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 2003, 2004
21 * Lock ordering in mm:
23 * inode->i_mutex (while writing or truncating, not reading or faulting)
24 * inode->i_alloc_sem (vmtruncate_range)
26 * page->flags PG_locked (lock_page)
27 * mapping->i_mmap_lock
29 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * inode_lock (in set_page_dirty's __mark_inode_dirty)
35 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
38 * within inode_lock in __sync_single_inode)
40 * (code doesn't rely on that order so it could be switched around)
42 * anon_vma->lock (memory_failure, collect_procs_anon)
47 #include <linux/pagemap.h>
48 #include <linux/swap.h>
49 #include <linux/swapops.h>
50 #include <linux/slab.h>
51 #include <linux/init.h>
52 #include <linux/rmap.h>
53 #include <linux/rcupdate.h>
54 #include <linux/module.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/migrate.h>
59 #include <asm/tlbflush.h>
63 static struct kmem_cache
*anon_vma_cachep
;
65 static inline struct anon_vma
*anon_vma_alloc(void)
67 return kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
70 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
72 kmem_cache_free(anon_vma_cachep
, anon_vma
);
76 * anon_vma_prepare - attach an anon_vma to a memory region
77 * @vma: the memory region in question
79 * This makes sure the memory mapping described by 'vma' has
80 * an 'anon_vma' attached to it, so that we can associate the
81 * anonymous pages mapped into it with that anon_vma.
83 * The common case will be that we already have one, but if
84 * if not we either need to find an adjacent mapping that we
85 * can re-use the anon_vma from (very common when the only
86 * reason for splitting a vma has been mprotect()), or we
89 * Anon-vma allocations are very subtle, because we may have
90 * optimistically looked up an anon_vma in page_lock_anon_vma()
91 * and that may actually touch the spinlock even in the newly
92 * allocated vma (it depends on RCU to make sure that the
93 * anon_vma isn't actually destroyed).
95 * As a result, we need to do proper anon_vma locking even
96 * for the new allocation. At the same time, we do not want
97 * to do any locking for the common case of already having
100 * This must be called with the mmap_sem held for reading.
102 int anon_vma_prepare(struct vm_area_struct
*vma
)
104 struct anon_vma
*anon_vma
= vma
->anon_vma
;
107 if (unlikely(!anon_vma
)) {
108 struct mm_struct
*mm
= vma
->vm_mm
;
109 struct anon_vma
*allocated
;
111 anon_vma
= find_mergeable_anon_vma(vma
);
114 anon_vma
= anon_vma_alloc();
115 if (unlikely(!anon_vma
))
117 allocated
= anon_vma
;
119 spin_lock(&anon_vma
->lock
);
121 /* page_table_lock to protect against threads */
122 spin_lock(&mm
->page_table_lock
);
123 if (likely(!vma
->anon_vma
)) {
124 vma
->anon_vma
= anon_vma
;
125 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
128 spin_unlock(&mm
->page_table_lock
);
130 spin_unlock(&anon_vma
->lock
);
131 if (unlikely(allocated
))
132 anon_vma_free(allocated
);
137 void __anon_vma_merge(struct vm_area_struct
*vma
, struct vm_area_struct
*next
)
139 BUG_ON(vma
->anon_vma
!= next
->anon_vma
);
140 list_del(&next
->anon_vma_node
);
143 void __anon_vma_link(struct vm_area_struct
*vma
)
145 struct anon_vma
*anon_vma
= vma
->anon_vma
;
148 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
151 void anon_vma_link(struct vm_area_struct
*vma
)
153 struct anon_vma
*anon_vma
= vma
->anon_vma
;
156 spin_lock(&anon_vma
->lock
);
157 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
158 spin_unlock(&anon_vma
->lock
);
162 void anon_vma_unlink(struct vm_area_struct
*vma
)
164 struct anon_vma
*anon_vma
= vma
->anon_vma
;
170 spin_lock(&anon_vma
->lock
);
171 list_del(&vma
->anon_vma_node
);
173 /* We must garbage collect the anon_vma if it's empty */
174 empty
= list_empty(&anon_vma
->head
);
175 spin_unlock(&anon_vma
->lock
);
178 anon_vma_free(anon_vma
);
181 static void anon_vma_ctor(void *data
)
183 struct anon_vma
*anon_vma
= data
;
185 spin_lock_init(&anon_vma
->lock
);
186 INIT_LIST_HEAD(&anon_vma
->head
);
189 void __init
anon_vma_init(void)
191 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
192 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
196 * Getting a lock on a stable anon_vma from a page off the LRU is
197 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
199 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
201 struct anon_vma
*anon_vma
;
202 unsigned long anon_mapping
;
205 anon_mapping
= (unsigned long) page
->mapping
;
206 if (!(anon_mapping
& PAGE_MAPPING_ANON
))
208 if (!page_mapped(page
))
211 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
212 spin_lock(&anon_vma
->lock
);
219 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
221 spin_unlock(&anon_vma
->lock
);
226 * At what user virtual address is page expected in @vma?
227 * Returns virtual address or -EFAULT if page's index/offset is not
228 * within the range mapped the @vma.
230 static inline unsigned long
231 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
233 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
234 unsigned long address
;
236 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
237 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
238 /* page should be within @vma mapping range */
245 * At what user virtual address is page expected in vma?
246 * checking that the page matches the vma.
248 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
250 if (PageAnon(page
)) {
251 if ((void *)vma
->anon_vma
!=
252 (void *)page
->mapping
- PAGE_MAPPING_ANON
)
254 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
256 vma
->vm_file
->f_mapping
!= page
->mapping
)
260 return vma_address(page
, vma
);
264 * Check that @page is mapped at @address into @mm.
266 * If @sync is false, page_check_address may perform a racy check to avoid
267 * the page table lock when the pte is not present (helpful when reclaiming
268 * highly shared pages).
270 * On success returns with pte mapped and locked.
272 pte_t
*page_check_address(struct page
*page
, struct mm_struct
*mm
,
273 unsigned long address
, spinlock_t
**ptlp
, int sync
)
281 pgd
= pgd_offset(mm
, address
);
282 if (!pgd_present(*pgd
))
285 pud
= pud_offset(pgd
, address
);
286 if (!pud_present(*pud
))
289 pmd
= pmd_offset(pud
, address
);
290 if (!pmd_present(*pmd
))
293 pte
= pte_offset_map(pmd
, address
);
294 /* Make a quick check before getting the lock */
295 if (!sync
&& !pte_present(*pte
)) {
300 ptl
= pte_lockptr(mm
, pmd
);
302 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
306 pte_unmap_unlock(pte
, ptl
);
311 * page_mapped_in_vma - check whether a page is really mapped in a VMA
312 * @page: the page to test
313 * @vma: the VMA to test
315 * Returns 1 if the page is mapped into the page tables of the VMA, 0
316 * if the page is not mapped into the page tables of this VMA. Only
317 * valid for normal file or anonymous VMAs.
319 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
321 unsigned long address
;
325 address
= vma_address(page
, vma
);
326 if (address
== -EFAULT
) /* out of vma range */
328 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
329 if (!pte
) /* the page is not in this mm */
331 pte_unmap_unlock(pte
, ptl
);
337 * Subfunctions of page_referenced: page_referenced_one called
338 * repeatedly from either page_referenced_anon or page_referenced_file.
340 static int page_referenced_one(struct page
*page
,
341 struct vm_area_struct
*vma
,
342 unsigned int *mapcount
,
343 unsigned long *vm_flags
)
345 struct mm_struct
*mm
= vma
->vm_mm
;
346 unsigned long address
;
351 address
= vma_address(page
, vma
);
352 if (address
== -EFAULT
)
355 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
360 * Don't want to elevate referenced for mlocked page that gets this far,
361 * in order that it progresses to try_to_unmap and is moved to the
364 if (vma
->vm_flags
& VM_LOCKED
) {
365 *mapcount
= 1; /* break early from loop */
366 *vm_flags
|= VM_LOCKED
;
370 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
372 * Don't treat a reference through a sequentially read
373 * mapping as such. If the page has been used in
374 * another mapping, we will catch it; if this other
375 * mapping is already gone, the unmap path will have
376 * set PG_referenced or activated the page.
378 if (likely(!VM_SequentialReadHint(vma
)))
382 /* Pretend the page is referenced if the task has the
383 swap token and is in the middle of a page fault. */
384 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
385 rwsem_is_locked(&mm
->mmap_sem
))
390 pte_unmap_unlock(pte
, ptl
);
393 *vm_flags
|= vma
->vm_flags
;
398 static int page_referenced_anon(struct page
*page
,
399 struct mem_cgroup
*mem_cont
,
400 unsigned long *vm_flags
)
402 unsigned int mapcount
;
403 struct anon_vma
*anon_vma
;
404 struct vm_area_struct
*vma
;
407 anon_vma
= page_lock_anon_vma(page
);
411 mapcount
= page_mapcount(page
);
412 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
414 * If we are reclaiming on behalf of a cgroup, skip
415 * counting on behalf of references from different
418 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
420 referenced
+= page_referenced_one(page
, vma
,
421 &mapcount
, vm_flags
);
426 page_unlock_anon_vma(anon_vma
);
431 * page_referenced_file - referenced check for object-based rmap
432 * @page: the page we're checking references on.
433 * @mem_cont: target memory controller
434 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
436 * For an object-based mapped page, find all the places it is mapped and
437 * check/clear the referenced flag. This is done by following the page->mapping
438 * pointer, then walking the chain of vmas it holds. It returns the number
439 * of references it found.
441 * This function is only called from page_referenced for object-based pages.
443 static int page_referenced_file(struct page
*page
,
444 struct mem_cgroup
*mem_cont
,
445 unsigned long *vm_flags
)
447 unsigned int mapcount
;
448 struct address_space
*mapping
= page
->mapping
;
449 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
450 struct vm_area_struct
*vma
;
451 struct prio_tree_iter iter
;
455 * The caller's checks on page->mapping and !PageAnon have made
456 * sure that this is a file page: the check for page->mapping
457 * excludes the case just before it gets set on an anon page.
459 BUG_ON(PageAnon(page
));
462 * The page lock not only makes sure that page->mapping cannot
463 * suddenly be NULLified by truncation, it makes sure that the
464 * structure at mapping cannot be freed and reused yet,
465 * so we can safely take mapping->i_mmap_lock.
467 BUG_ON(!PageLocked(page
));
469 spin_lock(&mapping
->i_mmap_lock
);
472 * i_mmap_lock does not stabilize mapcount at all, but mapcount
473 * is more likely to be accurate if we note it after spinning.
475 mapcount
= page_mapcount(page
);
477 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
479 * If we are reclaiming on behalf of a cgroup, skip
480 * counting on behalf of references from different
483 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
485 referenced
+= page_referenced_one(page
, vma
,
486 &mapcount
, vm_flags
);
491 spin_unlock(&mapping
->i_mmap_lock
);
496 * page_referenced - test if the page was referenced
497 * @page: the page to test
498 * @is_locked: caller holds lock on the page
499 * @mem_cont: target memory controller
500 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
502 * Quick test_and_clear_referenced for all mappings to a page,
503 * returns the number of ptes which referenced the page.
505 int page_referenced(struct page
*page
,
507 struct mem_cgroup
*mem_cont
,
508 unsigned long *vm_flags
)
512 if (TestClearPageReferenced(page
))
516 if (page_mapped(page
) && page
->mapping
) {
518 referenced
+= page_referenced_anon(page
, mem_cont
,
521 referenced
+= page_referenced_file(page
, mem_cont
,
523 else if (!trylock_page(page
))
527 referenced
+= page_referenced_file(page
,
533 if (page_test_and_clear_young(page
))
539 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
)
541 struct mm_struct
*mm
= vma
->vm_mm
;
542 unsigned long address
;
547 address
= vma_address(page
, vma
);
548 if (address
== -EFAULT
)
551 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
555 if (pte_dirty(*pte
) || pte_write(*pte
)) {
558 flush_cache_page(vma
, address
, pte_pfn(*pte
));
559 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
560 entry
= pte_wrprotect(entry
);
561 entry
= pte_mkclean(entry
);
562 set_pte_at(mm
, address
, pte
, entry
);
566 pte_unmap_unlock(pte
, ptl
);
571 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
573 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
574 struct vm_area_struct
*vma
;
575 struct prio_tree_iter iter
;
578 BUG_ON(PageAnon(page
));
580 spin_lock(&mapping
->i_mmap_lock
);
581 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
582 if (vma
->vm_flags
& VM_SHARED
)
583 ret
+= page_mkclean_one(page
, vma
);
585 spin_unlock(&mapping
->i_mmap_lock
);
589 int page_mkclean(struct page
*page
)
593 BUG_ON(!PageLocked(page
));
595 if (page_mapped(page
)) {
596 struct address_space
*mapping
= page_mapping(page
);
598 ret
= page_mkclean_file(mapping
, page
);
599 if (page_test_dirty(page
)) {
600 page_clear_dirty(page
);
608 EXPORT_SYMBOL_GPL(page_mkclean
);
611 * __page_set_anon_rmap - setup new anonymous rmap
612 * @page: the page to add the mapping to
613 * @vma: the vm area in which the mapping is added
614 * @address: the user virtual address mapped
616 static void __page_set_anon_rmap(struct page
*page
,
617 struct vm_area_struct
*vma
, unsigned long address
)
619 struct anon_vma
*anon_vma
= vma
->anon_vma
;
622 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
623 page
->mapping
= (struct address_space
*) anon_vma
;
625 page
->index
= linear_page_index(vma
, address
);
628 * nr_mapped state can be updated without turning off
629 * interrupts because it is not modified via interrupt.
631 __inc_zone_page_state(page
, NR_ANON_PAGES
);
635 * __page_check_anon_rmap - sanity check anonymous rmap addition
636 * @page: the page to add the mapping to
637 * @vma: the vm area in which the mapping is added
638 * @address: the user virtual address mapped
640 static void __page_check_anon_rmap(struct page
*page
,
641 struct vm_area_struct
*vma
, unsigned long address
)
643 #ifdef CONFIG_DEBUG_VM
645 * The page's anon-rmap details (mapping and index) are guaranteed to
646 * be set up correctly at this point.
648 * We have exclusion against page_add_anon_rmap because the caller
649 * always holds the page locked, except if called from page_dup_rmap,
650 * in which case the page is already known to be setup.
652 * We have exclusion against page_add_new_anon_rmap because those pages
653 * are initially only visible via the pagetables, and the pte is locked
654 * over the call to page_add_new_anon_rmap.
656 struct anon_vma
*anon_vma
= vma
->anon_vma
;
657 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
658 BUG_ON(page
->mapping
!= (struct address_space
*)anon_vma
);
659 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
664 * page_add_anon_rmap - add pte mapping to an anonymous page
665 * @page: the page to add the mapping to
666 * @vma: the vm area in which the mapping is added
667 * @address: the user virtual address mapped
669 * The caller needs to hold the pte lock and the page must be locked.
671 void page_add_anon_rmap(struct page
*page
,
672 struct vm_area_struct
*vma
, unsigned long address
)
674 VM_BUG_ON(!PageLocked(page
));
675 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
676 if (atomic_inc_and_test(&page
->_mapcount
))
677 __page_set_anon_rmap(page
, vma
, address
);
679 __page_check_anon_rmap(page
, vma
, address
);
683 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
684 * @page: the page to add the mapping to
685 * @vma: the vm area in which the mapping is added
686 * @address: the user virtual address mapped
688 * Same as page_add_anon_rmap but must only be called on *new* pages.
689 * This means the inc-and-test can be bypassed.
690 * Page does not have to be locked.
692 void page_add_new_anon_rmap(struct page
*page
,
693 struct vm_area_struct
*vma
, unsigned long address
)
695 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
696 SetPageSwapBacked(page
);
697 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
698 __page_set_anon_rmap(page
, vma
, address
);
699 if (page_evictable(page
, vma
))
700 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
702 add_page_to_unevictable_list(page
);
706 * page_add_file_rmap - add pte mapping to a file page
707 * @page: the page to add the mapping to
709 * The caller needs to hold the pte lock.
711 void page_add_file_rmap(struct page
*page
)
713 if (atomic_inc_and_test(&page
->_mapcount
)) {
714 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
715 mem_cgroup_update_mapped_file_stat(page
, 1);
720 * page_remove_rmap - take down pte mapping from a page
721 * @page: page to remove mapping from
723 * The caller needs to hold the pte lock.
725 void page_remove_rmap(struct page
*page
)
727 /* page still mapped by someone else? */
728 if (!atomic_add_negative(-1, &page
->_mapcount
))
732 * Now that the last pte has gone, s390 must transfer dirty
733 * flag from storage key to struct page. We can usually skip
734 * this if the page is anon, so about to be freed; but perhaps
735 * not if it's in swapcache - there might be another pte slot
736 * containing the swap entry, but page not yet written to swap.
738 if ((!PageAnon(page
) || PageSwapCache(page
)) && page_test_dirty(page
)) {
739 page_clear_dirty(page
);
740 set_page_dirty(page
);
742 if (PageAnon(page
)) {
743 mem_cgroup_uncharge_page(page
);
744 __dec_zone_page_state(page
, NR_ANON_PAGES
);
746 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
748 mem_cgroup_update_mapped_file_stat(page
, -1);
750 * It would be tidy to reset the PageAnon mapping here,
751 * but that might overwrite a racing page_add_anon_rmap
752 * which increments mapcount after us but sets mapping
753 * before us: so leave the reset to free_hot_cold_page,
754 * and remember that it's only reliable while mapped.
755 * Leaving it set also helps swapoff to reinstate ptes
756 * faster for those pages still in swapcache.
761 * Subfunctions of try_to_unmap: try_to_unmap_one called
762 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
764 static int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
765 enum ttu_flags flags
)
767 struct mm_struct
*mm
= vma
->vm_mm
;
768 unsigned long address
;
772 int ret
= SWAP_AGAIN
;
774 address
= vma_address(page
, vma
);
775 if (address
== -EFAULT
)
778 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
783 * If the page is mlock()d, we cannot swap it out.
784 * If it's recently referenced (perhaps page_referenced
785 * skipped over this mm) then we should reactivate it.
787 if (!(flags
& TTU_IGNORE_MLOCK
)) {
788 if (vma
->vm_flags
& VM_LOCKED
) {
793 if (!(flags
& TTU_IGNORE_ACCESS
)) {
794 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
800 /* Nuke the page table entry. */
801 flush_cache_page(vma
, address
, page_to_pfn(page
));
802 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
804 /* Move the dirty bit to the physical page now the pte is gone. */
805 if (pte_dirty(pteval
))
806 set_page_dirty(page
);
808 /* Update high watermark before we lower rss */
809 update_hiwater_rss(mm
);
811 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
813 dec_mm_counter(mm
, anon_rss
);
815 dec_mm_counter(mm
, file_rss
);
816 set_pte_at(mm
, address
, pte
,
817 swp_entry_to_pte(make_hwpoison_entry(page
)));
818 } else if (PageAnon(page
)) {
819 swp_entry_t entry
= { .val
= page_private(page
) };
821 if (PageSwapCache(page
)) {
823 * Store the swap location in the pte.
824 * See handle_pte_fault() ...
826 if (swap_duplicate(entry
) < 0) {
827 set_pte_at(mm
, address
, pte
, pteval
);
831 if (list_empty(&mm
->mmlist
)) {
832 spin_lock(&mmlist_lock
);
833 if (list_empty(&mm
->mmlist
))
834 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
835 spin_unlock(&mmlist_lock
);
837 dec_mm_counter(mm
, anon_rss
);
838 } else if (PAGE_MIGRATION
) {
840 * Store the pfn of the page in a special migration
841 * pte. do_swap_page() will wait until the migration
842 * pte is removed and then restart fault handling.
844 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
845 entry
= make_migration_entry(page
, pte_write(pteval
));
847 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
848 BUG_ON(pte_file(*pte
));
849 } else if (PAGE_MIGRATION
&& (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
850 /* Establish migration entry for a file page */
852 entry
= make_migration_entry(page
, pte_write(pteval
));
853 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
855 dec_mm_counter(mm
, file_rss
);
858 page_remove_rmap(page
);
859 page_cache_release(page
);
862 pte_unmap_unlock(pte
, ptl
);
868 * objrmap doesn't work for nonlinear VMAs because the assumption that
869 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
870 * Consequently, given a particular page and its ->index, we cannot locate the
871 * ptes which are mapping that page without an exhaustive linear search.
873 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
874 * maps the file to which the target page belongs. The ->vm_private_data field
875 * holds the current cursor into that scan. Successive searches will circulate
876 * around the vma's virtual address space.
878 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
879 * more scanning pressure is placed against them as well. Eventually pages
880 * will become fully unmapped and are eligible for eviction.
882 * For very sparsely populated VMAs this is a little inefficient - chances are
883 * there there won't be many ptes located within the scan cluster. In this case
884 * maybe we could scan further - to the end of the pte page, perhaps.
886 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
887 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
888 * rather than unmapping them. If we encounter the "check_page" that vmscan is
889 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
891 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
892 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
894 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
895 struct vm_area_struct
*vma
, struct page
*check_page
)
897 struct mm_struct
*mm
= vma
->vm_mm
;
905 unsigned long address
;
907 int ret
= SWAP_AGAIN
;
910 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
911 end
= address
+ CLUSTER_SIZE
;
912 if (address
< vma
->vm_start
)
913 address
= vma
->vm_start
;
914 if (end
> vma
->vm_end
)
917 pgd
= pgd_offset(mm
, address
);
918 if (!pgd_present(*pgd
))
921 pud
= pud_offset(pgd
, address
);
922 if (!pud_present(*pud
))
925 pmd
= pmd_offset(pud
, address
);
926 if (!pmd_present(*pmd
))
930 * MLOCK_PAGES => feature is configured.
931 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
932 * keep the sem while scanning the cluster for mlocking pages.
934 if (MLOCK_PAGES
&& down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
935 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
937 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
940 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
942 /* Update high watermark before we lower rss */
943 update_hiwater_rss(mm
);
945 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
946 if (!pte_present(*pte
))
948 page
= vm_normal_page(vma
, address
, *pte
);
949 BUG_ON(!page
|| PageAnon(page
));
952 mlock_vma_page(page
); /* no-op if already mlocked */
953 if (page
== check_page
)
955 continue; /* don't unmap */
958 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
961 /* Nuke the page table entry. */
962 flush_cache_page(vma
, address
, pte_pfn(*pte
));
963 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
965 /* If nonlinear, store the file page offset in the pte. */
966 if (page
->index
!= linear_page_index(vma
, address
))
967 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
969 /* Move the dirty bit to the physical page now the pte is gone. */
970 if (pte_dirty(pteval
))
971 set_page_dirty(page
);
973 page_remove_rmap(page
);
974 page_cache_release(page
);
975 dec_mm_counter(mm
, file_rss
);
978 pte_unmap_unlock(pte
- 1, ptl
);
980 up_read(&vma
->vm_mm
->mmap_sem
);
985 * common handling for pages mapped in VM_LOCKED vmas
987 static int try_to_mlock_page(struct page
*page
, struct vm_area_struct
*vma
)
991 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
992 if (vma
->vm_flags
& VM_LOCKED
) {
993 mlock_vma_page(page
);
994 mlocked
++; /* really mlocked the page */
996 up_read(&vma
->vm_mm
->mmap_sem
);
1002 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1004 * @page: the page to unmap/unlock
1005 * @flags: action and flags
1007 * Find all the mappings of a page using the mapping pointer and the vma chains
1008 * contained in the anon_vma struct it points to.
1010 * This function is only called from try_to_unmap/try_to_munlock for
1012 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1013 * where the page was found will be held for write. So, we won't recheck
1014 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1017 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1019 struct anon_vma
*anon_vma
;
1020 struct vm_area_struct
*vma
;
1021 unsigned int mlocked
= 0;
1022 int ret
= SWAP_AGAIN
;
1023 int unlock
= TTU_ACTION(flags
) == TTU_MUNLOCK
;
1025 if (MLOCK_PAGES
&& unlikely(unlock
))
1026 ret
= SWAP_SUCCESS
; /* default for try_to_munlock() */
1028 anon_vma
= page_lock_anon_vma(page
);
1032 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
1033 if (MLOCK_PAGES
&& unlikely(unlock
)) {
1034 if (!((vma
->vm_flags
& VM_LOCKED
) &&
1035 page_mapped_in_vma(page
, vma
)))
1036 continue; /* must visit all unlocked vmas */
1037 ret
= SWAP_MLOCK
; /* saw at least one mlocked vma */
1039 ret
= try_to_unmap_one(page
, vma
, flags
);
1040 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
1043 if (ret
== SWAP_MLOCK
) {
1044 mlocked
= try_to_mlock_page(page
, vma
);
1046 break; /* stop if actually mlocked page */
1050 page_unlock_anon_vma(anon_vma
);
1053 ret
= SWAP_MLOCK
; /* actually mlocked the page */
1054 else if (ret
== SWAP_MLOCK
)
1055 ret
= SWAP_AGAIN
; /* saw VM_LOCKED vma */
1061 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1062 * @page: the page to unmap/unlock
1063 * @flags: action and flags
1065 * Find all the mappings of a page using the mapping pointer and the vma chains
1066 * contained in the address_space struct it points to.
1068 * This function is only called from try_to_unmap/try_to_munlock for
1069 * object-based pages.
1070 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1071 * where the page was found will be held for write. So, we won't recheck
1072 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1075 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1077 struct address_space
*mapping
= page
->mapping
;
1078 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1079 struct vm_area_struct
*vma
;
1080 struct prio_tree_iter iter
;
1081 int ret
= SWAP_AGAIN
;
1082 unsigned long cursor
;
1083 unsigned long max_nl_cursor
= 0;
1084 unsigned long max_nl_size
= 0;
1085 unsigned int mapcount
;
1086 unsigned int mlocked
= 0;
1087 int unlock
= TTU_ACTION(flags
) == TTU_MUNLOCK
;
1089 if (MLOCK_PAGES
&& unlikely(unlock
))
1090 ret
= SWAP_SUCCESS
; /* default for try_to_munlock() */
1092 spin_lock(&mapping
->i_mmap_lock
);
1093 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1094 if (MLOCK_PAGES
&& unlikely(unlock
)) {
1095 if (!((vma
->vm_flags
& VM_LOCKED
) &&
1096 page_mapped_in_vma(page
, vma
)))
1097 continue; /* must visit all vmas */
1100 ret
= try_to_unmap_one(page
, vma
, flags
);
1101 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
1104 if (ret
== SWAP_MLOCK
) {
1105 mlocked
= try_to_mlock_page(page
, vma
);
1107 goto out
; /* stop if actually mlocked page */
1111 if (list_empty(&mapping
->i_mmap_nonlinear
))
1114 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1115 shared
.vm_set
.list
) {
1116 if (MLOCK_PAGES
&& unlikely(unlock
)) {
1117 if (!(vma
->vm_flags
& VM_LOCKED
))
1118 continue; /* must visit all vmas */
1119 ret
= SWAP_MLOCK
; /* leave mlocked == 0 */
1120 goto out
; /* no need to look further */
1122 if (!MLOCK_PAGES
&& !(flags
& TTU_IGNORE_MLOCK
) &&
1123 (vma
->vm_flags
& VM_LOCKED
))
1125 cursor
= (unsigned long) vma
->vm_private_data
;
1126 if (cursor
> max_nl_cursor
)
1127 max_nl_cursor
= cursor
;
1128 cursor
= vma
->vm_end
- vma
->vm_start
;
1129 if (cursor
> max_nl_size
)
1130 max_nl_size
= cursor
;
1133 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1139 * We don't try to search for this page in the nonlinear vmas,
1140 * and page_referenced wouldn't have found it anyway. Instead
1141 * just walk the nonlinear vmas trying to age and unmap some.
1142 * The mapcount of the page we came in with is irrelevant,
1143 * but even so use it as a guide to how hard we should try?
1145 mapcount
= page_mapcount(page
);
1148 cond_resched_lock(&mapping
->i_mmap_lock
);
1150 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1151 if (max_nl_cursor
== 0)
1152 max_nl_cursor
= CLUSTER_SIZE
;
1155 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1156 shared
.vm_set
.list
) {
1157 if (!MLOCK_PAGES
&& !(flags
& TTU_IGNORE_MLOCK
) &&
1158 (vma
->vm_flags
& VM_LOCKED
))
1160 cursor
= (unsigned long) vma
->vm_private_data
;
1161 while ( cursor
< max_nl_cursor
&&
1162 cursor
< vma
->vm_end
- vma
->vm_start
) {
1163 ret
= try_to_unmap_cluster(cursor
, &mapcount
,
1165 if (ret
== SWAP_MLOCK
)
1166 mlocked
= 2; /* to return below */
1167 cursor
+= CLUSTER_SIZE
;
1168 vma
->vm_private_data
= (void *) cursor
;
1169 if ((int)mapcount
<= 0)
1172 vma
->vm_private_data
= (void *) max_nl_cursor
;
1174 cond_resched_lock(&mapping
->i_mmap_lock
);
1175 max_nl_cursor
+= CLUSTER_SIZE
;
1176 } while (max_nl_cursor
<= max_nl_size
);
1179 * Don't loop forever (perhaps all the remaining pages are
1180 * in locked vmas). Reset cursor on all unreserved nonlinear
1181 * vmas, now forgetting on which ones it had fallen behind.
1183 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1184 vma
->vm_private_data
= NULL
;
1186 spin_unlock(&mapping
->i_mmap_lock
);
1188 ret
= SWAP_MLOCK
; /* actually mlocked the page */
1189 else if (ret
== SWAP_MLOCK
)
1190 ret
= SWAP_AGAIN
; /* saw VM_LOCKED vma */
1195 * try_to_unmap - try to remove all page table mappings to a page
1196 * @page: the page to get unmapped
1197 * @flags: action and flags
1199 * Tries to remove all the page table entries which are mapping this
1200 * page, used in the pageout path. Caller must hold the page lock.
1201 * Return values are:
1203 * SWAP_SUCCESS - we succeeded in removing all mappings
1204 * SWAP_AGAIN - we missed a mapping, try again later
1205 * SWAP_FAIL - the page is unswappable
1206 * SWAP_MLOCK - page is mlocked.
1208 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1212 BUG_ON(!PageLocked(page
));
1215 ret
= try_to_unmap_anon(page
, flags
);
1217 ret
= try_to_unmap_file(page
, flags
);
1218 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1224 * try_to_munlock - try to munlock a page
1225 * @page: the page to be munlocked
1227 * Called from munlock code. Checks all of the VMAs mapping the page
1228 * to make sure nobody else has this page mlocked. The page will be
1229 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1231 * Return values are:
1233 * SWAP_SUCCESS - no vma's holding page mlocked.
1234 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1235 * SWAP_MLOCK - page is now mlocked.
1237 int try_to_munlock(struct page
*page
)
1239 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1242 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1244 return try_to_unmap_file(page
, TTU_MUNLOCK
);