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 | |
b8072f09 | 35 | * mm->page_table_lock or pte_lock |
1da177e4 | 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 | |
ee498ed7 | 228 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
1da177e4 LT |
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)) { | |
ee498ed7 HD |
237 | if (!vma->vm_file || |
238 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
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 | * | |
b8072f09 | 248 | * On success returns with pte mapped and locked. |
81b4082d | 249 | */ |
ceffc078 | 250 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
c0718806 | 251 | unsigned long address, spinlock_t **ptlp) |
81b4082d ND |
252 | { |
253 | pgd_t *pgd; | |
254 | pud_t *pud; | |
255 | pmd_t *pmd; | |
256 | pte_t *pte; | |
c0718806 | 257 | spinlock_t *ptl; |
81b4082d | 258 | |
81b4082d | 259 | pgd = pgd_offset(mm, address); |
c0718806 HD |
260 | if (!pgd_present(*pgd)) |
261 | return NULL; | |
262 | ||
263 | pud = pud_offset(pgd, address); | |
264 | if (!pud_present(*pud)) | |
265 | return NULL; | |
266 | ||
267 | pmd = pmd_offset(pud, address); | |
268 | if (!pmd_present(*pmd)) | |
269 | return NULL; | |
270 | ||
271 | pte = pte_offset_map(pmd, address); | |
272 | /* Make a quick check before getting the lock */ | |
273 | if (!pte_present(*pte)) { | |
274 | pte_unmap(pte); | |
275 | return NULL; | |
276 | } | |
277 | ||
4c21e2f2 | 278 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
279 | spin_lock(ptl); |
280 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
281 | *ptlp = ptl; | |
282 | return pte; | |
81b4082d | 283 | } |
c0718806 HD |
284 | pte_unmap_unlock(pte, ptl); |
285 | return NULL; | |
81b4082d ND |
286 | } |
287 | ||
1da177e4 LT |
288 | /* |
289 | * Subfunctions of page_referenced: page_referenced_one called | |
290 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
291 | */ | |
292 | static int page_referenced_one(struct page *page, | |
f7b7fd8f | 293 | struct vm_area_struct *vma, unsigned int *mapcount) |
1da177e4 LT |
294 | { |
295 | struct mm_struct *mm = vma->vm_mm; | |
296 | unsigned long address; | |
1da177e4 | 297 | pte_t *pte; |
c0718806 | 298 | spinlock_t *ptl; |
1da177e4 LT |
299 | int referenced = 0; |
300 | ||
1da177e4 LT |
301 | address = vma_address(page, vma); |
302 | if (address == -EFAULT) | |
303 | goto out; | |
304 | ||
c0718806 HD |
305 | pte = page_check_address(page, mm, address, &ptl); |
306 | if (!pte) | |
307 | goto out; | |
1da177e4 | 308 | |
c0718806 HD |
309 | if (ptep_clear_flush_young(vma, address, pte)) |
310 | referenced++; | |
1da177e4 | 311 | |
c0718806 HD |
312 | /* Pretend the page is referenced if the task has the |
313 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 314 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
315 | rwsem_is_locked(&mm->mmap_sem)) |
316 | referenced++; | |
317 | ||
318 | (*mapcount)--; | |
319 | pte_unmap_unlock(pte, ptl); | |
1da177e4 LT |
320 | out: |
321 | return referenced; | |
322 | } | |
323 | ||
f7b7fd8f | 324 | static int page_referenced_anon(struct page *page) |
1da177e4 LT |
325 | { |
326 | unsigned int mapcount; | |
327 | struct anon_vma *anon_vma; | |
328 | struct vm_area_struct *vma; | |
329 | int referenced = 0; | |
330 | ||
331 | anon_vma = page_lock_anon_vma(page); | |
332 | if (!anon_vma) | |
333 | return referenced; | |
334 | ||
335 | mapcount = page_mapcount(page); | |
336 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
f7b7fd8f | 337 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
338 | if (!mapcount) |
339 | break; | |
340 | } | |
341 | spin_unlock(&anon_vma->lock); | |
342 | return referenced; | |
343 | } | |
344 | ||
345 | /** | |
346 | * page_referenced_file - referenced check for object-based rmap | |
347 | * @page: the page we're checking references on. | |
348 | * | |
349 | * For an object-based mapped page, find all the places it is mapped and | |
350 | * check/clear the referenced flag. This is done by following the page->mapping | |
351 | * pointer, then walking the chain of vmas it holds. It returns the number | |
352 | * of references it found. | |
353 | * | |
354 | * This function is only called from page_referenced for object-based pages. | |
355 | */ | |
f7b7fd8f | 356 | static int page_referenced_file(struct page *page) |
1da177e4 LT |
357 | { |
358 | unsigned int mapcount; | |
359 | struct address_space *mapping = page->mapping; | |
360 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
361 | struct vm_area_struct *vma; | |
362 | struct prio_tree_iter iter; | |
363 | int referenced = 0; | |
364 | ||
365 | /* | |
366 | * The caller's checks on page->mapping and !PageAnon have made | |
367 | * sure that this is a file page: the check for page->mapping | |
368 | * excludes the case just before it gets set on an anon page. | |
369 | */ | |
370 | BUG_ON(PageAnon(page)); | |
371 | ||
372 | /* | |
373 | * The page lock not only makes sure that page->mapping cannot | |
374 | * suddenly be NULLified by truncation, it makes sure that the | |
375 | * structure at mapping cannot be freed and reused yet, | |
376 | * so we can safely take mapping->i_mmap_lock. | |
377 | */ | |
378 | BUG_ON(!PageLocked(page)); | |
379 | ||
380 | spin_lock(&mapping->i_mmap_lock); | |
381 | ||
382 | /* | |
383 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
384 | * is more likely to be accurate if we note it after spinning. | |
385 | */ | |
386 | mapcount = page_mapcount(page); | |
387 | ||
388 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
389 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
390 | == (VM_LOCKED|VM_MAYSHARE)) { | |
391 | referenced++; | |
392 | break; | |
393 | } | |
f7b7fd8f | 394 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
395 | if (!mapcount) |
396 | break; | |
397 | } | |
398 | ||
399 | spin_unlock(&mapping->i_mmap_lock); | |
400 | return referenced; | |
401 | } | |
402 | ||
403 | /** | |
404 | * page_referenced - test if the page was referenced | |
405 | * @page: the page to test | |
406 | * @is_locked: caller holds lock on the page | |
407 | * | |
408 | * Quick test_and_clear_referenced for all mappings to a page, | |
409 | * returns the number of ptes which referenced the page. | |
410 | */ | |
f7b7fd8f | 411 | int page_referenced(struct page *page, int is_locked) |
1da177e4 LT |
412 | { |
413 | int referenced = 0; | |
414 | ||
1da177e4 LT |
415 | if (page_test_and_clear_young(page)) |
416 | referenced++; | |
417 | ||
418 | if (TestClearPageReferenced(page)) | |
419 | referenced++; | |
420 | ||
421 | if (page_mapped(page) && page->mapping) { | |
422 | if (PageAnon(page)) | |
f7b7fd8f | 423 | referenced += page_referenced_anon(page); |
1da177e4 | 424 | else if (is_locked) |
f7b7fd8f | 425 | referenced += page_referenced_file(page); |
1da177e4 LT |
426 | else if (TestSetPageLocked(page)) |
427 | referenced++; | |
428 | else { | |
429 | if (page->mapping) | |
f7b7fd8f | 430 | referenced += page_referenced_file(page); |
1da177e4 LT |
431 | unlock_page(page); |
432 | } | |
433 | } | |
434 | return referenced; | |
435 | } | |
436 | ||
9617d95e NP |
437 | /** |
438 | * page_set_anon_rmap - setup new anonymous rmap | |
439 | * @page: the page to add the mapping to | |
440 | * @vma: the vm area in which the mapping is added | |
441 | * @address: the user virtual address mapped | |
442 | */ | |
443 | static void __page_set_anon_rmap(struct page *page, | |
444 | struct vm_area_struct *vma, unsigned long address) | |
445 | { | |
446 | struct anon_vma *anon_vma = vma->anon_vma; | |
447 | ||
448 | BUG_ON(!anon_vma); | |
449 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
450 | page->mapping = (struct address_space *) anon_vma; | |
451 | ||
452 | page->index = linear_page_index(vma, address); | |
453 | ||
a74609fa NP |
454 | /* |
455 | * nr_mapped state can be updated without turning off | |
456 | * interrupts because it is not modified via interrupt. | |
457 | */ | |
458 | __inc_page_state(nr_mapped); | |
9617d95e NP |
459 | } |
460 | ||
1da177e4 LT |
461 | /** |
462 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
463 | * @page: the page to add the mapping to | |
464 | * @vma: the vm area in which the mapping is added | |
465 | * @address: the user virtual address mapped | |
466 | * | |
b8072f09 | 467 | * The caller needs to hold the pte lock. |
1da177e4 LT |
468 | */ |
469 | void page_add_anon_rmap(struct page *page, | |
470 | struct vm_area_struct *vma, unsigned long address) | |
471 | { | |
9617d95e NP |
472 | if (atomic_inc_and_test(&page->_mapcount)) |
473 | __page_set_anon_rmap(page, vma, address); | |
1da177e4 LT |
474 | /* else checking page index and mapping is racy */ |
475 | } | |
476 | ||
9617d95e NP |
477 | /* |
478 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
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 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
484 | * This means the inc-and-test can be bypassed. | |
485 | */ | |
486 | void page_add_new_anon_rmap(struct page *page, | |
487 | struct vm_area_struct *vma, unsigned long address) | |
488 | { | |
489 | atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ | |
490 | __page_set_anon_rmap(page, vma, address); | |
491 | } | |
492 | ||
1da177e4 LT |
493 | /** |
494 | * page_add_file_rmap - add pte mapping to a file page | |
495 | * @page: the page to add the mapping to | |
496 | * | |
b8072f09 | 497 | * The caller needs to hold the pte lock. |
1da177e4 LT |
498 | */ |
499 | void page_add_file_rmap(struct page *page) | |
500 | { | |
501 | BUG_ON(PageAnon(page)); | |
b5810039 | 502 | BUG_ON(!pfn_valid(page_to_pfn(page))); |
1da177e4 LT |
503 | |
504 | if (atomic_inc_and_test(&page->_mapcount)) | |
a74609fa | 505 | __inc_page_state(nr_mapped); |
1da177e4 LT |
506 | } |
507 | ||
508 | /** | |
509 | * page_remove_rmap - take down pte mapping from a page | |
510 | * @page: page to remove mapping from | |
511 | * | |
b8072f09 | 512 | * The caller needs to hold the pte lock. |
1da177e4 LT |
513 | */ |
514 | void page_remove_rmap(struct page *page) | |
515 | { | |
1da177e4 LT |
516 | if (atomic_add_negative(-1, &page->_mapcount)) { |
517 | BUG_ON(page_mapcount(page) < 0); | |
518 | /* | |
519 | * It would be tidy to reset the PageAnon mapping here, | |
520 | * but that might overwrite a racing page_add_anon_rmap | |
521 | * which increments mapcount after us but sets mapping | |
522 | * before us: so leave the reset to free_hot_cold_page, | |
523 | * and remember that it's only reliable while mapped. | |
524 | * Leaving it set also helps swapoff to reinstate ptes | |
525 | * faster for those pages still in swapcache. | |
526 | */ | |
527 | if (page_test_and_clear_dirty(page)) | |
528 | set_page_dirty(page); | |
a74609fa | 529 | __dec_page_state(nr_mapped); |
1da177e4 LT |
530 | } |
531 | } | |
532 | ||
533 | /* | |
534 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
535 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
536 | */ | |
537 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma) | |
538 | { | |
539 | struct mm_struct *mm = vma->vm_mm; | |
540 | unsigned long address; | |
1da177e4 LT |
541 | pte_t *pte; |
542 | pte_t pteval; | |
c0718806 | 543 | spinlock_t *ptl; |
1da177e4 LT |
544 | int ret = SWAP_AGAIN; |
545 | ||
1da177e4 LT |
546 | address = vma_address(page, vma); |
547 | if (address == -EFAULT) | |
548 | goto out; | |
549 | ||
c0718806 HD |
550 | pte = page_check_address(page, mm, address, &ptl); |
551 | if (!pte) | |
81b4082d | 552 | goto out; |
1da177e4 LT |
553 | |
554 | /* | |
555 | * If the page is mlock()d, we cannot swap it out. | |
556 | * If it's recently referenced (perhaps page_referenced | |
557 | * skipped over this mm) then we should reactivate it. | |
558 | */ | |
101d2be7 | 559 | if ((vma->vm_flags & VM_LOCKED) || |
1da177e4 LT |
560 | ptep_clear_flush_young(vma, address, pte)) { |
561 | ret = SWAP_FAIL; | |
562 | goto out_unmap; | |
563 | } | |
564 | ||
1da177e4 LT |
565 | /* Nuke the page table entry. */ |
566 | flush_cache_page(vma, address, page_to_pfn(page)); | |
567 | pteval = ptep_clear_flush(vma, address, pte); | |
568 | ||
569 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
570 | if (pte_dirty(pteval)) | |
571 | set_page_dirty(page); | |
572 | ||
365e9c87 HD |
573 | /* Update high watermark before we lower rss */ |
574 | update_hiwater_rss(mm); | |
575 | ||
1da177e4 | 576 | if (PageAnon(page)) { |
4c21e2f2 | 577 | swp_entry_t entry = { .val = page_private(page) }; |
1da177e4 LT |
578 | /* |
579 | * Store the swap location in the pte. | |
580 | * See handle_pte_fault() ... | |
581 | */ | |
582 | BUG_ON(!PageSwapCache(page)); | |
583 | swap_duplicate(entry); | |
584 | if (list_empty(&mm->mmlist)) { | |
585 | spin_lock(&mmlist_lock); | |
f412ac08 HD |
586 | if (list_empty(&mm->mmlist)) |
587 | list_add(&mm->mmlist, &init_mm.mmlist); | |
1da177e4 LT |
588 | spin_unlock(&mmlist_lock); |
589 | } | |
590 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
591 | BUG_ON(pte_file(*pte)); | |
592 | dec_mm_counter(mm, anon_rss); | |
4294621f HD |
593 | } else |
594 | dec_mm_counter(mm, file_rss); | |
1da177e4 | 595 | |
1da177e4 LT |
596 | page_remove_rmap(page); |
597 | page_cache_release(page); | |
598 | ||
599 | out_unmap: | |
c0718806 | 600 | pte_unmap_unlock(pte, ptl); |
1da177e4 LT |
601 | out: |
602 | return ret; | |
603 | } | |
604 | ||
605 | /* | |
606 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
607 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
608 | * Consequently, given a particular page and its ->index, we cannot locate the | |
609 | * ptes which are mapping that page without an exhaustive linear search. | |
610 | * | |
611 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
612 | * maps the file to which the target page belongs. The ->vm_private_data field | |
613 | * holds the current cursor into that scan. Successive searches will circulate | |
614 | * around the vma's virtual address space. | |
615 | * | |
616 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
617 | * more scanning pressure is placed against them as well. Eventually pages | |
618 | * will become fully unmapped and are eligible for eviction. | |
619 | * | |
620 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
621 | * there there won't be many ptes located within the scan cluster. In this case | |
622 | * maybe we could scan further - to the end of the pte page, perhaps. | |
623 | */ | |
624 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
625 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
626 | ||
627 | static void try_to_unmap_cluster(unsigned long cursor, | |
628 | unsigned int *mapcount, struct vm_area_struct *vma) | |
629 | { | |
630 | struct mm_struct *mm = vma->vm_mm; | |
631 | pgd_t *pgd; | |
632 | pud_t *pud; | |
633 | pmd_t *pmd; | |
c0718806 | 634 | pte_t *pte; |
1da177e4 | 635 | pte_t pteval; |
c0718806 | 636 | spinlock_t *ptl; |
1da177e4 LT |
637 | struct page *page; |
638 | unsigned long address; | |
639 | unsigned long end; | |
1da177e4 | 640 | |
1da177e4 LT |
641 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
642 | end = address + CLUSTER_SIZE; | |
643 | if (address < vma->vm_start) | |
644 | address = vma->vm_start; | |
645 | if (end > vma->vm_end) | |
646 | end = vma->vm_end; | |
647 | ||
648 | pgd = pgd_offset(mm, address); | |
649 | if (!pgd_present(*pgd)) | |
c0718806 | 650 | return; |
1da177e4 LT |
651 | |
652 | pud = pud_offset(pgd, address); | |
653 | if (!pud_present(*pud)) | |
c0718806 | 654 | return; |
1da177e4 LT |
655 | |
656 | pmd = pmd_offset(pud, address); | |
657 | if (!pmd_present(*pmd)) | |
c0718806 HD |
658 | return; |
659 | ||
660 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 661 | |
365e9c87 HD |
662 | /* Update high watermark before we lower rss */ |
663 | update_hiwater_rss(mm); | |
664 | ||
c0718806 | 665 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
666 | if (!pte_present(*pte)) |
667 | continue; | |
6aab341e LT |
668 | page = vm_normal_page(vma, address, *pte); |
669 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 LT |
670 | |
671 | if (ptep_clear_flush_young(vma, address, pte)) | |
672 | continue; | |
673 | ||
674 | /* Nuke the page table entry. */ | |
eca35133 | 675 | flush_cache_page(vma, address, pte_pfn(*pte)); |
1da177e4 LT |
676 | pteval = ptep_clear_flush(vma, address, pte); |
677 | ||
678 | /* If nonlinear, store the file page offset in the pte. */ | |
679 | if (page->index != linear_page_index(vma, address)) | |
680 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
681 | ||
682 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
683 | if (pte_dirty(pteval)) | |
684 | set_page_dirty(page); | |
685 | ||
686 | page_remove_rmap(page); | |
687 | page_cache_release(page); | |
4294621f | 688 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
689 | (*mapcount)--; |
690 | } | |
c0718806 | 691 | pte_unmap_unlock(pte - 1, ptl); |
1da177e4 LT |
692 | } |
693 | ||
694 | static int try_to_unmap_anon(struct page *page) | |
695 | { | |
696 | struct anon_vma *anon_vma; | |
697 | struct vm_area_struct *vma; | |
698 | int ret = SWAP_AGAIN; | |
699 | ||
700 | anon_vma = page_lock_anon_vma(page); | |
701 | if (!anon_vma) | |
702 | return ret; | |
703 | ||
704 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
705 | ret = try_to_unmap_one(page, vma); | |
706 | if (ret == SWAP_FAIL || !page_mapped(page)) | |
707 | break; | |
708 | } | |
709 | spin_unlock(&anon_vma->lock); | |
710 | return ret; | |
711 | } | |
712 | ||
713 | /** | |
714 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
715 | * @page: the page to unmap | |
716 | * | |
717 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
718 | * contained in the address_space struct it points to. | |
719 | * | |
720 | * This function is only called from try_to_unmap for object-based pages. | |
721 | */ | |
722 | static int try_to_unmap_file(struct page *page) | |
723 | { | |
724 | struct address_space *mapping = page->mapping; | |
725 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
726 | struct vm_area_struct *vma; | |
727 | struct prio_tree_iter iter; | |
728 | int ret = SWAP_AGAIN; | |
729 | unsigned long cursor; | |
730 | unsigned long max_nl_cursor = 0; | |
731 | unsigned long max_nl_size = 0; | |
732 | unsigned int mapcount; | |
733 | ||
734 | spin_lock(&mapping->i_mmap_lock); | |
735 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
736 | ret = try_to_unmap_one(page, vma); | |
737 | if (ret == SWAP_FAIL || !page_mapped(page)) | |
738 | goto out; | |
739 | } | |
740 | ||
741 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
742 | goto out; | |
743 | ||
744 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
745 | shared.vm_set.list) { | |
101d2be7 | 746 | if (vma->vm_flags & VM_LOCKED) |
1da177e4 LT |
747 | continue; |
748 | cursor = (unsigned long) vma->vm_private_data; | |
749 | if (cursor > max_nl_cursor) | |
750 | max_nl_cursor = cursor; | |
751 | cursor = vma->vm_end - vma->vm_start; | |
752 | if (cursor > max_nl_size) | |
753 | max_nl_size = cursor; | |
754 | } | |
755 | ||
756 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
757 | ret = SWAP_FAIL; | |
758 | goto out; | |
759 | } | |
760 | ||
761 | /* | |
762 | * We don't try to search for this page in the nonlinear vmas, | |
763 | * and page_referenced wouldn't have found it anyway. Instead | |
764 | * just walk the nonlinear vmas trying to age and unmap some. | |
765 | * The mapcount of the page we came in with is irrelevant, | |
766 | * but even so use it as a guide to how hard we should try? | |
767 | */ | |
768 | mapcount = page_mapcount(page); | |
769 | if (!mapcount) | |
770 | goto out; | |
771 | cond_resched_lock(&mapping->i_mmap_lock); | |
772 | ||
773 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
774 | if (max_nl_cursor == 0) | |
775 | max_nl_cursor = CLUSTER_SIZE; | |
776 | ||
777 | do { | |
778 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
779 | shared.vm_set.list) { | |
101d2be7 | 780 | if (vma->vm_flags & VM_LOCKED) |
1da177e4 LT |
781 | continue; |
782 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 783 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
784 | cursor < vma->vm_end - vma->vm_start) { |
785 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
786 | cursor += CLUSTER_SIZE; | |
787 | vma->vm_private_data = (void *) cursor; | |
788 | if ((int)mapcount <= 0) | |
789 | goto out; | |
790 | } | |
791 | vma->vm_private_data = (void *) max_nl_cursor; | |
792 | } | |
793 | cond_resched_lock(&mapping->i_mmap_lock); | |
794 | max_nl_cursor += CLUSTER_SIZE; | |
795 | } while (max_nl_cursor <= max_nl_size); | |
796 | ||
797 | /* | |
798 | * Don't loop forever (perhaps all the remaining pages are | |
799 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
800 | * vmas, now forgetting on which ones it had fallen behind. | |
801 | */ | |
101d2be7 HD |
802 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
803 | vma->vm_private_data = NULL; | |
1da177e4 LT |
804 | out: |
805 | spin_unlock(&mapping->i_mmap_lock); | |
806 | return ret; | |
807 | } | |
808 | ||
809 | /** | |
810 | * try_to_unmap - try to remove all page table mappings to a page | |
811 | * @page: the page to get unmapped | |
812 | * | |
813 | * Tries to remove all the page table entries which are mapping this | |
814 | * page, used in the pageout path. Caller must hold the page lock. | |
815 | * Return values are: | |
816 | * | |
817 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
818 | * SWAP_AGAIN - we missed a mapping, try again later | |
819 | * SWAP_FAIL - the page is unswappable | |
820 | */ | |
821 | int try_to_unmap(struct page *page) | |
822 | { | |
823 | int ret; | |
824 | ||
1da177e4 LT |
825 | BUG_ON(!PageLocked(page)); |
826 | ||
827 | if (PageAnon(page)) | |
828 | ret = try_to_unmap_anon(page); | |
829 | else | |
830 | ret = try_to_unmap_file(page); | |
831 | ||
832 | if (!page_mapped(page)) | |
833 | ret = SWAP_SUCCESS; | |
834 | return ret; | |
835 | } | |
81b4082d | 836 |