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 | 55 | |
1da177e4 LT |
56 | /* This must be called under the mmap_sem. */ |
57 | int anon_vma_prepare(struct vm_area_struct *vma) | |
58 | { | |
59 | struct anon_vma *anon_vma = vma->anon_vma; | |
60 | ||
61 | might_sleep(); | |
62 | if (unlikely(!anon_vma)) { | |
63 | struct mm_struct *mm = vma->vm_mm; | |
64 | struct anon_vma *allocated, *locked; | |
65 | ||
66 | anon_vma = find_mergeable_anon_vma(vma); | |
67 | if (anon_vma) { | |
68 | allocated = NULL; | |
69 | locked = anon_vma; | |
70 | spin_lock(&locked->lock); | |
71 | } else { | |
72 | anon_vma = anon_vma_alloc(); | |
73 | if (unlikely(!anon_vma)) | |
74 | return -ENOMEM; | |
75 | allocated = anon_vma; | |
76 | locked = NULL; | |
77 | } | |
78 | ||
79 | /* page_table_lock to protect against threads */ | |
80 | spin_lock(&mm->page_table_lock); | |
81 | if (likely(!vma->anon_vma)) { | |
82 | vma->anon_vma = anon_vma; | |
0697212a | 83 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
84 | allocated = NULL; |
85 | } | |
86 | spin_unlock(&mm->page_table_lock); | |
87 | ||
88 | if (locked) | |
89 | spin_unlock(&locked->lock); | |
90 | if (unlikely(allocated)) | |
91 | anon_vma_free(allocated); | |
92 | } | |
93 | return 0; | |
94 | } | |
95 | ||
96 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) | |
97 | { | |
98 | BUG_ON(vma->anon_vma != next->anon_vma); | |
99 | list_del(&next->anon_vma_node); | |
100 | } | |
101 | ||
102 | void __anon_vma_link(struct vm_area_struct *vma) | |
103 | { | |
104 | struct anon_vma *anon_vma = vma->anon_vma; | |
105 | ||
30acbaba | 106 | if (anon_vma) |
0697212a | 107 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
108 | } |
109 | ||
110 | void anon_vma_link(struct vm_area_struct *vma) | |
111 | { | |
112 | struct anon_vma *anon_vma = vma->anon_vma; | |
113 | ||
114 | if (anon_vma) { | |
115 | spin_lock(&anon_vma->lock); | |
0697212a | 116 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
117 | spin_unlock(&anon_vma->lock); |
118 | } | |
119 | } | |
120 | ||
121 | void anon_vma_unlink(struct vm_area_struct *vma) | |
122 | { | |
123 | struct anon_vma *anon_vma = vma->anon_vma; | |
124 | int empty; | |
125 | ||
126 | if (!anon_vma) | |
127 | return; | |
128 | ||
129 | spin_lock(&anon_vma->lock); | |
1da177e4 LT |
130 | list_del(&vma->anon_vma_node); |
131 | ||
132 | /* We must garbage collect the anon_vma if it's empty */ | |
133 | empty = list_empty(&anon_vma->head); | |
134 | spin_unlock(&anon_vma->lock); | |
135 | ||
136 | if (empty) | |
137 | anon_vma_free(anon_vma); | |
138 | } | |
139 | ||
4ba9b9d0 | 140 | static void anon_vma_ctor(struct kmem_cache *cachep, void *data) |
1da177e4 | 141 | { |
a35afb83 | 142 | struct anon_vma *anon_vma = data; |
1da177e4 | 143 | |
a35afb83 CL |
144 | spin_lock_init(&anon_vma->lock); |
145 | INIT_LIST_HEAD(&anon_vma->head); | |
1da177e4 LT |
146 | } |
147 | ||
148 | void __init anon_vma_init(void) | |
149 | { | |
150 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
20c2df83 | 151 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); |
1da177e4 LT |
152 | } |
153 | ||
154 | /* | |
155 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
156 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
157 | */ | |
158 | static struct anon_vma *page_lock_anon_vma(struct page *page) | |
159 | { | |
34bbd704 | 160 | struct anon_vma *anon_vma; |
1da177e4 LT |
161 | unsigned long anon_mapping; |
162 | ||
163 | rcu_read_lock(); | |
164 | anon_mapping = (unsigned long) page->mapping; | |
165 | if (!(anon_mapping & PAGE_MAPPING_ANON)) | |
166 | goto out; | |
167 | if (!page_mapped(page)) | |
168 | goto out; | |
169 | ||
170 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
171 | spin_lock(&anon_vma->lock); | |
34bbd704 | 172 | return anon_vma; |
1da177e4 LT |
173 | out: |
174 | rcu_read_unlock(); | |
34bbd704 ON |
175 | return NULL; |
176 | } | |
177 | ||
178 | static void page_unlock_anon_vma(struct anon_vma *anon_vma) | |
179 | { | |
180 | spin_unlock(&anon_vma->lock); | |
181 | rcu_read_unlock(); | |
1da177e4 LT |
182 | } |
183 | ||
184 | /* | |
3ad33b24 LS |
185 | * At what user virtual address is page expected in @vma? |
186 | * Returns virtual address or -EFAULT if page's index/offset is not | |
187 | * within the range mapped the @vma. | |
1da177e4 LT |
188 | */ |
189 | static inline unsigned long | |
190 | vma_address(struct page *page, struct vm_area_struct *vma) | |
191 | { | |
192 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
193 | unsigned long address; | |
194 | ||
195 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
196 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
3ad33b24 | 197 | /* page should be within @vma mapping range */ |
1da177e4 LT |
198 | return -EFAULT; |
199 | } | |
200 | return address; | |
201 | } | |
202 | ||
203 | /* | |
204 | * At what user virtual address is page expected in vma? checking that the | |
ee498ed7 | 205 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
1da177e4 LT |
206 | */ |
207 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
208 | { | |
209 | if (PageAnon(page)) { | |
210 | if ((void *)vma->anon_vma != | |
211 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
212 | return -EFAULT; | |
213 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
214 | if (!vma->vm_file || |
215 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
216 | return -EFAULT; |
217 | } else | |
218 | return -EFAULT; | |
219 | return vma_address(page, vma); | |
220 | } | |
221 | ||
81b4082d ND |
222 | /* |
223 | * Check that @page is mapped at @address into @mm. | |
224 | * | |
b8072f09 | 225 | * On success returns with pte mapped and locked. |
81b4082d | 226 | */ |
ceffc078 | 227 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
c0718806 | 228 | unsigned long address, spinlock_t **ptlp) |
81b4082d ND |
229 | { |
230 | pgd_t *pgd; | |
231 | pud_t *pud; | |
232 | pmd_t *pmd; | |
233 | pte_t *pte; | |
c0718806 | 234 | spinlock_t *ptl; |
81b4082d | 235 | |
81b4082d | 236 | pgd = pgd_offset(mm, address); |
c0718806 HD |
237 | if (!pgd_present(*pgd)) |
238 | return NULL; | |
239 | ||
240 | pud = pud_offset(pgd, address); | |
241 | if (!pud_present(*pud)) | |
242 | return NULL; | |
243 | ||
244 | pmd = pmd_offset(pud, address); | |
245 | if (!pmd_present(*pmd)) | |
246 | return NULL; | |
247 | ||
248 | pte = pte_offset_map(pmd, address); | |
249 | /* Make a quick check before getting the lock */ | |
250 | if (!pte_present(*pte)) { | |
251 | pte_unmap(pte); | |
252 | return NULL; | |
253 | } | |
254 | ||
4c21e2f2 | 255 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
256 | spin_lock(ptl); |
257 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
258 | *ptlp = ptl; | |
259 | return pte; | |
81b4082d | 260 | } |
c0718806 HD |
261 | pte_unmap_unlock(pte, ptl); |
262 | return NULL; | |
81b4082d ND |
263 | } |
264 | ||
1da177e4 LT |
265 | /* |
266 | * Subfunctions of page_referenced: page_referenced_one called | |
267 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
268 | */ | |
269 | static int page_referenced_one(struct page *page, | |
f7b7fd8f | 270 | struct vm_area_struct *vma, unsigned int *mapcount) |
1da177e4 LT |
271 | { |
272 | struct mm_struct *mm = vma->vm_mm; | |
273 | unsigned long address; | |
1da177e4 | 274 | pte_t *pte; |
c0718806 | 275 | spinlock_t *ptl; |
1da177e4 LT |
276 | int referenced = 0; |
277 | ||
1da177e4 LT |
278 | address = vma_address(page, vma); |
279 | if (address == -EFAULT) | |
280 | goto out; | |
281 | ||
c0718806 HD |
282 | pte = page_check_address(page, mm, address, &ptl); |
283 | if (!pte) | |
284 | goto out; | |
1da177e4 | 285 | |
5a9bbdcd HD |
286 | if (vma->vm_flags & VM_LOCKED) { |
287 | referenced++; | |
288 | *mapcount = 1; /* break early from loop */ | |
289 | } else if (ptep_clear_flush_young(vma, address, pte)) | |
c0718806 | 290 | referenced++; |
1da177e4 | 291 | |
c0718806 HD |
292 | /* Pretend the page is referenced if the task has the |
293 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 294 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
295 | rwsem_is_locked(&mm->mmap_sem)) |
296 | referenced++; | |
297 | ||
298 | (*mapcount)--; | |
299 | pte_unmap_unlock(pte, ptl); | |
1da177e4 LT |
300 | out: |
301 | return referenced; | |
302 | } | |
303 | ||
f7b7fd8f | 304 | static int page_referenced_anon(struct page *page) |
1da177e4 LT |
305 | { |
306 | unsigned int mapcount; | |
307 | struct anon_vma *anon_vma; | |
308 | struct vm_area_struct *vma; | |
309 | int referenced = 0; | |
310 | ||
311 | anon_vma = page_lock_anon_vma(page); | |
312 | if (!anon_vma) | |
313 | return referenced; | |
314 | ||
315 | mapcount = page_mapcount(page); | |
316 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
f7b7fd8f | 317 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
318 | if (!mapcount) |
319 | break; | |
320 | } | |
34bbd704 ON |
321 | |
322 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
323 | return referenced; |
324 | } | |
325 | ||
326 | /** | |
327 | * page_referenced_file - referenced check for object-based rmap | |
328 | * @page: the page we're checking references on. | |
329 | * | |
330 | * For an object-based mapped page, find all the places it is mapped and | |
331 | * check/clear the referenced flag. This is done by following the page->mapping | |
332 | * pointer, then walking the chain of vmas it holds. It returns the number | |
333 | * of references it found. | |
334 | * | |
335 | * This function is only called from page_referenced for object-based pages. | |
336 | */ | |
f7b7fd8f | 337 | static int page_referenced_file(struct page *page) |
1da177e4 LT |
338 | { |
339 | unsigned int mapcount; | |
340 | struct address_space *mapping = page->mapping; | |
341 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
342 | struct vm_area_struct *vma; | |
343 | struct prio_tree_iter iter; | |
344 | int referenced = 0; | |
345 | ||
346 | /* | |
347 | * The caller's checks on page->mapping and !PageAnon have made | |
348 | * sure that this is a file page: the check for page->mapping | |
349 | * excludes the case just before it gets set on an anon page. | |
350 | */ | |
351 | BUG_ON(PageAnon(page)); | |
352 | ||
353 | /* | |
354 | * The page lock not only makes sure that page->mapping cannot | |
355 | * suddenly be NULLified by truncation, it makes sure that the | |
356 | * structure at mapping cannot be freed and reused yet, | |
357 | * so we can safely take mapping->i_mmap_lock. | |
358 | */ | |
359 | BUG_ON(!PageLocked(page)); | |
360 | ||
361 | spin_lock(&mapping->i_mmap_lock); | |
362 | ||
363 | /* | |
364 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
365 | * is more likely to be accurate if we note it after spinning. | |
366 | */ | |
367 | mapcount = page_mapcount(page); | |
368 | ||
369 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
370 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
371 | == (VM_LOCKED|VM_MAYSHARE)) { | |
372 | referenced++; | |
373 | break; | |
374 | } | |
f7b7fd8f | 375 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
376 | if (!mapcount) |
377 | break; | |
378 | } | |
379 | ||
380 | spin_unlock(&mapping->i_mmap_lock); | |
381 | return referenced; | |
382 | } | |
383 | ||
384 | /** | |
385 | * page_referenced - test if the page was referenced | |
386 | * @page: the page to test | |
387 | * @is_locked: caller holds lock on the page | |
388 | * | |
389 | * Quick test_and_clear_referenced for all mappings to a page, | |
390 | * returns the number of ptes which referenced the page. | |
391 | */ | |
f7b7fd8f | 392 | int page_referenced(struct page *page, int is_locked) |
1da177e4 LT |
393 | { |
394 | int referenced = 0; | |
395 | ||
1da177e4 LT |
396 | if (page_test_and_clear_young(page)) |
397 | referenced++; | |
398 | ||
399 | if (TestClearPageReferenced(page)) | |
400 | referenced++; | |
401 | ||
402 | if (page_mapped(page) && page->mapping) { | |
403 | if (PageAnon(page)) | |
f7b7fd8f | 404 | referenced += page_referenced_anon(page); |
1da177e4 | 405 | else if (is_locked) |
f7b7fd8f | 406 | referenced += page_referenced_file(page); |
1da177e4 LT |
407 | else if (TestSetPageLocked(page)) |
408 | referenced++; | |
409 | else { | |
410 | if (page->mapping) | |
f7b7fd8f | 411 | referenced += page_referenced_file(page); |
1da177e4 LT |
412 | unlock_page(page); |
413 | } | |
414 | } | |
415 | return referenced; | |
416 | } | |
417 | ||
d08b3851 PZ |
418 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma) |
419 | { | |
420 | struct mm_struct *mm = vma->vm_mm; | |
421 | unsigned long address; | |
c2fda5fe | 422 | pte_t *pte; |
d08b3851 PZ |
423 | spinlock_t *ptl; |
424 | int ret = 0; | |
425 | ||
426 | address = vma_address(page, vma); | |
427 | if (address == -EFAULT) | |
428 | goto out; | |
429 | ||
430 | pte = page_check_address(page, mm, address, &ptl); | |
431 | if (!pte) | |
432 | goto out; | |
433 | ||
c2fda5fe PZ |
434 | if (pte_dirty(*pte) || pte_write(*pte)) { |
435 | pte_t entry; | |
d08b3851 | 436 | |
c2fda5fe PZ |
437 | flush_cache_page(vma, address, pte_pfn(*pte)); |
438 | entry = ptep_clear_flush(vma, address, pte); | |
439 | entry = pte_wrprotect(entry); | |
440 | entry = pte_mkclean(entry); | |
d6e88e67 | 441 | set_pte_at(mm, address, pte, entry); |
c2fda5fe PZ |
442 | ret = 1; |
443 | } | |
d08b3851 | 444 | |
d08b3851 PZ |
445 | pte_unmap_unlock(pte, ptl); |
446 | out: | |
447 | return ret; | |
448 | } | |
449 | ||
450 | static int page_mkclean_file(struct address_space *mapping, struct page *page) | |
451 | { | |
452 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
453 | struct vm_area_struct *vma; | |
454 | struct prio_tree_iter iter; | |
455 | int ret = 0; | |
456 | ||
457 | BUG_ON(PageAnon(page)); | |
458 | ||
459 | spin_lock(&mapping->i_mmap_lock); | |
460 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
461 | if (vma->vm_flags & VM_SHARED) | |
462 | ret += page_mkclean_one(page, vma); | |
463 | } | |
464 | spin_unlock(&mapping->i_mmap_lock); | |
465 | return ret; | |
466 | } | |
467 | ||
468 | int page_mkclean(struct page *page) | |
469 | { | |
470 | int ret = 0; | |
471 | ||
472 | BUG_ON(!PageLocked(page)); | |
473 | ||
474 | if (page_mapped(page)) { | |
475 | struct address_space *mapping = page_mapping(page); | |
ce7e9fae | 476 | if (mapping) { |
d08b3851 | 477 | ret = page_mkclean_file(mapping, page); |
ce7e9fae CB |
478 | if (page_test_dirty(page)) { |
479 | page_clear_dirty(page); | |
480 | ret = 1; | |
481 | } | |
6c210482 | 482 | } |
d08b3851 PZ |
483 | } |
484 | ||
485 | return ret; | |
486 | } | |
60b59bea | 487 | EXPORT_SYMBOL_GPL(page_mkclean); |
d08b3851 | 488 | |
9617d95e NP |
489 | /** |
490 | * page_set_anon_rmap - setup new anonymous rmap | |
491 | * @page: the page to add the mapping to | |
492 | * @vma: the vm area in which the mapping is added | |
493 | * @address: the user virtual address mapped | |
494 | */ | |
495 | static void __page_set_anon_rmap(struct page *page, | |
496 | struct vm_area_struct *vma, unsigned long address) | |
497 | { | |
498 | struct anon_vma *anon_vma = vma->anon_vma; | |
499 | ||
500 | BUG_ON(!anon_vma); | |
501 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
502 | page->mapping = (struct address_space *) anon_vma; | |
503 | ||
504 | page->index = linear_page_index(vma, address); | |
505 | ||
a74609fa NP |
506 | /* |
507 | * nr_mapped state can be updated without turning off | |
508 | * interrupts because it is not modified via interrupt. | |
509 | */ | |
f3dbd344 | 510 | __inc_zone_page_state(page, NR_ANON_PAGES); |
9617d95e NP |
511 | } |
512 | ||
c97a9e10 NP |
513 | /** |
514 | * page_set_anon_rmap - sanity check anonymous rmap addition | |
515 | * @page: the page to add the mapping to | |
516 | * @vma: the vm area in which the mapping is added | |
517 | * @address: the user virtual address mapped | |
518 | */ | |
519 | static void __page_check_anon_rmap(struct page *page, | |
520 | struct vm_area_struct *vma, unsigned long address) | |
521 | { | |
522 | #ifdef CONFIG_DEBUG_VM | |
523 | /* | |
524 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
525 | * be set up correctly at this point. | |
526 | * | |
527 | * We have exclusion against page_add_anon_rmap because the caller | |
528 | * always holds the page locked, except if called from page_dup_rmap, | |
529 | * in which case the page is already known to be setup. | |
530 | * | |
531 | * We have exclusion against page_add_new_anon_rmap because those pages | |
532 | * are initially only visible via the pagetables, and the pte is locked | |
533 | * over the call to page_add_new_anon_rmap. | |
534 | */ | |
535 | struct anon_vma *anon_vma = vma->anon_vma; | |
536 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
537 | BUG_ON(page->mapping != (struct address_space *)anon_vma); | |
538 | BUG_ON(page->index != linear_page_index(vma, address)); | |
539 | #endif | |
540 | } | |
541 | ||
1da177e4 LT |
542 | /** |
543 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
544 | * @page: the page to add the mapping to | |
545 | * @vma: the vm area in which the mapping is added | |
546 | * @address: the user virtual address mapped | |
547 | * | |
c97a9e10 | 548 | * The caller needs to hold the pte lock and the page must be locked. |
1da177e4 LT |
549 | */ |
550 | void page_add_anon_rmap(struct page *page, | |
551 | struct vm_area_struct *vma, unsigned long address) | |
552 | { | |
c97a9e10 NP |
553 | VM_BUG_ON(!PageLocked(page)); |
554 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
9617d95e NP |
555 | if (atomic_inc_and_test(&page->_mapcount)) |
556 | __page_set_anon_rmap(page, vma, address); | |
c97a9e10 NP |
557 | else |
558 | __page_check_anon_rmap(page, vma, address); | |
1da177e4 LT |
559 | } |
560 | ||
9617d95e NP |
561 | /* |
562 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
563 | * @page: the page to add the mapping to | |
564 | * @vma: the vm area in which the mapping is added | |
565 | * @address: the user virtual address mapped | |
566 | * | |
567 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
568 | * This means the inc-and-test can be bypassed. | |
c97a9e10 | 569 | * Page does not have to be locked. |
9617d95e NP |
570 | */ |
571 | void page_add_new_anon_rmap(struct page *page, | |
572 | struct vm_area_struct *vma, unsigned long address) | |
573 | { | |
c97a9e10 | 574 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
9617d95e NP |
575 | atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ |
576 | __page_set_anon_rmap(page, vma, address); | |
577 | } | |
578 | ||
1da177e4 LT |
579 | /** |
580 | * page_add_file_rmap - add pte mapping to a file page | |
581 | * @page: the page to add the mapping to | |
582 | * | |
b8072f09 | 583 | * The caller needs to hold the pte lock. |
1da177e4 LT |
584 | */ |
585 | void page_add_file_rmap(struct page *page) | |
586 | { | |
1da177e4 | 587 | if (atomic_inc_and_test(&page->_mapcount)) |
65ba55f5 | 588 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
1da177e4 LT |
589 | } |
590 | ||
c97a9e10 NP |
591 | #ifdef CONFIG_DEBUG_VM |
592 | /** | |
593 | * page_dup_rmap - duplicate pte mapping to a page | |
594 | * @page: the page to add the mapping to | |
595 | * | |
596 | * For copy_page_range only: minimal extract from page_add_file_rmap / | |
597 | * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's | |
598 | * quicker. | |
599 | * | |
600 | * The caller needs to hold the pte lock. | |
601 | */ | |
602 | void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address) | |
603 | { | |
604 | BUG_ON(page_mapcount(page) == 0); | |
605 | if (PageAnon(page)) | |
606 | __page_check_anon_rmap(page, vma, address); | |
607 | atomic_inc(&page->_mapcount); | |
608 | } | |
609 | #endif | |
610 | ||
1da177e4 LT |
611 | /** |
612 | * page_remove_rmap - take down pte mapping from a page | |
613 | * @page: page to remove mapping from | |
614 | * | |
b8072f09 | 615 | * The caller needs to hold the pte lock. |
1da177e4 | 616 | */ |
7de6b805 | 617 | void page_remove_rmap(struct page *page, struct vm_area_struct *vma) |
1da177e4 | 618 | { |
1da177e4 | 619 | if (atomic_add_negative(-1, &page->_mapcount)) { |
b7ab795b | 620 | if (unlikely(page_mapcount(page) < 0)) { |
ef2bf0dc | 621 | printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page)); |
7de6b805 | 622 | printk (KERN_EMERG " page pfn = %lx\n", page_to_pfn(page)); |
ef2bf0dc DJ |
623 | printk (KERN_EMERG " page->flags = %lx\n", page->flags); |
624 | printk (KERN_EMERG " page->count = %x\n", page_count(page)); | |
625 | printk (KERN_EMERG " page->mapping = %p\n", page->mapping); | |
7de6b805 | 626 | print_symbol (KERN_EMERG " vma->vm_ops = %s\n", (unsigned long)vma->vm_ops); |
54cb8821 | 627 | if (vma->vm_ops) { |
7de6b805 | 628 | print_symbol (KERN_EMERG " vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage); |
54cb8821 NP |
629 | print_symbol (KERN_EMERG " vma->vm_ops->fault = %s\n", (unsigned long)vma->vm_ops->fault); |
630 | } | |
7de6b805 NP |
631 | if (vma->vm_file && vma->vm_file->f_op) |
632 | print_symbol (KERN_EMERG " vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap); | |
b16bc64d | 633 | BUG(); |
ef2bf0dc | 634 | } |
b16bc64d | 635 | |
1da177e4 LT |
636 | /* |
637 | * It would be tidy to reset the PageAnon mapping here, | |
638 | * but that might overwrite a racing page_add_anon_rmap | |
639 | * which increments mapcount after us but sets mapping | |
640 | * before us: so leave the reset to free_hot_cold_page, | |
641 | * and remember that it's only reliable while mapped. | |
642 | * Leaving it set also helps swapoff to reinstate ptes | |
643 | * faster for those pages still in swapcache. | |
644 | */ | |
6c210482 MS |
645 | if (page_test_dirty(page)) { |
646 | page_clear_dirty(page); | |
1da177e4 | 647 | set_page_dirty(page); |
6c210482 | 648 | } |
f3dbd344 CL |
649 | __dec_zone_page_state(page, |
650 | PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED); | |
1da177e4 LT |
651 | } |
652 | } | |
653 | ||
654 | /* | |
655 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
656 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
657 | */ | |
a48d07af | 658 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
7352349a | 659 | int migration) |
1da177e4 LT |
660 | { |
661 | struct mm_struct *mm = vma->vm_mm; | |
662 | unsigned long address; | |
1da177e4 LT |
663 | pte_t *pte; |
664 | pte_t pteval; | |
c0718806 | 665 | spinlock_t *ptl; |
1da177e4 LT |
666 | int ret = SWAP_AGAIN; |
667 | ||
1da177e4 LT |
668 | address = vma_address(page, vma); |
669 | if (address == -EFAULT) | |
670 | goto out; | |
671 | ||
c0718806 HD |
672 | pte = page_check_address(page, mm, address, &ptl); |
673 | if (!pte) | |
81b4082d | 674 | goto out; |
1da177e4 LT |
675 | |
676 | /* | |
677 | * If the page is mlock()d, we cannot swap it out. | |
678 | * If it's recently referenced (perhaps page_referenced | |
679 | * skipped over this mm) then we should reactivate it. | |
680 | */ | |
e6a1530d CL |
681 | if (!migration && ((vma->vm_flags & VM_LOCKED) || |
682 | (ptep_clear_flush_young(vma, address, pte)))) { | |
1da177e4 LT |
683 | ret = SWAP_FAIL; |
684 | goto out_unmap; | |
685 | } | |
686 | ||
1da177e4 LT |
687 | /* Nuke the page table entry. */ |
688 | flush_cache_page(vma, address, page_to_pfn(page)); | |
689 | pteval = ptep_clear_flush(vma, address, pte); | |
690 | ||
691 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
692 | if (pte_dirty(pteval)) | |
693 | set_page_dirty(page); | |
694 | ||
365e9c87 HD |
695 | /* Update high watermark before we lower rss */ |
696 | update_hiwater_rss(mm); | |
697 | ||
1da177e4 | 698 | if (PageAnon(page)) { |
4c21e2f2 | 699 | swp_entry_t entry = { .val = page_private(page) }; |
0697212a CL |
700 | |
701 | if (PageSwapCache(page)) { | |
702 | /* | |
703 | * Store the swap location in the pte. | |
704 | * See handle_pte_fault() ... | |
705 | */ | |
706 | swap_duplicate(entry); | |
707 | if (list_empty(&mm->mmlist)) { | |
708 | spin_lock(&mmlist_lock); | |
709 | if (list_empty(&mm->mmlist)) | |
710 | list_add(&mm->mmlist, &init_mm.mmlist); | |
711 | spin_unlock(&mmlist_lock); | |
712 | } | |
442c9137 | 713 | dec_mm_counter(mm, anon_rss); |
04e62a29 | 714 | #ifdef CONFIG_MIGRATION |
0697212a CL |
715 | } else { |
716 | /* | |
717 | * Store the pfn of the page in a special migration | |
718 | * pte. do_swap_page() will wait until the migration | |
719 | * pte is removed and then restart fault handling. | |
720 | */ | |
721 | BUG_ON(!migration); | |
722 | entry = make_migration_entry(page, pte_write(pteval)); | |
04e62a29 | 723 | #endif |
1da177e4 LT |
724 | } |
725 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
726 | BUG_ON(pte_file(*pte)); | |
4294621f | 727 | } else |
04e62a29 CL |
728 | #ifdef CONFIG_MIGRATION |
729 | if (migration) { | |
730 | /* Establish migration entry for a file page */ | |
731 | swp_entry_t entry; | |
732 | entry = make_migration_entry(page, pte_write(pteval)); | |
733 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
734 | } else | |
735 | #endif | |
4294621f | 736 | dec_mm_counter(mm, file_rss); |
1da177e4 | 737 | |
04e62a29 | 738 | |
7de6b805 | 739 | page_remove_rmap(page, vma); |
1da177e4 LT |
740 | page_cache_release(page); |
741 | ||
742 | out_unmap: | |
c0718806 | 743 | pte_unmap_unlock(pte, ptl); |
1da177e4 LT |
744 | out: |
745 | return ret; | |
746 | } | |
747 | ||
748 | /* | |
749 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
750 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
751 | * Consequently, given a particular page and its ->index, we cannot locate the | |
752 | * ptes which are mapping that page without an exhaustive linear search. | |
753 | * | |
754 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
755 | * maps the file to which the target page belongs. The ->vm_private_data field | |
756 | * holds the current cursor into that scan. Successive searches will circulate | |
757 | * around the vma's virtual address space. | |
758 | * | |
759 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
760 | * more scanning pressure is placed against them as well. Eventually pages | |
761 | * will become fully unmapped and are eligible for eviction. | |
762 | * | |
763 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
764 | * there there won't be many ptes located within the scan cluster. In this case | |
765 | * maybe we could scan further - to the end of the pte page, perhaps. | |
766 | */ | |
767 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
768 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
769 | ||
770 | static void try_to_unmap_cluster(unsigned long cursor, | |
771 | unsigned int *mapcount, struct vm_area_struct *vma) | |
772 | { | |
773 | struct mm_struct *mm = vma->vm_mm; | |
774 | pgd_t *pgd; | |
775 | pud_t *pud; | |
776 | pmd_t *pmd; | |
c0718806 | 777 | pte_t *pte; |
1da177e4 | 778 | pte_t pteval; |
c0718806 | 779 | spinlock_t *ptl; |
1da177e4 LT |
780 | struct page *page; |
781 | unsigned long address; | |
782 | unsigned long end; | |
1da177e4 | 783 | |
1da177e4 LT |
784 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
785 | end = address + CLUSTER_SIZE; | |
786 | if (address < vma->vm_start) | |
787 | address = vma->vm_start; | |
788 | if (end > vma->vm_end) | |
789 | end = vma->vm_end; | |
790 | ||
791 | pgd = pgd_offset(mm, address); | |
792 | if (!pgd_present(*pgd)) | |
c0718806 | 793 | return; |
1da177e4 LT |
794 | |
795 | pud = pud_offset(pgd, address); | |
796 | if (!pud_present(*pud)) | |
c0718806 | 797 | return; |
1da177e4 LT |
798 | |
799 | pmd = pmd_offset(pud, address); | |
800 | if (!pmd_present(*pmd)) | |
c0718806 HD |
801 | return; |
802 | ||
803 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 804 | |
365e9c87 HD |
805 | /* Update high watermark before we lower rss */ |
806 | update_hiwater_rss(mm); | |
807 | ||
c0718806 | 808 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
809 | if (!pte_present(*pte)) |
810 | continue; | |
6aab341e LT |
811 | page = vm_normal_page(vma, address, *pte); |
812 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 LT |
813 | |
814 | if (ptep_clear_flush_young(vma, address, pte)) | |
815 | continue; | |
816 | ||
817 | /* Nuke the page table entry. */ | |
eca35133 | 818 | flush_cache_page(vma, address, pte_pfn(*pte)); |
1da177e4 LT |
819 | pteval = ptep_clear_flush(vma, address, pte); |
820 | ||
821 | /* If nonlinear, store the file page offset in the pte. */ | |
822 | if (page->index != linear_page_index(vma, address)) | |
823 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
824 | ||
825 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
826 | if (pte_dirty(pteval)) | |
827 | set_page_dirty(page); | |
828 | ||
7de6b805 | 829 | page_remove_rmap(page, vma); |
1da177e4 | 830 | page_cache_release(page); |
4294621f | 831 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
832 | (*mapcount)--; |
833 | } | |
c0718806 | 834 | pte_unmap_unlock(pte - 1, ptl); |
1da177e4 LT |
835 | } |
836 | ||
7352349a | 837 | static int try_to_unmap_anon(struct page *page, int migration) |
1da177e4 LT |
838 | { |
839 | struct anon_vma *anon_vma; | |
840 | struct vm_area_struct *vma; | |
841 | int ret = SWAP_AGAIN; | |
842 | ||
843 | anon_vma = page_lock_anon_vma(page); | |
844 | if (!anon_vma) | |
845 | return ret; | |
846 | ||
847 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
7352349a | 848 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
849 | if (ret == SWAP_FAIL || !page_mapped(page)) |
850 | break; | |
851 | } | |
34bbd704 ON |
852 | |
853 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
854 | return ret; |
855 | } | |
856 | ||
857 | /** | |
858 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
859 | * @page: the page to unmap | |
860 | * | |
861 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
862 | * contained in the address_space struct it points to. | |
863 | * | |
864 | * This function is only called from try_to_unmap for object-based pages. | |
865 | */ | |
7352349a | 866 | static int try_to_unmap_file(struct page *page, int migration) |
1da177e4 LT |
867 | { |
868 | struct address_space *mapping = page->mapping; | |
869 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
870 | struct vm_area_struct *vma; | |
871 | struct prio_tree_iter iter; | |
872 | int ret = SWAP_AGAIN; | |
873 | unsigned long cursor; | |
874 | unsigned long max_nl_cursor = 0; | |
875 | unsigned long max_nl_size = 0; | |
876 | unsigned int mapcount; | |
877 | ||
878 | spin_lock(&mapping->i_mmap_lock); | |
879 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
7352349a | 880 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
881 | if (ret == SWAP_FAIL || !page_mapped(page)) |
882 | goto out; | |
883 | } | |
884 | ||
885 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
886 | goto out; | |
887 | ||
888 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
889 | shared.vm_set.list) { | |
e6a1530d | 890 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
891 | continue; |
892 | cursor = (unsigned long) vma->vm_private_data; | |
893 | if (cursor > max_nl_cursor) | |
894 | max_nl_cursor = cursor; | |
895 | cursor = vma->vm_end - vma->vm_start; | |
896 | if (cursor > max_nl_size) | |
897 | max_nl_size = cursor; | |
898 | } | |
899 | ||
900 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
901 | ret = SWAP_FAIL; | |
902 | goto out; | |
903 | } | |
904 | ||
905 | /* | |
906 | * We don't try to search for this page in the nonlinear vmas, | |
907 | * and page_referenced wouldn't have found it anyway. Instead | |
908 | * just walk the nonlinear vmas trying to age and unmap some. | |
909 | * The mapcount of the page we came in with is irrelevant, | |
910 | * but even so use it as a guide to how hard we should try? | |
911 | */ | |
912 | mapcount = page_mapcount(page); | |
913 | if (!mapcount) | |
914 | goto out; | |
915 | cond_resched_lock(&mapping->i_mmap_lock); | |
916 | ||
917 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
918 | if (max_nl_cursor == 0) | |
919 | max_nl_cursor = CLUSTER_SIZE; | |
920 | ||
921 | do { | |
922 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
923 | shared.vm_set.list) { | |
e6a1530d | 924 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
925 | continue; |
926 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 927 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
928 | cursor < vma->vm_end - vma->vm_start) { |
929 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
930 | cursor += CLUSTER_SIZE; | |
931 | vma->vm_private_data = (void *) cursor; | |
932 | if ((int)mapcount <= 0) | |
933 | goto out; | |
934 | } | |
935 | vma->vm_private_data = (void *) max_nl_cursor; | |
936 | } | |
937 | cond_resched_lock(&mapping->i_mmap_lock); | |
938 | max_nl_cursor += CLUSTER_SIZE; | |
939 | } while (max_nl_cursor <= max_nl_size); | |
940 | ||
941 | /* | |
942 | * Don't loop forever (perhaps all the remaining pages are | |
943 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
944 | * vmas, now forgetting on which ones it had fallen behind. | |
945 | */ | |
101d2be7 HD |
946 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
947 | vma->vm_private_data = NULL; | |
1da177e4 LT |
948 | out: |
949 | spin_unlock(&mapping->i_mmap_lock); | |
950 | return ret; | |
951 | } | |
952 | ||
953 | /** | |
954 | * try_to_unmap - try to remove all page table mappings to a page | |
955 | * @page: the page to get unmapped | |
956 | * | |
957 | * Tries to remove all the page table entries which are mapping this | |
958 | * page, used in the pageout path. Caller must hold the page lock. | |
959 | * Return values are: | |
960 | * | |
961 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
962 | * SWAP_AGAIN - we missed a mapping, try again later | |
963 | * SWAP_FAIL - the page is unswappable | |
964 | */ | |
7352349a | 965 | int try_to_unmap(struct page *page, int migration) |
1da177e4 LT |
966 | { |
967 | int ret; | |
968 | ||
1da177e4 LT |
969 | BUG_ON(!PageLocked(page)); |
970 | ||
971 | if (PageAnon(page)) | |
7352349a | 972 | ret = try_to_unmap_anon(page, migration); |
1da177e4 | 973 | else |
7352349a | 974 | ret = try_to_unmap_file(page, migration); |
1da177e4 LT |
975 | |
976 | if (!page_mapped(page)) | |
977 | ret = SWAP_SUCCESS; | |
978 | return ret; | |
979 | } | |
81b4082d | 980 |