2 * mm/rmap.c - physical to virtual reverse mappings
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
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.
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
21 * Lock ordering in mm:
23 * inode->i_mutex (while writing or truncating, not reading or faulting)
24 * inode->i_alloc_sem (vmtruncate_range)
26 * page->flags PG_locked (lock_page)
27 * mapping->i_mmap_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)
42 #include <linux/pagemap.h>
43 #include <linux/swap.h>
44 #include <linux/swapops.h>
45 #include <linux/slab.h>
46 #include <linux/init.h>
47 #include <linux/rmap.h>
48 #include <linux/rcupdate.h>
49 #include <linux/module.h>
50 #include <linux/kallsyms.h>
52 #include <asm/tlbflush.h>
54 struct kmem_cache *anon_vma_cachep;
56 /* This must be called under the mmap_sem. */
57 int anon_vma_prepare(struct vm_area_struct *vma)
59 struct anon_vma *anon_vma = vma->anon_vma;
62 if (unlikely(!anon_vma)) {
63 struct mm_struct *mm = vma->vm_mm;
64 struct anon_vma *allocated, *locked;
66 anon_vma = find_mergeable_anon_vma(vma);
70 spin_lock(&locked->lock);
72 anon_vma = anon_vma_alloc();
73 if (unlikely(!anon_vma))
79 /* page_table_lock to protect against threads */
80 spin_lock(&mm->page_table_lock);
81 if (likely(!vma->anon_vma)) {
82 vma->anon_vma = anon_vma;
83 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
86 spin_unlock(&mm->page_table_lock);
89 spin_unlock(&locked->lock);
90 if (unlikely(allocated))
91 anon_vma_free(allocated);
96 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
98 BUG_ON(vma->anon_vma != next->anon_vma);
99 list_del(&next->anon_vma_node);
102 void __anon_vma_link(struct vm_area_struct *vma)
104 struct anon_vma *anon_vma = vma->anon_vma;
107 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
110 void anon_vma_link(struct vm_area_struct *vma)
112 struct anon_vma *anon_vma = vma->anon_vma;
115 spin_lock(&anon_vma->lock);
116 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
117 spin_unlock(&anon_vma->lock);
121 void anon_vma_unlink(struct vm_area_struct *vma)
123 struct anon_vma *anon_vma = vma->anon_vma;
129 spin_lock(&anon_vma->lock);
130 list_del(&vma->anon_vma_node);
132 /* We must garbage collect the anon_vma if it's empty */
133 empty = list_empty(&anon_vma->head);
134 spin_unlock(&anon_vma->lock);
137 anon_vma_free(anon_vma);
140 static void anon_vma_ctor(struct kmem_cache *cachep, void *data)
142 struct anon_vma *anon_vma = data;
144 spin_lock_init(&anon_vma->lock);
145 INIT_LIST_HEAD(&anon_vma->head);
148 void __init anon_vma_init(void)
150 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
151 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
155 * Getting a lock on a stable anon_vma from a page off the LRU is
156 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
158 static struct anon_vma *page_lock_anon_vma(struct page *page)
160 struct anon_vma *anon_vma;
161 unsigned long anon_mapping;
164 anon_mapping = (unsigned long) page->mapping;
165 if (!(anon_mapping & PAGE_MAPPING_ANON))
167 if (!page_mapped(page))
170 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
171 spin_lock(&anon_vma->lock);
178 static void page_unlock_anon_vma(struct anon_vma *anon_vma)
180 spin_unlock(&anon_vma->lock);
185 * At what user virtual address is page expected in @vma?
186 * Returns virtual address or -EFAULT if page's index/offset is not
187 * within the range mapped the @vma.
189 static inline unsigned long
190 vma_address(struct page *page, struct vm_area_struct *vma)
192 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
193 unsigned long address;
195 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
196 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
197 /* page should be within @vma mapping range */
204 * At what user virtual address is page expected in vma? checking that the
205 * page matches the vma: currently only used on anon pages, by unuse_vma;
207 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
209 if (PageAnon(page)) {
210 if ((void *)vma->anon_vma !=
211 (void *)page->mapping - PAGE_MAPPING_ANON)
213 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
215 vma->vm_file->f_mapping != page->mapping)
219 return vma_address(page, vma);
223 * Check that @page is mapped at @address into @mm.
225 * On success returns with pte mapped and locked.
227 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
228 unsigned long address, spinlock_t **ptlp)
236 pgd = pgd_offset(mm, address);
237 if (!pgd_present(*pgd))
240 pud = pud_offset(pgd, address);
241 if (!pud_present(*pud))
244 pmd = pmd_offset(pud, address);
245 if (!pmd_present(*pmd))
248 pte = pte_offset_map(pmd, address);
249 /* Make a quick check before getting the lock */
250 if (!pte_present(*pte)) {
255 ptl = pte_lockptr(mm, pmd);
257 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
261 pte_unmap_unlock(pte, ptl);
266 * Subfunctions of page_referenced: page_referenced_one called
267 * repeatedly from either page_referenced_anon or page_referenced_file.
269 static int page_referenced_one(struct page *page,
270 struct vm_area_struct *vma, unsigned int *mapcount)
272 struct mm_struct *mm = vma->vm_mm;
273 unsigned long address;
278 address = vma_address(page, vma);
279 if (address == -EFAULT)
282 pte = page_check_address(page, mm, address, &ptl);
286 if (vma->vm_flags & VM_LOCKED) {
288 *mapcount = 1; /* break early from loop */
289 } else if (ptep_clear_flush_young(vma, address, pte))
292 /* Pretend the page is referenced if the task has the
293 swap token and is in the middle of a page fault. */
294 if (mm != current->mm && has_swap_token(mm) &&
295 rwsem_is_locked(&mm->mmap_sem))
299 pte_unmap_unlock(pte, ptl);
304 static int page_referenced_anon(struct page *page)
306 unsigned int mapcount;
307 struct anon_vma *anon_vma;
308 struct vm_area_struct *vma;
311 anon_vma = page_lock_anon_vma(page);
315 mapcount = page_mapcount(page);
316 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
317 referenced += page_referenced_one(page, vma, &mapcount);
322 page_unlock_anon_vma(anon_vma);
327 * page_referenced_file - referenced check for object-based rmap
328 * @page: the page we're checking references on.
330 * For an object-based mapped page, find all the places it is mapped and
331 * check/clear the referenced flag. This is done by following the page->mapping
332 * pointer, then walking the chain of vmas it holds. It returns the number
333 * of references it found.
335 * This function is only called from page_referenced for object-based pages.
337 static int page_referenced_file(struct page *page)
339 unsigned int mapcount;
340 struct address_space *mapping = page->mapping;
341 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
342 struct vm_area_struct *vma;
343 struct prio_tree_iter iter;
347 * The caller's checks on page->mapping and !PageAnon have made
348 * sure that this is a file page: the check for page->mapping
349 * excludes the case just before it gets set on an anon page.
351 BUG_ON(PageAnon(page));
354 * The page lock not only makes sure that page->mapping cannot
355 * suddenly be NULLified by truncation, it makes sure that the
356 * structure at mapping cannot be freed and reused yet,
357 * so we can safely take mapping->i_mmap_lock.
359 BUG_ON(!PageLocked(page));
361 spin_lock(&mapping->i_mmap_lock);
364 * i_mmap_lock does not stabilize mapcount at all, but mapcount
365 * is more likely to be accurate if we note it after spinning.
367 mapcount = page_mapcount(page);
369 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
370 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
371 == (VM_LOCKED|VM_MAYSHARE)) {
375 referenced += page_referenced_one(page, vma, &mapcount);
380 spin_unlock(&mapping->i_mmap_lock);
385 * page_referenced - test if the page was referenced
386 * @page: the page to test
387 * @is_locked: caller holds lock on the page
389 * Quick test_and_clear_referenced for all mappings to a page,
390 * returns the number of ptes which referenced the page.
392 int page_referenced(struct page *page, int is_locked)
396 if (page_test_and_clear_young(page))
399 if (TestClearPageReferenced(page))
402 if (page_mapped(page) && page->mapping) {
404 referenced += page_referenced_anon(page);
406 referenced += page_referenced_file(page);
407 else if (TestSetPageLocked(page))
411 referenced += page_referenced_file(page);
418 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
420 struct mm_struct *mm = vma->vm_mm;
421 unsigned long address;
426 address = vma_address(page, vma);
427 if (address == -EFAULT)
430 pte = page_check_address(page, mm, address, &ptl);
434 if (pte_dirty(*pte) || pte_write(*pte)) {
437 flush_cache_page(vma, address, pte_pfn(*pte));
438 entry = ptep_clear_flush(vma, address, pte);
439 entry = pte_wrprotect(entry);
440 entry = pte_mkclean(entry);
441 set_pte_at(mm, address, pte, entry);
445 pte_unmap_unlock(pte, ptl);
450 static int page_mkclean_file(struct address_space *mapping, struct page *page)
452 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
453 struct vm_area_struct *vma;
454 struct prio_tree_iter iter;
457 BUG_ON(PageAnon(page));
459 spin_lock(&mapping->i_mmap_lock);
460 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
461 if (vma->vm_flags & VM_SHARED)
462 ret += page_mkclean_one(page, vma);
464 spin_unlock(&mapping->i_mmap_lock);
468 int page_mkclean(struct page *page)
472 BUG_ON(!PageLocked(page));
474 if (page_mapped(page)) {
475 struct address_space *mapping = page_mapping(page);
477 ret = page_mkclean_file(mapping, page);
478 if (page_test_dirty(page)) {
479 page_clear_dirty(page);
487 EXPORT_SYMBOL_GPL(page_mkclean);
490 * page_set_anon_rmap - setup new anonymous rmap
491 * @page: the page to add the mapping to
492 * @vma: the vm area in which the mapping is added
493 * @address: the user virtual address mapped
495 static void __page_set_anon_rmap(struct page *page,
496 struct vm_area_struct *vma, unsigned long address)
498 struct anon_vma *anon_vma = vma->anon_vma;
501 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
502 page->mapping = (struct address_space *) anon_vma;
504 page->index = linear_page_index(vma, address);
507 * nr_mapped state can be updated without turning off
508 * interrupts because it is not modified via interrupt.
510 __inc_zone_page_state(page, NR_ANON_PAGES);
514 * page_set_anon_rmap - sanity check anonymous rmap addition
515 * @page: the page to add the mapping to
516 * @vma: the vm area in which the mapping is added
517 * @address: the user virtual address mapped
519 static void __page_check_anon_rmap(struct page *page,
520 struct vm_area_struct *vma, unsigned long address)
522 #ifdef CONFIG_DEBUG_VM
524 * The page's anon-rmap details (mapping and index) are guaranteed to
525 * be set up correctly at this point.
527 * We have exclusion against page_add_anon_rmap because the caller
528 * always holds the page locked, except if called from page_dup_rmap,
529 * in which case the page is already known to be setup.
531 * We have exclusion against page_add_new_anon_rmap because those pages
532 * are initially only visible via the pagetables, and the pte is locked
533 * over the call to page_add_new_anon_rmap.
535 struct anon_vma *anon_vma = vma->anon_vma;
536 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
537 BUG_ON(page->mapping != (struct address_space *)anon_vma);
538 BUG_ON(page->index != linear_page_index(vma, address));
543 * page_add_anon_rmap - add pte mapping to an anonymous page
544 * @page: the page to add the mapping to
545 * @vma: the vm area in which the mapping is added
546 * @address: the user virtual address mapped
548 * The caller needs to hold the pte lock and the page must be locked.
550 void page_add_anon_rmap(struct page *page,
551 struct vm_area_struct *vma, unsigned long address)
553 VM_BUG_ON(!PageLocked(page));
554 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
555 if (atomic_inc_and_test(&page->_mapcount))
556 __page_set_anon_rmap(page, vma, address);
558 __page_check_anon_rmap(page, vma, address);
562 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
563 * @page: the page to add the mapping to
564 * @vma: the vm area in which the mapping is added
565 * @address: the user virtual address mapped
567 * Same as page_add_anon_rmap but must only be called on *new* pages.
568 * This means the inc-and-test can be bypassed.
569 * Page does not have to be locked.
571 void page_add_new_anon_rmap(struct page *page,
572 struct vm_area_struct *vma, unsigned long address)
574 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
575 atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */
576 __page_set_anon_rmap(page, vma, address);
580 * page_add_file_rmap - add pte mapping to a file page
581 * @page: the page to add the mapping to
583 * The caller needs to hold the pte lock.
585 void page_add_file_rmap(struct page *page)
587 if (atomic_inc_and_test(&page->_mapcount))
588 __inc_zone_page_state(page, NR_FILE_MAPPED);
591 #ifdef CONFIG_DEBUG_VM
593 * page_dup_rmap - duplicate pte mapping to a page
594 * @page: the page to add the mapping to
596 * For copy_page_range only: minimal extract from page_add_file_rmap /
597 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
600 * The caller needs to hold the pte lock.
602 void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address)
604 BUG_ON(page_mapcount(page) == 0);
606 __page_check_anon_rmap(page, vma, address);
607 atomic_inc(&page->_mapcount);
612 * page_remove_rmap - take down pte mapping from a page
613 * @page: page to remove mapping from
615 * The caller needs to hold the pte lock.
617 void page_remove_rmap(struct page *page, struct vm_area_struct *vma)
619 if (atomic_add_negative(-1, &page->_mapcount)) {
620 if (unlikely(page_mapcount(page) < 0)) {
621 printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page));
622 printk (KERN_EMERG " page pfn = %lx\n", page_to_pfn(page));
623 printk (KERN_EMERG " page->flags = %lx\n", page->flags);
624 printk (KERN_EMERG " page->count = %x\n", page_count(page));
625 printk (KERN_EMERG " page->mapping = %p\n", page->mapping);
626 print_symbol (KERN_EMERG " vma->vm_ops = %s\n", (unsigned long)vma->vm_ops);
628 print_symbol (KERN_EMERG " vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage);
629 print_symbol (KERN_EMERG " vma->vm_ops->fault = %s\n", (unsigned long)vma->vm_ops->fault);
631 if (vma->vm_file && vma->vm_file->f_op)
632 print_symbol (KERN_EMERG " vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap);
637 * It would be tidy to reset the PageAnon mapping here,
638 * but that might overwrite a racing page_add_anon_rmap
639 * which increments mapcount after us but sets mapping
640 * before us: so leave the reset to free_hot_cold_page,
641 * and remember that it's only reliable while mapped.
642 * Leaving it set also helps swapoff to reinstate ptes
643 * faster for those pages still in swapcache.
645 if (page_test_dirty(page)) {
646 page_clear_dirty(page);
647 set_page_dirty(page);
649 __dec_zone_page_state(page,
650 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
655 * Subfunctions of try_to_unmap: try_to_unmap_one called
656 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
658 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
661 struct mm_struct *mm = vma->vm_mm;
662 unsigned long address;
666 int ret = SWAP_AGAIN;
668 address = vma_address(page, vma);
669 if (address == -EFAULT)
672 pte = page_check_address(page, mm, address, &ptl);
677 * If the page is mlock()d, we cannot swap it out.
678 * If it's recently referenced (perhaps page_referenced
679 * skipped over this mm) then we should reactivate it.
681 if (!migration && ((vma->vm_flags & VM_LOCKED) ||
682 (ptep_clear_flush_young(vma, address, pte)))) {
687 /* Nuke the page table entry. */
688 flush_cache_page(vma, address, page_to_pfn(page));
689 pteval = ptep_clear_flush(vma, address, pte);
691 /* Move the dirty bit to the physical page now the pte is gone. */
692 if (pte_dirty(pteval))
693 set_page_dirty(page);
695 /* Update high watermark before we lower rss */
696 update_hiwater_rss(mm);
698 if (PageAnon(page)) {
699 swp_entry_t entry = { .val = page_private(page) };
701 if (PageSwapCache(page)) {
703 * Store the swap location in the pte.
704 * See handle_pte_fault() ...
706 swap_duplicate(entry);
707 if (list_empty(&mm->mmlist)) {
708 spin_lock(&mmlist_lock);
709 if (list_empty(&mm->mmlist))
710 list_add(&mm->mmlist, &init_mm.mmlist);
711 spin_unlock(&mmlist_lock);
713 dec_mm_counter(mm, anon_rss);
714 #ifdef CONFIG_MIGRATION
717 * Store the pfn of the page in a special migration
718 * pte. do_swap_page() will wait until the migration
719 * pte is removed and then restart fault handling.
722 entry = make_migration_entry(page, pte_write(pteval));
725 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
726 BUG_ON(pte_file(*pte));
728 #ifdef CONFIG_MIGRATION
730 /* Establish migration entry for a file page */
732 entry = make_migration_entry(page, pte_write(pteval));
733 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
736 dec_mm_counter(mm, file_rss);
739 page_remove_rmap(page, vma);
740 page_cache_release(page);
743 pte_unmap_unlock(pte, ptl);
749 * objrmap doesn't work for nonlinear VMAs because the assumption that
750 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
751 * Consequently, given a particular page and its ->index, we cannot locate the
752 * ptes which are mapping that page without an exhaustive linear search.
754 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
755 * maps the file to which the target page belongs. The ->vm_private_data field
756 * holds the current cursor into that scan. Successive searches will circulate
757 * around the vma's virtual address space.
759 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
760 * more scanning pressure is placed against them as well. Eventually pages
761 * will become fully unmapped and are eligible for eviction.
763 * For very sparsely populated VMAs this is a little inefficient - chances are
764 * there there won't be many ptes located within the scan cluster. In this case
765 * maybe we could scan further - to the end of the pte page, perhaps.
767 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
768 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
770 static void try_to_unmap_cluster(unsigned long cursor,
771 unsigned int *mapcount, struct vm_area_struct *vma)
773 struct mm_struct *mm = vma->vm_mm;
781 unsigned long address;
784 address = (vma->vm_start + cursor) & CLUSTER_MASK;
785 end = address + CLUSTER_SIZE;
786 if (address < vma->vm_start)
787 address = vma->vm_start;
788 if (end > vma->vm_end)
791 pgd = pgd_offset(mm, address);
792 if (!pgd_present(*pgd))
795 pud = pud_offset(pgd, address);
796 if (!pud_present(*pud))
799 pmd = pmd_offset(pud, address);
800 if (!pmd_present(*pmd))
803 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
805 /* Update high watermark before we lower rss */
806 update_hiwater_rss(mm);
808 for (; address < end; pte++, address += PAGE_SIZE) {
809 if (!pte_present(*pte))
811 page = vm_normal_page(vma, address, *pte);
812 BUG_ON(!page || PageAnon(page));
814 if (ptep_clear_flush_young(vma, address, pte))
817 /* Nuke the page table entry. */
818 flush_cache_page(vma, address, pte_pfn(*pte));
819 pteval = ptep_clear_flush(vma, address, pte);
821 /* If nonlinear, store the file page offset in the pte. */
822 if (page->index != linear_page_index(vma, address))
823 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
825 /* Move the dirty bit to the physical page now the pte is gone. */
826 if (pte_dirty(pteval))
827 set_page_dirty(page);
829 page_remove_rmap(page, vma);
830 page_cache_release(page);
831 dec_mm_counter(mm, file_rss);
834 pte_unmap_unlock(pte - 1, ptl);
837 static int try_to_unmap_anon(struct page *page, int migration)
839 struct anon_vma *anon_vma;
840 struct vm_area_struct *vma;
841 int ret = SWAP_AGAIN;
843 anon_vma = page_lock_anon_vma(page);
847 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
848 ret = try_to_unmap_one(page, vma, migration);
849 if (ret == SWAP_FAIL || !page_mapped(page))
853 page_unlock_anon_vma(anon_vma);
858 * try_to_unmap_file - unmap file page using the object-based rmap method
859 * @page: the page to unmap
861 * Find all the mappings of a page using the mapping pointer and the vma chains
862 * contained in the address_space struct it points to.
864 * This function is only called from try_to_unmap for object-based pages.
866 static int try_to_unmap_file(struct page *page, int migration)
868 struct address_space *mapping = page->mapping;
869 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
870 struct vm_area_struct *vma;
871 struct prio_tree_iter iter;
872 int ret = SWAP_AGAIN;
873 unsigned long cursor;
874 unsigned long max_nl_cursor = 0;
875 unsigned long max_nl_size = 0;
876 unsigned int mapcount;
878 spin_lock(&mapping->i_mmap_lock);
879 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
880 ret = try_to_unmap_one(page, vma, migration);
881 if (ret == SWAP_FAIL || !page_mapped(page))
885 if (list_empty(&mapping->i_mmap_nonlinear))
888 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
889 shared.vm_set.list) {
890 if ((vma->vm_flags & VM_LOCKED) && !migration)
892 cursor = (unsigned long) vma->vm_private_data;
893 if (cursor > max_nl_cursor)
894 max_nl_cursor = cursor;
895 cursor = vma->vm_end - vma->vm_start;
896 if (cursor > max_nl_size)
897 max_nl_size = cursor;
900 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
906 * We don't try to search for this page in the nonlinear vmas,
907 * and page_referenced wouldn't have found it anyway. Instead
908 * just walk the nonlinear vmas trying to age and unmap some.
909 * The mapcount of the page we came in with is irrelevant,
910 * but even so use it as a guide to how hard we should try?
912 mapcount = page_mapcount(page);
915 cond_resched_lock(&mapping->i_mmap_lock);
917 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
918 if (max_nl_cursor == 0)
919 max_nl_cursor = CLUSTER_SIZE;
922 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
923 shared.vm_set.list) {
924 if ((vma->vm_flags & VM_LOCKED) && !migration)
926 cursor = (unsigned long) vma->vm_private_data;
927 while ( cursor < max_nl_cursor &&
928 cursor < vma->vm_end - vma->vm_start) {
929 try_to_unmap_cluster(cursor, &mapcount, vma);
930 cursor += CLUSTER_SIZE;
931 vma->vm_private_data = (void *) cursor;
932 if ((int)mapcount <= 0)
935 vma->vm_private_data = (void *) max_nl_cursor;
937 cond_resched_lock(&mapping->i_mmap_lock);
938 max_nl_cursor += CLUSTER_SIZE;
939 } while (max_nl_cursor <= max_nl_size);
942 * Don't loop forever (perhaps all the remaining pages are
943 * in locked vmas). Reset cursor on all unreserved nonlinear
944 * vmas, now forgetting on which ones it had fallen behind.
946 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
947 vma->vm_private_data = NULL;
949 spin_unlock(&mapping->i_mmap_lock);
954 * try_to_unmap - try to remove all page table mappings to a page
955 * @page: the page to get unmapped
957 * Tries to remove all the page table entries which are mapping this
958 * page, used in the pageout path. Caller must hold the page lock.
961 * SWAP_SUCCESS - we succeeded in removing all mappings
962 * SWAP_AGAIN - we missed a mapping, try again later
963 * SWAP_FAIL - the page is unswappable
965 int try_to_unmap(struct page *page, int migration)
969 BUG_ON(!PageLocked(page));
972 ret = try_to_unmap_anon(page, migration);
974 ret = try_to_unmap_file(page, migration);
976 if (!page_mapped(page))