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>
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 BUG_ON(vma->anon_vma != next->anon_vma);
124 list_del(&next->anon_vma_node);
127 void __anon_vma_link(struct vm_area_struct *vma)
129 struct anon_vma *anon_vma = vma->anon_vma;
132 list_add(&vma->anon_vma_node, &anon_vma->head);
133 validate_anon_vma(vma);
137 void anon_vma_link(struct vm_area_struct *vma)
139 struct anon_vma *anon_vma = vma->anon_vma;
142 spin_lock(&anon_vma->lock);
143 list_add(&vma->anon_vma_node, &anon_vma->head);
144 validate_anon_vma(vma);
145 spin_unlock(&anon_vma->lock);
149 void anon_vma_unlink(struct vm_area_struct *vma)
151 struct anon_vma *anon_vma = vma->anon_vma;
157 spin_lock(&anon_vma->lock);
158 validate_anon_vma(vma);
159 list_del(&vma->anon_vma_node);
161 /* We must garbage collect the anon_vma if it's empty */
162 empty = list_empty(&anon_vma->head);
163 spin_unlock(&anon_vma->lock);
166 anon_vma_free(anon_vma);
169 static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
171 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
172 SLAB_CTOR_CONSTRUCTOR) {
173 struct anon_vma *anon_vma = data;
175 spin_lock_init(&anon_vma->lock);
176 INIT_LIST_HEAD(&anon_vma->head);
180 void __init anon_vma_init(void)
182 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
183 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
187 * Getting a lock on a stable anon_vma from a page off the LRU is
188 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
190 static struct anon_vma *page_lock_anon_vma(struct page *page)
192 struct anon_vma *anon_vma = NULL;
193 unsigned long anon_mapping;
196 anon_mapping = (unsigned long) page->mapping;
197 if (!(anon_mapping & PAGE_MAPPING_ANON))
199 if (!page_mapped(page))
202 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
203 spin_lock(&anon_vma->lock);
210 * At what user virtual address is page expected in vma?
212 static inline unsigned long
213 vma_address(struct page *page, struct vm_area_struct *vma)
215 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
216 unsigned long address;
218 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
219 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
220 /* page should be within any vma from prio_tree_next */
221 BUG_ON(!PageAnon(page));
228 * At what user virtual address is page expected in vma? checking that the
229 * page matches the vma: currently only used by unuse_process, on anon pages.
231 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
233 if (PageAnon(page)) {
234 if ((void *)vma->anon_vma !=
235 (void *)page->mapping - PAGE_MAPPING_ANON)
237 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
238 if (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 mapped pte and locked mm->page_table_lock.
250 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
251 unsigned long address)
259 * We need the page_table_lock to protect us from page faults,
260 * munmap, fork, etc...
262 spin_lock(&mm->page_table_lock);
263 pgd = pgd_offset(mm, address);
264 if (likely(pgd_present(*pgd))) {
265 pud = pud_offset(pgd, address);
266 if (likely(pud_present(*pud))) {
267 pmd = pmd_offset(pud, address);
268 if (likely(pmd_present(*pmd))) {
269 pte = pte_offset_map(pmd, address);
270 if (likely(pte_present(*pte) &&
271 page_to_pfn(page) == pte_pfn(*pte)))
277 spin_unlock(&mm->page_table_lock);
278 return ERR_PTR(-ENOENT);
282 * Subfunctions of page_referenced: page_referenced_one called
283 * repeatedly from either page_referenced_anon or page_referenced_file.
285 static int page_referenced_one(struct page *page,
286 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
288 struct mm_struct *mm = vma->vm_mm;
289 unsigned long address;
293 if (!get_mm_counter(mm, rss))
295 address = vma_address(page, vma);
296 if (address == -EFAULT)
299 pte = page_check_address(page, mm, address);
301 if (ptep_clear_flush_young(vma, address, pte))
304 if (mm != current->mm && !ignore_token && has_swap_token(mm))
309 spin_unlock(&mm->page_table_lock);
315 static int page_referenced_anon(struct page *page, int ignore_token)
317 unsigned int mapcount;
318 struct anon_vma *anon_vma;
319 struct vm_area_struct *vma;
322 anon_vma = page_lock_anon_vma(page);
326 mapcount = page_mapcount(page);
327 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
328 referenced += page_referenced_one(page, vma, &mapcount,
333 spin_unlock(&anon_vma->lock);
338 * page_referenced_file - referenced check for object-based rmap
339 * @page: the page we're checking references on.
341 * For an object-based mapped page, find all the places it is mapped and
342 * check/clear the referenced flag. This is done by following the page->mapping
343 * pointer, then walking the chain of vmas it holds. It returns the number
344 * of references it found.
346 * This function is only called from page_referenced for object-based pages.
348 static int page_referenced_file(struct page *page, int ignore_token)
350 unsigned int mapcount;
351 struct address_space *mapping = page->mapping;
352 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
353 struct vm_area_struct *vma;
354 struct prio_tree_iter iter;
358 * The caller's checks on page->mapping and !PageAnon have made
359 * sure that this is a file page: the check for page->mapping
360 * excludes the case just before it gets set on an anon page.
362 BUG_ON(PageAnon(page));
365 * The page lock not only makes sure that page->mapping cannot
366 * suddenly be NULLified by truncation, it makes sure that the
367 * structure at mapping cannot be freed and reused yet,
368 * so we can safely take mapping->i_mmap_lock.
370 BUG_ON(!PageLocked(page));
372 spin_lock(&mapping->i_mmap_lock);
375 * i_mmap_lock does not stabilize mapcount at all, but mapcount
376 * is more likely to be accurate if we note it after spinning.
378 mapcount = page_mapcount(page);
380 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
381 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
382 == (VM_LOCKED|VM_MAYSHARE)) {
386 referenced += page_referenced_one(page, vma, &mapcount,
392 spin_unlock(&mapping->i_mmap_lock);
397 * page_referenced - test if the page was referenced
398 * @page: the page to test
399 * @is_locked: caller holds lock on the page
401 * Quick test_and_clear_referenced for all mappings to a page,
402 * returns the number of ptes which referenced the page.
404 int page_referenced(struct page *page, int is_locked, int ignore_token)
408 if (!swap_token_default_timeout)
411 if (page_test_and_clear_young(page))
414 if (TestClearPageReferenced(page))
417 if (page_mapped(page) && page->mapping) {
419 referenced += page_referenced_anon(page, ignore_token);
421 referenced += page_referenced_file(page, ignore_token);
422 else if (TestSetPageLocked(page))
426 referenced += page_referenced_file(page,
435 * page_add_anon_rmap - add pte mapping to an anonymous page
436 * @page: the page to add the mapping to
437 * @vma: the vm area in which the mapping is added
438 * @address: the user virtual address mapped
440 * The caller needs to hold the mm->page_table_lock.
442 void page_add_anon_rmap(struct page *page,
443 struct vm_area_struct *vma, unsigned long address)
445 struct anon_vma *anon_vma = vma->anon_vma;
448 BUG_ON(PageReserved(page));
451 inc_mm_counter(vma->vm_mm, anon_rss);
453 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
454 index = (address - vma->vm_start) >> PAGE_SHIFT;
455 index += vma->vm_pgoff;
456 index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
458 if (atomic_inc_and_test(&page->_mapcount)) {
460 page->mapping = (struct address_space *) anon_vma;
461 inc_page_state(nr_mapped);
463 /* else checking page index and mapping is racy */
467 * page_add_file_rmap - add pte mapping to a file page
468 * @page: the page to add the mapping to
470 * The caller needs to hold the mm->page_table_lock.
472 void page_add_file_rmap(struct page *page)
474 BUG_ON(PageAnon(page));
475 if (!pfn_valid(page_to_pfn(page)) || PageReserved(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 * Caller needs to hold the mm->page_table_lock.
488 void page_remove_rmap(struct page *page)
490 BUG_ON(PageReserved(page));
492 if (atomic_add_negative(-1, &page->_mapcount)) {
493 BUG_ON(page_mapcount(page) < 0);
495 * It would be tidy to reset the PageAnon mapping here,
496 * but that might overwrite a racing page_add_anon_rmap
497 * which increments mapcount after us but sets mapping
498 * before us: so leave the reset to free_hot_cold_page,
499 * and remember that it's only reliable while mapped.
500 * Leaving it set also helps swapoff to reinstate ptes
501 * faster for those pages still in swapcache.
503 if (page_test_and_clear_dirty(page))
504 set_page_dirty(page);
505 dec_page_state(nr_mapped);
510 * Subfunctions of try_to_unmap: try_to_unmap_one called
511 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
513 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
515 struct mm_struct *mm = vma->vm_mm;
516 unsigned long address;
519 int ret = SWAP_AGAIN;
521 if (!get_mm_counter(mm, rss))
523 address = vma_address(page, vma);
524 if (address == -EFAULT)
527 pte = page_check_address(page, mm, address);
532 * If the page is mlock()d, we cannot swap it out.
533 * If it's recently referenced (perhaps page_referenced
534 * skipped over this mm) then we should reactivate it.
536 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
537 ptep_clear_flush_young(vma, address, pte)) {
542 /* Nuke the page table entry. */
543 flush_cache_page(vma, address, page_to_pfn(page));
544 pteval = ptep_clear_flush(vma, address, pte);
546 /* Move the dirty bit to the physical page now the pte is gone. */
547 if (pte_dirty(pteval))
548 set_page_dirty(page);
550 if (PageAnon(page)) {
551 swp_entry_t entry = { .val = page->private };
553 * Store the swap location in the pte.
554 * See handle_pte_fault() ...
556 BUG_ON(!PageSwapCache(page));
557 swap_duplicate(entry);
558 if (list_empty(&mm->mmlist)) {
559 spin_lock(&mmlist_lock);
560 list_add(&mm->mmlist, &init_mm.mmlist);
561 spin_unlock(&mmlist_lock);
563 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
564 BUG_ON(pte_file(*pte));
565 dec_mm_counter(mm, anon_rss);
568 dec_mm_counter(mm, rss);
569 page_remove_rmap(page);
570 page_cache_release(page);
574 spin_unlock(&mm->page_table_lock);
580 * objrmap doesn't work for nonlinear VMAs because the assumption that
581 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
582 * Consequently, given a particular page and its ->index, we cannot locate the
583 * ptes which are mapping that page without an exhaustive linear search.
585 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
586 * maps the file to which the target page belongs. The ->vm_private_data field
587 * holds the current cursor into that scan. Successive searches will circulate
588 * around the vma's virtual address space.
590 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
591 * more scanning pressure is placed against them as well. Eventually pages
592 * will become fully unmapped and are eligible for eviction.
594 * For very sparsely populated VMAs this is a little inefficient - chances are
595 * there there won't be many ptes located within the scan cluster. In this case
596 * maybe we could scan further - to the end of the pte page, perhaps.
598 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
599 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
601 static void try_to_unmap_cluster(unsigned long cursor,
602 unsigned int *mapcount, struct vm_area_struct *vma)
604 struct mm_struct *mm = vma->vm_mm;
608 pte_t *pte, *original_pte;
611 unsigned long address;
616 * We need the page_table_lock to protect us from page faults,
617 * munmap, fork, etc...
619 spin_lock(&mm->page_table_lock);
621 address = (vma->vm_start + cursor) & CLUSTER_MASK;
622 end = address + CLUSTER_SIZE;
623 if (address < vma->vm_start)
624 address = vma->vm_start;
625 if (end > vma->vm_end)
628 pgd = pgd_offset(mm, address);
629 if (!pgd_present(*pgd))
632 pud = pud_offset(pgd, address);
633 if (!pud_present(*pud))
636 pmd = pmd_offset(pud, address);
637 if (!pmd_present(*pmd))
640 for (original_pte = pte = pte_offset_map(pmd, address);
641 address < end; pte++, address += PAGE_SIZE) {
643 if (!pte_present(*pte))
650 page = pfn_to_page(pfn);
651 BUG_ON(PageAnon(page));
652 if (PageReserved(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, rss);
676 pte_unmap(original_pte);
678 spin_unlock(&mm->page_table_lock);
681 static int try_to_unmap_anon(struct page *page)
683 struct anon_vma *anon_vma;
684 struct vm_area_struct *vma;
685 int ret = SWAP_AGAIN;
687 anon_vma = page_lock_anon_vma(page);
691 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
692 ret = try_to_unmap_one(page, vma);
693 if (ret == SWAP_FAIL || !page_mapped(page))
696 spin_unlock(&anon_vma->lock);
701 * try_to_unmap_file - unmap file page using the object-based rmap method
702 * @page: the page to unmap
704 * Find all the mappings of a page using the mapping pointer and the vma chains
705 * contained in the address_space struct it points to.
707 * This function is only called from try_to_unmap for object-based pages.
709 static int try_to_unmap_file(struct page *page)
711 struct address_space *mapping = page->mapping;
712 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
713 struct vm_area_struct *vma;
714 struct prio_tree_iter iter;
715 int ret = SWAP_AGAIN;
716 unsigned long cursor;
717 unsigned long max_nl_cursor = 0;
718 unsigned long max_nl_size = 0;
719 unsigned int mapcount;
721 spin_lock(&mapping->i_mmap_lock);
722 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
723 ret = try_to_unmap_one(page, vma);
724 if (ret == SWAP_FAIL || !page_mapped(page))
728 if (list_empty(&mapping->i_mmap_nonlinear))
731 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
732 shared.vm_set.list) {
733 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
735 cursor = (unsigned long) vma->vm_private_data;
736 if (cursor > max_nl_cursor)
737 max_nl_cursor = cursor;
738 cursor = vma->vm_end - vma->vm_start;
739 if (cursor > max_nl_size)
740 max_nl_size = cursor;
743 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
749 * We don't try to search for this page in the nonlinear vmas,
750 * and page_referenced wouldn't have found it anyway. Instead
751 * just walk the nonlinear vmas trying to age and unmap some.
752 * The mapcount of the page we came in with is irrelevant,
753 * but even so use it as a guide to how hard we should try?
755 mapcount = page_mapcount(page);
758 cond_resched_lock(&mapping->i_mmap_lock);
760 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
761 if (max_nl_cursor == 0)
762 max_nl_cursor = CLUSTER_SIZE;
765 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
766 shared.vm_set.list) {
767 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
769 cursor = (unsigned long) vma->vm_private_data;
770 while (get_mm_counter(vma->vm_mm, rss) &&
771 cursor < max_nl_cursor &&
772 cursor < vma->vm_end - vma->vm_start) {
773 try_to_unmap_cluster(cursor, &mapcount, vma);
774 cursor += CLUSTER_SIZE;
775 vma->vm_private_data = (void *) cursor;
776 if ((int)mapcount <= 0)
779 vma->vm_private_data = (void *) max_nl_cursor;
781 cond_resched_lock(&mapping->i_mmap_lock);
782 max_nl_cursor += CLUSTER_SIZE;
783 } while (max_nl_cursor <= max_nl_size);
786 * Don't loop forever (perhaps all the remaining pages are
787 * in locked vmas). Reset cursor on all unreserved nonlinear
788 * vmas, now forgetting on which ones it had fallen behind.
790 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
791 shared.vm_set.list) {
792 if (!(vma->vm_flags & VM_RESERVED))
793 vma->vm_private_data = NULL;
796 spin_unlock(&mapping->i_mmap_lock);
801 * try_to_unmap - try to remove all page table mappings to a page
802 * @page: the page to get unmapped
804 * Tries to remove all the page table entries which are mapping this
805 * page, used in the pageout path. Caller must hold the page lock.
808 * SWAP_SUCCESS - we succeeded in removing all mappings
809 * SWAP_AGAIN - we missed a mapping, try again later
810 * SWAP_FAIL - the page is unswappable
812 int try_to_unmap(struct page *page)
816 BUG_ON(PageReserved(page));
817 BUG_ON(!PageLocked(page));
820 ret = try_to_unmap_anon(page);
822 ret = try_to_unmap_file(page);
824 if (!page_mapped(page))