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>
59 #include <linux/backing-dev.h>
61 #include <asm/tlbflush.h>
65 static struct kmem_cache *anon_vma_cachep;
66 static struct kmem_cache *anon_vma_chain_cachep;
68 static inline struct anon_vma *anon_vma_alloc(void)
70 struct anon_vma *anon_vma;
72 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
74 atomic_set(&anon_vma->refcount, 1);
75 anon_vma->degree = 1; /* Reference for first vma */
76 anon_vma->parent = anon_vma;
78 * Initialise the anon_vma root to point to itself. If called
79 * from fork, the root will be reset to the parents anon_vma.
81 anon_vma->root = anon_vma;
87 static inline void anon_vma_free(struct anon_vma *anon_vma)
89 VM_BUG_ON(atomic_read(&anon_vma->refcount));
92 * Synchronize against page_lock_anon_vma() such that
93 * we can safely hold the lock without the anon_vma getting
96 * Relies on the full mb implied by the atomic_dec_and_test() from
97 * put_anon_vma() against the acquire barrier implied by
98 * mutex_trylock() from page_lock_anon_vma(). This orders:
100 * page_lock_anon_vma() VS put_anon_vma()
101 * mutex_trylock() atomic_dec_and_test()
103 * atomic_read() mutex_is_locked()
105 * LOCK should suffice since the actual taking of the lock must
106 * happen _before_ what follows.
109 if (mutex_is_locked(&anon_vma->root->mutex)) {
110 anon_vma_lock(anon_vma);
111 anon_vma_unlock(anon_vma);
114 kmem_cache_free(anon_vma_cachep, anon_vma);
117 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
119 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
122 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
124 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
128 * anon_vma_prepare - attach an anon_vma to a memory region
129 * @vma: the memory region in question
131 * This makes sure the memory mapping described by 'vma' has
132 * an 'anon_vma' attached to it, so that we can associate the
133 * anonymous pages mapped into it with that anon_vma.
135 * The common case will be that we already have one, but if
136 * not we either need to find an adjacent mapping that we
137 * can re-use the anon_vma from (very common when the only
138 * reason for splitting a vma has been mprotect()), or we
139 * allocate a new one.
141 * Anon-vma allocations are very subtle, because we may have
142 * optimistically looked up an anon_vma in page_lock_anon_vma()
143 * and that may actually touch the spinlock even in the newly
144 * allocated vma (it depends on RCU to make sure that the
145 * anon_vma isn't actually destroyed).
147 * As a result, we need to do proper anon_vma locking even
148 * for the new allocation. At the same time, we do not want
149 * to do any locking for the common case of already having
152 * This must be called with the mmap_sem held for reading.
154 int anon_vma_prepare(struct vm_area_struct *vma)
156 struct anon_vma *anon_vma = vma->anon_vma;
157 struct anon_vma_chain *avc;
160 if (unlikely(!anon_vma)) {
161 struct mm_struct *mm = vma->vm_mm;
162 struct anon_vma *allocated;
164 avc = anon_vma_chain_alloc(GFP_KERNEL);
168 anon_vma = find_mergeable_anon_vma(vma);
171 anon_vma = anon_vma_alloc();
172 if (unlikely(!anon_vma))
173 goto out_enomem_free_avc;
174 allocated = anon_vma;
177 anon_vma_lock(anon_vma);
178 /* page_table_lock to protect against threads */
179 spin_lock(&mm->page_table_lock);
180 if (likely(!vma->anon_vma)) {
181 vma->anon_vma = anon_vma;
182 avc->anon_vma = anon_vma;
184 list_add(&avc->same_vma, &vma->anon_vma_chain);
185 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
186 /* vma reference or self-parent link for new root */
191 spin_unlock(&mm->page_table_lock);
192 anon_vma_unlock(anon_vma);
194 if (unlikely(allocated))
195 put_anon_vma(allocated);
197 anon_vma_chain_free(avc);
202 anon_vma_chain_free(avc);
208 * This is a useful helper function for locking the anon_vma root as
209 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
212 * Such anon_vma's should have the same root, so you'd expect to see
213 * just a single mutex_lock for the whole traversal.
215 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
217 struct anon_vma *new_root = anon_vma->root;
218 if (new_root != root) {
219 if (WARN_ON_ONCE(root))
220 mutex_unlock(&root->mutex);
222 mutex_lock(&root->mutex);
227 static inline void unlock_anon_vma_root(struct anon_vma *root)
230 mutex_unlock(&root->mutex);
233 static void anon_vma_chain_link(struct vm_area_struct *vma,
234 struct anon_vma_chain *avc,
235 struct anon_vma *anon_vma)
238 avc->anon_vma = anon_vma;
239 list_add(&avc->same_vma, &vma->anon_vma_chain);
242 * It's critical to add new vmas to the tail of the anon_vma,
243 * see comment in huge_memory.c:__split_huge_page().
245 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
249 * Attach the anon_vmas from src to dst.
250 * Returns 0 on success, -ENOMEM on failure.
252 * If dst->anon_vma is NULL this function tries to find and reuse existing
253 * anon_vma which has no vmas and only one child anon_vma. This prevents
254 * degradation of anon_vma hierarchy to endless linear chain in case of
255 * constantly forking task. On the other hand, an anon_vma with more than one
256 * child isn't reused even if there was no alive vma, thus rmap walker has a
257 * good chance of avoiding scanning the whole hierarchy when it searches where
260 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
262 struct anon_vma_chain *avc, *pavc;
263 struct anon_vma *root = NULL;
265 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
266 struct anon_vma *anon_vma;
268 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
269 if (unlikely(!avc)) {
270 unlock_anon_vma_root(root);
272 avc = anon_vma_chain_alloc(GFP_KERNEL);
276 anon_vma = pavc->anon_vma;
277 root = lock_anon_vma_root(root, anon_vma);
278 anon_vma_chain_link(dst, avc, anon_vma);
281 * Reuse existing anon_vma if its degree lower than two,
282 * that means it has no vma and only one anon_vma child.
284 * Do not chose parent anon_vma, otherwise first child
285 * will always reuse it. Root anon_vma is never reused:
286 * it has self-parent reference and at least one child.
288 if (!dst->anon_vma && anon_vma != src->anon_vma &&
289 anon_vma->degree < 2)
290 dst->anon_vma = anon_vma;
293 dst->anon_vma->degree++;
294 unlock_anon_vma_root(root);
298 unlink_anon_vmas(dst);
303 * Attach vma to its own anon_vma, as well as to the anon_vmas that
304 * the corresponding VMA in the parent process is attached to.
305 * Returns 0 on success, non-zero on failure.
307 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
309 struct anon_vma_chain *avc;
310 struct anon_vma *anon_vma;
312 /* Don't bother if the parent process has no anon_vma here. */
316 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
317 vma->anon_vma = NULL;
320 * First, attach the new VMA to the parent VMA's anon_vmas,
321 * so rmap can find non-COWed pages in child processes.
323 if (anon_vma_clone(vma, pvma))
326 /* An existing anon_vma has been reused, all done then. */
330 /* Then add our own anon_vma. */
331 anon_vma = anon_vma_alloc();
334 avc = anon_vma_chain_alloc(GFP_KERNEL);
336 goto out_error_free_anon_vma;
339 * The root anon_vma's spinlock is the lock actually used when we
340 * lock any of the anon_vmas in this anon_vma tree.
342 anon_vma->root = pvma->anon_vma->root;
343 anon_vma->parent = pvma->anon_vma;
345 * With refcounts, an anon_vma can stay around longer than the
346 * process it belongs to. The root anon_vma needs to be pinned until
347 * this anon_vma is freed, because the lock lives in the root.
349 get_anon_vma(anon_vma->root);
350 /* Mark this anon_vma as the one where our new (COWed) pages go. */
351 vma->anon_vma = anon_vma;
352 anon_vma_lock(anon_vma);
353 anon_vma_chain_link(vma, avc, anon_vma);
354 anon_vma->parent->degree++;
355 anon_vma_unlock(anon_vma);
359 out_error_free_anon_vma:
360 put_anon_vma(anon_vma);
362 unlink_anon_vmas(vma);
366 void unlink_anon_vmas(struct vm_area_struct *vma)
368 struct anon_vma_chain *avc, *next;
369 struct anon_vma *root = NULL;
372 * Unlink each anon_vma chained to the VMA. This list is ordered
373 * from newest to oldest, ensuring the root anon_vma gets freed last.
375 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
376 struct anon_vma *anon_vma = avc->anon_vma;
378 root = lock_anon_vma_root(root, anon_vma);
379 list_del(&avc->same_anon_vma);
382 * Leave empty anon_vmas on the list - we'll need
383 * to free them outside the lock.
385 if (list_empty(&anon_vma->head)) {
386 anon_vma->parent->degree--;
390 list_del(&avc->same_vma);
391 anon_vma_chain_free(avc);
394 vma->anon_vma->degree--;
395 unlock_anon_vma_root(root);
398 * Iterate the list once more, it now only contains empty and unlinked
399 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
400 * needing to acquire the anon_vma->root->mutex.
402 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
403 struct anon_vma *anon_vma = avc->anon_vma;
405 BUG_ON(anon_vma->degree);
406 put_anon_vma(anon_vma);
408 list_del(&avc->same_vma);
409 anon_vma_chain_free(avc);
413 static void anon_vma_ctor(void *data)
415 struct anon_vma *anon_vma = data;
417 mutex_init(&anon_vma->mutex);
418 atomic_set(&anon_vma->refcount, 0);
419 INIT_LIST_HEAD(&anon_vma->head);
422 void __init anon_vma_init(void)
424 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
425 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
426 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
430 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
432 * Since there is no serialization what so ever against page_remove_rmap()
433 * the best this function can do is return a locked anon_vma that might
434 * have been relevant to this page.
436 * The page might have been remapped to a different anon_vma or the anon_vma
437 * returned may already be freed (and even reused).
439 * In case it was remapped to a different anon_vma, the new anon_vma will be a
440 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
441 * ensure that any anon_vma obtained from the page will still be valid for as
442 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
444 * All users of this function must be very careful when walking the anon_vma
445 * chain and verify that the page in question is indeed mapped in it
446 * [ something equivalent to page_mapped_in_vma() ].
448 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
449 * that the anon_vma pointer from page->mapping is valid if there is a
450 * mapcount, we can dereference the anon_vma after observing those.
452 struct anon_vma *page_get_anon_vma(struct page *page)
454 struct anon_vma *anon_vma = NULL;
455 unsigned long anon_mapping;
458 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
459 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
461 if (!page_mapped(page))
464 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
465 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
471 * If this page is still mapped, then its anon_vma cannot have been
472 * freed. But if it has been unmapped, we have no security against the
473 * anon_vma structure being freed and reused (for another anon_vma:
474 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
475 * above cannot corrupt).
477 if (!page_mapped(page)) {
479 put_anon_vma(anon_vma);
489 * Similar to page_get_anon_vma() except it locks the anon_vma.
491 * Its a little more complex as it tries to keep the fast path to a single
492 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
493 * reference like with page_get_anon_vma() and then block on the mutex.
495 struct anon_vma *page_lock_anon_vma(struct page *page)
497 struct anon_vma *anon_vma = NULL;
498 struct anon_vma *root_anon_vma;
499 unsigned long anon_mapping;
502 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
503 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
505 if (!page_mapped(page))
508 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
509 root_anon_vma = ACCESS_ONCE(anon_vma->root);
510 if (mutex_trylock(&root_anon_vma->mutex)) {
512 * If the page is still mapped, then this anon_vma is still
513 * its anon_vma, and holding the mutex ensures that it will
514 * not go away, see anon_vma_free().
516 if (!page_mapped(page)) {
517 mutex_unlock(&root_anon_vma->mutex);
523 /* trylock failed, we got to sleep */
524 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
529 if (!page_mapped(page)) {
531 put_anon_vma(anon_vma);
535 /* we pinned the anon_vma, its safe to sleep */
537 anon_vma_lock(anon_vma);
539 if (atomic_dec_and_test(&anon_vma->refcount)) {
541 * Oops, we held the last refcount, release the lock
542 * and bail -- can't simply use put_anon_vma() because
543 * we'll deadlock on the anon_vma_lock() recursion.
545 anon_vma_unlock(anon_vma);
546 __put_anon_vma(anon_vma);
557 void page_unlock_anon_vma(struct anon_vma *anon_vma)
559 anon_vma_unlock(anon_vma);
563 * At what user virtual address is page expected in @vma?
564 * Returns virtual address or -EFAULT if page's index/offset is not
565 * within the range mapped the @vma.
568 vma_address(struct page *page, struct vm_area_struct *vma)
570 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
571 unsigned long address;
573 if (unlikely(is_vm_hugetlb_page(vma)))
574 pgoff = page->index << huge_page_order(page_hstate(page));
575 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
576 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
577 /* page should be within @vma mapping range */
584 * At what user virtual address is page expected in vma?
585 * Caller should check the page is actually part of the vma.
587 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
589 if (PageAnon(page)) {
590 struct anon_vma *page__anon_vma = page_anon_vma(page);
592 * Note: swapoff's unuse_vma() is more efficient with this
593 * check, and needs it to match anon_vma when KSM is active.
595 if (!vma->anon_vma || !page__anon_vma ||
596 vma->anon_vma->root != page__anon_vma->root)
598 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
600 vma->vm_file->f_mapping != page->mapping)
604 return vma_address(page, vma);
608 * Check that @page is mapped at @address into @mm.
610 * If @sync is false, page_check_address may perform a racy check to avoid
611 * the page table lock when the pte is not present (helpful when reclaiming
612 * highly shared pages).
614 * On success returns with pte mapped and locked.
616 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
617 unsigned long address, spinlock_t **ptlp, int sync)
625 if (unlikely(PageHuge(page))) {
626 /* when pud is not present, pte will be NULL */
627 pte = huge_pte_offset(mm, address);
631 ptl = &mm->page_table_lock;
635 pgd = pgd_offset(mm, address);
636 if (!pgd_present(*pgd))
639 pud = pud_offset(pgd, address);
640 if (!pud_present(*pud))
643 pmd = pmd_offset(pud, address);
644 if (!pmd_present(*pmd))
646 if (pmd_trans_huge(*pmd))
649 pte = pte_offset_map(pmd, address);
650 /* Make a quick check before getting the lock */
651 if (!sync && !pte_present(*pte)) {
656 ptl = pte_lockptr(mm, pmd);
659 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
663 pte_unmap_unlock(pte, ptl);
668 * page_mapped_in_vma - check whether a page is really mapped in a VMA
669 * @page: the page to test
670 * @vma: the VMA to test
672 * Returns 1 if the page is mapped into the page tables of the VMA, 0
673 * if the page is not mapped into the page tables of this VMA. Only
674 * valid for normal file or anonymous VMAs.
676 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
678 unsigned long address;
682 address = vma_address(page, vma);
683 if (address == -EFAULT) /* out of vma range */
685 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
686 if (!pte) /* the page is not in this mm */
688 pte_unmap_unlock(pte, ptl);
694 * Subfunctions of page_referenced: page_referenced_one called
695 * repeatedly from either page_referenced_anon or page_referenced_file.
697 int page_referenced_one(struct page *page, struct vm_area_struct *vma,
698 unsigned long address, unsigned int *mapcount,
699 unsigned long *vm_flags)
701 struct mm_struct *mm = vma->vm_mm;
704 if (unlikely(PageTransHuge(page))) {
707 spin_lock(&mm->page_table_lock);
709 * rmap might return false positives; we must filter
710 * these out using page_check_address_pmd().
712 pmd = page_check_address_pmd(page, mm, address,
713 PAGE_CHECK_ADDRESS_PMD_FLAG);
715 spin_unlock(&mm->page_table_lock);
719 if (vma->vm_flags & VM_LOCKED) {
720 spin_unlock(&mm->page_table_lock);
721 *mapcount = 0; /* break early from loop */
722 *vm_flags |= VM_LOCKED;
726 /* go ahead even if the pmd is pmd_trans_splitting() */
727 if (pmdp_clear_flush_young_notify(vma, address, pmd))
729 spin_unlock(&mm->page_table_lock);
735 * rmap might return false positives; we must filter
736 * these out using page_check_address().
738 pte = page_check_address(page, mm, address, &ptl, 0);
742 if (vma->vm_flags & VM_LOCKED) {
743 pte_unmap_unlock(pte, ptl);
744 *mapcount = 0; /* break early from loop */
745 *vm_flags |= VM_LOCKED;
749 if (ptep_clear_flush_young_notify(vma, address, pte)) {
751 * Don't treat a reference through a sequentially read
752 * mapping as such. If the page has been used in
753 * another mapping, we will catch it; if this other
754 * mapping is already gone, the unmap path will have
755 * set PG_referenced or activated the page.
757 if (likely(!VM_SequentialReadHint(vma)))
760 pte_unmap_unlock(pte, ptl);
763 /* Pretend the page is referenced if the task has the
764 swap token and is in the middle of a page fault. */
765 if (mm != current->mm && has_swap_token(mm) &&
766 rwsem_is_locked(&mm->mmap_sem))
772 *vm_flags |= vma->vm_flags;
777 static int page_referenced_anon(struct page *page,
778 struct mem_cgroup *mem_cont,
779 unsigned long *vm_flags)
781 unsigned int mapcount;
782 struct anon_vma *anon_vma;
783 struct anon_vma_chain *avc;
786 anon_vma = page_lock_anon_vma(page);
790 mapcount = page_mapcount(page);
791 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
792 struct vm_area_struct *vma = avc->vma;
793 unsigned long address = vma_address(page, vma);
794 if (address == -EFAULT)
797 * If we are reclaiming on behalf of a cgroup, skip
798 * counting on behalf of references from different
801 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
803 referenced += page_referenced_one(page, vma, address,
804 &mapcount, vm_flags);
809 page_unlock_anon_vma(anon_vma);
814 * page_referenced_file - referenced check for object-based rmap
815 * @page: the page we're checking references on.
816 * @mem_cont: target memory controller
817 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
819 * For an object-based mapped page, find all the places it is mapped and
820 * check/clear the referenced flag. This is done by following the page->mapping
821 * pointer, then walking the chain of vmas it holds. It returns the number
822 * of references it found.
824 * This function is only called from page_referenced for object-based pages.
826 static int page_referenced_file(struct page *page,
827 struct mem_cgroup *mem_cont,
828 unsigned long *vm_flags)
830 unsigned int mapcount;
831 struct address_space *mapping = page->mapping;
832 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
833 struct vm_area_struct *vma;
834 struct prio_tree_iter iter;
838 * The caller's checks on page->mapping and !PageAnon have made
839 * sure that this is a file page: the check for page->mapping
840 * excludes the case just before it gets set on an anon page.
842 BUG_ON(PageAnon(page));
845 * The page lock not only makes sure that page->mapping cannot
846 * suddenly be NULLified by truncation, it makes sure that the
847 * structure at mapping cannot be freed and reused yet,
848 * so we can safely take mapping->i_mmap_mutex.
850 BUG_ON(!PageLocked(page));
852 mutex_lock(&mapping->i_mmap_mutex);
855 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
856 * is more likely to be accurate if we note it after spinning.
858 mapcount = page_mapcount(page);
860 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
861 unsigned long address = vma_address(page, vma);
862 if (address == -EFAULT)
865 * If we are reclaiming on behalf of a cgroup, skip
866 * counting on behalf of references from different
869 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
871 referenced += page_referenced_one(page, vma, address,
872 &mapcount, vm_flags);
877 mutex_unlock(&mapping->i_mmap_mutex);
882 * page_referenced - test if the page was referenced
883 * @page: the page to test
884 * @is_locked: caller holds lock on the page
885 * @mem_cont: target memory controller
886 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
888 * Quick test_and_clear_referenced for all mappings to a page,
889 * returns the number of ptes which referenced the page.
891 int page_referenced(struct page *page,
893 struct mem_cgroup *mem_cont,
894 unsigned long *vm_flags)
900 if (page_mapped(page) && page_rmapping(page)) {
901 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
902 we_locked = trylock_page(page);
908 if (unlikely(PageKsm(page)))
909 referenced += page_referenced_ksm(page, mem_cont,
911 else if (PageAnon(page))
912 referenced += page_referenced_anon(page, mem_cont,
914 else if (page->mapping)
915 referenced += page_referenced_file(page, mem_cont,
920 if (page_test_and_clear_young(page_to_pfn(page)))
927 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
928 unsigned long address)
930 struct mm_struct *mm = vma->vm_mm;
935 pte = page_check_address(page, mm, address, &ptl, 1);
939 if (pte_dirty(*pte) || pte_write(*pte)) {
942 flush_cache_page(vma, address, pte_pfn(*pte));
943 entry = ptep_clear_flush_notify(vma, address, pte);
944 entry = pte_wrprotect(entry);
945 entry = pte_mkclean(entry);
946 set_pte_at(mm, address, pte, entry);
950 pte_unmap_unlock(pte, ptl);
955 static int page_mkclean_file(struct address_space *mapping, struct page *page)
957 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
958 struct vm_area_struct *vma;
959 struct prio_tree_iter iter;
962 BUG_ON(PageAnon(page));
964 mutex_lock(&mapping->i_mmap_mutex);
965 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
966 if (vma->vm_flags & VM_SHARED) {
967 unsigned long address = vma_address(page, vma);
968 if (address == -EFAULT)
970 ret += page_mkclean_one(page, vma, address);
973 mutex_unlock(&mapping->i_mmap_mutex);
977 int page_mkclean(struct page *page)
981 BUG_ON(!PageLocked(page));
983 if (page_mapped(page)) {
984 struct address_space *mapping = page_mapping(page);
986 ret = page_mkclean_file(mapping, page);
991 EXPORT_SYMBOL_GPL(page_mkclean);
994 * page_move_anon_rmap - move a page to our anon_vma
995 * @page: the page to move to our anon_vma
996 * @vma: the vma the page belongs to
997 * @address: the user virtual address mapped
999 * When a page belongs exclusively to one process after a COW event,
1000 * that page can be moved into the anon_vma that belongs to just that
1001 * process, so the rmap code will not search the parent or sibling
1004 void page_move_anon_rmap(struct page *page,
1005 struct vm_area_struct *vma, unsigned long address)
1007 struct anon_vma *anon_vma = vma->anon_vma;
1009 VM_BUG_ON(!PageLocked(page));
1010 VM_BUG_ON(!anon_vma);
1011 VM_BUG_ON(page->index != linear_page_index(vma, address));
1013 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1014 page->mapping = (struct address_space *) anon_vma;
1018 * __page_set_anon_rmap - set up new anonymous rmap
1019 * @page: Page to add to rmap
1020 * @vma: VM area to add page to.
1021 * @address: User virtual address of the mapping
1022 * @exclusive: the page is exclusively owned by the current process
1024 static void __page_set_anon_rmap(struct page *page,
1025 struct vm_area_struct *vma, unsigned long address, int exclusive)
1027 struct anon_vma *anon_vma = vma->anon_vma;
1035 * If the page isn't exclusively mapped into this vma,
1036 * we must use the _oldest_ possible anon_vma for the
1040 anon_vma = anon_vma->root;
1042 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1043 page->mapping = (struct address_space *) anon_vma;
1044 page->index = linear_page_index(vma, address);
1048 * __page_check_anon_rmap - sanity check anonymous rmap addition
1049 * @page: the page to add the mapping to
1050 * @vma: the vm area in which the mapping is added
1051 * @address: the user virtual address mapped
1053 static void __page_check_anon_rmap(struct page *page,
1054 struct vm_area_struct *vma, unsigned long address)
1056 #ifdef CONFIG_DEBUG_VM
1058 * The page's anon-rmap details (mapping and index) are guaranteed to
1059 * be set up correctly at this point.
1061 * We have exclusion against page_add_anon_rmap because the caller
1062 * always holds the page locked, except if called from page_dup_rmap,
1063 * in which case the page is already known to be setup.
1065 * We have exclusion against page_add_new_anon_rmap because those pages
1066 * are initially only visible via the pagetables, and the pte is locked
1067 * over the call to page_add_new_anon_rmap.
1069 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1070 BUG_ON(page->index != linear_page_index(vma, address));
1075 * page_add_anon_rmap - add pte mapping to an anonymous page
1076 * @page: the page to add the mapping to
1077 * @vma: the vm area in which the mapping is added
1078 * @address: the user virtual address mapped
1080 * The caller needs to hold the pte lock, and the page must be locked in
1081 * the anon_vma case: to serialize mapping,index checking after setting,
1082 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1083 * (but PageKsm is never downgraded to PageAnon).
1085 void page_add_anon_rmap(struct page *page,
1086 struct vm_area_struct *vma, unsigned long address)
1088 do_page_add_anon_rmap(page, vma, address, 0);
1092 * Special version of the above for do_swap_page, which often runs
1093 * into pages that are exclusively owned by the current process.
1094 * Everybody else should continue to use page_add_anon_rmap above.
1096 void do_page_add_anon_rmap(struct page *page,
1097 struct vm_area_struct *vma, unsigned long address, int exclusive)
1099 int first = atomic_inc_and_test(&page->_mapcount);
1101 if (!PageTransHuge(page))
1102 __inc_zone_page_state(page, NR_ANON_PAGES);
1104 __inc_zone_page_state(page,
1105 NR_ANON_TRANSPARENT_HUGEPAGES);
1107 if (unlikely(PageKsm(page)))
1110 VM_BUG_ON(!PageLocked(page));
1111 /* address might be in next vma when migration races vma_adjust */
1113 __page_set_anon_rmap(page, vma, address, exclusive);
1115 __page_check_anon_rmap(page, vma, address);
1119 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1120 * @page: the page to add the mapping to
1121 * @vma: the vm area in which the mapping is added
1122 * @address: the user virtual address mapped
1124 * Same as page_add_anon_rmap but must only be called on *new* pages.
1125 * This means the inc-and-test can be bypassed.
1126 * Page does not have to be locked.
1128 void page_add_new_anon_rmap(struct page *page,
1129 struct vm_area_struct *vma, unsigned long address)
1131 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1132 SetPageSwapBacked(page);
1133 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1134 if (!PageTransHuge(page))
1135 __inc_zone_page_state(page, NR_ANON_PAGES);
1137 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1138 __page_set_anon_rmap(page, vma, address, 1);
1139 if (page_evictable(page, vma))
1140 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1142 add_page_to_unevictable_list(page);
1146 * page_add_file_rmap - add pte mapping to a file page
1147 * @page: the page to add the mapping to
1149 * The caller needs to hold the pte lock.
1151 void page_add_file_rmap(struct page *page)
1153 if (atomic_inc_and_test(&page->_mapcount)) {
1154 __inc_zone_page_state(page, NR_FILE_MAPPED);
1155 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1160 * page_remove_rmap - take down pte mapping from a page
1161 * @page: page to remove mapping from
1163 * The caller needs to hold the pte lock.
1165 void page_remove_rmap(struct page *page)
1167 struct address_space *mapping = page_mapping(page);
1169 /* page still mapped by someone else? */
1170 if (!atomic_add_negative(-1, &page->_mapcount))
1174 * Now that the last pte has gone, s390 must transfer dirty
1175 * flag from storage key to struct page. We can usually skip
1176 * this if the page is anon, so about to be freed; but perhaps
1177 * not if it's in swapcache - there might be another pte slot
1178 * containing the swap entry, but page not yet written to swap.
1180 * And we can skip it on file pages, so long as the filesystem
1181 * participates in dirty tracking; but need to catch shm and tmpfs
1182 * and ramfs pages which have been modified since creation by read
1185 * Note that mapping must be decided above, before decrementing
1186 * mapcount (which luckily provides a barrier): once page is unmapped,
1187 * it could be truncated and page->mapping reset to NULL at any moment.
1188 * Note also that we are relying on page_mapping(page) to set mapping
1189 * to &swapper_space when PageSwapCache(page).
1191 if (mapping && !mapping_cap_account_dirty(mapping) &&
1192 page_test_and_clear_dirty(page_to_pfn(page), 1))
1193 set_page_dirty(page);
1195 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1196 * and not charged by memcg for now.
1198 if (unlikely(PageHuge(page)))
1200 if (PageAnon(page)) {
1201 mem_cgroup_uncharge_page(page);
1202 if (!PageTransHuge(page))
1203 __dec_zone_page_state(page, NR_ANON_PAGES);
1205 __dec_zone_page_state(page,
1206 NR_ANON_TRANSPARENT_HUGEPAGES);
1208 __dec_zone_page_state(page, NR_FILE_MAPPED);
1209 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1212 * It would be tidy to reset the PageAnon mapping here,
1213 * but that might overwrite a racing page_add_anon_rmap
1214 * which increments mapcount after us but sets mapping
1215 * before us: so leave the reset to free_hot_cold_page,
1216 * and remember that it's only reliable while mapped.
1217 * Leaving it set also helps swapoff to reinstate ptes
1218 * faster for those pages still in swapcache.
1223 * Subfunctions of try_to_unmap: try_to_unmap_one called
1224 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1226 int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1227 unsigned long address, enum ttu_flags flags)
1229 struct mm_struct *mm = vma->vm_mm;
1233 int ret = SWAP_AGAIN;
1235 pte = page_check_address(page, mm, address, &ptl, 0);
1240 * If the page is mlock()d, we cannot swap it out.
1241 * If it's recently referenced (perhaps page_referenced
1242 * skipped over this mm) then we should reactivate it.
1244 if (!(flags & TTU_IGNORE_MLOCK)) {
1245 if (vma->vm_flags & VM_LOCKED)
1248 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1251 if (!(flags & TTU_IGNORE_ACCESS)) {
1252 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1258 /* Nuke the page table entry. */
1259 flush_cache_page(vma, address, page_to_pfn(page));
1260 pteval = ptep_clear_flush_notify(vma, address, pte);
1262 /* Move the dirty bit to the physical page now the pte is gone. */
1263 if (pte_dirty(pteval))
1264 set_page_dirty(page);
1266 /* Update high watermark before we lower rss */
1267 update_hiwater_rss(mm);
1269 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1271 dec_mm_counter(mm, MM_ANONPAGES);
1273 dec_mm_counter(mm, MM_FILEPAGES);
1274 set_pte_at(mm, address, pte,
1275 swp_entry_to_pte(make_hwpoison_entry(page)));
1276 } else if (PageAnon(page)) {
1277 swp_entry_t entry = { .val = page_private(page) };
1279 if (PageSwapCache(page)) {
1281 * Store the swap location in the pte.
1282 * See handle_pte_fault() ...
1284 if (swap_duplicate(entry) < 0) {
1285 set_pte_at(mm, address, pte, pteval);
1289 if (list_empty(&mm->mmlist)) {
1290 spin_lock(&mmlist_lock);
1291 if (list_empty(&mm->mmlist))
1292 list_add(&mm->mmlist, &init_mm.mmlist);
1293 spin_unlock(&mmlist_lock);
1295 dec_mm_counter(mm, MM_ANONPAGES);
1296 inc_mm_counter(mm, MM_SWAPENTS);
1297 } else if (PAGE_MIGRATION) {
1299 * Store the pfn of the page in a special migration
1300 * pte. do_swap_page() will wait until the migration
1301 * pte is removed and then restart fault handling.
1303 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1304 entry = make_migration_entry(page, pte_write(pteval));
1306 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1307 BUG_ON(pte_file(*pte));
1308 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
1309 /* Establish migration entry for a file page */
1311 entry = make_migration_entry(page, pte_write(pteval));
1312 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1314 dec_mm_counter(mm, MM_FILEPAGES);
1316 page_remove_rmap(page);
1317 page_cache_release(page);
1320 pte_unmap_unlock(pte, ptl);
1325 pte_unmap_unlock(pte, ptl);
1329 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1330 * unstable result and race. Plus, We can't wait here because
1331 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1332 * if trylock failed, the page remain in evictable lru and later
1333 * vmscan could retry to move the page to unevictable lru if the
1334 * page is actually mlocked.
1336 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1337 if (vma->vm_flags & VM_LOCKED) {
1338 mlock_vma_page(page);
1341 up_read(&vma->vm_mm->mmap_sem);
1347 * objrmap doesn't work for nonlinear VMAs because the assumption that
1348 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1349 * Consequently, given a particular page and its ->index, we cannot locate the
1350 * ptes which are mapping that page without an exhaustive linear search.
1352 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1353 * maps the file to which the target page belongs. The ->vm_private_data field
1354 * holds the current cursor into that scan. Successive searches will circulate
1355 * around the vma's virtual address space.
1357 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1358 * more scanning pressure is placed against them as well. Eventually pages
1359 * will become fully unmapped and are eligible for eviction.
1361 * For very sparsely populated VMAs this is a little inefficient - chances are
1362 * there there won't be many ptes located within the scan cluster. In this case
1363 * maybe we could scan further - to the end of the pte page, perhaps.
1365 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1366 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1367 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1368 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1370 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1371 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1373 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1374 struct vm_area_struct *vma, struct page *check_page)
1376 struct mm_struct *mm = vma->vm_mm;
1384 unsigned long address;
1386 int ret = SWAP_AGAIN;
1389 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1390 end = address + CLUSTER_SIZE;
1391 if (address < vma->vm_start)
1392 address = vma->vm_start;
1393 if (end > vma->vm_end)
1396 pgd = pgd_offset(mm, address);
1397 if (!pgd_present(*pgd))
1400 pud = pud_offset(pgd, address);
1401 if (!pud_present(*pud))
1404 pmd = pmd_offset(pud, address);
1405 if (!pmd_present(*pmd))
1409 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1410 * keep the sem while scanning the cluster for mlocking pages.
1412 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1413 locked_vma = (vma->vm_flags & VM_LOCKED);
1415 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1418 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1420 /* Update high watermark before we lower rss */
1421 update_hiwater_rss(mm);
1423 for (; address < end; pte++, address += PAGE_SIZE) {
1424 if (!pte_present(*pte))
1426 page = vm_normal_page(vma, address, *pte);
1427 BUG_ON(!page || PageAnon(page));
1430 if (page == check_page) {
1431 /* we know we have check_page locked */
1432 mlock_vma_page(page);
1434 } else if (trylock_page(page)) {
1436 * If we can lock the page, perform mlock.
1437 * Otherwise leave the page alone, it will be
1438 * eventually encountered again later.
1440 mlock_vma_page(page);
1443 continue; /* don't unmap */
1446 if (ptep_clear_flush_young_notify(vma, address, pte))
1449 /* Nuke the page table entry. */
1450 flush_cache_page(vma, address, pte_pfn(*pte));
1451 pteval = ptep_clear_flush_notify(vma, address, pte);
1453 /* If nonlinear, store the file page offset in the pte. */
1454 if (page->index != linear_page_index(vma, address))
1455 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1457 /* Move the dirty bit to the physical page now the pte is gone. */
1458 if (pte_dirty(pteval))
1459 set_page_dirty(page);
1461 page_remove_rmap(page);
1462 page_cache_release(page);
1463 dec_mm_counter(mm, MM_FILEPAGES);
1466 pte_unmap_unlock(pte - 1, ptl);
1468 up_read(&vma->vm_mm->mmap_sem);
1472 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1474 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1479 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1480 VM_STACK_INCOMPLETE_SETUP)
1487 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1489 * @page: the page to unmap/unlock
1490 * @flags: action and flags
1492 * Find all the mappings of a page using the mapping pointer and the vma chains
1493 * contained in the anon_vma struct it points to.
1495 * This function is only called from try_to_unmap/try_to_munlock for
1497 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1498 * where the page was found will be held for write. So, we won't recheck
1499 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1502 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1504 struct anon_vma *anon_vma;
1505 struct anon_vma_chain *avc;
1506 int ret = SWAP_AGAIN;
1508 anon_vma = page_lock_anon_vma(page);
1512 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1513 struct vm_area_struct *vma = avc->vma;
1514 unsigned long address;
1517 * During exec, a temporary VMA is setup and later moved.
1518 * The VMA is moved under the anon_vma lock but not the
1519 * page tables leading to a race where migration cannot
1520 * find the migration ptes. Rather than increasing the
1521 * locking requirements of exec(), migration skips
1522 * temporary VMAs until after exec() completes.
1524 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1525 is_vma_temporary_stack(vma))
1528 address = vma_address(page, vma);
1529 if (address == -EFAULT)
1531 ret = try_to_unmap_one(page, vma, address, flags);
1532 if (ret != SWAP_AGAIN || !page_mapped(page))
1536 page_unlock_anon_vma(anon_vma);
1541 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1542 * @page: the page to unmap/unlock
1543 * @flags: action and flags
1545 * Find all the mappings of a page using the mapping pointer and the vma chains
1546 * contained in the address_space struct it points to.
1548 * This function is only called from try_to_unmap/try_to_munlock for
1549 * object-based pages.
1550 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1551 * where the page was found will be held for write. So, we won't recheck
1552 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1555 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1557 struct address_space *mapping = page->mapping;
1558 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1559 struct vm_area_struct *vma;
1560 struct prio_tree_iter iter;
1561 int ret = SWAP_AGAIN;
1562 unsigned long cursor;
1563 unsigned long max_nl_cursor = 0;
1564 unsigned long max_nl_size = 0;
1565 unsigned int mapcount;
1567 mutex_lock(&mapping->i_mmap_mutex);
1568 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1569 unsigned long address = vma_address(page, vma);
1570 if (address == -EFAULT)
1572 ret = try_to_unmap_one(page, vma, address, flags);
1573 if (ret != SWAP_AGAIN || !page_mapped(page))
1577 if (list_empty(&mapping->i_mmap_nonlinear))
1581 * We don't bother to try to find the munlocked page in nonlinears.
1582 * It's costly. Instead, later, page reclaim logic may call
1583 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1585 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1588 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1589 shared.vm_set.list) {
1590 cursor = (unsigned long) vma->vm_private_data;
1591 if (cursor > max_nl_cursor)
1592 max_nl_cursor = cursor;
1593 cursor = vma->vm_end - vma->vm_start;
1594 if (cursor > max_nl_size)
1595 max_nl_size = cursor;
1598 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1604 * We don't try to search for this page in the nonlinear vmas,
1605 * and page_referenced wouldn't have found it anyway. Instead
1606 * just walk the nonlinear vmas trying to age and unmap some.
1607 * The mapcount of the page we came in with is irrelevant,
1608 * but even so use it as a guide to how hard we should try?
1610 mapcount = page_mapcount(page);
1615 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1616 if (max_nl_cursor == 0)
1617 max_nl_cursor = CLUSTER_SIZE;
1620 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1621 shared.vm_set.list) {
1622 cursor = (unsigned long) vma->vm_private_data;
1623 while ( cursor < max_nl_cursor &&
1624 cursor < vma->vm_end - vma->vm_start) {
1625 if (try_to_unmap_cluster(cursor, &mapcount,
1626 vma, page) == SWAP_MLOCK)
1628 cursor += CLUSTER_SIZE;
1629 vma->vm_private_data = (void *) cursor;
1630 if ((int)mapcount <= 0)
1633 vma->vm_private_data = (void *) max_nl_cursor;
1636 max_nl_cursor += CLUSTER_SIZE;
1637 } while (max_nl_cursor <= max_nl_size);
1640 * Don't loop forever (perhaps all the remaining pages are
1641 * in locked vmas). Reset cursor on all unreserved nonlinear
1642 * vmas, now forgetting on which ones it had fallen behind.
1644 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1645 vma->vm_private_data = NULL;
1647 mutex_unlock(&mapping->i_mmap_mutex);
1652 * try_to_unmap - try to remove all page table mappings to a page
1653 * @page: the page to get unmapped
1654 * @flags: action and flags
1656 * Tries to remove all the page table entries which are mapping this
1657 * page, used in the pageout path. Caller must hold the page lock.
1658 * Return values are:
1660 * SWAP_SUCCESS - we succeeded in removing all mappings
1661 * SWAP_AGAIN - we missed a mapping, try again later
1662 * SWAP_FAIL - the page is unswappable
1663 * SWAP_MLOCK - page is mlocked.
1665 int try_to_unmap(struct page *page, enum ttu_flags flags)
1669 BUG_ON(!PageLocked(page));
1670 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1672 if (unlikely(PageKsm(page)))
1673 ret = try_to_unmap_ksm(page, flags);
1674 else if (PageAnon(page))
1675 ret = try_to_unmap_anon(page, flags);
1677 ret = try_to_unmap_file(page, flags);
1678 if (ret != SWAP_MLOCK && !page_mapped(page))
1684 * try_to_munlock - try to munlock a page
1685 * @page: the page to be munlocked
1687 * Called from munlock code. Checks all of the VMAs mapping the page
1688 * to make sure nobody else has this page mlocked. The page will be
1689 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1691 * Return values are:
1693 * SWAP_AGAIN - no vma is holding page mlocked, or,
1694 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1695 * SWAP_FAIL - page cannot be located at present
1696 * SWAP_MLOCK - page is now mlocked.
1698 int try_to_munlock(struct page *page)
1700 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1702 if (unlikely(PageKsm(page)))
1703 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1704 else if (PageAnon(page))
1705 return try_to_unmap_anon(page, TTU_MUNLOCK);
1707 return try_to_unmap_file(page, TTU_MUNLOCK);
1710 void __put_anon_vma(struct anon_vma *anon_vma)
1712 struct anon_vma *root = anon_vma->root;
1714 anon_vma_free(anon_vma);
1715 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1716 anon_vma_free(root);
1719 #ifdef CONFIG_MIGRATION
1721 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1722 * Called by migrate.c to remove migration ptes, but might be used more later.
1724 static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1725 struct vm_area_struct *, unsigned long, void *), void *arg)
1727 struct anon_vma *anon_vma;
1728 struct anon_vma_chain *avc;
1729 int ret = SWAP_AGAIN;
1732 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1733 * because that depends on page_mapped(); but not all its usages
1734 * are holding mmap_sem. Users without mmap_sem are required to
1735 * take a reference count to prevent the anon_vma disappearing
1737 anon_vma = page_anon_vma(page);
1740 anon_vma_lock(anon_vma);
1741 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1742 struct vm_area_struct *vma = avc->vma;
1743 unsigned long address = vma_address(page, vma);
1744 if (address == -EFAULT)
1746 ret = rmap_one(page, vma, address, arg);
1747 if (ret != SWAP_AGAIN)
1750 anon_vma_unlock(anon_vma);
1754 static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1755 struct vm_area_struct *, unsigned long, void *), void *arg)
1757 struct address_space *mapping = page->mapping;
1758 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1759 struct vm_area_struct *vma;
1760 struct prio_tree_iter iter;
1761 int ret = SWAP_AGAIN;
1765 mutex_lock(&mapping->i_mmap_mutex);
1766 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1767 unsigned long address = vma_address(page, vma);
1768 if (address == -EFAULT)
1770 ret = rmap_one(page, vma, address, arg);
1771 if (ret != SWAP_AGAIN)
1775 * No nonlinear handling: being always shared, nonlinear vmas
1776 * never contain migration ptes. Decide what to do about this
1777 * limitation to linear when we need rmap_walk() on nonlinear.
1779 mutex_unlock(&mapping->i_mmap_mutex);
1783 int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1784 struct vm_area_struct *, unsigned long, void *), void *arg)
1786 VM_BUG_ON(!PageLocked(page));
1788 if (unlikely(PageKsm(page)))
1789 return rmap_walk_ksm(page, rmap_one, arg);
1790 else if (PageAnon(page))
1791 return rmap_walk_anon(page, rmap_one, arg);
1793 return rmap_walk_file(page, rmap_one, arg);
1795 #endif /* CONFIG_MIGRATION */
1797 #ifdef CONFIG_HUGETLB_PAGE
1799 * The following three functions are for anonymous (private mapped) hugepages.
1800 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1801 * and no lru code, because we handle hugepages differently from common pages.
1803 static void __hugepage_set_anon_rmap(struct page *page,
1804 struct vm_area_struct *vma, unsigned long address, int exclusive)
1806 struct anon_vma *anon_vma = vma->anon_vma;
1813 anon_vma = anon_vma->root;
1815 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1816 page->mapping = (struct address_space *) anon_vma;
1817 page->index = linear_page_index(vma, address);
1820 void hugepage_add_anon_rmap(struct page *page,
1821 struct vm_area_struct *vma, unsigned long address)
1823 struct anon_vma *anon_vma = vma->anon_vma;
1826 BUG_ON(!PageLocked(page));
1828 /* address might be in next vma when migration races vma_adjust */
1829 first = atomic_inc_and_test(&page->_mapcount);
1831 __hugepage_set_anon_rmap(page, vma, address, 0);
1834 void hugepage_add_new_anon_rmap(struct page *page,
1835 struct vm_area_struct *vma, unsigned long address)
1837 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1838 atomic_set(&page->_mapcount, 0);
1839 __hugepage_set_anon_rmap(page, vma, address, 1);
1841 #endif /* CONFIG_HUGETLB_PAGE */