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);
124 * anon_vma_prepare - attach an anon_vma to a memory region
125 * @vma: the memory region in question
127 * This makes sure the memory mapping described by 'vma' has
128 * an 'anon_vma' attached to it, so that we can associate the
129 * anonymous pages mapped into it with that anon_vma.
131 * The common case will be that we already have one, but if
132 * not we either need to find an adjacent mapping that we
133 * can re-use the anon_vma from (very common when the only
134 * reason for splitting a vma has been mprotect()), or we
135 * allocate a new one.
137 * Anon-vma allocations are very subtle, because we may have
138 * optimistically looked up an anon_vma in page_lock_anon_vma()
139 * and that may actually touch the spinlock even in the newly
140 * allocated vma (it depends on RCU to make sure that the
141 * anon_vma isn't actually destroyed).
143 * As a result, we need to do proper anon_vma locking even
144 * for the new allocation. At the same time, we do not want
145 * to do any locking for the common case of already having
148 * This must be called with the mmap_sem held for reading.
150 int anon_vma_prepare(struct vm_area_struct *vma)
152 struct anon_vma *anon_vma = vma->anon_vma;
153 struct anon_vma_chain *avc;
156 if (unlikely(!anon_vma)) {
157 struct mm_struct *mm = vma->vm_mm;
158 struct anon_vma *allocated;
160 avc = anon_vma_chain_alloc(GFP_KERNEL);
164 anon_vma = find_mergeable_anon_vma(vma);
167 anon_vma = anon_vma_alloc();
168 if (unlikely(!anon_vma))
169 goto out_enomem_free_avc;
170 allocated = anon_vma;
173 anon_vma_lock(anon_vma);
174 /* page_table_lock to protect against threads */
175 spin_lock(&mm->page_table_lock);
176 if (likely(!vma->anon_vma)) {
177 vma->anon_vma = anon_vma;
178 avc->anon_vma = anon_vma;
180 list_add(&avc->same_vma, &vma->anon_vma_chain);
181 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
185 spin_unlock(&mm->page_table_lock);
186 anon_vma_unlock(anon_vma);
188 if (unlikely(allocated))
189 put_anon_vma(allocated);
191 anon_vma_chain_free(avc);
196 anon_vma_chain_free(avc);
202 * This is a useful helper function for locking the anon_vma root as
203 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
206 * Such anon_vma's should have the same root, so you'd expect to see
207 * just a single mutex_lock for the whole traversal.
209 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
211 struct anon_vma *new_root = anon_vma->root;
212 if (new_root != root) {
213 if (WARN_ON_ONCE(root))
214 mutex_unlock(&root->mutex);
216 mutex_lock(&root->mutex);
221 static inline void unlock_anon_vma_root(struct anon_vma *root)
224 mutex_unlock(&root->mutex);
227 static void anon_vma_chain_link(struct vm_area_struct *vma,
228 struct anon_vma_chain *avc,
229 struct anon_vma *anon_vma)
232 avc->anon_vma = anon_vma;
233 list_add(&avc->same_vma, &vma->anon_vma_chain);
236 * It's critical to add new vmas to the tail of the anon_vma,
237 * see comment in huge_memory.c:__split_huge_page().
239 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
243 * Attach the anon_vmas from src to dst.
244 * Returns 0 on success, -ENOMEM on failure.
246 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
248 struct anon_vma_chain *avc, *pavc;
249 struct anon_vma *root = NULL;
251 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
252 struct anon_vma *anon_vma;
254 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
255 if (unlikely(!avc)) {
256 unlock_anon_vma_root(root);
258 avc = anon_vma_chain_alloc(GFP_KERNEL);
262 anon_vma = pavc->anon_vma;
263 root = lock_anon_vma_root(root, anon_vma);
264 anon_vma_chain_link(dst, avc, anon_vma);
266 unlock_anon_vma_root(root);
270 unlink_anon_vmas(dst);
275 * Attach vma to its own anon_vma, as well as to the anon_vmas that
276 * the corresponding VMA in the parent process is attached to.
277 * Returns 0 on success, non-zero on failure.
279 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
281 struct anon_vma_chain *avc;
282 struct anon_vma *anon_vma;
284 /* Don't bother if the parent process has no anon_vma here. */
289 * First, attach the new VMA to the parent VMA's anon_vmas,
290 * so rmap can find non-COWed pages in child processes.
292 if (anon_vma_clone(vma, pvma))
295 /* Then add our own anon_vma. */
296 anon_vma = anon_vma_alloc();
299 avc = anon_vma_chain_alloc(GFP_KERNEL);
301 goto out_error_free_anon_vma;
304 * The root anon_vma's spinlock is the lock actually used when we
305 * lock any of the anon_vmas in this anon_vma tree.
307 anon_vma->root = pvma->anon_vma->root;
309 * With refcounts, an anon_vma can stay around longer than the
310 * process it belongs to. The root anon_vma needs to be pinned until
311 * this anon_vma is freed, because the lock lives in the root.
313 get_anon_vma(anon_vma->root);
314 /* Mark this anon_vma as the one where our new (COWed) pages go. */
315 vma->anon_vma = anon_vma;
316 anon_vma_lock(anon_vma);
317 anon_vma_chain_link(vma, avc, anon_vma);
318 anon_vma_unlock(anon_vma);
322 out_error_free_anon_vma:
323 put_anon_vma(anon_vma);
325 unlink_anon_vmas(vma);
329 void unlink_anon_vmas(struct vm_area_struct *vma)
331 struct anon_vma_chain *avc, *next;
332 struct anon_vma *root = NULL;
335 * Unlink each anon_vma chained to the VMA. This list is ordered
336 * from newest to oldest, ensuring the root anon_vma gets freed last.
338 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
339 struct anon_vma *anon_vma = avc->anon_vma;
341 root = lock_anon_vma_root(root, anon_vma);
342 list_del(&avc->same_anon_vma);
345 * Leave empty anon_vmas on the list - we'll need
346 * to free them outside the lock.
348 if (list_empty(&anon_vma->head))
351 list_del(&avc->same_vma);
352 anon_vma_chain_free(avc);
354 unlock_anon_vma_root(root);
357 * Iterate the list once more, it now only contains empty and unlinked
358 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
359 * needing to acquire the anon_vma->root->mutex.
361 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
362 struct anon_vma *anon_vma = avc->anon_vma;
364 put_anon_vma(anon_vma);
366 list_del(&avc->same_vma);
367 anon_vma_chain_free(avc);
371 static void anon_vma_ctor(void *data)
373 struct anon_vma *anon_vma = data;
375 mutex_init(&anon_vma->mutex);
376 atomic_set(&anon_vma->refcount, 0);
377 INIT_LIST_HEAD(&anon_vma->head);
380 void __init anon_vma_init(void)
382 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
383 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
384 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
388 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
390 * Since there is no serialization what so ever against page_remove_rmap()
391 * the best this function can do is return a locked anon_vma that might
392 * have been relevant to this page.
394 * The page might have been remapped to a different anon_vma or the anon_vma
395 * returned may already be freed (and even reused).
397 * In case it was remapped to a different anon_vma, the new anon_vma will be a
398 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
399 * ensure that any anon_vma obtained from the page will still be valid for as
400 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
402 * All users of this function must be very careful when walking the anon_vma
403 * chain and verify that the page in question is indeed mapped in it
404 * [ something equivalent to page_mapped_in_vma() ].
406 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
407 * that the anon_vma pointer from page->mapping is valid if there is a
408 * mapcount, we can dereference the anon_vma after observing those.
410 struct anon_vma *page_get_anon_vma(struct page *page)
412 struct anon_vma *anon_vma = NULL;
413 unsigned long anon_mapping;
416 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
417 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
419 if (!page_mapped(page))
422 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
423 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
429 * If this page is still mapped, then its anon_vma cannot have been
430 * freed. But if it has been unmapped, we have no security against the
431 * anon_vma structure being freed and reused (for another anon_vma:
432 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
433 * above cannot corrupt).
435 if (!page_mapped(page)) {
436 put_anon_vma(anon_vma);
446 * Similar to page_get_anon_vma() except it locks the anon_vma.
448 * Its a little more complex as it tries to keep the fast path to a single
449 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
450 * reference like with page_get_anon_vma() and then block on the mutex.
452 struct anon_vma *page_lock_anon_vma(struct page *page)
454 struct anon_vma *anon_vma = NULL;
455 struct anon_vma *root_anon_vma;
456 unsigned long anon_mapping;
459 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
460 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
462 if (!page_mapped(page))
465 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
466 root_anon_vma = ACCESS_ONCE(anon_vma->root);
467 if (mutex_trylock(&root_anon_vma->mutex)) {
469 * If the page is still mapped, then this anon_vma is still
470 * its anon_vma, and holding the mutex ensures that it will
471 * not go away, see anon_vma_free().
473 if (!page_mapped(page)) {
474 mutex_unlock(&root_anon_vma->mutex);
480 /* trylock failed, we got to sleep */
481 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
486 if (!page_mapped(page)) {
487 put_anon_vma(anon_vma);
492 /* we pinned the anon_vma, its safe to sleep */
494 anon_vma_lock(anon_vma);
496 if (atomic_dec_and_test(&anon_vma->refcount)) {
498 * Oops, we held the last refcount, release the lock
499 * and bail -- can't simply use put_anon_vma() because
500 * we'll deadlock on the anon_vma_lock() recursion.
502 anon_vma_unlock(anon_vma);
503 __put_anon_vma(anon_vma);
514 void page_unlock_anon_vma(struct anon_vma *anon_vma)
516 anon_vma_unlock(anon_vma);
520 * At what user virtual address is page expected in @vma?
521 * Returns virtual address or -EFAULT if page's index/offset is not
522 * within the range mapped the @vma.
525 vma_address(struct page *page, struct vm_area_struct *vma)
527 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
528 unsigned long address;
530 if (unlikely(is_vm_hugetlb_page(vma)))
531 pgoff = page->index << huge_page_order(page_hstate(page));
532 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
533 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
534 /* page should be within @vma mapping range */
541 * At what user virtual address is page expected in vma?
542 * Caller should check the page is actually part of the vma.
544 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
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 return vma_address(page, vma);
565 * Check that @page is mapped at @address into @mm.
567 * If @sync is false, page_check_address may perform a racy check to avoid
568 * the page table lock when the pte is not present (helpful when reclaiming
569 * highly shared pages).
571 * On success returns with pte mapped and locked.
573 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
574 unsigned long address, spinlock_t **ptlp, int sync)
582 if (unlikely(PageHuge(page))) {
583 pte = huge_pte_offset(mm, address);
584 ptl = &mm->page_table_lock;
588 pgd = pgd_offset(mm, address);
589 if (!pgd_present(*pgd))
592 pud = pud_offset(pgd, address);
593 if (!pud_present(*pud))
596 pmd = pmd_offset(pud, address);
597 if (!pmd_present(*pmd))
599 if (pmd_trans_huge(*pmd))
602 pte = pte_offset_map(pmd, address);
603 /* Make a quick check before getting the lock */
604 if (!sync && !pte_present(*pte)) {
609 ptl = pte_lockptr(mm, pmd);
612 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
616 pte_unmap_unlock(pte, ptl);
621 * page_mapped_in_vma - check whether a page is really mapped in a VMA
622 * @page: the page to test
623 * @vma: the VMA to test
625 * Returns 1 if the page is mapped into the page tables of the VMA, 0
626 * if the page is not mapped into the page tables of this VMA. Only
627 * valid for normal file or anonymous VMAs.
629 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
631 unsigned long address;
635 address = vma_address(page, vma);
636 if (address == -EFAULT) /* out of vma range */
638 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
639 if (!pte) /* the page is not in this mm */
641 pte_unmap_unlock(pte, ptl);
647 * Subfunctions of page_referenced: page_referenced_one called
648 * repeatedly from either page_referenced_anon or page_referenced_file.
650 int page_referenced_one(struct page *page, struct vm_area_struct *vma,
651 unsigned long address, unsigned int *mapcount,
652 unsigned long *vm_flags)
654 struct mm_struct *mm = vma->vm_mm;
657 if (unlikely(PageTransHuge(page))) {
660 spin_lock(&mm->page_table_lock);
662 * rmap might return false positives; we must filter
663 * these out using page_check_address_pmd().
665 pmd = page_check_address_pmd(page, mm, address,
666 PAGE_CHECK_ADDRESS_PMD_FLAG);
668 spin_unlock(&mm->page_table_lock);
672 if (vma->vm_flags & VM_LOCKED) {
673 spin_unlock(&mm->page_table_lock);
674 *mapcount = 0; /* break early from loop */
675 *vm_flags |= VM_LOCKED;
679 /* go ahead even if the pmd is pmd_trans_splitting() */
680 if (pmdp_clear_flush_young_notify(vma, address, pmd))
682 spin_unlock(&mm->page_table_lock);
688 * rmap might return false positives; we must filter
689 * these out using page_check_address().
691 pte = page_check_address(page, mm, address, &ptl, 0);
695 if (vma->vm_flags & VM_LOCKED) {
696 pte_unmap_unlock(pte, ptl);
697 *mapcount = 0; /* break early from loop */
698 *vm_flags |= VM_LOCKED;
702 if (ptep_clear_flush_young_notify(vma, address, pte)) {
704 * Don't treat a reference through a sequentially read
705 * mapping as such. If the page has been used in
706 * another mapping, we will catch it; if this other
707 * mapping is already gone, the unmap path will have
708 * set PG_referenced or activated the page.
710 if (likely(!VM_SequentialReadHint(vma)))
713 pte_unmap_unlock(pte, ptl);
716 /* Pretend the page is referenced if the task has the
717 swap token and is in the middle of a page fault. */
718 if (mm != current->mm && has_active_swap_token(mm))
724 *vm_flags |= vma->vm_flags;
729 static int page_referenced_anon(struct page *page,
730 struct mem_cgroup *mem_cont,
731 unsigned long *vm_flags)
733 unsigned int mapcount;
734 struct anon_vma *anon_vma;
735 struct anon_vma_chain *avc;
738 anon_vma = page_lock_anon_vma(page);
742 mapcount = page_mapcount(page);
743 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
744 struct vm_area_struct *vma = avc->vma;
745 unsigned long address = vma_address(page, vma);
746 if (address == -EFAULT)
749 * If we are reclaiming on behalf of a cgroup, skip
750 * counting on behalf of references from different
753 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
755 referenced += page_referenced_one(page, vma, address,
756 &mapcount, vm_flags);
761 page_unlock_anon_vma(anon_vma);
766 * page_referenced_file - referenced check for object-based rmap
767 * @page: the page we're checking references on.
768 * @mem_cont: target memory controller
769 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
771 * For an object-based mapped page, find all the places it is mapped and
772 * check/clear the referenced flag. This is done by following the page->mapping
773 * pointer, then walking the chain of vmas it holds. It returns the number
774 * of references it found.
776 * This function is only called from page_referenced for object-based pages.
778 static int page_referenced_file(struct page *page,
779 struct mem_cgroup *mem_cont,
780 unsigned long *vm_flags)
782 unsigned int mapcount;
783 struct address_space *mapping = page->mapping;
784 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
785 struct vm_area_struct *vma;
786 struct prio_tree_iter iter;
790 * The caller's checks on page->mapping and !PageAnon have made
791 * sure that this is a file page: the check for page->mapping
792 * excludes the case just before it gets set on an anon page.
794 BUG_ON(PageAnon(page));
797 * The page lock not only makes sure that page->mapping cannot
798 * suddenly be NULLified by truncation, it makes sure that the
799 * structure at mapping cannot be freed and reused yet,
800 * so we can safely take mapping->i_mmap_mutex.
802 BUG_ON(!PageLocked(page));
804 mutex_lock(&mapping->i_mmap_mutex);
807 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
808 * is more likely to be accurate if we note it after spinning.
810 mapcount = page_mapcount(page);
812 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
813 unsigned long address = vma_address(page, vma);
814 if (address == -EFAULT)
817 * If we are reclaiming on behalf of a cgroup, skip
818 * counting on behalf of references from different
821 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
823 referenced += page_referenced_one(page, vma, address,
824 &mapcount, vm_flags);
829 mutex_unlock(&mapping->i_mmap_mutex);
834 * page_referenced - test if the page was referenced
835 * @page: the page to test
836 * @is_locked: caller holds lock on the page
837 * @mem_cont: target memory controller
838 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
840 * Quick test_and_clear_referenced for all mappings to a page,
841 * returns the number of ptes which referenced the page.
843 int page_referenced(struct page *page,
845 struct mem_cgroup *mem_cont,
846 unsigned long *vm_flags)
852 if (page_mapped(page) && page_rmapping(page)) {
853 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
854 we_locked = trylock_page(page);
860 if (unlikely(PageKsm(page)))
861 referenced += page_referenced_ksm(page, mem_cont,
863 else if (PageAnon(page))
864 referenced += page_referenced_anon(page, mem_cont,
866 else if (page->mapping)
867 referenced += page_referenced_file(page, mem_cont,
872 if (page_test_and_clear_young(page_to_pfn(page)))
879 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
880 unsigned long address)
882 struct mm_struct *mm = vma->vm_mm;
887 pte = page_check_address(page, mm, address, &ptl, 1);
891 if (pte_dirty(*pte) || pte_write(*pte)) {
894 flush_cache_page(vma, address, pte_pfn(*pte));
895 entry = ptep_clear_flush_notify(vma, address, pte);
896 entry = pte_wrprotect(entry);
897 entry = pte_mkclean(entry);
898 set_pte_at(mm, address, pte, entry);
902 pte_unmap_unlock(pte, ptl);
907 static int page_mkclean_file(struct address_space *mapping, struct page *page)
909 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
910 struct vm_area_struct *vma;
911 struct prio_tree_iter iter;
914 BUG_ON(PageAnon(page));
916 mutex_lock(&mapping->i_mmap_mutex);
917 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
918 if (vma->vm_flags & VM_SHARED) {
919 unsigned long address = vma_address(page, vma);
920 if (address == -EFAULT)
922 ret += page_mkclean_one(page, vma, address);
925 mutex_unlock(&mapping->i_mmap_mutex);
929 int page_mkclean(struct page *page)
933 BUG_ON(!PageLocked(page));
935 if (page_mapped(page)) {
936 struct address_space *mapping = page_mapping(page);
938 ret = page_mkclean_file(mapping, page);
939 if (page_test_and_clear_dirty(page_to_pfn(page), 1))
946 EXPORT_SYMBOL_GPL(page_mkclean);
949 * page_move_anon_rmap - move a page to our anon_vma
950 * @page: the page to move to our anon_vma
951 * @vma: the vma the page belongs to
952 * @address: the user virtual address mapped
954 * When a page belongs exclusively to one process after a COW event,
955 * that page can be moved into the anon_vma that belongs to just that
956 * process, so the rmap code will not search the parent or sibling
959 void page_move_anon_rmap(struct page *page,
960 struct vm_area_struct *vma, unsigned long address)
962 struct anon_vma *anon_vma = vma->anon_vma;
964 VM_BUG_ON(!PageLocked(page));
965 VM_BUG_ON(!anon_vma);
966 VM_BUG_ON(page->index != linear_page_index(vma, address));
968 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
969 page->mapping = (struct address_space *) anon_vma;
973 * __page_set_anon_rmap - set up new anonymous rmap
974 * @page: Page to add to rmap
975 * @vma: VM area to add page to.
976 * @address: User virtual address of the mapping
977 * @exclusive: the page is exclusively owned by the current process
979 static void __page_set_anon_rmap(struct page *page,
980 struct vm_area_struct *vma, unsigned long address, int exclusive)
982 struct anon_vma *anon_vma = vma->anon_vma;
990 * If the page isn't exclusively mapped into this vma,
991 * we must use the _oldest_ possible anon_vma for the
995 anon_vma = anon_vma->root;
997 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
998 page->mapping = (struct address_space *) anon_vma;
999 page->index = linear_page_index(vma, address);
1003 * __page_check_anon_rmap - sanity check anonymous rmap addition
1004 * @page: the page to add the mapping to
1005 * @vma: the vm area in which the mapping is added
1006 * @address: the user virtual address mapped
1008 static void __page_check_anon_rmap(struct page *page,
1009 struct vm_area_struct *vma, unsigned long address)
1011 #ifdef CONFIG_DEBUG_VM
1013 * The page's anon-rmap details (mapping and index) are guaranteed to
1014 * be set up correctly at this point.
1016 * We have exclusion against page_add_anon_rmap because the caller
1017 * always holds the page locked, except if called from page_dup_rmap,
1018 * in which case the page is already known to be setup.
1020 * We have exclusion against page_add_new_anon_rmap because those pages
1021 * are initially only visible via the pagetables, and the pte is locked
1022 * over the call to page_add_new_anon_rmap.
1024 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1025 BUG_ON(page->index != linear_page_index(vma, address));
1030 * page_add_anon_rmap - add pte mapping to an anonymous page
1031 * @page: the page to add the mapping to
1032 * @vma: the vm area in which the mapping is added
1033 * @address: the user virtual address mapped
1035 * The caller needs to hold the pte lock, and the page must be locked in
1036 * the anon_vma case: to serialize mapping,index checking after setting,
1037 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1038 * (but PageKsm is never downgraded to PageAnon).
1040 void page_add_anon_rmap(struct page *page,
1041 struct vm_area_struct *vma, unsigned long address)
1043 do_page_add_anon_rmap(page, vma, address, 0);
1047 * Special version of the above for do_swap_page, which often runs
1048 * into pages that are exclusively owned by the current process.
1049 * Everybody else should continue to use page_add_anon_rmap above.
1051 void do_page_add_anon_rmap(struct page *page,
1052 struct vm_area_struct *vma, unsigned long address, int exclusive)
1054 int first = atomic_inc_and_test(&page->_mapcount);
1056 if (!PageTransHuge(page))
1057 __inc_zone_page_state(page, NR_ANON_PAGES);
1059 __inc_zone_page_state(page,
1060 NR_ANON_TRANSPARENT_HUGEPAGES);
1062 if (unlikely(PageKsm(page)))
1065 VM_BUG_ON(!PageLocked(page));
1066 /* address might be in next vma when migration races vma_adjust */
1068 __page_set_anon_rmap(page, vma, address, exclusive);
1070 __page_check_anon_rmap(page, vma, address);
1074 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1075 * @page: the page to add the mapping to
1076 * @vma: the vm area in which the mapping is added
1077 * @address: the user virtual address mapped
1079 * Same as page_add_anon_rmap but must only be called on *new* pages.
1080 * This means the inc-and-test can be bypassed.
1081 * Page does not have to be locked.
1083 void page_add_new_anon_rmap(struct page *page,
1084 struct vm_area_struct *vma, unsigned long address)
1086 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1087 SetPageSwapBacked(page);
1088 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1089 if (!PageTransHuge(page))
1090 __inc_zone_page_state(page, NR_ANON_PAGES);
1092 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1093 __page_set_anon_rmap(page, vma, address, 1);
1094 if (page_evictable(page, vma))
1095 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1097 add_page_to_unevictable_list(page);
1101 * page_add_file_rmap - add pte mapping to a file page
1102 * @page: the page to add the mapping to
1104 * The caller needs to hold the pte lock.
1106 void page_add_file_rmap(struct page *page)
1108 if (atomic_inc_and_test(&page->_mapcount)) {
1109 __inc_zone_page_state(page, NR_FILE_MAPPED);
1110 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1115 * page_remove_rmap - take down pte mapping from a page
1116 * @page: page to remove mapping from
1118 * The caller needs to hold the pte lock.
1120 void page_remove_rmap(struct page *page)
1122 /* page still mapped by someone else? */
1123 if (!atomic_add_negative(-1, &page->_mapcount))
1127 * Now that the last pte has gone, s390 must transfer dirty
1128 * flag from storage key to struct page. We can usually skip
1129 * this if the page is anon, so about to be freed; but perhaps
1130 * not if it's in swapcache - there might be another pte slot
1131 * containing the swap entry, but page not yet written to swap.
1133 if ((!PageAnon(page) || PageSwapCache(page)) &&
1134 page_test_and_clear_dirty(page_to_pfn(page), 1))
1135 set_page_dirty(page);
1137 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1138 * and not charged by memcg for now.
1140 if (unlikely(PageHuge(page)))
1142 if (PageAnon(page)) {
1143 mem_cgroup_uncharge_page(page);
1144 if (!PageTransHuge(page))
1145 __dec_zone_page_state(page, NR_ANON_PAGES);
1147 __dec_zone_page_state(page,
1148 NR_ANON_TRANSPARENT_HUGEPAGES);
1150 __dec_zone_page_state(page, NR_FILE_MAPPED);
1151 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1154 * It would be tidy to reset the PageAnon mapping here,
1155 * but that might overwrite a racing page_add_anon_rmap
1156 * which increments mapcount after us but sets mapping
1157 * before us: so leave the reset to free_hot_cold_page,
1158 * and remember that it's only reliable while mapped.
1159 * Leaving it set also helps swapoff to reinstate ptes
1160 * faster for those pages still in swapcache.
1165 * Subfunctions of try_to_unmap: try_to_unmap_one called
1166 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1168 int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1169 unsigned long address, enum ttu_flags flags)
1171 struct mm_struct *mm = vma->vm_mm;
1175 int ret = SWAP_AGAIN;
1177 pte = page_check_address(page, mm, address, &ptl, 0);
1182 * If the page is mlock()d, we cannot swap it out.
1183 * If it's recently referenced (perhaps page_referenced
1184 * skipped over this mm) then we should reactivate it.
1186 if (!(flags & TTU_IGNORE_MLOCK)) {
1187 if (vma->vm_flags & VM_LOCKED)
1190 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1193 if (!(flags & TTU_IGNORE_ACCESS)) {
1194 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1200 /* Nuke the page table entry. */
1201 flush_cache_page(vma, address, page_to_pfn(page));
1202 pteval = ptep_clear_flush_notify(vma, address, pte);
1204 /* Move the dirty bit to the physical page now the pte is gone. */
1205 if (pte_dirty(pteval))
1206 set_page_dirty(page);
1208 /* Update high watermark before we lower rss */
1209 update_hiwater_rss(mm);
1211 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1213 dec_mm_counter(mm, MM_ANONPAGES);
1215 dec_mm_counter(mm, MM_FILEPAGES);
1216 set_pte_at(mm, address, pte,
1217 swp_entry_to_pte(make_hwpoison_entry(page)));
1218 } else if (PageAnon(page)) {
1219 swp_entry_t entry = { .val = page_private(page) };
1221 if (PageSwapCache(page)) {
1223 * Store the swap location in the pte.
1224 * See handle_pte_fault() ...
1226 if (swap_duplicate(entry) < 0) {
1227 set_pte_at(mm, address, pte, pteval);
1231 if (list_empty(&mm->mmlist)) {
1232 spin_lock(&mmlist_lock);
1233 if (list_empty(&mm->mmlist))
1234 list_add(&mm->mmlist, &init_mm.mmlist);
1235 spin_unlock(&mmlist_lock);
1237 dec_mm_counter(mm, MM_ANONPAGES);
1238 inc_mm_counter(mm, MM_SWAPENTS);
1239 } else if (PAGE_MIGRATION) {
1241 * Store the pfn of the page in a special migration
1242 * pte. do_swap_page() will wait until the migration
1243 * pte is removed and then restart fault handling.
1245 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1246 entry = make_migration_entry(page, pte_write(pteval));
1248 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1249 BUG_ON(pte_file(*pte));
1250 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
1251 /* Establish migration entry for a file page */
1253 entry = make_migration_entry(page, pte_write(pteval));
1254 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1256 dec_mm_counter(mm, MM_FILEPAGES);
1258 page_remove_rmap(page);
1259 page_cache_release(page);
1262 pte_unmap_unlock(pte, ptl);
1267 pte_unmap_unlock(pte, ptl);
1271 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1272 * unstable result and race. Plus, We can't wait here because
1273 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1274 * if trylock failed, the page remain in evictable lru and later
1275 * vmscan could retry to move the page to unevictable lru if the
1276 * page is actually mlocked.
1278 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1279 if (vma->vm_flags & VM_LOCKED) {
1280 mlock_vma_page(page);
1283 up_read(&vma->vm_mm->mmap_sem);
1289 * objrmap doesn't work for nonlinear VMAs because the assumption that
1290 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1291 * Consequently, given a particular page and its ->index, we cannot locate the
1292 * ptes which are mapping that page without an exhaustive linear search.
1294 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1295 * maps the file to which the target page belongs. The ->vm_private_data field
1296 * holds the current cursor into that scan. Successive searches will circulate
1297 * around the vma's virtual address space.
1299 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1300 * more scanning pressure is placed against them as well. Eventually pages
1301 * will become fully unmapped and are eligible for eviction.
1303 * For very sparsely populated VMAs this is a little inefficient - chances are
1304 * there there won't be many ptes located within the scan cluster. In this case
1305 * maybe we could scan further - to the end of the pte page, perhaps.
1307 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1308 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1309 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1310 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1312 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1313 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1315 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1316 struct vm_area_struct *vma, struct page *check_page)
1318 struct mm_struct *mm = vma->vm_mm;
1326 unsigned long address;
1328 int ret = SWAP_AGAIN;
1331 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1332 end = address + CLUSTER_SIZE;
1333 if (address < vma->vm_start)
1334 address = vma->vm_start;
1335 if (end > vma->vm_end)
1338 pgd = pgd_offset(mm, address);
1339 if (!pgd_present(*pgd))
1342 pud = pud_offset(pgd, address);
1343 if (!pud_present(*pud))
1346 pmd = pmd_offset(pud, address);
1347 if (!pmd_present(*pmd))
1351 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1352 * keep the sem while scanning the cluster for mlocking pages.
1354 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1355 locked_vma = (vma->vm_flags & VM_LOCKED);
1357 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1360 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1362 /* Update high watermark before we lower rss */
1363 update_hiwater_rss(mm);
1365 for (; address < end; pte++, address += PAGE_SIZE) {
1366 if (!pte_present(*pte))
1368 page = vm_normal_page(vma, address, *pte);
1369 BUG_ON(!page || PageAnon(page));
1372 mlock_vma_page(page); /* no-op if already mlocked */
1373 if (page == check_page)
1375 continue; /* don't unmap */
1378 if (ptep_clear_flush_young_notify(vma, address, pte))
1381 /* Nuke the page table entry. */
1382 flush_cache_page(vma, address, pte_pfn(*pte));
1383 pteval = ptep_clear_flush_notify(vma, address, pte);
1385 /* If nonlinear, store the file page offset in the pte. */
1386 if (page->index != linear_page_index(vma, address))
1387 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1389 /* Move the dirty bit to the physical page now the pte is gone. */
1390 if (pte_dirty(pteval))
1391 set_page_dirty(page);
1393 page_remove_rmap(page);
1394 page_cache_release(page);
1395 dec_mm_counter(mm, MM_FILEPAGES);
1398 pte_unmap_unlock(pte - 1, ptl);
1400 up_read(&vma->vm_mm->mmap_sem);
1404 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1406 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1411 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1412 VM_STACK_INCOMPLETE_SETUP)
1419 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1421 * @page: the page to unmap/unlock
1422 * @flags: action and flags
1424 * Find all the mappings of a page using the mapping pointer and the vma chains
1425 * contained in the anon_vma struct it points to.
1427 * This function is only called from try_to_unmap/try_to_munlock for
1429 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1430 * where the page was found will be held for write. So, we won't recheck
1431 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1434 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1436 struct anon_vma *anon_vma;
1437 struct anon_vma_chain *avc;
1438 int ret = SWAP_AGAIN;
1440 anon_vma = page_lock_anon_vma(page);
1444 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1445 struct vm_area_struct *vma = avc->vma;
1446 unsigned long address;
1449 * During exec, a temporary VMA is setup and later moved.
1450 * The VMA is moved under the anon_vma lock but not the
1451 * page tables leading to a race where migration cannot
1452 * find the migration ptes. Rather than increasing the
1453 * locking requirements of exec(), migration skips
1454 * temporary VMAs until after exec() completes.
1456 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1457 is_vma_temporary_stack(vma))
1460 address = vma_address(page, vma);
1461 if (address == -EFAULT)
1463 ret = try_to_unmap_one(page, vma, address, flags);
1464 if (ret != SWAP_AGAIN || !page_mapped(page))
1468 page_unlock_anon_vma(anon_vma);
1473 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1474 * @page: the page to unmap/unlock
1475 * @flags: action and flags
1477 * Find all the mappings of a page using the mapping pointer and the vma chains
1478 * contained in the address_space struct it points to.
1480 * This function is only called from try_to_unmap/try_to_munlock for
1481 * object-based pages.
1482 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1483 * where the page was found will be held for write. So, we won't recheck
1484 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1487 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1489 struct address_space *mapping = page->mapping;
1490 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1491 struct vm_area_struct *vma;
1492 struct prio_tree_iter iter;
1493 int ret = SWAP_AGAIN;
1494 unsigned long cursor;
1495 unsigned long max_nl_cursor = 0;
1496 unsigned long max_nl_size = 0;
1497 unsigned int mapcount;
1499 mutex_lock(&mapping->i_mmap_mutex);
1500 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1501 unsigned long address = vma_address(page, vma);
1502 if (address == -EFAULT)
1504 ret = try_to_unmap_one(page, vma, address, flags);
1505 if (ret != SWAP_AGAIN || !page_mapped(page))
1509 if (list_empty(&mapping->i_mmap_nonlinear))
1513 * We don't bother to try to find the munlocked page in nonlinears.
1514 * It's costly. Instead, later, page reclaim logic may call
1515 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1517 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1520 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1521 shared.vm_set.list) {
1522 cursor = (unsigned long) vma->vm_private_data;
1523 if (cursor > max_nl_cursor)
1524 max_nl_cursor = cursor;
1525 cursor = vma->vm_end - vma->vm_start;
1526 if (cursor > max_nl_size)
1527 max_nl_size = cursor;
1530 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1536 * We don't try to search for this page in the nonlinear vmas,
1537 * and page_referenced wouldn't have found it anyway. Instead
1538 * just walk the nonlinear vmas trying to age and unmap some.
1539 * The mapcount of the page we came in with is irrelevant,
1540 * but even so use it as a guide to how hard we should try?
1542 mapcount = page_mapcount(page);
1547 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1548 if (max_nl_cursor == 0)
1549 max_nl_cursor = CLUSTER_SIZE;
1552 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1553 shared.vm_set.list) {
1554 cursor = (unsigned long) vma->vm_private_data;
1555 while ( cursor < max_nl_cursor &&
1556 cursor < vma->vm_end - vma->vm_start) {
1557 if (try_to_unmap_cluster(cursor, &mapcount,
1558 vma, page) == SWAP_MLOCK)
1560 cursor += CLUSTER_SIZE;
1561 vma->vm_private_data = (void *) cursor;
1562 if ((int)mapcount <= 0)
1565 vma->vm_private_data = (void *) max_nl_cursor;
1568 max_nl_cursor += CLUSTER_SIZE;
1569 } while (max_nl_cursor <= max_nl_size);
1572 * Don't loop forever (perhaps all the remaining pages are
1573 * in locked vmas). Reset cursor on all unreserved nonlinear
1574 * vmas, now forgetting on which ones it had fallen behind.
1576 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1577 vma->vm_private_data = NULL;
1579 mutex_unlock(&mapping->i_mmap_mutex);
1584 * try_to_unmap - try to remove all page table mappings to a page
1585 * @page: the page to get unmapped
1586 * @flags: action and flags
1588 * Tries to remove all the page table entries which are mapping this
1589 * page, used in the pageout path. Caller must hold the page lock.
1590 * Return values are:
1592 * SWAP_SUCCESS - we succeeded in removing all mappings
1593 * SWAP_AGAIN - we missed a mapping, try again later
1594 * SWAP_FAIL - the page is unswappable
1595 * SWAP_MLOCK - page is mlocked.
1597 int try_to_unmap(struct page *page, enum ttu_flags flags)
1601 BUG_ON(!PageLocked(page));
1602 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1604 if (unlikely(PageKsm(page)))
1605 ret = try_to_unmap_ksm(page, flags);
1606 else if (PageAnon(page))
1607 ret = try_to_unmap_anon(page, flags);
1609 ret = try_to_unmap_file(page, flags);
1610 if (ret != SWAP_MLOCK && !page_mapped(page))
1616 * try_to_munlock - try to munlock a page
1617 * @page: the page to be munlocked
1619 * Called from munlock code. Checks all of the VMAs mapping the page
1620 * to make sure nobody else has this page mlocked. The page will be
1621 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1623 * Return values are:
1625 * SWAP_AGAIN - no vma is holding page mlocked, or,
1626 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1627 * SWAP_FAIL - page cannot be located at present
1628 * SWAP_MLOCK - page is now mlocked.
1630 int try_to_munlock(struct page *page)
1632 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1634 if (unlikely(PageKsm(page)))
1635 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1636 else if (PageAnon(page))
1637 return try_to_unmap_anon(page, TTU_MUNLOCK);
1639 return try_to_unmap_file(page, TTU_MUNLOCK);
1642 void __put_anon_vma(struct anon_vma *anon_vma)
1644 struct anon_vma *root = anon_vma->root;
1646 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1647 anon_vma_free(root);
1649 anon_vma_free(anon_vma);
1652 #ifdef CONFIG_MIGRATION
1654 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1655 * Called by migrate.c to remove migration ptes, but might be used more later.
1657 static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1658 struct vm_area_struct *, unsigned long, void *), void *arg)
1660 struct anon_vma *anon_vma;
1661 struct anon_vma_chain *avc;
1662 int ret = SWAP_AGAIN;
1665 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1666 * because that depends on page_mapped(); but not all its usages
1667 * are holding mmap_sem. Users without mmap_sem are required to
1668 * take a reference count to prevent the anon_vma disappearing
1670 anon_vma = page_anon_vma(page);
1673 anon_vma_lock(anon_vma);
1674 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1675 struct vm_area_struct *vma = avc->vma;
1676 unsigned long address = vma_address(page, vma);
1677 if (address == -EFAULT)
1679 ret = rmap_one(page, vma, address, arg);
1680 if (ret != SWAP_AGAIN)
1683 anon_vma_unlock(anon_vma);
1687 static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1688 struct vm_area_struct *, unsigned long, void *), void *arg)
1690 struct address_space *mapping = page->mapping;
1691 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1692 struct vm_area_struct *vma;
1693 struct prio_tree_iter iter;
1694 int ret = SWAP_AGAIN;
1698 mutex_lock(&mapping->i_mmap_mutex);
1699 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1700 unsigned long address = vma_address(page, vma);
1701 if (address == -EFAULT)
1703 ret = rmap_one(page, vma, address, arg);
1704 if (ret != SWAP_AGAIN)
1708 * No nonlinear handling: being always shared, nonlinear vmas
1709 * never contain migration ptes. Decide what to do about this
1710 * limitation to linear when we need rmap_walk() on nonlinear.
1712 mutex_unlock(&mapping->i_mmap_mutex);
1716 int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1717 struct vm_area_struct *, unsigned long, void *), void *arg)
1719 VM_BUG_ON(!PageLocked(page));
1721 if (unlikely(PageKsm(page)))
1722 return rmap_walk_ksm(page, rmap_one, arg);
1723 else if (PageAnon(page))
1724 return rmap_walk_anon(page, rmap_one, arg);
1726 return rmap_walk_file(page, rmap_one, arg);
1728 #endif /* CONFIG_MIGRATION */
1730 #ifdef CONFIG_HUGETLB_PAGE
1732 * The following three functions are for anonymous (private mapped) hugepages.
1733 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1734 * and no lru code, because we handle hugepages differently from common pages.
1736 static void __hugepage_set_anon_rmap(struct page *page,
1737 struct vm_area_struct *vma, unsigned long address, int exclusive)
1739 struct anon_vma *anon_vma = vma->anon_vma;
1746 anon_vma = anon_vma->root;
1748 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1749 page->mapping = (struct address_space *) anon_vma;
1750 page->index = linear_page_index(vma, address);
1753 void hugepage_add_anon_rmap(struct page *page,
1754 struct vm_area_struct *vma, unsigned long address)
1756 struct anon_vma *anon_vma = vma->anon_vma;
1759 BUG_ON(!PageLocked(page));
1761 /* address might be in next vma when migration races vma_adjust */
1762 first = atomic_inc_and_test(&page->_mapcount);
1764 __hugepage_set_anon_rmap(page, vma, address, 0);
1767 void hugepage_add_new_anon_rmap(struct page *page,
1768 struct vm_area_struct *vma, unsigned long address)
1770 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1771 atomic_set(&page->_mapcount, 0);
1772 __hugepage_set_anon_rmap(page, vma, address, 1);
1774 #endif /* CONFIG_HUGETLB_PAGE */