Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * mm/rmap.c - physical to virtual reverse mappings | |
3 | * | |
4 | * Copyright 2001, Rik van Riel <riel@conectiva.com.br> | |
5 | * Released under the General Public License (GPL). | |
6 | * | |
7 | * Simple, low overhead reverse mapping scheme. | |
8 | * Please try to keep this thing as modular as possible. | |
9 | * | |
10 | * Provides methods for unmapping each kind of mapped page: | |
11 | * the anon methods track anonymous pages, and | |
12 | * the file methods track pages belonging to an inode. | |
13 | * | |
14 | * Original design by Rik van Riel <riel@conectiva.com.br> 2001 | |
15 | * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 | |
16 | * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 | |
17 | * Contributions by Hugh Dickins <hugh@veritas.com> 2003, 2004 | |
18 | */ | |
19 | ||
20 | /* | |
21 | * Lock ordering in mm: | |
22 | * | |
1b1dcc1b | 23 | * inode->i_mutex (while writing or truncating, not reading or faulting) |
82591e6e NP |
24 | * inode->i_alloc_sem (vmtruncate_range) |
25 | * mm->mmap_sem | |
26 | * page->flags PG_locked (lock_page) | |
27 | * mapping->i_mmap_lock | |
28 | * anon_vma->lock | |
29 | * mm->page_table_lock or pte_lock | |
30 | * zone->lru_lock (in mark_page_accessed, isolate_lru_page) | |
31 | * swap_lock (in swap_duplicate, swap_info_get) | |
32 | * mmlist_lock (in mmput, drain_mmlist and others) | |
33 | * mapping->private_lock (in __set_page_dirty_buffers) | |
34 | * inode_lock (in set_page_dirty's __mark_inode_dirty) | |
35 | * sb_lock (within inode_lock in fs/fs-writeback.c) | |
36 | * mapping->tree_lock (widely used, in set_page_dirty, | |
37 | * in arch-dependent flush_dcache_mmap_lock, | |
38 | * within inode_lock in __sync_single_inode) | |
1da177e4 LT |
39 | */ |
40 | ||
41 | #include <linux/mm.h> | |
42 | #include <linux/pagemap.h> | |
43 | #include <linux/swap.h> | |
44 | #include <linux/swapops.h> | |
45 | #include <linux/slab.h> | |
46 | #include <linux/init.h> | |
47 | #include <linux/rmap.h> | |
48 | #include <linux/rcupdate.h> | |
a48d07af | 49 | #include <linux/module.h> |
7de6b805 | 50 | #include <linux/kallsyms.h> |
1da177e4 LT |
51 | |
52 | #include <asm/tlbflush.h> | |
53 | ||
fcc234f8 | 54 | struct kmem_cache *anon_vma_cachep; |
1da177e4 LT |
55 | |
56 | static inline void validate_anon_vma(struct vm_area_struct *find_vma) | |
57 | { | |
b7ab795b | 58 | #ifdef CONFIG_DEBUG_VM |
1da177e4 LT |
59 | struct anon_vma *anon_vma = find_vma->anon_vma; |
60 | struct vm_area_struct *vma; | |
61 | unsigned int mapcount = 0; | |
62 | int found = 0; | |
63 | ||
64 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
65 | mapcount++; | |
66 | BUG_ON(mapcount > 100000); | |
67 | if (vma == find_vma) | |
68 | found = 1; | |
69 | } | |
70 | BUG_ON(!found); | |
71 | #endif | |
72 | } | |
73 | ||
74 | /* This must be called under the mmap_sem. */ | |
75 | int anon_vma_prepare(struct vm_area_struct *vma) | |
76 | { | |
77 | struct anon_vma *anon_vma = vma->anon_vma; | |
78 | ||
79 | might_sleep(); | |
80 | if (unlikely(!anon_vma)) { | |
81 | struct mm_struct *mm = vma->vm_mm; | |
82 | struct anon_vma *allocated, *locked; | |
83 | ||
84 | anon_vma = find_mergeable_anon_vma(vma); | |
85 | if (anon_vma) { | |
86 | allocated = NULL; | |
87 | locked = anon_vma; | |
88 | spin_lock(&locked->lock); | |
89 | } else { | |
90 | anon_vma = anon_vma_alloc(); | |
91 | if (unlikely(!anon_vma)) | |
92 | return -ENOMEM; | |
93 | allocated = anon_vma; | |
94 | locked = NULL; | |
95 | } | |
96 | ||
97 | /* page_table_lock to protect against threads */ | |
98 | spin_lock(&mm->page_table_lock); | |
99 | if (likely(!vma->anon_vma)) { | |
100 | vma->anon_vma = anon_vma; | |
0697212a | 101 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
102 | allocated = NULL; |
103 | } | |
104 | spin_unlock(&mm->page_table_lock); | |
105 | ||
106 | if (locked) | |
107 | spin_unlock(&locked->lock); | |
108 | if (unlikely(allocated)) | |
109 | anon_vma_free(allocated); | |
110 | } | |
111 | return 0; | |
112 | } | |
113 | ||
114 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) | |
115 | { | |
116 | BUG_ON(vma->anon_vma != next->anon_vma); | |
117 | list_del(&next->anon_vma_node); | |
118 | } | |
119 | ||
120 | void __anon_vma_link(struct vm_area_struct *vma) | |
121 | { | |
122 | struct anon_vma *anon_vma = vma->anon_vma; | |
123 | ||
124 | if (anon_vma) { | |
0697212a | 125 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
126 | validate_anon_vma(vma); |
127 | } | |
128 | } | |
129 | ||
130 | void anon_vma_link(struct vm_area_struct *vma) | |
131 | { | |
132 | struct anon_vma *anon_vma = vma->anon_vma; | |
133 | ||
134 | if (anon_vma) { | |
135 | spin_lock(&anon_vma->lock); | |
0697212a | 136 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
137 | validate_anon_vma(vma); |
138 | spin_unlock(&anon_vma->lock); | |
139 | } | |
140 | } | |
141 | ||
142 | void anon_vma_unlink(struct vm_area_struct *vma) | |
143 | { | |
144 | struct anon_vma *anon_vma = vma->anon_vma; | |
145 | int empty; | |
146 | ||
147 | if (!anon_vma) | |
148 | return; | |
149 | ||
150 | spin_lock(&anon_vma->lock); | |
151 | validate_anon_vma(vma); | |
152 | list_del(&vma->anon_vma_node); | |
153 | ||
154 | /* We must garbage collect the anon_vma if it's empty */ | |
155 | empty = list_empty(&anon_vma->head); | |
156 | spin_unlock(&anon_vma->lock); | |
157 | ||
158 | if (empty) | |
159 | anon_vma_free(anon_vma); | |
160 | } | |
161 | ||
fcc234f8 PE |
162 | static void anon_vma_ctor(void *data, struct kmem_cache *cachep, |
163 | unsigned long flags) | |
1da177e4 | 164 | { |
a35afb83 | 165 | struct anon_vma *anon_vma = data; |
1da177e4 | 166 | |
a35afb83 CL |
167 | spin_lock_init(&anon_vma->lock); |
168 | INIT_LIST_HEAD(&anon_vma->head); | |
1da177e4 LT |
169 | } |
170 | ||
171 | void __init anon_vma_init(void) | |
172 | { | |
173 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
174 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL); | |
175 | } | |
176 | ||
177 | /* | |
178 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
179 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
180 | */ | |
181 | static struct anon_vma *page_lock_anon_vma(struct page *page) | |
182 | { | |
34bbd704 | 183 | struct anon_vma *anon_vma; |
1da177e4 LT |
184 | unsigned long anon_mapping; |
185 | ||
186 | rcu_read_lock(); | |
187 | anon_mapping = (unsigned long) page->mapping; | |
188 | if (!(anon_mapping & PAGE_MAPPING_ANON)) | |
189 | goto out; | |
190 | if (!page_mapped(page)) | |
191 | goto out; | |
192 | ||
193 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
194 | spin_lock(&anon_vma->lock); | |
34bbd704 | 195 | return anon_vma; |
1da177e4 LT |
196 | out: |
197 | rcu_read_unlock(); | |
34bbd704 ON |
198 | return NULL; |
199 | } | |
200 | ||
201 | static void page_unlock_anon_vma(struct anon_vma *anon_vma) | |
202 | { | |
203 | spin_unlock(&anon_vma->lock); | |
204 | rcu_read_unlock(); | |
1da177e4 LT |
205 | } |
206 | ||
207 | /* | |
208 | * At what user virtual address is page expected in vma? | |
209 | */ | |
210 | static inline unsigned long | |
211 | vma_address(struct page *page, struct vm_area_struct *vma) | |
212 | { | |
213 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
214 | unsigned long address; | |
215 | ||
216 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
217 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
218 | /* page should be within any vma from prio_tree_next */ | |
219 | BUG_ON(!PageAnon(page)); | |
220 | return -EFAULT; | |
221 | } | |
222 | return address; | |
223 | } | |
224 | ||
225 | /* | |
226 | * At what user virtual address is page expected in vma? checking that the | |
ee498ed7 | 227 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
1da177e4 LT |
228 | */ |
229 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
230 | { | |
231 | if (PageAnon(page)) { | |
232 | if ((void *)vma->anon_vma != | |
233 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
234 | return -EFAULT; | |
235 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
236 | if (!vma->vm_file || |
237 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
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 | * | |
b8072f09 | 247 | * On success returns with pte mapped and locked. |
81b4082d | 248 | */ |
ceffc078 | 249 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
c0718806 | 250 | unsigned long address, spinlock_t **ptlp) |
81b4082d ND |
251 | { |
252 | pgd_t *pgd; | |
253 | pud_t *pud; | |
254 | pmd_t *pmd; | |
255 | pte_t *pte; | |
c0718806 | 256 | spinlock_t *ptl; |
81b4082d | 257 | |
81b4082d | 258 | pgd = pgd_offset(mm, address); |
c0718806 HD |
259 | if (!pgd_present(*pgd)) |
260 | return NULL; | |
261 | ||
262 | pud = pud_offset(pgd, address); | |
263 | if (!pud_present(*pud)) | |
264 | return NULL; | |
265 | ||
266 | pmd = pmd_offset(pud, address); | |
267 | if (!pmd_present(*pmd)) | |
268 | return NULL; | |
269 | ||
270 | pte = pte_offset_map(pmd, address); | |
271 | /* Make a quick check before getting the lock */ | |
272 | if (!pte_present(*pte)) { | |
273 | pte_unmap(pte); | |
274 | return NULL; | |
275 | } | |
276 | ||
4c21e2f2 | 277 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
278 | spin_lock(ptl); |
279 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
280 | *ptlp = ptl; | |
281 | return pte; | |
81b4082d | 282 | } |
c0718806 HD |
283 | pte_unmap_unlock(pte, ptl); |
284 | return NULL; | |
81b4082d ND |
285 | } |
286 | ||
1da177e4 LT |
287 | /* |
288 | * Subfunctions of page_referenced: page_referenced_one called | |
289 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
290 | */ | |
291 | static int page_referenced_one(struct page *page, | |
f7b7fd8f | 292 | struct vm_area_struct *vma, unsigned int *mapcount) |
1da177e4 LT |
293 | { |
294 | struct mm_struct *mm = vma->vm_mm; | |
295 | unsigned long address; | |
1da177e4 | 296 | pte_t *pte; |
c0718806 | 297 | spinlock_t *ptl; |
1da177e4 LT |
298 | int referenced = 0; |
299 | ||
1da177e4 LT |
300 | address = vma_address(page, vma); |
301 | if (address == -EFAULT) | |
302 | goto out; | |
303 | ||
c0718806 HD |
304 | pte = page_check_address(page, mm, address, &ptl); |
305 | if (!pte) | |
306 | goto out; | |
1da177e4 | 307 | |
c0718806 HD |
308 | if (ptep_clear_flush_young(vma, address, pte)) |
309 | referenced++; | |
1da177e4 | 310 | |
c0718806 HD |
311 | /* Pretend the page is referenced if the task has the |
312 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 313 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
314 | rwsem_is_locked(&mm->mmap_sem)) |
315 | referenced++; | |
316 | ||
317 | (*mapcount)--; | |
318 | pte_unmap_unlock(pte, ptl); | |
1da177e4 LT |
319 | out: |
320 | return referenced; | |
321 | } | |
322 | ||
f7b7fd8f | 323 | static int page_referenced_anon(struct page *page) |
1da177e4 LT |
324 | { |
325 | unsigned int mapcount; | |
326 | struct anon_vma *anon_vma; | |
327 | struct vm_area_struct *vma; | |
328 | int referenced = 0; | |
329 | ||
330 | anon_vma = page_lock_anon_vma(page); | |
331 | if (!anon_vma) | |
332 | return referenced; | |
333 | ||
334 | mapcount = page_mapcount(page); | |
335 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
f7b7fd8f | 336 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
337 | if (!mapcount) |
338 | break; | |
339 | } | |
34bbd704 ON |
340 | |
341 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
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 | ||
d08b3851 PZ |
437 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma) |
438 | { | |
439 | struct mm_struct *mm = vma->vm_mm; | |
440 | unsigned long address; | |
c2fda5fe | 441 | pte_t *pte; |
d08b3851 PZ |
442 | spinlock_t *ptl; |
443 | int ret = 0; | |
444 | ||
445 | address = vma_address(page, vma); | |
446 | if (address == -EFAULT) | |
447 | goto out; | |
448 | ||
449 | pte = page_check_address(page, mm, address, &ptl); | |
450 | if (!pte) | |
451 | goto out; | |
452 | ||
c2fda5fe PZ |
453 | if (pte_dirty(*pte) || pte_write(*pte)) { |
454 | pte_t entry; | |
d08b3851 | 455 | |
c2fda5fe PZ |
456 | flush_cache_page(vma, address, pte_pfn(*pte)); |
457 | entry = ptep_clear_flush(vma, address, pte); | |
458 | entry = pte_wrprotect(entry); | |
459 | entry = pte_mkclean(entry); | |
d6e88e67 | 460 | set_pte_at(mm, address, pte, entry); |
c2fda5fe PZ |
461 | lazy_mmu_prot_update(entry); |
462 | ret = 1; | |
463 | } | |
d08b3851 | 464 | |
d08b3851 PZ |
465 | pte_unmap_unlock(pte, ptl); |
466 | out: | |
467 | return ret; | |
468 | } | |
469 | ||
470 | static int page_mkclean_file(struct address_space *mapping, struct page *page) | |
471 | { | |
472 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
473 | struct vm_area_struct *vma; | |
474 | struct prio_tree_iter iter; | |
475 | int ret = 0; | |
476 | ||
477 | BUG_ON(PageAnon(page)); | |
478 | ||
479 | spin_lock(&mapping->i_mmap_lock); | |
480 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
481 | if (vma->vm_flags & VM_SHARED) | |
482 | ret += page_mkclean_one(page, vma); | |
483 | } | |
484 | spin_unlock(&mapping->i_mmap_lock); | |
485 | return ret; | |
486 | } | |
487 | ||
488 | int page_mkclean(struct page *page) | |
489 | { | |
490 | int ret = 0; | |
491 | ||
492 | BUG_ON(!PageLocked(page)); | |
493 | ||
494 | if (page_mapped(page)) { | |
495 | struct address_space *mapping = page_mapping(page); | |
496 | if (mapping) | |
497 | ret = page_mkclean_file(mapping, page); | |
6c210482 MS |
498 | if (page_test_dirty(page)) { |
499 | page_clear_dirty(page); | |
6e1beb3c | 500 | ret = 1; |
6c210482 | 501 | } |
d08b3851 PZ |
502 | } |
503 | ||
504 | return ret; | |
505 | } | |
60b59bea | 506 | EXPORT_SYMBOL_GPL(page_mkclean); |
d08b3851 | 507 | |
9617d95e NP |
508 | /** |
509 | * page_set_anon_rmap - setup new anonymous rmap | |
510 | * @page: the page to add the mapping to | |
511 | * @vma: the vm area in which the mapping is added | |
512 | * @address: the user virtual address mapped | |
513 | */ | |
514 | static void __page_set_anon_rmap(struct page *page, | |
515 | struct vm_area_struct *vma, unsigned long address) | |
516 | { | |
517 | struct anon_vma *anon_vma = vma->anon_vma; | |
518 | ||
519 | BUG_ON(!anon_vma); | |
520 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
521 | page->mapping = (struct address_space *) anon_vma; | |
522 | ||
523 | page->index = linear_page_index(vma, address); | |
524 | ||
a74609fa NP |
525 | /* |
526 | * nr_mapped state can be updated without turning off | |
527 | * interrupts because it is not modified via interrupt. | |
528 | */ | |
f3dbd344 | 529 | __inc_zone_page_state(page, NR_ANON_PAGES); |
9617d95e NP |
530 | } |
531 | ||
1da177e4 LT |
532 | /** |
533 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
534 | * @page: the page to add the mapping to | |
535 | * @vma: the vm area in which the mapping is added | |
536 | * @address: the user virtual address mapped | |
537 | * | |
b8072f09 | 538 | * The caller needs to hold the pte lock. |
1da177e4 LT |
539 | */ |
540 | void page_add_anon_rmap(struct page *page, | |
541 | struct vm_area_struct *vma, unsigned long address) | |
542 | { | |
9617d95e NP |
543 | if (atomic_inc_and_test(&page->_mapcount)) |
544 | __page_set_anon_rmap(page, vma, address); | |
1da177e4 LT |
545 | /* else checking page index and mapping is racy */ |
546 | } | |
547 | ||
9617d95e NP |
548 | /* |
549 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
550 | * @page: the page to add the mapping to | |
551 | * @vma: the vm area in which the mapping is added | |
552 | * @address: the user virtual address mapped | |
553 | * | |
554 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
555 | * This means the inc-and-test can be bypassed. | |
556 | */ | |
557 | void page_add_new_anon_rmap(struct page *page, | |
558 | struct vm_area_struct *vma, unsigned long address) | |
559 | { | |
560 | atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ | |
561 | __page_set_anon_rmap(page, vma, address); | |
562 | } | |
563 | ||
1da177e4 LT |
564 | /** |
565 | * page_add_file_rmap - add pte mapping to a file page | |
566 | * @page: the page to add the mapping to | |
567 | * | |
b8072f09 | 568 | * The caller needs to hold the pte lock. |
1da177e4 LT |
569 | */ |
570 | void page_add_file_rmap(struct page *page) | |
571 | { | |
1da177e4 | 572 | if (atomic_inc_and_test(&page->_mapcount)) |
65ba55f5 | 573 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
1da177e4 LT |
574 | } |
575 | ||
576 | /** | |
577 | * page_remove_rmap - take down pte mapping from a page | |
578 | * @page: page to remove mapping from | |
579 | * | |
b8072f09 | 580 | * The caller needs to hold the pte lock. |
1da177e4 | 581 | */ |
7de6b805 | 582 | void page_remove_rmap(struct page *page, struct vm_area_struct *vma) |
1da177e4 | 583 | { |
1da177e4 | 584 | if (atomic_add_negative(-1, &page->_mapcount)) { |
b7ab795b | 585 | if (unlikely(page_mapcount(page) < 0)) { |
ef2bf0dc | 586 | printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page)); |
7de6b805 | 587 | printk (KERN_EMERG " page pfn = %lx\n", page_to_pfn(page)); |
ef2bf0dc DJ |
588 | printk (KERN_EMERG " page->flags = %lx\n", page->flags); |
589 | printk (KERN_EMERG " page->count = %x\n", page_count(page)); | |
590 | printk (KERN_EMERG " page->mapping = %p\n", page->mapping); | |
7de6b805 NP |
591 | print_symbol (KERN_EMERG " vma->vm_ops = %s\n", (unsigned long)vma->vm_ops); |
592 | if (vma->vm_ops) | |
593 | print_symbol (KERN_EMERG " vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage); | |
594 | if (vma->vm_file && vma->vm_file->f_op) | |
595 | print_symbol (KERN_EMERG " vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap); | |
b16bc64d | 596 | BUG(); |
ef2bf0dc | 597 | } |
b16bc64d | 598 | |
1da177e4 LT |
599 | /* |
600 | * It would be tidy to reset the PageAnon mapping here, | |
601 | * but that might overwrite a racing page_add_anon_rmap | |
602 | * which increments mapcount after us but sets mapping | |
603 | * before us: so leave the reset to free_hot_cold_page, | |
604 | * and remember that it's only reliable while mapped. | |
605 | * Leaving it set also helps swapoff to reinstate ptes | |
606 | * faster for those pages still in swapcache. | |
607 | */ | |
6c210482 MS |
608 | if (page_test_dirty(page)) { |
609 | page_clear_dirty(page); | |
1da177e4 | 610 | set_page_dirty(page); |
6c210482 | 611 | } |
f3dbd344 CL |
612 | __dec_zone_page_state(page, |
613 | PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED); | |
1da177e4 LT |
614 | } |
615 | } | |
616 | ||
617 | /* | |
618 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
619 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
620 | */ | |
a48d07af | 621 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
7352349a | 622 | int migration) |
1da177e4 LT |
623 | { |
624 | struct mm_struct *mm = vma->vm_mm; | |
625 | unsigned long address; | |
1da177e4 LT |
626 | pte_t *pte; |
627 | pte_t pteval; | |
c0718806 | 628 | spinlock_t *ptl; |
1da177e4 LT |
629 | int ret = SWAP_AGAIN; |
630 | ||
1da177e4 LT |
631 | address = vma_address(page, vma); |
632 | if (address == -EFAULT) | |
633 | goto out; | |
634 | ||
c0718806 HD |
635 | pte = page_check_address(page, mm, address, &ptl); |
636 | if (!pte) | |
81b4082d | 637 | goto out; |
1da177e4 LT |
638 | |
639 | /* | |
640 | * If the page is mlock()d, we cannot swap it out. | |
641 | * If it's recently referenced (perhaps page_referenced | |
642 | * skipped over this mm) then we should reactivate it. | |
643 | */ | |
e6a1530d CL |
644 | if (!migration && ((vma->vm_flags & VM_LOCKED) || |
645 | (ptep_clear_flush_young(vma, address, pte)))) { | |
1da177e4 LT |
646 | ret = SWAP_FAIL; |
647 | goto out_unmap; | |
648 | } | |
649 | ||
1da177e4 LT |
650 | /* Nuke the page table entry. */ |
651 | flush_cache_page(vma, address, page_to_pfn(page)); | |
652 | pteval = ptep_clear_flush(vma, address, pte); | |
653 | ||
654 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
655 | if (pte_dirty(pteval)) | |
656 | set_page_dirty(page); | |
657 | ||
365e9c87 HD |
658 | /* Update high watermark before we lower rss */ |
659 | update_hiwater_rss(mm); | |
660 | ||
1da177e4 | 661 | if (PageAnon(page)) { |
4c21e2f2 | 662 | swp_entry_t entry = { .val = page_private(page) }; |
0697212a CL |
663 | |
664 | if (PageSwapCache(page)) { | |
665 | /* | |
666 | * Store the swap location in the pte. | |
667 | * See handle_pte_fault() ... | |
668 | */ | |
669 | swap_duplicate(entry); | |
670 | if (list_empty(&mm->mmlist)) { | |
671 | spin_lock(&mmlist_lock); | |
672 | if (list_empty(&mm->mmlist)) | |
673 | list_add(&mm->mmlist, &init_mm.mmlist); | |
674 | spin_unlock(&mmlist_lock); | |
675 | } | |
442c9137 | 676 | dec_mm_counter(mm, anon_rss); |
04e62a29 | 677 | #ifdef CONFIG_MIGRATION |
0697212a CL |
678 | } else { |
679 | /* | |
680 | * Store the pfn of the page in a special migration | |
681 | * pte. do_swap_page() will wait until the migration | |
682 | * pte is removed and then restart fault handling. | |
683 | */ | |
684 | BUG_ON(!migration); | |
685 | entry = make_migration_entry(page, pte_write(pteval)); | |
04e62a29 | 686 | #endif |
1da177e4 LT |
687 | } |
688 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
689 | BUG_ON(pte_file(*pte)); | |
4294621f | 690 | } else |
04e62a29 CL |
691 | #ifdef CONFIG_MIGRATION |
692 | if (migration) { | |
693 | /* Establish migration entry for a file page */ | |
694 | swp_entry_t entry; | |
695 | entry = make_migration_entry(page, pte_write(pteval)); | |
696 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
697 | } else | |
698 | #endif | |
4294621f | 699 | dec_mm_counter(mm, file_rss); |
1da177e4 | 700 | |
04e62a29 | 701 | |
7de6b805 | 702 | page_remove_rmap(page, vma); |
1da177e4 LT |
703 | page_cache_release(page); |
704 | ||
705 | out_unmap: | |
c0718806 | 706 | pte_unmap_unlock(pte, ptl); |
1da177e4 LT |
707 | out: |
708 | return ret; | |
709 | } | |
710 | ||
711 | /* | |
712 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
713 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
714 | * Consequently, given a particular page and its ->index, we cannot locate the | |
715 | * ptes which are mapping that page without an exhaustive linear search. | |
716 | * | |
717 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
718 | * maps the file to which the target page belongs. The ->vm_private_data field | |
719 | * holds the current cursor into that scan. Successive searches will circulate | |
720 | * around the vma's virtual address space. | |
721 | * | |
722 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
723 | * more scanning pressure is placed against them as well. Eventually pages | |
724 | * will become fully unmapped and are eligible for eviction. | |
725 | * | |
726 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
727 | * there there won't be many ptes located within the scan cluster. In this case | |
728 | * maybe we could scan further - to the end of the pte page, perhaps. | |
729 | */ | |
730 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
731 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
732 | ||
733 | static void try_to_unmap_cluster(unsigned long cursor, | |
734 | unsigned int *mapcount, struct vm_area_struct *vma) | |
735 | { | |
736 | struct mm_struct *mm = vma->vm_mm; | |
737 | pgd_t *pgd; | |
738 | pud_t *pud; | |
739 | pmd_t *pmd; | |
c0718806 | 740 | pte_t *pte; |
1da177e4 | 741 | pte_t pteval; |
c0718806 | 742 | spinlock_t *ptl; |
1da177e4 LT |
743 | struct page *page; |
744 | unsigned long address; | |
745 | unsigned long end; | |
1da177e4 | 746 | |
1da177e4 LT |
747 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
748 | end = address + CLUSTER_SIZE; | |
749 | if (address < vma->vm_start) | |
750 | address = vma->vm_start; | |
751 | if (end > vma->vm_end) | |
752 | end = vma->vm_end; | |
753 | ||
754 | pgd = pgd_offset(mm, address); | |
755 | if (!pgd_present(*pgd)) | |
c0718806 | 756 | return; |
1da177e4 LT |
757 | |
758 | pud = pud_offset(pgd, address); | |
759 | if (!pud_present(*pud)) | |
c0718806 | 760 | return; |
1da177e4 LT |
761 | |
762 | pmd = pmd_offset(pud, address); | |
763 | if (!pmd_present(*pmd)) | |
c0718806 HD |
764 | return; |
765 | ||
766 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 767 | |
365e9c87 HD |
768 | /* Update high watermark before we lower rss */ |
769 | update_hiwater_rss(mm); | |
770 | ||
c0718806 | 771 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
772 | if (!pte_present(*pte)) |
773 | continue; | |
6aab341e LT |
774 | page = vm_normal_page(vma, address, *pte); |
775 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 LT |
776 | |
777 | if (ptep_clear_flush_young(vma, address, pte)) | |
778 | continue; | |
779 | ||
780 | /* Nuke the page table entry. */ | |
eca35133 | 781 | flush_cache_page(vma, address, pte_pfn(*pte)); |
1da177e4 LT |
782 | pteval = ptep_clear_flush(vma, address, pte); |
783 | ||
784 | /* If nonlinear, store the file page offset in the pte. */ | |
785 | if (page->index != linear_page_index(vma, address)) | |
786 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
787 | ||
788 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
789 | if (pte_dirty(pteval)) | |
790 | set_page_dirty(page); | |
791 | ||
7de6b805 | 792 | page_remove_rmap(page, vma); |
1da177e4 | 793 | page_cache_release(page); |
4294621f | 794 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
795 | (*mapcount)--; |
796 | } | |
c0718806 | 797 | pte_unmap_unlock(pte - 1, ptl); |
1da177e4 LT |
798 | } |
799 | ||
7352349a | 800 | static int try_to_unmap_anon(struct page *page, int migration) |
1da177e4 LT |
801 | { |
802 | struct anon_vma *anon_vma; | |
803 | struct vm_area_struct *vma; | |
804 | int ret = SWAP_AGAIN; | |
805 | ||
806 | anon_vma = page_lock_anon_vma(page); | |
807 | if (!anon_vma) | |
808 | return ret; | |
809 | ||
810 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
7352349a | 811 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
812 | if (ret == SWAP_FAIL || !page_mapped(page)) |
813 | break; | |
814 | } | |
34bbd704 ON |
815 | |
816 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
817 | return ret; |
818 | } | |
819 | ||
820 | /** | |
821 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
822 | * @page: the page to unmap | |
823 | * | |
824 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
825 | * contained in the address_space struct it points to. | |
826 | * | |
827 | * This function is only called from try_to_unmap for object-based pages. | |
828 | */ | |
7352349a | 829 | static int try_to_unmap_file(struct page *page, int migration) |
1da177e4 LT |
830 | { |
831 | struct address_space *mapping = page->mapping; | |
832 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
833 | struct vm_area_struct *vma; | |
834 | struct prio_tree_iter iter; | |
835 | int ret = SWAP_AGAIN; | |
836 | unsigned long cursor; | |
837 | unsigned long max_nl_cursor = 0; | |
838 | unsigned long max_nl_size = 0; | |
839 | unsigned int mapcount; | |
840 | ||
841 | spin_lock(&mapping->i_mmap_lock); | |
842 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
7352349a | 843 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
844 | if (ret == SWAP_FAIL || !page_mapped(page)) |
845 | goto out; | |
846 | } | |
847 | ||
848 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
849 | goto out; | |
850 | ||
851 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
852 | shared.vm_set.list) { | |
e6a1530d | 853 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
854 | continue; |
855 | cursor = (unsigned long) vma->vm_private_data; | |
856 | if (cursor > max_nl_cursor) | |
857 | max_nl_cursor = cursor; | |
858 | cursor = vma->vm_end - vma->vm_start; | |
859 | if (cursor > max_nl_size) | |
860 | max_nl_size = cursor; | |
861 | } | |
862 | ||
863 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
864 | ret = SWAP_FAIL; | |
865 | goto out; | |
866 | } | |
867 | ||
868 | /* | |
869 | * We don't try to search for this page in the nonlinear vmas, | |
870 | * and page_referenced wouldn't have found it anyway. Instead | |
871 | * just walk the nonlinear vmas trying to age and unmap some. | |
872 | * The mapcount of the page we came in with is irrelevant, | |
873 | * but even so use it as a guide to how hard we should try? | |
874 | */ | |
875 | mapcount = page_mapcount(page); | |
876 | if (!mapcount) | |
877 | goto out; | |
878 | cond_resched_lock(&mapping->i_mmap_lock); | |
879 | ||
880 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
881 | if (max_nl_cursor == 0) | |
882 | max_nl_cursor = CLUSTER_SIZE; | |
883 | ||
884 | do { | |
885 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
886 | shared.vm_set.list) { | |
e6a1530d | 887 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
888 | continue; |
889 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 890 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
891 | cursor < vma->vm_end - vma->vm_start) { |
892 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
893 | cursor += CLUSTER_SIZE; | |
894 | vma->vm_private_data = (void *) cursor; | |
895 | if ((int)mapcount <= 0) | |
896 | goto out; | |
897 | } | |
898 | vma->vm_private_data = (void *) max_nl_cursor; | |
899 | } | |
900 | cond_resched_lock(&mapping->i_mmap_lock); | |
901 | max_nl_cursor += CLUSTER_SIZE; | |
902 | } while (max_nl_cursor <= max_nl_size); | |
903 | ||
904 | /* | |
905 | * Don't loop forever (perhaps all the remaining pages are | |
906 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
907 | * vmas, now forgetting on which ones it had fallen behind. | |
908 | */ | |
101d2be7 HD |
909 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
910 | vma->vm_private_data = NULL; | |
1da177e4 LT |
911 | out: |
912 | spin_unlock(&mapping->i_mmap_lock); | |
913 | return ret; | |
914 | } | |
915 | ||
916 | /** | |
917 | * try_to_unmap - try to remove all page table mappings to a page | |
918 | * @page: the page to get unmapped | |
919 | * | |
920 | * Tries to remove all the page table entries which are mapping this | |
921 | * page, used in the pageout path. Caller must hold the page lock. | |
922 | * Return values are: | |
923 | * | |
924 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
925 | * SWAP_AGAIN - we missed a mapping, try again later | |
926 | * SWAP_FAIL - the page is unswappable | |
927 | */ | |
7352349a | 928 | int try_to_unmap(struct page *page, int migration) |
1da177e4 LT |
929 | { |
930 | int ret; | |
931 | ||
1da177e4 LT |
932 | BUG_ON(!PageLocked(page)); |
933 | ||
934 | if (PageAnon(page)) | |
7352349a | 935 | ret = try_to_unmap_anon(page, migration); |
1da177e4 | 936 | else |
7352349a | 937 | ret = try_to_unmap_file(page, migration); |
1da177e4 LT |
938 | |
939 | if (!page_mapped(page)) | |
940 | ret = SWAP_SUCCESS; | |
941 | return ret; | |
942 | } | |
81b4082d | 943 |