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) | |
7a405079 | 39 | * zone->lock (within radix tree node alloc) |
1da177e4 LT |
40 | */ |
41 | ||
42 | #include <linux/mm.h> | |
43 | #include <linux/pagemap.h> | |
44 | #include <linux/swap.h> | |
45 | #include <linux/swapops.h> | |
46 | #include <linux/slab.h> | |
47 | #include <linux/init.h> | |
48 | #include <linux/rmap.h> | |
49 | #include <linux/rcupdate.h> | |
a48d07af | 50 | #include <linux/module.h> |
7de6b805 | 51 | #include <linux/kallsyms.h> |
1da177e4 LT |
52 | |
53 | #include <asm/tlbflush.h> | |
54 | ||
fcc234f8 | 55 | struct kmem_cache *anon_vma_cachep; |
1da177e4 | 56 | |
1da177e4 LT |
57 | /* This must be called under the mmap_sem. */ |
58 | int anon_vma_prepare(struct vm_area_struct *vma) | |
59 | { | |
60 | struct anon_vma *anon_vma = vma->anon_vma; | |
61 | ||
62 | might_sleep(); | |
63 | if (unlikely(!anon_vma)) { | |
64 | struct mm_struct *mm = vma->vm_mm; | |
65 | struct anon_vma *allocated, *locked; | |
66 | ||
67 | anon_vma = find_mergeable_anon_vma(vma); | |
68 | if (anon_vma) { | |
69 | allocated = NULL; | |
70 | locked = anon_vma; | |
71 | spin_lock(&locked->lock); | |
72 | } else { | |
73 | anon_vma = anon_vma_alloc(); | |
74 | if (unlikely(!anon_vma)) | |
75 | return -ENOMEM; | |
76 | allocated = anon_vma; | |
77 | locked = NULL; | |
78 | } | |
79 | ||
80 | /* page_table_lock to protect against threads */ | |
81 | spin_lock(&mm->page_table_lock); | |
82 | if (likely(!vma->anon_vma)) { | |
83 | vma->anon_vma = anon_vma; | |
0697212a | 84 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
85 | allocated = NULL; |
86 | } | |
87 | spin_unlock(&mm->page_table_lock); | |
88 | ||
89 | if (locked) | |
90 | spin_unlock(&locked->lock); | |
91 | if (unlikely(allocated)) | |
92 | anon_vma_free(allocated); | |
93 | } | |
94 | return 0; | |
95 | } | |
96 | ||
97 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) | |
98 | { | |
99 | BUG_ON(vma->anon_vma != next->anon_vma); | |
100 | list_del(&next->anon_vma_node); | |
101 | } | |
102 | ||
103 | void __anon_vma_link(struct vm_area_struct *vma) | |
104 | { | |
105 | struct anon_vma *anon_vma = vma->anon_vma; | |
106 | ||
30acbaba | 107 | if (anon_vma) |
0697212a | 108 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
109 | } |
110 | ||
111 | void anon_vma_link(struct vm_area_struct *vma) | |
112 | { | |
113 | struct anon_vma *anon_vma = vma->anon_vma; | |
114 | ||
115 | if (anon_vma) { | |
116 | spin_lock(&anon_vma->lock); | |
0697212a | 117 | list_add_tail(&vma->anon_vma_node, &anon_vma->head); |
1da177e4 LT |
118 | spin_unlock(&anon_vma->lock); |
119 | } | |
120 | } | |
121 | ||
122 | void anon_vma_unlink(struct vm_area_struct *vma) | |
123 | { | |
124 | struct anon_vma *anon_vma = vma->anon_vma; | |
125 | int empty; | |
126 | ||
127 | if (!anon_vma) | |
128 | return; | |
129 | ||
130 | spin_lock(&anon_vma->lock); | |
1da177e4 LT |
131 | list_del(&vma->anon_vma_node); |
132 | ||
133 | /* We must garbage collect the anon_vma if it's empty */ | |
134 | empty = list_empty(&anon_vma->head); | |
135 | spin_unlock(&anon_vma->lock); | |
136 | ||
137 | if (empty) | |
138 | anon_vma_free(anon_vma); | |
139 | } | |
140 | ||
4ba9b9d0 | 141 | static void anon_vma_ctor(struct kmem_cache *cachep, void *data) |
1da177e4 | 142 | { |
a35afb83 | 143 | struct anon_vma *anon_vma = data; |
1da177e4 | 144 | |
a35afb83 CL |
145 | spin_lock_init(&anon_vma->lock); |
146 | INIT_LIST_HEAD(&anon_vma->head); | |
1da177e4 LT |
147 | } |
148 | ||
149 | void __init anon_vma_init(void) | |
150 | { | |
151 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
20c2df83 | 152 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); |
1da177e4 LT |
153 | } |
154 | ||
155 | /* | |
156 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
157 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
158 | */ | |
159 | static struct anon_vma *page_lock_anon_vma(struct page *page) | |
160 | { | |
34bbd704 | 161 | struct anon_vma *anon_vma; |
1da177e4 LT |
162 | unsigned long anon_mapping; |
163 | ||
164 | rcu_read_lock(); | |
165 | anon_mapping = (unsigned long) page->mapping; | |
166 | if (!(anon_mapping & PAGE_MAPPING_ANON)) | |
167 | goto out; | |
168 | if (!page_mapped(page)) | |
169 | goto out; | |
170 | ||
171 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
172 | spin_lock(&anon_vma->lock); | |
34bbd704 | 173 | return anon_vma; |
1da177e4 LT |
174 | out: |
175 | rcu_read_unlock(); | |
34bbd704 ON |
176 | return NULL; |
177 | } | |
178 | ||
179 | static void page_unlock_anon_vma(struct anon_vma *anon_vma) | |
180 | { | |
181 | spin_unlock(&anon_vma->lock); | |
182 | rcu_read_unlock(); | |
1da177e4 LT |
183 | } |
184 | ||
185 | /* | |
3ad33b24 LS |
186 | * At what user virtual address is page expected in @vma? |
187 | * Returns virtual address or -EFAULT if page's index/offset is not | |
188 | * within the range mapped the @vma. | |
1da177e4 LT |
189 | */ |
190 | static inline unsigned long | |
191 | vma_address(struct page *page, struct vm_area_struct *vma) | |
192 | { | |
193 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
194 | unsigned long address; | |
195 | ||
196 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
197 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
3ad33b24 | 198 | /* page should be within @vma mapping range */ |
1da177e4 LT |
199 | return -EFAULT; |
200 | } | |
201 | return address; | |
202 | } | |
203 | ||
204 | /* | |
205 | * At what user virtual address is page expected in vma? checking that the | |
ee498ed7 | 206 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
1da177e4 LT |
207 | */ |
208 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
209 | { | |
210 | if (PageAnon(page)) { | |
211 | if ((void *)vma->anon_vma != | |
212 | (void *)page->mapping - PAGE_MAPPING_ANON) | |
213 | return -EFAULT; | |
214 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
215 | if (!vma->vm_file || |
216 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
217 | return -EFAULT; |
218 | } else | |
219 | return -EFAULT; | |
220 | return vma_address(page, vma); | |
221 | } | |
222 | ||
81b4082d ND |
223 | /* |
224 | * Check that @page is mapped at @address into @mm. | |
225 | * | |
b8072f09 | 226 | * On success returns with pte mapped and locked. |
81b4082d | 227 | */ |
ceffc078 | 228 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
c0718806 | 229 | unsigned long address, spinlock_t **ptlp) |
81b4082d ND |
230 | { |
231 | pgd_t *pgd; | |
232 | pud_t *pud; | |
233 | pmd_t *pmd; | |
234 | pte_t *pte; | |
c0718806 | 235 | spinlock_t *ptl; |
81b4082d | 236 | |
81b4082d | 237 | pgd = pgd_offset(mm, address); |
c0718806 HD |
238 | if (!pgd_present(*pgd)) |
239 | return NULL; | |
240 | ||
241 | pud = pud_offset(pgd, address); | |
242 | if (!pud_present(*pud)) | |
243 | return NULL; | |
244 | ||
245 | pmd = pmd_offset(pud, address); | |
246 | if (!pmd_present(*pmd)) | |
247 | return NULL; | |
248 | ||
249 | pte = pte_offset_map(pmd, address); | |
250 | /* Make a quick check before getting the lock */ | |
251 | if (!pte_present(*pte)) { | |
252 | pte_unmap(pte); | |
253 | return NULL; | |
254 | } | |
255 | ||
4c21e2f2 | 256 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
257 | spin_lock(ptl); |
258 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
259 | *ptlp = ptl; | |
260 | return pte; | |
81b4082d | 261 | } |
c0718806 HD |
262 | pte_unmap_unlock(pte, ptl); |
263 | return NULL; | |
81b4082d ND |
264 | } |
265 | ||
1da177e4 LT |
266 | /* |
267 | * Subfunctions of page_referenced: page_referenced_one called | |
268 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
269 | */ | |
270 | static int page_referenced_one(struct page *page, | |
f7b7fd8f | 271 | struct vm_area_struct *vma, unsigned int *mapcount) |
1da177e4 LT |
272 | { |
273 | struct mm_struct *mm = vma->vm_mm; | |
274 | unsigned long address; | |
1da177e4 | 275 | pte_t *pte; |
c0718806 | 276 | spinlock_t *ptl; |
1da177e4 LT |
277 | int referenced = 0; |
278 | ||
1da177e4 LT |
279 | address = vma_address(page, vma); |
280 | if (address == -EFAULT) | |
281 | goto out; | |
282 | ||
c0718806 HD |
283 | pte = page_check_address(page, mm, address, &ptl); |
284 | if (!pte) | |
285 | goto out; | |
1da177e4 | 286 | |
c0718806 HD |
287 | if (ptep_clear_flush_young(vma, address, pte)) |
288 | referenced++; | |
1da177e4 | 289 | |
c0718806 HD |
290 | /* Pretend the page is referenced if the task has the |
291 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 292 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
293 | rwsem_is_locked(&mm->mmap_sem)) |
294 | referenced++; | |
295 | ||
296 | (*mapcount)--; | |
297 | pte_unmap_unlock(pte, ptl); | |
1da177e4 LT |
298 | out: |
299 | return referenced; | |
300 | } | |
301 | ||
f7b7fd8f | 302 | static int page_referenced_anon(struct page *page) |
1da177e4 LT |
303 | { |
304 | unsigned int mapcount; | |
305 | struct anon_vma *anon_vma; | |
306 | struct vm_area_struct *vma; | |
307 | int referenced = 0; | |
308 | ||
309 | anon_vma = page_lock_anon_vma(page); | |
310 | if (!anon_vma) | |
311 | return referenced; | |
312 | ||
313 | mapcount = page_mapcount(page); | |
314 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
f7b7fd8f | 315 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
316 | if (!mapcount) |
317 | break; | |
318 | } | |
34bbd704 ON |
319 | |
320 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
321 | return referenced; |
322 | } | |
323 | ||
324 | /** | |
325 | * page_referenced_file - referenced check for object-based rmap | |
326 | * @page: the page we're checking references on. | |
327 | * | |
328 | * For an object-based mapped page, find all the places it is mapped and | |
329 | * check/clear the referenced flag. This is done by following the page->mapping | |
330 | * pointer, then walking the chain of vmas it holds. It returns the number | |
331 | * of references it found. | |
332 | * | |
333 | * This function is only called from page_referenced for object-based pages. | |
334 | */ | |
f7b7fd8f | 335 | static int page_referenced_file(struct page *page) |
1da177e4 LT |
336 | { |
337 | unsigned int mapcount; | |
338 | struct address_space *mapping = page->mapping; | |
339 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
340 | struct vm_area_struct *vma; | |
341 | struct prio_tree_iter iter; | |
342 | int referenced = 0; | |
343 | ||
344 | /* | |
345 | * The caller's checks on page->mapping and !PageAnon have made | |
346 | * sure that this is a file page: the check for page->mapping | |
347 | * excludes the case just before it gets set on an anon page. | |
348 | */ | |
349 | BUG_ON(PageAnon(page)); | |
350 | ||
351 | /* | |
352 | * The page lock not only makes sure that page->mapping cannot | |
353 | * suddenly be NULLified by truncation, it makes sure that the | |
354 | * structure at mapping cannot be freed and reused yet, | |
355 | * so we can safely take mapping->i_mmap_lock. | |
356 | */ | |
357 | BUG_ON(!PageLocked(page)); | |
358 | ||
359 | spin_lock(&mapping->i_mmap_lock); | |
360 | ||
361 | /* | |
362 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
363 | * is more likely to be accurate if we note it after spinning. | |
364 | */ | |
365 | mapcount = page_mapcount(page); | |
366 | ||
367 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
368 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) | |
369 | == (VM_LOCKED|VM_MAYSHARE)) { | |
370 | referenced++; | |
371 | break; | |
372 | } | |
f7b7fd8f | 373 | referenced += page_referenced_one(page, vma, &mapcount); |
1da177e4 LT |
374 | if (!mapcount) |
375 | break; | |
376 | } | |
377 | ||
378 | spin_unlock(&mapping->i_mmap_lock); | |
379 | return referenced; | |
380 | } | |
381 | ||
382 | /** | |
383 | * page_referenced - test if the page was referenced | |
384 | * @page: the page to test | |
385 | * @is_locked: caller holds lock on the page | |
386 | * | |
387 | * Quick test_and_clear_referenced for all mappings to a page, | |
388 | * returns the number of ptes which referenced the page. | |
389 | */ | |
f7b7fd8f | 390 | int page_referenced(struct page *page, int is_locked) |
1da177e4 LT |
391 | { |
392 | int referenced = 0; | |
393 | ||
1da177e4 LT |
394 | if (page_test_and_clear_young(page)) |
395 | referenced++; | |
396 | ||
397 | if (TestClearPageReferenced(page)) | |
398 | referenced++; | |
399 | ||
400 | if (page_mapped(page) && page->mapping) { | |
401 | if (PageAnon(page)) | |
f7b7fd8f | 402 | referenced += page_referenced_anon(page); |
1da177e4 | 403 | else if (is_locked) |
f7b7fd8f | 404 | referenced += page_referenced_file(page); |
1da177e4 LT |
405 | else if (TestSetPageLocked(page)) |
406 | referenced++; | |
407 | else { | |
408 | if (page->mapping) | |
f7b7fd8f | 409 | referenced += page_referenced_file(page); |
1da177e4 LT |
410 | unlock_page(page); |
411 | } | |
412 | } | |
413 | return referenced; | |
414 | } | |
415 | ||
d08b3851 PZ |
416 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma) |
417 | { | |
418 | struct mm_struct *mm = vma->vm_mm; | |
419 | unsigned long address; | |
c2fda5fe | 420 | pte_t *pte; |
d08b3851 PZ |
421 | spinlock_t *ptl; |
422 | int ret = 0; | |
423 | ||
424 | address = vma_address(page, vma); | |
425 | if (address == -EFAULT) | |
426 | goto out; | |
427 | ||
428 | pte = page_check_address(page, mm, address, &ptl); | |
429 | if (!pte) | |
430 | goto out; | |
431 | ||
c2fda5fe PZ |
432 | if (pte_dirty(*pte) || pte_write(*pte)) { |
433 | pte_t entry; | |
d08b3851 | 434 | |
c2fda5fe PZ |
435 | flush_cache_page(vma, address, pte_pfn(*pte)); |
436 | entry = ptep_clear_flush(vma, address, pte); | |
437 | entry = pte_wrprotect(entry); | |
438 | entry = pte_mkclean(entry); | |
d6e88e67 | 439 | set_pte_at(mm, address, pte, entry); |
c2fda5fe PZ |
440 | ret = 1; |
441 | } | |
d08b3851 | 442 | |
d08b3851 PZ |
443 | pte_unmap_unlock(pte, ptl); |
444 | out: | |
445 | return ret; | |
446 | } | |
447 | ||
448 | static int page_mkclean_file(struct address_space *mapping, struct page *page) | |
449 | { | |
450 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
451 | struct vm_area_struct *vma; | |
452 | struct prio_tree_iter iter; | |
453 | int ret = 0; | |
454 | ||
455 | BUG_ON(PageAnon(page)); | |
456 | ||
457 | spin_lock(&mapping->i_mmap_lock); | |
458 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
459 | if (vma->vm_flags & VM_SHARED) | |
460 | ret += page_mkclean_one(page, vma); | |
461 | } | |
462 | spin_unlock(&mapping->i_mmap_lock); | |
463 | return ret; | |
464 | } | |
465 | ||
466 | int page_mkclean(struct page *page) | |
467 | { | |
468 | int ret = 0; | |
469 | ||
470 | BUG_ON(!PageLocked(page)); | |
471 | ||
472 | if (page_mapped(page)) { | |
473 | struct address_space *mapping = page_mapping(page); | |
ce7e9fae | 474 | if (mapping) { |
d08b3851 | 475 | ret = page_mkclean_file(mapping, page); |
ce7e9fae CB |
476 | if (page_test_dirty(page)) { |
477 | page_clear_dirty(page); | |
478 | ret = 1; | |
479 | } | |
6c210482 | 480 | } |
d08b3851 PZ |
481 | } |
482 | ||
483 | return ret; | |
484 | } | |
60b59bea | 485 | EXPORT_SYMBOL_GPL(page_mkclean); |
d08b3851 | 486 | |
9617d95e NP |
487 | /** |
488 | * page_set_anon_rmap - setup new anonymous rmap | |
489 | * @page: the page to add the mapping to | |
490 | * @vma: the vm area in which the mapping is added | |
491 | * @address: the user virtual address mapped | |
492 | */ | |
493 | static void __page_set_anon_rmap(struct page *page, | |
494 | struct vm_area_struct *vma, unsigned long address) | |
495 | { | |
496 | struct anon_vma *anon_vma = vma->anon_vma; | |
497 | ||
498 | BUG_ON(!anon_vma); | |
499 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
500 | page->mapping = (struct address_space *) anon_vma; | |
501 | ||
502 | page->index = linear_page_index(vma, address); | |
503 | ||
a74609fa NP |
504 | /* |
505 | * nr_mapped state can be updated without turning off | |
506 | * interrupts because it is not modified via interrupt. | |
507 | */ | |
f3dbd344 | 508 | __inc_zone_page_state(page, NR_ANON_PAGES); |
9617d95e NP |
509 | } |
510 | ||
c97a9e10 NP |
511 | /** |
512 | * page_set_anon_rmap - sanity check anonymous rmap addition | |
513 | * @page: the page to add the mapping to | |
514 | * @vma: the vm area in which the mapping is added | |
515 | * @address: the user virtual address mapped | |
516 | */ | |
517 | static void __page_check_anon_rmap(struct page *page, | |
518 | struct vm_area_struct *vma, unsigned long address) | |
519 | { | |
520 | #ifdef CONFIG_DEBUG_VM | |
521 | /* | |
522 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
523 | * be set up correctly at this point. | |
524 | * | |
525 | * We have exclusion against page_add_anon_rmap because the caller | |
526 | * always holds the page locked, except if called from page_dup_rmap, | |
527 | * in which case the page is already known to be setup. | |
528 | * | |
529 | * We have exclusion against page_add_new_anon_rmap because those pages | |
530 | * are initially only visible via the pagetables, and the pte is locked | |
531 | * over the call to page_add_new_anon_rmap. | |
532 | */ | |
533 | struct anon_vma *anon_vma = vma->anon_vma; | |
534 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
535 | BUG_ON(page->mapping != (struct address_space *)anon_vma); | |
536 | BUG_ON(page->index != linear_page_index(vma, address)); | |
537 | #endif | |
538 | } | |
539 | ||
1da177e4 LT |
540 | /** |
541 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
542 | * @page: the page to add the mapping to | |
543 | * @vma: the vm area in which the mapping is added | |
544 | * @address: the user virtual address mapped | |
545 | * | |
c97a9e10 | 546 | * The caller needs to hold the pte lock and the page must be locked. |
1da177e4 LT |
547 | */ |
548 | void page_add_anon_rmap(struct page *page, | |
549 | struct vm_area_struct *vma, unsigned long address) | |
550 | { | |
c97a9e10 NP |
551 | VM_BUG_ON(!PageLocked(page)); |
552 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
9617d95e NP |
553 | if (atomic_inc_and_test(&page->_mapcount)) |
554 | __page_set_anon_rmap(page, vma, address); | |
c97a9e10 NP |
555 | else |
556 | __page_check_anon_rmap(page, vma, address); | |
1da177e4 LT |
557 | } |
558 | ||
9617d95e NP |
559 | /* |
560 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
561 | * @page: the page to add the mapping to | |
562 | * @vma: the vm area in which the mapping is added | |
563 | * @address: the user virtual address mapped | |
564 | * | |
565 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
566 | * This means the inc-and-test can be bypassed. | |
c97a9e10 | 567 | * Page does not have to be locked. |
9617d95e NP |
568 | */ |
569 | void page_add_new_anon_rmap(struct page *page, | |
570 | struct vm_area_struct *vma, unsigned long address) | |
571 | { | |
c97a9e10 | 572 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
9617d95e NP |
573 | atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ |
574 | __page_set_anon_rmap(page, vma, address); | |
575 | } | |
576 | ||
1da177e4 LT |
577 | /** |
578 | * page_add_file_rmap - add pte mapping to a file page | |
579 | * @page: the page to add the mapping to | |
580 | * | |
b8072f09 | 581 | * The caller needs to hold the pte lock. |
1da177e4 LT |
582 | */ |
583 | void page_add_file_rmap(struct page *page) | |
584 | { | |
1da177e4 | 585 | if (atomic_inc_and_test(&page->_mapcount)) |
65ba55f5 | 586 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
1da177e4 LT |
587 | } |
588 | ||
c97a9e10 NP |
589 | #ifdef CONFIG_DEBUG_VM |
590 | /** | |
591 | * page_dup_rmap - duplicate pte mapping to a page | |
592 | * @page: the page to add the mapping to | |
593 | * | |
594 | * For copy_page_range only: minimal extract from page_add_file_rmap / | |
595 | * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's | |
596 | * quicker. | |
597 | * | |
598 | * The caller needs to hold the pte lock. | |
599 | */ | |
600 | void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address) | |
601 | { | |
602 | BUG_ON(page_mapcount(page) == 0); | |
603 | if (PageAnon(page)) | |
604 | __page_check_anon_rmap(page, vma, address); | |
605 | atomic_inc(&page->_mapcount); | |
606 | } | |
607 | #endif | |
608 | ||
1da177e4 LT |
609 | /** |
610 | * page_remove_rmap - take down pte mapping from a page | |
611 | * @page: page to remove mapping from | |
612 | * | |
b8072f09 | 613 | * The caller needs to hold the pte lock. |
1da177e4 | 614 | */ |
7de6b805 | 615 | void page_remove_rmap(struct page *page, struct vm_area_struct *vma) |
1da177e4 | 616 | { |
1da177e4 | 617 | if (atomic_add_negative(-1, &page->_mapcount)) { |
b7ab795b | 618 | if (unlikely(page_mapcount(page) < 0)) { |
ef2bf0dc | 619 | printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page)); |
7de6b805 | 620 | printk (KERN_EMERG " page pfn = %lx\n", page_to_pfn(page)); |
ef2bf0dc DJ |
621 | printk (KERN_EMERG " page->flags = %lx\n", page->flags); |
622 | printk (KERN_EMERG " page->count = %x\n", page_count(page)); | |
623 | printk (KERN_EMERG " page->mapping = %p\n", page->mapping); | |
7de6b805 | 624 | print_symbol (KERN_EMERG " vma->vm_ops = %s\n", (unsigned long)vma->vm_ops); |
54cb8821 | 625 | if (vma->vm_ops) { |
7de6b805 | 626 | print_symbol (KERN_EMERG " vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage); |
54cb8821 NP |
627 | print_symbol (KERN_EMERG " vma->vm_ops->fault = %s\n", (unsigned long)vma->vm_ops->fault); |
628 | } | |
7de6b805 NP |
629 | if (vma->vm_file && vma->vm_file->f_op) |
630 | print_symbol (KERN_EMERG " vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap); | |
b16bc64d | 631 | BUG(); |
ef2bf0dc | 632 | } |
b16bc64d | 633 | |
1da177e4 LT |
634 | /* |
635 | * It would be tidy to reset the PageAnon mapping here, | |
636 | * but that might overwrite a racing page_add_anon_rmap | |
637 | * which increments mapcount after us but sets mapping | |
638 | * before us: so leave the reset to free_hot_cold_page, | |
639 | * and remember that it's only reliable while mapped. | |
640 | * Leaving it set also helps swapoff to reinstate ptes | |
641 | * faster for those pages still in swapcache. | |
642 | */ | |
6c210482 MS |
643 | if (page_test_dirty(page)) { |
644 | page_clear_dirty(page); | |
1da177e4 | 645 | set_page_dirty(page); |
6c210482 | 646 | } |
f3dbd344 CL |
647 | __dec_zone_page_state(page, |
648 | PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED); | |
1da177e4 LT |
649 | } |
650 | } | |
651 | ||
652 | /* | |
653 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
654 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
655 | */ | |
a48d07af | 656 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
7352349a | 657 | int migration) |
1da177e4 LT |
658 | { |
659 | struct mm_struct *mm = vma->vm_mm; | |
660 | unsigned long address; | |
1da177e4 LT |
661 | pte_t *pte; |
662 | pte_t pteval; | |
c0718806 | 663 | spinlock_t *ptl; |
1da177e4 LT |
664 | int ret = SWAP_AGAIN; |
665 | ||
1da177e4 LT |
666 | address = vma_address(page, vma); |
667 | if (address == -EFAULT) | |
668 | goto out; | |
669 | ||
c0718806 HD |
670 | pte = page_check_address(page, mm, address, &ptl); |
671 | if (!pte) | |
81b4082d | 672 | goto out; |
1da177e4 LT |
673 | |
674 | /* | |
675 | * If the page is mlock()d, we cannot swap it out. | |
676 | * If it's recently referenced (perhaps page_referenced | |
677 | * skipped over this mm) then we should reactivate it. | |
678 | */ | |
e6a1530d CL |
679 | if (!migration && ((vma->vm_flags & VM_LOCKED) || |
680 | (ptep_clear_flush_young(vma, address, pte)))) { | |
1da177e4 LT |
681 | ret = SWAP_FAIL; |
682 | goto out_unmap; | |
683 | } | |
684 | ||
1da177e4 LT |
685 | /* Nuke the page table entry. */ |
686 | flush_cache_page(vma, address, page_to_pfn(page)); | |
687 | pteval = ptep_clear_flush(vma, address, pte); | |
688 | ||
689 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
690 | if (pte_dirty(pteval)) | |
691 | set_page_dirty(page); | |
692 | ||
365e9c87 HD |
693 | /* Update high watermark before we lower rss */ |
694 | update_hiwater_rss(mm); | |
695 | ||
1da177e4 | 696 | if (PageAnon(page)) { |
4c21e2f2 | 697 | swp_entry_t entry = { .val = page_private(page) }; |
0697212a CL |
698 | |
699 | if (PageSwapCache(page)) { | |
700 | /* | |
701 | * Store the swap location in the pte. | |
702 | * See handle_pte_fault() ... | |
703 | */ | |
704 | swap_duplicate(entry); | |
705 | if (list_empty(&mm->mmlist)) { | |
706 | spin_lock(&mmlist_lock); | |
707 | if (list_empty(&mm->mmlist)) | |
708 | list_add(&mm->mmlist, &init_mm.mmlist); | |
709 | spin_unlock(&mmlist_lock); | |
710 | } | |
442c9137 | 711 | dec_mm_counter(mm, anon_rss); |
04e62a29 | 712 | #ifdef CONFIG_MIGRATION |
0697212a CL |
713 | } else { |
714 | /* | |
715 | * Store the pfn of the page in a special migration | |
716 | * pte. do_swap_page() will wait until the migration | |
717 | * pte is removed and then restart fault handling. | |
718 | */ | |
719 | BUG_ON(!migration); | |
720 | entry = make_migration_entry(page, pte_write(pteval)); | |
04e62a29 | 721 | #endif |
1da177e4 LT |
722 | } |
723 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
724 | BUG_ON(pte_file(*pte)); | |
4294621f | 725 | } else |
04e62a29 CL |
726 | #ifdef CONFIG_MIGRATION |
727 | if (migration) { | |
728 | /* Establish migration entry for a file page */ | |
729 | swp_entry_t entry; | |
730 | entry = make_migration_entry(page, pte_write(pteval)); | |
731 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
732 | } else | |
733 | #endif | |
4294621f | 734 | dec_mm_counter(mm, file_rss); |
1da177e4 | 735 | |
04e62a29 | 736 | |
7de6b805 | 737 | page_remove_rmap(page, vma); |
1da177e4 LT |
738 | page_cache_release(page); |
739 | ||
740 | out_unmap: | |
c0718806 | 741 | pte_unmap_unlock(pte, ptl); |
1da177e4 LT |
742 | out: |
743 | return ret; | |
744 | } | |
745 | ||
746 | /* | |
747 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
748 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
749 | * Consequently, given a particular page and its ->index, we cannot locate the | |
750 | * ptes which are mapping that page without an exhaustive linear search. | |
751 | * | |
752 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
753 | * maps the file to which the target page belongs. The ->vm_private_data field | |
754 | * holds the current cursor into that scan. Successive searches will circulate | |
755 | * around the vma's virtual address space. | |
756 | * | |
757 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
758 | * more scanning pressure is placed against them as well. Eventually pages | |
759 | * will become fully unmapped and are eligible for eviction. | |
760 | * | |
761 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
762 | * there there won't be many ptes located within the scan cluster. In this case | |
763 | * maybe we could scan further - to the end of the pte page, perhaps. | |
764 | */ | |
765 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
766 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
767 | ||
768 | static void try_to_unmap_cluster(unsigned long cursor, | |
769 | unsigned int *mapcount, struct vm_area_struct *vma) | |
770 | { | |
771 | struct mm_struct *mm = vma->vm_mm; | |
772 | pgd_t *pgd; | |
773 | pud_t *pud; | |
774 | pmd_t *pmd; | |
c0718806 | 775 | pte_t *pte; |
1da177e4 | 776 | pte_t pteval; |
c0718806 | 777 | spinlock_t *ptl; |
1da177e4 LT |
778 | struct page *page; |
779 | unsigned long address; | |
780 | unsigned long end; | |
1da177e4 | 781 | |
1da177e4 LT |
782 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
783 | end = address + CLUSTER_SIZE; | |
784 | if (address < vma->vm_start) | |
785 | address = vma->vm_start; | |
786 | if (end > vma->vm_end) | |
787 | end = vma->vm_end; | |
788 | ||
789 | pgd = pgd_offset(mm, address); | |
790 | if (!pgd_present(*pgd)) | |
c0718806 | 791 | return; |
1da177e4 LT |
792 | |
793 | pud = pud_offset(pgd, address); | |
794 | if (!pud_present(*pud)) | |
c0718806 | 795 | return; |
1da177e4 LT |
796 | |
797 | pmd = pmd_offset(pud, address); | |
798 | if (!pmd_present(*pmd)) | |
c0718806 HD |
799 | return; |
800 | ||
801 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 802 | |
365e9c87 HD |
803 | /* Update high watermark before we lower rss */ |
804 | update_hiwater_rss(mm); | |
805 | ||
c0718806 | 806 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
807 | if (!pte_present(*pte)) |
808 | continue; | |
6aab341e LT |
809 | page = vm_normal_page(vma, address, *pte); |
810 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 LT |
811 | |
812 | if (ptep_clear_flush_young(vma, address, pte)) | |
813 | continue; | |
814 | ||
815 | /* Nuke the page table entry. */ | |
eca35133 | 816 | flush_cache_page(vma, address, pte_pfn(*pte)); |
1da177e4 LT |
817 | pteval = ptep_clear_flush(vma, address, pte); |
818 | ||
819 | /* If nonlinear, store the file page offset in the pte. */ | |
820 | if (page->index != linear_page_index(vma, address)) | |
821 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
822 | ||
823 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
824 | if (pte_dirty(pteval)) | |
825 | set_page_dirty(page); | |
826 | ||
7de6b805 | 827 | page_remove_rmap(page, vma); |
1da177e4 | 828 | page_cache_release(page); |
4294621f | 829 | dec_mm_counter(mm, file_rss); |
1da177e4 LT |
830 | (*mapcount)--; |
831 | } | |
c0718806 | 832 | pte_unmap_unlock(pte - 1, ptl); |
1da177e4 LT |
833 | } |
834 | ||
7352349a | 835 | static int try_to_unmap_anon(struct page *page, int migration) |
1da177e4 LT |
836 | { |
837 | struct anon_vma *anon_vma; | |
838 | struct vm_area_struct *vma; | |
839 | int ret = SWAP_AGAIN; | |
840 | ||
841 | anon_vma = page_lock_anon_vma(page); | |
842 | if (!anon_vma) | |
843 | return ret; | |
844 | ||
845 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { | |
7352349a | 846 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
847 | if (ret == SWAP_FAIL || !page_mapped(page)) |
848 | break; | |
849 | } | |
34bbd704 ON |
850 | |
851 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
852 | return ret; |
853 | } | |
854 | ||
855 | /** | |
856 | * try_to_unmap_file - unmap file page using the object-based rmap method | |
857 | * @page: the page to unmap | |
858 | * | |
859 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
860 | * contained in the address_space struct it points to. | |
861 | * | |
862 | * This function is only called from try_to_unmap for object-based pages. | |
863 | */ | |
7352349a | 864 | static int try_to_unmap_file(struct page *page, int migration) |
1da177e4 LT |
865 | { |
866 | struct address_space *mapping = page->mapping; | |
867 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
868 | struct vm_area_struct *vma; | |
869 | struct prio_tree_iter iter; | |
870 | int ret = SWAP_AGAIN; | |
871 | unsigned long cursor; | |
872 | unsigned long max_nl_cursor = 0; | |
873 | unsigned long max_nl_size = 0; | |
874 | unsigned int mapcount; | |
875 | ||
876 | spin_lock(&mapping->i_mmap_lock); | |
877 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
7352349a | 878 | ret = try_to_unmap_one(page, vma, migration); |
1da177e4 LT |
879 | if (ret == SWAP_FAIL || !page_mapped(page)) |
880 | goto out; | |
881 | } | |
882 | ||
883 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
884 | goto out; | |
885 | ||
886 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
887 | shared.vm_set.list) { | |
e6a1530d | 888 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
889 | continue; |
890 | cursor = (unsigned long) vma->vm_private_data; | |
891 | if (cursor > max_nl_cursor) | |
892 | max_nl_cursor = cursor; | |
893 | cursor = vma->vm_end - vma->vm_start; | |
894 | if (cursor > max_nl_size) | |
895 | max_nl_size = cursor; | |
896 | } | |
897 | ||
898 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ | |
899 | ret = SWAP_FAIL; | |
900 | goto out; | |
901 | } | |
902 | ||
903 | /* | |
904 | * We don't try to search for this page in the nonlinear vmas, | |
905 | * and page_referenced wouldn't have found it anyway. Instead | |
906 | * just walk the nonlinear vmas trying to age and unmap some. | |
907 | * The mapcount of the page we came in with is irrelevant, | |
908 | * but even so use it as a guide to how hard we should try? | |
909 | */ | |
910 | mapcount = page_mapcount(page); | |
911 | if (!mapcount) | |
912 | goto out; | |
913 | cond_resched_lock(&mapping->i_mmap_lock); | |
914 | ||
915 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
916 | if (max_nl_cursor == 0) | |
917 | max_nl_cursor = CLUSTER_SIZE; | |
918 | ||
919 | do { | |
920 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
921 | shared.vm_set.list) { | |
e6a1530d | 922 | if ((vma->vm_flags & VM_LOCKED) && !migration) |
1da177e4 LT |
923 | continue; |
924 | cursor = (unsigned long) vma->vm_private_data; | |
839b9685 | 925 | while ( cursor < max_nl_cursor && |
1da177e4 LT |
926 | cursor < vma->vm_end - vma->vm_start) { |
927 | try_to_unmap_cluster(cursor, &mapcount, vma); | |
928 | cursor += CLUSTER_SIZE; | |
929 | vma->vm_private_data = (void *) cursor; | |
930 | if ((int)mapcount <= 0) | |
931 | goto out; | |
932 | } | |
933 | vma->vm_private_data = (void *) max_nl_cursor; | |
934 | } | |
935 | cond_resched_lock(&mapping->i_mmap_lock); | |
936 | max_nl_cursor += CLUSTER_SIZE; | |
937 | } while (max_nl_cursor <= max_nl_size); | |
938 | ||
939 | /* | |
940 | * Don't loop forever (perhaps all the remaining pages are | |
941 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
942 | * vmas, now forgetting on which ones it had fallen behind. | |
943 | */ | |
101d2be7 HD |
944 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
945 | vma->vm_private_data = NULL; | |
1da177e4 LT |
946 | out: |
947 | spin_unlock(&mapping->i_mmap_lock); | |
948 | return ret; | |
949 | } | |
950 | ||
951 | /** | |
952 | * try_to_unmap - try to remove all page table mappings to a page | |
953 | * @page: the page to get unmapped | |
954 | * | |
955 | * Tries to remove all the page table entries which are mapping this | |
956 | * page, used in the pageout path. Caller must hold the page lock. | |
957 | * Return values are: | |
958 | * | |
959 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
960 | * SWAP_AGAIN - we missed a mapping, try again later | |
961 | * SWAP_FAIL - the page is unswappable | |
962 | */ | |
7352349a | 963 | int try_to_unmap(struct page *page, int migration) |
1da177e4 LT |
964 | { |
965 | int ret; | |
966 | ||
1da177e4 LT |
967 | BUG_ON(!PageLocked(page)); |
968 | ||
969 | if (PageAnon(page)) | |
7352349a | 970 | ret = try_to_unmap_anon(page, migration); |
1da177e4 | 971 | else |
7352349a | 972 | ret = try_to_unmap_file(page, migration); |
1da177e4 LT |
973 | |
974 | if (!page_mapped(page)) | |
975 | ret = SWAP_SUCCESS; | |
976 | return ret; | |
977 | } | |
81b4082d | 978 |