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/rmap.h>
55 #include <linux/rcupdate.h>
56 #include <linux/vs_memory.h>
57 #include <linux/rcupdate.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 if (!vma->anon_vma) {
126 BUG_ON(!next->anon_vma);
127 vma->anon_vma = next->anon_vma;
128 list_add(&vma->anon_vma_node, &next->anon_vma_node);
130 /* if they're both non-null they must be the same */
131 BUG_ON(vma->anon_vma != next->anon_vma);
133 list_del(&next->anon_vma_node);
136 void __anon_vma_link(struct vm_area_struct *vma)
138 struct anon_vma *anon_vma = vma->anon_vma;
141 list_add(&vma->anon_vma_node, &anon_vma->head);
142 validate_anon_vma(vma);
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(&vma->anon_vma_node, &anon_vma->head);
153 validate_anon_vma(vma);
154 spin_unlock(&anon_vma->lock);
158 void anon_vma_unlink(struct vm_area_struct *vma)
160 struct anon_vma *anon_vma = vma->anon_vma;
166 spin_lock(&anon_vma->lock);
167 validate_anon_vma(vma);
168 list_del(&vma->anon_vma_node);
170 /* We must garbage collect the anon_vma if it's empty */
171 empty = list_empty(&anon_vma->head);
172 spin_unlock(&anon_vma->lock);
175 anon_vma_free(anon_vma);
178 static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
180 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
181 SLAB_CTOR_CONSTRUCTOR) {
182 struct anon_vma *anon_vma = data;
184 spin_lock_init(&anon_vma->lock);
185 INIT_LIST_HEAD(&anon_vma->head);
189 void __init anon_vma_init(void)
191 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
192 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
196 * Getting a lock on a stable anon_vma from a page off the LRU is
197 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
199 static struct anon_vma *page_lock_anon_vma(struct page *page)
201 struct anon_vma *anon_vma = NULL;
202 unsigned long anon_mapping;
205 anon_mapping = (unsigned long) page->mapping;
206 if (!(anon_mapping & PAGE_MAPPING_ANON))
208 if (!page_mapped(page))
211 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
212 spin_lock(&anon_vma->lock);
219 * At what user virtual address is page expected in vma?
221 static inline unsigned long
222 vma_address(struct page *page, struct vm_area_struct *vma)
224 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
225 unsigned long address;
227 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
228 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
229 /* page should be within any vma from prio_tree_next */
230 BUG_ON(!PageAnon(page));
237 * At what user virtual address is page expected in vma? checking that the
238 * page matches the vma: currently only used by unuse_process, on anon pages.
240 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
242 if (PageAnon(page)) {
243 if ((void *)vma->anon_vma !=
244 (void *)page->mapping - PAGE_MAPPING_ANON)
246 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
247 if (vma->vm_file->f_mapping != page->mapping)
251 return vma_address(page, vma);
255 * Subfunctions of page_referenced: page_referenced_one called
256 * repeatedly from either page_referenced_anon or page_referenced_file.
258 static int page_referenced_one(struct page *page,
259 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
261 struct mm_struct *mm = vma->vm_mm;
262 unsigned long address;
270 address = vma_address(page, vma);
271 if (address == -EFAULT)
274 spin_lock(&mm->page_table_lock);
276 pgd = pgd_offset(mm, address);
277 if (!pgd_present(*pgd))
280 pmd = pmd_offset(pgd, address);
281 if (!pmd_present(*pmd))
284 pte = pte_offset_map(pmd, address);
285 if (!pte_present(*pte))
288 if (page_to_pfn(page) != pte_pfn(*pte))
291 if (ptep_clear_flush_young(vma, address, pte))
294 if (mm != current->mm && !ignore_token && has_swap_token(mm))
302 spin_unlock(&mm->page_table_lock);
307 static int page_referenced_anon(struct page *page, int ignore_token)
309 unsigned int mapcount;
310 struct anon_vma *anon_vma;
311 struct vm_area_struct *vma;
314 anon_vma = page_lock_anon_vma(page);
318 mapcount = page_mapcount(page);
319 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
320 referenced += page_referenced_one(page, vma, &mapcount,
325 spin_unlock(&anon_vma->lock);
330 * page_referenced_file - referenced check for object-based rmap
331 * @page: the page we're checking references on.
333 * For an object-based mapped page, find all the places it is mapped and
334 * check/clear the referenced flag. This is done by following the page->mapping
335 * pointer, then walking the chain of vmas it holds. It returns the number
336 * of references it found.
338 * This function is only called from page_referenced for object-based pages.
340 static int page_referenced_file(struct page *page, int ignore_token)
342 unsigned int mapcount;
343 struct address_space *mapping = page->mapping;
344 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
345 struct vm_area_struct *vma;
346 struct prio_tree_iter iter;
350 * The caller's checks on page->mapping and !PageAnon have made
351 * sure that this is a file page: the check for page->mapping
352 * excludes the case just before it gets set on an anon page.
354 BUG_ON(PageAnon(page));
357 * The page lock not only makes sure that page->mapping cannot
358 * suddenly be NULLified by truncation, it makes sure that the
359 * structure at mapping cannot be freed and reused yet,
360 * so we can safely take mapping->i_mmap_lock.
362 BUG_ON(!PageLocked(page));
364 spin_lock(&mapping->i_mmap_lock);
367 * i_mmap_lock does not stabilize mapcount at all, but mapcount
368 * is more likely to be accurate if we note it after spinning.
370 mapcount = page_mapcount(page);
372 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
373 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
374 == (VM_LOCKED|VM_MAYSHARE)) {
378 referenced += page_referenced_one(page, vma, &mapcount,
384 spin_unlock(&mapping->i_mmap_lock);
389 * page_referenced - test if the page was referenced
390 * @page: the page to test
391 * @is_locked: caller holds lock on the page
393 * Quick test_and_clear_referenced for all mappings to a page,
394 * returns the number of ptes which referenced the page.
396 int page_referenced(struct page *page, int is_locked, int ignore_token)
400 if (!swap_token_default_timeout)
403 if (page_test_and_clear_young(page))
406 if (TestClearPageReferenced(page))
409 if (page_mapped(page) && page->mapping) {
411 referenced += page_referenced_anon(page, ignore_token);
413 referenced += page_referenced_file(page, ignore_token);
414 else if (TestSetPageLocked(page))
418 referenced += page_referenced_file(page,
427 * page_add_anon_rmap - add pte mapping to an anonymous page
428 * @page: the page to add the mapping to
429 * @vma: the vm area in which the mapping is added
430 * @address: the user virtual address mapped
432 * The caller needs to hold the mm->page_table_lock.
434 void page_add_anon_rmap(struct page *page,
435 struct vm_area_struct *vma, unsigned long address)
437 struct anon_vma *anon_vma = vma->anon_vma;
440 BUG_ON(PageReserved(page));
443 vma->vm_mm->anon_rss++;
445 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
446 index = (address - vma->vm_start) >> PAGE_SHIFT;
447 index += vma->vm_pgoff;
448 index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
450 if (atomic_inc_and_test(&page->_mapcount)) {
452 page->mapping = (struct address_space *) anon_vma;
453 inc_page_state(nr_mapped);
455 /* else checking page index and mapping is racy */
459 * page_add_file_rmap - add pte mapping to a file page
460 * @page: the page to add the mapping to
462 * The caller needs to hold the mm->page_table_lock.
464 void page_add_file_rmap(struct page *page)
466 BUG_ON(PageAnon(page));
467 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
470 if (atomic_inc_and_test(&page->_mapcount))
471 inc_page_state(nr_mapped);
475 * page_remove_rmap - take down pte mapping from a page
476 * @page: page to remove mapping from
478 * Caller needs to hold the mm->page_table_lock.
480 void page_remove_rmap(struct page *page)
482 BUG_ON(PageReserved(page));
484 if (atomic_add_negative(-1, &page->_mapcount)) {
485 BUG_ON(page_mapcount(page) < 0);
487 * It would be tidy to reset the PageAnon mapping here,
488 * but that might overwrite a racing page_add_anon_rmap
489 * which increments mapcount after us but sets mapping
490 * before us: so leave the reset to free_hot_cold_page,
491 * and remember that it's only reliable while mapped.
492 * Leaving it set also helps swapoff to reinstate ptes
493 * faster for those pages still in swapcache.
495 if (page_test_and_clear_dirty(page))
496 set_page_dirty(page);
497 dec_page_state(nr_mapped);
502 * Subfunctions of try_to_unmap: try_to_unmap_one called
503 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
505 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
507 struct mm_struct *mm = vma->vm_mm;
508 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 pmd = pmd_offset(pgd, address);
532 if (!pmd_present(*pmd))
535 pte = pte_offset_map(pmd, address);
536 if (!pte_present(*pte))
539 if (page_to_pfn(page) != pte_pfn(*pte))
543 * If the page is mlock()d, we cannot swap it out.
544 * If it's recently referenced (perhaps page_referenced
545 * skipped over this mm) then we should reactivate it.
547 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
548 ptep_clear_flush_young(vma, address, pte)) {
554 * Don't pull an anonymous page out from under get_user_pages.
555 * GUP carefully breaks COW and raises page count (while holding
556 * page_table_lock, as we have here) to make sure that the page
557 * cannot be freed. If we unmap that page here, a user write
558 * access to the virtual address will bring back the page, but
559 * its raised count will (ironically) be taken to mean it's not
560 * an exclusive swap page, do_wp_page will replace it by a copy
561 * page, and the user never get to see the data GUP was holding
562 * the original page for.
564 * This test is also useful for when swapoff (unuse_process) has
565 * to drop page lock: its reference to the page stops existing
566 * ptes from being unmapped, so swapoff can make progress.
568 if (PageSwapCache(page) &&
569 page_count(page) != page_mapcount(page) + 2) {
574 /* Nuke the page table entry. */
575 flush_cache_page(vma, address);
576 pteval = ptep_clear_flush(vma, address, pte);
578 /* Move the dirty bit to the physical page now the pte is gone. */
579 if (pte_dirty(pteval))
580 set_page_dirty(page);
582 if (PageAnon(page)) {
583 swp_entry_t entry = { .val = page->private };
585 * Store the swap location in the pte.
586 * See handle_pte_fault() ...
588 BUG_ON(!PageSwapCache(page));
589 swap_duplicate(entry);
590 if (list_empty(&mm->mmlist)) {
591 spin_lock(&mmlist_lock);
592 list_add(&mm->mmlist, &init_mm.mmlist);
593 spin_unlock(&mmlist_lock);
595 set_pte(pte, swp_entry_to_pte(entry));
596 BUG_ON(pte_file(*pte));
602 page_remove_rmap(page);
603 page_cache_release(page);
608 spin_unlock(&mm->page_table_lock);
614 * objrmap doesn't work for nonlinear VMAs because the assumption that
615 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
616 * Consequently, given a particular page and its ->index, we cannot locate the
617 * ptes which are mapping that page without an exhaustive linear search.
619 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
620 * maps the file to which the target page belongs. The ->vm_private_data field
621 * holds the current cursor into that scan. Successive searches will circulate
622 * around the vma's virtual address space.
624 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
625 * more scanning pressure is placed against them as well. Eventually pages
626 * will become fully unmapped and are eligible for eviction.
628 * For very sparsely populated VMAs this is a little inefficient - chances are
629 * there there won't be many ptes located within the scan cluster. In this case
630 * maybe we could scan further - to the end of the pte page, perhaps.
632 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
633 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
635 static void try_to_unmap_cluster(unsigned long cursor,
636 unsigned int *mapcount, struct vm_area_struct *vma)
638 struct mm_struct *mm = vma->vm_mm;
644 unsigned long address;
649 * We need the page_table_lock to protect us from page faults,
650 * munmap, fork, etc...
652 spin_lock(&mm->page_table_lock);
654 address = (vma->vm_start + cursor) & CLUSTER_MASK;
655 end = address + CLUSTER_SIZE;
656 if (address < vma->vm_start)
657 address = vma->vm_start;
658 if (end > vma->vm_end)
661 pgd = pgd_offset(mm, address);
662 if (!pgd_present(*pgd))
665 pmd = pmd_offset(pgd, address);
666 if (!pmd_present(*pmd))
669 for (pte = pte_offset_map(pmd, address);
670 address < end; pte++, address += PAGE_SIZE) {
672 if (!pte_present(*pte))
679 page = pfn_to_page(pfn);
680 BUG_ON(PageAnon(page));
681 if (PageReserved(page))
684 if (ptep_clear_flush_young(vma, address, pte))
687 /* Nuke the page table entry. */
688 flush_cache_page(vma, address);
689 pteval = ptep_clear_flush(vma, address, pte);
691 /* If nonlinear, store the file page offset in the pte. */
692 if (page->index != linear_page_index(vma, address))
693 set_pte(pte, pgoff_to_pte(page->index));
695 /* Move the dirty bit to the physical page now the pte is gone. */
696 if (pte_dirty(pteval))
697 set_page_dirty(page);
699 page_remove_rmap(page);
700 page_cache_release(page);
709 spin_unlock(&mm->page_table_lock);
712 static int try_to_unmap_anon(struct page *page)
714 struct anon_vma *anon_vma;
715 struct vm_area_struct *vma;
716 int ret = SWAP_AGAIN;
718 anon_vma = page_lock_anon_vma(page);
722 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
723 ret = try_to_unmap_one(page, vma);
724 if (ret == SWAP_FAIL || !page_mapped(page))
727 spin_unlock(&anon_vma->lock);
732 * try_to_unmap_file - unmap file page using the object-based rmap method
733 * @page: the page to unmap
735 * Find all the mappings of a page using the mapping pointer and the vma chains
736 * contained in the address_space struct it points to.
738 * This function is only called from try_to_unmap for object-based pages.
740 static int try_to_unmap_file(struct page *page)
742 struct address_space *mapping = page->mapping;
743 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
744 struct vm_area_struct *vma;
745 struct prio_tree_iter iter;
746 int ret = SWAP_AGAIN;
747 unsigned long cursor;
748 unsigned long max_nl_cursor = 0;
749 unsigned long max_nl_size = 0;
750 unsigned int mapcount;
752 spin_lock(&mapping->i_mmap_lock);
753 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
754 ret = try_to_unmap_one(page, vma);
755 if (ret == SWAP_FAIL || !page_mapped(page))
759 if (list_empty(&mapping->i_mmap_nonlinear))
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 if (cursor > max_nl_cursor)
768 max_nl_cursor = cursor;
769 cursor = vma->vm_end - vma->vm_start;
770 if (cursor > max_nl_size)
771 max_nl_size = cursor;
774 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
780 * We don't try to search for this page in the nonlinear vmas,
781 * and page_referenced wouldn't have found it anyway. Instead
782 * just walk the nonlinear vmas trying to age and unmap some.
783 * The mapcount of the page we came in with is irrelevant,
784 * but even so use it as a guide to how hard we should try?
786 mapcount = page_mapcount(page);
789 cond_resched_lock(&mapping->i_mmap_lock);
791 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
792 if (max_nl_cursor == 0)
793 max_nl_cursor = CLUSTER_SIZE;
796 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
797 shared.vm_set.list) {
798 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
800 cursor = (unsigned long) vma->vm_private_data;
801 while (vma->vm_mm->rss &&
802 cursor < max_nl_cursor &&
803 cursor < vma->vm_end - vma->vm_start) {
804 try_to_unmap_cluster(cursor, &mapcount, vma);
805 cursor += CLUSTER_SIZE;
806 vma->vm_private_data = (void *) cursor;
807 if ((int)mapcount <= 0)
810 vma->vm_private_data = (void *) max_nl_cursor;
812 cond_resched_lock(&mapping->i_mmap_lock);
813 max_nl_cursor += CLUSTER_SIZE;
814 } while (max_nl_cursor <= max_nl_size);
817 * Don't loop forever (perhaps all the remaining pages are
818 * in locked vmas). Reset cursor on all unreserved nonlinear
819 * vmas, now forgetting on which ones it had fallen behind.
821 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
822 shared.vm_set.list) {
823 if (!(vma->vm_flags & VM_RESERVED))
824 vma->vm_private_data = NULL;
827 spin_unlock(&mapping->i_mmap_lock);
832 * try_to_unmap - try to remove all page table mappings to a page
833 * @page: the page to get unmapped
835 * Tries to remove all the page table entries which are mapping this
836 * page, used in the pageout path. Caller must hold the page lock.
839 * SWAP_SUCCESS - we succeeded in removing all mappings
840 * SWAP_AGAIN - we missed a mapping, try again later
841 * SWAP_FAIL - the page is unswappable
843 int try_to_unmap(struct page *page)
847 BUG_ON(PageReserved(page));
848 BUG_ON(!PageLocked(page));
851 ret = try_to_unmap_anon(page);
853 ret = try_to_unmap_file(page);
855 if (!page_mapped(page))