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)
25 * page->flags PG_locked (lock_page)
26 * mapping->i_mmap_mutex
28 * mm->page_table_lock or pte_lock
29 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30 * swap_lock (in swap_duplicate, swap_info_get)
31 * mmlist_lock (in mmput, drain_mmlist and others)
32 * mapping->private_lock (in __set_page_dirty_buffers)
33 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
34 * bdi.wb->list_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 bdi.wb->list_lock in __sync_single_inode)
40 * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon)
46 #include <linux/pagemap.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/slab.h>
50 #include <linux/init.h>
51 #include <linux/ksm.h>
52 #include <linux/rmap.h>
53 #include <linux/rcupdate.h>
54 #include <linux/export.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/migrate.h>
58 #include <linux/hugetlb.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 struct anon_vma *anon_vma;
71 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
73 atomic_set(&anon_vma->refcount, 1);
75 * Initialise the anon_vma root to point to itself. If called
76 * from fork, the root will be reset to the parents anon_vma.
78 anon_vma->root = anon_vma;
84 static inline void anon_vma_free(struct anon_vma *anon_vma)
86 VM_BUG_ON(atomic_read(&anon_vma->refcount));
89 * Synchronize against page_lock_anon_vma() such that
90 * we can safely hold the lock without the anon_vma getting
93 * Relies on the full mb implied by the atomic_dec_and_test() from
94 * put_anon_vma() against the acquire barrier implied by
95 * mutex_trylock() from page_lock_anon_vma(). This orders:
97 * page_lock_anon_vma() VS put_anon_vma()
98 * mutex_trylock() atomic_dec_and_test()
100 * atomic_read() mutex_is_locked()
102 * LOCK should suffice since the actual taking of the lock must
103 * happen _before_ what follows.
105 if (mutex_is_locked(&anon_vma->root->mutex)) {
106 anon_vma_lock(anon_vma);
107 anon_vma_unlock(anon_vma);
110 kmem_cache_free(anon_vma_cachep, anon_vma);
113 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
115 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
118 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
120 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
123 static void anon_vma_chain_link(struct vm_area_struct *vma,
124 struct anon_vma_chain *avc,
125 struct anon_vma *anon_vma)
128 avc->anon_vma = anon_vma;
129 list_add(&avc->same_vma, &vma->anon_vma_chain);
130 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
134 * anon_vma_prepare - attach an anon_vma to a memory region
135 * @vma: the memory region in question
137 * This makes sure the memory mapping described by 'vma' has
138 * an 'anon_vma' attached to it, so that we can associate the
139 * anonymous pages mapped into it with that anon_vma.
141 * The common case will be that we already have one, but if
142 * not we either need to find an adjacent mapping that we
143 * can re-use the anon_vma from (very common when the only
144 * reason for splitting a vma has been mprotect()), or we
145 * allocate a new one.
147 * Anon-vma allocations are very subtle, because we may have
148 * optimistically looked up an anon_vma in page_lock_anon_vma()
149 * and that may actually touch the spinlock even in the newly
150 * allocated vma (it depends on RCU to make sure that the
151 * anon_vma isn't actually destroyed).
153 * As a result, we need to do proper anon_vma locking even
154 * for the new allocation. At the same time, we do not want
155 * to do any locking for the common case of already having
158 * This must be called with the mmap_sem held for reading.
160 int anon_vma_prepare(struct vm_area_struct *vma)
162 struct anon_vma *anon_vma = vma->anon_vma;
163 struct anon_vma_chain *avc;
166 if (unlikely(!anon_vma)) {
167 struct mm_struct *mm = vma->vm_mm;
168 struct anon_vma *allocated;
170 avc = anon_vma_chain_alloc(GFP_KERNEL);
174 anon_vma = find_mergeable_anon_vma(vma);
177 anon_vma = anon_vma_alloc();
178 if (unlikely(!anon_vma))
179 goto out_enomem_free_avc;
180 allocated = anon_vma;
183 anon_vma_lock(anon_vma);
184 /* page_table_lock to protect against threads */
185 spin_lock(&mm->page_table_lock);
186 if (likely(!vma->anon_vma)) {
187 vma->anon_vma = anon_vma;
188 anon_vma_chain_link(vma, avc, anon_vma);
192 spin_unlock(&mm->page_table_lock);
193 anon_vma_unlock(anon_vma);
195 if (unlikely(allocated))
196 put_anon_vma(allocated);
198 anon_vma_chain_free(avc);
203 anon_vma_chain_free(avc);
209 * This is a useful helper function for locking the anon_vma root as
210 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
213 * Such anon_vma's should have the same root, so you'd expect to see
214 * just a single mutex_lock for the whole traversal.
216 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
218 struct anon_vma *new_root = anon_vma->root;
219 if (new_root != root) {
220 if (WARN_ON_ONCE(root))
221 mutex_unlock(&root->mutex);
223 mutex_lock(&root->mutex);
228 static inline void unlock_anon_vma_root(struct anon_vma *root)
231 mutex_unlock(&root->mutex);
235 * Attach the anon_vmas from src to dst.
236 * Returns 0 on success, -ENOMEM on failure.
238 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
240 struct anon_vma_chain *avc, *pavc;
241 struct anon_vma *root = NULL;
243 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
244 struct anon_vma *anon_vma;
246 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
247 if (unlikely(!avc)) {
248 unlock_anon_vma_root(root);
250 avc = anon_vma_chain_alloc(GFP_KERNEL);
254 anon_vma = pavc->anon_vma;
255 root = lock_anon_vma_root(root, anon_vma);
256 anon_vma_chain_link(dst, avc, anon_vma);
258 unlock_anon_vma_root(root);
262 unlink_anon_vmas(dst);
267 * Attach vma to its own anon_vma, as well as to the anon_vmas that
268 * the corresponding VMA in the parent process is attached to.
269 * Returns 0 on success, non-zero on failure.
271 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
273 struct anon_vma_chain *avc;
274 struct anon_vma *anon_vma;
276 /* Don't bother if the parent process has no anon_vma here. */
281 * First, attach the new VMA to the parent VMA's anon_vmas,
282 * so rmap can find non-COWed pages in child processes.
284 if (anon_vma_clone(vma, pvma))
287 /* Then add our own anon_vma. */
288 anon_vma = anon_vma_alloc();
291 avc = anon_vma_chain_alloc(GFP_KERNEL);
293 goto out_error_free_anon_vma;
296 * The root anon_vma's spinlock is the lock actually used when we
297 * lock any of the anon_vmas in this anon_vma tree.
299 anon_vma->root = pvma->anon_vma->root;
301 * With refcounts, an anon_vma can stay around longer than the
302 * process it belongs to. The root anon_vma needs to be pinned until
303 * this anon_vma is freed, because the lock lives in the root.
305 get_anon_vma(anon_vma->root);
306 /* Mark this anon_vma as the one where our new (COWed) pages go. */
307 vma->anon_vma = anon_vma;
308 anon_vma_lock(anon_vma);
309 anon_vma_chain_link(vma, avc, anon_vma);
310 anon_vma_unlock(anon_vma);
314 out_error_free_anon_vma:
315 put_anon_vma(anon_vma);
317 unlink_anon_vmas(vma);
321 void unlink_anon_vmas(struct vm_area_struct *vma)
323 struct anon_vma_chain *avc, *next;
324 struct anon_vma *root = NULL;
327 * Unlink each anon_vma chained to the VMA. This list is ordered
328 * from newest to oldest, ensuring the root anon_vma gets freed last.
330 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
331 struct anon_vma *anon_vma = avc->anon_vma;
333 root = lock_anon_vma_root(root, anon_vma);
334 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
337 * Leave empty anon_vmas on the list - we'll need
338 * to free them outside the lock.
340 if (RB_EMPTY_ROOT(&anon_vma->rb_root))
343 list_del(&avc->same_vma);
344 anon_vma_chain_free(avc);
346 unlock_anon_vma_root(root);
349 * Iterate the list once more, it now only contains empty and unlinked
350 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
351 * needing to acquire the anon_vma->root->mutex.
353 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
354 struct anon_vma *anon_vma = avc->anon_vma;
356 put_anon_vma(anon_vma);
358 list_del(&avc->same_vma);
359 anon_vma_chain_free(avc);
363 static void anon_vma_ctor(void *data)
365 struct anon_vma *anon_vma = data;
367 mutex_init(&anon_vma->mutex);
368 atomic_set(&anon_vma->refcount, 0);
370 atomic_set(&anon_vma->swapra_miss, 0);
372 anon_vma->rb_root = RB_ROOT;
375 void __init anon_vma_init(void)
377 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
378 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
379 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
383 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
385 * Since there is no serialization what so ever against page_remove_rmap()
386 * the best this function can do is return a locked anon_vma that might
387 * have been relevant to this page.
389 * The page might have been remapped to a different anon_vma or the anon_vma
390 * returned may already be freed (and even reused).
392 * In case it was remapped to a different anon_vma, the new anon_vma will be a
393 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
394 * ensure that any anon_vma obtained from the page will still be valid for as
395 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
397 * All users of this function must be very careful when walking the anon_vma
398 * chain and verify that the page in question is indeed mapped in it
399 * [ something equivalent to page_mapped_in_vma() ].
401 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
402 * that the anon_vma pointer from page->mapping is valid if there is a
403 * mapcount, we can dereference the anon_vma after observing those.
405 struct anon_vma *page_get_anon_vma(struct page *page)
407 struct anon_vma *anon_vma = NULL;
408 unsigned long anon_mapping;
411 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
412 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
414 if (!page_mapped(page))
417 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
418 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
424 * If this page is still mapped, then its anon_vma cannot have been
425 * freed. But if it has been unmapped, we have no security against the
426 * anon_vma structure being freed and reused (for another anon_vma:
427 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
428 * above cannot corrupt).
430 if (!page_mapped(page)) {
431 put_anon_vma(anon_vma);
441 * Similar to page_get_anon_vma() except it locks the anon_vma.
443 * Its a little more complex as it tries to keep the fast path to a single
444 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
445 * reference like with page_get_anon_vma() and then block on the mutex.
447 struct anon_vma *page_lock_anon_vma(struct page *page)
449 struct anon_vma *anon_vma = NULL;
450 struct anon_vma *root_anon_vma;
451 unsigned long anon_mapping;
454 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
455 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
457 if (!page_mapped(page))
460 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
461 root_anon_vma = ACCESS_ONCE(anon_vma->root);
462 if (mutex_trylock(&root_anon_vma->mutex)) {
464 * If the page is still mapped, then this anon_vma is still
465 * its anon_vma, and holding the mutex ensures that it will
466 * not go away, see anon_vma_free().
468 if (!page_mapped(page)) {
469 mutex_unlock(&root_anon_vma->mutex);
475 /* trylock failed, we got to sleep */
476 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
481 if (!page_mapped(page)) {
482 put_anon_vma(anon_vma);
487 /* we pinned the anon_vma, its safe to sleep */
489 anon_vma_lock(anon_vma);
491 if (atomic_dec_and_test(&anon_vma->refcount)) {
493 * Oops, we held the last refcount, release the lock
494 * and bail -- can't simply use put_anon_vma() because
495 * we'll deadlock on the anon_vma_lock() recursion.
497 anon_vma_unlock(anon_vma);
498 __put_anon_vma(anon_vma);
509 void page_unlock_anon_vma(struct anon_vma *anon_vma)
511 anon_vma_unlock(anon_vma);
515 * At what user virtual address is page expected in @vma?
517 static inline unsigned long
518 __vma_address(struct page *page, struct vm_area_struct *vma)
520 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
522 if (unlikely(is_vm_hugetlb_page(vma)))
523 pgoff = page->index << huge_page_order(page_hstate(page));
525 return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
529 vma_address(struct page *page, struct vm_area_struct *vma)
531 unsigned long address = __vma_address(page, vma);
533 /* page should be within @vma mapping range */
534 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
540 * At what user virtual address is page expected in vma?
541 * Caller should check the page is actually part of the vma.
543 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
545 unsigned long address;
546 if (PageAnon(page)) {
547 struct anon_vma *page__anon_vma = page_anon_vma(page);
549 * Note: swapoff's unuse_vma() is more efficient with this
550 * check, and needs it to match anon_vma when KSM is active.
552 if (!vma->anon_vma || !page__anon_vma ||
553 vma->anon_vma->root != page__anon_vma->root)
555 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
557 vma->vm_file->f_mapping != page->mapping)
561 address = __vma_address(page, vma);
562 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
568 * Check that @page is mapped at @address into @mm.
570 * If @sync is false, page_check_address may perform a racy check to avoid
571 * the page table lock when the pte is not present (helpful when reclaiming
572 * highly shared pages).
574 * On success returns with pte mapped and locked.
576 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
577 unsigned long address, spinlock_t **ptlp, int sync)
585 if (unlikely(PageHuge(page))) {
586 pte = huge_pte_offset(mm, address);
587 ptl = &mm->page_table_lock;
591 pgd = pgd_offset(mm, address);
592 if (!pgd_present(*pgd))
595 pud = pud_offset(pgd, address);
596 if (!pud_present(*pud))
599 pmd = pmd_offset(pud, address);
600 if (!pmd_present(*pmd))
602 if (pmd_trans_huge(*pmd))
605 pte = pte_offset_map(pmd, address);
606 /* Make a quick check before getting the lock */
607 if (!sync && !pte_present(*pte)) {
612 ptl = pte_lockptr(mm, pmd);
615 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
619 pte_unmap_unlock(pte, ptl);
624 * page_mapped_in_vma - check whether a page is really mapped in a VMA
625 * @page: the page to test
626 * @vma: the VMA to test
628 * Returns 1 if the page is mapped into the page tables of the VMA, 0
629 * if the page is not mapped into the page tables of this VMA. Only
630 * valid for normal file or anonymous VMAs.
632 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
634 unsigned long address;
638 address = __vma_address(page, vma);
639 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
641 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
642 if (!pte) /* the page is not in this mm */
644 pte_unmap_unlock(pte, ptl);
650 * Subfunctions of page_referenced: page_referenced_one called
651 * repeatedly from either page_referenced_anon or page_referenced_file.
653 int page_referenced_one(struct page *page, struct vm_area_struct *vma,
654 unsigned long address, unsigned int *mapcount,
655 unsigned long *vm_flags)
657 struct mm_struct *mm = vma->vm_mm;
660 if (unlikely(PageTransHuge(page))) {
663 spin_lock(&mm->page_table_lock);
665 * rmap might return false positives; we must filter
666 * these out using page_check_address_pmd().
668 pmd = page_check_address_pmd(page, mm, address,
669 PAGE_CHECK_ADDRESS_PMD_FLAG);
671 spin_unlock(&mm->page_table_lock);
675 if (vma->vm_flags & VM_LOCKED) {
676 spin_unlock(&mm->page_table_lock);
677 *mapcount = 0; /* break early from loop */
678 *vm_flags |= VM_LOCKED;
682 /* go ahead even if the pmd is pmd_trans_splitting() */
683 if (pmdp_clear_flush_young_notify(vma, address, pmd))
685 spin_unlock(&mm->page_table_lock);
691 * rmap might return false positives; we must filter
692 * these out using page_check_address().
694 pte = page_check_address(page, mm, address, &ptl, 0);
698 if (vma->vm_flags & VM_LOCKED) {
699 pte_unmap_unlock(pte, ptl);
700 *mapcount = 0; /* break early from loop */
701 *vm_flags |= VM_LOCKED;
705 if (ptep_clear_flush_young_notify(vma, address, pte)) {
707 * Don't treat a reference through a sequentially read
708 * mapping as such. If the page has been used in
709 * another mapping, we will catch it; if this other
710 * mapping is already gone, the unmap path will have
711 * set PG_referenced or activated the page.
713 if (likely(!VM_SequentialReadHint(vma)))
716 pte_unmap_unlock(pte, ptl);
722 *vm_flags |= vma->vm_flags;
727 static int page_referenced_anon(struct page *page,
728 struct mem_cgroup *memcg,
729 unsigned long *vm_flags)
731 unsigned int mapcount;
732 struct anon_vma *anon_vma;
734 struct anon_vma_chain *avc;
737 anon_vma = page_lock_anon_vma(page);
741 mapcount = page_mapcount(page);
742 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
743 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
744 struct vm_area_struct *vma = avc->vma;
745 unsigned long address = vma_address(page, vma);
747 * If we are reclaiming on behalf of a cgroup, skip
748 * counting on behalf of references from different
751 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
753 referenced += page_referenced_one(page, vma, address,
754 &mapcount, vm_flags);
759 page_unlock_anon_vma(anon_vma);
764 * page_referenced_file - referenced check for object-based rmap
765 * @page: the page we're checking references on.
766 * @memcg: target memory control group
767 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
769 * For an object-based mapped page, find all the places it is mapped and
770 * check/clear the referenced flag. This is done by following the page->mapping
771 * pointer, then walking the chain of vmas it holds. It returns the number
772 * of references it found.
774 * This function is only called from page_referenced for object-based pages.
776 static int page_referenced_file(struct page *page,
777 struct mem_cgroup *memcg,
778 unsigned long *vm_flags)
780 unsigned int mapcount;
781 struct address_space *mapping = page->mapping;
782 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
783 struct vm_area_struct *vma;
787 * The caller's checks on page->mapping and !PageAnon have made
788 * sure that this is a file page: the check for page->mapping
789 * excludes the case just before it gets set on an anon page.
791 BUG_ON(PageAnon(page));
794 * The page lock not only makes sure that page->mapping cannot
795 * suddenly be NULLified by truncation, it makes sure that the
796 * structure at mapping cannot be freed and reused yet,
797 * so we can safely take mapping->i_mmap_mutex.
799 BUG_ON(!PageLocked(page));
801 mutex_lock(&mapping->i_mmap_mutex);
804 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
805 * is more likely to be accurate if we note it after spinning.
807 mapcount = page_mapcount(page);
809 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
810 unsigned long address = vma_address(page, vma);
812 * If we are reclaiming on behalf of a cgroup, skip
813 * counting on behalf of references from different
816 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
818 referenced += page_referenced_one(page, vma, address,
819 &mapcount, vm_flags);
824 mutex_unlock(&mapping->i_mmap_mutex);
829 * page_referenced - test if the page was referenced
830 * @page: the page to test
831 * @is_locked: caller holds lock on the page
832 * @memcg: target memory cgroup
833 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
835 * Quick test_and_clear_referenced for all mappings to a page,
836 * returns the number of ptes which referenced the page.
838 int page_referenced(struct page *page,
840 struct mem_cgroup *memcg,
841 unsigned long *vm_flags)
847 if (page_mapped(page) && page_rmapping(page)) {
848 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
849 we_locked = trylock_page(page);
855 if (unlikely(PageKsm(page)))
856 referenced += page_referenced_ksm(page, memcg,
858 else if (PageAnon(page))
859 referenced += page_referenced_anon(page, memcg,
861 else if (page->mapping)
862 referenced += page_referenced_file(page, memcg,
867 if (page_test_and_clear_young(page_to_pfn(page)))
874 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
875 unsigned long address)
877 struct mm_struct *mm = vma->vm_mm;
882 pte = page_check_address(page, mm, address, &ptl, 1);
886 if (pte_dirty(*pte) || pte_write(*pte)) {
889 flush_cache_page(vma, address, pte_pfn(*pte));
890 entry = ptep_clear_flush_notify(vma, address, pte);
891 entry = pte_wrprotect(entry);
892 entry = pte_mkclean(entry);
893 set_pte_at(mm, address, pte, entry);
897 pte_unmap_unlock(pte, ptl);
902 static int page_mkclean_file(struct address_space *mapping, struct page *page)
904 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
905 struct vm_area_struct *vma;
908 BUG_ON(PageAnon(page));
910 mutex_lock(&mapping->i_mmap_mutex);
911 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
912 if (vma->vm_flags & VM_SHARED) {
913 unsigned long address = vma_address(page, vma);
914 ret += page_mkclean_one(page, vma, address);
917 mutex_unlock(&mapping->i_mmap_mutex);
921 int page_mkclean(struct page *page)
925 BUG_ON(!PageLocked(page));
927 if (page_mapped(page)) {
928 struct address_space *mapping = page_mapping(page);
930 ret = page_mkclean_file(mapping, page);
931 if (page_test_and_clear_dirty(page_to_pfn(page), 1))
938 EXPORT_SYMBOL_GPL(page_mkclean);
941 * page_move_anon_rmap - move a page to our anon_vma
942 * @page: the page to move to our anon_vma
943 * @vma: the vma the page belongs to
944 * @address: the user virtual address mapped
946 * When a page belongs exclusively to one process after a COW event,
947 * that page can be moved into the anon_vma that belongs to just that
948 * process, so the rmap code will not search the parent or sibling
951 void page_move_anon_rmap(struct page *page,
952 struct vm_area_struct *vma, unsigned long address)
954 struct anon_vma *anon_vma = vma->anon_vma;
956 VM_BUG_ON(!PageLocked(page));
957 VM_BUG_ON(!anon_vma);
958 VM_BUG_ON(page->index != linear_page_index(vma, address));
960 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
961 page->mapping = (struct address_space *) anon_vma;
965 * __page_set_anon_rmap - set up new anonymous rmap
966 * @page: Page to add to rmap
967 * @vma: VM area to add page to.
968 * @address: User virtual address of the mapping
969 * @exclusive: the page is exclusively owned by the current process
971 static void __page_set_anon_rmap(struct page *page,
972 struct vm_area_struct *vma, unsigned long address, int exclusive)
974 struct anon_vma *anon_vma = vma->anon_vma;
982 * If the page isn't exclusively mapped into this vma,
983 * we must use the _oldest_ possible anon_vma for the
987 anon_vma = anon_vma->root;
989 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
990 page->mapping = (struct address_space *) anon_vma;
991 page->index = linear_page_index(vma, address);
995 * __page_check_anon_rmap - sanity check anonymous rmap addition
996 * @page: the page to add the mapping to
997 * @vma: the vm area in which the mapping is added
998 * @address: the user virtual address mapped
1000 static void __page_check_anon_rmap(struct page *page,
1001 struct vm_area_struct *vma, unsigned long address)
1003 #ifdef CONFIG_DEBUG_VM
1005 * The page's anon-rmap details (mapping and index) are guaranteed to
1006 * be set up correctly at this point.
1008 * We have exclusion against page_add_anon_rmap because the caller
1009 * always holds the page locked, except if called from page_dup_rmap,
1010 * in which case the page is already known to be setup.
1012 * We have exclusion against page_add_new_anon_rmap because those pages
1013 * are initially only visible via the pagetables, and the pte is locked
1014 * over the call to page_add_new_anon_rmap.
1016 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1017 BUG_ON(page->index != linear_page_index(vma, address));
1022 * page_add_anon_rmap - add pte mapping to an anonymous page
1023 * @page: the page to add the mapping to
1024 * @vma: the vm area in which the mapping is added
1025 * @address: the user virtual address mapped
1027 * The caller needs to hold the pte lock, and the page must be locked in
1028 * the anon_vma case: to serialize mapping,index checking after setting,
1029 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1030 * (but PageKsm is never downgraded to PageAnon).
1032 void page_add_anon_rmap(struct page *page,
1033 struct vm_area_struct *vma, unsigned long address)
1035 do_page_add_anon_rmap(page, vma, address, 0);
1039 * Special version of the above for do_swap_page, which often runs
1040 * into pages that are exclusively owned by the current process.
1041 * Everybody else should continue to use page_add_anon_rmap above.
1043 void do_page_add_anon_rmap(struct page *page,
1044 struct vm_area_struct *vma, unsigned long address, int exclusive)
1046 int first = atomic_inc_and_test(&page->_mapcount);
1048 if (!PageTransHuge(page))
1049 __inc_zone_page_state(page, NR_ANON_PAGES);
1051 __inc_zone_page_state(page,
1052 NR_ANON_TRANSPARENT_HUGEPAGES);
1054 if (unlikely(PageKsm(page)))
1057 VM_BUG_ON(!PageLocked(page));
1058 /* address might be in next vma when migration races vma_adjust */
1060 __page_set_anon_rmap(page, vma, address, exclusive);
1062 __page_check_anon_rmap(page, vma, address);
1066 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1067 * @page: the page to add the mapping to
1068 * @vma: the vm area in which the mapping is added
1069 * @address: the user virtual address mapped
1071 * Same as page_add_anon_rmap but must only be called on *new* pages.
1072 * This means the inc-and-test can be bypassed.
1073 * Page does not have to be locked.
1075 void page_add_new_anon_rmap(struct page *page,
1076 struct vm_area_struct *vma, unsigned long address)
1078 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1079 SetPageSwapBacked(page);
1080 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1081 if (!PageTransHuge(page))
1082 __inc_zone_page_state(page, NR_ANON_PAGES);
1084 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1085 __page_set_anon_rmap(page, vma, address, 1);
1086 if (!mlocked_vma_newpage(vma, page))
1087 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1089 add_page_to_unevictable_list(page);
1093 * page_add_file_rmap - add pte mapping to a file page
1094 * @page: the page to add the mapping to
1096 * The caller needs to hold the pte lock.
1098 void page_add_file_rmap(struct page *page)
1101 unsigned long flags;
1103 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1104 if (atomic_inc_and_test(&page->_mapcount)) {
1105 __inc_zone_page_state(page, NR_FILE_MAPPED);
1106 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1108 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1112 * page_remove_rmap - take down pte mapping from a page
1113 * @page: page to remove mapping from
1115 * The caller needs to hold the pte lock.
1117 void page_remove_rmap(struct page *page)
1119 bool anon = PageAnon(page);
1121 unsigned long flags;
1124 * The anon case has no mem_cgroup page_stat to update; but may
1125 * uncharge_page() below, where the lock ordering can deadlock if
1126 * we hold the lock against page_stat move: so avoid it on anon.
1129 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1131 /* page still mapped by someone else? */
1132 if (!atomic_add_negative(-1, &page->_mapcount))
1136 * Now that the last pte has gone, s390 must transfer dirty
1137 * flag from storage key to struct page. We can usually skip
1138 * this if the page is anon, so about to be freed; but perhaps
1139 * not if it's in swapcache - there might be another pte slot
1140 * containing the swap entry, but page not yet written to swap.
1142 if ((!anon || PageSwapCache(page)) &&
1143 page_test_and_clear_dirty(page_to_pfn(page), 1))
1144 set_page_dirty(page);
1146 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1147 * and not charged by memcg for now.
1149 if (unlikely(PageHuge(page)))
1152 mem_cgroup_uncharge_page(page);
1153 if (!PageTransHuge(page))
1154 __dec_zone_page_state(page, NR_ANON_PAGES);
1156 __dec_zone_page_state(page,
1157 NR_ANON_TRANSPARENT_HUGEPAGES);
1159 __dec_zone_page_state(page, NR_FILE_MAPPED);
1160 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1161 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1163 if (unlikely(PageMlocked(page)))
1164 clear_page_mlock(page);
1166 * It would be tidy to reset the PageAnon mapping here,
1167 * but that might overwrite a racing page_add_anon_rmap
1168 * which increments mapcount after us but sets mapping
1169 * before us: so leave the reset to free_hot_cold_page,
1170 * and remember that it's only reliable while mapped.
1171 * Leaving it set also helps swapoff to reinstate ptes
1172 * faster for those pages still in swapcache.
1177 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1181 * Subfunctions of try_to_unmap: try_to_unmap_one called
1182 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1184 int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1185 unsigned long address, enum ttu_flags flags)
1187 struct mm_struct *mm = vma->vm_mm;
1191 int ret = SWAP_AGAIN;
1193 pte = page_check_address(page, mm, address, &ptl, 0);
1198 * If the page is mlock()d, we cannot swap it out.
1199 * If it's recently referenced (perhaps page_referenced
1200 * skipped over this mm) then we should reactivate it.
1202 if (!(flags & TTU_IGNORE_MLOCK)) {
1203 if (vma->vm_flags & VM_LOCKED)
1206 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1209 if (!(flags & TTU_IGNORE_ACCESS)) {
1210 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1216 /* Nuke the page table entry. */
1217 flush_cache_page(vma, address, page_to_pfn(page));
1218 pteval = ptep_clear_flush_notify(vma, address, pte);
1220 /* Move the dirty bit to the physical page now the pte is gone. */
1221 if (pte_dirty(pteval))
1222 set_page_dirty(page);
1224 /* Update high watermark before we lower rss */
1225 update_hiwater_rss(mm);
1227 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1229 dec_mm_counter(mm, MM_ANONPAGES);
1231 dec_mm_counter(mm, MM_FILEPAGES);
1232 set_pte_at(mm, address, pte,
1233 swp_entry_to_pte(make_hwpoison_entry(page)));
1234 } else if (PageAnon(page)) {
1235 swp_entry_t entry = { .val = page_private(page) };
1237 if (PageSwapCache(page)) {
1239 * Store the swap location in the pte.
1240 * See handle_pte_fault() ...
1242 if (swap_duplicate(entry) < 0) {
1243 set_pte_at(mm, address, pte, pteval);
1247 if (list_empty(&mm->mmlist)) {
1248 spin_lock(&mmlist_lock);
1249 if (list_empty(&mm->mmlist))
1250 list_add(&mm->mmlist, &init_mm.mmlist);
1251 spin_unlock(&mmlist_lock);
1253 dec_mm_counter(mm, MM_ANONPAGES);
1254 inc_mm_counter(mm, MM_SWAPENTS);
1255 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1257 * Store the pfn of the page in a special migration
1258 * pte. do_swap_page() will wait until the migration
1259 * pte is removed and then restart fault handling.
1261 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1262 entry = make_migration_entry(page, pte_write(pteval));
1264 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1265 BUG_ON(pte_file(*pte));
1266 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1267 (TTU_ACTION(flags) == TTU_MIGRATION)) {
1268 /* Establish migration entry for a file page */
1270 entry = make_migration_entry(page, pte_write(pteval));
1271 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1273 dec_mm_counter(mm, MM_FILEPAGES);
1275 page_remove_rmap(page);
1276 page_cache_release(page);
1279 pte_unmap_unlock(pte, ptl);
1284 pte_unmap_unlock(pte, ptl);
1288 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1289 * unstable result and race. Plus, We can't wait here because
1290 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1291 * if trylock failed, the page remain in evictable lru and later
1292 * vmscan could retry to move the page to unevictable lru if the
1293 * page is actually mlocked.
1295 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1296 if (vma->vm_flags & VM_LOCKED) {
1297 mlock_vma_page(page);
1300 up_read(&vma->vm_mm->mmap_sem);
1306 * objrmap doesn't work for nonlinear VMAs because the assumption that
1307 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1308 * Consequently, given a particular page and its ->index, we cannot locate the
1309 * ptes which are mapping that page without an exhaustive linear search.
1311 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1312 * maps the file to which the target page belongs. The ->vm_private_data field
1313 * holds the current cursor into that scan. Successive searches will circulate
1314 * around the vma's virtual address space.
1316 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1317 * more scanning pressure is placed against them as well. Eventually pages
1318 * will become fully unmapped and are eligible for eviction.
1320 * For very sparsely populated VMAs this is a little inefficient - chances are
1321 * there there won't be many ptes located within the scan cluster. In this case
1322 * maybe we could scan further - to the end of the pte page, perhaps.
1324 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1325 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1326 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1327 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1329 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1330 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1332 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1333 struct vm_area_struct *vma, struct page *check_page)
1335 struct mm_struct *mm = vma->vm_mm;
1343 unsigned long address;
1345 int ret = SWAP_AGAIN;
1348 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1349 end = address + CLUSTER_SIZE;
1350 if (address < vma->vm_start)
1351 address = vma->vm_start;
1352 if (end > vma->vm_end)
1355 pgd = pgd_offset(mm, address);
1356 if (!pgd_present(*pgd))
1359 pud = pud_offset(pgd, address);
1360 if (!pud_present(*pud))
1363 pmd = pmd_offset(pud, address);
1364 if (!pmd_present(*pmd))
1368 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1369 * keep the sem while scanning the cluster for mlocking pages.
1371 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1372 locked_vma = (vma->vm_flags & VM_LOCKED);
1374 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1377 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1379 /* Update high watermark before we lower rss */
1380 update_hiwater_rss(mm);
1382 for (; address < end; pte++, address += PAGE_SIZE) {
1383 if (!pte_present(*pte))
1385 page = vm_normal_page(vma, address, *pte);
1386 BUG_ON(!page || PageAnon(page));
1389 mlock_vma_page(page); /* no-op if already mlocked */
1390 if (page == check_page)
1392 continue; /* don't unmap */
1395 if (ptep_clear_flush_young_notify(vma, address, pte))
1398 /* Nuke the page table entry. */
1399 flush_cache_page(vma, address, pte_pfn(*pte));
1400 pteval = ptep_clear_flush_notify(vma, address, pte);
1402 /* If nonlinear, store the file page offset in the pte. */
1403 if (page->index != linear_page_index(vma, address))
1404 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1406 /* Move the dirty bit to the physical page now the pte is gone. */
1407 if (pte_dirty(pteval))
1408 set_page_dirty(page);
1410 page_remove_rmap(page);
1411 page_cache_release(page);
1412 dec_mm_counter(mm, MM_FILEPAGES);
1415 pte_unmap_unlock(pte - 1, ptl);
1417 up_read(&vma->vm_mm->mmap_sem);
1421 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1423 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1428 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1429 VM_STACK_INCOMPLETE_SETUP)
1436 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1438 * @page: the page to unmap/unlock
1439 * @flags: action and flags
1441 * Find all the mappings of a page using the mapping pointer and the vma chains
1442 * contained in the anon_vma struct it points to.
1444 * This function is only called from try_to_unmap/try_to_munlock for
1446 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1447 * where the page was found will be held for write. So, we won't recheck
1448 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1451 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1453 struct anon_vma *anon_vma;
1455 struct anon_vma_chain *avc;
1456 int ret = SWAP_AGAIN;
1458 anon_vma = page_lock_anon_vma(page);
1462 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1463 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1464 struct vm_area_struct *vma = avc->vma;
1465 unsigned long address;
1468 * During exec, a temporary VMA is setup and later moved.
1469 * The VMA is moved under the anon_vma lock but not the
1470 * page tables leading to a race where migration cannot
1471 * find the migration ptes. Rather than increasing the
1472 * locking requirements of exec(), migration skips
1473 * temporary VMAs until after exec() completes.
1475 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1476 is_vma_temporary_stack(vma))
1479 address = vma_address(page, vma);
1480 ret = try_to_unmap_one(page, vma, address, flags);
1481 if (ret != SWAP_AGAIN || !page_mapped(page))
1485 page_unlock_anon_vma(anon_vma);
1490 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1491 * @page: the page to unmap/unlock
1492 * @flags: action and flags
1494 * Find all the mappings of a page using the mapping pointer and the vma chains
1495 * contained in the address_space struct it points to.
1497 * This function is only called from try_to_unmap/try_to_munlock for
1498 * object-based pages.
1499 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1500 * where the page was found will be held for write. So, we won't recheck
1501 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1504 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1506 struct address_space *mapping = page->mapping;
1507 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1508 struct vm_area_struct *vma;
1509 int ret = SWAP_AGAIN;
1510 unsigned long cursor;
1511 unsigned long max_nl_cursor = 0;
1512 unsigned long max_nl_size = 0;
1513 unsigned int mapcount;
1515 mutex_lock(&mapping->i_mmap_mutex);
1516 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1517 unsigned long address = vma_address(page, vma);
1518 ret = try_to_unmap_one(page, vma, address, flags);
1519 if (ret != SWAP_AGAIN || !page_mapped(page))
1523 if (list_empty(&mapping->i_mmap_nonlinear))
1527 * We don't bother to try to find the munlocked page in nonlinears.
1528 * It's costly. Instead, later, page reclaim logic may call
1529 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1531 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1534 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1536 cursor = (unsigned long) vma->vm_private_data;
1537 if (cursor > max_nl_cursor)
1538 max_nl_cursor = cursor;
1539 cursor = vma->vm_end - vma->vm_start;
1540 if (cursor > max_nl_size)
1541 max_nl_size = cursor;
1544 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1550 * We don't try to search for this page in the nonlinear vmas,
1551 * and page_referenced wouldn't have found it anyway. Instead
1552 * just walk the nonlinear vmas trying to age and unmap some.
1553 * The mapcount of the page we came in with is irrelevant,
1554 * but even so use it as a guide to how hard we should try?
1556 mapcount = page_mapcount(page);
1561 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1562 if (max_nl_cursor == 0)
1563 max_nl_cursor = CLUSTER_SIZE;
1566 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1568 cursor = (unsigned long) vma->vm_private_data;
1569 while ( cursor < max_nl_cursor &&
1570 cursor < vma->vm_end - vma->vm_start) {
1571 if (try_to_unmap_cluster(cursor, &mapcount,
1572 vma, page) == SWAP_MLOCK)
1574 cursor += CLUSTER_SIZE;
1575 vma->vm_private_data = (void *) cursor;
1576 if ((int)mapcount <= 0)
1579 vma->vm_private_data = (void *) max_nl_cursor;
1582 max_nl_cursor += CLUSTER_SIZE;
1583 } while (max_nl_cursor <= max_nl_size);
1586 * Don't loop forever (perhaps all the remaining pages are
1587 * in locked vmas). Reset cursor on all unreserved nonlinear
1588 * vmas, now forgetting on which ones it had fallen behind.
1590 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear)
1591 vma->vm_private_data = NULL;
1593 mutex_unlock(&mapping->i_mmap_mutex);
1598 * try_to_unmap - try to remove all page table mappings to a page
1599 * @page: the page to get unmapped
1600 * @flags: action and flags
1602 * Tries to remove all the page table entries which are mapping this
1603 * page, used in the pageout path. Caller must hold the page lock.
1604 * Return values are:
1606 * SWAP_SUCCESS - we succeeded in removing all mappings
1607 * SWAP_AGAIN - we missed a mapping, try again later
1608 * SWAP_FAIL - the page is unswappable
1609 * SWAP_MLOCK - page is mlocked.
1611 int try_to_unmap(struct page *page, enum ttu_flags flags)
1615 BUG_ON(!PageLocked(page));
1616 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1618 if (unlikely(PageKsm(page)))
1619 ret = try_to_unmap_ksm(page, flags);
1620 else if (PageAnon(page))
1621 ret = try_to_unmap_anon(page, flags);
1623 ret = try_to_unmap_file(page, flags);
1624 if (ret != SWAP_MLOCK && !page_mapped(page))
1630 * try_to_munlock - try to munlock a page
1631 * @page: the page to be munlocked
1633 * Called from munlock code. Checks all of the VMAs mapping the page
1634 * to make sure nobody else has this page mlocked. The page will be
1635 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1637 * Return values are:
1639 * SWAP_AGAIN - no vma is holding page mlocked, or,
1640 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1641 * SWAP_FAIL - page cannot be located at present
1642 * SWAP_MLOCK - page is now mlocked.
1644 int try_to_munlock(struct page *page)
1646 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1648 if (unlikely(PageKsm(page)))
1649 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1650 else if (PageAnon(page))
1651 return try_to_unmap_anon(page, TTU_MUNLOCK);
1653 return try_to_unmap_file(page, TTU_MUNLOCK);
1656 void __put_anon_vma(struct anon_vma *anon_vma)
1658 struct anon_vma *root = anon_vma->root;
1660 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1661 anon_vma_free(root);
1663 anon_vma_free(anon_vma);
1666 #ifdef CONFIG_MIGRATION
1668 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1669 * Called by migrate.c to remove migration ptes, but might be used more later.
1671 static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1672 struct vm_area_struct *, unsigned long, void *), void *arg)
1674 struct anon_vma *anon_vma;
1675 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1676 struct anon_vma_chain *avc;
1677 int ret = SWAP_AGAIN;
1680 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1681 * because that depends on page_mapped(); but not all its usages
1682 * are holding mmap_sem. Users without mmap_sem are required to
1683 * take a reference count to prevent the anon_vma disappearing
1685 anon_vma = page_anon_vma(page);
1688 anon_vma_lock(anon_vma);
1689 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1690 struct vm_area_struct *vma = avc->vma;
1691 unsigned long address = vma_address(page, vma);
1692 ret = rmap_one(page, vma, address, arg);
1693 if (ret != SWAP_AGAIN)
1696 anon_vma_unlock(anon_vma);
1700 static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1701 struct vm_area_struct *, unsigned long, void *), void *arg)
1703 struct address_space *mapping = page->mapping;
1704 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1705 struct vm_area_struct *vma;
1706 int ret = SWAP_AGAIN;
1710 mutex_lock(&mapping->i_mmap_mutex);
1711 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1712 unsigned long address = vma_address(page, vma);
1713 ret = rmap_one(page, vma, address, arg);
1714 if (ret != SWAP_AGAIN)
1718 * No nonlinear handling: being always shared, nonlinear vmas
1719 * never contain migration ptes. Decide what to do about this
1720 * limitation to linear when we need rmap_walk() on nonlinear.
1722 mutex_unlock(&mapping->i_mmap_mutex);
1726 int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1727 struct vm_area_struct *, unsigned long, void *), void *arg)
1729 VM_BUG_ON(!PageLocked(page));
1731 if (unlikely(PageKsm(page)))
1732 return rmap_walk_ksm(page, rmap_one, arg);
1733 else if (PageAnon(page))
1734 return rmap_walk_anon(page, rmap_one, arg);
1736 return rmap_walk_file(page, rmap_one, arg);
1738 #endif /* CONFIG_MIGRATION */
1740 #ifdef CONFIG_HUGETLB_PAGE
1742 * The following three functions are for anonymous (private mapped) hugepages.
1743 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1744 * and no lru code, because we handle hugepages differently from common pages.
1746 static void __hugepage_set_anon_rmap(struct page *page,
1747 struct vm_area_struct *vma, unsigned long address, int exclusive)
1749 struct anon_vma *anon_vma = vma->anon_vma;
1756 anon_vma = anon_vma->root;
1758 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1759 page->mapping = (struct address_space *) anon_vma;
1760 page->index = linear_page_index(vma, address);
1763 void hugepage_add_anon_rmap(struct page *page,
1764 struct vm_area_struct *vma, unsigned long address)
1766 struct anon_vma *anon_vma = vma->anon_vma;
1769 BUG_ON(!PageLocked(page));
1771 /* address might be in next vma when migration races vma_adjust */
1772 first = atomic_inc_and_test(&page->_mapcount);
1774 __hugepage_set_anon_rmap(page, vma, address, 0);
1777 void hugepage_add_new_anon_rmap(struct page *page,
1778 struct vm_area_struct *vma, unsigned long address)
1780 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1781 atomic_set(&page->_mapcount, 0);
1782 __hugepage_set_anon_rmap(page, vma, address, 1);
1784 #endif /* CONFIG_HUGETLB_PAGE */