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 * swap_device_lock (in swap_duplicate, swap_info_get)
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>
55 #include <linux/vs_memory.h>
57 #include <asm/tlbflush.h>
59 //#define RMAP_DEBUG /* can be enabled only for debugging */
61 kmem_cache_t *anon_vma_cachep;
63 static inline void validate_anon_vma(struct vm_area_struct *find_vma)
66 struct anon_vma *anon_vma = find_vma->anon_vma;
67 struct vm_area_struct *vma;
68 unsigned int mapcount = 0;
71 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
73 BUG_ON(mapcount > 100000);
81 /* This must be called under the mmap_sem. */
82 int anon_vma_prepare(struct vm_area_struct *vma)
84 struct anon_vma *anon_vma = vma->anon_vma;
87 if (unlikely(!anon_vma)) {
88 struct mm_struct *mm = vma->vm_mm;
89 struct anon_vma *allocated, *locked;
91 anon_vma = find_mergeable_anon_vma(vma);
95 spin_lock(&locked->lock);
97 anon_vma = anon_vma_alloc();
98 if (unlikely(!anon_vma))
100 allocated = anon_vma;
104 /* page_table_lock to protect against threads */
105 spin_lock(&mm->page_table_lock);
106 if (likely(!vma->anon_vma)) {
107 vma->anon_vma = anon_vma;
108 list_add(&vma->anon_vma_node, &anon_vma->head);
111 spin_unlock(&mm->page_table_lock);
114 spin_unlock(&locked->lock);
115 if (unlikely(allocated))
116 anon_vma_free(allocated);
121 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
123 if (!vma->anon_vma) {
124 BUG_ON(!next->anon_vma);
125 vma->anon_vma = next->anon_vma;
126 list_add(&vma->anon_vma_node, &next->anon_vma_node);
128 /* if they're both non-null they must be the same */
129 BUG_ON(vma->anon_vma != next->anon_vma);
131 list_del(&next->anon_vma_node);
134 void __anon_vma_link(struct vm_area_struct *vma)
136 struct anon_vma *anon_vma = vma->anon_vma;
139 list_add(&vma->anon_vma_node, &anon_vma->head);
140 validate_anon_vma(vma);
144 void anon_vma_link(struct vm_area_struct *vma)
146 struct anon_vma *anon_vma = vma->anon_vma;
149 spin_lock(&anon_vma->lock);
150 list_add(&vma->anon_vma_node, &anon_vma->head);
151 validate_anon_vma(vma);
152 spin_unlock(&anon_vma->lock);
156 void anon_vma_unlink(struct vm_area_struct *vma)
158 struct anon_vma *anon_vma = vma->anon_vma;
164 spin_lock(&anon_vma->lock);
165 validate_anon_vma(vma);
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, kmem_cache_t *cachep, unsigned long flags)
178 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
179 SLAB_CTOR_CONSTRUCTOR) {
180 struct anon_vma *anon_vma = data;
182 spin_lock_init(&anon_vma->lock);
183 INIT_LIST_HEAD(&anon_vma->head);
187 void __init anon_vma_init(void)
189 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
190 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
194 * Getting a lock on a stable anon_vma from a page off the LRU is
195 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
197 static struct anon_vma *page_lock_anon_vma(struct page *page)
199 struct anon_vma *anon_vma = NULL;
200 unsigned long anon_mapping;
203 anon_mapping = (unsigned long) page->mapping;
204 if (!(anon_mapping & PAGE_MAPPING_ANON))
206 if (!page_mapped(page))
209 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
210 spin_lock(&anon_vma->lock);
217 * At what user virtual address is page expected in vma?
219 static inline unsigned long
220 vma_address(struct page *page, struct vm_area_struct *vma)
222 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
223 unsigned long address;
225 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
226 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
227 /* page should be within any vma from prio_tree_next */
228 BUG_ON(!PageAnon(page));
235 * At what user virtual address is page expected in vma? checking that the
236 * page matches the vma: currently only used by unuse_process, on anon pages.
238 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
240 if (PageAnon(page)) {
241 if ((void *)vma->anon_vma !=
242 (void *)page->mapping - PAGE_MAPPING_ANON)
244 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
245 if (vma->vm_file->f_mapping != page->mapping)
249 return vma_address(page, vma);
253 * Subfunctions of page_referenced: page_referenced_one called
254 * repeatedly from either page_referenced_anon or page_referenced_file.
256 static int page_referenced_one(struct page *page,
257 struct vm_area_struct *vma, unsigned int *mapcount)
259 struct mm_struct *mm = vma->vm_mm;
260 unsigned long address;
268 address = vma_address(page, vma);
269 if (address == -EFAULT)
272 spin_lock(&mm->page_table_lock);
274 pgd = pgd_offset(mm, address);
275 if (!pgd_present(*pgd))
278 pmd = pmd_offset(pgd, 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 && has_swap_token(mm))
300 spin_unlock(&mm->page_table_lock);
305 static int page_referenced_anon(struct page *page)
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);
322 spin_unlock(&anon_vma->lock);
327 * page_referenced_file - referenced check for object-based rmap
328 * @page: the page we're checking references on.
330 * For an object-based mapped page, find all the places it is mapped and
331 * check/clear the referenced flag. This is done by following the page->mapping
332 * pointer, then walking the chain of vmas it holds. It returns the number
333 * of references it found.
335 * This function is only called from page_referenced for object-based pages.
337 static int page_referenced_file(struct page *page)
339 unsigned int mapcount;
340 struct address_space *mapping = page->mapping;
341 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
342 struct vm_area_struct *vma;
343 struct prio_tree_iter iter;
347 * The caller's checks on page->mapping and !PageAnon have made
348 * sure that this is a file page: the check for page->mapping
349 * excludes the case just before it gets set on an anon page.
351 BUG_ON(PageAnon(page));
354 * The page lock not only makes sure that page->mapping cannot
355 * suddenly be NULLified by truncation, it makes sure that the
356 * structure at mapping cannot be freed and reused yet,
357 * so we can safely take mapping->i_mmap_lock.
359 BUG_ON(!PageLocked(page));
361 spin_lock(&mapping->i_mmap_lock);
364 * i_mmap_lock does not stabilize mapcount at all, but mapcount
365 * is more likely to be accurate if we note it after spinning.
367 mapcount = page_mapcount(page);
369 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
370 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
371 == (VM_LOCKED|VM_MAYSHARE)) {
375 referenced += page_referenced_one(page, vma, &mapcount);
380 spin_unlock(&mapping->i_mmap_lock);
385 * page_referenced - test if the page was referenced
386 * @page: the page to test
387 * @is_locked: caller holds lock on the page
389 * Quick test_and_clear_referenced for all mappings to a page,
390 * returns the number of ptes which referenced the page.
392 int page_referenced(struct page *page, int is_locked)
396 if (page_test_and_clear_young(page))
399 if (TestClearPageReferenced(page))
402 if (page_mapped(page) && page->mapping) {
404 referenced += page_referenced_anon(page);
406 referenced += page_referenced_file(page);
407 else if (TestSetPageLocked(page))
411 referenced += page_referenced_file(page);
419 * page_add_anon_rmap - add pte mapping to an anonymous page
420 * @page: the page to add the mapping to
421 * @vma: the vm area in which the mapping is added
422 * @address: the user virtual address mapped
424 * The caller needs to hold the mm->page_table_lock.
426 void page_add_anon_rmap(struct page *page,
427 struct vm_area_struct *vma, unsigned long address)
429 struct anon_vma *anon_vma = vma->anon_vma;
432 BUG_ON(PageReserved(page));
435 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
436 index = (address - vma->vm_start) >> PAGE_SHIFT;
437 index += vma->vm_pgoff;
438 index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
440 if (atomic_inc_and_test(&page->_mapcount)) {
442 page->mapping = (struct address_space *) anon_vma;
443 inc_page_state(nr_mapped);
445 /* else checking page index and mapping is racy */
449 * page_add_file_rmap - add pte mapping to a file page
450 * @page: the page to add the mapping to
452 * The caller needs to hold the mm->page_table_lock.
454 void page_add_file_rmap(struct page *page)
456 BUG_ON(PageAnon(page));
457 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
460 if (atomic_inc_and_test(&page->_mapcount))
461 inc_page_state(nr_mapped);
465 * page_remove_rmap - take down pte mapping from a page
466 * @page: page to remove mapping from
468 * Caller needs to hold the mm->page_table_lock.
470 void page_remove_rmap(struct page *page)
472 BUG_ON(PageReserved(page));
474 if (atomic_add_negative(-1, &page->_mapcount)) {
475 BUG_ON(page_mapcount(page) < 0);
477 * It would be tidy to reset the PageAnon mapping here,
478 * but that might overwrite a racing page_add_anon_rmap
479 * which increments mapcount after us but sets mapping
480 * before us: so leave the reset to free_hot_cold_page,
481 * and remember that it's only reliable while mapped.
482 * Leaving it set also helps swapoff to reinstate ptes
483 * faster for those pages still in swapcache.
485 if (page_test_and_clear_dirty(page))
486 set_page_dirty(page);
487 dec_page_state(nr_mapped);
492 * Subfunctions of try_to_unmap: try_to_unmap_one called
493 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
495 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
497 struct mm_struct *mm = vma->vm_mm;
498 unsigned long address;
503 int ret = SWAP_AGAIN;
507 address = vma_address(page, vma);
508 if (address == -EFAULT)
512 * We need the page_table_lock to protect us from page faults,
513 * munmap, fork, etc...
515 spin_lock(&mm->page_table_lock);
517 pgd = pgd_offset(mm, address);
518 if (!pgd_present(*pgd))
521 pmd = pmd_offset(pgd, address);
522 if (!pmd_present(*pmd))
525 pte = pte_offset_map(pmd, address);
526 if (!pte_present(*pte))
529 if (page_to_pfn(page) != pte_pfn(*pte))
533 * If the page is mlock()d, we cannot swap it out.
534 * If it's recently referenced (perhaps page_referenced
535 * skipped over this mm) then we should reactivate it.
537 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
538 ptep_clear_flush_young(vma, address, pte)) {
544 * Don't pull an anonymous page out from under get_user_pages.
545 * GUP carefully breaks COW and raises page count (while holding
546 * page_table_lock, as we have here) to make sure that the page
547 * cannot be freed. If we unmap that page here, a user write
548 * access to the virtual address will bring back the page, but
549 * its raised count will (ironically) be taken to mean it's not
550 * an exclusive swap page, do_wp_page will replace it by a copy
551 * page, and the user never get to see the data GUP was holding
552 * the original page for.
554 * This test is also useful for when swapoff (unuse_process) has
555 * to drop page lock: its reference to the page stops existing
556 * ptes from being unmapped, so swapoff can make progress.
558 if (PageSwapCache(page) &&
559 page_count(page) != page_mapcount(page) + 2) {
564 /* Nuke the page table entry. */
565 flush_cache_page(vma, address);
566 pteval = ptep_clear_flush(vma, address, pte);
568 /* Move the dirty bit to the physical page now the pte is gone. */
569 if (pte_dirty(pteval))
570 set_page_dirty(page);
572 if (PageAnon(page)) {
573 swp_entry_t entry = { .val = page->private };
575 * Store the swap location in the pte.
576 * See handle_pte_fault() ...
578 BUG_ON(!PageSwapCache(page));
579 swap_duplicate(entry);
580 set_pte(pte, swp_entry_to_pte(entry));
581 BUG_ON(pte_file(*pte));
586 page_remove_rmap(page);
587 page_cache_release(page);
592 spin_unlock(&mm->page_table_lock);
598 * objrmap doesn't work for nonlinear VMAs because the assumption that
599 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
600 * Consequently, given a particular page and its ->index, we cannot locate the
601 * ptes which are mapping that page without an exhaustive linear search.
603 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
604 * maps the file to which the target page belongs. The ->vm_private_data field
605 * holds the current cursor into that scan. Successive searches will circulate
606 * around the vma's virtual address space.
608 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
609 * more scanning pressure is placed against them as well. Eventually pages
610 * will become fully unmapped and are eligible for eviction.
612 * For very sparsely populated VMAs this is a little inefficient - chances are
613 * there there won't be many ptes located within the scan cluster. In this case
614 * maybe we could scan further - to the end of the pte page, perhaps.
616 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
617 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
619 static void try_to_unmap_cluster(unsigned long cursor,
620 unsigned int *mapcount, struct vm_area_struct *vma)
622 struct mm_struct *mm = vma->vm_mm;
628 unsigned long address;
633 * We need the page_table_lock to protect us from page faults,
634 * munmap, fork, etc...
636 spin_lock(&mm->page_table_lock);
638 address = (vma->vm_start + cursor) & CLUSTER_MASK;
639 end = address + CLUSTER_SIZE;
640 if (address < vma->vm_start)
641 address = vma->vm_start;
642 if (end > vma->vm_end)
645 pgd = pgd_offset(mm, address);
646 if (!pgd_present(*pgd))
649 pmd = pmd_offset(pgd, address);
650 if (!pmd_present(*pmd))
653 for (pte = pte_offset_map(pmd, address);
654 address < end; pte++, address += PAGE_SIZE) {
656 if (!pte_present(*pte))
663 page = pfn_to_page(pfn);
664 BUG_ON(PageAnon(page));
665 if (PageReserved(page))
668 if (ptep_clear_flush_young(vma, address, pte))
671 /* Nuke the page table entry. */
672 flush_cache_page(vma, address);
673 pteval = ptep_clear_flush(vma, address, pte);
675 /* If nonlinear, store the file page offset in the pte. */
676 if (page->index != linear_page_index(vma, address))
677 set_pte(pte, pgoff_to_pte(page->index));
679 /* Move the dirty bit to the physical page now the pte is gone. */
680 if (pte_dirty(pteval))
681 set_page_dirty(page);
683 page_remove_rmap(page);
684 page_cache_release(page);
693 spin_unlock(&mm->page_table_lock);
696 static int try_to_unmap_anon(struct page *page)
698 struct anon_vma *anon_vma;
699 struct vm_area_struct *vma;
700 int ret = SWAP_AGAIN;
702 anon_vma = page_lock_anon_vma(page);
706 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
707 ret = try_to_unmap_one(page, vma);
708 if (ret == SWAP_FAIL || !page_mapped(page))
711 spin_unlock(&anon_vma->lock);
716 * try_to_unmap_file - unmap file page using the object-based rmap method
717 * @page: the page to unmap
719 * Find all the mappings of a page using the mapping pointer and the vma chains
720 * contained in the address_space struct it points to.
722 * This function is only called from try_to_unmap for object-based pages.
724 static int try_to_unmap_file(struct page *page)
726 struct address_space *mapping = page->mapping;
727 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
728 struct vm_area_struct *vma;
729 struct prio_tree_iter iter;
730 int ret = SWAP_AGAIN;
731 unsigned long cursor;
732 unsigned long max_nl_cursor = 0;
733 unsigned long max_nl_size = 0;
734 unsigned int mapcount;
736 spin_lock(&mapping->i_mmap_lock);
737 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
738 ret = try_to_unmap_one(page, vma);
739 if (ret == SWAP_FAIL || !page_mapped(page))
743 if (list_empty(&mapping->i_mmap_nonlinear))
746 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
747 shared.vm_set.list) {
748 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
750 cursor = (unsigned long) vma->vm_private_data;
751 if (cursor > max_nl_cursor)
752 max_nl_cursor = cursor;
753 cursor = vma->vm_end - vma->vm_start;
754 if (cursor > max_nl_size)
755 max_nl_size = cursor;
758 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
764 * We don't try to search for this page in the nonlinear vmas,
765 * and page_referenced wouldn't have found it anyway. Instead
766 * just walk the nonlinear vmas trying to age and unmap some.
767 * The mapcount of the page we came in with is irrelevant,
768 * but even so use it as a guide to how hard we should try?
770 mapcount = page_mapcount(page);
773 cond_resched_lock(&mapping->i_mmap_lock);
775 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
776 if (max_nl_cursor == 0)
777 max_nl_cursor = CLUSTER_SIZE;
780 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
781 shared.vm_set.list) {
782 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
784 cursor = (unsigned long) vma->vm_private_data;
785 while (vma->vm_mm->rss &&
786 cursor < max_nl_cursor &&
787 cursor < vma->vm_end - vma->vm_start) {
788 try_to_unmap_cluster(cursor, &mapcount, vma);
789 cursor += CLUSTER_SIZE;
790 vma->vm_private_data = (void *) cursor;
791 if ((int)mapcount <= 0)
794 vma->vm_private_data = (void *) max_nl_cursor;
796 cond_resched_lock(&mapping->i_mmap_lock);
797 max_nl_cursor += CLUSTER_SIZE;
798 } while (max_nl_cursor <= max_nl_size);
801 * Don't loop forever (perhaps all the remaining pages are
802 * in locked vmas). Reset cursor on all unreserved nonlinear
803 * vmas, now forgetting on which ones it had fallen behind.
805 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
806 shared.vm_set.list) {
807 if (!(vma->vm_flags & VM_RESERVED))
808 vma->vm_private_data = NULL;
811 spin_unlock(&mapping->i_mmap_lock);
816 * try_to_unmap - try to remove all page table mappings to a page
817 * @page: the page to get unmapped
819 * Tries to remove all the page table entries which are mapping this
820 * page, used in the pageout path. Caller must hold the page lock.
823 * SWAP_SUCCESS - we succeeded in removing all mappings
824 * SWAP_AGAIN - we missed a mapping, try again later
825 * SWAP_FAIL - the page is unswappable
827 int try_to_unmap(struct page *page)
831 BUG_ON(PageReserved(page));
832 BUG_ON(!PageLocked(page));
835 ret = try_to_unmap_anon(page);
837 ret = try_to_unmap_file(page);
839 if (!page_mapped(page))