blob: 1fc559e09ca8c75793ff1dc1b5d6f8ddb7ee2abe [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
Rik Van Rielfcdae292005-10-29 18:15:46 -0700301 /* Pretend the page is referenced if the task has the
302 swap token and is in the middle of a page fault. */
303 if (mm != current->mm && !ignore_token &&
304 has_swap_token(mm) &&
305 rwsem_is_locked(&mm->mmap_sem))
Nikita Danilov81b40822005-05-01 08:58:36 -0700306 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307
Nikita Danilov81b40822005-05-01 08:58:36 -0700308 (*mapcount)--;
309 pte_unmap(pte);
310 spin_unlock(&mm->page_table_lock);
311 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312out:
313 return referenced;
314}
315
316static int page_referenced_anon(struct page *page, int ignore_token)
317{
318 unsigned int mapcount;
319 struct anon_vma *anon_vma;
320 struct vm_area_struct *vma;
321 int referenced = 0;
322
323 anon_vma = page_lock_anon_vma(page);
324 if (!anon_vma)
325 return referenced;
326
327 mapcount = page_mapcount(page);
328 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
329 referenced += page_referenced_one(page, vma, &mapcount,
330 ignore_token);
331 if (!mapcount)
332 break;
333 }
334 spin_unlock(&anon_vma->lock);
335 return referenced;
336}
337
338/**
339 * page_referenced_file - referenced check for object-based rmap
340 * @page: the page we're checking references on.
341 *
342 * For an object-based mapped page, find all the places it is mapped and
343 * check/clear the referenced flag. This is done by following the page->mapping
344 * pointer, then walking the chain of vmas it holds. It returns the number
345 * of references it found.
346 *
347 * This function is only called from page_referenced for object-based pages.
348 */
349static int page_referenced_file(struct page *page, int ignore_token)
350{
351 unsigned int mapcount;
352 struct address_space *mapping = page->mapping;
353 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
354 struct vm_area_struct *vma;
355 struct prio_tree_iter iter;
356 int referenced = 0;
357
358 /*
359 * The caller's checks on page->mapping and !PageAnon have made
360 * sure that this is a file page: the check for page->mapping
361 * excludes the case just before it gets set on an anon page.
362 */
363 BUG_ON(PageAnon(page));
364
365 /*
366 * The page lock not only makes sure that page->mapping cannot
367 * suddenly be NULLified by truncation, it makes sure that the
368 * structure at mapping cannot be freed and reused yet,
369 * so we can safely take mapping->i_mmap_lock.
370 */
371 BUG_ON(!PageLocked(page));
372
373 spin_lock(&mapping->i_mmap_lock);
374
375 /*
376 * i_mmap_lock does not stabilize mapcount at all, but mapcount
377 * is more likely to be accurate if we note it after spinning.
378 */
379 mapcount = page_mapcount(page);
380
381 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
382 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
383 == (VM_LOCKED|VM_MAYSHARE)) {
384 referenced++;
385 break;
386 }
387 referenced += page_referenced_one(page, vma, &mapcount,
388 ignore_token);
389 if (!mapcount)
390 break;
391 }
392
393 spin_unlock(&mapping->i_mmap_lock);
394 return referenced;
395}
396
397/**
398 * page_referenced - test if the page was referenced
399 * @page: the page to test
400 * @is_locked: caller holds lock on the page
401 *
402 * Quick test_and_clear_referenced for all mappings to a page,
403 * returns the number of ptes which referenced the page.
404 */
405int page_referenced(struct page *page, int is_locked, int ignore_token)
406{
407 int referenced = 0;
408
409 if (!swap_token_default_timeout)
410 ignore_token = 1;
411
412 if (page_test_and_clear_young(page))
413 referenced++;
414
415 if (TestClearPageReferenced(page))
416 referenced++;
417
418 if (page_mapped(page) && page->mapping) {
419 if (PageAnon(page))
420 referenced += page_referenced_anon(page, ignore_token);
421 else if (is_locked)
422 referenced += page_referenced_file(page, ignore_token);
423 else if (TestSetPageLocked(page))
424 referenced++;
425 else {
426 if (page->mapping)
427 referenced += page_referenced_file(page,
428 ignore_token);
429 unlock_page(page);
430 }
431 }
432 return referenced;
433}
434
435/**
436 * page_add_anon_rmap - add pte mapping to an anonymous page
437 * @page: the page to add the mapping to
438 * @vma: the vm area in which the mapping is added
439 * @address: the user virtual address mapped
440 *
441 * The caller needs to hold the mm->page_table_lock.
442 */
443void page_add_anon_rmap(struct page *page,
444 struct vm_area_struct *vma, unsigned long address)
445{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 BUG_ON(PageReserved(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447
448 inc_mm_counter(vma->vm_mm, anon_rss);
449
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 if (atomic_inc_and_test(&page->_mapcount)) {
Nick Piggin2822c1a2005-09-03 15:54:47 -0700451 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin2822c1a2005-09-03 15:54:47 -0700452
453 BUG_ON(!anon_vma);
454 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 page->mapping = (struct address_space *) anon_vma;
Nick Piggin2822c1a2005-09-03 15:54:47 -0700456
Nick Piggin4d7670e2005-09-03 15:54:48 -0700457 page->index = linear_page_index(vma, address);
Nick Piggin2822c1a2005-09-03 15:54:47 -0700458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 inc_page_state(nr_mapped);
460 }
461 /* else checking page index and mapping is racy */
462}
463
464/**
465 * page_add_file_rmap - add pte mapping to a file page
466 * @page: the page to add the mapping to
467 *
468 * The caller needs to hold the mm->page_table_lock.
469 */
470void page_add_file_rmap(struct page *page)
471{
472 BUG_ON(PageAnon(page));
473 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
474 return;
475
476 if (atomic_inc_and_test(&page->_mapcount))
477 inc_page_state(nr_mapped);
478}
479
480/**
481 * page_remove_rmap - take down pte mapping from a page
482 * @page: page to remove mapping from
483 *
484 * Caller needs to hold the mm->page_table_lock.
485 */
486void page_remove_rmap(struct page *page)
487{
488 BUG_ON(PageReserved(page));
489
490 if (atomic_add_negative(-1, &page->_mapcount)) {
491 BUG_ON(page_mapcount(page) < 0);
492 /*
493 * It would be tidy to reset the PageAnon mapping here,
494 * but that might overwrite a racing page_add_anon_rmap
495 * which increments mapcount after us but sets mapping
496 * before us: so leave the reset to free_hot_cold_page,
497 * and remember that it's only reliable while mapped.
498 * Leaving it set also helps swapoff to reinstate ptes
499 * faster for those pages still in swapcache.
500 */
501 if (page_test_and_clear_dirty(page))
502 set_page_dirty(page);
503 dec_page_state(nr_mapped);
504 }
505}
506
507/*
508 * Subfunctions of try_to_unmap: try_to_unmap_one called
509 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
510 */
511static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
512{
513 struct mm_struct *mm = vma->vm_mm;
514 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 pte_t *pte;
516 pte_t pteval;
517 int ret = SWAP_AGAIN;
518
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 address = vma_address(page, vma);
520 if (address == -EFAULT)
521 goto out;
522
Nikita Danilov81b40822005-05-01 08:58:36 -0700523 pte = page_check_address(page, mm, address);
524 if (IS_ERR(pte))
525 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526
527 /*
528 * If the page is mlock()d, we cannot swap it out.
529 * If it's recently referenced (perhaps page_referenced
530 * skipped over this mm) then we should reactivate it.
Nick Pigginc3dce2d2005-09-03 15:54:46 -0700531 *
532 * Pages belonging to VM_RESERVED regions should not happen here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 */
534 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
535 ptep_clear_flush_young(vma, address, pte)) {
536 ret = SWAP_FAIL;
537 goto out_unmap;
538 }
539
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 /* Nuke the page table entry. */
541 flush_cache_page(vma, address, page_to_pfn(page));
542 pteval = ptep_clear_flush(vma, address, pte);
543
544 /* Move the dirty bit to the physical page now the pte is gone. */
545 if (pte_dirty(pteval))
546 set_page_dirty(page);
547
548 if (PageAnon(page)) {
549 swp_entry_t entry = { .val = page->private };
550 /*
551 * Store the swap location in the pte.
552 * See handle_pte_fault() ...
553 */
554 BUG_ON(!PageSwapCache(page));
555 swap_duplicate(entry);
556 if (list_empty(&mm->mmlist)) {
557 spin_lock(&mmlist_lock);
558 list_add(&mm->mmlist, &init_mm.mmlist);
559 spin_unlock(&mmlist_lock);
560 }
561 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
562 BUG_ON(pte_file(*pte));
563 dec_mm_counter(mm, anon_rss);
564 }
565
Bjorn Steinbrink202d1822005-05-16 21:53:17 -0700566 dec_mm_counter(mm, rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 page_remove_rmap(page);
568 page_cache_release(page);
569
570out_unmap:
571 pte_unmap(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 spin_unlock(&mm->page_table_lock);
573out:
574 return ret;
575}
576
577/*
578 * objrmap doesn't work for nonlinear VMAs because the assumption that
579 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
580 * Consequently, given a particular page and its ->index, we cannot locate the
581 * ptes which are mapping that page without an exhaustive linear search.
582 *
583 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
584 * maps the file to which the target page belongs. The ->vm_private_data field
585 * holds the current cursor into that scan. Successive searches will circulate
586 * around the vma's virtual address space.
587 *
588 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
589 * more scanning pressure is placed against them as well. Eventually pages
590 * will become fully unmapped and are eligible for eviction.
591 *
592 * For very sparsely populated VMAs this is a little inefficient - chances are
593 * there there won't be many ptes located within the scan cluster. In this case
594 * maybe we could scan further - to the end of the pte page, perhaps.
595 */
596#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
597#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
598
599static void try_to_unmap_cluster(unsigned long cursor,
600 unsigned int *mapcount, struct vm_area_struct *vma)
601{
602 struct mm_struct *mm = vma->vm_mm;
603 pgd_t *pgd;
604 pud_t *pud;
605 pmd_t *pmd;
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700606 pte_t *pte, *original_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 pte_t pteval;
608 struct page *page;
609 unsigned long address;
610 unsigned long end;
611 unsigned long pfn;
612
613 /*
614 * We need the page_table_lock to protect us from page faults,
615 * munmap, fork, etc...
616 */
617 spin_lock(&mm->page_table_lock);
618
619 address = (vma->vm_start + cursor) & CLUSTER_MASK;
620 end = address + CLUSTER_SIZE;
621 if (address < vma->vm_start)
622 address = vma->vm_start;
623 if (end > vma->vm_end)
624 end = vma->vm_end;
625
626 pgd = pgd_offset(mm, address);
627 if (!pgd_present(*pgd))
628 goto out_unlock;
629
630 pud = pud_offset(pgd, address);
631 if (!pud_present(*pud))
632 goto out_unlock;
633
634 pmd = pmd_offset(pud, address);
635 if (!pmd_present(*pmd))
636 goto out_unlock;
637
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700638 for (original_pte = pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 address < end; pte++, address += PAGE_SIZE) {
640
641 if (!pte_present(*pte))
642 continue;
643
644 pfn = pte_pfn(*pte);
645 if (!pfn_valid(pfn))
646 continue;
647
648 page = pfn_to_page(pfn);
649 BUG_ON(PageAnon(page));
650 if (PageReserved(page))
651 continue;
652
653 if (ptep_clear_flush_young(vma, address, pte))
654 continue;
655
656 /* Nuke the page table entry. */
657 flush_cache_page(vma, address, pfn);
658 pteval = ptep_clear_flush(vma, address, pte);
659
660 /* If nonlinear, store the file page offset in the pte. */
661 if (page->index != linear_page_index(vma, address))
662 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
663
664 /* Move the dirty bit to the physical page now the pte is gone. */
665 if (pte_dirty(pteval))
666 set_page_dirty(page);
667
668 page_remove_rmap(page);
669 page_cache_release(page);
670 dec_mm_counter(mm, rss);
671 (*mapcount)--;
672 }
673
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700674 pte_unmap(original_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675out_unlock:
676 spin_unlock(&mm->page_table_lock);
677}
678
679static int try_to_unmap_anon(struct page *page)
680{
681 struct anon_vma *anon_vma;
682 struct vm_area_struct *vma;
683 int ret = SWAP_AGAIN;
684
685 anon_vma = page_lock_anon_vma(page);
686 if (!anon_vma)
687 return ret;
688
689 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
690 ret = try_to_unmap_one(page, vma);
691 if (ret == SWAP_FAIL || !page_mapped(page))
692 break;
693 }
694 spin_unlock(&anon_vma->lock);
695 return ret;
696}
697
698/**
699 * try_to_unmap_file - unmap file page using the object-based rmap method
700 * @page: the page to unmap
701 *
702 * Find all the mappings of a page using the mapping pointer and the vma chains
703 * contained in the address_space struct it points to.
704 *
705 * This function is only called from try_to_unmap for object-based pages.
706 */
707static int try_to_unmap_file(struct page *page)
708{
709 struct address_space *mapping = page->mapping;
710 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
711 struct vm_area_struct *vma;
712 struct prio_tree_iter iter;
713 int ret = SWAP_AGAIN;
714 unsigned long cursor;
715 unsigned long max_nl_cursor = 0;
716 unsigned long max_nl_size = 0;
717 unsigned int mapcount;
718
719 spin_lock(&mapping->i_mmap_lock);
720 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
721 ret = try_to_unmap_one(page, vma);
722 if (ret == SWAP_FAIL || !page_mapped(page))
723 goto out;
724 }
725
726 if (list_empty(&mapping->i_mmap_nonlinear))
727 goto out;
728
729 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
730 shared.vm_set.list) {
731 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
732 continue;
733 cursor = (unsigned long) vma->vm_private_data;
734 if (cursor > max_nl_cursor)
735 max_nl_cursor = cursor;
736 cursor = vma->vm_end - vma->vm_start;
737 if (cursor > max_nl_size)
738 max_nl_size = cursor;
739 }
740
741 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
742 ret = SWAP_FAIL;
743 goto out;
744 }
745
746 /*
747 * We don't try to search for this page in the nonlinear vmas,
748 * and page_referenced wouldn't have found it anyway. Instead
749 * just walk the nonlinear vmas trying to age and unmap some.
750 * The mapcount of the page we came in with is irrelevant,
751 * but even so use it as a guide to how hard we should try?
752 */
753 mapcount = page_mapcount(page);
754 if (!mapcount)
755 goto out;
756 cond_resched_lock(&mapping->i_mmap_lock);
757
758 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
759 if (max_nl_cursor == 0)
760 max_nl_cursor = CLUSTER_SIZE;
761
762 do {
763 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
764 shared.vm_set.list) {
765 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
766 continue;
767 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -0700768 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 cursor < vma->vm_end - vma->vm_start) {
770 try_to_unmap_cluster(cursor, &mapcount, vma);
771 cursor += CLUSTER_SIZE;
772 vma->vm_private_data = (void *) cursor;
773 if ((int)mapcount <= 0)
774 goto out;
775 }
776 vma->vm_private_data = (void *) max_nl_cursor;
777 }
778 cond_resched_lock(&mapping->i_mmap_lock);
779 max_nl_cursor += CLUSTER_SIZE;
780 } while (max_nl_cursor <= max_nl_size);
781
782 /*
783 * Don't loop forever (perhaps all the remaining pages are
784 * in locked vmas). Reset cursor on all unreserved nonlinear
785 * vmas, now forgetting on which ones it had fallen behind.
786 */
787 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
788 shared.vm_set.list) {
789 if (!(vma->vm_flags & VM_RESERVED))
790 vma->vm_private_data = NULL;
791 }
792out:
793 spin_unlock(&mapping->i_mmap_lock);
794 return ret;
795}
796
797/**
798 * try_to_unmap - try to remove all page table mappings to a page
799 * @page: the page to get unmapped
800 *
801 * Tries to remove all the page table entries which are mapping this
802 * page, used in the pageout path. Caller must hold the page lock.
803 * Return values are:
804 *
805 * SWAP_SUCCESS - we succeeded in removing all mappings
806 * SWAP_AGAIN - we missed a mapping, try again later
807 * SWAP_FAIL - the page is unswappable
808 */
809int try_to_unmap(struct page *page)
810{
811 int ret;
812
813 BUG_ON(PageReserved(page));
814 BUG_ON(!PageLocked(page));
815
816 if (PageAnon(page))
817 ret = try_to_unmap_anon(page);
818 else
819 ret = try_to_unmap_file(page);
820
821 if (!page_mapped(page))
822 ret = SWAP_SUCCESS;
823 return ret;
824}
Nikita Danilov81b40822005-05-01 08:58:36 -0700825