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 by unuse_process, on anon pages.
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)) {
237 if (vma->vm_file->f_mapping != page->mapping)
241 return vma_address(page, vma);
245 * Check that @page is mapped at @address into @mm.
247 * On success returns with pte mapped and locked.
249 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
250 unsigned long address, spinlock_t **ptlp)
258 pgd = pgd_offset(mm, address);
259 if (!pgd_present(*pgd))
262 pud = pud_offset(pgd, address);
263 if (!pud_present(*pud))
266 pmd = pmd_offset(pud, address);
267 if (!pmd_present(*pmd))
270 pte = pte_offset_map(pmd, address);
271 /* Make a quick check before getting the lock */
272 if (!pte_present(*pte)) {
277 ptl = pte_lockptr(mm, pmd);
279 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
283 pte_unmap_unlock(pte, ptl);
288 * Subfunctions of page_referenced: page_referenced_one called
289 * repeatedly from either page_referenced_anon or page_referenced_file.
291 static int page_referenced_one(struct page *page,
292 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
294 struct mm_struct *mm = vma->vm_mm;
295 unsigned long address;
300 address = vma_address(page, vma);
301 if (address == -EFAULT)
304 pte = page_check_address(page, mm, address, &ptl);
308 if (ptep_clear_flush_young(vma, address, pte))
311 /* Pretend the page is referenced if the task has the
312 swap token and is in the middle of a page fault. */
313 if (mm != current->mm && !ignore_token && has_swap_token(mm) &&
314 rwsem_is_locked(&mm->mmap_sem))
318 pte_unmap_unlock(pte, ptl);
323 static int page_referenced_anon(struct page *page, int ignore_token)
325 unsigned int mapcount;
326 struct anon_vma *anon_vma;
327 struct vm_area_struct *vma;
330 anon_vma = page_lock_anon_vma(page);
334 mapcount = page_mapcount(page);
335 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
336 referenced += page_referenced_one(page, vma, &mapcount,
341 spin_unlock(&anon_vma->lock);
346 * page_referenced_file - referenced check for object-based rmap
347 * @page: the page we're checking references on.
349 * For an object-based mapped page, find all the places it is mapped and
350 * check/clear the referenced flag. This is done by following the page->mapping
351 * pointer, then walking the chain of vmas it holds. It returns the number
352 * of references it found.
354 * This function is only called from page_referenced for object-based pages.
356 static int page_referenced_file(struct page *page, int ignore_token)
358 unsigned int mapcount;
359 struct address_space *mapping = page->mapping;
360 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
361 struct vm_area_struct *vma;
362 struct prio_tree_iter iter;
366 * The caller's checks on page->mapping and !PageAnon have made
367 * sure that this is a file page: the check for page->mapping
368 * excludes the case just before it gets set on an anon page.
370 BUG_ON(PageAnon(page));
373 * The page lock not only makes sure that page->mapping cannot
374 * suddenly be NULLified by truncation, it makes sure that the
375 * structure at mapping cannot be freed and reused yet,
376 * so we can safely take mapping->i_mmap_lock.
378 BUG_ON(!PageLocked(page));
380 spin_lock(&mapping->i_mmap_lock);
383 * i_mmap_lock does not stabilize mapcount at all, but mapcount
384 * is more likely to be accurate if we note it after spinning.
386 mapcount = page_mapcount(page);
388 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
389 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
390 == (VM_LOCKED|VM_MAYSHARE)) {
394 referenced += page_referenced_one(page, vma, &mapcount,
400 spin_unlock(&mapping->i_mmap_lock);
405 * page_referenced - test if the page was referenced
406 * @page: the page to test
407 * @is_locked: caller holds lock on the page
409 * Quick test_and_clear_referenced for all mappings to a page,
410 * returns the number of ptes which referenced the page.
412 int page_referenced(struct page *page, int is_locked, int ignore_token)
416 if (!swap_token_default_timeout)
419 if (page_test_and_clear_young(page))
422 if (TestClearPageReferenced(page))
425 if (page_mapped(page) && page->mapping) {
427 referenced += page_referenced_anon(page, ignore_token);
429 referenced += page_referenced_file(page, ignore_token);
430 else if (TestSetPageLocked(page))
434 referenced += page_referenced_file(page,
443 * page_add_anon_rmap - add pte mapping to an anonymous page
444 * @page: the page to add the mapping to
445 * @vma: the vm area in which the mapping is added
446 * @address: the user virtual address mapped
448 * The caller needs to hold the pte lock.
450 void page_add_anon_rmap(struct page *page,
451 struct vm_area_struct *vma, unsigned long address)
453 if (atomic_inc_and_test(&page->_mapcount)) {
454 struct anon_vma *anon_vma = vma->anon_vma;
457 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
458 page->mapping = (struct address_space *) anon_vma;
460 page->index = linear_page_index(vma, address);
462 inc_page_state(nr_mapped);
464 /* else checking page index and mapping is racy */
468 * page_add_file_rmap - add pte mapping to a file page
469 * @page: the page to add the mapping to
471 * The caller needs to hold the pte lock.
473 void page_add_file_rmap(struct page *page)
475 BUG_ON(PageAnon(page));
476 BUG_ON(!pfn_valid(page_to_pfn(page)));
478 if (atomic_inc_and_test(&page->_mapcount))
479 inc_page_state(nr_mapped);
483 * page_remove_rmap - take down pte mapping from a page
484 * @page: page to remove mapping from
486 * The caller needs to hold the pte lock.
488 void page_remove_rmap(struct page *page)
490 if (atomic_add_negative(-1, &page->_mapcount)) {
491 BUG_ON(page_mapcount(page) < 0);
493 * It would be tidy to reset the PageAnon mapping here,
494 * but that might overwrite a racing page_add_anon_rmap
495 * which increments mapcount after us but sets mapping
496 * before us: so leave the reset to free_hot_cold_page,
497 * and remember that it's only reliable while mapped.
498 * Leaving it set also helps swapoff to reinstate ptes
499 * faster for those pages still in swapcache.
501 if (page_test_and_clear_dirty(page))
502 set_page_dirty(page);
503 dec_page_state(nr_mapped);
508 * Subfunctions of try_to_unmap: try_to_unmap_one called
509 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
511 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
513 struct mm_struct *mm = vma->vm_mm;
514 unsigned long address;
518 int ret = SWAP_AGAIN;
520 address = vma_address(page, vma);
521 if (address == -EFAULT)
524 pte = page_check_address(page, mm, address, &ptl);
529 * If the page is mlock()d, we cannot swap it out.
530 * If it's recently referenced (perhaps page_referenced
531 * skipped over this mm) then we should reactivate it.
533 * Pages belonging to VM_RESERVED regions should not happen here.
535 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
536 ptep_clear_flush_young(vma, address, pte)) {
541 /* Nuke the page table entry. */
542 flush_cache_page(vma, address, page_to_pfn(page));
543 pteval = ptep_clear_flush(vma, address, pte);
545 /* Move the dirty bit to the physical page now the pte is gone. */
546 if (pte_dirty(pteval))
547 set_page_dirty(page);
549 /* Update high watermark before we lower rss */
550 update_hiwater_rss(mm);
552 if (PageAnon(page)) {
553 swp_entry_t entry = { .val = page_private(page) };
555 * Store the swap location in the pte.
556 * See handle_pte_fault() ...
558 BUG_ON(!PageSwapCache(page));
559 swap_duplicate(entry);
560 if (list_empty(&mm->mmlist)) {
561 spin_lock(&mmlist_lock);
562 if (list_empty(&mm->mmlist))
563 list_add(&mm->mmlist, &init_mm.mmlist);
564 spin_unlock(&mmlist_lock);
566 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
567 BUG_ON(pte_file(*pte));
568 dec_mm_counter(mm, anon_rss);
570 dec_mm_counter(mm, file_rss);
572 page_remove_rmap(page);
573 page_cache_release(page);
576 pte_unmap_unlock(pte, ptl);
582 * objrmap doesn't work for nonlinear VMAs because the assumption that
583 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
584 * Consequently, given a particular page and its ->index, we cannot locate the
585 * ptes which are mapping that page without an exhaustive linear search.
587 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
588 * maps the file to which the target page belongs. The ->vm_private_data field
589 * holds the current cursor into that scan. Successive searches will circulate
590 * around the vma's virtual address space.
592 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
593 * more scanning pressure is placed against them as well. Eventually pages
594 * will become fully unmapped and are eligible for eviction.
596 * For very sparsely populated VMAs this is a little inefficient - chances are
597 * there there won't be many ptes located within the scan cluster. In this case
598 * maybe we could scan further - to the end of the pte page, perhaps.
600 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
601 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
603 static void try_to_unmap_cluster(unsigned long cursor,
604 unsigned int *mapcount, struct vm_area_struct *vma)
606 struct mm_struct *mm = vma->vm_mm;
614 unsigned long address;
618 address = (vma->vm_start + cursor) & CLUSTER_MASK;
619 end = address + CLUSTER_SIZE;
620 if (address < vma->vm_start)
621 address = vma->vm_start;
622 if (end > vma->vm_end)
625 pgd = pgd_offset(mm, address);
626 if (!pgd_present(*pgd))
629 pud = pud_offset(pgd, address);
630 if (!pud_present(*pud))
633 pmd = pmd_offset(pud, address);
634 if (!pmd_present(*pmd))
637 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
639 /* Update high watermark before we lower rss */
640 update_hiwater_rss(mm);
642 for (; address < end; pte++, address += PAGE_SIZE) {
643 if (!pte_present(*pte))
647 if (unlikely(!pfn_valid(pfn))) {
648 print_bad_pte(vma, *pte, address);
652 page = pfn_to_page(pfn);
653 BUG_ON(PageAnon(page));
655 if (ptep_clear_flush_young(vma, address, pte))
658 /* Nuke the page table entry. */
659 flush_cache_page(vma, address, pfn);
660 pteval = ptep_clear_flush(vma, address, pte);
662 /* If nonlinear, store the file page offset in the pte. */
663 if (page->index != linear_page_index(vma, address))
664 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
666 /* Move the dirty bit to the physical page now the pte is gone. */
667 if (pte_dirty(pteval))
668 set_page_dirty(page);
670 page_remove_rmap(page);
671 page_cache_release(page);
672 dec_mm_counter(mm, file_rss);
675 pte_unmap_unlock(pte - 1, ptl);
678 static int try_to_unmap_anon(struct page *page)
680 struct anon_vma *anon_vma;
681 struct vm_area_struct *vma;
682 int ret = SWAP_AGAIN;
684 anon_vma = page_lock_anon_vma(page);
688 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
689 ret = try_to_unmap_one(page, vma);
690 if (ret == SWAP_FAIL || !page_mapped(page))
693 spin_unlock(&anon_vma->lock);
698 * try_to_unmap_file - unmap file page using the object-based rmap method
699 * @page: the page to unmap
701 * Find all the mappings of a page using the mapping pointer and the vma chains
702 * contained in the address_space struct it points to.
704 * This function is only called from try_to_unmap for object-based pages.
706 static int try_to_unmap_file(struct page *page)
708 struct address_space *mapping = page->mapping;
709 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
710 struct vm_area_struct *vma;
711 struct prio_tree_iter iter;
712 int ret = SWAP_AGAIN;
713 unsigned long cursor;
714 unsigned long max_nl_cursor = 0;
715 unsigned long max_nl_size = 0;
716 unsigned int mapcount;
718 spin_lock(&mapping->i_mmap_lock);
719 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
720 ret = try_to_unmap_one(page, vma);
721 if (ret == SWAP_FAIL || !page_mapped(page))
725 if (list_empty(&mapping->i_mmap_nonlinear))
728 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
729 shared.vm_set.list) {
730 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
732 cursor = (unsigned long) vma->vm_private_data;
733 if (cursor > max_nl_cursor)
734 max_nl_cursor = cursor;
735 cursor = vma->vm_end - vma->vm_start;
736 if (cursor > max_nl_size)
737 max_nl_size = cursor;
740 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
746 * We don't try to search for this page in the nonlinear vmas,
747 * and page_referenced wouldn't have found it anyway. Instead
748 * just walk the nonlinear vmas trying to age and unmap some.
749 * The mapcount of the page we came in with is irrelevant,
750 * but even so use it as a guide to how hard we should try?
752 mapcount = page_mapcount(page);
755 cond_resched_lock(&mapping->i_mmap_lock);
757 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
758 if (max_nl_cursor == 0)
759 max_nl_cursor = CLUSTER_SIZE;
762 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
763 shared.vm_set.list) {
764 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
766 cursor = (unsigned long) vma->vm_private_data;
767 while ( cursor < max_nl_cursor &&
768 cursor < vma->vm_end - vma->vm_start) {
769 try_to_unmap_cluster(cursor, &mapcount, vma);
770 cursor += CLUSTER_SIZE;
771 vma->vm_private_data = (void *) cursor;
772 if ((int)mapcount <= 0)
775 vma->vm_private_data = (void *) max_nl_cursor;
777 cond_resched_lock(&mapping->i_mmap_lock);
778 max_nl_cursor += CLUSTER_SIZE;
779 } while (max_nl_cursor <= max_nl_size);
782 * Don't loop forever (perhaps all the remaining pages are
783 * in locked vmas). Reset cursor on all unreserved nonlinear
784 * vmas, now forgetting on which ones it had fallen behind.
786 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
787 shared.vm_set.list) {
788 if (!(vma->vm_flags & VM_RESERVED))
789 vma->vm_private_data = NULL;
792 spin_unlock(&mapping->i_mmap_lock);
797 * try_to_unmap - try to remove all page table mappings to a page
798 * @page: the page to get unmapped
800 * Tries to remove all the page table entries which are mapping this
801 * page, used in the pageout path. Caller must hold the page lock.
804 * SWAP_SUCCESS - we succeeded in removing all mappings
805 * SWAP_AGAIN - we missed a mapping, try again later
806 * SWAP_FAIL - the page is unswappable
808 int try_to_unmap(struct page *page)
812 BUG_ON(!PageLocked(page));
815 ret = try_to_unmap_anon(page);
817 ret = try_to_unmap_file(page);
819 if (!page_mapped(page))