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 * This VMA had no anon_vma yet. This anon_vma is
137 * the root of any anon_vma tree that might form.
139 anon_vma->root = anon_vma;
142 anon_vma_lock(anon_vma);
143 /* page_table_lock to protect against threads */
144 spin_lock(&mm->page_table_lock);
145 if (likely(!vma->anon_vma)) {
146 vma->anon_vma = anon_vma;
147 avc->anon_vma = anon_vma;
149 list_add(&avc->same_vma, &vma->anon_vma_chain);
150 list_add(&avc->same_anon_vma, &anon_vma->head);
154 spin_unlock(&mm->page_table_lock);
155 anon_vma_unlock(anon_vma);
157 if (unlikely(allocated))
158 anon_vma_free(allocated);
160 anon_vma_chain_free(avc);
165 anon_vma_chain_free(avc);
170 static void anon_vma_chain_link(struct vm_area_struct *vma,
171 struct anon_vma_chain *avc,
172 struct anon_vma *anon_vma)
175 avc->anon_vma = anon_vma;
176 list_add(&avc->same_vma, &vma->anon_vma_chain);
178 anon_vma_lock(anon_vma);
179 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
180 anon_vma_unlock(anon_vma);
184 * Attach the anon_vmas from src to dst.
185 * Returns 0 on success, -ENOMEM on failure.
187 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
189 struct anon_vma_chain *avc, *pavc;
191 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
192 avc = anon_vma_chain_alloc();
195 anon_vma_chain_link(dst, avc, pavc->anon_vma);
200 unlink_anon_vmas(dst);
205 * Attach vma to its own anon_vma, as well as to the anon_vmas that
206 * the corresponding VMA in the parent process is attached to.
207 * Returns 0 on success, non-zero on failure.
209 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
211 struct anon_vma_chain *avc;
212 struct anon_vma *anon_vma;
214 /* Don't bother if the parent process has no anon_vma here. */
219 * First, attach the new VMA to the parent VMA's anon_vmas,
220 * so rmap can find non-COWed pages in child processes.
222 if (anon_vma_clone(vma, pvma))
225 /* Then add our own anon_vma. */
226 anon_vma = anon_vma_alloc();
229 avc = anon_vma_chain_alloc();
231 goto out_error_free_anon_vma;
234 * The root anon_vma's spinlock is the lock actually used when we
235 * lock any of the anon_vmas in this anon_vma tree.
237 anon_vma->root = pvma->anon_vma->root;
238 /* Mark this anon_vma as the one where our new (COWed) pages go. */
239 vma->anon_vma = anon_vma;
240 anon_vma_chain_link(vma, avc, anon_vma);
244 out_error_free_anon_vma:
245 anon_vma_free(anon_vma);
247 unlink_anon_vmas(vma);
251 static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
253 struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
256 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
260 anon_vma_lock(anon_vma);
261 list_del(&anon_vma_chain->same_anon_vma);
263 /* We must garbage collect the anon_vma if it's empty */
264 empty = list_empty(&anon_vma->head) && !anonvma_external_refcount(anon_vma);
265 anon_vma_unlock(anon_vma);
268 anon_vma_free(anon_vma);
271 void unlink_anon_vmas(struct vm_area_struct *vma)
273 struct anon_vma_chain *avc, *next;
276 * Unlink each anon_vma chained to the VMA. This list is ordered
277 * from newest to oldest, ensuring the root anon_vma gets freed last.
279 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
280 anon_vma_unlink(avc);
281 list_del(&avc->same_vma);
282 anon_vma_chain_free(avc);
286 static void anon_vma_ctor(void *data)
288 struct anon_vma *anon_vma = data;
290 spin_lock_init(&anon_vma->lock);
291 anonvma_external_refcount_init(anon_vma);
292 INIT_LIST_HEAD(&anon_vma->head);
295 void __init anon_vma_init(void)
297 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
298 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
299 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
303 * Getting a lock on a stable anon_vma from a page off the LRU is
304 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
306 struct anon_vma *page_lock_anon_vma(struct page *page)
308 struct anon_vma *anon_vma;
309 unsigned long anon_mapping;
312 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
313 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
315 if (!page_mapped(page))
318 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
319 anon_vma_lock(anon_vma);
326 void page_unlock_anon_vma(struct anon_vma *anon_vma)
328 anon_vma_unlock(anon_vma);
333 * At what user virtual address is page expected in @vma?
334 * Returns virtual address or -EFAULT if page's index/offset is not
335 * within the range mapped the @vma.
337 static inline unsigned long
338 vma_address(struct page *page, struct vm_area_struct *vma)
340 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
341 unsigned long address;
343 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
344 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
345 /* page should be within @vma mapping range */
352 * At what user virtual address is page expected in vma?
353 * Caller should check the page is actually part of the vma.
355 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
359 else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
361 vma->vm_file->f_mapping != page->mapping)
365 return vma_address(page, vma);
369 * Check that @page is mapped at @address into @mm.
371 * If @sync is false, page_check_address may perform a racy check to avoid
372 * the page table lock when the pte is not present (helpful when reclaiming
373 * highly shared pages).
375 * On success returns with pte mapped and locked.
377 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
378 unsigned long address, spinlock_t **ptlp, int sync)
386 pgd = pgd_offset(mm, address);
387 if (!pgd_present(*pgd))
390 pud = pud_offset(pgd, address);
391 if (!pud_present(*pud))
394 pmd = pmd_offset(pud, address);
395 if (!pmd_present(*pmd))
398 pte = pte_offset_map(pmd, address);
399 /* Make a quick check before getting the lock */
400 if (!sync && !pte_present(*pte)) {
405 ptl = pte_lockptr(mm, pmd);
407 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
411 pte_unmap_unlock(pte, ptl);
416 * page_mapped_in_vma - check whether a page is really mapped in a VMA
417 * @page: the page to test
418 * @vma: the VMA to test
420 * Returns 1 if the page is mapped into the page tables of the VMA, 0
421 * if the page is not mapped into the page tables of this VMA. Only
422 * valid for normal file or anonymous VMAs.
424 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
426 unsigned long address;
430 address = vma_address(page, vma);
431 if (address == -EFAULT) /* out of vma range */
433 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
434 if (!pte) /* the page is not in this mm */
436 pte_unmap_unlock(pte, ptl);
442 * Subfunctions of page_referenced: page_referenced_one called
443 * repeatedly from either page_referenced_anon or page_referenced_file.
445 int page_referenced_one(struct page *page, struct vm_area_struct *vma,
446 unsigned long address, unsigned int *mapcount,
447 unsigned long *vm_flags)
449 struct mm_struct *mm = vma->vm_mm;
454 pte = page_check_address(page, mm, address, &ptl, 0);
459 * Don't want to elevate referenced for mlocked page that gets this far,
460 * in order that it progresses to try_to_unmap and is moved to the
463 if (vma->vm_flags & VM_LOCKED) {
464 *mapcount = 1; /* break early from loop */
465 *vm_flags |= VM_LOCKED;
469 if (ptep_clear_flush_young_notify(vma, address, pte)) {
471 * Don't treat a reference through a sequentially read
472 * mapping as such. If the page has been used in
473 * another mapping, we will catch it; if this other
474 * mapping is already gone, the unmap path will have
475 * set PG_referenced or activated the page.
477 if (likely(!VM_SequentialReadHint(vma)))
481 /* Pretend the page is referenced if the task has the
482 swap token and is in the middle of a page fault. */
483 if (mm != current->mm && has_swap_token(mm) &&
484 rwsem_is_locked(&mm->mmap_sem))
489 pte_unmap_unlock(pte, ptl);
492 *vm_flags |= vma->vm_flags;
497 static int page_referenced_anon(struct page *page,
498 struct mem_cgroup *mem_cont,
499 unsigned long *vm_flags)
501 unsigned int mapcount;
502 struct anon_vma *anon_vma;
503 struct anon_vma_chain *avc;
506 anon_vma = page_lock_anon_vma(page);
510 mapcount = page_mapcount(page);
511 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
512 struct vm_area_struct *vma = avc->vma;
513 unsigned long address = vma_address(page, vma);
514 if (address == -EFAULT)
517 * If we are reclaiming on behalf of a cgroup, skip
518 * counting on behalf of references from different
521 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
523 referenced += page_referenced_one(page, vma, address,
524 &mapcount, vm_flags);
529 page_unlock_anon_vma(anon_vma);
534 * page_referenced_file - referenced check for object-based rmap
535 * @page: the page we're checking references on.
536 * @mem_cont: target memory controller
537 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
539 * For an object-based mapped page, find all the places it is mapped and
540 * check/clear the referenced flag. This is done by following the page->mapping
541 * pointer, then walking the chain of vmas it holds. It returns the number
542 * of references it found.
544 * This function is only called from page_referenced for object-based pages.
546 static int page_referenced_file(struct page *page,
547 struct mem_cgroup *mem_cont,
548 unsigned long *vm_flags)
550 unsigned int mapcount;
551 struct address_space *mapping = page->mapping;
552 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
553 struct vm_area_struct *vma;
554 struct prio_tree_iter iter;
558 * The caller's checks on page->mapping and !PageAnon have made
559 * sure that this is a file page: the check for page->mapping
560 * excludes the case just before it gets set on an anon page.
562 BUG_ON(PageAnon(page));
565 * The page lock not only makes sure that page->mapping cannot
566 * suddenly be NULLified by truncation, it makes sure that the
567 * structure at mapping cannot be freed and reused yet,
568 * so we can safely take mapping->i_mmap_lock.
570 BUG_ON(!PageLocked(page));
572 spin_lock(&mapping->i_mmap_lock);
575 * i_mmap_lock does not stabilize mapcount at all, but mapcount
576 * is more likely to be accurate if we note it after spinning.
578 mapcount = page_mapcount(page);
580 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
581 unsigned long address = vma_address(page, vma);
582 if (address == -EFAULT)
585 * If we are reclaiming on behalf of a cgroup, skip
586 * counting on behalf of references from different
589 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
591 referenced += page_referenced_one(page, vma, address,
592 &mapcount, vm_flags);
597 spin_unlock(&mapping->i_mmap_lock);
602 * page_referenced - test if the page was referenced
603 * @page: the page to test
604 * @is_locked: caller holds lock on the page
605 * @mem_cont: target memory controller
606 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
608 * Quick test_and_clear_referenced for all mappings to a page,
609 * returns the number of ptes which referenced the page.
611 int page_referenced(struct page *page,
613 struct mem_cgroup *mem_cont,
614 unsigned long *vm_flags)
620 if (page_mapped(page) && page_rmapping(page)) {
621 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
622 we_locked = trylock_page(page);
628 if (unlikely(PageKsm(page)))
629 referenced += page_referenced_ksm(page, mem_cont,
631 else if (PageAnon(page))
632 referenced += page_referenced_anon(page, mem_cont,
634 else if (page->mapping)
635 referenced += page_referenced_file(page, mem_cont,
641 if (page_test_and_clear_young(page))
647 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
648 unsigned long address)
650 struct mm_struct *mm = vma->vm_mm;
655 pte = page_check_address(page, mm, address, &ptl, 1);
659 if (pte_dirty(*pte) || pte_write(*pte)) {
662 flush_cache_page(vma, address, pte_pfn(*pte));
663 entry = ptep_clear_flush_notify(vma, address, pte);
664 entry = pte_wrprotect(entry);
665 entry = pte_mkclean(entry);
666 set_pte_at(mm, address, pte, entry);
670 pte_unmap_unlock(pte, ptl);
675 static int page_mkclean_file(struct address_space *mapping, struct page *page)
677 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
678 struct vm_area_struct *vma;
679 struct prio_tree_iter iter;
682 BUG_ON(PageAnon(page));
684 spin_lock(&mapping->i_mmap_lock);
685 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
686 if (vma->vm_flags & VM_SHARED) {
687 unsigned long address = vma_address(page, vma);
688 if (address == -EFAULT)
690 ret += page_mkclean_one(page, vma, address);
693 spin_unlock(&mapping->i_mmap_lock);
697 int page_mkclean(struct page *page)
701 BUG_ON(!PageLocked(page));
703 if (page_mapped(page)) {
704 struct address_space *mapping = page_mapping(page);
706 ret = page_mkclean_file(mapping, page);
707 if (page_test_dirty(page)) {
708 page_clear_dirty(page);
716 EXPORT_SYMBOL_GPL(page_mkclean);
719 * page_move_anon_rmap - move a page to our anon_vma
720 * @page: the page to move to our anon_vma
721 * @vma: the vma the page belongs to
722 * @address: the user virtual address mapped
724 * When a page belongs exclusively to one process after a COW event,
725 * that page can be moved into the anon_vma that belongs to just that
726 * process, so the rmap code will not search the parent or sibling
729 void page_move_anon_rmap(struct page *page,
730 struct vm_area_struct *vma, unsigned long address)
732 struct anon_vma *anon_vma = vma->anon_vma;
734 VM_BUG_ON(!PageLocked(page));
735 VM_BUG_ON(!anon_vma);
736 VM_BUG_ON(page->index != linear_page_index(vma, address));
738 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
739 page->mapping = (struct address_space *) anon_vma;
743 * __page_set_anon_rmap - setup new anonymous rmap
744 * @page: the page to add the mapping to
745 * @vma: the vm area in which the mapping is added
746 * @address: the user virtual address mapped
747 * @exclusive: the page is exclusively owned by the current process
749 static void __page_set_anon_rmap(struct page *page,
750 struct vm_area_struct *vma, unsigned long address, int exclusive)
752 struct anon_vma *anon_vma = vma->anon_vma;
757 * If the page isn't exclusively mapped into this vma,
758 * we must use the _oldest_ possible anon_vma for the
761 * So take the last AVC chain entry in the vma, which is
762 * the deepest ancestor, and use the anon_vma from that.
765 struct anon_vma_chain *avc;
766 avc = list_entry(vma->anon_vma_chain.prev, struct anon_vma_chain, same_vma);
767 anon_vma = avc->anon_vma;
770 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
771 page->mapping = (struct address_space *) anon_vma;
772 page->index = linear_page_index(vma, address);
776 * __page_check_anon_rmap - sanity check anonymous rmap addition
777 * @page: the page to add the mapping to
778 * @vma: the vm area in which the mapping is added
779 * @address: the user virtual address mapped
781 static void __page_check_anon_rmap(struct page *page,
782 struct vm_area_struct *vma, unsigned long address)
784 #ifdef CONFIG_DEBUG_VM
786 * The page's anon-rmap details (mapping and index) are guaranteed to
787 * be set up correctly at this point.
789 * We have exclusion against page_add_anon_rmap because the caller
790 * always holds the page locked, except if called from page_dup_rmap,
791 * in which case the page is already known to be setup.
793 * We have exclusion against page_add_new_anon_rmap because those pages
794 * are initially only visible via the pagetables, and the pte is locked
795 * over the call to page_add_new_anon_rmap.
797 BUG_ON(page->index != linear_page_index(vma, address));
802 * page_add_anon_rmap - add pte mapping to an 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 * The caller needs to hold the pte lock, and the page must be locked in
808 * the anon_vma case: to serialize mapping,index checking after setting,
809 * and to ensure that PageAnon is not being upgraded racily to PageKsm
810 * (but PageKsm is never downgraded to PageAnon).
812 void page_add_anon_rmap(struct page *page,
813 struct vm_area_struct *vma, unsigned long address)
815 int first = atomic_inc_and_test(&page->_mapcount);
817 __inc_zone_page_state(page, NR_ANON_PAGES);
818 if (unlikely(PageKsm(page)))
821 VM_BUG_ON(!PageLocked(page));
822 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
824 __page_set_anon_rmap(page, vma, address, 0);
826 __page_check_anon_rmap(page, vma, address);
830 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
831 * @page: the page to add the mapping to
832 * @vma: the vm area in which the mapping is added
833 * @address: the user virtual address mapped
835 * Same as page_add_anon_rmap but must only be called on *new* pages.
836 * This means the inc-and-test can be bypassed.
837 * Page does not have to be locked.
839 void page_add_new_anon_rmap(struct page *page,
840 struct vm_area_struct *vma, unsigned long address)
842 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
843 SetPageSwapBacked(page);
844 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
845 __inc_zone_page_state(page, NR_ANON_PAGES);
846 __page_set_anon_rmap(page, vma, address, 1);
847 if (page_evictable(page, vma))
848 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
850 add_page_to_unevictable_list(page);
854 * page_add_file_rmap - add pte mapping to a file page
855 * @page: the page to add the mapping to
857 * The caller needs to hold the pte lock.
859 void page_add_file_rmap(struct page *page)
861 if (atomic_inc_and_test(&page->_mapcount)) {
862 __inc_zone_page_state(page, NR_FILE_MAPPED);
863 mem_cgroup_update_file_mapped(page, 1);
868 * page_remove_rmap - take down pte mapping from a page
869 * @page: page to remove mapping from
871 * The caller needs to hold the pte lock.
873 void page_remove_rmap(struct page *page)
875 /* page still mapped by someone else? */
876 if (!atomic_add_negative(-1, &page->_mapcount))
880 * Now that the last pte has gone, s390 must transfer dirty
881 * flag from storage key to struct page. We can usually skip
882 * this if the page is anon, so about to be freed; but perhaps
883 * not if it's in swapcache - there might be another pte slot
884 * containing the swap entry, but page not yet written to swap.
886 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
887 page_clear_dirty(page);
888 set_page_dirty(page);
890 if (PageAnon(page)) {
891 mem_cgroup_uncharge_page(page);
892 __dec_zone_page_state(page, NR_ANON_PAGES);
894 __dec_zone_page_state(page, NR_FILE_MAPPED);
895 mem_cgroup_update_file_mapped(page, -1);
898 * It would be tidy to reset the PageAnon mapping here,
899 * but that might overwrite a racing page_add_anon_rmap
900 * which increments mapcount after us but sets mapping
901 * before us: so leave the reset to free_hot_cold_page,
902 * and remember that it's only reliable while mapped.
903 * Leaving it set also helps swapoff to reinstate ptes
904 * faster for those pages still in swapcache.
909 * Subfunctions of try_to_unmap: try_to_unmap_one called
910 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
912 int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
913 unsigned long address, enum ttu_flags flags)
915 struct mm_struct *mm = vma->vm_mm;
919 int ret = SWAP_AGAIN;
921 pte = page_check_address(page, mm, address, &ptl, 0);
926 * If the page is mlock()d, we cannot swap it out.
927 * If it's recently referenced (perhaps page_referenced
928 * skipped over this mm) then we should reactivate it.
930 if (!(flags & TTU_IGNORE_MLOCK)) {
931 if (vma->vm_flags & VM_LOCKED)
934 if (TTU_ACTION(flags) == TTU_MUNLOCK)
937 if (!(flags & TTU_IGNORE_ACCESS)) {
938 if (ptep_clear_flush_young_notify(vma, address, pte)) {
944 /* Nuke the page table entry. */
945 flush_cache_page(vma, address, page_to_pfn(page));
946 pteval = ptep_clear_flush_notify(vma, address, pte);
948 /* Move the dirty bit to the physical page now the pte is gone. */
949 if (pte_dirty(pteval))
950 set_page_dirty(page);
952 /* Update high watermark before we lower rss */
953 update_hiwater_rss(mm);
955 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
957 dec_mm_counter(mm, MM_ANONPAGES);
959 dec_mm_counter(mm, MM_FILEPAGES);
960 set_pte_at(mm, address, pte,
961 swp_entry_to_pte(make_hwpoison_entry(page)));
962 } else if (PageAnon(page)) {
963 swp_entry_t entry = { .val = page_private(page) };
965 if (PageSwapCache(page)) {
967 * Store the swap location in the pte.
968 * See handle_pte_fault() ...
970 if (swap_duplicate(entry) < 0) {
971 set_pte_at(mm, address, pte, pteval);
975 if (list_empty(&mm->mmlist)) {
976 spin_lock(&mmlist_lock);
977 if (list_empty(&mm->mmlist))
978 list_add(&mm->mmlist, &init_mm.mmlist);
979 spin_unlock(&mmlist_lock);
981 dec_mm_counter(mm, MM_ANONPAGES);
982 inc_mm_counter(mm, MM_SWAPENTS);
983 } else if (PAGE_MIGRATION) {
985 * Store the pfn of the page in a special migration
986 * pte. do_swap_page() will wait until the migration
987 * pte is removed and then restart fault handling.
989 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
990 entry = make_migration_entry(page, pte_write(pteval));
992 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
993 BUG_ON(pte_file(*pte));
994 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
995 /* Establish migration entry for a file page */
997 entry = make_migration_entry(page, pte_write(pteval));
998 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1000 dec_mm_counter(mm, MM_FILEPAGES);
1002 page_remove_rmap(page);
1003 page_cache_release(page);
1006 pte_unmap_unlock(pte, ptl);
1011 pte_unmap_unlock(pte, ptl);
1015 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1016 * unstable result and race. Plus, We can't wait here because
1017 * we now hold anon_vma->lock or mapping->i_mmap_lock.
1018 * if trylock failed, the page remain in evictable lru and later
1019 * vmscan could retry to move the page to unevictable lru if the
1020 * page is actually mlocked.
1022 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1023 if (vma->vm_flags & VM_LOCKED) {
1024 mlock_vma_page(page);
1027 up_read(&vma->vm_mm->mmap_sem);
1033 * objrmap doesn't work for nonlinear VMAs because the assumption that
1034 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1035 * Consequently, given a particular page and its ->index, we cannot locate the
1036 * ptes which are mapping that page without an exhaustive linear search.
1038 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1039 * maps the file to which the target page belongs. The ->vm_private_data field
1040 * holds the current cursor into that scan. Successive searches will circulate
1041 * around the vma's virtual address space.
1043 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1044 * more scanning pressure is placed against them as well. Eventually pages
1045 * will become fully unmapped and are eligible for eviction.
1047 * For very sparsely populated VMAs this is a little inefficient - chances are
1048 * there there won't be many ptes located within the scan cluster. In this case
1049 * maybe we could scan further - to the end of the pte page, perhaps.
1051 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1052 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1053 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1054 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1056 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1057 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1059 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1060 struct vm_area_struct *vma, struct page *check_page)
1062 struct mm_struct *mm = vma->vm_mm;
1070 unsigned long address;
1072 int ret = SWAP_AGAIN;
1075 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1076 end = address + CLUSTER_SIZE;
1077 if (address < vma->vm_start)
1078 address = vma->vm_start;
1079 if (end > vma->vm_end)
1082 pgd = pgd_offset(mm, address);
1083 if (!pgd_present(*pgd))
1086 pud = pud_offset(pgd, address);
1087 if (!pud_present(*pud))
1090 pmd = pmd_offset(pud, address);
1091 if (!pmd_present(*pmd))
1095 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1096 * keep the sem while scanning the cluster for mlocking pages.
1098 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1099 locked_vma = (vma->vm_flags & VM_LOCKED);
1101 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1104 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1106 /* Update high watermark before we lower rss */
1107 update_hiwater_rss(mm);
1109 for (; address < end; pte++, address += PAGE_SIZE) {
1110 if (!pte_present(*pte))
1112 page = vm_normal_page(vma, address, *pte);
1113 BUG_ON(!page || PageAnon(page));
1116 mlock_vma_page(page); /* no-op if already mlocked */
1117 if (page == check_page)
1119 continue; /* don't unmap */
1122 if (ptep_clear_flush_young_notify(vma, address, pte))
1125 /* Nuke the page table entry. */
1126 flush_cache_page(vma, address, pte_pfn(*pte));
1127 pteval = ptep_clear_flush_notify(vma, address, pte);
1129 /* If nonlinear, store the file page offset in the pte. */
1130 if (page->index != linear_page_index(vma, address))
1131 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1133 /* Move the dirty bit to the physical page now the pte is gone. */
1134 if (pte_dirty(pteval))
1135 set_page_dirty(page);
1137 page_remove_rmap(page);
1138 page_cache_release(page);
1139 dec_mm_counter(mm, MM_FILEPAGES);
1142 pte_unmap_unlock(pte - 1, ptl);
1144 up_read(&vma->vm_mm->mmap_sem);
1148 static bool is_vma_temporary_stack(struct vm_area_struct *vma)
1150 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1155 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1156 VM_STACK_INCOMPLETE_SETUP)
1163 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1165 * @page: the page to unmap/unlock
1166 * @flags: action and flags
1168 * Find all the mappings of a page using the mapping pointer and the vma chains
1169 * contained in the anon_vma struct it points to.
1171 * This function is only called from try_to_unmap/try_to_munlock for
1173 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1174 * where the page was found will be held for write. So, we won't recheck
1175 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1178 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1180 struct anon_vma *anon_vma;
1181 struct anon_vma_chain *avc;
1182 int ret = SWAP_AGAIN;
1184 anon_vma = page_lock_anon_vma(page);
1188 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1189 struct vm_area_struct *vma = avc->vma;
1190 unsigned long address;
1193 * During exec, a temporary VMA is setup and later moved.
1194 * The VMA is moved under the anon_vma lock but not the
1195 * page tables leading to a race where migration cannot
1196 * find the migration ptes. Rather than increasing the
1197 * locking requirements of exec(), migration skips
1198 * temporary VMAs until after exec() completes.
1200 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1201 is_vma_temporary_stack(vma))
1204 address = vma_address(page, vma);
1205 if (address == -EFAULT)
1207 ret = try_to_unmap_one(page, vma, address, flags);
1208 if (ret != SWAP_AGAIN || !page_mapped(page))
1212 page_unlock_anon_vma(anon_vma);
1217 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1218 * @page: the page to unmap/unlock
1219 * @flags: action and flags
1221 * Find all the mappings of a page using the mapping pointer and the vma chains
1222 * contained in the address_space struct it points to.
1224 * This function is only called from try_to_unmap/try_to_munlock for
1225 * object-based pages.
1226 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1227 * where the page was found will be held for write. So, we won't recheck
1228 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1231 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1233 struct address_space *mapping = page->mapping;
1234 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1235 struct vm_area_struct *vma;
1236 struct prio_tree_iter iter;
1237 int ret = SWAP_AGAIN;
1238 unsigned long cursor;
1239 unsigned long max_nl_cursor = 0;
1240 unsigned long max_nl_size = 0;
1241 unsigned int mapcount;
1243 spin_lock(&mapping->i_mmap_lock);
1244 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1245 unsigned long address = vma_address(page, vma);
1246 if (address == -EFAULT)
1248 ret = try_to_unmap_one(page, vma, address, flags);
1249 if (ret != SWAP_AGAIN || !page_mapped(page))
1253 if (list_empty(&mapping->i_mmap_nonlinear))
1257 * We don't bother to try to find the munlocked page in nonlinears.
1258 * It's costly. Instead, later, page reclaim logic may call
1259 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1261 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1264 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1265 shared.vm_set.list) {
1266 cursor = (unsigned long) vma->vm_private_data;
1267 if (cursor > max_nl_cursor)
1268 max_nl_cursor = cursor;
1269 cursor = vma->vm_end - vma->vm_start;
1270 if (cursor > max_nl_size)
1271 max_nl_size = cursor;
1274 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1280 * We don't try to search for this page in the nonlinear vmas,
1281 * and page_referenced wouldn't have found it anyway. Instead
1282 * just walk the nonlinear vmas trying to age and unmap some.
1283 * The mapcount of the page we came in with is irrelevant,
1284 * but even so use it as a guide to how hard we should try?
1286 mapcount = page_mapcount(page);
1289 cond_resched_lock(&mapping->i_mmap_lock);
1291 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1292 if (max_nl_cursor == 0)
1293 max_nl_cursor = CLUSTER_SIZE;
1296 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1297 shared.vm_set.list) {
1298 cursor = (unsigned long) vma->vm_private_data;
1299 while ( cursor < max_nl_cursor &&
1300 cursor < vma->vm_end - vma->vm_start) {
1301 if (try_to_unmap_cluster(cursor, &mapcount,
1302 vma, page) == SWAP_MLOCK)
1304 cursor += CLUSTER_SIZE;
1305 vma->vm_private_data = (void *) cursor;
1306 if ((int)mapcount <= 0)
1309 vma->vm_private_data = (void *) max_nl_cursor;
1311 cond_resched_lock(&mapping->i_mmap_lock);
1312 max_nl_cursor += CLUSTER_SIZE;
1313 } while (max_nl_cursor <= max_nl_size);
1316 * Don't loop forever (perhaps all the remaining pages are
1317 * in locked vmas). Reset cursor on all unreserved nonlinear
1318 * vmas, now forgetting on which ones it had fallen behind.
1320 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1321 vma->vm_private_data = NULL;
1323 spin_unlock(&mapping->i_mmap_lock);
1328 * try_to_unmap - try to remove all page table mappings to a page
1329 * @page: the page to get unmapped
1330 * @flags: action and flags
1332 * Tries to remove all the page table entries which are mapping this
1333 * page, used in the pageout path. Caller must hold the page lock.
1334 * Return values are:
1336 * SWAP_SUCCESS - we succeeded in removing all mappings
1337 * SWAP_AGAIN - we missed a mapping, try again later
1338 * SWAP_FAIL - the page is unswappable
1339 * SWAP_MLOCK - page is mlocked.
1341 int try_to_unmap(struct page *page, enum ttu_flags flags)
1345 BUG_ON(!PageLocked(page));
1347 if (unlikely(PageKsm(page)))
1348 ret = try_to_unmap_ksm(page, flags);
1349 else if (PageAnon(page))
1350 ret = try_to_unmap_anon(page, flags);
1352 ret = try_to_unmap_file(page, flags);
1353 if (ret != SWAP_MLOCK && !page_mapped(page))
1359 * try_to_munlock - try to munlock a page
1360 * @page: the page to be munlocked
1362 * Called from munlock code. Checks all of the VMAs mapping the page
1363 * to make sure nobody else has this page mlocked. The page will be
1364 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1366 * Return values are:
1368 * SWAP_AGAIN - no vma is holding page mlocked, or,
1369 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1370 * SWAP_FAIL - page cannot be located at present
1371 * SWAP_MLOCK - page is now mlocked.
1373 int try_to_munlock(struct page *page)
1375 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1377 if (unlikely(PageKsm(page)))
1378 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1379 else if (PageAnon(page))
1380 return try_to_unmap_anon(page, TTU_MUNLOCK);
1382 return try_to_unmap_file(page, TTU_MUNLOCK);
1385 #ifdef CONFIG_MIGRATION
1387 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1388 * Called by migrate.c to remove migration ptes, but might be used more later.
1390 static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1391 struct vm_area_struct *, unsigned long, void *), void *arg)
1393 struct anon_vma *anon_vma;
1394 struct anon_vma_chain *avc;
1395 int ret = SWAP_AGAIN;
1398 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1399 * because that depends on page_mapped(); but not all its usages
1400 * are holding mmap_sem. Users without mmap_sem are required to
1401 * take a reference count to prevent the anon_vma disappearing
1403 anon_vma = page_anon_vma(page);
1406 anon_vma_lock(anon_vma);
1407 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1408 struct vm_area_struct *vma = avc->vma;
1409 unsigned long address = vma_address(page, vma);
1410 if (address == -EFAULT)
1412 ret = rmap_one(page, vma, address, arg);
1413 if (ret != SWAP_AGAIN)
1416 anon_vma_unlock(anon_vma);
1420 static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1421 struct vm_area_struct *, unsigned long, void *), void *arg)
1423 struct address_space *mapping = page->mapping;
1424 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1425 struct vm_area_struct *vma;
1426 struct prio_tree_iter iter;
1427 int ret = SWAP_AGAIN;
1431 spin_lock(&mapping->i_mmap_lock);
1432 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1433 unsigned long address = vma_address(page, vma);
1434 if (address == -EFAULT)
1436 ret = rmap_one(page, vma, address, arg);
1437 if (ret != SWAP_AGAIN)
1441 * No nonlinear handling: being always shared, nonlinear vmas
1442 * never contain migration ptes. Decide what to do about this
1443 * limitation to linear when we need rmap_walk() on nonlinear.
1445 spin_unlock(&mapping->i_mmap_lock);
1449 int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1450 struct vm_area_struct *, unsigned long, void *), void *arg)
1452 VM_BUG_ON(!PageLocked(page));
1454 if (unlikely(PageKsm(page)))
1455 return rmap_walk_ksm(page, rmap_one, arg);
1456 else if (PageAnon(page))
1457 return rmap_walk_anon(page, rmap_one, arg);
1459 return rmap_walk_file(page, rmap_one, arg);
1461 #endif /* CONFIG_MIGRATION */