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
35 * mm->page_table_lock or pte_lock
36 * zone->lru_lock (in mark_page_accessed)
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>
56 #include <asm/tlbflush.h>
58 //#define RMAP_DEBUG /* can be enabled only for debugging */
60 kmem_cache_t
*anon_vma_cachep
;
62 static inline void validate_anon_vma(struct vm_area_struct
*find_vma
)
65 struct anon_vma
*anon_vma
= find_vma
->anon_vma
;
66 struct vm_area_struct
*vma
;
67 unsigned int mapcount
= 0;
70 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
72 BUG_ON(mapcount
> 100000);
80 /* This must be called under the mmap_sem. */
81 int anon_vma_prepare(struct vm_area_struct
*vma
)
83 struct anon_vma
*anon_vma
= vma
->anon_vma
;
86 if (unlikely(!anon_vma
)) {
87 struct mm_struct
*mm
= vma
->vm_mm
;
88 struct anon_vma
*allocated
, *locked
;
90 anon_vma
= find_mergeable_anon_vma(vma
);
94 spin_lock(&locked
->lock
);
96 anon_vma
= anon_vma_alloc();
97 if (unlikely(!anon_vma
))
103 /* page_table_lock to protect against threads */
104 spin_lock(&mm
->page_table_lock
);
105 if (likely(!vma
->anon_vma
)) {
106 vma
->anon_vma
= anon_vma
;
107 list_add(&vma
->anon_vma_node
, &anon_vma
->head
);
110 spin_unlock(&mm
->page_table_lock
);
113 spin_unlock(&locked
->lock
);
114 if (unlikely(allocated
))
115 anon_vma_free(allocated
);
120 void __anon_vma_merge(struct vm_area_struct
*vma
, struct vm_area_struct
*next
)
122 BUG_ON(vma
->anon_vma
!= next
->anon_vma
);
123 list_del(&next
->anon_vma_node
);
126 void __anon_vma_link(struct vm_area_struct
*vma
)
128 struct anon_vma
*anon_vma
= vma
->anon_vma
;
131 list_add(&vma
->anon_vma_node
, &anon_vma
->head
);
132 validate_anon_vma(vma
);
136 void anon_vma_link(struct vm_area_struct
*vma
)
138 struct anon_vma
*anon_vma
= vma
->anon_vma
;
141 spin_lock(&anon_vma
->lock
);
142 list_add(&vma
->anon_vma_node
, &anon_vma
->head
);
143 validate_anon_vma(vma
);
144 spin_unlock(&anon_vma
->lock
);
148 void anon_vma_unlink(struct vm_area_struct
*vma
)
150 struct anon_vma
*anon_vma
= vma
->anon_vma
;
156 spin_lock(&anon_vma
->lock
);
157 validate_anon_vma(vma
);
158 list_del(&vma
->anon_vma_node
);
160 /* We must garbage collect the anon_vma if it's empty */
161 empty
= list_empty(&anon_vma
->head
);
162 spin_unlock(&anon_vma
->lock
);
165 anon_vma_free(anon_vma
);
168 static void anon_vma_ctor(void *data
, kmem_cache_t
*cachep
, unsigned long flags
)
170 if ((flags
& (SLAB_CTOR_VERIFY
|SLAB_CTOR_CONSTRUCTOR
)) ==
171 SLAB_CTOR_CONSTRUCTOR
) {
172 struct anon_vma
*anon_vma
= data
;
174 spin_lock_init(&anon_vma
->lock
);
175 INIT_LIST_HEAD(&anon_vma
->head
);
179 void __init
anon_vma_init(void)
181 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
182 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
, NULL
);
186 * Getting a lock on a stable anon_vma from a page off the LRU is
187 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
189 static struct anon_vma
*page_lock_anon_vma(struct page
*page
)
191 struct anon_vma
*anon_vma
= NULL
;
192 unsigned long anon_mapping
;
195 anon_mapping
= (unsigned long) page
->mapping
;
196 if (!(anon_mapping
& PAGE_MAPPING_ANON
))
198 if (!page_mapped(page
))
201 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
202 spin_lock(&anon_vma
->lock
);
209 * At what user virtual address is page expected in vma?
211 static inline unsigned long
212 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
214 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
215 unsigned long address
;
217 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
218 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
219 /* page should be within any vma from prio_tree_next */
220 BUG_ON(!PageAnon(page
));
227 * At what user virtual address is page expected in vma? checking that the
228 * page matches the vma: currently only used on anon pages, by unuse_vma;
229 * and by extraordinary checks on anon pages in VM_UNPAGED vmas, taking
230 * care that an mmap of /dev/mem might window free and foreign pages.
232 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
234 if (PageAnon(page
)) {
235 if ((void *)vma
->anon_vma
!=
236 (void *)page
->mapping
- PAGE_MAPPING_ANON
)
238 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
240 vma
->vm_file
->f_mapping
!= page
->mapping
)
244 return vma_address(page
, vma
);
248 * Check that @page is mapped at @address into @mm.
250 * On success returns with pte mapped and locked.
252 pte_t
*page_check_address(struct page
*page
, struct mm_struct
*mm
,
253 unsigned long address
, spinlock_t
**ptlp
)
261 pgd
= pgd_offset(mm
, address
);
262 if (!pgd_present(*pgd
))
265 pud
= pud_offset(pgd
, address
);
266 if (!pud_present(*pud
))
269 pmd
= pmd_offset(pud
, address
);
270 if (!pmd_present(*pmd
))
273 pte
= pte_offset_map(pmd
, address
);
274 /* Make a quick check before getting the lock */
275 if (!pte_present(*pte
)) {
280 ptl
= pte_lockptr(mm
, pmd
);
282 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
286 pte_unmap_unlock(pte
, ptl
);
291 * Subfunctions of page_referenced: page_referenced_one called
292 * repeatedly from either page_referenced_anon or page_referenced_file.
294 static int page_referenced_one(struct page
*page
,
295 struct vm_area_struct
*vma
, unsigned int *mapcount
, int ignore_token
)
297 struct mm_struct
*mm
= vma
->vm_mm
;
298 unsigned long address
;
303 address
= vma_address(page
, vma
);
304 if (address
== -EFAULT
)
307 pte
= page_check_address(page
, mm
, address
, &ptl
);
311 if (ptep_clear_flush_young(vma
, address
, pte
))
314 /* Pretend the page is referenced if the task has the
315 swap token and is in the middle of a page fault. */
316 if (mm
!= current
->mm
&& !ignore_token
&& has_swap_token(mm
) &&
317 rwsem_is_locked(&mm
->mmap_sem
))
321 pte_unmap_unlock(pte
, ptl
);
326 static int page_referenced_anon(struct page
*page
, int ignore_token
)
328 unsigned int mapcount
;
329 struct anon_vma
*anon_vma
;
330 struct vm_area_struct
*vma
;
333 anon_vma
= page_lock_anon_vma(page
);
337 mapcount
= page_mapcount(page
);
338 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
339 referenced
+= page_referenced_one(page
, vma
, &mapcount
,
344 spin_unlock(&anon_vma
->lock
);
349 * page_referenced_file - referenced check for object-based rmap
350 * @page: the page we're checking references on.
352 * For an object-based mapped page, find all the places it is mapped and
353 * check/clear the referenced flag. This is done by following the page->mapping
354 * pointer, then walking the chain of vmas it holds. It returns the number
355 * of references it found.
357 * This function is only called from page_referenced for object-based pages.
359 static int page_referenced_file(struct page
*page
, int ignore_token
)
361 unsigned int mapcount
;
362 struct address_space
*mapping
= page
->mapping
;
363 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
364 struct vm_area_struct
*vma
;
365 struct prio_tree_iter iter
;
369 * The caller's checks on page->mapping and !PageAnon have made
370 * sure that this is a file page: the check for page->mapping
371 * excludes the case just before it gets set on an anon page.
373 BUG_ON(PageAnon(page
));
376 * The page lock not only makes sure that page->mapping cannot
377 * suddenly be NULLified by truncation, it makes sure that the
378 * structure at mapping cannot be freed and reused yet,
379 * so we can safely take mapping->i_mmap_lock.
381 BUG_ON(!PageLocked(page
));
383 spin_lock(&mapping
->i_mmap_lock
);
386 * i_mmap_lock does not stabilize mapcount at all, but mapcount
387 * is more likely to be accurate if we note it after spinning.
389 mapcount
= page_mapcount(page
);
391 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
392 if ((vma
->vm_flags
& (VM_LOCKED
|VM_MAYSHARE
))
393 == (VM_LOCKED
|VM_MAYSHARE
)) {
397 referenced
+= page_referenced_one(page
, vma
, &mapcount
,
403 spin_unlock(&mapping
->i_mmap_lock
);
408 * page_referenced - test if the page was referenced
409 * @page: the page to test
410 * @is_locked: caller holds lock on the page
412 * Quick test_and_clear_referenced for all mappings to a page,
413 * returns the number of ptes which referenced the page.
415 int page_referenced(struct page
*page
, int is_locked
, int ignore_token
)
419 if (!swap_token_default_timeout
)
422 if (page_test_and_clear_young(page
))
425 if (TestClearPageReferenced(page
))
428 if (page_mapped(page
) && page
->mapping
) {
430 referenced
+= page_referenced_anon(page
, ignore_token
);
432 referenced
+= page_referenced_file(page
, ignore_token
);
433 else if (TestSetPageLocked(page
))
437 referenced
+= page_referenced_file(page
,
446 * page_add_anon_rmap - add pte mapping to an anonymous page
447 * @page: the page to add the mapping to
448 * @vma: the vm area in which the mapping is added
449 * @address: the user virtual address mapped
451 * The caller needs to hold the pte lock.
453 void page_add_anon_rmap(struct page
*page
,
454 struct vm_area_struct
*vma
, unsigned long address
)
456 if (atomic_inc_and_test(&page
->_mapcount
)) {
457 struct anon_vma
*anon_vma
= vma
->anon_vma
;
460 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
461 page
->mapping
= (struct address_space
*) anon_vma
;
463 page
->index
= linear_page_index(vma
, address
);
465 inc_page_state(nr_mapped
);
467 /* else checking page index and mapping is racy */
471 * page_add_file_rmap - add pte mapping to a file page
472 * @page: the page to add the mapping to
474 * The caller needs to hold the pte lock.
476 void page_add_file_rmap(struct page
*page
)
478 BUG_ON(PageAnon(page
));
479 BUG_ON(!pfn_valid(page_to_pfn(page
)));
481 if (atomic_inc_and_test(&page
->_mapcount
))
482 inc_page_state(nr_mapped
);
486 * page_remove_rmap - take down pte mapping from a page
487 * @page: page to remove mapping from
489 * The caller needs to hold the pte lock.
491 void page_remove_rmap(struct page
*page
)
493 if (atomic_add_negative(-1, &page
->_mapcount
)) {
494 BUG_ON(page_mapcount(page
) < 0);
496 * It would be tidy to reset the PageAnon mapping here,
497 * but that might overwrite a racing page_add_anon_rmap
498 * which increments mapcount after us but sets mapping
499 * before us: so leave the reset to free_hot_cold_page,
500 * and remember that it's only reliable while mapped.
501 * Leaving it set also helps swapoff to reinstate ptes
502 * faster for those pages still in swapcache.
504 if (page_test_and_clear_dirty(page
))
505 set_page_dirty(page
);
506 dec_page_state(nr_mapped
);
511 * Subfunctions of try_to_unmap: try_to_unmap_one called
512 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
514 static int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
)
516 struct mm_struct
*mm
= vma
->vm_mm
;
517 unsigned long address
;
521 int ret
= SWAP_AGAIN
;
523 address
= vma_address(page
, vma
);
524 if (address
== -EFAULT
)
527 pte
= page_check_address(page
, mm
, address
, &ptl
);
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
) ||
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 /* Update high watermark before we lower rss */
551 update_hiwater_rss(mm
);
553 if (PageAnon(page
)) {
554 swp_entry_t entry
= { .val
= page_private(page
) };
556 * Store the swap location in the pte.
557 * See handle_pte_fault() ...
559 BUG_ON(!PageSwapCache(page
));
560 swap_duplicate(entry
);
561 if (list_empty(&mm
->mmlist
)) {
562 spin_lock(&mmlist_lock
);
563 if (list_empty(&mm
->mmlist
))
564 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
565 spin_unlock(&mmlist_lock
);
567 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
568 BUG_ON(pte_file(*pte
));
569 dec_mm_counter(mm
, anon_rss
);
571 dec_mm_counter(mm
, file_rss
);
573 page_remove_rmap(page
);
574 page_cache_release(page
);
577 pte_unmap_unlock(pte
, ptl
);
583 * objrmap doesn't work for nonlinear VMAs because the assumption that
584 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
585 * Consequently, given a particular page and its ->index, we cannot locate the
586 * ptes which are mapping that page without an exhaustive linear search.
588 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
589 * maps the file to which the target page belongs. The ->vm_private_data field
590 * holds the current cursor into that scan. Successive searches will circulate
591 * around the vma's virtual address space.
593 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
594 * more scanning pressure is placed against them as well. Eventually pages
595 * will become fully unmapped and are eligible for eviction.
597 * For very sparsely populated VMAs this is a little inefficient - chances are
598 * there there won't be many ptes located within the scan cluster. In this case
599 * maybe we could scan further - to the end of the pte page, perhaps.
601 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
602 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
604 static void try_to_unmap_cluster(unsigned long cursor
,
605 unsigned int *mapcount
, struct vm_area_struct
*vma
)
607 struct mm_struct
*mm
= vma
->vm_mm
;
615 unsigned long address
;
619 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
620 end
= address
+ CLUSTER_SIZE
;
621 if (address
< vma
->vm_start
)
622 address
= vma
->vm_start
;
623 if (end
> vma
->vm_end
)
626 pgd
= pgd_offset(mm
, address
);
627 if (!pgd_present(*pgd
))
630 pud
= pud_offset(pgd
, address
);
631 if (!pud_present(*pud
))
634 pmd
= pmd_offset(pud
, address
);
635 if (!pmd_present(*pmd
))
638 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
640 /* Update high watermark before we lower rss */
641 update_hiwater_rss(mm
);
643 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
644 if (!pte_present(*pte
))
648 if (unlikely(!pfn_valid(pfn
))) {
649 print_bad_pte(vma
, *pte
, address
);
653 page
= pfn_to_page(pfn
);
654 BUG_ON(PageAnon(page
));
656 if (ptep_clear_flush_young(vma
, address
, pte
))
659 /* Nuke the page table entry. */
660 flush_cache_page(vma
, address
, pfn
);
661 pteval
= ptep_clear_flush(vma
, address
, pte
);
663 /* If nonlinear, store the file page offset in the pte. */
664 if (page
->index
!= linear_page_index(vma
, address
))
665 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
667 /* Move the dirty bit to the physical page now the pte is gone. */
668 if (pte_dirty(pteval
))
669 set_page_dirty(page
);
671 page_remove_rmap(page
);
672 page_cache_release(page
);
673 dec_mm_counter(mm
, file_rss
);
676 pte_unmap_unlock(pte
- 1, ptl
);
679 static int try_to_unmap_anon(struct page
*page
)
681 struct anon_vma
*anon_vma
;
682 struct vm_area_struct
*vma
;
683 int ret
= SWAP_AGAIN
;
685 anon_vma
= page_lock_anon_vma(page
);
689 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
690 ret
= try_to_unmap_one(page
, vma
);
691 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
694 spin_unlock(&anon_vma
->lock
);
699 * try_to_unmap_file - unmap file page using the object-based rmap method
700 * @page: the page to unmap
702 * Find all the mappings of a page using the mapping pointer and the vma chains
703 * contained in the address_space struct it points to.
705 * This function is only called from try_to_unmap for object-based pages.
707 static int try_to_unmap_file(struct page
*page
)
709 struct address_space
*mapping
= page
->mapping
;
710 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
711 struct vm_area_struct
*vma
;
712 struct prio_tree_iter iter
;
713 int ret
= SWAP_AGAIN
;
714 unsigned long cursor
;
715 unsigned long max_nl_cursor
= 0;
716 unsigned long max_nl_size
= 0;
717 unsigned int mapcount
;
719 spin_lock(&mapping
->i_mmap_lock
);
720 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
721 ret
= try_to_unmap_one(page
, vma
);
722 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
726 if (list_empty(&mapping
->i_mmap_nonlinear
))
729 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
730 shared
.vm_set
.list
) {
731 if (vma
->vm_flags
& VM_LOCKED
)
733 cursor
= (unsigned long) vma
->vm_private_data
;
734 if (cursor
> max_nl_cursor
)
735 max_nl_cursor
= cursor
;
736 cursor
= vma
->vm_end
- vma
->vm_start
;
737 if (cursor
> max_nl_size
)
738 max_nl_size
= cursor
;
741 if (max_nl_size
== 0) { /* any nonlinears locked or reserved */
747 * We don't try to search for this page in the nonlinear vmas,
748 * and page_referenced wouldn't have found it anyway. Instead
749 * just walk the nonlinear vmas trying to age and unmap some.
750 * The mapcount of the page we came in with is irrelevant,
751 * but even so use it as a guide to how hard we should try?
753 mapcount
= page_mapcount(page
);
756 cond_resched_lock(&mapping
->i_mmap_lock
);
758 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
759 if (max_nl_cursor
== 0)
760 max_nl_cursor
= CLUSTER_SIZE
;
763 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
764 shared
.vm_set
.list
) {
765 if (vma
->vm_flags
& VM_LOCKED
)
767 cursor
= (unsigned long) vma
->vm_private_data
;
768 while ( cursor
< max_nl_cursor
&&
769 cursor
< vma
->vm_end
- vma
->vm_start
) {
770 try_to_unmap_cluster(cursor
, &mapcount
, vma
);
771 cursor
+= CLUSTER_SIZE
;
772 vma
->vm_private_data
= (void *) cursor
;
773 if ((int)mapcount
<= 0)
776 vma
->vm_private_data
= (void *) max_nl_cursor
;
778 cond_resched_lock(&mapping
->i_mmap_lock
);
779 max_nl_cursor
+= CLUSTER_SIZE
;
780 } while (max_nl_cursor
<= max_nl_size
);
783 * Don't loop forever (perhaps all the remaining pages are
784 * in locked vmas). Reset cursor on all unreserved nonlinear
785 * vmas, now forgetting on which ones it had fallen behind.
787 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
788 vma
->vm_private_data
= NULL
;
790 spin_unlock(&mapping
->i_mmap_lock
);
795 * try_to_unmap - try to remove all page table mappings to a page
796 * @page: the page to get unmapped
798 * Tries to remove all the page table entries which are mapping this
799 * page, used in the pageout path. Caller must hold the page lock.
802 * SWAP_SUCCESS - we succeeded in removing all mappings
803 * SWAP_AGAIN - we missed a mapping, try again later
804 * SWAP_FAIL - the page is unswappable
806 int try_to_unmap(struct page
*page
)
810 BUG_ON(!PageLocked(page
));
813 ret
= try_to_unmap_anon(page
);
815 ret
= try_to_unmap_file(page
);
817 if (!page_mapped(page
))