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
36 * zone->lru_lock (in mark_page_accessed)
37 * swap_list_lock (in swap_free etc's swap_info_get)
38 * mmlist_lock (in mmput, drain_mmlist and others)
39 * swap_device_lock (in swap_duplicate, swap_info_get)
40 * mapping->private_lock (in __set_page_dirty_buffers)
41 * inode_lock (in set_page_dirty's __mark_inode_dirty)
42 * sb_lock (within inode_lock in fs/fs-writeback.c)
43 * mapping->tree_lock (widely used, in set_page_dirty,
44 * in arch-dependent flush_dcache_mmap_lock,
45 * within inode_lock in __sync_single_inode)
49 #include <linux/pagemap.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/slab.h>
53 #include <linux/init.h>
54 #include <linux/acct.h>
55 #include <linux/rmap.h>
56 #include <linux/rcupdate.h>
57 #include <linux/vs_memory.h>
59 #include <asm/tlbflush.h>
61 //#define RMAP_DEBUG /* can be enabled only for debugging */
63 kmem_cache_t *anon_vma_cachep;
65 static inline void validate_anon_vma(struct vm_area_struct *find_vma)
68 struct anon_vma *anon_vma = find_vma->anon_vma;
69 struct vm_area_struct *vma;
70 unsigned int mapcount = 0;
73 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
75 BUG_ON(mapcount > 100000);
83 /* This must be called under the mmap_sem. */
84 int anon_vma_prepare(struct vm_area_struct *vma)
86 struct anon_vma *anon_vma = vma->anon_vma;
89 if (unlikely(!anon_vma)) {
90 struct mm_struct *mm = vma->vm_mm;
91 struct anon_vma *allocated, *locked;
93 anon_vma = find_mergeable_anon_vma(vma);
97 spin_lock(&locked->lock);
99 anon_vma = anon_vma_alloc();
100 if (unlikely(!anon_vma))
102 allocated = anon_vma;
106 /* page_table_lock to protect against threads */
107 spin_lock(&mm->page_table_lock);
108 if (likely(!vma->anon_vma)) {
109 vma->anon_vma = anon_vma;
110 list_add(&vma->anon_vma_node, &anon_vma->head);
113 spin_unlock(&mm->page_table_lock);
116 spin_unlock(&locked->lock);
117 if (unlikely(allocated))
118 anon_vma_free(allocated);
123 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
125 BUG_ON(vma->anon_vma != next->anon_vma);
126 list_del(&next->anon_vma_node);
129 void __anon_vma_link(struct vm_area_struct *vma)
131 struct anon_vma *anon_vma = vma->anon_vma;
134 list_add(&vma->anon_vma_node, &anon_vma->head);
135 validate_anon_vma(vma);
139 void anon_vma_link(struct vm_area_struct *vma)
141 struct anon_vma *anon_vma = vma->anon_vma;
144 spin_lock(&anon_vma->lock);
145 list_add(&vma->anon_vma_node, &anon_vma->head);
146 validate_anon_vma(vma);
147 spin_unlock(&anon_vma->lock);
151 void anon_vma_unlink(struct vm_area_struct *vma)
153 struct anon_vma *anon_vma = vma->anon_vma;
159 spin_lock(&anon_vma->lock);
160 validate_anon_vma(vma);
161 list_del(&vma->anon_vma_node);
163 /* We must garbage collect the anon_vma if it's empty */
164 empty = list_empty(&anon_vma->head);
165 spin_unlock(&anon_vma->lock);
168 anon_vma_free(anon_vma);
171 static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
173 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
174 SLAB_CTOR_CONSTRUCTOR) {
175 struct anon_vma *anon_vma = data;
177 spin_lock_init(&anon_vma->lock);
178 INIT_LIST_HEAD(&anon_vma->head);
182 void __init anon_vma_init(void)
184 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
185 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
189 * Getting a lock on a stable anon_vma from a page off the LRU is
190 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
192 static struct anon_vma *page_lock_anon_vma(struct page *page)
194 struct anon_vma *anon_vma = NULL;
195 unsigned long anon_mapping;
198 anon_mapping = (unsigned long) page->mapping;
199 if (!(anon_mapping & PAGE_MAPPING_ANON))
201 if (!page_mapped(page))
204 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
205 spin_lock(&anon_vma->lock);
212 * At what user virtual address is page expected in vma?
214 static inline unsigned long
215 vma_address(struct page *page, struct vm_area_struct *vma)
217 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
218 unsigned long address;
220 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
221 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
222 /* page should be within any vma from prio_tree_next */
223 BUG_ON(!PageAnon(page));
230 * At what user virtual address is page expected in vma? checking that the
231 * page matches the vma: currently only used by unuse_process, on anon pages.
233 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
235 if (PageAnon(page)) {
236 if ((void *)vma->anon_vma !=
237 (void *)page->mapping - PAGE_MAPPING_ANON)
239 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
240 if (vma->vm_file->f_mapping != page->mapping)
244 return vma_address(page, vma);
248 * Subfunctions of page_referenced: page_referenced_one called
249 * repeatedly from either page_referenced_anon or page_referenced_file.
251 static int page_referenced_one(struct page *page,
252 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
254 struct mm_struct *mm = vma->vm_mm;
255 unsigned long address;
264 address = vma_address(page, vma);
265 if (address == -EFAULT)
268 spin_lock(&mm->page_table_lock);
270 pgd = pgd_offset(mm, address);
271 if (!pgd_present(*pgd))
274 pud = pud_offset(pgd, address);
275 if (!pud_present(*pud))
278 pmd = pmd_offset(pud, address);
279 if (!pmd_present(*pmd))
282 pte = pte_offset_map(pmd, address);
283 if (!pte_present(*pte))
286 if (page_to_pfn(page) != pte_pfn(*pte))
289 if (ptep_clear_flush_young(vma, address, pte))
292 if (mm != current->mm && !ignore_token && has_swap_token(mm))
300 spin_unlock(&mm->page_table_lock);
305 static int page_referenced_anon(struct page *page, int ignore_token)
307 unsigned int mapcount;
308 struct anon_vma *anon_vma;
309 struct vm_area_struct *vma;
312 anon_vma = page_lock_anon_vma(page);
316 mapcount = page_mapcount(page);
317 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
318 referenced += page_referenced_one(page, vma, &mapcount,
323 spin_unlock(&anon_vma->lock);
328 * page_referenced_file - referenced check for object-based rmap
329 * @page: the page we're checking references on.
331 * For an object-based mapped page, find all the places it is mapped and
332 * check/clear the referenced flag. This is done by following the page->mapping
333 * pointer, then walking the chain of vmas it holds. It returns the number
334 * of references it found.
336 * This function is only called from page_referenced for object-based pages.
338 static int page_referenced_file(struct page *page, int ignore_token)
340 unsigned int mapcount;
341 struct address_space *mapping = page->mapping;
342 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
343 struct vm_area_struct *vma;
344 struct prio_tree_iter iter;
348 * The caller's checks on page->mapping and !PageAnon have made
349 * sure that this is a file page: the check for page->mapping
350 * excludes the case just before it gets set on an anon page.
352 BUG_ON(PageAnon(page));
355 * The page lock not only makes sure that page->mapping cannot
356 * suddenly be NULLified by truncation, it makes sure that the
357 * structure at mapping cannot be freed and reused yet,
358 * so we can safely take mapping->i_mmap_lock.
360 BUG_ON(!PageLocked(page));
362 spin_lock(&mapping->i_mmap_lock);
365 * i_mmap_lock does not stabilize mapcount at all, but mapcount
366 * is more likely to be accurate if we note it after spinning.
368 mapcount = page_mapcount(page);
370 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
371 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
372 == (VM_LOCKED|VM_MAYSHARE)) {
376 referenced += page_referenced_one(page, vma, &mapcount,
382 spin_unlock(&mapping->i_mmap_lock);
387 * page_referenced - test if the page was referenced
388 * @page: the page to test
389 * @is_locked: caller holds lock on the page
391 * Quick test_and_clear_referenced for all mappings to a page,
392 * returns the number of ptes which referenced the page.
394 int page_referenced(struct page *page, int is_locked, int ignore_token)
398 if (!swap_token_default_timeout)
401 if (page_test_and_clear_young(page))
404 if (TestClearPageReferenced(page))
407 if (page_mapped(page) && page->mapping) {
409 referenced += page_referenced_anon(page, ignore_token);
411 referenced += page_referenced_file(page, ignore_token);
412 else if (TestSetPageLocked(page))
416 referenced += page_referenced_file(page,
425 * page_add_anon_rmap - add pte mapping to an anonymous page
426 * @page: the page to add the mapping to
427 * @vma: the vm area in which the mapping is added
428 * @address: the user virtual address mapped
430 * The caller needs to hold the mm->page_table_lock.
432 void page_add_anon_rmap(struct page *page,
433 struct vm_area_struct *vma, unsigned long address)
435 struct anon_vma *anon_vma = vma->anon_vma;
438 BUG_ON(PageReserved(page));
441 // vma->vm_mm->anon_rss++;
442 vx_anonpages_inc(vma->vm_mm);
444 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
445 index = (address - vma->vm_start) >> PAGE_SHIFT;
446 index += vma->vm_pgoff;
447 index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
449 if (atomic_inc_and_test(&page->_mapcount)) {
451 page->mapping = (struct address_space *) anon_vma;
452 inc_page_state(nr_mapped);
454 /* else checking page index and mapping is racy */
458 * page_add_file_rmap - add pte mapping to a file page
459 * @page: the page to add the mapping to
461 * The caller needs to hold the mm->page_table_lock.
463 void page_add_file_rmap(struct page *page)
465 BUG_ON(PageAnon(page));
466 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
469 if (atomic_inc_and_test(&page->_mapcount))
470 inc_page_state(nr_mapped);
474 * page_remove_rmap - take down pte mapping from a page
475 * @page: page to remove mapping from
477 * Caller needs to hold the mm->page_table_lock.
479 void page_remove_rmap(struct page *page)
481 BUG_ON(PageReserved(page));
483 if (atomic_add_negative(-1, &page->_mapcount)) {
484 BUG_ON(page_mapcount(page) < 0);
486 * It would be tidy to reset the PageAnon mapping here,
487 * but that might overwrite a racing page_add_anon_rmap
488 * which increments mapcount after us but sets mapping
489 * before us: so leave the reset to free_hot_cold_page,
490 * and remember that it's only reliable while mapped.
491 * Leaving it set also helps swapoff to reinstate ptes
492 * faster for those pages still in swapcache.
494 if (page_test_and_clear_dirty(page))
495 set_page_dirty(page);
496 dec_page_state(nr_mapped);
501 * Subfunctions of try_to_unmap: try_to_unmap_one called
502 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
504 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
506 struct mm_struct *mm = vma->vm_mm;
507 unsigned long address;
513 int ret = SWAP_AGAIN;
517 address = vma_address(page, vma);
518 if (address == -EFAULT)
522 * We need the page_table_lock to protect us from page faults,
523 * munmap, fork, etc...
525 spin_lock(&mm->page_table_lock);
527 pgd = pgd_offset(mm, address);
528 if (!pgd_present(*pgd))
531 pud = pud_offset(pgd, address);
532 if (!pud_present(*pud))
535 pmd = pmd_offset(pud, address);
536 if (!pmd_present(*pmd))
539 pte = pte_offset_map(pmd, address);
540 if (!pte_present(*pte))
543 if (page_to_pfn(page) != pte_pfn(*pte))
547 * If the page is mlock()d, we cannot swap it out.
548 * If it's recently referenced (perhaps page_referenced
549 * skipped over this mm) then we should reactivate it.
551 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
552 ptep_clear_flush_young(vma, address, pte)) {
558 * Don't pull an anonymous page out from under get_user_pages.
559 * GUP carefully breaks COW and raises page count (while holding
560 * page_table_lock, as we have here) to make sure that the page
561 * cannot be freed. If we unmap that page here, a user write
562 * access to the virtual address will bring back the page, but
563 * its raised count will (ironically) be taken to mean it's not
564 * an exclusive swap page, do_wp_page will replace it by a copy
565 * page, and the user never get to see the data GUP was holding
566 * the original page for.
568 * This test is also useful for when swapoff (unuse_process) has
569 * to drop page lock: its reference to the page stops existing
570 * ptes from being unmapped, so swapoff can make progress.
572 if (PageSwapCache(page) &&
573 page_count(page) != page_mapcount(page) + 2) {
578 /* Nuke the page table entry. */
579 flush_cache_page(vma, address);
580 pteval = ptep_clear_flush(vma, address, pte);
582 /* Move the dirty bit to the physical page now the pte is gone. */
583 if (pte_dirty(pteval))
584 set_page_dirty(page);
586 if (PageAnon(page)) {
587 swp_entry_t entry = { .val = page->private };
589 * Store the swap location in the pte.
590 * See handle_pte_fault() ...
592 BUG_ON(!PageSwapCache(page));
593 swap_duplicate(entry);
594 if (list_empty(&mm->mmlist)) {
595 spin_lock(&mmlist_lock);
596 list_add(&mm->mmlist, &init_mm.mmlist);
597 spin_unlock(&mmlist_lock);
599 set_pte(pte, swp_entry_to_pte(entry));
600 BUG_ON(pte_file(*pte));
602 vx_anonpages_dec(mm);
607 acct_update_integrals();
608 page_remove_rmap(page);
609 page_cache_release(page);
614 spin_unlock(&mm->page_table_lock);
620 * objrmap doesn't work for nonlinear VMAs because the assumption that
621 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
622 * Consequently, given a particular page and its ->index, we cannot locate the
623 * ptes which are mapping that page without an exhaustive linear search.
625 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
626 * maps the file to which the target page belongs. The ->vm_private_data field
627 * holds the current cursor into that scan. Successive searches will circulate
628 * around the vma's virtual address space.
630 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
631 * more scanning pressure is placed against them as well. Eventually pages
632 * will become fully unmapped and are eligible for eviction.
634 * For very sparsely populated VMAs this is a little inefficient - chances are
635 * there there won't be many ptes located within the scan cluster. In this case
636 * maybe we could scan further - to the end of the pte page, perhaps.
638 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
639 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
641 static void try_to_unmap_cluster(unsigned long cursor,
642 unsigned int *mapcount, struct vm_area_struct *vma)
644 struct mm_struct *mm = vma->vm_mm;
651 unsigned long address;
656 * We need the page_table_lock to protect us from page faults,
657 * munmap, fork, etc...
659 spin_lock(&mm->page_table_lock);
661 address = (vma->vm_start + cursor) & CLUSTER_MASK;
662 end = address + CLUSTER_SIZE;
663 if (address < vma->vm_start)
664 address = vma->vm_start;
665 if (end > vma->vm_end)
668 pgd = pgd_offset(mm, address);
669 if (!pgd_present(*pgd))
672 pud = pud_offset(pgd, address);
673 if (!pud_present(*pud))
676 pmd = pmd_offset(pud, address);
677 if (!pmd_present(*pmd))
680 for (pte = pte_offset_map(pmd, address);
681 address < end; pte++, address += PAGE_SIZE) {
683 if (!pte_present(*pte))
690 page = pfn_to_page(pfn);
691 BUG_ON(PageAnon(page));
692 if (PageReserved(page))
695 if (ptep_clear_flush_young(vma, address, pte))
698 /* Nuke the page table entry. */
699 flush_cache_page(vma, address);
700 pteval = ptep_clear_flush(vma, address, pte);
702 /* If nonlinear, store the file page offset in the pte. */
703 if (page->index != linear_page_index(vma, address))
704 set_pte(pte, pgoff_to_pte(page->index));
706 /* Move the dirty bit to the physical page now the pte is gone. */
707 if (pte_dirty(pteval))
708 set_page_dirty(page);
710 page_remove_rmap(page);
711 page_cache_release(page);
712 acct_update_integrals();
721 spin_unlock(&mm->page_table_lock);
724 static int try_to_unmap_anon(struct page *page)
726 struct anon_vma *anon_vma;
727 struct vm_area_struct *vma;
728 int ret = SWAP_AGAIN;
730 anon_vma = page_lock_anon_vma(page);
734 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
735 ret = try_to_unmap_one(page, vma);
736 if (ret == SWAP_FAIL || !page_mapped(page))
739 spin_unlock(&anon_vma->lock);
744 * try_to_unmap_file - unmap file page using the object-based rmap method
745 * @page: the page to unmap
747 * Find all the mappings of a page using the mapping pointer and the vma chains
748 * contained in the address_space struct it points to.
750 * This function is only called from try_to_unmap for object-based pages.
752 static int try_to_unmap_file(struct page *page)
754 struct address_space *mapping = page->mapping;
755 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
756 struct vm_area_struct *vma;
757 struct prio_tree_iter iter;
758 int ret = SWAP_AGAIN;
759 unsigned long cursor;
760 unsigned long max_nl_cursor = 0;
761 unsigned long max_nl_size = 0;
762 unsigned int mapcount;
764 spin_lock(&mapping->i_mmap_lock);
765 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
766 ret = try_to_unmap_one(page, vma);
767 if (ret == SWAP_FAIL || !page_mapped(page))
771 if (list_empty(&mapping->i_mmap_nonlinear))
774 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
775 shared.vm_set.list) {
776 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
778 cursor = (unsigned long) vma->vm_private_data;
779 if (cursor > max_nl_cursor)
780 max_nl_cursor = cursor;
781 cursor = vma->vm_end - vma->vm_start;
782 if (cursor > max_nl_size)
783 max_nl_size = cursor;
786 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
792 * We don't try to search for this page in the nonlinear vmas,
793 * and page_referenced wouldn't have found it anyway. Instead
794 * just walk the nonlinear vmas trying to age and unmap some.
795 * The mapcount of the page we came in with is irrelevant,
796 * but even so use it as a guide to how hard we should try?
798 mapcount = page_mapcount(page);
801 cond_resched_lock(&mapping->i_mmap_lock);
803 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
804 if (max_nl_cursor == 0)
805 max_nl_cursor = CLUSTER_SIZE;
808 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
809 shared.vm_set.list) {
810 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
812 cursor = (unsigned long) vma->vm_private_data;
813 while (vma->vm_mm->rss &&
814 cursor < max_nl_cursor &&
815 cursor < vma->vm_end - vma->vm_start) {
816 try_to_unmap_cluster(cursor, &mapcount, vma);
817 cursor += CLUSTER_SIZE;
818 vma->vm_private_data = (void *) cursor;
819 if ((int)mapcount <= 0)
822 vma->vm_private_data = (void *) max_nl_cursor;
824 cond_resched_lock(&mapping->i_mmap_lock);
825 max_nl_cursor += CLUSTER_SIZE;
826 } while (max_nl_cursor <= max_nl_size);
829 * Don't loop forever (perhaps all the remaining pages are
830 * in locked vmas). Reset cursor on all unreserved nonlinear
831 * vmas, now forgetting on which ones it had fallen behind.
833 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
834 shared.vm_set.list) {
835 if (!(vma->vm_flags & VM_RESERVED))
836 vma->vm_private_data = NULL;
839 spin_unlock(&mapping->i_mmap_lock);
844 * try_to_unmap - try to remove all page table mappings to a page
845 * @page: the page to get unmapped
847 * Tries to remove all the page table entries which are mapping this
848 * page, used in the pageout path. Caller must hold the page lock.
851 * SWAP_SUCCESS - we succeeded in removing all mappings
852 * SWAP_AGAIN - we missed a mapping, try again later
853 * SWAP_FAIL - the page is unswappable
855 int try_to_unmap(struct page *page)
859 BUG_ON(PageReserved(page));
860 BUG_ON(!PageLocked(page));
863 ret = try_to_unmap_anon(page);
865 ret = try_to_unmap_file(page);
867 if (!page_mapped(page))