
530 | goto out_unmap; | |
531 | } | |
532 | ||
1da177e4 LT |
533 | /* Nuke the page table entry. */ |
534 | flush_cache_page(vma, address, page_to_pfn(page)); | |
535 | pteval = ptep_clear_flush(vma, address, pte); | |
536 | ||
537 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
538 | if (pte_dirty(pteval)) | |
539 | set_page_dirty(page); | |
540 | ||
541 | if (PageAnon(page)) { | |
542 | swp_entry_t entry = { .val = page->private }; | |
543 | /* | |
544 | * Store the swap location in the pte. | |
545 | * See handle_pte_fault() ... | |
546 | */ | |
547 | BUG_ON(!PageSwapCache(page)); | |
548 | swap_duplicate(entry); | |
549 | if (list_empty(&mm->mmlist)) { | |
550 | spin_lock(&mmlist_lock); | |
551 | list_add(&mm->mmlist, &init_mm.mmlist); | |
552 | spin_unlock(&mmlist_lock); | |
553 | } | |
554 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
555 | BUG_ON(pte_file(*pte)); | |
556 | dec_mm_counter(mm, anon_rss); | |
4294621f HD |
557 | } else |
558 | dec_mm_counter(mm, file_rss); | |
1da177e4 | 559 | |
1da177e4 LT |
560 | page_remove_rmap(page); |
561 | page_cache_release(page); | |
562 | ||
563 | out_unmap: | |
564 | pte_unmap(pte); | |
1da177e4 LT |
565 | spin_unlock(&mm->page_table_lock); |
566 | out: | |
567 | return ret; | |
568 | } | |
569 | ||
570 | /* | |
571 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
572 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
573 | * Consequently, given a particular page and its ->index, we cannot locate the | |
574 | * ptes which are mapping that page without an exhaustive linear search. | |
575 | * | |
576 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
577 | * maps the file to which the target page belongs. The ->vm_private_data field | |
578 | * holds the current cursor into that scan. Successive searches will circulate | |
579 | * around the vma's virtual address space. | |
580 | * | |
581 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
582 | * more scanning pressure is placed against them as well. Eventually pages | |
583 | * will become fully unmapped and are eligible for eviction. | |
584 | * | |
585 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
586 | * there there won't be many ptes located within the scan cluster. In this case | |
587 | * maybe we could scan further - to the end of the pte page, perhaps. | |
588 | */ | |
589 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
590 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
591 | ||
592 | static void try_to_unmap_cluster(unsigned long cursor, | |
593 | unsigned int *mapcount, struct vm_area_struct *vma) | |
594 | { | |
595 | struct mm_struct *mm = vma->vm_mm; | |
596 | pgd_t *pgd; | |
597 | pud_t *pud; | |
598 | pmd_t *pmd; | |
cafdd8ba | 599 | pte_t *pte, *original_pte; |
1da177e4 LT |
600 | pte_t pteval; |
601 | struct page *page; | |
602 | unsigned long address; | |
603 | unsigned long end; | |
604 | unsigned long pfn; | |
605 | ||
606 | /* | |
607 | * We need the page_table_lock to protect us from page faults, | |
608 | * munmap, fork, etc... | |
609 | */ | |
610 | spin_lock(&mm->page_table_lock); | |
611 | ||
612 | address = (vma->vm_start + cursor) & CLUSTER_MASK; | |
613 | end = address + CLUSTER_SIZE; | |
614 | if (address < vma->vm_start) | |
615 | address = vma->vm_start; | |
616 | if (end > vma->vm_end) | |
617 | end = vma->vm_end; | |
618 | ||
619 | pgd = pgd_offset(mm, address); | |
620 | if (!pgd_present(*pgd)) | |
621 | goto out_unlock; | |
622 | ||
623 | pud = pud_offset(pgd, address); | |
624 | if (!pud_present(*pud)) | |
625 | goto out_unlock; | |
626 | ||
627 | pmd = pmd_offset(pud, address); | |
628 | if (!pmd_present(*pmd)) | |
629 | goto out_unlock; | |
630 | ||
cafdd8ba | 631 | for (original_pte = pte = pte_offset_map(pmd, address); |
1da177e4 LT |
632 | address < end; pte++, address += PAGE_SIZE) { |
633 | ||
634 | if (!pte_present(*pte)) | |
635 | continue; | |
636 | ||
637 | pfn = pte_pfn(*pte); | |
b5810039 NP |
638 | if (unlikely(!pfn_valid(pfn))) { |
639 | print_bad_pte(vma, *pte, address); | |
1da177e4 | 640 | continue; |
b5810039 | 641 | } |
1da177e4 LT |
642 | |
643 | page = pfn_to_page(pfn); | |
644 | BUG_ON(PageAnon(page)); | |
1da177e4 LT |
645 | |
646 | if (ptep_clear_flush_young(vma, address, pte)) | |
647 | continue; | |
648 | ||
649 | /* Nuke the page table entry. */ | |
650 | flush_cache_page(vma, address, pfn); | |
651 | pteval = ptep_clear_flush(vma, address, pte); | |
652 | ||
653 | /* If nonlinear, store the file page offset in the pte. */ | |
654 | if (page->index != linear_page_index(vma, address)) | |
655 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
656 | ||
657 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
658 | if (pte_dirty(pteval)) | |
659 | set_page_dirty(page); | |
660 | ||
661 | page_remove_rmap(page); | |
662 | page_cache_release(page); | |
4294621f | 663 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
664 | (*mapcount)--; |
665 | } | |
666 | ||
cafdd8ba | 667 | pte_unmap(original_pte); |
1da177e4 LT |
668 | out_unlock: |
669 | spin_unlock(&mm->page_table_lock); | |
670 | } | |
671 | ||
672 | static int try_to_unmap_anon(struct page *page) | |
673 | { | |
674 | struct anon_vma *anon_vma; | |
675 | struct vm_area_struct *vma; | |
676 | int ret = SWAP_AGAIN; | |
677 | ||
678 | anon_vma = page_lock_anon_vma(page); | |
679 | if (!anon_vma) | |
680 | return ret; | |
681 | ||
682 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
683 | ret = try_to_unmap_one(page, vma); | |
684 | if (ret == SWAP_FAIL || !page_mapped(page)) | |
685 | break; | |
686 | } | |
687 | spin_unlock(&anon_vma->lock); | |
688 | return ret; | |
689 | } | |
690 | ||
691 | /** | |
692 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
693 | * @page: the page to unmap | |
694 | * | |
695 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
696 | * contained in the address_space struct it points to. | |
697 | * | |
698 | * This function is only called from try_to_unmap for object-based pages. | |
699 | */ | |
700 | static int try_to_unmap_file(struct page *page) | |
701 | { | |
702 | struct address_space *mapping = page->mapping; | |
703 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
704 | struct vm_area_struct *vma; | |
705 | struct prio_tree_iter iter; | |
706 | int ret = SWAP_AGAIN; | |
707 | unsigned long cursor; | |
708 | unsigned long max_nl_cursor = 0; | |
709 | unsigned long max_nl_size = 0; | |
710 | unsigned int mapcount; | |
711 | ||
712 | spin_lock(&mapping->i_mmap_lock); | |
713 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
714 | ret = try_to_unmap_one(page, vma); | |
715 | if (ret == SWAP_FAIL || !page_mapped(page)) | |
716 | goto out; | |
717 | } | |
718 | ||
719 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
720 | goto out; | |
721 | ||
722 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
723 | shared.vm_set.list) { | |
724 | if (vma->vm_flags & (VM_LOCKED|VM_RESERVED)) | |
725 | continue; | |
726 | cursor = (unsigned long) vma->vm_private_data; | |
727 | if (cursor > max_nl_cursor) | |
728 | max_nl_cursor = cursor; | |
729 | cursor = vma->vm_end - vma->vm_start; | |
730 | if (cursor > max_nl_size) | |
731 | max_nl_size = cursor; | |
732 | } | |
733 | ||
734 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
735 | ret = SWAP_FAIL; | |
736 | goto out; | |
737 | } | |
738 | ||
739 | /* | |
740 | * We don't try to search for this page in the nonlinear vmas, | |
741 | * and page_referenced wouldn't have found it anyway. Instead | |
742 | * just walk the nonlinear vmas trying to age and unmap some. | |
743 | * The mapcount of the page we came in with is irrelevant, | |
744 | * but even so use it as a guide to how hard we should try? | |
745 | */ | |
746 | mapcount = page_mapcount(page); | |
747 | if (!mapcount) | |
748 | goto out; | |
749 | cond_resched_lock(&mapping->i_mmap_lock); | |
750 | ||
751 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
752 | if (max_nl_cursor == 0) | |
753 | max_nl_cursor = CLUSTER_SIZE; | |
754 | ||
755 | do { | |
756 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
757 | shared.vm_set.list) { | |
758 | if (vma->vm_flags & (VM_LOCKED|VM_RESERVED)) | |
759 | continue; | |
760 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 761 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
762 | cursor < vma->vm_end - vma->vm_start) { |
763 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
764 | cursor += CLUSTER_SIZE; | |
765 | vma->vm_private_data = (void *) cursor; | |
766 | if ((int)mapcount <= 0) | |
767 | goto out; | |
768 | } | |
769 | vma->vm_private_data = (void *) max_nl_cursor; | |
770 | } | |
771 | cond_resched_lock(&mapping->i_mmap_lock); | |
772 | max_nl_cursor += CLUSTER_SIZE; | |
773 | } while (max_nl_cursor <= max_nl_size); | |
774 | ||
775 | /* | |
776 | * Don't loop forever (perhaps all the remaining pages are | |
777 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
778 | * vmas, now forgetting on which ones it had fallen behind. | |
779 | */ | |
780 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
781 | shared.vm_set.list) { | |
782 | if (!(vma->vm_flags & VM_RESERVED)) | |
783 | vma->vm_private_data = NULL; | |
784 | } | |
785 | out: | |
786 | spin_unlock(&mapping->i_mmap_lock); | |
787 | return ret; | |
788 | } | |
789 | ||
790 | /** | |
791 | * try_to_unmap - try to remove all page table mappings to a page | |
792 | * @page: the page to get unmapped | |
793 | * | |
794 | * Tries to remove all the page table entries which are mapping this | |
795 | * page, used in the pageout path. Caller must hold the page lock. | |
796 | * Return values are: | |
797 | * | |
798 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
799 | * SWAP_AGAIN - we missed a mapping, try again later | |
800 | * SWAP_FAIL - the page is unswappable | |
801 | */ | |
802 | int try_to_unmap(struct page *page) | |
803 | { | |
804 | int ret; | |
805 | ||
1da177e4 LT |
806 | BUG_ON(!PageLocked(page)); |
807 | ||
808 | if (PageAnon(page)) | |
809 | ret = try_to_unmap_anon(page); | |
810 | else | |
811 | ret = try_to_unmap_file(page); | |
812 | ||
813 | if (!page_mapped(page)) | |
814 | ret = SWAP_SUCCESS; | |
815 | return ret; | |
816 | } | |
81b4082d | 817 |