blob: facb8cdca665cc9b80debc37c7c8ba5d6ad7d763 [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
292 if (!get_mm_counter(mm, rss))
293 goto out;
294 address = vma_address(page, vma);
295 if (address == -EFAULT)
296 goto out;
297
Nikita Danilov81b40822005-05-01 08:58:36 -0700298 pte = page_check_address(page, mm, address);
299 if (!IS_ERR(pte)) {
300 if (ptep_clear_flush_young(vma, address, pte))
301 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302
Nikita Danilov81b40822005-05-01 08:58:36 -0700303 if (mm != current->mm && !ignore_token && has_swap_token(mm))
304 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305
Nikita Danilov81b40822005-05-01 08:58:36 -0700306 (*mapcount)--;
307 pte_unmap(pte);
308 spin_unlock(&mm->page_table_lock);
309 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310out:
311 return referenced;
312}
313
314static int page_referenced_anon(struct page *page, int ignore_token)
315{
316 unsigned int mapcount;
317 struct anon_vma *anon_vma;
318 struct vm_area_struct *vma;
319 int referenced = 0;
320
321 anon_vma = page_lock_anon_vma(page);
322 if (!anon_vma)
323 return referenced;
324
325 mapcount = page_mapcount(page);
326 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
327 referenced += page_referenced_one(page, vma, &mapcount,
328 ignore_token);
329 if (!mapcount)
330 break;
331 }
332 spin_unlock(&anon_vma->lock);
333 return referenced;
334}
335
336/**
337 * page_referenced_file - referenced check for object-based rmap
338 * @page: the page we're checking references on.
339 *
340 * For an object-based mapped page, find all the places it is mapped and
341 * check/clear the referenced flag. This is done by following the page->mapping
342 * pointer, then walking the chain of vmas it holds. It returns the number
343 * of references it found.
344 *
345 * This function is only called from page_referenced for object-based pages.
346 */
347static int page_referenced_file(struct page *page, int ignore_token)
348{
349 unsigned int mapcount;
350 struct address_space *mapping = page->mapping;
351 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
352 struct vm_area_struct *vma;
353 struct prio_tree_iter iter;
354 int referenced = 0;
355
356 /*
357 * The caller's checks on page->mapping and !PageAnon have made
358 * sure that this is a file page: the check for page->mapping
359 * excludes the case just before it gets set on an anon page.
360 */
361 BUG_ON(PageAnon(page));
362
363 /*
364 * The page lock not only makes sure that page->mapping cannot
365 * suddenly be NULLified by truncation, it makes sure that the
366 * structure at mapping cannot be freed and reused yet,
367 * so we can safely take mapping->i_mmap_lock.
368 */
369 BUG_ON(!PageLocked(page));
370
371 spin_lock(&mapping->i_mmap_lock);
372
373 /*
374 * i_mmap_lock does not stabilize mapcount at all, but mapcount
375 * is more likely to be accurate if we note it after spinning.
376 */
377 mapcount = page_mapcount(page);
378
379 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
380 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
381 == (VM_LOCKED|VM_MAYSHARE)) {
382 referenced++;
383 break;
384 }
385 referenced += page_referenced_one(page, vma, &mapcount,
386 ignore_token);
387 if (!mapcount)
388 break;
389 }
390
391 spin_unlock(&mapping->i_mmap_lock);
392 return referenced;
393}
394
395/**
396 * page_referenced - test if the page was referenced
397 * @page: the page to test
398 * @is_locked: caller holds lock on the page
399 *
400 * Quick test_and_clear_referenced for all mappings to a page,
401 * returns the number of ptes which referenced the page.
402 */
403int page_referenced(struct page *page, int is_locked, int ignore_token)
404{
405 int referenced = 0;
406
407 if (!swap_token_default_timeout)
408 ignore_token = 1;
409
410 if (page_test_and_clear_young(page))
411 referenced++;
412
413 if (TestClearPageReferenced(page))
414 referenced++;
415
416 if (page_mapped(page) && page->mapping) {
417 if (PageAnon(page))
418 referenced += page_referenced_anon(page, ignore_token);
419 else if (is_locked)
420 referenced += page_referenced_file(page, ignore_token);
421 else if (TestSetPageLocked(page))
422 referenced++;
423 else {
424 if (page->mapping)
425 referenced += page_referenced_file(page,
426 ignore_token);
427 unlock_page(page);
428 }
429 }
430 return referenced;
431}
432
433/**
434 * page_add_anon_rmap - add pte mapping to an anonymous page
435 * @page: the page to add the mapping to
436 * @vma: the vm area in which the mapping is added
437 * @address: the user virtual address mapped
438 *
439 * The caller needs to hold the mm->page_table_lock.
440 */
441void page_add_anon_rmap(struct page *page,
442 struct vm_area_struct *vma, unsigned long address)
443{
444 struct anon_vma *anon_vma = vma->anon_vma;
445 pgoff_t index;
446
447 BUG_ON(PageReserved(page));
448 BUG_ON(!anon_vma);
449
450 inc_mm_counter(vma->vm_mm, anon_rss);
451
452 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
453 index = (address - vma->vm_start) >> PAGE_SHIFT;
454 index += vma->vm_pgoff;
455 index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
456
457 if (atomic_inc_and_test(&page->_mapcount)) {
458 page->index = index;
459 page->mapping = (struct address_space *) anon_vma;
460 inc_page_state(nr_mapped);
461 }
462 /* else checking page index and mapping is racy */
463}
464
465/**
466 * page_add_file_rmap - add pte mapping to a file page
467 * @page: the page to add the mapping to
468 *
469 * The caller needs to hold the mm->page_table_lock.
470 */
471void page_add_file_rmap(struct page *page)
472{
473 BUG_ON(PageAnon(page));
474 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
475 return;
476
477 if (atomic_inc_and_test(&page->_mapcount))
478 inc_page_state(nr_mapped);
479}
480
481/**
482 * page_remove_rmap - take down pte mapping from a page
483 * @page: page to remove mapping from
484 *
485 * Caller needs to hold the mm->page_table_lock.
486 */
487void page_remove_rmap(struct page *page)
488{
489 BUG_ON(PageReserved(page));
490
491 if (atomic_add_negative(-1, &page->_mapcount)) {
492 BUG_ON(page_mapcount(page) < 0);
493 /*
494 * It would be tidy to reset the PageAnon mapping here,
495 * but that might overwrite a racing page_add_anon_rmap
496 * which increments mapcount after us but sets mapping
497 * before us: so leave the reset to free_hot_cold_page,
498 * and remember that it's only reliable while mapped.
499 * Leaving it set also helps swapoff to reinstate ptes
500 * faster for those pages still in swapcache.
501 */
502 if (page_test_and_clear_dirty(page))
503 set_page_dirty(page);
504 dec_page_state(nr_mapped);
505 }
506}
507
508/*
509 * Subfunctions of try_to_unmap: try_to_unmap_one called
510 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
511 */
512static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
513{
514 struct mm_struct *mm = vma->vm_mm;
515 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 pte_t *pte;
517 pte_t pteval;
518 int ret = SWAP_AGAIN;
519
520 if (!get_mm_counter(mm, rss))
521 goto out;
522 address = vma_address(page, vma);
523 if (address == -EFAULT)
524 goto out;
525
Nikita Danilov81b40822005-05-01 08:58:36 -0700526 pte = page_check_address(page, mm, address);
527 if (IS_ERR(pte))
528 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 /*
531 * If the page is mlock()d, we cannot swap it out.
532 * If it's recently referenced (perhaps page_referenced
533 * skipped over this mm) then we should reactivate it.
534 */
535 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
536 ptep_clear_flush_young(vma, address, pte)) {
537 ret = SWAP_FAIL;
538 goto out_unmap;
539 }
540
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 /* Nuke the page table entry. */
542 flush_cache_page(vma, address, page_to_pfn(page));
543 pteval = ptep_clear_flush(vma, address, pte);
544
545 /* Move the dirty bit to the physical page now the pte is gone. */
546 if (pte_dirty(pteval))
547 set_page_dirty(page);
548
549 if (PageAnon(page)) {
550 swp_entry_t entry = { .val = page->private };
551 /*
552 * Store the swap location in the pte.
553 * See handle_pte_fault() ...
554 */
555 BUG_ON(!PageSwapCache(page));
556 swap_duplicate(entry);
557 if (list_empty(&mm->mmlist)) {
558 spin_lock(&mmlist_lock);
559 list_add(&mm->mmlist, &init_mm.mmlist);
560 spin_unlock(&mmlist_lock);
561 }
562 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
563 BUG_ON(pte_file(*pte));
564 dec_mm_counter(mm, anon_rss);
565 }
566
Bjorn Steinbrink202d1822005-05-16 21:53:17 -0700567 dec_mm_counter(mm, rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 page_remove_rmap(page);
569 page_cache_release(page);
570
571out_unmap:
572 pte_unmap(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 spin_unlock(&mm->page_table_lock);
574out:
575 return ret;
576}
577
578/*
579 * objrmap doesn't work for nonlinear VMAs because the assumption that
580 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
581 * Consequently, given a particular page and its ->index, we cannot locate the
582 * ptes which are mapping that page without an exhaustive linear search.
583 *
584 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
585 * maps the file to which the target page belongs. The ->vm_private_data field
586 * holds the current cursor into that scan. Successive searches will circulate
587 * around the vma's virtual address space.
588 *
589 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
590 * more scanning pressure is placed against them as well. Eventually pages
591 * will become fully unmapped and are eligible for eviction.
592 *
593 * For very sparsely populated VMAs this is a little inefficient - chances are
594 * there there won't be many ptes located within the scan cluster. In this case
595 * maybe we could scan further - to the end of the pte page, perhaps.
596 */
597#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
598#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
599
600static void try_to_unmap_cluster(unsigned long cursor,
601 unsigned int *mapcount, struct vm_area_struct *vma)
602{
603 struct mm_struct *mm = vma->vm_mm;
604 pgd_t *pgd;
605 pud_t *pud;
606 pmd_t *pmd;
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700607 pte_t *pte, *original_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 pte_t pteval;
609 struct page *page;
610 unsigned long address;
611 unsigned long end;
612 unsigned long pfn;
613
614 /*
615 * We need the page_table_lock to protect us from page faults,
616 * munmap, fork, etc...
617 */
618 spin_lock(&mm->page_table_lock);
619
620 address = (vma->vm_start + cursor) & CLUSTER_MASK;
621 end = address + CLUSTER_SIZE;
622 if (address < vma->vm_start)
623 address = vma->vm_start;
624 if (end > vma->vm_end)
625 end = vma->vm_end;
626
627 pgd = pgd_offset(mm, address);
628 if (!pgd_present(*pgd))
629 goto out_unlock;
630
631 pud = pud_offset(pgd, address);
632 if (!pud_present(*pud))
633 goto out_unlock;
634
635 pmd = pmd_offset(pud, address);
636 if (!pmd_present(*pmd))
637 goto out_unlock;
638
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700639 for (original_pte = pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 address < end; pte++, address += PAGE_SIZE) {
641
642 if (!pte_present(*pte))
643 continue;
644
645 pfn = pte_pfn(*pte);
646 if (!pfn_valid(pfn))
647 continue;
648
649 page = pfn_to_page(pfn);
650 BUG_ON(PageAnon(page));
651 if (PageReserved(page))
652 continue;
653
654 if (ptep_clear_flush_young(vma, address, pte))
655 continue;
656
657 /* Nuke the page table entry. */
658 flush_cache_page(vma, address, pfn);
659 pteval = ptep_clear_flush(vma, address, pte);
660
661 /* If nonlinear, store the file page offset in the pte. */
662 if (page->index != linear_page_index(vma, address))
663 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
664
665 /* Move the dirty bit to the physical page now the pte is gone. */
666 if (pte_dirty(pteval))
667 set_page_dirty(page);
668
669 page_remove_rmap(page);
670 page_cache_release(page);
671 dec_mm_counter(mm, rss);
672 (*mapcount)--;
673 }
674
William Lee Irwin IIIcafdd8b2005-05-24 19:31:09 -0700675 pte_unmap(original_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676out_unlock:
677 spin_unlock(&mm->page_table_lock);
678}
679
680static int try_to_unmap_anon(struct page *page)
681{
682 struct anon_vma *anon_vma;
683 struct vm_area_struct *vma;
684 int ret = SWAP_AGAIN;
685
686 anon_vma = page_lock_anon_vma(page);
687 if (!anon_vma)
688 return ret;
689
690 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
691 ret = try_to_unmap_one(page, vma);
692 if (ret == SWAP_FAIL || !page_mapped(page))
693 break;
694 }
695 spin_unlock(&anon_vma->lock);
696 return ret;
697}
698
699/**
700 * try_to_unmap_file - unmap file page using the object-based rmap method
701 * @page: the page to unmap
702 *
703 * Find all the mappings of a page using the mapping pointer and the vma chains
704 * contained in the address_space struct it points to.
705 *
706 * This function is only called from try_to_unmap for object-based pages.
707 */
708static int try_to_unmap_file(struct page *page)
709{
710 struct address_space *mapping = page->mapping;
711 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
712 struct vm_area_struct *vma;
713 struct prio_tree_iter iter;
714 int ret = SWAP_AGAIN;
715 unsigned long cursor;
716 unsigned long max_nl_cursor = 0;
717 unsigned long max_nl_size = 0;
718 unsigned int mapcount;
719
720 spin_lock(&mapping->i_mmap_lock);
721 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
722 ret = try_to_unmap_one(page, vma);
723 if (ret == SWAP_FAIL || !page_mapped(page))
724 goto out;
725 }
726
727 if (list_empty(&mapping->i_mmap_nonlinear))
728 goto out;
729
730 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
731 shared.vm_set.list) {
732 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
733 continue;
734 cursor = (unsigned long) vma->vm_private_data;
735 if (cursor > max_nl_cursor)
736 max_nl_cursor = cursor;
737 cursor = vma->vm_end - vma->vm_start;
738 if (cursor > max_nl_size)
739 max_nl_size = cursor;
740 }
741
742 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
743 ret = SWAP_FAIL;
744 goto out;
745 }
746
747 /*
748 * We don't try to search for this page in the nonlinear vmas,
749 * and page_referenced wouldn't have found it anyway. Instead
750 * just walk the nonlinear vmas trying to age and unmap some.
751 * The mapcount of the page we came in with is irrelevant,
752 * but even so use it as a guide to how hard we should try?
753 */
754 mapcount = page_mapcount(page);
755 if (!mapcount)
756 goto out;
757 cond_resched_lock(&mapping->i_mmap_lock);
758
759 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
760 if (max_nl_cursor == 0)
761 max_nl_cursor = CLUSTER_SIZE;
762
763 do {
764 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
765 shared.vm_set.list) {
766 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
767 continue;
768 cursor = (unsigned long) vma->vm_private_data;
769 while (get_mm_counter(vma->vm_mm, rss) &&
770 cursor < max_nl_cursor &&
771 cursor < vma->vm_end - vma->vm_start) {
772 try_to_unmap_cluster(cursor, &mapcount, vma);
773 cursor += CLUSTER_SIZE;
774 vma->vm_private_data = (void *) cursor;
775 if ((int)mapcount <= 0)
776 goto out;
777 }
778 vma->vm_private_data = (void *) max_nl_cursor;
779 }
780 cond_resched_lock(&mapping->i_mmap_lock);
781 max_nl_cursor += CLUSTER_SIZE;
782 } while (max_nl_cursor <= max_nl_size);
783
784 /*
785 * Don't loop forever (perhaps all the remaining pages are
786 * in locked vmas). Reset cursor on all unreserved nonlinear
787 * vmas, now forgetting on which ones it had fallen behind.
788 */
789 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
790 shared.vm_set.list) {
791 if (!(vma->vm_flags & VM_RESERVED))
792 vma->vm_private_data = NULL;
793 }
794out:
795 spin_unlock(&mapping->i_mmap_lock);
796 return ret;
797}
798
799/**
800 * try_to_unmap - try to remove all page table mappings to a page
801 * @page: the page to get unmapped
802 *
803 * Tries to remove all the page table entries which are mapping this
804 * page, used in the pageout path. Caller must hold the page lock.
805 * Return values are:
806 *
807 * SWAP_SUCCESS - we succeeded in removing all mappings
808 * SWAP_AGAIN - we missed a mapping, try again later
809 * SWAP_FAIL - the page is unswappable
810 */
811int try_to_unmap(struct page *page)
812{
813 int ret;
814
815 BUG_ON(PageReserved(page));
816 BUG_ON(!PageLocked(page));
817
818 if (PageAnon(page))
819 ret = try_to_unmap_anon(page);
820 else
821 ret = try_to_unmap_file(page);
822
823 if (!page_mapped(page))
824 ret = SWAP_SUCCESS;
825 return ret;
826}
Nikita Danilov81b40822005-05-01 08:58:36 -0700827