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_sem (while writing or truncating, not reading or faulting)
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.
32 * page->flags PG_locked (lock_page)
33 * mapping->i_mmap_lock
35 * mm->page_table_lock or pte_lock
36 * zone->lru_lock (in mark_page_accessed)
37 * swap_lock (in swap_duplicate, swap_info_get)
38 * mmlist_lock (in mmput, drain_mmlist and others)
39 * 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)
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>
56 #include <asm/tlbflush.h>
58 //#define RMAP_DEBUG /* can be enabled only for debugging */
60 kmem_cache_t *anon_vma_cachep;
62 static inline void validate_anon_vma(struct vm_area_struct *find_vma)
65 struct anon_vma *anon_vma = find_vma->anon_vma;
66 struct vm_area_struct *vma;
67 unsigned int mapcount = 0;
70 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
72 BUG_ON(mapcount > 100000);
80 /* This must be called under the mmap_sem. */
81 int anon_vma_prepare(struct vm_area_struct *vma)
83 struct anon_vma *anon_vma = vma->anon_vma;
86 if (unlikely(!anon_vma)) {
87 struct mm_struct *mm = vma->vm_mm;
88 struct anon_vma *allocated, *locked;
90 anon_vma = find_mergeable_anon_vma(vma);
94 spin_lock(&locked->lock);
96 anon_vma = anon_vma_alloc();
97 if (unlikely(!anon_vma))
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);
110 spin_unlock(&mm->page_table_lock);
113 spin_unlock(&locked->lock);
114 if (unlikely(allocated))
115 anon_vma_free(allocated);
120 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
122 BUG_ON(vma->anon_vma != next->anon_vma);
123 list_del(&next->anon_vma_node);
126 void __anon_vma_link(struct vm_area_struct *vma)
128 struct anon_vma *anon_vma = vma->anon_vma;
131 list_add(&vma->anon_vma_node, &anon_vma->head);
132 validate_anon_vma(vma);
136 void anon_vma_link(struct vm_area_struct *vma)
138 struct anon_vma *anon_vma = vma->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);
148 void anon_vma_unlink(struct vm_area_struct *vma)
150 struct anon_vma *anon_vma = vma->anon_vma;
156 spin_lock(&anon_vma->lock);
157 validate_anon_vma(vma);
158 list_del(&vma->anon_vma_node);
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);
165 anon_vma_free(anon_vma);
168 static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
170 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
171 SLAB_CTOR_CONSTRUCTOR) {
172 struct anon_vma *anon_vma = data;
174 spin_lock_init(&anon_vma->lock);
175 INIT_LIST_HEAD(&anon_vma->head);
179 void __init anon_vma_init(void)
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);
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.
189 static struct anon_vma *page_lock_anon_vma(struct page *page)
191 struct anon_vma *anon_vma = NULL;
192 unsigned long anon_mapping;
195 anon_mapping = (unsigned long) page->mapping;
196 if (!(anon_mapping & PAGE_MAPPING_ANON))
198 if (!page_mapped(page))
201 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
202 spin_lock(&anon_vma->lock);
209 * At what user virtual address is page expected in vma?
211 static inline unsigned long
212 vma_address(struct page *page, struct vm_area_struct *vma)
214 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
215 unsigned long address;
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));
227 * At what user virtual address is page expected in vma? checking that the
228 * page matches the vma: currently only used on anon pages, by unuse_vma;
230 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
232 if (PageAnon(page)) {
233 if ((void *)vma->anon_vma !=
234 (void *)page->mapping - PAGE_MAPPING_ANON)
236 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
238 vma->vm_file->f_mapping != page->mapping)
242 return vma_address(page, vma);
246 * Check that @page is mapped at @address into @mm.
248 * On success returns with pte mapped and locked.
250 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
251 unsigned long address, spinlock_t **ptlp)
259 pgd = pgd_offset(mm, address);
260 if (!pgd_present(*pgd))
263 pud = pud_offset(pgd, address);
264 if (!pud_present(*pud))
267 pmd = pmd_offset(pud, address);
268 if (!pmd_present(*pmd))
271 pte = pte_offset_map(pmd, address);
272 /* Make a quick check before getting the lock */
273 if (!pte_present(*pte)) {
278 ptl = pte_lockptr(mm, pmd);
280 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
284 pte_unmap_unlock(pte, ptl);
289 * Subfunctions of page_referenced: page_referenced_one called
290 * repeatedly from either page_referenced_anon or page_referenced_file.
292 static int page_referenced_one(struct page *page,
293 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
295 struct mm_struct *mm = vma->vm_mm;
296 unsigned long address;
301 address = vma_address(page, vma);
302 if (address == -EFAULT)
305 pte = page_check_address(page, mm, address, &ptl);
309 if (ptep_clear_flush_young(vma, address, pte))
312 /* Pretend the page is referenced if the task has the
313 swap token and is in the middle of a page fault. */
314 if (mm != current->mm && !ignore_token && has_swap_token(mm) &&
315 rwsem_is_locked(&mm->mmap_sem))
319 pte_unmap_unlock(pte, ptl);
324 static int page_referenced_anon(struct page *page, int ignore_token)
326 unsigned int mapcount;
327 struct anon_vma *anon_vma;
328 struct vm_area_struct *vma;
331 anon_vma = page_lock_anon_vma(page);
335 mapcount = page_mapcount(page);
336 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
337 referenced += page_referenced_one(page, vma, &mapcount,
342 spin_unlock(&anon_vma->lock);
347 * page_referenced_file - referenced check for object-based rmap
348 * @page: the page we're checking references on.
350 * For an object-based mapped page, find all the places it is mapped and
351 * check/clear the referenced flag. This is done by following the page->mapping
352 * pointer, then walking the chain of vmas it holds. It returns the number
353 * of references it found.
355 * This function is only called from page_referenced for object-based pages.
357 static int page_referenced_file(struct page *page, int ignore_token)
359 unsigned int mapcount;
360 struct address_space *mapping = page->mapping;
361 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
362 struct vm_area_struct *vma;
363 struct prio_tree_iter iter;
367 * The caller's checks on page->mapping and !PageAnon have made
368 * sure that this is a file page: the check for page->mapping
369 * excludes the case just before it gets set on an anon page.
371 BUG_ON(PageAnon(page));
374 * The page lock not only makes sure that page->mapping cannot
375 * suddenly be NULLified by truncation, it makes sure that the
376 * structure at mapping cannot be freed and reused yet,
377 * so we can safely take mapping->i_mmap_lock.
379 BUG_ON(!PageLocked(page));
381 spin_lock(&mapping->i_mmap_lock);
384 * i_mmap_lock does not stabilize mapcount at all, but mapcount
385 * is more likely to be accurate if we note it after spinning.
387 mapcount = page_mapcount(page);
389 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
390 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
391 == (VM_LOCKED|VM_MAYSHARE)) {
395 referenced += page_referenced_one(page, vma, &mapcount,
401 spin_unlock(&mapping->i_mmap_lock);
406 * page_referenced - test if the page was referenced
407 * @page: the page to test
408 * @is_locked: caller holds lock on the page
410 * Quick test_and_clear_referenced for all mappings to a page,
411 * returns the number of ptes which referenced the page.
413 int page_referenced(struct page *page, int is_locked, int ignore_token)
417 if (!swap_token_default_timeout)
420 if (page_test_and_clear_young(page))
423 if (TestClearPageReferenced(page))
426 if (page_mapped(page) && page->mapping) {
428 referenced += page_referenced_anon(page, ignore_token);
430 referenced += page_referenced_file(page, ignore_token);
431 else if (TestSetPageLocked(page))
435 referenced += page_referenced_file(page,
444 * page_add_anon_rmap - add pte mapping to an anonymous page
445 * @page: the page to add the mapping to
446 * @vma: the vm area in which the mapping is added
447 * @address: the user virtual address mapped
449 * The caller needs to hold the pte lock.
451 void page_add_anon_rmap(struct page *page,
452 struct vm_area_struct *vma, unsigned long address)
454 if (atomic_inc_and_test(&page->_mapcount)) {
455 struct anon_vma *anon_vma = vma->anon_vma;
458 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
459 page->mapping = (struct address_space *) anon_vma;
461 page->index = linear_page_index(vma, address);
463 inc_page_state(nr_mapped);
465 /* else checking page index and mapping is racy */
469 * page_add_file_rmap - add pte mapping to a file page
470 * @page: the page to add the mapping to
472 * The caller needs to hold the pte lock.
474 void page_add_file_rmap(struct page *page)
476 BUG_ON(PageAnon(page));
477 BUG_ON(!pfn_valid(page_to_pfn(page)));
479 if (atomic_inc_and_test(&page->_mapcount))
480 inc_page_state(nr_mapped);
484 * page_remove_rmap - take down pte mapping from a page
485 * @page: page to remove mapping from
487 * The caller needs to hold the pte lock.
489 void page_remove_rmap(struct page *page)
491 if (atomic_add_negative(-1, &page->_mapcount)) {
492 BUG_ON(page_mapcount(page) < 0);
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.
502 if (page_test_and_clear_dirty(page))
503 set_page_dirty(page);
504 dec_page_state(nr_mapped);
509 * Subfunctions of try_to_unmap: try_to_unmap_one called
510 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
512 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
514 struct mm_struct *mm = vma->vm_mm;
515 unsigned long address;
519 int ret = SWAP_AGAIN;
521 address = vma_address(page, vma);
522 if (address == -EFAULT)
525 pte = page_check_address(page, mm, address, &ptl);
530 * If the page is mlock()d, we cannot swap it out.
531 * If it's recently referenced (perhaps page_referenced
532 * skipped over this mm) then we should reactivate it.
534 if ((vma->vm_flags & VM_LOCKED) ||
535 ptep_clear_flush_young(vma, address, pte)) {
540 /* Nuke the page table entry. */
541 flush_cache_page(vma, address, page_to_pfn(page));
542 pteval = ptep_clear_flush(vma, address, pte);
544 /* Move the dirty bit to the physical page now the pte is gone. */
545 if (pte_dirty(pteval))
546 set_page_dirty(page);
548 /* Update high watermark before we lower rss */
549 update_hiwater_rss(mm);
551 if (PageAnon(page)) {
552 swp_entry_t entry = { .val = page_private(page) };
554 * Store the swap location in the pte.
555 * See handle_pte_fault() ...
557 BUG_ON(!PageSwapCache(page));
558 swap_duplicate(entry);
559 if (list_empty(&mm->mmlist)) {
560 spin_lock(&mmlist_lock);
561 if (list_empty(&mm->mmlist))
562 list_add(&mm->mmlist, &init_mm.mmlist);
563 spin_unlock(&mmlist_lock);
565 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
566 BUG_ON(pte_file(*pte));
567 dec_mm_counter(mm, anon_rss);
569 dec_mm_counter(mm, file_rss);
571 page_remove_rmap(page);
572 page_cache_release(page);
575 pte_unmap_unlock(pte, ptl);
581 * objrmap doesn't work for nonlinear VMAs because the assumption that
582 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
583 * Consequently, given a particular page and its ->index, we cannot locate the
584 * ptes which are mapping that page without an exhaustive linear search.
586 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
587 * maps the file to which the target page belongs. The ->vm_private_data field
588 * holds the current cursor into that scan. Successive searches will circulate
589 * around the vma's virtual address space.
591 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
592 * more scanning pressure is placed against them as well. Eventually pages
593 * will become fully unmapped and are eligible for eviction.
595 * For very sparsely populated VMAs this is a little inefficient - chances are
596 * there there won't be many ptes located within the scan cluster. In this case
597 * maybe we could scan further - to the end of the pte page, perhaps.
599 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
600 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
602 static void try_to_unmap_cluster(unsigned long cursor,
603 unsigned int *mapcount, struct vm_area_struct *vma)
605 struct mm_struct *mm = vma->vm_mm;
613 unsigned long address;
616 address = (vma->vm_start + cursor) & CLUSTER_MASK;
617 end = address + CLUSTER_SIZE;
618 if (address < vma->vm_start)
619 address = vma->vm_start;
620 if (end > vma->vm_end)
623 pgd = pgd_offset(mm, address);
624 if (!pgd_present(*pgd))
627 pud = pud_offset(pgd, address);
628 if (!pud_present(*pud))
631 pmd = pmd_offset(pud, address);
632 if (!pmd_present(*pmd))
635 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
637 /* Update high watermark before we lower rss */
638 update_hiwater_rss(mm);
640 for (; address < end; pte++, address += PAGE_SIZE) {
641 if (!pte_present(*pte))
643 page = vm_normal_page(vma, address, *pte);
644 BUG_ON(!page || PageAnon(page));
646 if (ptep_clear_flush_young(vma, address, pte))
649 /* Nuke the page table entry. */
650 flush_cache_page(vma, address, pfn);
651 pteval = ptep_clear_flush(vma, address, pte);
653 /* If nonlinear, store the file page offset in the pte. */
654 if (page->index != linear_page_index(vma, address))
655 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
657 /* Move the dirty bit to the physical page now the pte is gone. */
658 if (pte_dirty(pteval))
659 set_page_dirty(page);
661 page_remove_rmap(page);
662 page_cache_release(page);
663 dec_mm_counter(mm, file_rss);
666 pte_unmap_unlock(pte - 1, ptl);
669 static int try_to_unmap_anon(struct page *page)
671 struct anon_vma *anon_vma;
672 struct vm_area_struct *vma;
673 int ret = SWAP_AGAIN;
675 anon_vma = page_lock_anon_vma(page);
679 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
680 ret = try_to_unmap_one(page, vma);
681 if (ret == SWAP_FAIL || !page_mapped(page))
684 spin_unlock(&anon_vma->lock);
689 * try_to_unmap_file - unmap file page using the object-based rmap method
690 * @page: the page to unmap
692 * Find all the mappings of a page using the mapping pointer and the vma chains
693 * contained in the address_space struct it points to.
695 * This function is only called from try_to_unmap for object-based pages.
697 static int try_to_unmap_file(struct page *page)
699 struct address_space *mapping = page->mapping;
700 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
701 struct vm_area_struct *vma;
702 struct prio_tree_iter iter;
703 int ret = SWAP_AGAIN;
704 unsigned long cursor;
705 unsigned long max_nl_cursor = 0;
706 unsigned long max_nl_size = 0;
707 unsigned int mapcount;
709 spin_lock(&mapping->i_mmap_lock);
710 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
711 ret = try_to_unmap_one(page, vma);
712 if (ret == SWAP_FAIL || !page_mapped(page))
716 if (list_empty(&mapping->i_mmap_nonlinear))
719 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
720 shared.vm_set.list) {
721 if (vma->vm_flags & VM_LOCKED)
723 cursor = (unsigned long) vma->vm_private_data;
724 if (cursor > max_nl_cursor)
725 max_nl_cursor = cursor;
726 cursor = vma->vm_end - vma->vm_start;
727 if (cursor > max_nl_size)
728 max_nl_size = cursor;
731 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
737 * We don't try to search for this page in the nonlinear vmas,
738 * and page_referenced wouldn't have found it anyway. Instead
739 * just walk the nonlinear vmas trying to age and unmap some.
740 * The mapcount of the page we came in with is irrelevant,
741 * but even so use it as a guide to how hard we should try?
743 mapcount = page_mapcount(page);
746 cond_resched_lock(&mapping->i_mmap_lock);
748 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
749 if (max_nl_cursor == 0)
750 max_nl_cursor = CLUSTER_SIZE;
753 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
754 shared.vm_set.list) {
755 if (vma->vm_flags & VM_LOCKED)
757 cursor = (unsigned long) vma->vm_private_data;
758 while ( cursor < max_nl_cursor &&
759 cursor < vma->vm_end - vma->vm_start) {
760 try_to_unmap_cluster(cursor, &mapcount, vma);
761 cursor += CLUSTER_SIZE;
762 vma->vm_private_data = (void *) cursor;
763 if ((int)mapcount <= 0)
766 vma->vm_private_data = (void *) max_nl_cursor;
768 cond_resched_lock(&mapping->i_mmap_lock);
769 max_nl_cursor += CLUSTER_SIZE;
770 } while (max_nl_cursor <= max_nl_size);
773 * Don't loop forever (perhaps all the remaining pages are
774 * in locked vmas). Reset cursor on all unreserved nonlinear
775 * vmas, now forgetting on which ones it had fallen behind.
777 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
778 vma->vm_private_data = NULL;
780 spin_unlock(&mapping->i_mmap_lock);
785 * try_to_unmap - try to remove all page table mappings to a page
786 * @page: the page to get unmapped
788 * Tries to remove all the page table entries which are mapping this
789 * page, used in the pageout path. Caller must hold the page lock.
792 * SWAP_SUCCESS - we succeeded in removing all mappings
793 * SWAP_AGAIN - we missed a mapping, try again later
794 * SWAP_FAIL - the page is unswappable
796 int try_to_unmap(struct page *page)
800 BUG_ON(!PageLocked(page));
803 ret = try_to_unmap_anon(page);
805 ret = try_to_unmap_file(page);
807 if (!page_mapped(page))