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_mutex (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_mutex; in sys_msync, i_mutex nests within
28 * down_read of mmap_sem; i_mutex and down_write of mmap_sem are never
29 * taken together; in truncation, i_mutex is taken outermost.
32 * page->flags PG_locked (lock_page)
33 * mapping->i_mmap_lock
35 * mm->page_table_lock or pte_lock
36 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
37 * swap_lock (in swap_duplicate, swap_info_get)
38 * mmlist_lock (in mmput, drain_mmlist and others)
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/module.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 on anon pages, by unuse_vma;
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)) {
239 vma->vm_file->f_mapping != page->mapping)
243 return vma_address(page, vma);
247 * Check that @page is mapped at @address into @mm.
249 * On success returns with pte mapped and locked.
251 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
252 unsigned long address, spinlock_t **ptlp)
260 pgd = pgd_offset(mm, address);
261 if (!pgd_present(*pgd))
264 pud = pud_offset(pgd, address);
265 if (!pud_present(*pud))
268 pmd = pmd_offset(pud, address);
269 if (!pmd_present(*pmd))
272 pte = pte_offset_map(pmd, address);
273 /* Make a quick check before getting the lock */
274 if (!pte_present(*pte)) {
279 ptl = pte_lockptr(mm, pmd);
281 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
285 pte_unmap_unlock(pte, ptl);
290 * Subfunctions of page_referenced: page_referenced_one called
291 * repeatedly from either page_referenced_anon or page_referenced_file.
293 static int page_referenced_one(struct page *page,
294 struct vm_area_struct *vma, unsigned int *mapcount)
296 struct mm_struct *mm = vma->vm_mm;
297 unsigned long address;
302 address = vma_address(page, vma);
303 if (address == -EFAULT)
306 pte = page_check_address(page, mm, address, &ptl);
310 if (ptep_clear_flush_young(vma, address, pte))
313 /* Pretend the page is referenced if the task has the
314 swap token and is in the middle of a page fault. */
315 if (mm != current->mm && has_swap_token(mm) &&
316 rwsem_is_locked(&mm->mmap_sem))
320 pte_unmap_unlock(pte, ptl);
325 static int page_referenced_anon(struct page *page)
327 unsigned int mapcount;
328 struct anon_vma *anon_vma;
329 struct vm_area_struct *vma;
332 anon_vma = page_lock_anon_vma(page);
336 mapcount = page_mapcount(page);
337 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
338 referenced += page_referenced_one(page, vma, &mapcount);
342 spin_unlock(&anon_vma->lock);
347 * page_referenced_file - referenced check for object-based rmap
348 * @page: the page we're checking references on.
350 * For an object-based mapped page, find all the places it is mapped and
351 * check/clear the referenced flag. This is done by following the page->mapping
352 * pointer, then walking the chain of vmas it holds. It returns the number
353 * of references it found.
355 * This function is only called from page_referenced for object-based pages.
357 static int page_referenced_file(struct page *page)
359 unsigned int mapcount;
360 struct address_space *mapping = page->mapping;
361 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
362 struct vm_area_struct *vma;
363 struct prio_tree_iter iter;
367 * The caller's checks on page->mapping and !PageAnon have made
368 * sure that this is a file page: the check for page->mapping
369 * excludes the case just before it gets set on an anon page.
371 BUG_ON(PageAnon(page));
374 * The page lock not only makes sure that page->mapping cannot
375 * suddenly be NULLified by truncation, it makes sure that the
376 * structure at mapping cannot be freed and reused yet,
377 * so we can safely take mapping->i_mmap_lock.
379 BUG_ON(!PageLocked(page));
381 spin_lock(&mapping->i_mmap_lock);
384 * i_mmap_lock does not stabilize mapcount at all, but mapcount
385 * is more likely to be accurate if we note it after spinning.
387 mapcount = page_mapcount(page);
389 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
390 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
391 == (VM_LOCKED|VM_MAYSHARE)) {
395 referenced += page_referenced_one(page, vma, &mapcount);
400 spin_unlock(&mapping->i_mmap_lock);
405 * page_referenced - test if the page was referenced
406 * @page: the page to test
407 * @is_locked: caller holds lock on the page
409 * Quick test_and_clear_referenced for all mappings to a page,
410 * returns the number of ptes which referenced the page.
412 int page_referenced(struct page *page, int is_locked)
416 if (page_test_and_clear_young(page))
419 if (TestClearPageReferenced(page))
422 if (page_mapped(page) && page->mapping) {
424 referenced += page_referenced_anon(page);
426 referenced += page_referenced_file(page);
427 else if (TestSetPageLocked(page))
431 referenced += page_referenced_file(page);
439 * page_set_anon_rmap - setup new anonymous rmap
440 * @page: the page to add the mapping to
441 * @vma: the vm area in which the mapping is added
442 * @address: the user virtual address mapped
444 static void __page_set_anon_rmap(struct page *page,
445 struct vm_area_struct *vma, unsigned long address)
447 struct anon_vma *anon_vma = vma->anon_vma;
450 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
451 page->mapping = (struct address_space *) anon_vma;
453 page->index = linear_page_index(vma, address);
456 * nr_mapped state can be updated without turning off
457 * interrupts because it is not modified via interrupt.
459 __inc_page_state(nr_mapped);
463 * page_add_anon_rmap - add pte mapping to an anonymous page
464 * @page: the page to add the mapping to
465 * @vma: the vm area in which the mapping is added
466 * @address: the user virtual address mapped
468 * The caller needs to hold the pte lock.
470 void page_add_anon_rmap(struct page *page,
471 struct vm_area_struct *vma, unsigned long address)
473 if (atomic_inc_and_test(&page->_mapcount))
474 __page_set_anon_rmap(page, vma, address);
475 /* else checking page index and mapping is racy */
479 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
480 * @page: the page to add the mapping to
481 * @vma: the vm area in which the mapping is added
482 * @address: the user virtual address mapped
484 * Same as page_add_anon_rmap but must only be called on *new* pages.
485 * This means the inc-and-test can be bypassed.
487 void page_add_new_anon_rmap(struct page *page,
488 struct vm_area_struct *vma, unsigned long address)
490 atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */
491 __page_set_anon_rmap(page, vma, address);
495 * page_add_file_rmap - add pte mapping to a file page
496 * @page: the page to add the mapping to
498 * The caller needs to hold the pte lock.
500 void page_add_file_rmap(struct page *page)
502 BUG_ON(PageAnon(page));
503 BUG_ON(!pfn_valid(page_to_pfn(page)));
505 if (atomic_inc_and_test(&page->_mapcount))
506 __inc_page_state(nr_mapped);
510 * page_remove_rmap - take down pte mapping from a page
511 * @page: page to remove mapping from
513 * The caller needs to hold the pte lock.
515 void page_remove_rmap(struct page *page)
517 if (atomic_add_negative(-1, &page->_mapcount)) {
518 if (page_mapcount(page) < 0) {
519 printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page));
520 printk (KERN_EMERG " page->flags = %lx\n", page->flags);
521 printk (KERN_EMERG " page->count = %x\n", page_count(page));
522 printk (KERN_EMERG " page->mapping = %p\n", page->mapping);
525 BUG_ON(page_mapcount(page) < 0);
527 * It would be tidy to reset the PageAnon mapping here,
528 * but that might overwrite a racing page_add_anon_rmap
529 * which increments mapcount after us but sets mapping
530 * before us: so leave the reset to free_hot_cold_page,
531 * and remember that it's only reliable while mapped.
532 * Leaving it set also helps swapoff to reinstate ptes
533 * faster for those pages still in swapcache.
535 if (page_test_and_clear_dirty(page))
536 set_page_dirty(page);
537 __dec_page_state(nr_mapped);
542 * Subfunctions of try_to_unmap: try_to_unmap_one called
543 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
545 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
548 struct mm_struct *mm = vma->vm_mm;
549 unsigned long address;
553 int ret = SWAP_AGAIN;
555 address = vma_address(page, vma);
556 if (address == -EFAULT)
559 pte = page_check_address(page, mm, address, &ptl);
564 * If the page is mlock()d, we cannot swap it out.
565 * If it's recently referenced (perhaps page_referenced
566 * skipped over this mm) then we should reactivate it.
568 if ((vma->vm_flags & VM_LOCKED) ||
569 (ptep_clear_flush_young(vma, address, pte)
575 /* Nuke the page table entry. */
576 flush_cache_page(vma, address, page_to_pfn(page));
577 pteval = ptep_clear_flush(vma, address, pte);
579 /* Move the dirty bit to the physical page now the pte is gone. */
580 if (pte_dirty(pteval))
581 set_page_dirty(page);
583 /* Update high watermark before we lower rss */
584 update_hiwater_rss(mm);
586 if (PageAnon(page)) {
587 swp_entry_t entry = { .val = page_private(page) };
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 if (list_empty(&mm->mmlist))
597 list_add(&mm->mmlist, &init_mm.mmlist);
598 spin_unlock(&mmlist_lock);
600 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
601 BUG_ON(pte_file(*pte));
602 dec_mm_counter(mm, anon_rss);
604 dec_mm_counter(mm, file_rss);
606 page_remove_rmap(page);
607 page_cache_release(page);
610 pte_unmap_unlock(pte, ptl);
616 * objrmap doesn't work for nonlinear VMAs because the assumption that
617 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
618 * Consequently, given a particular page and its ->index, we cannot locate the
619 * ptes which are mapping that page without an exhaustive linear search.
621 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
622 * maps the file to which the target page belongs. The ->vm_private_data field
623 * holds the current cursor into that scan. Successive searches will circulate
624 * around the vma's virtual address space.
626 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
627 * more scanning pressure is placed against them as well. Eventually pages
628 * will become fully unmapped and are eligible for eviction.
630 * For very sparsely populated VMAs this is a little inefficient - chances are
631 * there there won't be many ptes located within the scan cluster. In this case
632 * maybe we could scan further - to the end of the pte page, perhaps.
634 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
635 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
637 static void try_to_unmap_cluster(unsigned long cursor,
638 unsigned int *mapcount, struct vm_area_struct *vma)
640 struct mm_struct *mm = vma->vm_mm;
648 unsigned long address;
651 address = (vma->vm_start + cursor) & CLUSTER_MASK;
652 end = address + CLUSTER_SIZE;
653 if (address < vma->vm_start)
654 address = vma->vm_start;
655 if (end > vma->vm_end)
658 pgd = pgd_offset(mm, address);
659 if (!pgd_present(*pgd))
662 pud = pud_offset(pgd, address);
663 if (!pud_present(*pud))
666 pmd = pmd_offset(pud, address);
667 if (!pmd_present(*pmd))
670 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
672 /* Update high watermark before we lower rss */
673 update_hiwater_rss(mm);
675 for (; address < end; pte++, address += PAGE_SIZE) {
676 if (!pte_present(*pte))
678 page = vm_normal_page(vma, address, *pte);
679 BUG_ON(!page || PageAnon(page));
681 if (ptep_clear_flush_young(vma, address, pte))
684 /* Nuke the page table entry. */
685 flush_cache_page(vma, address, pte_pfn(*pte));
686 pteval = ptep_clear_flush(vma, address, pte);
688 /* If nonlinear, store the file page offset in the pte. */
689 if (page->index != linear_page_index(vma, address))
690 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
692 /* Move the dirty bit to the physical page now the pte is gone. */
693 if (pte_dirty(pteval))
694 set_page_dirty(page);
696 page_remove_rmap(page);
697 page_cache_release(page);
698 dec_mm_counter(mm, file_rss);
701 pte_unmap_unlock(pte - 1, ptl);
704 static int try_to_unmap_anon(struct page *page, int ignore_refs)
706 struct anon_vma *anon_vma;
707 struct vm_area_struct *vma;
708 int ret = SWAP_AGAIN;
710 anon_vma = page_lock_anon_vma(page);
714 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
715 ret = try_to_unmap_one(page, vma, ignore_refs);
716 if (ret == SWAP_FAIL || !page_mapped(page))
719 spin_unlock(&anon_vma->lock);
724 * try_to_unmap_file - unmap file page using the object-based rmap method
725 * @page: the page to unmap
727 * Find all the mappings of a page using the mapping pointer and the vma chains
728 * contained in the address_space struct it points to.
730 * This function is only called from try_to_unmap for object-based pages.
732 static int try_to_unmap_file(struct page *page, int ignore_refs)
734 struct address_space *mapping = page->mapping;
735 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
736 struct vm_area_struct *vma;
737 struct prio_tree_iter iter;
738 int ret = SWAP_AGAIN;
739 unsigned long cursor;
740 unsigned long max_nl_cursor = 0;
741 unsigned long max_nl_size = 0;
742 unsigned int mapcount;
744 spin_lock(&mapping->i_mmap_lock);
745 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
746 ret = try_to_unmap_one(page, vma, ignore_refs);
747 if (ret == SWAP_FAIL || !page_mapped(page))
751 if (list_empty(&mapping->i_mmap_nonlinear))
754 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
755 shared.vm_set.list) {
756 if (vma->vm_flags & VM_LOCKED)
758 cursor = (unsigned long) vma->vm_private_data;
759 if (cursor > max_nl_cursor)
760 max_nl_cursor = cursor;
761 cursor = vma->vm_end - vma->vm_start;
762 if (cursor > max_nl_size)
763 max_nl_size = cursor;
766 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
772 * We don't try to search for this page in the nonlinear vmas,
773 * and page_referenced wouldn't have found it anyway. Instead
774 * just walk the nonlinear vmas trying to age and unmap some.
775 * The mapcount of the page we came in with is irrelevant,
776 * but even so use it as a guide to how hard we should try?
778 mapcount = page_mapcount(page);
781 cond_resched_lock(&mapping->i_mmap_lock);
783 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
784 if (max_nl_cursor == 0)
785 max_nl_cursor = CLUSTER_SIZE;
788 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
789 shared.vm_set.list) {
790 if (vma->vm_flags & VM_LOCKED)
792 cursor = (unsigned long) vma->vm_private_data;
793 while ( 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, shared.vm_set.list)
813 vma->vm_private_data = NULL;
815 spin_unlock(&mapping->i_mmap_lock);
820 * try_to_unmap - try to remove all page table mappings to a page
821 * @page: the page to get unmapped
823 * Tries to remove all the page table entries which are mapping this
824 * page, used in the pageout path. Caller must hold the page lock.
827 * SWAP_SUCCESS - we succeeded in removing all mappings
828 * SWAP_AGAIN - we missed a mapping, try again later
829 * SWAP_FAIL - the page is unswappable
831 int try_to_unmap(struct page *page, int ignore_refs)
835 BUG_ON(!PageLocked(page));
838 ret = try_to_unmap_anon(page, ignore_refs);
840 ret = try_to_unmap_file(page, ignore_refs);
842 if (!page_mapped(page))