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>
56 #include <linux/rcupdate.h>
58 #include <asm/tlbflush.h>
60 //#define RMAP_DEBUG /* can be enabled only for debugging */
62 kmem_cache_t *anon_vma_cachep;
64 static inline void validate_anon_vma(struct vm_area_struct *find_vma)
67 struct anon_vma *anon_vma = find_vma->anon_vma;
68 struct vm_area_struct *vma;
69 unsigned int mapcount = 0;
72 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
74 BUG_ON(mapcount > 100000);
82 /* This must be called under the mmap_sem. */
83 int anon_vma_prepare(struct vm_area_struct *vma)
85 struct anon_vma *anon_vma = vma->anon_vma;
88 if (unlikely(!anon_vma)) {
89 struct mm_struct *mm = vma->vm_mm;
90 struct anon_vma *allocated, *locked;
92 anon_vma = find_mergeable_anon_vma(vma);
96 spin_lock(&locked->lock);
98 anon_vma = anon_vma_alloc();
99 if (unlikely(!anon_vma))
101 allocated = anon_vma;
105 /* page_table_lock to protect against threads */
106 spin_lock(&mm->page_table_lock);
107 if (likely(!vma->anon_vma)) {
108 vma->anon_vma = anon_vma;
109 list_add(&vma->anon_vma_node, &anon_vma->head);
112 spin_unlock(&mm->page_table_lock);
115 spin_unlock(&locked->lock);
116 if (unlikely(allocated))
117 anon_vma_free(allocated);
122 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
124 if (!vma->anon_vma) {
125 BUG_ON(!next->anon_vma);
126 vma->anon_vma = next->anon_vma;
127 list_add(&vma->anon_vma_node, &next->anon_vma_node);
129 /* if they're both non-null they must be the same */
130 BUG_ON(vma->anon_vma != next->anon_vma);
132 list_del(&next->anon_vma_node);
135 void __anon_vma_link(struct vm_area_struct *vma)
137 struct anon_vma *anon_vma = vma->anon_vma;
140 list_add(&vma->anon_vma_node, &anon_vma->head);
141 validate_anon_vma(vma);
145 void anon_vma_link(struct vm_area_struct *vma)
147 struct anon_vma *anon_vma = vma->anon_vma;
150 spin_lock(&anon_vma->lock);
151 list_add(&vma->anon_vma_node, &anon_vma->head);
152 validate_anon_vma(vma);
153 spin_unlock(&anon_vma->lock);
157 void anon_vma_unlink(struct vm_area_struct *vma)
159 struct anon_vma *anon_vma = vma->anon_vma;
165 spin_lock(&anon_vma->lock);
166 validate_anon_vma(vma);
167 list_del(&vma->anon_vma_node);
169 /* We must garbage collect the anon_vma if it's empty */
170 empty = list_empty(&anon_vma->head);
171 spin_unlock(&anon_vma->lock);
174 anon_vma_free(anon_vma);
177 static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
179 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
180 SLAB_CTOR_CONSTRUCTOR) {
181 struct anon_vma *anon_vma = data;
183 spin_lock_init(&anon_vma->lock);
184 INIT_LIST_HEAD(&anon_vma->head);
188 void __init anon_vma_init(void)
190 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
191 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
195 * Getting a lock on a stable anon_vma from a page off the LRU is
196 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
198 static struct anon_vma *page_lock_anon_vma(struct page *page)
200 struct anon_vma *anon_vma = NULL;
201 unsigned long anon_mapping;
204 anon_mapping = (unsigned long) page->mapping;
205 if (!(anon_mapping & PAGE_MAPPING_ANON))
207 if (!page_mapped(page))
210 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
211 spin_lock(&anon_vma->lock);
218 * At what user virtual address is page expected in vma?
220 static inline unsigned long
221 vma_address(struct page *page, struct vm_area_struct *vma)
223 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
224 unsigned long address;
226 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
227 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
228 /* page should be within any vma from prio_tree_next */
229 BUG_ON(!PageAnon(page));
236 * At what user virtual address is page expected in vma? checking that the
237 * page matches the vma: currently only used by unuse_process, on anon pages.
239 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
241 if (PageAnon(page)) {
242 if ((void *)vma->anon_vma !=
243 (void *)page->mapping - PAGE_MAPPING_ANON)
245 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
246 if (vma->vm_file->f_mapping != page->mapping)
250 return vma_address(page, vma);
254 * Subfunctions of page_referenced: page_referenced_one called
255 * repeatedly from either page_referenced_anon or page_referenced_file.
257 static int page_referenced_one(struct page *page,
258 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
260 struct mm_struct *mm = vma->vm_mm;
261 unsigned long address;
269 address = vma_address(page, vma);
270 if (address == -EFAULT)
273 spin_lock(&mm->page_table_lock);
275 pgd = pgd_offset(mm, address);
276 if (!pgd_present(*pgd))
279 pmd = pmd_offset(pgd, address);
280 if (!pmd_present(*pmd))
283 pte = pte_offset_map(pmd, address);
284 if (!pte_present(*pte))
287 if (page_to_pfn(page) != pte_pfn(*pte))
290 if (ptep_clear_flush_young(vma, address, pte))
293 if (mm != current->mm && !ignore_token && has_swap_token(mm))
301 spin_unlock(&mm->page_table_lock);
306 static int page_referenced_anon(struct page *page, int ignore_token)
308 unsigned int mapcount;
309 struct anon_vma *anon_vma;
310 struct vm_area_struct *vma;
313 anon_vma = page_lock_anon_vma(page);
317 mapcount = page_mapcount(page);
318 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
319 referenced += page_referenced_one(page, vma, &mapcount,
324 spin_unlock(&anon_vma->lock);
329 * page_referenced_file - referenced check for object-based rmap
330 * @page: the page we're checking references on.
332 * For an object-based mapped page, find all the places it is mapped and
333 * check/clear the referenced flag. This is done by following the page->mapping
334 * pointer, then walking the chain of vmas it holds. It returns the number
335 * of references it found.
337 * This function is only called from page_referenced for object-based pages.
339 static int page_referenced_file(struct page *page, int ignore_token)
341 unsigned int mapcount;
342 struct address_space *mapping = page->mapping;
343 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
344 struct vm_area_struct *vma;
345 struct prio_tree_iter iter;
349 * The caller's checks on page->mapping and !PageAnon have made
350 * sure that this is a file page: the check for page->mapping
351 * excludes the case just before it gets set on an anon page.
353 BUG_ON(PageAnon(page));
356 * The page lock not only makes sure that page->mapping cannot
357 * suddenly be NULLified by truncation, it makes sure that the
358 * structure at mapping cannot be freed and reused yet,
359 * so we can safely take mapping->i_mmap_lock.
361 BUG_ON(!PageLocked(page));
363 spin_lock(&mapping->i_mmap_lock);
366 * i_mmap_lock does not stabilize mapcount at all, but mapcount
367 * is more likely to be accurate if we note it after spinning.
369 mapcount = page_mapcount(page);
371 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
372 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
373 == (VM_LOCKED|VM_MAYSHARE)) {
377 referenced += page_referenced_one(page, vma, &mapcount,
383 spin_unlock(&mapping->i_mmap_lock);
388 * page_referenced - test if the page was referenced
389 * @page: the page to test
390 * @is_locked: caller holds lock on the page
392 * Quick test_and_clear_referenced for all mappings to a page,
393 * returns the number of ptes which referenced the page.
395 int page_referenced(struct page *page, int is_locked, int ignore_token)
399 if (page_test_and_clear_young(page))
402 if (TestClearPageReferenced(page))
405 if (page_mapped(page) && page->mapping) {
407 referenced += page_referenced_anon(page, ignore_token);
409 referenced += page_referenced_file(page, ignore_token);
410 else if (TestSetPageLocked(page))
414 referenced += page_referenced_file(page,
423 * page_add_anon_rmap - add pte mapping to an anonymous page
424 * @page: the page to add the mapping to
425 * @vma: the vm area in which the mapping is added
426 * @address: the user virtual address mapped
428 * The caller needs to hold the mm->page_table_lock.
430 void page_add_anon_rmap(struct page *page,
431 struct vm_area_struct *vma, unsigned long address)
433 struct anon_vma *anon_vma = vma->anon_vma;
436 BUG_ON(PageReserved(page));
439 vma->vm_mm->anon_rss++;
441 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
442 index = (address - vma->vm_start) >> PAGE_SHIFT;
443 index += vma->vm_pgoff;
444 index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
446 if (atomic_inc_and_test(&page->_mapcount)) {
448 page->mapping = (struct address_space *) anon_vma;
449 inc_page_state(nr_mapped);
451 /* else checking page index and mapping is racy */
455 * page_add_file_rmap - add pte mapping to a file page
456 * @page: the page to add the mapping to
458 * The caller needs to hold the mm->page_table_lock.
460 void page_add_file_rmap(struct page *page)
462 BUG_ON(PageAnon(page));
463 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
466 if (atomic_inc_and_test(&page->_mapcount))
467 inc_page_state(nr_mapped);
471 * page_remove_rmap - take down pte mapping from a page
472 * @page: page to remove mapping from
474 * Caller needs to hold the mm->page_table_lock.
476 void page_remove_rmap(struct page *page)
478 BUG_ON(PageReserved(page));
480 if (atomic_add_negative(-1, &page->_mapcount)) {
481 BUG_ON(page_mapcount(page) < 0);
483 * It would be tidy to reset the PageAnon mapping here,
484 * but that might overwrite a racing page_add_anon_rmap
485 * which increments mapcount after us but sets mapping
486 * before us: so leave the reset to free_hot_cold_page,
487 * and remember that it's only reliable while mapped.
488 * Leaving it set also helps swapoff to reinstate ptes
489 * faster for those pages still in swapcache.
491 if (page_test_and_clear_dirty(page))
492 set_page_dirty(page);
493 dec_page_state(nr_mapped);
498 * Subfunctions of try_to_unmap: try_to_unmap_one called
499 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
501 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
503 struct mm_struct *mm = vma->vm_mm;
504 unsigned long address;
509 int ret = SWAP_AGAIN;
513 address = vma_address(page, vma);
514 if (address == -EFAULT)
518 * We need the page_table_lock to protect us from page faults,
519 * munmap, fork, etc...
521 spin_lock(&mm->page_table_lock);
523 pgd = pgd_offset(mm, address);
524 if (!pgd_present(*pgd))
527 pmd = pmd_offset(pgd, address);
528 if (!pmd_present(*pmd))
531 pte = pte_offset_map(pmd, address);
532 if (!pte_present(*pte))
535 if (page_to_pfn(page) != pte_pfn(*pte))
539 * If the page is mlock()d, we cannot swap it out.
540 * If it's recently referenced (perhaps page_referenced
541 * skipped over this mm) then we should reactivate it.
543 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
544 ptep_clear_flush_young(vma, address, pte)) {
550 * Don't pull an anonymous page out from under get_user_pages.
551 * GUP carefully breaks COW and raises page count (while holding
552 * page_table_lock, as we have here) to make sure that the page
553 * cannot be freed. If we unmap that page here, a user write
554 * access to the virtual address will bring back the page, but
555 * its raised count will (ironically) be taken to mean it's not
556 * an exclusive swap page, do_wp_page will replace it by a copy
557 * page, and the user never get to see the data GUP was holding
558 * the original page for.
560 * This test is also useful for when swapoff (unuse_process) has
561 * to drop page lock: its reference to the page stops existing
562 * ptes from being unmapped, so swapoff can make progress.
564 if (PageSwapCache(page) &&
565 page_count(page) != page_mapcount(page) + 2) {
570 /* Nuke the page table entry. */
571 flush_cache_page(vma, address);
572 pteval = ptep_clear_flush(vma, address, pte);
574 /* Move the dirty bit to the physical page now the pte is gone. */
575 if (pte_dirty(pteval))
576 set_page_dirty(page);
578 if (PageAnon(page)) {
579 swp_entry_t entry = { .val = page->private };
581 * Store the swap location in the pte.
582 * See handle_pte_fault() ...
584 BUG_ON(!PageSwapCache(page));
585 swap_duplicate(entry);
586 set_pte(pte, swp_entry_to_pte(entry));
587 BUG_ON(pte_file(*pte));
593 page_remove_rmap(page);
594 page_cache_release(page);
599 spin_unlock(&mm->page_table_lock);
605 * objrmap doesn't work for nonlinear VMAs because the assumption that
606 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
607 * Consequently, given a particular page and its ->index, we cannot locate the
608 * ptes which are mapping that page without an exhaustive linear search.
610 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
611 * maps the file to which the target page belongs. The ->vm_private_data field
612 * holds the current cursor into that scan. Successive searches will circulate
613 * around the vma's virtual address space.
615 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
616 * more scanning pressure is placed against them as well. Eventually pages
617 * will become fully unmapped and are eligible for eviction.
619 * For very sparsely populated VMAs this is a little inefficient - chances are
620 * there there won't be many ptes located within the scan cluster. In this case
621 * maybe we could scan further - to the end of the pte page, perhaps.
623 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
624 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
626 static void try_to_unmap_cluster(unsigned long cursor,
627 unsigned int *mapcount, struct vm_area_struct *vma)
629 struct mm_struct *mm = vma->vm_mm;
635 unsigned long address;
640 * We need the page_table_lock to protect us from page faults,
641 * munmap, fork, etc...
643 spin_lock(&mm->page_table_lock);
645 address = (vma->vm_start + cursor) & CLUSTER_MASK;
646 end = address + CLUSTER_SIZE;
647 if (address < vma->vm_start)
648 address = vma->vm_start;
649 if (end > vma->vm_end)
652 pgd = pgd_offset(mm, address);
653 if (!pgd_present(*pgd))
656 pmd = pmd_offset(pgd, address);
657 if (!pmd_present(*pmd))
660 for (pte = pte_offset_map(pmd, address);
661 address < end; pte++, address += PAGE_SIZE) {
663 if (!pte_present(*pte))
670 page = pfn_to_page(pfn);
671 BUG_ON(PageAnon(page));
672 if (PageReserved(page))
675 if (ptep_clear_flush_young(vma, address, pte))
678 /* Nuke the page table entry. */
679 flush_cache_page(vma, address);
680 pteval = ptep_clear_flush(vma, address, pte);
682 /* If nonlinear, store the file page offset in the pte. */
683 if (page->index != linear_page_index(vma, address))
684 set_pte(pte, pgoff_to_pte(page->index));
686 /* Move the dirty bit to the physical page now the pte is gone. */
687 if (pte_dirty(pteval))
688 set_page_dirty(page);
690 page_remove_rmap(page);
691 page_cache_release(page);
700 spin_unlock(&mm->page_table_lock);
703 static int try_to_unmap_anon(struct page *page)
705 struct anon_vma *anon_vma;
706 struct vm_area_struct *vma;
707 int ret = SWAP_AGAIN;
709 anon_vma = page_lock_anon_vma(page);
713 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
714 ret = try_to_unmap_one(page, vma);
715 if (ret == SWAP_FAIL || !page_mapped(page))
718 spin_unlock(&anon_vma->lock);
723 * try_to_unmap_file - unmap file page using the object-based rmap method
724 * @page: the page to unmap
726 * Find all the mappings of a page using the mapping pointer and the vma chains
727 * contained in the address_space struct it points to.
729 * This function is only called from try_to_unmap for object-based pages.
731 static int try_to_unmap_file(struct page *page)
733 struct address_space *mapping = page->mapping;
734 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
735 struct vm_area_struct *vma;
736 struct prio_tree_iter iter;
737 int ret = SWAP_AGAIN;
738 unsigned long cursor;
739 unsigned long max_nl_cursor = 0;
740 unsigned long max_nl_size = 0;
741 unsigned int mapcount;
743 spin_lock(&mapping->i_mmap_lock);
744 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
745 ret = try_to_unmap_one(page, vma);
746 if (ret == SWAP_FAIL || !page_mapped(page))
750 if (list_empty(&mapping->i_mmap_nonlinear))
753 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
754 shared.vm_set.list) {
755 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
757 cursor = (unsigned long) vma->vm_private_data;
758 if (cursor > max_nl_cursor)
759 max_nl_cursor = cursor;
760 cursor = vma->vm_end - vma->vm_start;
761 if (cursor > max_nl_size)
762 max_nl_size = cursor;
765 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
771 * We don't try to search for this page in the nonlinear vmas,
772 * and page_referenced wouldn't have found it anyway. Instead
773 * just walk the nonlinear vmas trying to age and unmap some.
774 * The mapcount of the page we came in with is irrelevant,
775 * but even so use it as a guide to how hard we should try?
777 mapcount = page_mapcount(page);
780 cond_resched_lock(&mapping->i_mmap_lock);
782 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
783 if (max_nl_cursor == 0)
784 max_nl_cursor = CLUSTER_SIZE;
787 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
788 shared.vm_set.list) {
789 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
791 cursor = (unsigned long) vma->vm_private_data;
792 while (vma->vm_mm->rss &&
793 cursor < max_nl_cursor &&
794 cursor < vma->vm_end - vma->vm_start) {
795 try_to_unmap_cluster(cursor, &mapcount, vma);
796 cursor += CLUSTER_SIZE;
797 vma->vm_private_data = (void *) cursor;
798 if ((int)mapcount <= 0)
801 vma->vm_private_data = (void *) max_nl_cursor;
803 cond_resched_lock(&mapping->i_mmap_lock);
804 max_nl_cursor += CLUSTER_SIZE;
805 } while (max_nl_cursor <= max_nl_size);
808 * Don't loop forever (perhaps all the remaining pages are
809 * in locked vmas). Reset cursor on all unreserved nonlinear
810 * vmas, now forgetting on which ones it had fallen behind.
812 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
813 shared.vm_set.list) {
814 if (!(vma->vm_flags & VM_RESERVED))
815 vma->vm_private_data = NULL;
818 spin_unlock(&mapping->i_mmap_lock);
823 * try_to_unmap - try to remove all page table mappings to a page
824 * @page: the page to get unmapped
826 * Tries to remove all the page table entries which are mapping this
827 * page, used in the pageout path. Caller must hold the page lock.
830 * SWAP_SUCCESS - we succeeded in removing all mappings
831 * SWAP_AGAIN - we missed a mapping, try again later
832 * SWAP_FAIL - the page is unswappable
834 int try_to_unmap(struct page *page)
838 BUG_ON(PageReserved(page));
839 BUG_ON(!PageLocked(page));
842 ret = try_to_unmap_anon(page);
844 ret = try_to_unmap_file(page);
846 if (!page_mapped(page))