blob: 450f5241b5a525282ae7eb150fe322423b01e570 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * mm/rmap.c - physical to virtual reverse mappings
3 *
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
6 *
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
9 *
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
13 *
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17 * Contributions by Hugh Dickins <hugh@veritas.com> 2003, 2004
18 */
19
20/*
21 * Lock ordering in mm:
22 *
23 * inode->i_sem (while writing or truncating, not reading or faulting)
24 * inode->i_alloc_sem
25 *
26 * When a page fault occurs in writing from user to file, down_read
27 * of mmap_sem nests within i_sem; in sys_msync, i_sem nests within
28 * down_read of mmap_sem; i_sem and down_write of mmap_sem are never
29 * taken together; in truncation, i_sem is taken outermost.
30 *
31 * mm->mmap_sem
32 * page->flags PG_locked (lock_page)
33 * mapping->i_mmap_lock
34 * anon_vma->lock
35 * mm->page_table_lock
36 * zone->lru_lock (in mark_page_accessed)
Hugh Dickins5d337b92005-09-03 15:54:41 -070037 * swap_lock (in swap_duplicate, swap_info_get)
Linus Torvalds1da177e2005-04-16 15:20:36 -070038 * mmlist_lock (in mmput, drain_mmlist and others)
Linus Torvalds1da177e2005-04-16 15:20:36 -070039 * mapping->private_lock (in __set_page_dirty_buffers)
40 * inode_lock (in set_page_dirty's __mark_inode_dirty)
41 * sb_lock (within inode_lock in fs/fs-writeback.c)
42 * mapping->tree_lock (widely used, in set_page_dirty,
43 * in arch-dependent flush_dcache_mmap_lock,
44 * within inode_lock in __sync_single_inode)
45 */
46
47#include <linux/mm.h>
48#include <linux/pagemap.h>
49#include <linux/swap.h>
50#include <linux/swapops.h>
51#include <linux/slab.h>
52#include <linux/init.h>
53#include <linux/rmap.h>
54#include <linux/rcupdate.h>
55
56#include <asm/tlbflush.h>
57
58//#define RMAP_DEBUG /* can be enabled only for debugging */
59
60kmem_cache_t *anon_vma_cachep;
61
62static inline void validate_anon_vma(struct vm_area_struct *find_vma)
63{
64#ifdef RMAP_DEBUG
65 struct anon_vma *anon_vma = find_vma->anon_vma;
66 struct vm_area_struct *vma;
67 unsigned int mapcount = 0;
68 int found = 0;
69
70 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
71 mapcount++;
72 BUG_ON(mapcount > 100000);
73 if (vma == find_vma)
74 found = 1;
75 }
76 BUG_ON(!found);
77#endif
78}
79
80/* This must be called under the mmap_sem. */
81int anon_vma_prepare(struct vm_area_struct *vma)
82{
83 struct anon_vma *anon_vma = vma->anon_vma;
84
85 might_sleep();
86 if (unlikely(!anon_vma)) {
87 struct mm_struct *mm = vma->vm_mm;
88 struct anon_vma *allocated, *locked;
89
90 anon_vma = find_mergeable_anon_vma(vma);
91 if (anon_vma) {
92 allocated = NULL;
93 locked = anon_vma;
94 spin_lock(&locked->lock);
95 } else {
96 anon_vma = anon_vma_alloc();
97 if (unlikely(!anon_vma))
98 return -ENOMEM;
99 allocated = anon_vma;
100 locked = NULL;
101 }
102
103 /* page_table_lock to protect against threads */
104 spin_lock(&mm->page_table_lock);
105 if (likely(!vma->anon_vma)) {
106 vma->anon_vma = anon_vma;
107 list_add(&vma->anon_vma_node, &anon_vma->head);
108 allocated = NULL;
109 }
110 spin_unlock(&mm->page_table_lock);
111
112 if (locked)
113 spin_unlock(&locked->lock);
114 if (unlikely(allocated))
115 anon_vma_free(allocated);
116 }
117 return 0;
118}
119
120void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
121{
122 BUG_ON(vma->anon_vma != next->anon_vma);
123 list_del(&next->anon_vma_node);
124}
125
126void __anon_vma_link(struct vm_area_struct *vma)
127{
128 struct anon_vma *anon_vma = vma->anon_vma;
129
130 if (anon_vma) {
131 list_add(&vma->anon_vma_node, &anon_vma->head);
132 validate_anon_vma(vma);
133 }
134}
135
136void anon_vma_link(struct vm_area_struct *vma)
137{
138 struct anon_vma *anon_vma = vma->anon_vma;
139
140 if (anon_vma) {
141 spin_lock(&anon_vma->lock);
142 list_add(&vma->anon_vma_node, &anon_vma->head);
143 validate_anon_vma(vma);
144 spin_unlock(&anon_vma->lock);
145 }
146}
147
148void anon_vma_unlink(struct vm_area_struct *vma)
149{
150 struct anon_vma *anon_vma = vma->anon_vma;
151 int empty;
152
153 if (!anon_vma)
154 return;
155
156 spin_lock(&anon_vma->lock);
157 validate_anon_vma(vma);
158 list_del(&vma->anon_vma_node);
159
160 /* We must garbage collect the anon_vma if it's empty */
161 empty = list_empty(&anon_vma->head);
162 spin_unlock(&anon_vma->lock);
163
164 if (empty)
165 anon_vma_free(anon_vma);
166}
167
168static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
169{
170 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
171 SLAB_CTOR_CONSTRUCTOR) {
172 struct anon_vma *anon_vma = data;
173
174 spin_lock_init(&anon_vma->lock);
175 INIT_LIST_HEAD(&anon_vma->head);
176 }
177}
178
179void __init anon_vma_init(void)
180{
181 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
182 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
183}
184
185/*
186 * Getting a lock on a stable anon_vma from a page off the LRU is
187 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
188 */
189static struct anon_vma *page_lock_anon_vma(struct page *page)
190{
191 struct anon_vma *anon_vma = NULL;
192 unsigned long anon_mapping;
193
194 rcu_read_lock();
195 anon_mapping = (unsigned long) page->mapping;
196 if (!(anon_mapping & PAGE_MAPPING_ANON))
197 goto out;
198 if (!page_mapped(page))
199 goto out;
200
201 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
202 spin_lock(&anon_vma->lock);
203out:
204 rcu_read_unlock();
205 return anon_vma;
206}
207
208/*
209 * At what user virtual address is page expected in vma?
210 */
211static inline unsigned long
212vma_address(struct page *page, struct vm_area_struct *vma)
213{
214 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
215 unsigned long address;
216
217 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
218 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
219 /* page should be within any vma from prio_tree_next */
220 BUG_ON(!PageAnon(page));
221 return -EFAULT;
222 }
223 return address;
224}
225
226/*
227 * At what user virtual address is page expected in vma? checking that the
228 * page matches the vma: currently only used by unuse_process, on anon pages.
229 */
230unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
231{
232 if (PageAnon(page)) {
233 if ((void *)vma->anon_vma !=
234 (void *)page->mapping - PAGE_MAPPING_ANON)
235 return -EFAULT;
236 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
237 if (vma->vm_file->f_mapping != page->mapping)
238 return -EFAULT;
239 } else
240 return -EFAULT;
241 return vma_address(page, vma);
242}
243
244/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700245 * Check that @page is mapped at @address into @mm.
246 *
247 * On success returns with mapped pte and locked mm->page_table_lock.
248 */
Carsten Otteceffc072005-06-23 22:05:25 -0700249pte_t *page_check_address(struct page *page, struct mm_struct *mm,
250 unsigned long address)
Nikita Danilov81b40822005-05-01 08:58:36 -0700251{
252 pgd_t *pgd;
253 pud_t *pud;
254 pmd_t *pmd;
255 pte_t *pte;
256
257 /*
258 * We need the page_table_lock to protect us from page faults,
259 * munmap, fork, etc...
260 */
261 spin_lock(&mm->page_table_lock);
262 pgd = pgd_offset(mm, address);
263 if (likely(pgd_present(*pgd))) {
264 pud = pud_offset(pgd, address);
265 if (likely(pud_present(*pud))) {
266 pmd = pmd_offset(pud, address);
267 if (likely(pmd_present(*pmd))) {
268 pte = pte_offset_map(pmd, address);
269 if (likely(pte_present(*pte) &&
270 page_to_pfn(page) == pte_pfn(*pte)))
271 return pte;
272 pte_unmap(pte);
273 }
274 }
275 }
276 spin_unlock(&mm->page_table_lock);
277 return ERR_PTR(-ENOENT);
278}
279
280/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 * Subfunctions of page_referenced: page_referenced_one called
282 * repeatedly from either page_referenced_anon or page_referenced_file.
283 */
284static int page_referenced_one(struct page *page,
285 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
286{
287 struct mm_struct *mm = vma->vm_mm;
288 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 pte_t *pte;
290 int referenced = 0;
291
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 address = vma_address(page, vma);
293 if (address == -EFAULT)
294 goto out;
295
Nikita Danilov81b40822005-05-01 08:58:36 -0700296 pte = page_check_address(page, mm, address);
297 if (!IS_ERR(pte)) {
298 if (ptep_clear_flush_young(vma, address, pte))
299 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
Nikita Danilov81b40822005-05-01 08:58:36 -0700301 if (mm != current->mm && !ignore_token && has_swap_token(mm))
302 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Nikita Danilov81b40822005-05-01 08:58:36 -0700304 (*mapcount)--;
305 pte_unmap(pte);
306 spin_unlock(&mm->page_table_lock);
307 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308out:
309 return referenced;
310}
311
312static int page_referenced_anon(struct page *page, int ignore_token)
313{
314 unsigned int mapcount;
315 struct anon_vma *anon_vma;
316 struct vm_area_struct *vma;
317 int referenced = 0;
318
319 anon_vma = page_lock_anon_vma(page);
320 if (!anon_vma)
321 return referenced;
322
323 mapcount = page_mapcount(page);
324 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
325 referenced += page_referenced_one(page, vma, &mapcount,
326 ignore_token);
327 if (!mapcount)
328 break;
329 }
330 spin_unlock(&anon_vma->lock);
331 return referenced;
332}
333
334/**
335 * page_referenced_file - referenced check for object-based rmap
336 * @page: the page we're checking references on.
337 *
338 * For an object-based mapped page, find all the places it is mapped and
339 * check/clear the referenced flag. This is done by following the page->mapping
340 * pointer, then walking the chain of vmas it holds. It returns the number
341 * of references it found.
342 *
343 * This function is only called from page_referenced for object-based pages.
344 */
345static int page_referenced_file(struct page *page, int ignore_token)
346{
347 unsigned int mapcount;
348 struct address_space *mapping = page->mapping;
349 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
350 struct vm_area_struct *vma;
351 struct prio_tree_iter iter;
352 int referenced = 0;
353
354 /*
355 * The caller's checks on page->mapping and !PageAnon have made
356 * sure that this is a file page: the check for page->mapping
357 * excludes the case just before it gets set on an anon page.
358 */
359 BUG_ON(PageAnon(page));
360
361 /*
362 * The page lock not only makes sure that page->mapping cannot
363 * suddenly be NULLified by truncation, it makes sure that the
364 * structure at mapping cannot be freed and reused yet,
365 * so we can safely take mapping->i_mmap_lock.
366 */
367 BUG_ON(!PageLocked(page));
368
369 spin_lock(&mapping->i_mmap_lock);
370
371 /*
372 * i_mmap_lock does not stabilize mapcount at all, but mapcount
373 * is more likely to be accurate if we note it after spinning.
374 */
375 mapcount = page_mapcount(page);
376
377 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
378 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
379 == (VM_LOCKED|VM_MAYSHARE)) {
380 referenced++;
381 break;
382 }
383 referenced += page_referenced_one(page, vma, &mapcount,
384 ignore_token);
385 if (!mapcount)
386 break;
387 }
388
389 spin_unlock(&mapping->i_mmap_lock);
390 return referenced;
391}
392
393/**
394 * page_referenced - test if the page was referenced
395 * @page: the page to test
396 * @is_locked: caller holds lock on the page
397 *
398 * Quick test_and_clear_referenced for all mappings to a page,
399 * returns the number of ptes which referenced the page.
400 */
401int page_referenced(struct page *page, int is_locked, int ignore_token)
402{
403 int referenced = 0;
404
405 if (!swap_token_default_timeout)
406 ignore_token = 1;
407
408 if (page_test_and_clear_young(page))
409 referenced++;
410
411 if (TestClearPageReferenced(page))
412 referenced++;
413
414 if (page_mapped(page) && page->mapping) {
415 if (PageAnon(page))
416 referenced += page_referenced_anon(page, ignore_token);
417 else if (is_locked)
418 referenced += page_referenced_file(page, ignore_token);
419 else if (TestSetPageLocked(page))
420 referenced++;
421 else {
422 if (page->mapping)
423 referenced += page_referenced_file(page,
424 ignore_token);
425 unlock_page(page);
426 }
427 }
428 return referenced;
429}
430
431/**
432 * page_add_anon_rmap - add pte mapping to an anonymous page
433 * @page: the page to add the mapping to
434 * @vma: the vm area in which the mapping is added
435 * @address: the user virtual address mapped
436 *
437 * The caller needs to hold the mm->page_table_lock.
438 */
439void page_add_anon_rmap(struct page *page,
440 struct vm_area_struct *vma, unsigned long address)
441{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 BUG_ON(PageReserved(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
444 inc_mm_counter(vma->vm_mm, anon_rss);
445
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 if (atomic_inc_and_test(&page->_mapcount)) {
Nick Piggin2822c1a2005-09-03 15:54:47 -0700447 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin2822c1a2005-09-03 15:54:47 -0700448
449 BUG_ON(!anon_vma);
450 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 page->mapping = (struct address_space *) anon_vma;
Nick Piggin2822c1a2005-09-03 15:54:47 -0700452
Nick Piggin4d7670e2005-09-03 15:54:48 -0700453 page->index = linear_page_index(vma, address);
Nick Piggin2822c1a2005-09-03 15:54:47 -0700454
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 inc_page_state(nr_mapped);
456 }
457 /* else checking page index and mapping is racy */
458}
459
460/**
461 * page_add_file_rmap - add pte mapping to a file page
462 * @page: the page to add the mapping to
463 *
464 * The caller needs to hold the mm->page_table_lock.
465 */
466void page_add_file_rmap(struct page *page)
467{
468 BUG_ON(PageAnon(page));
469 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
470 return;
471
472 if (atomic_inc_and_test(&page->_mapcount))
473 inc_page_state(nr_mapped);
474}
475
476/**
477 * page_remove_rmap - take down pte mapping from a page
478 * @page: page to remove mapping from
479 *
480 * Caller needs to hold the mm->page_table_lock.
481 */
482void page_remove_rmap(struct page *page)
483{
484 BUG_ON(PageReserved(page));
485
486 if (atomic_add_negative(-1, &page->_mapcount)) {
487 BUG_ON(page_mapcount(page) < 0);
488 /*
489 * It would be tidy to reset the PageAnon mapping here,
490 * but that might overwrite a racing page_add_anon_rmap
491 * which increments mapcount after us but sets mapping
492 * before us: so leave the reset to free_hot_cold_page,
493 * and remember that it's only reliable while mapped.
494 * Leaving it set also helps swapoff to reinstate ptes
495 * faster for those pages still in swapcache.
496 */
497 if (page_test_and_clear_dirty(page))
498 set_page_dirty(page);
499 dec_page_state(nr_mapped);
500 }
501}
502
503/*
504 * Subfunctions of try_to_unmap: try_to_unmap_one called
505 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
506 */
507static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
508{
509 struct mm_struct *mm = vma->vm_mm;
510 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 pte_t *pte;
512 pte_t pteval;
513 int ret = SWAP_AGAIN;
514
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 address = vma_address(page, vma);
516 if (address == -EFAULT)
517 goto out;
518
Nikita Danilov81b40822005-05-01 08:58:36 -0700519 pte = page_check_address(page, mm, address);
520 if (IS_ERR(pte))
521 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
523 /*
524 * If the page is mlock()d, we cannot swap it out.
525 * If it's recently referenced (perhaps page_referenced
526 * skipped over this mm) then we should reactivate it.
Nick Pigginc3dce2d2005-09-03 15:54:46 -0700527 *
528 * Pages belonging to VM_RESERVED regions should not happen here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 */
530 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
531 ptep_clear_flush_young(vma, address, pte)) {
532 ret = SWAP_FAIL;
533 goto out_unmap;
534 }
535
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 /* Nuke the page table entry. */
537 flush_cache_page(vma, address, page_to_pfn(page));
538 pteval = ptep_clear_flush(vma, address, pte);
539
540 /* Move the dirty bit to the physical page now the pte is gone. */
541 if (pte_dirty(pteval))
542 set_page_dirty(page);
543
544 if (PageAnon(page)) {
545 swp_entry_t entry = { .val = page->private };
546 /*
547 * Store the swap location in the pte.
548 * See handle_pte_fault() ...
549 */
550 BUG_ON(!PageSwapCache(page));
551 swap_duplicate(entry);
552 if (list_empty(&mm->mmlist)) {
553 spin_lock(&mmlist_lock);
554 list_add(&mm->mmlist, &init_mm.mmlist);
555 spin_unlock(&mmlist_lock);
556 }
557 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
558 BUG_ON(pte_file(*pte));
559 dec_mm_counter(mm, anon_rss);
560 }
561
Bjorn Steinbrink202d1822005-05-16 21:53:17 -0700562 dec_mm_counter(mm, rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 page_remove_rmap(page);
564 page_cache_release(page);
565
566out_unmap:
567 pte_unmap(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 spin_unlock(&mm->page_table_lock);
569out:
570 return ret;
571}
572
573/*
574 * objrmap doesn't work for nonlinear VMAs because the assumption that
575 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
576 * Consequently, given a particular page and its ->index, we cannot locate the
577 * ptes which are mapping that page without an exhaustive linear search.
578 *
579 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
580 * maps the file to which the target page belongs. The ->vm_private_data field
581 * holds the current cursor into that scan. Successive searches will circulate
582 * around the vma's virtual address space.
583 *
584 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
585 * more scanning pressure is placed against them as well. Eventually pages
586 * will become fully unmapped and are eligible for eviction.
587 *
588 * For very sparsely populated VMAs this is a little inefficient - chances are
589 * there there won't be many ptes located within the scan cluster. In this case
590 * maybe we could scan further - to the end of the pte page, perhaps.
591 */
592#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
593#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
594
595static void try_to_unmap_cluster(unsigned long cursor,
596 unsigned int *mapcount, struct vm_area_struct *vma)
597{
598 struct mm_struct *mm = vma->vm_mm;
599 pgd_t *pgd;
600 pud_t *pud;
601 pmd_t *pmd;
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700602 pte_t *pte, *original_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 pte_t pteval;
604 struct page *page;
605 unsigned long address;
606 unsigned long end;
607 unsigned long pfn;
608
609 /*
610 * We need the page_table_lock to protect us from page faults,
611 * munmap, fork, etc...
612 */
613 spin_lock(&mm->page_table_lock);
614
615 address = (vma->vm_start + cursor) & CLUSTER_MASK;
616 end = address + CLUSTER_SIZE;
617 if (address < vma->vm_start)
618 address = vma->vm_start;
619 if (end > vma->vm_end)
620 end = vma->vm_end;
621
622 pgd = pgd_offset(mm, address);
623 if (!pgd_present(*pgd))
624 goto out_unlock;
625
626 pud = pud_offset(pgd, address);
627 if (!pud_present(*pud))
628 goto out_unlock;
629
630 pmd = pmd_offset(pud, address);
631 if (!pmd_present(*pmd))
632 goto out_unlock;
633
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700634 for (original_pte = pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 address < end; pte++, address += PAGE_SIZE) {
636
637 if (!pte_present(*pte))
638 continue;
639
640 pfn = pte_pfn(*pte);
641 if (!pfn_valid(pfn))
642 continue;
643
644 page = pfn_to_page(pfn);
645 BUG_ON(PageAnon(page));
646 if (PageReserved(page))
647 continue;
648
649 if (ptep_clear_flush_young(vma, address, pte))
650 continue;
651
652 /* Nuke the page table entry. */
653 flush_cache_page(vma, address, pfn);
654 pteval = ptep_clear_flush(vma, address, pte);
655
656 /* If nonlinear, store the file page offset in the pte. */
657 if (page->index != linear_page_index(vma, address))
658 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
659
660 /* Move the dirty bit to the physical page now the pte is gone. */
661 if (pte_dirty(pteval))
662 set_page_dirty(page);
663
664 page_remove_rmap(page);
665 page_cache_release(page);
666 dec_mm_counter(mm, rss);
667 (*mapcount)--;
668 }
669
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700670 pte_unmap(original_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671out_unlock:
672 spin_unlock(&mm->page_table_lock);
673}
674
675static int try_to_unmap_anon(struct page *page)
676{
677 struct anon_vma *anon_vma;
678 struct vm_area_struct *vma;
679 int ret = SWAP_AGAIN;
680
681 anon_vma = page_lock_anon_vma(page);
682 if (!anon_vma)
683 return ret;
684
685 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
686 ret = try_to_unmap_one(page, vma);
687 if (ret == SWAP_FAIL || !page_mapped(page))
688 break;
689 }
690 spin_unlock(&anon_vma->lock);
691 return ret;
692}
693
694/**
695 * try_to_unmap_file - unmap file page using the object-based rmap method
696 * @page: the page to unmap
697 *
698 * Find all the mappings of a page using the mapping pointer and the vma chains
699 * contained in the address_space struct it points to.
700 *
701 * This function is only called from try_to_unmap for object-based pages.
702 */
703static int try_to_unmap_file(struct page *page)
704{
705 struct address_space *mapping = page->mapping;
706 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
707 struct vm_area_struct *vma;
708 struct prio_tree_iter iter;
709 int ret = SWAP_AGAIN;
710 unsigned long cursor;
711 unsigned long max_nl_cursor = 0;
712 unsigned long max_nl_size = 0;
713 unsigned int mapcount;
714
715 spin_lock(&mapping->i_mmap_lock);
716 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
717 ret = try_to_unmap_one(page, vma);
718 if (ret == SWAP_FAIL || !page_mapped(page))
719 goto out;
720 }
721
722 if (list_empty(&mapping->i_mmap_nonlinear))
723 goto out;
724
725 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
726 shared.vm_set.list) {
727 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
728 continue;
729 cursor = (unsigned long) vma->vm_private_data;
730 if (cursor > max_nl_cursor)
731 max_nl_cursor = cursor;
732 cursor = vma->vm_end - vma->vm_start;
733 if (cursor > max_nl_size)
734 max_nl_size = cursor;
735 }
736
737 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
738 ret = SWAP_FAIL;
739 goto out;
740 }
741
742 /*
743 * We don't try to search for this page in the nonlinear vmas,
744 * and page_referenced wouldn't have found it anyway. Instead
745 * just walk the nonlinear vmas trying to age and unmap some.
746 * The mapcount of the page we came in with is irrelevant,
747 * but even so use it as a guide to how hard we should try?
748 */
749 mapcount = page_mapcount(page);
750 if (!mapcount)
751 goto out;
752 cond_resched_lock(&mapping->i_mmap_lock);
753
754 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
755 if (max_nl_cursor == 0)
756 max_nl_cursor = CLUSTER_SIZE;
757
758 do {
759 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
760 shared.vm_set.list) {
761 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
762 continue;
763 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -0700764 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 cursor < vma->vm_end - vma->vm_start) {
766 try_to_unmap_cluster(cursor, &mapcount, vma);
767 cursor += CLUSTER_SIZE;
768 vma->vm_private_data = (void *) cursor;
769 if ((int)mapcount <= 0)
770 goto out;
771 }
772 vma->vm_private_data = (void *) max_nl_cursor;
773 }
774 cond_resched_lock(&mapping->i_mmap_lock);
775 max_nl_cursor += CLUSTER_SIZE;
776 } while (max_nl_cursor <= max_nl_size);
777
778 /*
779 * Don't loop forever (perhaps all the remaining pages are
780 * in locked vmas). Reset cursor on all unreserved nonlinear
781 * vmas, now forgetting on which ones it had fallen behind.
782 */
783 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
784 shared.vm_set.list) {
785 if (!(vma->vm_flags & VM_RESERVED))
786 vma->vm_private_data = NULL;
787 }
788out:
789 spin_unlock(&mapping->i_mmap_lock);
790 return ret;
791}
792
793/**
794 * try_to_unmap - try to remove all page table mappings to a page
795 * @page: the page to get unmapped
796 *
797 * Tries to remove all the page table entries which are mapping this
798 * page, used in the pageout path. Caller must hold the page lock.
799 * Return values are:
800 *
801 * SWAP_SUCCESS - we succeeded in removing all mappings
802 * SWAP_AGAIN - we missed a mapping, try again later
803 * SWAP_FAIL - the page is unswappable
804 */
805int try_to_unmap(struct page *page)
806{
807 int ret;
808
809 BUG_ON(PageReserved(page));
810 BUG_ON(!PageLocked(page));
811
812 if (PageAnon(page))
813 ret = try_to_unmap_anon(page);
814 else
815 ret = try_to_unmap_file(page);
816
817 if (!page_mapped(page))
818 ret = SWAP_SUCCESS;
819 return ret;
820}
Nikita Danilov81b40822005-05-01 08:58:36 -0700821