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 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)
40 * (code doesn't rely on that order so it could be switched around)
42 * anon_vma->lock (memory_failure, collect_procs_anon)
47 #include <linux/pagemap.h>
48 #include <linux/swap.h>
49 #include <linux/swapops.h>
50 #include <linux/slab.h>
51 #include <linux/init.h>
52 #include <linux/ksm.h>
53 #include <linux/rmap.h>
54 #include <linux/rcupdate.h>
55 #include <linux/module.h>
56 #include <linux/memcontrol.h>
57 #include <linux/mmu_notifier.h>
58 #include <linux/migrate.h>
60 #include <asm/tlbflush.h>
64 static struct kmem_cache *anon_vma_cachep;
66 static inline struct anon_vma *anon_vma_alloc(void)
68 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
71 static inline void anon_vma_free(struct anon_vma *anon_vma)
73 kmem_cache_free(anon_vma_cachep, anon_vma);
77 * anon_vma_prepare - attach an anon_vma to a memory region
78 * @vma: the memory region in question
80 * This makes sure the memory mapping described by 'vma' has
81 * an 'anon_vma' attached to it, so that we can associate the
82 * anonymous pages mapped into it with that anon_vma.
84 * The common case will be that we already have one, but if
85 * if not we either need to find an adjacent mapping that we
86 * can re-use the anon_vma from (very common when the only
87 * reason for splitting a vma has been mprotect()), or we
90 * Anon-vma allocations are very subtle, because we may have
91 * optimistically looked up an anon_vma in page_lock_anon_vma()
92 * and that may actually touch the spinlock even in the newly
93 * allocated vma (it depends on RCU to make sure that the
94 * anon_vma isn't actually destroyed).
96 * As a result, we need to do proper anon_vma locking even
97 * for the new allocation. At the same time, we do not want
98 * to do any locking for the common case of already having
101 * This must be called with the mmap_sem held for reading.
103 int anon_vma_prepare(struct vm_area_struct *vma)
105 struct anon_vma *anon_vma = vma->anon_vma;
108 if (unlikely(!anon_vma)) {
109 struct mm_struct *mm = vma->vm_mm;
110 struct anon_vma *allocated;
112 anon_vma = find_mergeable_anon_vma(vma);
115 anon_vma = anon_vma_alloc();
116 if (unlikely(!anon_vma))
118 allocated = anon_vma;
120 spin_lock(&anon_vma->lock);
122 /* page_table_lock to protect against threads */
123 spin_lock(&mm->page_table_lock);
124 if (likely(!vma->anon_vma)) {
125 vma->anon_vma = anon_vma;
126 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
129 spin_unlock(&mm->page_table_lock);
131 spin_unlock(&anon_vma->lock);
132 if (unlikely(allocated))
133 anon_vma_free(allocated);
138 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
140 BUG_ON(vma->anon_vma != next->anon_vma);
141 list_del(&next->anon_vma_node);
144 void __anon_vma_link(struct vm_area_struct *vma)
146 struct anon_vma *anon_vma = vma->anon_vma;
149 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
152 void anon_vma_link(struct vm_area_struct *vma)
154 struct anon_vma *anon_vma = vma->anon_vma;
157 spin_lock(&anon_vma->lock);
158 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
159 spin_unlock(&anon_vma->lock);
163 void anon_vma_unlink(struct vm_area_struct *vma)
165 struct anon_vma *anon_vma = vma->anon_vma;
171 spin_lock(&anon_vma->lock);
172 list_del(&vma->anon_vma_node);
174 /* We must garbage collect the anon_vma if it's empty */
175 empty = list_empty(&anon_vma->head);
176 spin_unlock(&anon_vma->lock);
179 anon_vma_free(anon_vma);
182 static void anon_vma_ctor(void *data)
184 struct anon_vma *anon_vma = data;
186 spin_lock_init(&anon_vma->lock);
187 INIT_LIST_HEAD(&anon_vma->head);
190 void __init anon_vma_init(void)
192 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
193 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
197 * Getting a lock on a stable anon_vma from a page off the LRU is
198 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
200 struct anon_vma *page_lock_anon_vma(struct page *page)
202 struct anon_vma *anon_vma;
203 unsigned long anon_mapping;
206 anon_mapping = (unsigned long) page->mapping;
207 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
209 if (!page_mapped(page))
212 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
213 spin_lock(&anon_vma->lock);
220 void page_unlock_anon_vma(struct anon_vma *anon_vma)
222 spin_unlock(&anon_vma->lock);
227 * At what user virtual address is page expected in @vma?
228 * Returns virtual address or -EFAULT if page's index/offset is not
229 * within the range mapped the @vma.
231 static inline unsigned long
232 vma_address(struct page *page, struct vm_area_struct *vma)
234 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
235 unsigned long address;
237 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
238 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
239 /* page should be within @vma mapping range */
246 * At what user virtual address is page expected in vma?
247 * checking that the page matches the vma.
249 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
251 if (PageAnon(page)) {
252 if (vma->anon_vma != page_anon_vma(page))
254 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
256 vma->vm_file->f_mapping != page->mapping)
260 return vma_address(page, vma);
264 * Check that @page is mapped at @address into @mm.
266 * If @sync is false, page_check_address may perform a racy check to avoid
267 * the page table lock when the pte is not present (helpful when reclaiming
268 * highly shared pages).
270 * On success returns with pte mapped and locked.
272 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
273 unsigned long address, spinlock_t **ptlp, int sync)
281 pgd = pgd_offset(mm, address);
282 if (!pgd_present(*pgd))
285 pud = pud_offset(pgd, address);
286 if (!pud_present(*pud))
289 pmd = pmd_offset(pud, address);
290 if (!pmd_present(*pmd))
293 pte = pte_offset_map(pmd, address);
294 /* Make a quick check before getting the lock */
295 if (!sync && !pte_present(*pte)) {
300 ptl = pte_lockptr(mm, pmd);
302 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
306 pte_unmap_unlock(pte, ptl);
311 * page_mapped_in_vma - check whether a page is really mapped in a VMA
312 * @page: the page to test
313 * @vma: the VMA to test
315 * Returns 1 if the page is mapped into the page tables of the VMA, 0
316 * if the page is not mapped into the page tables of this VMA. Only
317 * valid for normal file or anonymous VMAs.
319 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
321 unsigned long address;
325 address = vma_address(page, vma);
326 if (address == -EFAULT) /* out of vma range */
328 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
329 if (!pte) /* the page is not in this mm */
331 pte_unmap_unlock(pte, ptl);
337 * Subfunctions of page_referenced: page_referenced_one called
338 * repeatedly from either page_referenced_anon or page_referenced_file.
340 int page_referenced_one(struct page *page, struct vm_area_struct *vma,
341 unsigned long address, unsigned int *mapcount,
342 unsigned long *vm_flags)
344 struct mm_struct *mm = vma->vm_mm;
349 pte = page_check_address(page, mm, address, &ptl, 0);
354 * Don't want to elevate referenced for mlocked page that gets this far,
355 * in order that it progresses to try_to_unmap and is moved to the
358 if (vma->vm_flags & VM_LOCKED) {
359 *mapcount = 1; /* break early from loop */
360 *vm_flags |= VM_LOCKED;
364 if (ptep_clear_flush_young_notify(vma, address, pte)) {
366 * Don't treat a reference through a sequentially read
367 * mapping as such. If the page has been used in
368 * another mapping, we will catch it; if this other
369 * mapping is already gone, the unmap path will have
370 * set PG_referenced or activated the page.
372 if (likely(!VM_SequentialReadHint(vma)))
376 /* Pretend the page is referenced if the task has the
377 swap token and is in the middle of a page fault. */
378 if (mm != current->mm && has_swap_token(mm) &&
379 rwsem_is_locked(&mm->mmap_sem))
384 pte_unmap_unlock(pte, ptl);
387 *vm_flags |= vma->vm_flags;
392 static int page_referenced_anon(struct page *page,
393 struct mem_cgroup *mem_cont,
394 unsigned long *vm_flags)
396 unsigned int mapcount;
397 struct anon_vma *anon_vma;
398 struct vm_area_struct *vma;
401 anon_vma = page_lock_anon_vma(page);
405 mapcount = page_mapcount(page);
406 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
407 unsigned long address = vma_address(page, vma);
408 if (address == -EFAULT)
411 * If we are reclaiming on behalf of a cgroup, skip
412 * counting on behalf of references from different
415 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
417 referenced += page_referenced_one(page, vma, address,
418 &mapcount, vm_flags);
423 page_unlock_anon_vma(anon_vma);
428 * page_referenced_file - referenced check for object-based rmap
429 * @page: the page we're checking references on.
430 * @mem_cont: target memory controller
431 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
433 * For an object-based mapped page, find all the places it is mapped and
434 * check/clear the referenced flag. This is done by following the page->mapping
435 * pointer, then walking the chain of vmas it holds. It returns the number
436 * of references it found.
438 * This function is only called from page_referenced for object-based pages.
440 static int page_referenced_file(struct page *page,
441 struct mem_cgroup *mem_cont,
442 unsigned long *vm_flags)
444 unsigned int mapcount;
445 struct address_space *mapping = page->mapping;
446 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
447 struct vm_area_struct *vma;
448 struct prio_tree_iter iter;
452 * The caller's checks on page->mapping and !PageAnon have made
453 * sure that this is a file page: the check for page->mapping
454 * excludes the case just before it gets set on an anon page.
456 BUG_ON(PageAnon(page));
459 * The page lock not only makes sure that page->mapping cannot
460 * suddenly be NULLified by truncation, it makes sure that the
461 * structure at mapping cannot be freed and reused yet,
462 * so we can safely take mapping->i_mmap_lock.
464 BUG_ON(!PageLocked(page));
466 spin_lock(&mapping->i_mmap_lock);
469 * i_mmap_lock does not stabilize mapcount at all, but mapcount
470 * is more likely to be accurate if we note it after spinning.
472 mapcount = page_mapcount(page);
474 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
475 unsigned long address = vma_address(page, vma);
476 if (address == -EFAULT)
479 * If we are reclaiming on behalf of a cgroup, skip
480 * counting on behalf of references from different
483 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
485 referenced += page_referenced_one(page, vma, address,
486 &mapcount, vm_flags);
491 spin_unlock(&mapping->i_mmap_lock);
496 * page_referenced - test if the page was referenced
497 * @page: the page to test
498 * @is_locked: caller holds lock on the page
499 * @mem_cont: target memory controller
500 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
502 * Quick test_and_clear_referenced for all mappings to a page,
503 * returns the number of ptes which referenced the page.
505 int page_referenced(struct page *page,
507 struct mem_cgroup *mem_cont,
508 unsigned long *vm_flags)
513 if (TestClearPageReferenced(page))
517 if (page_mapped(page) && page_rmapping(page)) {
518 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
519 we_locked = trylock_page(page);
525 if (unlikely(PageKsm(page)))
526 referenced += page_referenced_ksm(page, mem_cont,
528 else if (PageAnon(page))
529 referenced += page_referenced_anon(page, mem_cont,
531 else if (page->mapping)
532 referenced += page_referenced_file(page, mem_cont,
538 if (page_test_and_clear_young(page))
544 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
545 unsigned long address)
547 struct mm_struct *mm = vma->vm_mm;
552 pte = page_check_address(page, mm, address, &ptl, 1);
556 if (pte_dirty(*pte) || pte_write(*pte)) {
559 flush_cache_page(vma, address, pte_pfn(*pte));
560 entry = ptep_clear_flush_notify(vma, address, pte);
561 entry = pte_wrprotect(entry);
562 entry = pte_mkclean(entry);
563 set_pte_at(mm, address, pte, entry);
567 pte_unmap_unlock(pte, ptl);
572 static int page_mkclean_file(struct address_space *mapping, struct page *page)
574 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
575 struct vm_area_struct *vma;
576 struct prio_tree_iter iter;
579 BUG_ON(PageAnon(page));
581 spin_lock(&mapping->i_mmap_lock);
582 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
583 if (vma->vm_flags & VM_SHARED) {
584 unsigned long address = vma_address(page, vma);
585 if (address == -EFAULT)
587 ret += page_mkclean_one(page, vma, address);
590 spin_unlock(&mapping->i_mmap_lock);
594 int page_mkclean(struct page *page)
598 BUG_ON(!PageLocked(page));
600 if (page_mapped(page)) {
601 struct address_space *mapping = page_mapping(page);
603 ret = page_mkclean_file(mapping, page);
604 if (page_test_dirty(page)) {
605 page_clear_dirty(page);
613 EXPORT_SYMBOL_GPL(page_mkclean);
616 * __page_set_anon_rmap - setup new anonymous rmap
617 * @page: the page to add the mapping to
618 * @vma: the vm area in which the mapping is added
619 * @address: the user virtual address mapped
621 static void __page_set_anon_rmap(struct page *page,
622 struct vm_area_struct *vma, unsigned long address)
624 struct anon_vma *anon_vma = vma->anon_vma;
627 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
628 page->mapping = (struct address_space *) anon_vma;
629 page->index = linear_page_index(vma, address);
633 * __page_check_anon_rmap - sanity check anonymous rmap addition
634 * @page: the page to add the mapping to
635 * @vma: the vm area in which the mapping is added
636 * @address: the user virtual address mapped
638 static void __page_check_anon_rmap(struct page *page,
639 struct vm_area_struct *vma, unsigned long address)
641 #ifdef CONFIG_DEBUG_VM
643 * The page's anon-rmap details (mapping and index) are guaranteed to
644 * be set up correctly at this point.
646 * We have exclusion against page_add_anon_rmap because the caller
647 * always holds the page locked, except if called from page_dup_rmap,
648 * in which case the page is already known to be setup.
650 * We have exclusion against page_add_new_anon_rmap because those pages
651 * are initially only visible via the pagetables, and the pte is locked
652 * over the call to page_add_new_anon_rmap.
654 struct anon_vma *anon_vma = vma->anon_vma;
655 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
656 BUG_ON(page->mapping != (struct address_space *)anon_vma);
657 BUG_ON(page->index != linear_page_index(vma, address));
662 * page_add_anon_rmap - add pte mapping to an anonymous page
663 * @page: the page to add the mapping to
664 * @vma: the vm area in which the mapping is added
665 * @address: the user virtual address mapped
667 * The caller needs to hold the pte lock, and the page must be locked in
668 * the anon_vma case: to serialize mapping,index checking after setting.
670 void page_add_anon_rmap(struct page *page,
671 struct vm_area_struct *vma, unsigned long address)
673 int first = atomic_inc_and_test(&page->_mapcount);
675 __inc_zone_page_state(page, NR_ANON_PAGES);
676 if (unlikely(PageKsm(page)))
679 VM_BUG_ON(!PageLocked(page));
680 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
682 __page_set_anon_rmap(page, vma, address);
684 __page_check_anon_rmap(page, vma, address);
688 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
689 * @page: the page to add the mapping to
690 * @vma: the vm area in which the mapping is added
691 * @address: the user virtual address mapped
693 * Same as page_add_anon_rmap but must only be called on *new* pages.
694 * This means the inc-and-test can be bypassed.
695 * Page does not have to be locked.
697 void page_add_new_anon_rmap(struct page *page,
698 struct vm_area_struct *vma, unsigned long address)
700 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
701 SetPageSwapBacked(page);
702 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
703 __inc_zone_page_state(page, NR_ANON_PAGES);
704 __page_set_anon_rmap(page, vma, address);
705 if (page_evictable(page, vma))
706 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
708 add_page_to_unevictable_list(page);
712 * page_add_file_rmap - add pte mapping to a file page
713 * @page: the page to add the mapping to
715 * The caller needs to hold the pte lock.
717 void page_add_file_rmap(struct page *page)
719 if (atomic_inc_and_test(&page->_mapcount)) {
720 __inc_zone_page_state(page, NR_FILE_MAPPED);
721 mem_cgroup_update_mapped_file_stat(page, 1);
726 * page_remove_rmap - take down pte mapping from a page
727 * @page: page to remove mapping from
729 * The caller needs to hold the pte lock.
731 void page_remove_rmap(struct page *page)
733 /* page still mapped by someone else? */
734 if (!atomic_add_negative(-1, &page->_mapcount))
738 * Now that the last pte has gone, s390 must transfer dirty
739 * flag from storage key to struct page. We can usually skip
740 * this if the page is anon, so about to be freed; but perhaps
741 * not if it's in swapcache - there might be another pte slot
742 * containing the swap entry, but page not yet written to swap.
744 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
745 page_clear_dirty(page);
746 set_page_dirty(page);
748 if (PageAnon(page)) {
749 mem_cgroup_uncharge_page(page);
750 __dec_zone_page_state(page, NR_ANON_PAGES);
752 __dec_zone_page_state(page, NR_FILE_MAPPED);
754 mem_cgroup_update_mapped_file_stat(page, -1);
756 * It would be tidy to reset the PageAnon mapping here,
757 * but that might overwrite a racing page_add_anon_rmap
758 * which increments mapcount after us but sets mapping
759 * before us: so leave the reset to free_hot_cold_page,
760 * and remember that it's only reliable while mapped.
761 * Leaving it set also helps swapoff to reinstate ptes
762 * faster for those pages still in swapcache.
767 * Subfunctions of try_to_unmap: try_to_unmap_one called
768 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
770 int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
771 unsigned long address, enum ttu_flags flags)
773 struct mm_struct *mm = vma->vm_mm;
777 int ret = SWAP_AGAIN;
779 pte = page_check_address(page, mm, address, &ptl, 0);
784 * If the page is mlock()d, we cannot swap it out.
785 * If it's recently referenced (perhaps page_referenced
786 * skipped over this mm) then we should reactivate it.
788 if (!(flags & TTU_IGNORE_MLOCK)) {
789 if (vma->vm_flags & VM_LOCKED) {
793 if (TTU_ACTION(flags) == TTU_MUNLOCK)
796 if (!(flags & TTU_IGNORE_ACCESS)) {
797 if (ptep_clear_flush_young_notify(vma, address, pte)) {
803 /* Nuke the page table entry. */
804 flush_cache_page(vma, address, page_to_pfn(page));
805 pteval = ptep_clear_flush_notify(vma, address, pte);
807 /* Move the dirty bit to the physical page now the pte is gone. */
808 if (pte_dirty(pteval))
809 set_page_dirty(page);
811 /* Update high watermark before we lower rss */
812 update_hiwater_rss(mm);
814 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
816 dec_mm_counter(mm, anon_rss);
818 dec_mm_counter(mm, file_rss);
819 set_pte_at(mm, address, pte,
820 swp_entry_to_pte(make_hwpoison_entry(page)));
821 } else if (PageAnon(page)) {
822 swp_entry_t entry = { .val = page_private(page) };
824 if (PageSwapCache(page)) {
826 * Store the swap location in the pte.
827 * See handle_pte_fault() ...
829 if (swap_duplicate(entry) < 0) {
830 set_pte_at(mm, address, pte, pteval);
834 if (list_empty(&mm->mmlist)) {
835 spin_lock(&mmlist_lock);
836 if (list_empty(&mm->mmlist))
837 list_add(&mm->mmlist, &init_mm.mmlist);
838 spin_unlock(&mmlist_lock);
840 dec_mm_counter(mm, anon_rss);
841 } else if (PAGE_MIGRATION) {
843 * Store the pfn of the page in a special migration
844 * pte. do_swap_page() will wait until the migration
845 * pte is removed and then restart fault handling.
847 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
848 entry = make_migration_entry(page, pte_write(pteval));
850 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
851 BUG_ON(pte_file(*pte));
852 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
853 /* Establish migration entry for a file page */
855 entry = make_migration_entry(page, pte_write(pteval));
856 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
858 dec_mm_counter(mm, file_rss);
860 page_remove_rmap(page);
861 page_cache_release(page);
864 pte_unmap_unlock(pte, ptl);
866 if (ret == SWAP_MLOCK) {
868 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
869 if (vma->vm_flags & VM_LOCKED) {
870 mlock_vma_page(page);
873 up_read(&vma->vm_mm->mmap_sem);
881 * objrmap doesn't work for nonlinear VMAs because the assumption that
882 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
883 * Consequently, given a particular page and its ->index, we cannot locate the
884 * ptes which are mapping that page without an exhaustive linear search.
886 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
887 * maps the file to which the target page belongs. The ->vm_private_data field
888 * holds the current cursor into that scan. Successive searches will circulate
889 * around the vma's virtual address space.
891 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
892 * more scanning pressure is placed against them as well. Eventually pages
893 * will become fully unmapped and are eligible for eviction.
895 * For very sparsely populated VMAs this is a little inefficient - chances are
896 * there there won't be many ptes located within the scan cluster. In this case
897 * maybe we could scan further - to the end of the pte page, perhaps.
899 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
900 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
901 * rather than unmapping them. If we encounter the "check_page" that vmscan is
902 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
904 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
905 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
907 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
908 struct vm_area_struct *vma, struct page *check_page)
910 struct mm_struct *mm = vma->vm_mm;
918 unsigned long address;
920 int ret = SWAP_AGAIN;
923 address = (vma->vm_start + cursor) & CLUSTER_MASK;
924 end = address + CLUSTER_SIZE;
925 if (address < vma->vm_start)
926 address = vma->vm_start;
927 if (end > vma->vm_end)
930 pgd = pgd_offset(mm, address);
931 if (!pgd_present(*pgd))
934 pud = pud_offset(pgd, address);
935 if (!pud_present(*pud))
938 pmd = pmd_offset(pud, address);
939 if (!pmd_present(*pmd))
943 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
944 * keep the sem while scanning the cluster for mlocking pages.
946 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
947 locked_vma = (vma->vm_flags & VM_LOCKED);
949 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
952 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
954 /* Update high watermark before we lower rss */
955 update_hiwater_rss(mm);
957 for (; address < end; pte++, address += PAGE_SIZE) {
958 if (!pte_present(*pte))
960 page = vm_normal_page(vma, address, *pte);
961 BUG_ON(!page || PageAnon(page));
964 mlock_vma_page(page); /* no-op if already mlocked */
965 if (page == check_page)
967 continue; /* don't unmap */
970 if (ptep_clear_flush_young_notify(vma, address, pte))
973 /* Nuke the page table entry. */
974 flush_cache_page(vma, address, pte_pfn(*pte));
975 pteval = ptep_clear_flush_notify(vma, address, pte);
977 /* If nonlinear, store the file page offset in the pte. */
978 if (page->index != linear_page_index(vma, address))
979 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
981 /* Move the dirty bit to the physical page now the pte is gone. */
982 if (pte_dirty(pteval))
983 set_page_dirty(page);
985 page_remove_rmap(page);
986 page_cache_release(page);
987 dec_mm_counter(mm, file_rss);
990 pte_unmap_unlock(pte - 1, ptl);
992 up_read(&vma->vm_mm->mmap_sem);
997 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
999 * @page: the page to unmap/unlock
1000 * @flags: action and flags
1002 * Find all the mappings of a page using the mapping pointer and the vma chains
1003 * contained in the anon_vma struct it points to.
1005 * This function is only called from try_to_unmap/try_to_munlock for
1007 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1008 * where the page was found will be held for write. So, we won't recheck
1009 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1012 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1014 struct anon_vma *anon_vma;
1015 struct vm_area_struct *vma;
1016 int ret = SWAP_AGAIN;
1018 anon_vma = page_lock_anon_vma(page);
1022 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1023 unsigned long address = vma_address(page, vma);
1024 if (address == -EFAULT)
1026 ret = try_to_unmap_one(page, vma, address, flags);
1027 if (ret != SWAP_AGAIN || !page_mapped(page))
1031 page_unlock_anon_vma(anon_vma);
1036 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1037 * @page: the page to unmap/unlock
1038 * @flags: action and flags
1040 * Find all the mappings of a page using the mapping pointer and the vma chains
1041 * contained in the address_space struct it points to.
1043 * This function is only called from try_to_unmap/try_to_munlock for
1044 * object-based pages.
1045 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1046 * where the page was found will be held for write. So, we won't recheck
1047 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1050 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1052 struct address_space *mapping = page->mapping;
1053 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1054 struct vm_area_struct *vma;
1055 struct prio_tree_iter iter;
1056 int ret = SWAP_AGAIN;
1057 unsigned long cursor;
1058 unsigned long max_nl_cursor = 0;
1059 unsigned long max_nl_size = 0;
1060 unsigned int mapcount;
1062 spin_lock(&mapping->i_mmap_lock);
1063 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1064 unsigned long address = vma_address(page, vma);
1065 if (address == -EFAULT)
1067 ret = try_to_unmap_one(page, vma, address, flags);
1068 if (ret != SWAP_AGAIN || !page_mapped(page))
1072 if (list_empty(&mapping->i_mmap_nonlinear))
1076 * We don't bother to try to find the munlocked page in nonlinears.
1077 * It's costly. Instead, later, page reclaim logic may call
1078 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1080 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1083 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1084 shared.vm_set.list) {
1085 cursor = (unsigned long) vma->vm_private_data;
1086 if (cursor > max_nl_cursor)
1087 max_nl_cursor = cursor;
1088 cursor = vma->vm_end - vma->vm_start;
1089 if (cursor > max_nl_size)
1090 max_nl_size = cursor;
1093 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1099 * We don't try to search for this page in the nonlinear vmas,
1100 * and page_referenced wouldn't have found it anyway. Instead
1101 * just walk the nonlinear vmas trying to age and unmap some.
1102 * The mapcount of the page we came in with is irrelevant,
1103 * but even so use it as a guide to how hard we should try?
1105 mapcount = page_mapcount(page);
1108 cond_resched_lock(&mapping->i_mmap_lock);
1110 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1111 if (max_nl_cursor == 0)
1112 max_nl_cursor = CLUSTER_SIZE;
1115 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1116 shared.vm_set.list) {
1117 cursor = (unsigned long) vma->vm_private_data;
1118 while ( cursor < max_nl_cursor &&
1119 cursor < vma->vm_end - vma->vm_start) {
1120 if (try_to_unmap_cluster(cursor, &mapcount,
1121 vma, page) == SWAP_MLOCK)
1123 cursor += CLUSTER_SIZE;
1124 vma->vm_private_data = (void *) cursor;
1125 if ((int)mapcount <= 0)
1128 vma->vm_private_data = (void *) max_nl_cursor;
1130 cond_resched_lock(&mapping->i_mmap_lock);
1131 max_nl_cursor += CLUSTER_SIZE;
1132 } while (max_nl_cursor <= max_nl_size);
1135 * Don't loop forever (perhaps all the remaining pages are
1136 * in locked vmas). Reset cursor on all unreserved nonlinear
1137 * vmas, now forgetting on which ones it had fallen behind.
1139 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1140 vma->vm_private_data = NULL;
1142 spin_unlock(&mapping->i_mmap_lock);
1147 * try_to_unmap - try to remove all page table mappings to a page
1148 * @page: the page to get unmapped
1149 * @flags: action and flags
1151 * Tries to remove all the page table entries which are mapping this
1152 * page, used in the pageout path. Caller must hold the page lock.
1153 * Return values are:
1155 * SWAP_SUCCESS - we succeeded in removing all mappings
1156 * SWAP_AGAIN - we missed a mapping, try again later
1157 * SWAP_FAIL - the page is unswappable
1158 * SWAP_MLOCK - page is mlocked.
1160 int try_to_unmap(struct page *page, enum ttu_flags flags)
1164 BUG_ON(!PageLocked(page));
1166 if (unlikely(PageKsm(page)))
1167 ret = try_to_unmap_ksm(page, flags);
1168 else if (PageAnon(page))
1169 ret = try_to_unmap_anon(page, flags);
1171 ret = try_to_unmap_file(page, flags);
1172 if (ret != SWAP_MLOCK && !page_mapped(page))
1178 * try_to_munlock - try to munlock a page
1179 * @page: the page to be munlocked
1181 * Called from munlock code. Checks all of the VMAs mapping the page
1182 * to make sure nobody else has this page mlocked. The page will be
1183 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1185 * Return values are:
1187 * SWAP_AGAIN - no vma is holding page mlocked, or,
1188 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1189 * SWAP_FAIL - page cannot be located at present
1190 * SWAP_MLOCK - page is now mlocked.
1192 int try_to_munlock(struct page *page)
1194 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1196 if (unlikely(PageKsm(page)))
1197 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1198 else if (PageAnon(page))
1199 return try_to_unmap_anon(page, TTU_MUNLOCK);
1201 return try_to_unmap_file(page, TTU_MUNLOCK);