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);
369 anon_vma->rb_root = RB_ROOT;
372 void __init anon_vma_init(void)
374 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
375 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
376 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
380 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
382 * Since there is no serialization what so ever against page_remove_rmap()
383 * the best this function can do is return a locked anon_vma that might
384 * have been relevant to this page.
386 * The page might have been remapped to a different anon_vma or the anon_vma
387 * returned may already be freed (and even reused).
389 * In case it was remapped to a different anon_vma, the new anon_vma will be a
390 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
391 * ensure that any anon_vma obtained from the page will still be valid for as
392 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
394 * All users of this function must be very careful when walking the anon_vma
395 * chain and verify that the page in question is indeed mapped in it
396 * [ something equivalent to page_mapped_in_vma() ].
398 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
399 * that the anon_vma pointer from page->mapping is valid if there is a
400 * mapcount, we can dereference the anon_vma after observing those.
402 struct anon_vma *page_get_anon_vma(struct page *page)
404 struct anon_vma *anon_vma = NULL;
405 unsigned long anon_mapping;
408 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
409 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
411 if (!page_mapped(page))
414 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
415 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
421 * If this page is still mapped, then its anon_vma cannot have been
422 * freed. But if it has been unmapped, we have no security against the
423 * anon_vma structure being freed and reused (for another anon_vma:
424 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
425 * above cannot corrupt).
427 if (!page_mapped(page)) {
428 put_anon_vma(anon_vma);
438 * Similar to page_get_anon_vma() except it locks the anon_vma.
440 * Its a little more complex as it tries to keep the fast path to a single
441 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
442 * reference like with page_get_anon_vma() and then block on the mutex.
444 struct anon_vma *page_lock_anon_vma(struct page *page)
446 struct anon_vma *anon_vma = NULL;
447 struct anon_vma *root_anon_vma;
448 unsigned long anon_mapping;
451 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
452 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
454 if (!page_mapped(page))
457 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
458 root_anon_vma = ACCESS_ONCE(anon_vma->root);
459 if (mutex_trylock(&root_anon_vma->mutex)) {
461 * If the page is still mapped, then this anon_vma is still
462 * its anon_vma, and holding the mutex ensures that it will
463 * not go away, see anon_vma_free().
465 if (!page_mapped(page)) {
466 mutex_unlock(&root_anon_vma->mutex);
472 /* trylock failed, we got to sleep */
473 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
478 if (!page_mapped(page)) {
479 put_anon_vma(anon_vma);
484 /* we pinned the anon_vma, its safe to sleep */
486 anon_vma_lock(anon_vma);
488 if (atomic_dec_and_test(&anon_vma->refcount)) {
490 * Oops, we held the last refcount, release the lock
491 * and bail -- can't simply use put_anon_vma() because
492 * we'll deadlock on the anon_vma_lock() recursion.
494 anon_vma_unlock(anon_vma);
495 __put_anon_vma(anon_vma);
506 void page_unlock_anon_vma(struct anon_vma *anon_vma)
508 anon_vma_unlock(anon_vma);
512 * At what user virtual address is page expected in @vma?
513 * Returns virtual address or -EFAULT if page's index/offset is not
514 * within the range mapped the @vma.
517 vma_address(struct page *page, struct vm_area_struct *vma)
519 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
520 unsigned long address;
522 if (unlikely(is_vm_hugetlb_page(vma)))
523 pgoff = page->index << huge_page_order(page_hstate(page));
524 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
525 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
526 /* page should be within @vma mapping range */
533 * At what user virtual address is page expected in vma?
534 * Caller should check the page is actually part of the vma.
536 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
538 if (PageAnon(page)) {
539 struct anon_vma *page__anon_vma = page_anon_vma(page);
541 * Note: swapoff's unuse_vma() is more efficient with this
542 * check, and needs it to match anon_vma when KSM is active.
544 if (!vma->anon_vma || !page__anon_vma ||
545 vma->anon_vma->root != page__anon_vma->root)
547 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
549 vma->vm_file->f_mapping != page->mapping)
553 return vma_address(page, vma);
557 * Check that @page is mapped at @address into @mm.
559 * If @sync is false, page_check_address may perform a racy check to avoid
560 * the page table lock when the pte is not present (helpful when reclaiming
561 * highly shared pages).
563 * On success returns with pte mapped and locked.
565 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
566 unsigned long address, spinlock_t **ptlp, int sync)
574 if (unlikely(PageHuge(page))) {
575 pte = huge_pte_offset(mm, address);
576 ptl = &mm->page_table_lock;
580 pgd = pgd_offset(mm, address);
581 if (!pgd_present(*pgd))
584 pud = pud_offset(pgd, address);
585 if (!pud_present(*pud))
588 pmd = pmd_offset(pud, address);
589 if (!pmd_present(*pmd))
591 if (pmd_trans_huge(*pmd))
594 pte = pte_offset_map(pmd, address);
595 /* Make a quick check before getting the lock */
596 if (!sync && !pte_present(*pte)) {
601 ptl = pte_lockptr(mm, pmd);
604 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
608 pte_unmap_unlock(pte, ptl);
613 * page_mapped_in_vma - check whether a page is really mapped in a VMA
614 * @page: the page to test
615 * @vma: the VMA to test
617 * Returns 1 if the page is mapped into the page tables of the VMA, 0
618 * if the page is not mapped into the page tables of this VMA. Only
619 * valid for normal file or anonymous VMAs.
621 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
623 unsigned long address;
627 address = vma_address(page, vma);
628 if (address == -EFAULT) /* out of vma range */
630 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
631 if (!pte) /* the page is not in this mm */
633 pte_unmap_unlock(pte, ptl);
639 * Subfunctions of page_referenced: page_referenced_one called
640 * repeatedly from either page_referenced_anon or page_referenced_file.
642 int page_referenced_one(struct page *page, struct vm_area_struct *vma,
643 unsigned long address, unsigned int *mapcount,
644 unsigned long *vm_flags)
646 struct mm_struct *mm = vma->vm_mm;
649 if (unlikely(PageTransHuge(page))) {
652 spin_lock(&mm->page_table_lock);
654 * rmap might return false positives; we must filter
655 * these out using page_check_address_pmd().
657 pmd = page_check_address_pmd(page, mm, address,
658 PAGE_CHECK_ADDRESS_PMD_FLAG);
660 spin_unlock(&mm->page_table_lock);
664 if (vma->vm_flags & VM_LOCKED) {
665 spin_unlock(&mm->page_table_lock);
666 *mapcount = 0; /* break early from loop */
667 *vm_flags |= VM_LOCKED;
671 /* go ahead even if the pmd is pmd_trans_splitting() */
672 if (pmdp_clear_flush_young_notify(vma, address, pmd))
674 spin_unlock(&mm->page_table_lock);
680 * rmap might return false positives; we must filter
681 * these out using page_check_address().
683 pte = page_check_address(page, mm, address, &ptl, 0);
687 if (vma->vm_flags & VM_LOCKED) {
688 pte_unmap_unlock(pte, ptl);
689 *mapcount = 0; /* break early from loop */
690 *vm_flags |= VM_LOCKED;
694 if (ptep_clear_flush_young_notify(vma, address, pte)) {
696 * Don't treat a reference through a sequentially read
697 * mapping as such. If the page has been used in
698 * another mapping, we will catch it; if this other
699 * mapping is already gone, the unmap path will have
700 * set PG_referenced or activated the page.
702 if (likely(!VM_SequentialReadHint(vma)))
705 pte_unmap_unlock(pte, ptl);
711 *vm_flags |= vma->vm_flags;
716 static int page_referenced_anon(struct page *page,
717 struct mem_cgroup *memcg,
718 unsigned long *vm_flags)
720 unsigned int mapcount;
721 struct anon_vma *anon_vma;
723 struct anon_vma_chain *avc;
726 anon_vma = page_lock_anon_vma(page);
730 mapcount = page_mapcount(page);
731 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
732 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
733 struct vm_area_struct *vma = avc->vma;
734 unsigned long address = vma_address(page, vma);
735 if (address == -EFAULT)
738 * If we are reclaiming on behalf of a cgroup, skip
739 * counting on behalf of references from different
742 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
744 referenced += page_referenced_one(page, vma, address,
745 &mapcount, vm_flags);
750 page_unlock_anon_vma(anon_vma);
755 * page_referenced_file - referenced check for object-based rmap
756 * @page: the page we're checking references on.
757 * @memcg: target memory control group
758 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
760 * For an object-based mapped page, find all the places it is mapped and
761 * check/clear the referenced flag. This is done by following the page->mapping
762 * pointer, then walking the chain of vmas it holds. It returns the number
763 * of references it found.
765 * This function is only called from page_referenced for object-based pages.
767 static int page_referenced_file(struct page *page,
768 struct mem_cgroup *memcg,
769 unsigned long *vm_flags)
771 unsigned int mapcount;
772 struct address_space *mapping = page->mapping;
773 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
774 struct vm_area_struct *vma;
778 * The caller's checks on page->mapping and !PageAnon have made
779 * sure that this is a file page: the check for page->mapping
780 * excludes the case just before it gets set on an anon page.
782 BUG_ON(PageAnon(page));
785 * The page lock not only makes sure that page->mapping cannot
786 * suddenly be NULLified by truncation, it makes sure that the
787 * structure at mapping cannot be freed and reused yet,
788 * so we can safely take mapping->i_mmap_mutex.
790 BUG_ON(!PageLocked(page));
792 mutex_lock(&mapping->i_mmap_mutex);
795 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
796 * is more likely to be accurate if we note it after spinning.
798 mapcount = page_mapcount(page);
800 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
801 unsigned long address = vma_address(page, vma);
802 if (address == -EFAULT)
805 * If we are reclaiming on behalf of a cgroup, skip
806 * counting on behalf of references from different
809 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
811 referenced += page_referenced_one(page, vma, address,
812 &mapcount, vm_flags);
817 mutex_unlock(&mapping->i_mmap_mutex);
822 * page_referenced - test if the page was referenced
823 * @page: the page to test
824 * @is_locked: caller holds lock on the page
825 * @memcg: target memory cgroup
826 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
828 * Quick test_and_clear_referenced for all mappings to a page,
829 * returns the number of ptes which referenced the page.
831 int page_referenced(struct page *page,
833 struct mem_cgroup *memcg,
834 unsigned long *vm_flags)
840 if (page_mapped(page) && page_rmapping(page)) {
841 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
842 we_locked = trylock_page(page);
848 if (unlikely(PageKsm(page)))
849 referenced += page_referenced_ksm(page, memcg,
851 else if (PageAnon(page))
852 referenced += page_referenced_anon(page, memcg,
854 else if (page->mapping)
855 referenced += page_referenced_file(page, memcg,
860 if (page_test_and_clear_young(page_to_pfn(page)))
867 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
868 unsigned long address)
870 struct mm_struct *mm = vma->vm_mm;
875 pte = page_check_address(page, mm, address, &ptl, 1);
879 if (pte_dirty(*pte) || pte_write(*pte)) {
882 flush_cache_page(vma, address, pte_pfn(*pte));
883 entry = ptep_clear_flush_notify(vma, address, pte);
884 entry = pte_wrprotect(entry);
885 entry = pte_mkclean(entry);
886 set_pte_at(mm, address, pte, entry);
890 pte_unmap_unlock(pte, ptl);
895 static int page_mkclean_file(struct address_space *mapping, struct page *page)
897 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
898 struct vm_area_struct *vma;
901 BUG_ON(PageAnon(page));
903 mutex_lock(&mapping->i_mmap_mutex);
904 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
905 if (vma->vm_flags & VM_SHARED) {
906 unsigned long address = vma_address(page, vma);
907 if (address == -EFAULT)
909 ret += page_mkclean_one(page, vma, address);
912 mutex_unlock(&mapping->i_mmap_mutex);
916 int page_mkclean(struct page *page)
920 BUG_ON(!PageLocked(page));
922 if (page_mapped(page)) {
923 struct address_space *mapping = page_mapping(page);
925 ret = page_mkclean_file(mapping, page);
926 if (page_test_and_clear_dirty(page_to_pfn(page), 1))
933 EXPORT_SYMBOL_GPL(page_mkclean);
936 * page_move_anon_rmap - move a page to our anon_vma
937 * @page: the page to move to our anon_vma
938 * @vma: the vma the page belongs to
939 * @address: the user virtual address mapped
941 * When a page belongs exclusively to one process after a COW event,
942 * that page can be moved into the anon_vma that belongs to just that
943 * process, so the rmap code will not search the parent or sibling
946 void page_move_anon_rmap(struct page *page,
947 struct vm_area_struct *vma, unsigned long address)
949 struct anon_vma *anon_vma = vma->anon_vma;
951 VM_BUG_ON(!PageLocked(page));
952 VM_BUG_ON(!anon_vma);
953 VM_BUG_ON(page->index != linear_page_index(vma, address));
955 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
956 page->mapping = (struct address_space *) anon_vma;
960 * __page_set_anon_rmap - set up new anonymous rmap
961 * @page: Page to add to rmap
962 * @vma: VM area to add page to.
963 * @address: User virtual address of the mapping
964 * @exclusive: the page is exclusively owned by the current process
966 static void __page_set_anon_rmap(struct page *page,
967 struct vm_area_struct *vma, unsigned long address, int exclusive)
969 struct anon_vma *anon_vma = vma->anon_vma;
977 * If the page isn't exclusively mapped into this vma,
978 * we must use the _oldest_ possible anon_vma for the
982 anon_vma = anon_vma->root;
984 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
985 page->mapping = (struct address_space *) anon_vma;
986 page->index = linear_page_index(vma, address);
990 * __page_check_anon_rmap - sanity check anonymous rmap addition
991 * @page: the page to add the mapping to
992 * @vma: the vm area in which the mapping is added
993 * @address: the user virtual address mapped
995 static void __page_check_anon_rmap(struct page *page,
996 struct vm_area_struct *vma, unsigned long address)
998 #ifdef CONFIG_DEBUG_VM
1000 * The page's anon-rmap details (mapping and index) are guaranteed to
1001 * be set up correctly at this point.
1003 * We have exclusion against page_add_anon_rmap because the caller
1004 * always holds the page locked, except if called from page_dup_rmap,
1005 * in which case the page is already known to be setup.
1007 * We have exclusion against page_add_new_anon_rmap because those pages
1008 * are initially only visible via the pagetables, and the pte is locked
1009 * over the call to page_add_new_anon_rmap.
1011 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1012 BUG_ON(page->index != linear_page_index(vma, address));
1017 * page_add_anon_rmap - add pte mapping to an anonymous page
1018 * @page: the page to add the mapping to
1019 * @vma: the vm area in which the mapping is added
1020 * @address: the user virtual address mapped
1022 * The caller needs to hold the pte lock, and the page must be locked in
1023 * the anon_vma case: to serialize mapping,index checking after setting,
1024 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1025 * (but PageKsm is never downgraded to PageAnon).
1027 void page_add_anon_rmap(struct page *page,
1028 struct vm_area_struct *vma, unsigned long address)
1030 do_page_add_anon_rmap(page, vma, address, 0);
1034 * Special version of the above for do_swap_page, which often runs
1035 * into pages that are exclusively owned by the current process.
1036 * Everybody else should continue to use page_add_anon_rmap above.
1038 void do_page_add_anon_rmap(struct page *page,
1039 struct vm_area_struct *vma, unsigned long address, int exclusive)
1041 int first = atomic_inc_and_test(&page->_mapcount);
1043 if (!PageTransHuge(page))
1044 __inc_zone_page_state(page, NR_ANON_PAGES);
1046 __inc_zone_page_state(page,
1047 NR_ANON_TRANSPARENT_HUGEPAGES);
1049 if (unlikely(PageKsm(page)))
1052 VM_BUG_ON(!PageLocked(page));
1053 /* address might be in next vma when migration races vma_adjust */
1055 __page_set_anon_rmap(page, vma, address, exclusive);
1057 __page_check_anon_rmap(page, vma, address);
1061 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1062 * @page: the page to add the mapping to
1063 * @vma: the vm area in which the mapping is added
1064 * @address: the user virtual address mapped
1066 * Same as page_add_anon_rmap but must only be called on *new* pages.
1067 * This means the inc-and-test can be bypassed.
1068 * Page does not have to be locked.
1070 void page_add_new_anon_rmap(struct page *page,
1071 struct vm_area_struct *vma, unsigned long address)
1073 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1074 SetPageSwapBacked(page);
1075 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1076 if (!PageTransHuge(page))
1077 __inc_zone_page_state(page, NR_ANON_PAGES);
1079 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1080 __page_set_anon_rmap(page, vma, address, 1);
1081 if (page_evictable(page, vma))
1082 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1084 add_page_to_unevictable_list(page);
1088 * page_add_file_rmap - add pte mapping to a file page
1089 * @page: the page to add the mapping to
1091 * The caller needs to hold the pte lock.
1093 void page_add_file_rmap(struct page *page)
1096 unsigned long flags;
1098 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1099 if (atomic_inc_and_test(&page->_mapcount)) {
1100 __inc_zone_page_state(page, NR_FILE_MAPPED);
1101 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1103 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1107 * page_remove_rmap - take down pte mapping from a page
1108 * @page: page to remove mapping from
1110 * The caller needs to hold the pte lock.
1112 void page_remove_rmap(struct page *page)
1114 bool anon = PageAnon(page);
1116 unsigned long flags;
1119 * The anon case has no mem_cgroup page_stat to update; but may
1120 * uncharge_page() below, where the lock ordering can deadlock if
1121 * we hold the lock against page_stat move: so avoid it on anon.
1124 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1126 /* page still mapped by someone else? */
1127 if (!atomic_add_negative(-1, &page->_mapcount))
1131 * Now that the last pte has gone, s390 must transfer dirty
1132 * flag from storage key to struct page. We can usually skip
1133 * this if the page is anon, so about to be freed; but perhaps
1134 * not if it's in swapcache - there might be another pte slot
1135 * containing the swap entry, but page not yet written to swap.
1137 if ((!anon || PageSwapCache(page)) &&
1138 page_test_and_clear_dirty(page_to_pfn(page), 1))
1139 set_page_dirty(page);
1141 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1142 * and not charged by memcg for now.
1144 if (unlikely(PageHuge(page)))
1147 mem_cgroup_uncharge_page(page);
1148 if (!PageTransHuge(page))
1149 __dec_zone_page_state(page, NR_ANON_PAGES);
1151 __dec_zone_page_state(page,
1152 NR_ANON_TRANSPARENT_HUGEPAGES);
1154 __dec_zone_page_state(page, NR_FILE_MAPPED);
1155 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1158 * It would be tidy to reset the PageAnon mapping here,
1159 * but that might overwrite a racing page_add_anon_rmap
1160 * which increments mapcount after us but sets mapping
1161 * before us: so leave the reset to free_hot_cold_page,
1162 * and remember that it's only reliable while mapped.
1163 * Leaving it set also helps swapoff to reinstate ptes
1164 * faster for those pages still in swapcache.
1168 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1172 * Subfunctions of try_to_unmap: try_to_unmap_one called
1173 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1175 int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1176 unsigned long address, enum ttu_flags flags)
1178 struct mm_struct *mm = vma->vm_mm;
1182 int ret = SWAP_AGAIN;
1184 pte = page_check_address(page, mm, address, &ptl, 0);
1189 * If the page is mlock()d, we cannot swap it out.
1190 * If it's recently referenced (perhaps page_referenced
1191 * skipped over this mm) then we should reactivate it.
1193 if (!(flags & TTU_IGNORE_MLOCK)) {
1194 if (vma->vm_flags & VM_LOCKED)
1197 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1200 if (!(flags & TTU_IGNORE_ACCESS)) {
1201 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1207 /* Nuke the page table entry. */
1208 flush_cache_page(vma, address, page_to_pfn(page));
1209 pteval = ptep_clear_flush_notify(vma, address, pte);
1211 /* Move the dirty bit to the physical page now the pte is gone. */
1212 if (pte_dirty(pteval))
1213 set_page_dirty(page);
1215 /* Update high watermark before we lower rss */
1216 update_hiwater_rss(mm);
1218 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1220 dec_mm_counter(mm, MM_ANONPAGES);
1222 dec_mm_counter(mm, MM_FILEPAGES);
1223 set_pte_at(mm, address, pte,
1224 swp_entry_to_pte(make_hwpoison_entry(page)));
1225 } else if (PageAnon(page)) {
1226 swp_entry_t entry = { .val = page_private(page) };
1228 if (PageSwapCache(page)) {
1230 * Store the swap location in the pte.
1231 * See handle_pte_fault() ...
1233 if (swap_duplicate(entry) < 0) {
1234 set_pte_at(mm, address, pte, pteval);
1238 if (list_empty(&mm->mmlist)) {
1239 spin_lock(&mmlist_lock);
1240 if (list_empty(&mm->mmlist))
1241 list_add(&mm->mmlist, &init_mm.mmlist);
1242 spin_unlock(&mmlist_lock);
1244 dec_mm_counter(mm, MM_ANONPAGES);
1245 inc_mm_counter(mm, MM_SWAPENTS);
1246 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1248 * Store the pfn of the page in a special migration
1249 * pte. do_swap_page() will wait until the migration
1250 * pte is removed and then restart fault handling.
1252 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1253 entry = make_migration_entry(page, pte_write(pteval));
1255 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1256 BUG_ON(pte_file(*pte));
1257 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1258 (TTU_ACTION(flags) == TTU_MIGRATION)) {
1259 /* Establish migration entry for a file page */
1261 entry = make_migration_entry(page, pte_write(pteval));
1262 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1264 dec_mm_counter(mm, MM_FILEPAGES);
1266 page_remove_rmap(page);
1267 page_cache_release(page);
1270 pte_unmap_unlock(pte, ptl);
1275 pte_unmap_unlock(pte, ptl);
1279 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1280 * unstable result and race. Plus, We can't wait here because
1281 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1282 * if trylock failed, the page remain in evictable lru and later
1283 * vmscan could retry to move the page to unevictable lru if the
1284 * page is actually mlocked.
1286 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1287 if (vma->vm_flags & VM_LOCKED) {
1288 mlock_vma_page(page);
1291 up_read(&vma->vm_mm->mmap_sem);
1297 * objrmap doesn't work for nonlinear VMAs because the assumption that
1298 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1299 * Consequently, given a particular page and its ->index, we cannot locate the
1300 * ptes which are mapping that page without an exhaustive linear search.
1302 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1303 * maps the file to which the target page belongs. The ->vm_private_data field
1304 * holds the current cursor into that scan. Successive searches will circulate
1305 * around the vma's virtual address space.
1307 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1308 * more scanning pressure is placed against them as well. Eventually pages
1309 * will become fully unmapped and are eligible for eviction.
1311 * For very sparsely populated VMAs this is a little inefficient - chances are
1312 * there there won't be many ptes located within the scan cluster. In this case
1313 * maybe we could scan further - to the end of the pte page, perhaps.
1315 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1316 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1317 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1318 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1320 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1321 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1323 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1324 struct vm_area_struct *vma, struct page *check_page)
1326 struct mm_struct *mm = vma->vm_mm;
1334 unsigned long address;
1336 int ret = SWAP_AGAIN;
1339 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1340 end = address + CLUSTER_SIZE;
1341 if (address < vma->vm_start)
1342 address = vma->vm_start;
1343 if (end > vma->vm_end)
1346 pgd = pgd_offset(mm, address);
1347 if (!pgd_present(*pgd))
1350 pud = pud_offset(pgd, address);
1351 if (!pud_present(*pud))
1354 pmd = pmd_offset(pud, address);
1355 if (!pmd_present(*pmd))
1359 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1360 * keep the sem while scanning the cluster for mlocking pages.
1362 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1363 locked_vma = (vma->vm_flags & VM_LOCKED);
1365 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1368 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1370 /* Update high watermark before we lower rss */
1371 update_hiwater_rss(mm);
1373 for (; address < end; pte++, address += PAGE_SIZE) {
1374 if (!pte_present(*pte))
1376 page = vm_normal_page(vma, address, *pte);
1377 BUG_ON(!page || PageAnon(page));
1380 mlock_vma_page(page); /* no-op if already mlocked */
1381 if (page == check_page)
1383 continue; /* don't unmap */
1386 if (ptep_clear_flush_young_notify(vma, address, pte))
1389 /* Nuke the page table entry. */
1390 flush_cache_page(vma, address, pte_pfn(*pte));
1391 pteval = ptep_clear_flush_notify(vma, address, pte);
1393 /* If nonlinear, store the file page offset in the pte. */
1394 if (page->index != linear_page_index(vma, address))
1395 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1397 /* Move the dirty bit to the physical page now the pte is gone. */
1398 if (pte_dirty(pteval))
1399 set_page_dirty(page);
1401 page_remove_rmap(page);
1402 page_cache_release(page);
1403 dec_mm_counter(mm, MM_FILEPAGES);
1406 pte_unmap_unlock(pte - 1, ptl);
1408 up_read(&vma->vm_mm->mmap_sem);
1412 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1414 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1419 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1420 VM_STACK_INCOMPLETE_SETUP)
1427 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1429 * @page: the page to unmap/unlock
1430 * @flags: action and flags
1432 * Find all the mappings of a page using the mapping pointer and the vma chains
1433 * contained in the anon_vma struct it points to.
1435 * This function is only called from try_to_unmap/try_to_munlock for
1437 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1438 * where the page was found will be held for write. So, we won't recheck
1439 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1442 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1444 struct anon_vma *anon_vma;
1446 struct anon_vma_chain *avc;
1447 int ret = SWAP_AGAIN;
1449 anon_vma = page_lock_anon_vma(page);
1453 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1454 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1455 struct vm_area_struct *vma = avc->vma;
1456 unsigned long address;
1459 * During exec, a temporary VMA is setup and later moved.
1460 * The VMA is moved under the anon_vma lock but not the
1461 * page tables leading to a race where migration cannot
1462 * find the migration ptes. Rather than increasing the
1463 * locking requirements of exec(), migration skips
1464 * temporary VMAs until after exec() completes.
1466 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1467 is_vma_temporary_stack(vma))
1470 address = vma_address(page, vma);
1471 if (address == -EFAULT)
1473 ret = try_to_unmap_one(page, vma, address, flags);
1474 if (ret != SWAP_AGAIN || !page_mapped(page))
1478 page_unlock_anon_vma(anon_vma);
1483 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1484 * @page: the page to unmap/unlock
1485 * @flags: action and flags
1487 * Find all the mappings of a page using the mapping pointer and the vma chains
1488 * contained in the address_space struct it points to.
1490 * This function is only called from try_to_unmap/try_to_munlock for
1491 * object-based pages.
1492 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1493 * where the page was found will be held for write. So, we won't recheck
1494 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1497 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1499 struct address_space *mapping = page->mapping;
1500 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1501 struct vm_area_struct *vma;
1502 int ret = SWAP_AGAIN;
1503 unsigned long cursor;
1504 unsigned long max_nl_cursor = 0;
1505 unsigned long max_nl_size = 0;
1506 unsigned int mapcount;
1508 mutex_lock(&mapping->i_mmap_mutex);
1509 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1510 unsigned long address = vma_address(page, vma);
1511 if (address == -EFAULT)
1513 ret = try_to_unmap_one(page, vma, address, flags);
1514 if (ret != SWAP_AGAIN || !page_mapped(page))
1518 if (list_empty(&mapping->i_mmap_nonlinear))
1522 * We don't bother to try to find the munlocked page in nonlinears.
1523 * It's costly. Instead, later, page reclaim logic may call
1524 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1526 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1529 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1531 cursor = (unsigned long) vma->vm_private_data;
1532 if (cursor > max_nl_cursor)
1533 max_nl_cursor = cursor;
1534 cursor = vma->vm_end - vma->vm_start;
1535 if (cursor > max_nl_size)
1536 max_nl_size = cursor;
1539 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1545 * We don't try to search for this page in the nonlinear vmas,
1546 * and page_referenced wouldn't have found it anyway. Instead
1547 * just walk the nonlinear vmas trying to age and unmap some.
1548 * The mapcount of the page we came in with is irrelevant,
1549 * but even so use it as a guide to how hard we should try?
1551 mapcount = page_mapcount(page);
1556 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1557 if (max_nl_cursor == 0)
1558 max_nl_cursor = CLUSTER_SIZE;
1561 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1563 cursor = (unsigned long) vma->vm_private_data;
1564 while ( cursor < max_nl_cursor &&
1565 cursor < vma->vm_end - vma->vm_start) {
1566 if (try_to_unmap_cluster(cursor, &mapcount,
1567 vma, page) == SWAP_MLOCK)
1569 cursor += CLUSTER_SIZE;
1570 vma->vm_private_data = (void *) cursor;
1571 if ((int)mapcount <= 0)
1574 vma->vm_private_data = (void *) max_nl_cursor;
1577 max_nl_cursor += CLUSTER_SIZE;
1578 } while (max_nl_cursor <= max_nl_size);
1581 * Don't loop forever (perhaps all the remaining pages are
1582 * in locked vmas). Reset cursor on all unreserved nonlinear
1583 * vmas, now forgetting on which ones it had fallen behind.
1585 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear)
1586 vma->vm_private_data = NULL;
1588 mutex_unlock(&mapping->i_mmap_mutex);
1593 * try_to_unmap - try to remove all page table mappings to a page
1594 * @page: the page to get unmapped
1595 * @flags: action and flags
1597 * Tries to remove all the page table entries which are mapping this
1598 * page, used in the pageout path. Caller must hold the page lock.
1599 * Return values are:
1601 * SWAP_SUCCESS - we succeeded in removing all mappings
1602 * SWAP_AGAIN - we missed a mapping, try again later
1603 * SWAP_FAIL - the page is unswappable
1604 * SWAP_MLOCK - page is mlocked.
1606 int try_to_unmap(struct page *page, enum ttu_flags flags)
1610 BUG_ON(!PageLocked(page));
1611 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1613 if (unlikely(PageKsm(page)))
1614 ret = try_to_unmap_ksm(page, flags);
1615 else if (PageAnon(page))
1616 ret = try_to_unmap_anon(page, flags);
1618 ret = try_to_unmap_file(page, flags);
1619 if (ret != SWAP_MLOCK && !page_mapped(page))
1625 * try_to_munlock - try to munlock a page
1626 * @page: the page to be munlocked
1628 * Called from munlock code. Checks all of the VMAs mapping the page
1629 * to make sure nobody else has this page mlocked. The page will be
1630 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1632 * Return values are:
1634 * SWAP_AGAIN - no vma is holding page mlocked, or,
1635 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1636 * SWAP_FAIL - page cannot be located at present
1637 * SWAP_MLOCK - page is now mlocked.
1639 int try_to_munlock(struct page *page)
1641 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1643 if (unlikely(PageKsm(page)))
1644 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1645 else if (PageAnon(page))
1646 return try_to_unmap_anon(page, TTU_MUNLOCK);
1648 return try_to_unmap_file(page, TTU_MUNLOCK);
1651 void __put_anon_vma(struct anon_vma *anon_vma)
1653 struct anon_vma *root = anon_vma->root;
1655 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1656 anon_vma_free(root);
1658 anon_vma_free(anon_vma);
1661 #ifdef CONFIG_MIGRATION
1663 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1664 * Called by migrate.c to remove migration ptes, but might be used more later.
1666 static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1667 struct vm_area_struct *, unsigned long, void *), void *arg)
1669 struct anon_vma *anon_vma;
1670 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1671 struct anon_vma_chain *avc;
1672 int ret = SWAP_AGAIN;
1675 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1676 * because that depends on page_mapped(); but not all its usages
1677 * are holding mmap_sem. Users without mmap_sem are required to
1678 * take a reference count to prevent the anon_vma disappearing
1680 anon_vma = page_anon_vma(page);
1683 anon_vma_lock(anon_vma);
1684 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1685 struct vm_area_struct *vma = avc->vma;
1686 unsigned long address = vma_address(page, vma);
1687 if (address == -EFAULT)
1689 ret = rmap_one(page, vma, address, arg);
1690 if (ret != SWAP_AGAIN)
1693 anon_vma_unlock(anon_vma);
1697 static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1698 struct vm_area_struct *, unsigned long, void *), void *arg)
1700 struct address_space *mapping = page->mapping;
1701 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1702 struct vm_area_struct *vma;
1703 int ret = SWAP_AGAIN;
1707 mutex_lock(&mapping->i_mmap_mutex);
1708 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1709 unsigned long address = vma_address(page, vma);
1710 if (address == -EFAULT)
1712 ret = rmap_one(page, vma, address, arg);
1713 if (ret != SWAP_AGAIN)
1717 * No nonlinear handling: being always shared, nonlinear vmas
1718 * never contain migration ptes. Decide what to do about this
1719 * limitation to linear when we need rmap_walk() on nonlinear.
1721 mutex_unlock(&mapping->i_mmap_mutex);
1725 int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1726 struct vm_area_struct *, unsigned long, void *), void *arg)
1728 VM_BUG_ON(!PageLocked(page));
1730 if (unlikely(PageKsm(page)))
1731 return rmap_walk_ksm(page, rmap_one, arg);
1732 else if (PageAnon(page))
1733 return rmap_walk_anon(page, rmap_one, arg);
1735 return rmap_walk_file(page, rmap_one, arg);
1737 #endif /* CONFIG_MIGRATION */
1739 #ifdef CONFIG_HUGETLB_PAGE
1741 * The following three functions are for anonymous (private mapped) hugepages.
1742 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1743 * and no lru code, because we handle hugepages differently from common pages.
1745 static void __hugepage_set_anon_rmap(struct page *page,
1746 struct vm_area_struct *vma, unsigned long address, int exclusive)
1748 struct anon_vma *anon_vma = vma->anon_vma;
1755 anon_vma = anon_vma->root;
1757 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1758 page->mapping = (struct address_space *) anon_vma;
1759 page->index = linear_page_index(vma, address);
1762 void hugepage_add_anon_rmap(struct page *page,
1763 struct vm_area_struct *vma, unsigned long address)
1765 struct anon_vma *anon_vma = vma->anon_vma;
1768 BUG_ON(!PageLocked(page));
1770 /* address might be in next vma when migration races vma_adjust */
1771 first = atomic_inc_and_test(&page->_mapcount);
1773 __hugepage_set_anon_rmap(page, vma, address, 0);
1776 void hugepage_add_new_anon_rmap(struct page *page,
1777 struct vm_area_struct *vma, unsigned long address)
1779 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1780 atomic_set(&page->_mapcount, 0);
1781 __hugepage_set_anon_rmap(page, vma, address, 1);
1783 #endif /* CONFIG_HUGETLB_PAGE */