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/ksm.h>
53 #include <linux/rmap.h>
54 #include <linux/rcupdate.h>
55 #include <linux/module.h>
56 #include <linux/memcontrol.h>
57 #include <linux/mmu_notifier.h>
58 #include <linux/migrate.h>
60 #include <asm/tlbflush.h>
64 static struct kmem_cache *anon_vma_cachep;
65 static struct kmem_cache *anon_vma_chain_cachep;
67 static inline struct anon_vma *anon_vma_alloc(void)
69 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
72 void anon_vma_free(struct anon_vma *anon_vma)
74 kmem_cache_free(anon_vma_cachep, anon_vma);
77 static inline struct anon_vma_chain *anon_vma_chain_alloc(void)
79 return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL);
82 void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
84 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
88 * anon_vma_prepare - attach an anon_vma to a memory region
89 * @vma: the memory region in question
91 * This makes sure the memory mapping described by 'vma' has
92 * an 'anon_vma' attached to it, so that we can associate the
93 * anonymous pages mapped into it with that anon_vma.
95 * The common case will be that we already have one, but if
96 * if not we either need to find an adjacent mapping that we
97 * can re-use the anon_vma from (very common when the only
98 * reason for splitting a vma has been mprotect()), or we
101 * Anon-vma allocations are very subtle, because we may have
102 * optimistically looked up an anon_vma in page_lock_anon_vma()
103 * and that may actually touch the spinlock even in the newly
104 * allocated vma (it depends on RCU to make sure that the
105 * anon_vma isn't actually destroyed).
107 * As a result, we need to do proper anon_vma locking even
108 * for the new allocation. At the same time, we do not want
109 * to do any locking for the common case of already having
112 * This must be called with the mmap_sem held for reading.
114 int anon_vma_prepare(struct vm_area_struct *vma)
116 struct anon_vma *anon_vma = vma->anon_vma;
117 struct anon_vma_chain *avc;
120 if (unlikely(!anon_vma)) {
121 struct mm_struct *mm = vma->vm_mm;
122 struct anon_vma *allocated;
124 avc = anon_vma_chain_alloc();
128 anon_vma = find_mergeable_anon_vma(vma);
131 anon_vma = anon_vma_alloc();
132 if (unlikely(!anon_vma))
133 goto out_enomem_free_avc;
134 allocated = anon_vma;
136 spin_lock(&anon_vma->lock);
138 /* page_table_lock to protect against threads */
139 spin_lock(&mm->page_table_lock);
140 if (likely(!vma->anon_vma)) {
141 vma->anon_vma = anon_vma;
142 avc->anon_vma = anon_vma;
144 list_add(&avc->same_vma, &vma->anon_vma_chain);
145 list_add(&avc->same_anon_vma, &anon_vma->head);
148 spin_unlock(&mm->page_table_lock);
150 spin_unlock(&anon_vma->lock);
151 if (unlikely(allocated)) {
152 anon_vma_free(allocated);
153 anon_vma_chain_free(avc);
159 anon_vma_chain_free(avc);
164 static void anon_vma_chain_link(struct vm_area_struct *vma,
165 struct anon_vma_chain *avc,
166 struct anon_vma *anon_vma)
169 avc->anon_vma = anon_vma;
170 list_add(&avc->same_vma, &vma->anon_vma_chain);
172 spin_lock(&anon_vma->lock);
173 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
174 spin_unlock(&anon_vma->lock);
178 * Attach the anon_vmas from src to dst.
179 * Returns 0 on success, -ENOMEM on failure.
181 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
183 struct anon_vma_chain *avc, *pavc;
185 list_for_each_entry(pavc, &src->anon_vma_chain, same_vma) {
186 avc = anon_vma_chain_alloc();
189 anon_vma_chain_link(dst, avc, pavc->anon_vma);
194 unlink_anon_vmas(dst);
199 * Attach vma to its own anon_vma, as well as to the anon_vmas that
200 * the corresponding VMA in the parent process is attached to.
201 * Returns 0 on success, non-zero on failure.
203 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
205 struct anon_vma_chain *avc;
206 struct anon_vma *anon_vma;
208 /* Don't bother if the parent process has no anon_vma here. */
213 * First, attach the new VMA to the parent VMA's anon_vmas,
214 * so rmap can find non-COWed pages in child processes.
216 if (anon_vma_clone(vma, pvma))
219 /* Then add our own anon_vma. */
220 anon_vma = anon_vma_alloc();
223 avc = anon_vma_chain_alloc();
225 goto out_error_free_anon_vma;
226 anon_vma_chain_link(vma, avc, anon_vma);
227 /* Mark this anon_vma as the one where our new (COWed) pages go. */
228 vma->anon_vma = anon_vma;
232 out_error_free_anon_vma:
233 anon_vma_free(anon_vma);
238 static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
240 struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
243 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
247 spin_lock(&anon_vma->lock);
248 list_del(&anon_vma_chain->same_anon_vma);
250 /* We must garbage collect the anon_vma if it's empty */
251 empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma);
252 spin_unlock(&anon_vma->lock);
255 anon_vma_free(anon_vma);
258 void unlink_anon_vmas(struct vm_area_struct *vma)
260 struct anon_vma_chain *avc, *next;
262 /* Unlink each anon_vma chained to the VMA. */
263 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
264 anon_vma_unlink(avc);
265 list_del(&avc->same_vma);
266 anon_vma_chain_free(avc);
270 static void anon_vma_ctor(void *data)
272 struct anon_vma *anon_vma = data;
274 spin_lock_init(&anon_vma->lock);
275 ksm_refcount_init(anon_vma);
276 INIT_LIST_HEAD(&anon_vma->head);
279 void __init anon_vma_init(void)
281 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
282 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
283 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
287 * Getting a lock on a stable anon_vma from a page off the LRU is
288 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
290 struct anon_vma *page_lock_anon_vma(struct page *page)
292 struct anon_vma *anon_vma;
293 unsigned long anon_mapping;
296 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
297 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
299 if (!page_mapped(page))
302 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
303 spin_lock(&anon_vma->lock);
310 void page_unlock_anon_vma(struct anon_vma *anon_vma)
312 spin_unlock(&anon_vma->lock);
317 * At what user virtual address is page expected in @vma?
318 * Returns virtual address or -EFAULT if page's index/offset is not
319 * within the range mapped the @vma.
321 static inline unsigned long
322 vma_address(struct page *page, struct vm_area_struct *vma)
324 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
325 unsigned long address;
327 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
328 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
329 /* page should be within @vma mapping range */
336 * At what user virtual address is page expected in vma?
337 * checking that the page matches the vma.
339 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
341 if (PageAnon(page)) {
342 if (vma->anon_vma != page_anon_vma(page))
344 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
346 vma->vm_file->f_mapping != page->mapping)
350 return vma_address(page, vma);
354 * Check that @page is mapped at @address into @mm.
356 * If @sync is false, page_check_address may perform a racy check to avoid
357 * the page table lock when the pte is not present (helpful when reclaiming
358 * highly shared pages).
360 * On success returns with pte mapped and locked.
362 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
363 unsigned long address, spinlock_t **ptlp, int sync)
371 pgd = pgd_offset(mm, address);
372 if (!pgd_present(*pgd))
375 pud = pud_offset(pgd, address);
376 if (!pud_present(*pud))
379 pmd = pmd_offset(pud, address);
380 if (!pmd_present(*pmd))
383 pte = pte_offset_map(pmd, address);
384 /* Make a quick check before getting the lock */
385 if (!sync && !pte_present(*pte)) {
390 ptl = pte_lockptr(mm, pmd);
392 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
396 pte_unmap_unlock(pte, ptl);
401 * page_mapped_in_vma - check whether a page is really mapped in a VMA
402 * @page: the page to test
403 * @vma: the VMA to test
405 * Returns 1 if the page is mapped into the page tables of the VMA, 0
406 * if the page is not mapped into the page tables of this VMA. Only
407 * valid for normal file or anonymous VMAs.
409 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
411 unsigned long address;
415 address = vma_address(page, vma);
416 if (address == -EFAULT) /* out of vma range */
418 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
419 if (!pte) /* the page is not in this mm */
421 pte_unmap_unlock(pte, ptl);
427 * Subfunctions of page_referenced: page_referenced_one called
428 * repeatedly from either page_referenced_anon or page_referenced_file.
430 int page_referenced_one(struct page *page, struct vm_area_struct *vma,
431 unsigned long address, unsigned int *mapcount,
432 unsigned long *vm_flags)
434 struct mm_struct *mm = vma->vm_mm;
439 pte = page_check_address(page, mm, address, &ptl, 0);
444 * Don't want to elevate referenced for mlocked page that gets this far,
445 * in order that it progresses to try_to_unmap and is moved to the
448 if (vma->vm_flags & VM_LOCKED) {
449 *mapcount = 1; /* break early from loop */
450 *vm_flags |= VM_LOCKED;
454 if (ptep_clear_flush_young_notify(vma, address, pte)) {
456 * Don't treat a reference through a sequentially read
457 * mapping as such. If the page has been used in
458 * another mapping, we will catch it; if this other
459 * mapping is already gone, the unmap path will have
460 * set PG_referenced or activated the page.
462 if (likely(!VM_SequentialReadHint(vma)))
466 /* Pretend the page is referenced if the task has the
467 swap token and is in the middle of a page fault. */
468 if (mm != current->mm && has_swap_token(mm) &&
469 rwsem_is_locked(&mm->mmap_sem))
474 pte_unmap_unlock(pte, ptl);
477 *vm_flags |= vma->vm_flags;
482 static int page_referenced_anon(struct page *page,
483 struct mem_cgroup *mem_cont,
484 unsigned long *vm_flags)
486 unsigned int mapcount;
487 struct anon_vma *anon_vma;
488 struct anon_vma_chain *avc;
491 anon_vma = page_lock_anon_vma(page);
495 mapcount = page_mapcount(page);
496 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
497 struct vm_area_struct *vma = avc->vma;
498 unsigned long address = vma_address(page, vma);
499 if (address == -EFAULT)
502 * If we are reclaiming on behalf of a cgroup, skip
503 * counting on behalf of references from different
506 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
508 referenced += page_referenced_one(page, vma, address,
509 &mapcount, vm_flags);
514 page_unlock_anon_vma(anon_vma);
519 * page_referenced_file - referenced check for object-based rmap
520 * @page: the page we're checking references on.
521 * @mem_cont: target memory controller
522 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
524 * For an object-based mapped page, find all the places it is mapped and
525 * check/clear the referenced flag. This is done by following the page->mapping
526 * pointer, then walking the chain of vmas it holds. It returns the number
527 * of references it found.
529 * This function is only called from page_referenced for object-based pages.
531 static int page_referenced_file(struct page *page,
532 struct mem_cgroup *mem_cont,
533 unsigned long *vm_flags)
535 unsigned int mapcount;
536 struct address_space *mapping = page->mapping;
537 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
538 struct vm_area_struct *vma;
539 struct prio_tree_iter iter;
543 * The caller's checks on page->mapping and !PageAnon have made
544 * sure that this is a file page: the check for page->mapping
545 * excludes the case just before it gets set on an anon page.
547 BUG_ON(PageAnon(page));
550 * The page lock not only makes sure that page->mapping cannot
551 * suddenly be NULLified by truncation, it makes sure that the
552 * structure at mapping cannot be freed and reused yet,
553 * so we can safely take mapping->i_mmap_lock.
555 BUG_ON(!PageLocked(page));
557 spin_lock(&mapping->i_mmap_lock);
560 * i_mmap_lock does not stabilize mapcount at all, but mapcount
561 * is more likely to be accurate if we note it after spinning.
563 mapcount = page_mapcount(page);
565 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
566 unsigned long address = vma_address(page, vma);
567 if (address == -EFAULT)
570 * If we are reclaiming on behalf of a cgroup, skip
571 * counting on behalf of references from different
574 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
576 referenced += page_referenced_one(page, vma, address,
577 &mapcount, vm_flags);
582 spin_unlock(&mapping->i_mmap_lock);
587 * page_referenced - test if the page was referenced
588 * @page: the page to test
589 * @is_locked: caller holds lock on the page
590 * @mem_cont: target memory controller
591 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
593 * Quick test_and_clear_referenced for all mappings to a page,
594 * returns the number of ptes which referenced the page.
596 int page_referenced(struct page *page,
598 struct mem_cgroup *mem_cont,
599 unsigned long *vm_flags)
604 if (TestClearPageReferenced(page))
608 if (page_mapped(page) && page_rmapping(page)) {
609 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
610 we_locked = trylock_page(page);
616 if (unlikely(PageKsm(page)))
617 referenced += page_referenced_ksm(page, mem_cont,
619 else if (PageAnon(page))
620 referenced += page_referenced_anon(page, mem_cont,
622 else if (page->mapping)
623 referenced += page_referenced_file(page, mem_cont,
629 if (page_test_and_clear_young(page))
635 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
636 unsigned long address)
638 struct mm_struct *mm = vma->vm_mm;
643 pte = page_check_address(page, mm, address, &ptl, 1);
647 if (pte_dirty(*pte) || pte_write(*pte)) {
650 flush_cache_page(vma, address, pte_pfn(*pte));
651 entry = ptep_clear_flush_notify(vma, address, pte);
652 entry = pte_wrprotect(entry);
653 entry = pte_mkclean(entry);
654 set_pte_at(mm, address, pte, entry);
658 pte_unmap_unlock(pte, ptl);
663 static int page_mkclean_file(struct address_space *mapping, struct page *page)
665 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
666 struct vm_area_struct *vma;
667 struct prio_tree_iter iter;
670 BUG_ON(PageAnon(page));
672 spin_lock(&mapping->i_mmap_lock);
673 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
674 if (vma->vm_flags & VM_SHARED) {
675 unsigned long address = vma_address(page, vma);
676 if (address == -EFAULT)
678 ret += page_mkclean_one(page, vma, address);
681 spin_unlock(&mapping->i_mmap_lock);
685 int page_mkclean(struct page *page)
689 BUG_ON(!PageLocked(page));
691 if (page_mapped(page)) {
692 struct address_space *mapping = page_mapping(page);
694 ret = page_mkclean_file(mapping, page);
695 if (page_test_dirty(page)) {
696 page_clear_dirty(page);
704 EXPORT_SYMBOL_GPL(page_mkclean);
707 * page_move_anon_rmap - move a page to our anon_vma
708 * @page: the page to move to our anon_vma
709 * @vma: the vma the page belongs to
710 * @address: the user virtual address mapped
712 * When a page belongs exclusively to one process after a COW event,
713 * that page can be moved into the anon_vma that belongs to just that
714 * process, so the rmap code will not search the parent or sibling
717 void page_move_anon_rmap(struct page *page,
718 struct vm_area_struct *vma, unsigned long address)
720 struct anon_vma *anon_vma = vma->anon_vma;
722 VM_BUG_ON(!PageLocked(page));
723 VM_BUG_ON(!anon_vma);
724 VM_BUG_ON(page->index != linear_page_index(vma, address));
726 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
727 page->mapping = (struct address_space *) anon_vma;
731 * __page_set_anon_rmap - setup new anonymous rmap
732 * @page: the page to add the mapping to
733 * @vma: the vm area in which the mapping is added
734 * @address: the user virtual address mapped
736 static void __page_set_anon_rmap(struct page *page,
737 struct vm_area_struct *vma, unsigned long address)
739 struct anon_vma *anon_vma = vma->anon_vma;
742 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
743 page->mapping = (struct address_space *) anon_vma;
744 page->index = linear_page_index(vma, address);
748 * __page_check_anon_rmap - sanity check anonymous rmap addition
749 * @page: the page to add the mapping to
750 * @vma: the vm area in which the mapping is added
751 * @address: the user virtual address mapped
753 static void __page_check_anon_rmap(struct page *page,
754 struct vm_area_struct *vma, unsigned long address)
756 #ifdef CONFIG_DEBUG_VM
758 * The page's anon-rmap details (mapping and index) are guaranteed to
759 * be set up correctly at this point.
761 * We have exclusion against page_add_anon_rmap because the caller
762 * always holds the page locked, except if called from page_dup_rmap,
763 * in which case the page is already known to be setup.
765 * We have exclusion against page_add_new_anon_rmap because those pages
766 * are initially only visible via the pagetables, and the pte is locked
767 * over the call to page_add_new_anon_rmap.
769 BUG_ON(page->index != linear_page_index(vma, address));
774 * page_add_anon_rmap - add pte mapping to an anonymous page
775 * @page: the page to add the mapping to
776 * @vma: the vm area in which the mapping is added
777 * @address: the user virtual address mapped
779 * The caller needs to hold the pte lock, and the page must be locked in
780 * the anon_vma case: to serialize mapping,index checking after setting,
781 * and to ensure that PageAnon is not being upgraded racily to PageKsm
782 * (but PageKsm is never downgraded to PageAnon).
784 void page_add_anon_rmap(struct page *page,
785 struct vm_area_struct *vma, unsigned long address)
787 int first = atomic_inc_and_test(&page->_mapcount);
789 __inc_zone_page_state(page, NR_ANON_PAGES);
790 if (unlikely(PageKsm(page)))
793 VM_BUG_ON(!PageLocked(page));
794 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
796 __page_set_anon_rmap(page, vma, address);
798 __page_check_anon_rmap(page, vma, address);
802 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
803 * @page: the page to add the mapping to
804 * @vma: the vm area in which the mapping is added
805 * @address: the user virtual address mapped
807 * Same as page_add_anon_rmap but must only be called on *new* pages.
808 * This means the inc-and-test can be bypassed.
809 * Page does not have to be locked.
811 void page_add_new_anon_rmap(struct page *page,
812 struct vm_area_struct *vma, unsigned long address)
814 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
815 SetPageSwapBacked(page);
816 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
817 __inc_zone_page_state(page, NR_ANON_PAGES);
818 __page_set_anon_rmap(page, vma, address);
819 if (page_evictable(page, vma))
820 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
822 add_page_to_unevictable_list(page);
826 * page_add_file_rmap - add pte mapping to a file page
827 * @page: the page to add the mapping to
829 * The caller needs to hold the pte lock.
831 void page_add_file_rmap(struct page *page)
833 if (atomic_inc_and_test(&page->_mapcount)) {
834 __inc_zone_page_state(page, NR_FILE_MAPPED);
835 mem_cgroup_update_file_mapped(page, 1);
840 * page_remove_rmap - take down pte mapping from a page
841 * @page: page to remove mapping from
843 * The caller needs to hold the pte lock.
845 void page_remove_rmap(struct page *page)
847 /* page still mapped by someone else? */
848 if (!atomic_add_negative(-1, &page->_mapcount))
852 * Now that the last pte has gone, s390 must transfer dirty
853 * flag from storage key to struct page. We can usually skip
854 * this if the page is anon, so about to be freed; but perhaps
855 * not if it's in swapcache - there might be another pte slot
856 * containing the swap entry, but page not yet written to swap.
858 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
859 page_clear_dirty(page);
860 set_page_dirty(page);
862 if (PageAnon(page)) {
863 mem_cgroup_uncharge_page(page);
864 __dec_zone_page_state(page, NR_ANON_PAGES);
866 __dec_zone_page_state(page, NR_FILE_MAPPED);
867 mem_cgroup_update_file_mapped(page, -1);
870 * It would be tidy to reset the PageAnon mapping here,
871 * but that might overwrite a racing page_add_anon_rmap
872 * which increments mapcount after us but sets mapping
873 * before us: so leave the reset to free_hot_cold_page,
874 * and remember that it's only reliable while mapped.
875 * Leaving it set also helps swapoff to reinstate ptes
876 * faster for those pages still in swapcache.
881 * Subfunctions of try_to_unmap: try_to_unmap_one called
882 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
884 int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
885 unsigned long address, enum ttu_flags flags)
887 struct mm_struct *mm = vma->vm_mm;
891 int ret = SWAP_AGAIN;
893 pte = page_check_address(page, mm, address, &ptl, 0);
898 * If the page is mlock()d, we cannot swap it out.
899 * If it's recently referenced (perhaps page_referenced
900 * skipped over this mm) then we should reactivate it.
902 if (!(flags & TTU_IGNORE_MLOCK)) {
903 if (vma->vm_flags & VM_LOCKED)
906 if (TTU_ACTION(flags) == TTU_MUNLOCK)
909 if (!(flags & TTU_IGNORE_ACCESS)) {
910 if (ptep_clear_flush_young_notify(vma, address, pte)) {
916 /* Nuke the page table entry. */
917 flush_cache_page(vma, address, page_to_pfn(page));
918 pteval = ptep_clear_flush_notify(vma, address, pte);
920 /* Move the dirty bit to the physical page now the pte is gone. */
921 if (pte_dirty(pteval))
922 set_page_dirty(page);
924 /* Update high watermark before we lower rss */
925 update_hiwater_rss(mm);
927 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
929 dec_mm_counter(mm, MM_ANONPAGES);
931 dec_mm_counter(mm, MM_FILEPAGES);
932 set_pte_at(mm, address, pte,
933 swp_entry_to_pte(make_hwpoison_entry(page)));
934 } else if (PageAnon(page)) {
935 swp_entry_t entry = { .val = page_private(page) };
937 if (PageSwapCache(page)) {
939 * Store the swap location in the pte.
940 * See handle_pte_fault() ...
942 if (swap_duplicate(entry) < 0) {
943 set_pte_at(mm, address, pte, pteval);
947 if (list_empty(&mm->mmlist)) {
948 spin_lock(&mmlist_lock);
949 if (list_empty(&mm->mmlist))
950 list_add(&mm->mmlist, &init_mm.mmlist);
951 spin_unlock(&mmlist_lock);
953 dec_mm_counter(mm, MM_ANONPAGES);
954 inc_mm_counter(mm, MM_SWAPENTS);
955 } else if (PAGE_MIGRATION) {
957 * Store the pfn of the page in a special migration
958 * pte. do_swap_page() will wait until the migration
959 * pte is removed and then restart fault handling.
961 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
962 entry = make_migration_entry(page, pte_write(pteval));
964 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
965 BUG_ON(pte_file(*pte));
966 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
967 /* Establish migration entry for a file page */
969 entry = make_migration_entry(page, pte_write(pteval));
970 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
972 dec_mm_counter(mm, MM_FILEPAGES);
974 page_remove_rmap(page);
975 page_cache_release(page);
978 pte_unmap_unlock(pte, ptl);
983 pte_unmap_unlock(pte, ptl);
987 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
988 * unstable result and race. Plus, We can't wait here because
989 * we now hold anon_vma->lock or mapping->i_mmap_lock.
990 * if trylock failed, the page remain in evictable lru and later
991 * vmscan could retry to move the page to unevictable lru if the
992 * page is actually mlocked.
994 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
995 if (vma->vm_flags & VM_LOCKED) {
996 mlock_vma_page(page);
999 up_read(&vma->vm_mm->mmap_sem);
1005 * objrmap doesn't work for nonlinear VMAs because the assumption that
1006 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1007 * Consequently, given a particular page and its ->index, we cannot locate the
1008 * ptes which are mapping that page without an exhaustive linear search.
1010 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1011 * maps the file to which the target page belongs. The ->vm_private_data field
1012 * holds the current cursor into that scan. Successive searches will circulate
1013 * around the vma's virtual address space.
1015 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1016 * more scanning pressure is placed against them as well. Eventually pages
1017 * will become fully unmapped and are eligible for eviction.
1019 * For very sparsely populated VMAs this is a little inefficient - chances are
1020 * there there won't be many ptes located within the scan cluster. In this case
1021 * maybe we could scan further - to the end of the pte page, perhaps.
1023 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1024 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1025 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1026 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1028 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1029 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1031 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1032 struct vm_area_struct *vma, struct page *check_page)
1034 struct mm_struct *mm = vma->vm_mm;
1042 unsigned long address;
1044 int ret = SWAP_AGAIN;
1047 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1048 end = address + CLUSTER_SIZE;
1049 if (address < vma->vm_start)
1050 address = vma->vm_start;
1051 if (end > vma->vm_end)
1054 pgd = pgd_offset(mm, address);
1055 if (!pgd_present(*pgd))
1058 pud = pud_offset(pgd, address);
1059 if (!pud_present(*pud))
1062 pmd = pmd_offset(pud, address);
1063 if (!pmd_present(*pmd))
1067 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1068 * keep the sem while scanning the cluster for mlocking pages.
1070 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1071 locked_vma = (vma->vm_flags & VM_LOCKED);
1073 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1076 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1078 /* Update high watermark before we lower rss */
1079 update_hiwater_rss(mm);
1081 for (; address < end; pte++, address += PAGE_SIZE) {
1082 if (!pte_present(*pte))
1084 page = vm_normal_page(vma, address, *pte);
1085 BUG_ON(!page || PageAnon(page));
1088 mlock_vma_page(page); /* no-op if already mlocked */
1089 if (page == check_page)
1091 continue; /* don't unmap */
1094 if (ptep_clear_flush_young_notify(vma, address, pte))
1097 /* Nuke the page table entry. */
1098 flush_cache_page(vma, address, pte_pfn(*pte));
1099 pteval = ptep_clear_flush_notify(vma, address, pte);
1101 /* If nonlinear, store the file page offset in the pte. */
1102 if (page->index != linear_page_index(vma, address))
1103 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1105 /* Move the dirty bit to the physical page now the pte is gone. */
1106 if (pte_dirty(pteval))
1107 set_page_dirty(page);
1109 page_remove_rmap(page);
1110 page_cache_release(page);
1111 dec_mm_counter(mm, MM_FILEPAGES);
1114 pte_unmap_unlock(pte - 1, ptl);
1116 up_read(&vma->vm_mm->mmap_sem);
1121 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1123 * @page: the page to unmap/unlock
1124 * @flags: action and flags
1126 * Find all the mappings of a page using the mapping pointer and the vma chains
1127 * contained in the anon_vma struct it points to.
1129 * This function is only called from try_to_unmap/try_to_munlock for
1131 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1132 * where the page was found will be held for write. So, we won't recheck
1133 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1136 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1138 struct anon_vma *anon_vma;
1139 struct anon_vma_chain *avc;
1140 int ret = SWAP_AGAIN;
1142 anon_vma = page_lock_anon_vma(page);
1146 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1147 struct vm_area_struct *vma = avc->vma;
1148 unsigned long address = vma_address(page, vma);
1149 if (address == -EFAULT)
1151 ret = try_to_unmap_one(page, vma, address, flags);
1152 if (ret != SWAP_AGAIN || !page_mapped(page))
1156 page_unlock_anon_vma(anon_vma);
1161 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1162 * @page: the page to unmap/unlock
1163 * @flags: action and flags
1165 * Find all the mappings of a page using the mapping pointer and the vma chains
1166 * contained in the address_space struct it points to.
1168 * This function is only called from try_to_unmap/try_to_munlock for
1169 * object-based pages.
1170 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1171 * where the page was found will be held for write. So, we won't recheck
1172 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1175 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1177 struct address_space *mapping = page->mapping;
1178 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1179 struct vm_area_struct *vma;
1180 struct prio_tree_iter iter;
1181 int ret = SWAP_AGAIN;
1182 unsigned long cursor;
1183 unsigned long max_nl_cursor = 0;
1184 unsigned long max_nl_size = 0;
1185 unsigned int mapcount;
1187 spin_lock(&mapping->i_mmap_lock);
1188 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1189 unsigned long address = vma_address(page, vma);
1190 if (address == -EFAULT)
1192 ret = try_to_unmap_one(page, vma, address, flags);
1193 if (ret != SWAP_AGAIN || !page_mapped(page))
1197 if (list_empty(&mapping->i_mmap_nonlinear))
1201 * We don't bother to try to find the munlocked page in nonlinears.
1202 * It's costly. Instead, later, page reclaim logic may call
1203 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1205 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1208 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1209 shared.vm_set.list) {
1210 cursor = (unsigned long) vma->vm_private_data;
1211 if (cursor > max_nl_cursor)
1212 max_nl_cursor = cursor;
1213 cursor = vma->vm_end - vma->vm_start;
1214 if (cursor > max_nl_size)
1215 max_nl_size = cursor;
1218 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1224 * We don't try to search for this page in the nonlinear vmas,
1225 * and page_referenced wouldn't have found it anyway. Instead
1226 * just walk the nonlinear vmas trying to age and unmap some.
1227 * The mapcount of the page we came in with is irrelevant,
1228 * but even so use it as a guide to how hard we should try?
1230 mapcount = page_mapcount(page);
1233 cond_resched_lock(&mapping->i_mmap_lock);
1235 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1236 if (max_nl_cursor == 0)
1237 max_nl_cursor = CLUSTER_SIZE;
1240 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1241 shared.vm_set.list) {
1242 cursor = (unsigned long) vma->vm_private_data;
1243 while ( cursor < max_nl_cursor &&
1244 cursor < vma->vm_end - vma->vm_start) {
1245 if (try_to_unmap_cluster(cursor, &mapcount,
1246 vma, page) == SWAP_MLOCK)
1248 cursor += CLUSTER_SIZE;
1249 vma->vm_private_data = (void *) cursor;
1250 if ((int)mapcount <= 0)
1253 vma->vm_private_data = (void *) max_nl_cursor;
1255 cond_resched_lock(&mapping->i_mmap_lock);
1256 max_nl_cursor += CLUSTER_SIZE;
1257 } while (max_nl_cursor <= max_nl_size);
1260 * Don't loop forever (perhaps all the remaining pages are
1261 * in locked vmas). Reset cursor on all unreserved nonlinear
1262 * vmas, now forgetting on which ones it had fallen behind.
1264 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1265 vma->vm_private_data = NULL;
1267 spin_unlock(&mapping->i_mmap_lock);
1272 * try_to_unmap - try to remove all page table mappings to a page
1273 * @page: the page to get unmapped
1274 * @flags: action and flags
1276 * Tries to remove all the page table entries which are mapping this
1277 * page, used in the pageout path. Caller must hold the page lock.
1278 * Return values are:
1280 * SWAP_SUCCESS - we succeeded in removing all mappings
1281 * SWAP_AGAIN - we missed a mapping, try again later
1282 * SWAP_FAIL - the page is unswappable
1283 * SWAP_MLOCK - page is mlocked.
1285 int try_to_unmap(struct page *page, enum ttu_flags flags)
1289 BUG_ON(!PageLocked(page));
1291 if (unlikely(PageKsm(page)))
1292 ret = try_to_unmap_ksm(page, flags);
1293 else if (PageAnon(page))
1294 ret = try_to_unmap_anon(page, flags);
1296 ret = try_to_unmap_file(page, flags);
1297 if (ret != SWAP_MLOCK && !page_mapped(page))
1303 * try_to_munlock - try to munlock a page
1304 * @page: the page to be munlocked
1306 * Called from munlock code. Checks all of the VMAs mapping the page
1307 * to make sure nobody else has this page mlocked. The page will be
1308 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1310 * Return values are:
1312 * SWAP_AGAIN - no vma is holding page mlocked, or,
1313 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1314 * SWAP_FAIL - page cannot be located at present
1315 * SWAP_MLOCK - page is now mlocked.
1317 int try_to_munlock(struct page *page)
1319 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1321 if (unlikely(PageKsm(page)))
1322 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1323 else if (PageAnon(page))
1324 return try_to_unmap_anon(page, TTU_MUNLOCK);
1326 return try_to_unmap_file(page, TTU_MUNLOCK);
1329 #ifdef CONFIG_MIGRATION
1331 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1332 * Called by migrate.c to remove migration ptes, but might be used more later.
1334 static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1335 struct vm_area_struct *, unsigned long, void *), void *arg)
1337 struct anon_vma *anon_vma;
1338 struct anon_vma_chain *avc;
1339 int ret = SWAP_AGAIN;
1342 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1343 * because that depends on page_mapped(); but not all its usages
1344 * are holding mmap_sem, which also gave the necessary guarantee
1345 * (that this anon_vma's slab has not already been destroyed).
1346 * This needs to be reviewed later: avoiding page_lock_anon_vma()
1347 * is risky, and currently limits the usefulness of rmap_walk().
1349 anon_vma = page_anon_vma(page);
1352 spin_lock(&anon_vma->lock);
1353 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1354 struct vm_area_struct *vma = avc->vma;
1355 unsigned long address = vma_address(page, vma);
1356 if (address == -EFAULT)
1358 ret = rmap_one(page, vma, address, arg);
1359 if (ret != SWAP_AGAIN)
1362 spin_unlock(&anon_vma->lock);
1366 static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1367 struct vm_area_struct *, unsigned long, void *), void *arg)
1369 struct address_space *mapping = page->mapping;
1370 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1371 struct vm_area_struct *vma;
1372 struct prio_tree_iter iter;
1373 int ret = SWAP_AGAIN;
1377 spin_lock(&mapping->i_mmap_lock);
1378 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1379 unsigned long address = vma_address(page, vma);
1380 if (address == -EFAULT)
1382 ret = rmap_one(page, vma, address, arg);
1383 if (ret != SWAP_AGAIN)
1387 * No nonlinear handling: being always shared, nonlinear vmas
1388 * never contain migration ptes. Decide what to do about this
1389 * limitation to linear when we need rmap_walk() on nonlinear.
1391 spin_unlock(&mapping->i_mmap_lock);
1395 int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1396 struct vm_area_struct *, unsigned long, void *), void *arg)
1398 VM_BUG_ON(!PageLocked(page));
1400 if (unlikely(PageKsm(page)))
1401 return rmap_walk_ksm(page, rmap_one, arg);
1402 else if (PageAnon(page))
1403 return rmap_walk_anon(page, rmap_one, arg);
1405 return rmap_walk_file(page, rmap_one, arg);
1407 #endif /* CONFIG_MIGRATION */