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)
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/memcontrol.h>
51 #include <linux/mmu_notifier.h>
52 #include <linux/migrate.h>
54 #include <asm/tlbflush.h>
58 static struct kmem_cache *anon_vma_cachep;
60 static inline struct anon_vma *anon_vma_alloc(void)
62 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
65 static inline void anon_vma_free(struct anon_vma *anon_vma)
67 kmem_cache_free(anon_vma_cachep, anon_vma);
71 * anon_vma_prepare - attach an anon_vma to a memory region
72 * @vma: the memory region in question
74 * This makes sure the memory mapping described by 'vma' has
75 * an 'anon_vma' attached to it, so that we can associate the
76 * anonymous pages mapped into it with that anon_vma.
78 * The common case will be that we already have one, but if
79 * if not we either need to find an adjacent mapping that we
80 * can re-use the anon_vma from (very common when the only
81 * reason for splitting a vma has been mprotect()), or we
84 * Anon-vma allocations are very subtle, because we may have
85 * optimistically looked up an anon_vma in page_lock_anon_vma()
86 * and that may actually touch the spinlock even in the newly
87 * allocated vma (it depends on RCU to make sure that the
88 * anon_vma isn't actually destroyed).
90 * As a result, we need to do proper anon_vma locking even
91 * for the new allocation. At the same time, we do not want
92 * to do any locking for the common case of already having
95 * This must be called with the mmap_sem held for reading.
97 int anon_vma_prepare(struct vm_area_struct *vma)
99 struct anon_vma *anon_vma = vma->anon_vma;
102 if (unlikely(!anon_vma)) {
103 struct mm_struct *mm = vma->vm_mm;
104 struct anon_vma *allocated;
106 anon_vma = find_mergeable_anon_vma(vma);
109 anon_vma = anon_vma_alloc();
110 if (unlikely(!anon_vma))
112 allocated = anon_vma;
114 spin_lock(&anon_vma->lock);
116 /* page_table_lock to protect against threads */
117 spin_lock(&mm->page_table_lock);
118 if (likely(!vma->anon_vma)) {
119 vma->anon_vma = anon_vma;
120 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
123 spin_unlock(&mm->page_table_lock);
125 spin_unlock(&anon_vma->lock);
126 if (unlikely(allocated))
127 anon_vma_free(allocated);
132 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
134 BUG_ON(vma->anon_vma != next->anon_vma);
135 list_del(&next->anon_vma_node);
138 void __anon_vma_link(struct vm_area_struct *vma)
140 struct anon_vma *anon_vma = vma->anon_vma;
143 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
146 void anon_vma_link(struct vm_area_struct *vma)
148 struct anon_vma *anon_vma = vma->anon_vma;
151 spin_lock(&anon_vma->lock);
152 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
153 spin_unlock(&anon_vma->lock);
157 void anon_vma_unlink(struct vm_area_struct *vma)
159 struct anon_vma *anon_vma = vma->anon_vma;
165 spin_lock(&anon_vma->lock);
166 list_del(&vma->anon_vma_node);
168 /* We must garbage collect the anon_vma if it's empty */
169 empty = list_empty(&anon_vma->head);
170 spin_unlock(&anon_vma->lock);
173 anon_vma_free(anon_vma);
176 static void anon_vma_ctor(void *data)
178 struct anon_vma *anon_vma = data;
180 spin_lock_init(&anon_vma->lock);
181 INIT_LIST_HEAD(&anon_vma->head);
184 void __init anon_vma_init(void)
186 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
187 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
191 * Getting a lock on a stable anon_vma from a page off the LRU is
192 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
194 struct anon_vma *page_lock_anon_vma(struct page *page)
196 struct anon_vma *anon_vma;
197 unsigned long anon_mapping;
200 anon_mapping = (unsigned long) page->mapping;
201 if (!(anon_mapping & PAGE_MAPPING_ANON))
203 if (!page_mapped(page))
206 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
207 spin_lock(&anon_vma->lock);
214 void page_unlock_anon_vma(struct anon_vma *anon_vma)
216 spin_unlock(&anon_vma->lock);
221 * At what user virtual address is page expected in @vma?
222 * Returns virtual address or -EFAULT if page's index/offset is not
223 * within the range mapped the @vma.
225 static inline unsigned long
226 vma_address(struct page *page, struct vm_area_struct *vma)
228 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
229 unsigned long address;
231 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
232 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
233 /* page should be within @vma mapping range */
240 * At what user virtual address is page expected in vma? checking that the
241 * page matches the vma: currently only used on anon pages, by unuse_vma;
243 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
245 if (PageAnon(page)) {
246 if ((void *)vma->anon_vma !=
247 (void *)page->mapping - PAGE_MAPPING_ANON)
249 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
251 vma->vm_file->f_mapping != page->mapping)
255 return vma_address(page, vma);
259 * Check that @page is mapped at @address into @mm.
261 * If @sync is false, page_check_address may perform a racy check to avoid
262 * the page table lock when the pte is not present (helpful when reclaiming
263 * highly shared pages).
265 * On success returns with pte mapped and locked.
267 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
268 unsigned long address, spinlock_t **ptlp, int sync)
276 pgd = pgd_offset(mm, address);
277 if (!pgd_present(*pgd))
280 pud = pud_offset(pgd, address);
281 if (!pud_present(*pud))
284 pmd = pmd_offset(pud, address);
285 if (!pmd_present(*pmd))
288 pte = pte_offset_map(pmd, address);
289 /* Make a quick check before getting the lock */
290 if (!sync && !pte_present(*pte)) {
295 ptl = pte_lockptr(mm, pmd);
297 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
301 pte_unmap_unlock(pte, ptl);
306 * page_mapped_in_vma - check whether a page is really mapped in a VMA
307 * @page: the page to test
308 * @vma: the VMA to test
310 * Returns 1 if the page is mapped into the page tables of the VMA, 0
311 * if the page is not mapped into the page tables of this VMA. Only
312 * valid for normal file or anonymous VMAs.
314 static int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
316 unsigned long address;
320 address = vma_address(page, vma);
321 if (address == -EFAULT) /* out of vma range */
323 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
324 if (!pte) /* the page is not in this mm */
326 pte_unmap_unlock(pte, ptl);
332 * Subfunctions of page_referenced: page_referenced_one called
333 * repeatedly from either page_referenced_anon or page_referenced_file.
335 static int page_referenced_one(struct page *page,
336 struct vm_area_struct *vma,
337 unsigned int *mapcount,
338 unsigned long *vm_flags)
340 struct mm_struct *mm = vma->vm_mm;
341 unsigned long address;
346 address = vma_address(page, vma);
347 if (address == -EFAULT)
350 pte = page_check_address(page, mm, address, &ptl, 0);
355 * Don't want to elevate referenced for mlocked page that gets this far,
356 * in order that it progresses to try_to_unmap and is moved to the
359 if (vma->vm_flags & VM_LOCKED) {
360 *mapcount = 1; /* break early from loop */
361 *vm_flags |= VM_LOCKED;
365 if (ptep_clear_flush_young_notify(vma, address, pte)) {
367 * Don't treat a reference through a sequentially read
368 * mapping as such. If the page has been used in
369 * another mapping, we will catch it; if this other
370 * mapping is already gone, the unmap path will have
371 * set PG_referenced or activated the page.
373 if (likely(!VM_SequentialReadHint(vma)))
377 /* Pretend the page is referenced if the task has the
378 swap token and is in the middle of a page fault. */
379 if (mm != current->mm && has_swap_token(mm) &&
380 rwsem_is_locked(&mm->mmap_sem))
385 pte_unmap_unlock(pte, ptl);
388 *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) {
408 * If we are reclaiming on behalf of a cgroup, skip
409 * counting on behalf of references from different
412 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
414 referenced += page_referenced_one(page, vma,
415 &mapcount, vm_flags);
420 page_unlock_anon_vma(anon_vma);
425 * page_referenced_file - referenced check for object-based rmap
426 * @page: the page we're checking references on.
427 * @mem_cont: target memory controller
428 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
430 * For an object-based mapped page, find all the places it is mapped and
431 * check/clear the referenced flag. This is done by following the page->mapping
432 * pointer, then walking the chain of vmas it holds. It returns the number
433 * of references it found.
435 * This function is only called from page_referenced for object-based pages.
437 static int page_referenced_file(struct page *page,
438 struct mem_cgroup *mem_cont,
439 unsigned long *vm_flags)
441 unsigned int mapcount;
442 struct address_space *mapping = page->mapping;
443 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
444 struct vm_area_struct *vma;
445 struct prio_tree_iter iter;
449 * The caller's checks on page->mapping and !PageAnon have made
450 * sure that this is a file page: the check for page->mapping
451 * excludes the case just before it gets set on an anon page.
453 BUG_ON(PageAnon(page));
456 * The page lock not only makes sure that page->mapping cannot
457 * suddenly be NULLified by truncation, it makes sure that the
458 * structure at mapping cannot be freed and reused yet,
459 * so we can safely take mapping->i_mmap_lock.
461 BUG_ON(!PageLocked(page));
463 spin_lock(&mapping->i_mmap_lock);
466 * i_mmap_lock does not stabilize mapcount at all, but mapcount
467 * is more likely to be accurate if we note it after spinning.
469 mapcount = page_mapcount(page);
471 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
473 * If we are reclaiming on behalf of a cgroup, skip
474 * counting on behalf of references from different
477 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
479 referenced += page_referenced_one(page, vma,
480 &mapcount, vm_flags);
485 spin_unlock(&mapping->i_mmap_lock);
490 * page_referenced - test if the page was referenced
491 * @page: the page to test
492 * @is_locked: caller holds lock on the page
493 * @mem_cont: target memory controller
494 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
496 * Quick test_and_clear_referenced for all mappings to a page,
497 * returns the number of ptes which referenced the page.
499 int page_referenced(struct page *page,
501 struct mem_cgroup *mem_cont,
502 unsigned long *vm_flags)
506 if (TestClearPageReferenced(page))
510 if (page_mapped(page) && page->mapping) {
512 referenced += page_referenced_anon(page, mem_cont,
515 referenced += page_referenced_file(page, mem_cont,
517 else if (!trylock_page(page))
521 referenced += page_referenced_file(page,
527 if (page_test_and_clear_young(page))
533 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
535 struct mm_struct *mm = vma->vm_mm;
536 unsigned long address;
541 address = vma_address(page, vma);
542 if (address == -EFAULT)
545 pte = page_check_address(page, mm, address, &ptl, 1);
549 if (pte_dirty(*pte) || pte_write(*pte)) {
552 flush_cache_page(vma, address, pte_pfn(*pte));
553 entry = ptep_clear_flush_notify(vma, address, pte);
554 entry = pte_wrprotect(entry);
555 entry = pte_mkclean(entry);
556 set_pte_at(mm, address, pte, entry);
560 pte_unmap_unlock(pte, ptl);
565 static int page_mkclean_file(struct address_space *mapping, struct page *page)
567 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
568 struct vm_area_struct *vma;
569 struct prio_tree_iter iter;
572 BUG_ON(PageAnon(page));
574 spin_lock(&mapping->i_mmap_lock);
575 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
576 if (vma->vm_flags & VM_SHARED)
577 ret += page_mkclean_one(page, vma);
579 spin_unlock(&mapping->i_mmap_lock);
583 int page_mkclean(struct page *page)
587 BUG_ON(!PageLocked(page));
589 if (page_mapped(page)) {
590 struct address_space *mapping = page_mapping(page);
592 ret = page_mkclean_file(mapping, page);
593 if (page_test_dirty(page)) {
594 page_clear_dirty(page);
602 EXPORT_SYMBOL_GPL(page_mkclean);
605 * __page_set_anon_rmap - setup new anonymous rmap
606 * @page: the page to add the mapping to
607 * @vma: the vm area in which the mapping is added
608 * @address: the user virtual address mapped
610 static void __page_set_anon_rmap(struct page *page,
611 struct vm_area_struct *vma, unsigned long address)
613 struct anon_vma *anon_vma = vma->anon_vma;
616 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
617 page->mapping = (struct address_space *) anon_vma;
619 page->index = linear_page_index(vma, address);
622 * nr_mapped state can be updated without turning off
623 * interrupts because it is not modified via interrupt.
625 __inc_zone_page_state(page, NR_ANON_PAGES);
629 * __page_check_anon_rmap - sanity check anonymous rmap addition
630 * @page: the page to add the mapping to
631 * @vma: the vm area in which the mapping is added
632 * @address: the user virtual address mapped
634 static void __page_check_anon_rmap(struct page *page,
635 struct vm_area_struct *vma, unsigned long address)
637 #ifdef CONFIG_DEBUG_VM
639 * The page's anon-rmap details (mapping and index) are guaranteed to
640 * be set up correctly at this point.
642 * We have exclusion against page_add_anon_rmap because the caller
643 * always holds the page locked, except if called from page_dup_rmap,
644 * in which case the page is already known to be setup.
646 * We have exclusion against page_add_new_anon_rmap because those pages
647 * are initially only visible via the pagetables, and the pte is locked
648 * over the call to page_add_new_anon_rmap.
650 struct anon_vma *anon_vma = vma->anon_vma;
651 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
652 BUG_ON(page->mapping != (struct address_space *)anon_vma);
653 BUG_ON(page->index != linear_page_index(vma, address));
658 * page_add_anon_rmap - add pte mapping to an anonymous page
659 * @page: the page to add the mapping to
660 * @vma: the vm area in which the mapping is added
661 * @address: the user virtual address mapped
663 * The caller needs to hold the pte lock and the page must be locked.
665 void page_add_anon_rmap(struct page *page,
666 struct vm_area_struct *vma, unsigned long address)
668 VM_BUG_ON(!PageLocked(page));
669 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
670 if (atomic_inc_and_test(&page->_mapcount))
671 __page_set_anon_rmap(page, vma, address);
673 __page_check_anon_rmap(page, vma, address);
677 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
678 * @page: the page to add the mapping to
679 * @vma: the vm area in which the mapping is added
680 * @address: the user virtual address mapped
682 * Same as page_add_anon_rmap but must only be called on *new* pages.
683 * This means the inc-and-test can be bypassed.
684 * Page does not have to be locked.
686 void page_add_new_anon_rmap(struct page *page,
687 struct vm_area_struct *vma, unsigned long address)
689 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
690 SetPageSwapBacked(page);
691 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
692 __page_set_anon_rmap(page, vma, address);
693 if (page_evictable(page, vma))
694 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
696 add_page_to_unevictable_list(page);
700 * page_add_file_rmap - add pte mapping to a file page
701 * @page: the page to add the mapping to
703 * The caller needs to hold the pte lock.
705 void page_add_file_rmap(struct page *page)
707 if (atomic_inc_and_test(&page->_mapcount)) {
708 __inc_zone_page_state(page, NR_FILE_MAPPED);
709 mem_cgroup_update_mapped_file_stat(page, 1);
713 #ifdef CONFIG_DEBUG_VM
715 * page_dup_rmap - duplicate pte mapping to a page
716 * @page: the page to add the mapping to
717 * @vma: the vm area being duplicated
718 * @address: the user virtual address mapped
720 * For copy_page_range only: minimal extract from page_add_file_rmap /
721 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
724 * The caller needs to hold the pte lock.
726 void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address)
729 __page_check_anon_rmap(page, vma, address);
730 atomic_inc(&page->_mapcount);
735 * page_remove_rmap - take down pte mapping from a page
736 * @page: page to remove mapping from
738 * The caller needs to hold the pte lock.
740 void page_remove_rmap(struct page *page)
742 if (atomic_add_negative(-1, &page->_mapcount)) {
744 * Now that the last pte has gone, s390 must transfer dirty
745 * flag from storage key to struct page. We can usually skip
746 * this if the page is anon, so about to be freed; but perhaps
747 * not if it's in swapcache - there might be another pte slot
748 * containing the swap entry, but page not yet written to swap.
750 if ((!PageAnon(page) || PageSwapCache(page)) &&
751 page_test_dirty(page)) {
752 page_clear_dirty(page);
753 set_page_dirty(page);
756 mem_cgroup_uncharge_page(page);
757 __dec_zone_page_state(page,
758 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
759 mem_cgroup_update_mapped_file_stat(page, -1);
761 * It would be tidy to reset the PageAnon mapping here,
762 * but that might overwrite a racing page_add_anon_rmap
763 * which increments mapcount after us but sets mapping
764 * before us: so leave the reset to free_hot_cold_page,
765 * and remember that it's only reliable while mapped.
766 * Leaving it set also helps swapoff to reinstate ptes
767 * faster for those pages still in swapcache.
773 * Subfunctions of try_to_unmap: try_to_unmap_one called
774 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
776 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
777 enum ttu_flags flags)
779 struct mm_struct *mm = vma->vm_mm;
780 unsigned long address;
784 int ret = SWAP_AGAIN;
786 address = vma_address(page, vma);
787 if (address == -EFAULT)
790 pte = page_check_address(page, mm, address, &ptl, 0);
795 * If the page is mlock()d, we cannot swap it out.
796 * If it's recently referenced (perhaps page_referenced
797 * skipped over this mm) then we should reactivate it.
799 if (!(flags & TTU_IGNORE_MLOCK)) {
800 if (vma->vm_flags & VM_LOCKED) {
805 if (!(flags & TTU_IGNORE_ACCESS)) {
806 if (ptep_clear_flush_young_notify(vma, address, pte)) {
812 /* Nuke the page table entry. */
813 flush_cache_page(vma, address, page_to_pfn(page));
814 pteval = ptep_clear_flush_notify(vma, address, pte);
816 /* Move the dirty bit to the physical page now the pte is gone. */
817 if (pte_dirty(pteval))
818 set_page_dirty(page);
820 /* Update high watermark before we lower rss */
821 update_hiwater_rss(mm);
823 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
825 dec_mm_counter(mm, anon_rss);
827 dec_mm_counter(mm, file_rss);
828 set_pte_at(mm, address, pte,
829 swp_entry_to_pte(make_hwpoison_entry(page)));
830 } else if (PageAnon(page)) {
831 swp_entry_t entry = { .val = page_private(page) };
833 if (PageSwapCache(page)) {
835 * Store the swap location in the pte.
836 * See handle_pte_fault() ...
838 swap_duplicate(entry);
839 if (list_empty(&mm->mmlist)) {
840 spin_lock(&mmlist_lock);
841 if (list_empty(&mm->mmlist))
842 list_add(&mm->mmlist, &init_mm.mmlist);
843 spin_unlock(&mmlist_lock);
845 dec_mm_counter(mm, anon_rss);
846 } else if (PAGE_MIGRATION) {
848 * Store the pfn of the page in a special migration
849 * pte. do_swap_page() will wait until the migration
850 * pte is removed and then restart fault handling.
852 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
853 entry = make_migration_entry(page, pte_write(pteval));
855 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
856 BUG_ON(pte_file(*pte));
857 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
858 /* Establish migration entry for a file page */
860 entry = make_migration_entry(page, pte_write(pteval));
861 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
863 dec_mm_counter(mm, file_rss);
866 page_remove_rmap(page);
867 page_cache_release(page);
870 pte_unmap_unlock(pte, ptl);
876 * objrmap doesn't work for nonlinear VMAs because the assumption that
877 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
878 * Consequently, given a particular page and its ->index, we cannot locate the
879 * ptes which are mapping that page without an exhaustive linear search.
881 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
882 * maps the file to which the target page belongs. The ->vm_private_data field
883 * holds the current cursor into that scan. Successive searches will circulate
884 * around the vma's virtual address space.
886 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
887 * more scanning pressure is placed against them as well. Eventually pages
888 * will become fully unmapped and are eligible for eviction.
890 * For very sparsely populated VMAs this is a little inefficient - chances are
891 * there there won't be many ptes located within the scan cluster. In this case
892 * maybe we could scan further - to the end of the pte page, perhaps.
894 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
895 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
896 * rather than unmapping them. If we encounter the "check_page" that vmscan is
897 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
899 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
900 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
902 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
903 struct vm_area_struct *vma, struct page *check_page)
905 struct mm_struct *mm = vma->vm_mm;
913 unsigned long address;
915 int ret = SWAP_AGAIN;
918 address = (vma->vm_start + cursor) & CLUSTER_MASK;
919 end = address + CLUSTER_SIZE;
920 if (address < vma->vm_start)
921 address = vma->vm_start;
922 if (end > vma->vm_end)
925 pgd = pgd_offset(mm, address);
926 if (!pgd_present(*pgd))
929 pud = pud_offset(pgd, address);
930 if (!pud_present(*pud))
933 pmd = pmd_offset(pud, address);
934 if (!pmd_present(*pmd))
938 * MLOCK_PAGES => feature is configured.
939 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
940 * keep the sem while scanning the cluster for mlocking pages.
942 if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
943 locked_vma = (vma->vm_flags & VM_LOCKED);
945 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
948 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
950 /* Update high watermark before we lower rss */
951 update_hiwater_rss(mm);
953 for (; address < end; pte++, address += PAGE_SIZE) {
954 if (!pte_present(*pte))
956 page = vm_normal_page(vma, address, *pte);
957 BUG_ON(!page || PageAnon(page));
960 mlock_vma_page(page); /* no-op if already mlocked */
961 if (page == check_page)
963 continue; /* don't unmap */
966 if (ptep_clear_flush_young_notify(vma, address, pte))
969 /* Nuke the page table entry. */
970 flush_cache_page(vma, address, pte_pfn(*pte));
971 pteval = ptep_clear_flush_notify(vma, address, pte);
973 /* If nonlinear, store the file page offset in the pte. */
974 if (page->index != linear_page_index(vma, address))
975 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
977 /* Move the dirty bit to the physical page now the pte is gone. */
978 if (pte_dirty(pteval))
979 set_page_dirty(page);
981 page_remove_rmap(page);
982 page_cache_release(page);
983 dec_mm_counter(mm, file_rss);
986 pte_unmap_unlock(pte - 1, ptl);
988 up_read(&vma->vm_mm->mmap_sem);
993 * common handling for pages mapped in VM_LOCKED vmas
995 static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
999 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1000 if (vma->vm_flags & VM_LOCKED) {
1001 mlock_vma_page(page);
1002 mlocked++; /* really mlocked the page */
1004 up_read(&vma->vm_mm->mmap_sem);
1010 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1012 * @page: the page to unmap/unlock
1013 * @unlock: request for unlock rather than unmap [unlikely]
1014 * @migration: unmapping for migration - ignored if @unlock
1016 * Find all the mappings of a page using the mapping pointer and the vma chains
1017 * contained in the anon_vma struct it points to.
1019 * This function is only called from try_to_unmap/try_to_munlock for
1021 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1022 * where the page was found will be held for write. So, we won't recheck
1023 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1026 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1028 struct anon_vma *anon_vma;
1029 struct vm_area_struct *vma;
1030 unsigned int mlocked = 0;
1031 int ret = SWAP_AGAIN;
1032 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
1034 if (MLOCK_PAGES && unlikely(unlock))
1035 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1037 anon_vma = page_lock_anon_vma(page);
1041 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1042 if (MLOCK_PAGES && unlikely(unlock)) {
1043 if (!((vma->vm_flags & VM_LOCKED) &&
1044 page_mapped_in_vma(page, vma)))
1045 continue; /* must visit all unlocked vmas */
1046 ret = SWAP_MLOCK; /* saw at least one mlocked vma */
1048 ret = try_to_unmap_one(page, vma, flags);
1049 if (ret == SWAP_FAIL || !page_mapped(page))
1052 if (ret == SWAP_MLOCK) {
1053 mlocked = try_to_mlock_page(page, vma);
1055 break; /* stop if actually mlocked page */
1059 page_unlock_anon_vma(anon_vma);
1062 ret = SWAP_MLOCK; /* actually mlocked the page */
1063 else if (ret == SWAP_MLOCK)
1064 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1070 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1071 * @page: the page to unmap/unlock
1072 * @flags: action and flags
1074 * Find all the mappings of a page using the mapping pointer and the vma chains
1075 * contained in the address_space struct it points to.
1077 * This function is only called from try_to_unmap/try_to_munlock for
1078 * object-based pages.
1079 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1080 * where the page was found will be held for write. So, we won't recheck
1081 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1084 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1086 struct address_space *mapping = page->mapping;
1087 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1088 struct vm_area_struct *vma;
1089 struct prio_tree_iter iter;
1090 int ret = SWAP_AGAIN;
1091 unsigned long cursor;
1092 unsigned long max_nl_cursor = 0;
1093 unsigned long max_nl_size = 0;
1094 unsigned int mapcount;
1095 unsigned int mlocked = 0;
1096 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
1098 if (MLOCK_PAGES && unlikely(unlock))
1099 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1101 spin_lock(&mapping->i_mmap_lock);
1102 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1103 if (MLOCK_PAGES && unlikely(unlock)) {
1104 if (!((vma->vm_flags & VM_LOCKED) &&
1105 page_mapped_in_vma(page, vma)))
1106 continue; /* must visit all vmas */
1109 ret = try_to_unmap_one(page, vma, flags);
1110 if (ret == SWAP_FAIL || !page_mapped(page))
1113 if (ret == SWAP_MLOCK) {
1114 mlocked = try_to_mlock_page(page, vma);
1116 break; /* stop if actually mlocked page */
1123 if (list_empty(&mapping->i_mmap_nonlinear))
1126 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1127 shared.vm_set.list) {
1128 if (MLOCK_PAGES && unlikely(unlock)) {
1129 if (!(vma->vm_flags & VM_LOCKED))
1130 continue; /* must visit all vmas */
1131 ret = SWAP_MLOCK; /* leave mlocked == 0 */
1132 goto out; /* no need to look further */
1134 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
1135 (vma->vm_flags & VM_LOCKED))
1137 cursor = (unsigned long) vma->vm_private_data;
1138 if (cursor > max_nl_cursor)
1139 max_nl_cursor = cursor;
1140 cursor = vma->vm_end - vma->vm_start;
1141 if (cursor > max_nl_size)
1142 max_nl_size = cursor;
1145 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1151 * We don't try to search for this page in the nonlinear vmas,
1152 * and page_referenced wouldn't have found it anyway. Instead
1153 * just walk the nonlinear vmas trying to age and unmap some.
1154 * The mapcount of the page we came in with is irrelevant,
1155 * but even so use it as a guide to how hard we should try?
1157 mapcount = page_mapcount(page);
1160 cond_resched_lock(&mapping->i_mmap_lock);
1162 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1163 if (max_nl_cursor == 0)
1164 max_nl_cursor = CLUSTER_SIZE;
1167 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1168 shared.vm_set.list) {
1169 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
1170 (vma->vm_flags & VM_LOCKED))
1172 cursor = (unsigned long) vma->vm_private_data;
1173 while ( cursor < max_nl_cursor &&
1174 cursor < vma->vm_end - vma->vm_start) {
1175 ret = try_to_unmap_cluster(cursor, &mapcount,
1177 if (ret == SWAP_MLOCK)
1178 mlocked = 2; /* to return below */
1179 cursor += CLUSTER_SIZE;
1180 vma->vm_private_data = (void *) cursor;
1181 if ((int)mapcount <= 0)
1184 vma->vm_private_data = (void *) max_nl_cursor;
1186 cond_resched_lock(&mapping->i_mmap_lock);
1187 max_nl_cursor += CLUSTER_SIZE;
1188 } while (max_nl_cursor <= max_nl_size);
1191 * Don't loop forever (perhaps all the remaining pages are
1192 * in locked vmas). Reset cursor on all unreserved nonlinear
1193 * vmas, now forgetting on which ones it had fallen behind.
1195 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1196 vma->vm_private_data = NULL;
1198 spin_unlock(&mapping->i_mmap_lock);
1200 ret = SWAP_MLOCK; /* actually mlocked the page */
1201 else if (ret == SWAP_MLOCK)
1202 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1207 * try_to_unmap - try to remove all page table mappings to a page
1208 * @page: the page to get unmapped
1209 * @flags: action and flags
1211 * Tries to remove all the page table entries which are mapping this
1212 * page, used in the pageout path. Caller must hold the page lock.
1213 * Return values are:
1215 * SWAP_SUCCESS - we succeeded in removing all mappings
1216 * SWAP_AGAIN - we missed a mapping, try again later
1217 * SWAP_FAIL - the page is unswappable
1218 * SWAP_MLOCK - page is mlocked.
1220 int try_to_unmap(struct page *page, enum ttu_flags flags)
1224 BUG_ON(!PageLocked(page));
1227 ret = try_to_unmap_anon(page, flags);
1229 ret = try_to_unmap_file(page, flags);
1230 if (ret != SWAP_MLOCK && !page_mapped(page))
1236 * try_to_munlock - try to munlock a page
1237 * @page: the page to be munlocked
1239 * Called from munlock code. Checks all of the VMAs mapping the page
1240 * to make sure nobody else has this page mlocked. The page will be
1241 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1243 * Return values are:
1245 * SWAP_SUCCESS - no vma's holding page mlocked.
1246 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1247 * SWAP_MLOCK - page is now mlocked.
1249 int try_to_munlock(struct page *page)
1251 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1254 return try_to_unmap_anon(page, TTU_MUNLOCK);
1256 return try_to_unmap_file(page, TTU_MUNLOCK);