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) | |
37 | * swap_list_lock (in swap_free etc's swap_info_get) | |
38 | * mmlist_lock (in mmput, drain_mmlist and others) | |
39 | * swap_device_lock (in swap_duplicate, swap_info_get) | |
40 | * mapping->private_lock (in __set_page_dirty_buffers) | |
41 | * inode_lock (in set_page_dirty's __mark_inode_dirty) | |
42 | * sb_lock (within inode_lock in fs/fs-writeback.c) | |
43 | * mapping->tree_lock (widely used, in set_page_dirty, | |
44 | * in arch-dependent flush_dcache_mmap_lock, | |
45 | * within inode_lock in __sync_single_inode) | |
46 | */ | |
47 | ||
48 | #include <linux/mm.h> | |
49 | #include <linux/pagemap.h> | |
50 | #include <linux/swap.h> | |
51 | #include <linux/swapops.h> | |
52 | #include <linux/slab.h> | |
53 | #include <linux/init.h> | |
54 | #include <linux/rmap.h> | |
55 | #include <linux/rcupdate.h> | |
56 | ||
57 | #include <asm/tlbflush.h> | |
58 | ||
59 | //#define RMAP_DEBUG /* can be enabled only for debugging */ | |
60 | ||
61 | kmem_cache_t *anon_vma_cachep; | |
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 | ||
169 | static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags) | |
170 | { | |
171 | if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == | |
172 | SLAB_CTOR_CONSTRUCTOR) { | |
173 | struct anon_vma *anon_vma = data; | |
174 | ||
175 | spin_lock_init(&anon_vma->lock); | |
176 | INIT_LIST_HEAD(&anon_vma->head); | |
177 | } | |
178 | } | |
179 | ||
180 | void __init anon_vma_init(void) | |
181 | { | |
182 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
183 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL); | |
184 | } | |
185 | ||
186 | /* | |
187 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
188 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
189 | */ | |
190 | static struct anon_vma *page_lock_anon_vma(struct page *page) | |
191 | { | |
192 | struct anon_vma *anon_vma = NULL; | |
193 | unsigned long anon_mapping; | |
194 | ||
195 | rcu_read_lock(); | |
196 | anon_mapping = (unsigned long) page->mapping; | |
197 | if (!(anon_mapping & PAGE_MAPPING_ANON)) | |
198 | goto out; | |
199 | if (!page_mapped(page)) | |
200 | goto out; | |
201 | ||
202 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
203 | spin_lock(&anon_vma->lock); | |
204 | out: | |
205 | rcu_read_unlock(); | |
206 | return anon_vma; | |
207 | } | |
208 | ||
209 | /* | |
210 | * At what user virtual address is page expected in vma? | |
211 | */ | |
212 | static inline unsigned long | |
213 | vma_address(struct page *page, struct vm_area_struct *vma) | |
214 | { | |
215 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
216 | unsigned long address; | |
217 | ||
218 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
219 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
220 | /* page should be within any vma from prio_tree_next */ | |
221 | BUG_ON(!PageAnon(page)); | |
222 | return -EFAULT; | |
223 | } | |
224 | return address; | |
225 | } | |
226 | ||
227 | /* | |
228 | * At what user virtual address is page expected in vma? checking that the | |
229 | * page matches the vma: currently only used by unuse_process, on anon pages. | |
230 | */ | |
231 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
232 | { | |
233 | if (PageAnon(page)) { | |
234 | if ((void *)vma->anon_vma != | |
235 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
236 | return -EFAULT; | |
237 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
238 | if (vma->vm_file->f_mapping != page->mapping) | |
239 | return -EFAULT; | |
240 | } else | |
241 | return -EFAULT; | |
242 | return vma_address(page, vma); | |
243 | } | |
244 | ||
81b4082d ND |
245 | /* |
246 | * Check that @page is mapped at @address into @mm. | |
247 | * | |
248 | * On success returns with mapped pte and locked mm->page_table_lock. | |
249 | */ | |
ceffc078 CO |
250 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
251 | unsigned long address) | |
81b4082d ND |
252 | { |
253 | pgd_t *pgd; | |
254 | pud_t *pud; | |
255 | pmd_t *pmd; | |
256 | pte_t *pte; | |
257 | ||
258 | /* | |
259 | * We need the page_table_lock to protect us from page faults, | |
260 | * munmap, fork, etc... | |
261 | */ | |
262 | spin_lock(&mm->page_table_lock); | |
263 | pgd = pgd_offset(mm, address); | |
264 | if (likely(pgd_present(*pgd))) { | |
265 | pud = pud_offset(pgd, address); | |
266 | if (likely(pud_present(*pud))) { | |
267 | pmd = pmd_offset(pud, address); | |
268 | if (likely(pmd_present(*pmd))) { | |
269 | pte = pte_offset_map(pmd, address); | |
270 | if (likely(pte_present(*pte) && | |
271 | page_to_pfn(page) == pte_pfn(*pte))) | |
272 | return pte; | |
273 | pte_unmap(pte); | |
274 | } | |
275 | } | |
276 | } | |
277 | spin_unlock(&mm->page_table_lock); | |
278 | return ERR_PTR(-ENOENT); | |
279 | } | |
280 | ||
1da177e4 LT |
281 | /* |
282 | * Subfunctions of page_referenced: page_referenced_one called | |
283 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
284 | */ | |
285 | static int page_referenced_one(struct page *page, | |
286 | struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token) | |
287 | { | |
288 | struct mm_struct *mm = vma->vm_mm; | |
289 | unsigned long address; | |
1da177e4 LT |
290 | pte_t *pte; |
291 | int referenced = 0; | |
292 | ||
293 | if (!get_mm_counter(mm, rss)) | |
294 | goto out; | |
295 | address = vma_address(page, vma); | |
296 | if (address == -EFAULT) | |
297 | goto out; | |
298 | ||
81b4082d ND |
299 | pte = page_check_address(page, mm, address); |
300 | if (!IS_ERR(pte)) { | |
301 | if (ptep_clear_flush_young(vma, address, pte)) | |
302 | referenced++; | |
1da177e4 | 303 | |
81b4082d ND |
304 | if (mm != current->mm && !ignore_token && has_swap_token(mm)) |
305 | referenced++; | |
1da177e4 | 306 | |
81b4082d ND |
307 | (*mapcount)--; |
308 | pte_unmap(pte); | |
309 | spin_unlock(&mm->page_table_lock); | |
310 | } | |
1da177e4 LT |
311 | out: |
312 | return referenced; | |
313 | } | |
314 | ||
315 | static int page_referenced_anon(struct page *page, int ignore_token) | |
316 | { | |
317 | unsigned int mapcount; | |
318 | struct anon_vma *anon_vma; | |
319 | struct vm_area_struct *vma; | |
320 | int referenced = 0; | |
321 | ||
322 | anon_vma = page_lock_anon_vma(page); | |
323 | if (!anon_vma) | |
324 | return referenced; | |
325 | ||
326 | mapcount = page_mapcount(page); | |
327 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
328 | referenced += page_referenced_one(page, vma, &mapcount, | |
329 | ignore_token); | |
330 | if (!mapcount) | |
331 | break; | |
332 | } | |
333 | spin_unlock(&anon_vma->lock); | |
334 | return referenced; | |
335 | } | |
336 | ||
337 | /** | |
338 | * page_referenced_file - referenced check for object-based rmap | |
339 | * @page: the page we're checking references on. | |
340 | * | |
341 | * For an object-based mapped page, find all the places it is mapped and | |
342 | * check/clear the referenced flag. This is done by following the page->mapping | |
343 | * pointer, then walking the chain of vmas it holds. It returns the number | |
344 | * of references it found. | |
345 | * | |
346 | * This function is only called from page_referenced for object-based pages. | |
347 | */ | |
348 | static int page_referenced_file(struct page *page, int ignore_token) | |
349 | { | |
350 | unsigned int mapcount; | |
351 | struct address_space *mapping = page->mapping; | |
352 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
353 | struct vm_area_struct *vma; | |
354 | struct prio_tree_iter iter; | |
355 | int referenced = 0; | |
356 | ||
357 | /* | |
358 | * The caller's checks on page->mapping and !PageAnon have made | |
359 | * sure that this is a file page: the check for page->mapping | |
360 | * excludes the case just before it gets set on an anon page. | |
361 | */ | |
362 | BUG_ON(PageAnon(page)); | |
363 | ||
364 | /* | |
365 | * The page lock not only makes sure that page->mapping cannot | |
366 | * suddenly be NULLified by truncation, it makes sure that the | |
367 | * structure at mapping cannot be freed and reused yet, | |
368 | * so we can safely take mapping->i_mmap_lock. | |
369 | */ | |
370 | BUG_ON(!PageLocked(page)); | |
371 | ||
372 | spin_lock(&mapping->i_mmap_lock); | |
373 | ||
374 | /* | |
375 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
376 | * is more likely to be accurate if we note it after spinning. | |
377 | */ | |
378 | mapcount = page_mapcount(page); | |
379 | ||
380 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
381 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
382 | == (VM_LOCKED|VM_MAYSHARE)) { | |
383 | referenced++; | |
384 | break; | |
385 | } | |
386 | referenced += page_referenced_one(page, vma, &mapcount, | |
387 | ignore_token); | |
388 | if (!mapcount) | |
389 | break; | |
390 | } | |
391 | ||
392 | spin_unlock(&mapping->i_mmap_lock); | |
393 | return referenced; | |
394 | } | |
395 | ||
396 | /** | |
397 | * page_referenced - test if the page was referenced | |
398 | * @page: the page to test | |
399 | * @is_locked: caller holds lock on the page | |
400 | * | |
401 | * Quick test_and_clear_referenced for all mappings to a page, | |
402 | * returns the number of ptes which referenced the page. | |
403 | */ | |
404 | int page_referenced(struct page *page, int is_locked, int ignore_token) | |
405 | { | |
406 | int referenced = 0; | |
407 | ||
408 | if (!swap_token_default_timeout) | |
409 | ignore_token = 1; | |
410 | ||
411 | if (page_test_and_clear_young(page)) | |
412 | referenced++; | |
413 | ||
414 | if (TestClearPageReferenced(page)) | |
415 | referenced++; | |
416 | ||
417 | if (page_mapped(page) && page->mapping) { | |
418 | if (PageAnon(page)) | |
419 | referenced += page_referenced_anon(page, ignore_token); | |
420 | else if (is_locked) | |
421 | referenced += page_referenced_file(page, ignore_token); | |
422 | else if (TestSetPageLocked(page)) | |
423 | referenced++; | |
424 | else { | |
425 | if (page->mapping) | |
426 | referenced += page_referenced_file(page, | |
427 | ignore_token); | |
428 | unlock_page(page); | |
429 | } | |
430 | } | |
431 | return referenced; | |
432 | } | |
433 | ||
434 | /** | |
435 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
436 | * @page: the page to add the mapping to | |
437 | * @vma: the vm area in which the mapping is added | |
438 | * @address: the user virtual address mapped | |
439 | * | |
440 | * The caller needs to hold the mm->page_table_lock. | |
441 | */ | |
442 | void page_add_anon_rmap(struct page *page, | |
443 | struct vm_area_struct *vma, unsigned long address) | |
444 | { | |
445 | struct anon_vma *anon_vma = vma->anon_vma; | |
446 | pgoff_t index; | |
447 | ||
448 | BUG_ON(PageReserved(page)); | |
449 | BUG_ON(!anon_vma); | |
450 | ||
451 | inc_mm_counter(vma->vm_mm, anon_rss); | |
452 | ||
453 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
454 | index = (address - vma->vm_start) >> PAGE_SHIFT; | |
455 | index += vma->vm_pgoff; | |
456 | index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT; | |
457 | ||
458 | if (atomic_inc_and_test(&page->_mapcount)) { | |
459 | page->index = index; | |
460 | page->mapping = (struct address_space *) anon_vma; | |
461 | inc_page_state(nr_mapped); | |
462 | } | |
463 | /* else checking page index and mapping is racy */ | |
464 | } | |
465 | ||
466 | /** | |
467 | * page_add_file_rmap - add pte mapping to a file page | |
468 | * @page: the page to add the mapping to | |
469 | * | |
470 | * The caller needs to hold the mm->page_table_lock. | |
471 | */ | |
472 | void page_add_file_rmap(struct page *page) | |
473 | { | |
474 | BUG_ON(PageAnon(page)); | |
475 | if (!pfn_valid(page_to_pfn(page)) || PageReserved(page)) | |
476 | return; | |
477 | ||
478 | if (atomic_inc_and_test(&page->_mapcount)) | |
479 | inc_page_state(nr_mapped); | |
480 | } | |
481 | ||
482 | /** | |
483 | * page_remove_rmap - take down pte mapping from a page | |
484 | * @page: page to remove mapping from | |
485 | * | |
486 | * Caller needs to hold the mm->page_table_lock. | |
487 | */ | |
488 | void page_remove_rmap(struct page *page) | |
489 | { | |
490 | BUG_ON(PageReserved(page)); | |
491 | ||
492 | if (atomic_add_negative(-1, &page->_mapcount)) { | |
493 | BUG_ON(page_mapcount(page) < 0); | |
494 | /* | |
495 | * It would be tidy to reset the PageAnon mapping here, | |
496 | * but that might overwrite a racing page_add_anon_rmap | |
497 | * which increments mapcount after us but sets mapping | |
498 | * before us: so leave the reset to free_hot_cold_page, | |
499 | * and remember that it's only reliable while mapped. | |
500 | * Leaving it set also helps swapoff to reinstate ptes | |
501 | * faster for those pages still in swapcache. | |
502 | */ | |
503 | if (page_test_and_clear_dirty(page)) | |
504 | set_page_dirty(page); | |
505 | dec_page_state(nr_mapped); | |
506 | } | |
507 | } | |
508 | ||
509 | /* | |
510 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
511 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
512 | */ | |
513 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma) | |
514 | { | |
515 | struct mm_struct *mm = vma->vm_mm; | |
516 | unsigned long address; | |
1da177e4 LT |
517 | pte_t *pte; |
518 | pte_t pteval; | |
519 | int ret = SWAP_AGAIN; | |
520 | ||
521 | if (!get_mm_counter(mm, rss)) | |
522 | goto out; | |
523 | address = vma_address(page, vma); | |
524 | if (address == -EFAULT) | |
525 | goto out; | |
526 | ||
81b4082d ND |
527 | pte = page_check_address(page, mm, address); |
528 | if (IS_ERR(pte)) | |
529 | goto out; | |
1da177e4 LT |
530 | |
531 | /* | |
532 | * If the page is mlock()d, we cannot swap it out. | |
533 | * If it's recently referenced (perhaps page_referenced | |
534 | * skipped over this mm) then we should reactivate it. | |
535 | */ | |
536 | if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) || | |
537 | ptep_clear_flush_young(vma, address, pte)) { | |
538 | ret = SWAP_FAIL; | |
539 | goto out_unmap; | |
540 | } | |
541 | ||
1da177e4 LT |
542 | /* Nuke the page table entry. */ |
543 | flush_cache_page(vma, address, page_to_pfn(page)); | |
544 | pteval = ptep_clear_flush(vma, address, pte); | |
545 | ||
546 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
547 | if (pte_dirty(pteval)) | |
548 | set_page_dirty(page); | |
549 | ||
550 | if (PageAnon(page)) { | |
551 | swp_entry_t entry = { .val = page->private }; | |
552 | /* | |
553 | * Store the swap location in the pte. | |
554 | * See handle_pte_fault() ... | |
555 | */ | |
556 | BUG_ON(!PageSwapCache(page)); | |
557 | swap_duplicate(entry); | |
558 | if (list_empty(&mm->mmlist)) { | |
559 | spin_lock(&mmlist_lock); | |
560 | list_add(&mm->mmlist, &init_mm.mmlist); | |
561 | spin_unlock(&mmlist_lock); | |
562 | } | |
563 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
564 | BUG_ON(pte_file(*pte)); | |
565 | dec_mm_counter(mm, anon_rss); | |
566 | } | |
567 | ||
202d182a | 568 | dec_mm_counter(mm, rss); |
1da177e4 LT |
569 | page_remove_rmap(page); |
570 | page_cache_release(page); | |
571 | ||
572 | out_unmap: | |
573 | pte_unmap(pte); | |
1da177e4 LT |
574 | spin_unlock(&mm->page_table_lock); |
575 | out: | |
576 | return ret; | |
577 | } | |
578 | ||
579 | /* | |
580 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
581 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
582 | * Consequently, given a particular page and its ->index, we cannot locate the | |
583 | * ptes which are mapping that page without an exhaustive linear search. | |
584 | * | |
585 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
586 | * maps the file to which the target page belongs. The ->vm_private_data field | |
587 | * holds the current cursor into that scan. Successive searches will circulate | |
588 | * around the vma's virtual address space. | |
589 | * | |
590 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
591 | * more scanning pressure is placed against them as well. Eventually pages | |
592 | * will become fully unmapped and are eligible for eviction. | |
593 | * | |
594 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
595 | * there there won't be many ptes located within the scan cluster. In this case | |
596 | * maybe we could scan further - to the end of the pte page, perhaps. | |
597 | */ | |
598 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
599 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
600 | ||
601 | static void try_to_unmap_cluster(unsigned long cursor, | |
602 | unsigned int *mapcount, struct vm_area_struct *vma) | |
603 | { | |
604 | struct mm_struct *mm = vma->vm_mm; | |
605 | pgd_t *pgd; | |
606 | pud_t *pud; | |
607 | pmd_t *pmd; | |
cafdd8ba | 608 | pte_t *pte, *original_pte; |
1da177e4 LT |
609 | pte_t pteval; |
610 | struct page *page; | |
611 | unsigned long address; | |
612 | unsigned long end; | |
613 | unsigned long pfn; | |
614 | ||
615 | /* | |
616 | * We need the page_table_lock to protect us from page faults, | |
617 | * munmap, fork, etc... | |
618 | */ | |
619 | spin_lock(&mm->page_table_lock); | |
620 | ||
621 | address = (vma->vm_start + cursor) & CLUSTER_MASK; | |
622 | end = address + CLUSTER_SIZE; | |
623 | if (address < vma->vm_start) | |
624 | address = vma->vm_start; | |
625 | if (end > vma->vm_end) | |
626 | end = vma->vm_end; | |
627 | ||
628 | pgd = pgd_offset(mm, address); | |
629 | if (!pgd_present(*pgd)) | |
630 | goto out_unlock; | |
631 | ||
632 | pud = pud_offset(pgd, address); | |
633 | if (!pud_present(*pud)) | |
634 | goto out_unlock; | |
635 | ||
636 | pmd = pmd_offset(pud, address); | |
637 | if (!pmd_present(*pmd)) | |
638 | goto out_unlock; | |
639 | ||
cafdd8ba | 640 | for (original_pte = pte = pte_offset_map(pmd, address); |
1da177e4 LT |
641 | address < end; pte++, address += PAGE_SIZE) { |
642 | ||
643 | if (!pte_present(*pte)) | |
644 | continue; | |
645 | ||
646 | pfn = pte_pfn(*pte); | |
647 | if (!pfn_valid(pfn)) | |
648 | continue; | |
649 | ||
650 | page = pfn_to_page(pfn); | |
651 | BUG_ON(PageAnon(page)); | |
652 | if (PageReserved(page)) | |
653 | continue; | |
654 | ||
655 | if (ptep_clear_flush_young(vma, address, pte)) | |
656 | continue; | |
657 | ||
658 | /* Nuke the page table entry. */ | |
659 | flush_cache_page(vma, address, pfn); | |
660 | pteval = ptep_clear_flush(vma, address, pte); | |
661 | ||
662 | /* If nonlinear, store the file page offset in the pte. */ | |
663 | if (page->index != linear_page_index(vma, address)) | |
664 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
665 | ||
666 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
667 | if (pte_dirty(pteval)) | |
668 | set_page_dirty(page); | |
669 | ||
670 | page_remove_rmap(page); | |
671 | page_cache_release(page); | |
672 | dec_mm_counter(mm, rss); | |
673 | (*mapcount)--; | |
674 | } | |
675 | ||
cafdd8ba | 676 | pte_unmap(original_pte); |
1da177e4 LT |
677 | out_unlock: |
678 | spin_unlock(&mm->page_table_lock); | |
679 | } | |
680 | ||
681 | static int try_to_unmap_anon(struct page *page) | |
682 | { | |
683 | struct anon_vma *anon_vma; | |
684 | struct vm_area_struct *vma; | |
685 | int ret = SWAP_AGAIN; | |
686 | ||
687 | anon_vma = page_lock_anon_vma(page); | |
688 | if (!anon_vma) | |
689 | return ret; | |
690 | ||
691 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
692 | ret = try_to_unmap_one(page, vma); | |
693 | if (ret == SWAP_FAIL || !page_mapped(page)) | |
694 | break; | |
695 | } | |
696 | spin_unlock(&anon_vma->lock); | |
697 | return ret; | |
698 | } | |
699 | ||
700 | /** | |
701 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
702 | * @page: the page to unmap | |
703 | * | |
704 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
705 | * contained in the address_space struct it points to. | |
706 | * | |
707 | * This function is only called from try_to_unmap for object-based pages. | |
708 | */ | |
709 | static int try_to_unmap_file(struct page *page) | |
710 | { | |
711 | struct address_space *mapping = page->mapping; | |
712 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
713 | struct vm_area_struct *vma; | |
714 | struct prio_tree_iter iter; | |
715 | int ret = SWAP_AGAIN; | |
716 | unsigned long cursor; | |
717 | unsigned long max_nl_cursor = 0; | |
718 | unsigned long max_nl_size = 0; | |
719 | unsigned int mapcount; | |
720 | ||
721 | spin_lock(&mapping->i_mmap_lock); | |
722 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
723 | ret = try_to_unmap_one(page, vma); | |
724 | if (ret == SWAP_FAIL || !page_mapped(page)) | |
725 | goto out; | |
726 | } | |
727 | ||
728 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
729 | goto out; | |
730 | ||
731 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
732 | shared.vm_set.list) { | |
733 | if (vma->vm_flags & (VM_LOCKED|VM_RESERVED)) | |
734 | continue; | |
735 | cursor = (unsigned long) vma->vm_private_data; | |
736 | if (cursor > max_nl_cursor) | |
737 | max_nl_cursor = cursor; | |
738 | cursor = vma->vm_end - vma->vm_start; | |
739 | if (cursor > max_nl_size) | |
740 | max_nl_size = cursor; | |
741 | } | |
742 | ||
743 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
744 | ret = SWAP_FAIL; | |
745 | goto out; | |
746 | } | |
747 | ||
748 | /* | |
749 | * We don't try to search for this page in the nonlinear vmas, | |
750 | * and page_referenced wouldn't have found it anyway. Instead | |
751 | * just walk the nonlinear vmas trying to age and unmap some. | |
752 | * The mapcount of the page we came in with is irrelevant, | |
753 | * but even so use it as a guide to how hard we should try? | |
754 | */ | |
755 | mapcount = page_mapcount(page); | |
756 | if (!mapcount) | |
757 | goto out; | |
758 | cond_resched_lock(&mapping->i_mmap_lock); | |
759 | ||
760 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
761 | if (max_nl_cursor == 0) | |
762 | max_nl_cursor = CLUSTER_SIZE; | |
763 | ||
764 | do { | |
765 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
766 | shared.vm_set.list) { | |
767 | if (vma->vm_flags & (VM_LOCKED|VM_RESERVED)) | |
768 | continue; | |
769 | cursor = (unsigned long) vma->vm_private_data; | |
770 | while (get_mm_counter(vma->vm_mm, rss) && | |
771 | cursor < max_nl_cursor && | |
772 | cursor < vma->vm_end - vma->vm_start) { | |
773 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
774 | cursor += CLUSTER_SIZE; | |
775 | vma->vm_private_data = (void *) cursor; | |
776 | if ((int)mapcount <= 0) | |
777 | goto out; | |
778 | } | |
779 | vma->vm_private_data = (void *) max_nl_cursor; | |
780 | } | |
781 | cond_resched_lock(&mapping->i_mmap_lock); | |
782 | max_nl_cursor += CLUSTER_SIZE; | |
783 | } while (max_nl_cursor <= max_nl_size); | |
784 | ||
785 | /* | |
786 | * Don't loop forever (perhaps all the remaining pages are | |
787 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
788 | * vmas, now forgetting on which ones it had fallen behind. | |
789 | */ | |
790 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
791 | shared.vm_set.list) { | |
792 | if (!(vma->vm_flags & VM_RESERVED)) | |
793 | vma->vm_private_data = NULL; | |
794 | } | |
795 | out: | |
796 | spin_unlock(&mapping->i_mmap_lock); | |
797 | return ret; | |
798 | } | |
799 | ||
800 | /** | |
801 | * try_to_unmap - try to remove all page table mappings to a page | |
802 | * @page: the page to get unmapped | |
803 | * | |
804 | * Tries to remove all the page table entries which are mapping this | |
805 | * page, used in the pageout path. Caller must hold the page lock. | |
806 | * Return values are: | |
807 | * | |
808 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
809 | * SWAP_AGAIN - we missed a mapping, try again later | |
810 | * SWAP_FAIL - the page is unswappable | |
811 | */ | |
812 | int try_to_unmap(struct page *page) | |
813 | { | |
814 | int ret; | |
815 | ||
816 | BUG_ON(PageReserved(page)); | |
817 | BUG_ON(!PageLocked(page)); | |
818 | ||
819 | if (PageAnon(page)) | |
820 | ret = try_to_unmap_anon(page); | |
821 | else | |
822 | ret = try_to_unmap_file(page); | |
823 | ||
824 | if (!page_mapped(page)) | |
825 | ret = SWAP_SUCCESS; | |
826 | return ret; | |
827 | } | |
81b4082d | 828 |