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 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
29 * mm->page_table_lock or pte_lock
30 * pgdat->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
35 * i_pages lock (widely used)
36 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
37 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
38 * sb_lock (within inode_lock in fs/fs-writeback.c)
39 * i_pages lock (widely used, in set_page_dirty,
40 * in arch-dependent flush_dcache_mmap_lock,
41 * within bdi.wb->list_lock in __sync_single_inode)
43 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
49 #include <linux/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/pagemap.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/slab.h>
55 #include <linux/init.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/rcupdate.h>
59 #include <linux/export.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/huge_mm.h>
65 #include <linux/backing-dev.h>
66 #include <linux/page_idle.h>
67 #include <linux/memremap.h>
68 #include <linux/userfaultfd_k.h>
70 #include <asm/tlbflush.h>
72 #include <trace/events/tlb.h>
76 static struct kmem_cache *anon_vma_cachep;
77 static struct kmem_cache *anon_vma_chain_cachep;
79 static inline struct anon_vma *anon_vma_alloc(void)
81 struct anon_vma *anon_vma;
83 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
85 atomic_set(&anon_vma->refcount, 1);
86 anon_vma->num_children = 0;
87 anon_vma->num_active_vmas = 0;
88 anon_vma->parent = anon_vma;
90 * Initialise the anon_vma root to point to itself. If called
91 * from fork, the root will be reset to the parents anon_vma.
93 anon_vma->root = anon_vma;
99 static inline void anon_vma_free(struct anon_vma *anon_vma)
101 VM_BUG_ON(atomic_read(&anon_vma->refcount));
104 * Synchronize against page_lock_anon_vma_read() such that
105 * we can safely hold the lock without the anon_vma getting
108 * Relies on the full mb implied by the atomic_dec_and_test() from
109 * put_anon_vma() against the acquire barrier implied by
110 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
112 * page_lock_anon_vma_read() VS put_anon_vma()
113 * down_read_trylock() atomic_dec_and_test()
115 * atomic_read() rwsem_is_locked()
117 * LOCK should suffice since the actual taking of the lock must
118 * happen _before_ what follows.
121 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
122 anon_vma_lock_write(anon_vma);
123 anon_vma_unlock_write(anon_vma);
126 kmem_cache_free(anon_vma_cachep, anon_vma);
129 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
131 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
134 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
136 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
139 static void anon_vma_chain_link(struct vm_area_struct *vma,
140 struct anon_vma_chain *avc,
141 struct anon_vma *anon_vma)
144 avc->anon_vma = anon_vma;
145 list_add(&avc->same_vma, &vma->anon_vma_chain);
146 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
150 * __anon_vma_prepare - attach an anon_vma to a memory region
151 * @vma: the memory region in question
153 * This makes sure the memory mapping described by 'vma' has
154 * an 'anon_vma' attached to it, so that we can associate the
155 * anonymous pages mapped into it with that anon_vma.
157 * The common case will be that we already have one, which
158 * is handled inline by anon_vma_prepare(). But if
159 * not we either need to find an adjacent mapping that we
160 * can re-use the anon_vma from (very common when the only
161 * reason for splitting a vma has been mprotect()), or we
162 * allocate a new one.
164 * Anon-vma allocations are very subtle, because we may have
165 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
166 * and that may actually touch the spinlock even in the newly
167 * allocated vma (it depends on RCU to make sure that the
168 * anon_vma isn't actually destroyed).
170 * As a result, we need to do proper anon_vma locking even
171 * for the new allocation. At the same time, we do not want
172 * to do any locking for the common case of already having
175 * This must be called with the mmap_sem held for reading.
177 int __anon_vma_prepare(struct vm_area_struct *vma)
179 struct mm_struct *mm = vma->vm_mm;
180 struct anon_vma *anon_vma, *allocated;
181 struct anon_vma_chain *avc;
185 avc = anon_vma_chain_alloc(GFP_KERNEL);
189 anon_vma = find_mergeable_anon_vma(vma);
192 anon_vma = anon_vma_alloc();
193 if (unlikely(!anon_vma))
194 goto out_enomem_free_avc;
195 anon_vma->num_children++; /* self-parent link for new root */
196 allocated = anon_vma;
199 anon_vma_lock_write(anon_vma);
200 /* page_table_lock to protect against threads */
201 spin_lock(&mm->page_table_lock);
202 if (likely(!vma->anon_vma)) {
203 vma->anon_vma = anon_vma;
204 anon_vma_chain_link(vma, avc, anon_vma);
205 anon_vma->num_active_vmas++;
209 spin_unlock(&mm->page_table_lock);
210 anon_vma_unlock_write(anon_vma);
212 if (unlikely(allocated))
213 put_anon_vma(allocated);
215 anon_vma_chain_free(avc);
220 anon_vma_chain_free(avc);
226 * This is a useful helper function for locking the anon_vma root as
227 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
230 * Such anon_vma's should have the same root, so you'd expect to see
231 * just a single mutex_lock for the whole traversal.
233 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
235 struct anon_vma *new_root = anon_vma->root;
236 if (new_root != root) {
237 if (WARN_ON_ONCE(root))
238 up_write(&root->rwsem);
240 down_write(&root->rwsem);
245 static inline void unlock_anon_vma_root(struct anon_vma *root)
248 up_write(&root->rwsem);
252 * Attach the anon_vmas from src to dst.
253 * Returns 0 on success, -ENOMEM on failure.
255 * If dst->anon_vma is NULL this function tries to find and reuse existing
256 * anon_vma which has no vmas and only one child anon_vma. This prevents
257 * degradation of anon_vma hierarchy to endless linear chain in case of
258 * constantly forking task. On the other hand, an anon_vma with more than one
259 * child isn't reused even if there was no alive vma, thus rmap walker has a
260 * good chance of avoiding scanning the whole hierarchy when it searches where
263 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
265 struct anon_vma_chain *avc, *pavc;
266 struct anon_vma *root = NULL;
268 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
269 struct anon_vma *anon_vma;
271 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
272 if (unlikely(!avc)) {
273 unlock_anon_vma_root(root);
275 avc = anon_vma_chain_alloc(GFP_KERNEL);
279 anon_vma = pavc->anon_vma;
280 root = lock_anon_vma_root(root, anon_vma);
281 anon_vma_chain_link(dst, avc, anon_vma);
284 * Reuse existing anon_vma if it has no vma and only one
287 * Root anon_vma is never reused:
288 * it has self-parent reference and at least one child.
290 if (!dst->anon_vma &&
291 anon_vma->num_children < 2 &&
292 anon_vma->num_active_vmas == 0)
293 dst->anon_vma = anon_vma;
296 dst->anon_vma->num_active_vmas++;
297 unlock_anon_vma_root(root);
302 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
303 * decremented in unlink_anon_vmas().
304 * We can safely do this because callers of anon_vma_clone() don't care
305 * about dst->anon_vma if anon_vma_clone() failed.
307 dst->anon_vma = NULL;
308 unlink_anon_vmas(dst);
313 * Attach vma to its own anon_vma, as well as to the anon_vmas that
314 * the corresponding VMA in the parent process is attached to.
315 * Returns 0 on success, non-zero on failure.
317 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
319 struct anon_vma_chain *avc;
320 struct anon_vma *anon_vma;
323 /* Don't bother if the parent process has no anon_vma here. */
327 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
328 vma->anon_vma = NULL;
331 * First, attach the new VMA to the parent VMA's anon_vmas,
332 * so rmap can find non-COWed pages in child processes.
334 error = anon_vma_clone(vma, pvma);
338 /* An existing anon_vma has been reused, all done then. */
342 /* Then add our own anon_vma. */
343 anon_vma = anon_vma_alloc();
346 anon_vma->num_active_vmas++;
347 avc = anon_vma_chain_alloc(GFP_KERNEL);
349 goto out_error_free_anon_vma;
352 * The root anon_vma's spinlock is the lock actually used when we
353 * lock any of the anon_vmas in this anon_vma tree.
355 anon_vma->root = pvma->anon_vma->root;
356 anon_vma->parent = pvma->anon_vma;
358 * With refcounts, an anon_vma can stay around longer than the
359 * process it belongs to. The root anon_vma needs to be pinned until
360 * this anon_vma is freed, because the lock lives in the root.
362 get_anon_vma(anon_vma->root);
363 /* Mark this anon_vma as the one where our new (COWed) pages go. */
364 vma->anon_vma = anon_vma;
365 anon_vma_lock_write(anon_vma);
366 anon_vma_chain_link(vma, avc, anon_vma);
367 anon_vma->parent->num_children++;
368 anon_vma_unlock_write(anon_vma);
372 out_error_free_anon_vma:
373 put_anon_vma(anon_vma);
375 unlink_anon_vmas(vma);
379 void unlink_anon_vmas(struct vm_area_struct *vma)
381 struct anon_vma_chain *avc, *next;
382 struct anon_vma *root = NULL;
385 * Unlink each anon_vma chained to the VMA. This list is ordered
386 * from newest to oldest, ensuring the root anon_vma gets freed last.
388 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
389 struct anon_vma *anon_vma = avc->anon_vma;
391 root = lock_anon_vma_root(root, anon_vma);
392 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
395 * Leave empty anon_vmas on the list - we'll need
396 * to free them outside the lock.
398 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
399 anon_vma->parent->num_children--;
403 list_del(&avc->same_vma);
404 anon_vma_chain_free(avc);
407 vma->anon_vma->num_active_vmas--;
408 unlock_anon_vma_root(root);
411 * Iterate the list once more, it now only contains empty and unlinked
412 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
413 * needing to write-acquire the anon_vma->root->rwsem.
415 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
416 struct anon_vma *anon_vma = avc->anon_vma;
418 VM_WARN_ON(anon_vma->num_children);
419 VM_WARN_ON(anon_vma->num_active_vmas);
420 put_anon_vma(anon_vma);
422 list_del(&avc->same_vma);
423 anon_vma_chain_free(avc);
427 static void anon_vma_ctor(void *data)
429 struct anon_vma *anon_vma = data;
431 init_rwsem(&anon_vma->rwsem);
432 atomic_set(&anon_vma->refcount, 0);
433 anon_vma->rb_root = RB_ROOT_CACHED;
436 void __init anon_vma_init(void)
438 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
439 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
441 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
442 SLAB_PANIC|SLAB_ACCOUNT);
446 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
448 * Since there is no serialization what so ever against page_remove_rmap()
449 * the best this function can do is return a locked anon_vma that might
450 * have been relevant to this page.
452 * The page might have been remapped to a different anon_vma or the anon_vma
453 * returned may already be freed (and even reused).
455 * In case it was remapped to a different anon_vma, the new anon_vma will be a
456 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
457 * ensure that any anon_vma obtained from the page will still be valid for as
458 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
460 * All users of this function must be very careful when walking the anon_vma
461 * chain and verify that the page in question is indeed mapped in it
462 * [ something equivalent to page_mapped_in_vma() ].
464 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
465 * that the anon_vma pointer from page->mapping is valid if there is a
466 * mapcount, we can dereference the anon_vma after observing those.
468 struct anon_vma *page_get_anon_vma(struct page *page)
470 struct anon_vma *anon_vma = NULL;
471 unsigned long anon_mapping;
474 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
475 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
477 if (!page_mapped(page))
480 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
481 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
487 * If this page is still mapped, then its anon_vma cannot have been
488 * freed. But if it has been unmapped, we have no security against the
489 * anon_vma structure being freed and reused (for another anon_vma:
490 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
491 * above cannot corrupt).
493 if (!page_mapped(page)) {
495 put_anon_vma(anon_vma);
505 * Similar to page_get_anon_vma() except it locks the anon_vma.
507 * Its a little more complex as it tries to keep the fast path to a single
508 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
509 * reference like with page_get_anon_vma() and then block on the mutex.
511 struct anon_vma *page_lock_anon_vma_read(struct page *page)
513 struct anon_vma *anon_vma = NULL;
514 struct anon_vma *root_anon_vma;
515 unsigned long anon_mapping;
518 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
519 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
521 if (!page_mapped(page))
524 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
525 root_anon_vma = READ_ONCE(anon_vma->root);
526 if (down_read_trylock(&root_anon_vma->rwsem)) {
528 * If the page is still mapped, then this anon_vma is still
529 * its anon_vma, and holding the mutex ensures that it will
530 * not go away, see anon_vma_free().
532 if (!page_mapped(page)) {
533 up_read(&root_anon_vma->rwsem);
539 /* trylock failed, we got to sleep */
540 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
545 if (!page_mapped(page)) {
547 put_anon_vma(anon_vma);
551 /* we pinned the anon_vma, its safe to sleep */
553 anon_vma_lock_read(anon_vma);
555 if (atomic_dec_and_test(&anon_vma->refcount)) {
557 * Oops, we held the last refcount, release the lock
558 * and bail -- can't simply use put_anon_vma() because
559 * we'll deadlock on the anon_vma_lock_write() recursion.
561 anon_vma_unlock_read(anon_vma);
562 __put_anon_vma(anon_vma);
573 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
575 anon_vma_unlock_read(anon_vma);
578 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
580 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
581 * important if a PTE was dirty when it was unmapped that it's flushed
582 * before any IO is initiated on the page to prevent lost writes. Similarly,
583 * it must be flushed before freeing to prevent data leakage.
585 void try_to_unmap_flush(void)
587 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
589 if (!tlb_ubc->flush_required)
592 arch_tlbbatch_flush(&tlb_ubc->arch);
593 tlb_ubc->flush_required = false;
594 tlb_ubc->writable = false;
597 /* Flush iff there are potentially writable TLB entries that can race with IO */
598 void try_to_unmap_flush_dirty(void)
600 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
602 if (tlb_ubc->writable)
603 try_to_unmap_flush();
606 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
608 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
610 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
611 tlb_ubc->flush_required = true;
614 * Ensure compiler does not re-order the setting of tlb_flush_batched
615 * before the PTE is cleared.
618 mm->tlb_flush_batched = true;
621 * If the PTE was dirty then it's best to assume it's writable. The
622 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
623 * before the page is queued for IO.
626 tlb_ubc->writable = true;
630 * Returns true if the TLB flush should be deferred to the end of a batch of
631 * unmap operations to reduce IPIs.
633 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
635 bool should_defer = false;
637 if (!(flags & TTU_BATCH_FLUSH))
640 /* If remote CPUs need to be flushed then defer batch the flush */
641 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
649 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
650 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
651 * operation such as mprotect or munmap to race between reclaim unmapping
652 * the page and flushing the page. If this race occurs, it potentially allows
653 * access to data via a stale TLB entry. Tracking all mm's that have TLB
654 * batching in flight would be expensive during reclaim so instead track
655 * whether TLB batching occurred in the past and if so then do a flush here
656 * if required. This will cost one additional flush per reclaim cycle paid
657 * by the first operation at risk such as mprotect and mumap.
659 * This must be called under the PTL so that an access to tlb_flush_batched
660 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
663 void flush_tlb_batched_pending(struct mm_struct *mm)
665 if (mm->tlb_flush_batched) {
669 * Do not allow the compiler to re-order the clearing of
670 * tlb_flush_batched before the tlb is flushed.
673 mm->tlb_flush_batched = false;
677 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
681 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
685 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
688 * At what user virtual address is page expected in vma?
689 * Caller should check the page is actually part of the vma.
691 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
693 if (PageAnon(page)) {
694 struct anon_vma *page__anon_vma = page_anon_vma(page);
696 * Note: swapoff's unuse_vma() is more efficient with this
697 * check, and needs it to match anon_vma when KSM is active.
699 if (!vma->anon_vma || !page__anon_vma ||
700 vma->anon_vma->root != page__anon_vma->root)
702 } else if (!vma->vm_file) {
704 } else if (vma->vm_file->f_mapping != compound_head(page)->mapping) {
708 return vma_address(page, vma);
711 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
719 pgd = pgd_offset(mm, address);
720 if (!pgd_present(*pgd))
723 p4d = p4d_offset(pgd, address);
724 if (!p4d_present(*p4d))
727 pud = pud_offset(p4d, address);
728 if (!pud_present(*pud))
731 pmd = pmd_offset(pud, address);
733 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
734 * without holding anon_vma lock for write. So when looking for a
735 * genuine pmde (in which to find pte), test present and !THP together.
739 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
745 struct page_referenced_arg {
748 unsigned long vm_flags;
749 struct mem_cgroup *memcg;
752 * arg: page_referenced_arg will be passed
754 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
755 unsigned long address, void *arg)
757 struct page_referenced_arg *pra = arg;
758 struct page_vma_mapped_walk pvmw = {
765 while (page_vma_mapped_walk(&pvmw)) {
766 address = pvmw.address;
768 if (vma->vm_flags & VM_LOCKED) {
769 page_vma_mapped_walk_done(&pvmw);
770 pra->vm_flags |= VM_LOCKED;
771 return false; /* To break the loop */
775 if (ptep_clear_flush_young_notify(vma, address,
778 * Don't treat a reference through
779 * a sequentially read mapping as such.
780 * If the page has been used in another mapping,
781 * we will catch it; if this other mapping is
782 * already gone, the unmap path will have set
783 * PG_referenced or activated the page.
785 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
788 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
789 if (pmdp_clear_flush_young_notify(vma, address,
793 /* unexpected pmd-mapped page? */
801 clear_page_idle(page);
802 if (test_and_clear_page_young(page))
807 pra->vm_flags |= vma->vm_flags;
811 return false; /* To break the loop */
816 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
818 struct page_referenced_arg *pra = arg;
819 struct mem_cgroup *memcg = pra->memcg;
821 if (!mm_match_cgroup(vma->vm_mm, memcg))
828 * page_referenced - test if the page was referenced
829 * @page: the page to test
830 * @is_locked: caller holds lock on the page
831 * @memcg: target memory cgroup
832 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
834 * Quick test_and_clear_referenced for all mappings to a page,
835 * returns the number of ptes which referenced the page.
837 int page_referenced(struct page *page,
839 struct mem_cgroup *memcg,
840 unsigned long *vm_flags)
843 struct page_referenced_arg pra = {
844 .mapcount = total_mapcount(page),
847 struct rmap_walk_control rwc = {
848 .rmap_one = page_referenced_one,
850 .anon_lock = page_lock_anon_vma_read,
857 if (!page_rmapping(page))
860 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
861 we_locked = trylock_page(page);
867 * If we are reclaiming on behalf of a cgroup, skip
868 * counting on behalf of references from different
872 rwc.invalid_vma = invalid_page_referenced_vma;
875 rmap_walk(page, &rwc);
876 *vm_flags = pra.vm_flags;
881 return pra.referenced;
884 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
885 unsigned long address, void *arg)
887 struct page_vma_mapped_walk pvmw = {
893 struct mmu_notifier_range range;
897 * We have to assume the worse case ie pmd for invalidation. Note that
898 * the page can not be free from this function.
900 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
901 0, vma, vma->vm_mm, address,
902 vma_address_end(page, vma));
903 mmu_notifier_invalidate_range_start(&range);
905 while (page_vma_mapped_walk(&pvmw)) {
908 address = pvmw.address;
911 pte_t *pte = pvmw.pte;
913 if (!pte_dirty(*pte) && !pte_write(*pte))
916 flush_cache_page(vma, address, pte_pfn(*pte));
917 entry = ptep_clear_flush(vma, address, pte);
918 entry = pte_wrprotect(entry);
919 entry = pte_mkclean(entry);
920 set_pte_at(vma->vm_mm, address, pte, entry);
923 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
924 pmd_t *pmd = pvmw.pmd;
927 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
930 flush_cache_page(vma, address, page_to_pfn(page));
931 entry = pmdp_invalidate(vma, address, pmd);
932 entry = pmd_wrprotect(entry);
933 entry = pmd_mkclean(entry);
934 set_pmd_at(vma->vm_mm, address, pmd, entry);
937 /* unexpected pmd-mapped page? */
943 * No need to call mmu_notifier_invalidate_range() as we are
944 * downgrading page table protection not changing it to point
947 * See Documentation/vm/mmu_notifier.rst
953 mmu_notifier_invalidate_range_end(&range);
958 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
960 if (vma->vm_flags & VM_SHARED)
966 int page_mkclean(struct page *page)
969 struct address_space *mapping;
970 struct rmap_walk_control rwc = {
971 .arg = (void *)&cleaned,
972 .rmap_one = page_mkclean_one,
973 .invalid_vma = invalid_mkclean_vma,
976 BUG_ON(!PageLocked(page));
978 if (!page_mapped(page))
981 mapping = page_mapping(page);
985 rmap_walk(page, &rwc);
989 EXPORT_SYMBOL_GPL(page_mkclean);
992 * page_move_anon_rmap - move a page to our anon_vma
993 * @page: the page to move to our anon_vma
994 * @vma: the vma the page belongs to
996 * When a page belongs exclusively to one process after a COW event,
997 * that page can be moved into the anon_vma that belongs to just that
998 * process, so the rmap code will not search the parent or sibling
1001 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1003 struct anon_vma *anon_vma = vma->anon_vma;
1005 page = compound_head(page);
1007 VM_BUG_ON_PAGE(!PageLocked(page), page);
1008 VM_BUG_ON_VMA(!anon_vma, vma);
1010 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1012 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1013 * simultaneously, so a concurrent reader (eg page_referenced()'s
1014 * PageAnon()) will not see one without the other.
1016 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1020 * __page_set_anon_rmap - set up new anonymous rmap
1021 * @page: Page or Hugepage to add to rmap
1022 * @vma: VM area to add page to.
1023 * @address: User virtual address of the mapping
1024 * @exclusive: the page is exclusively owned by the current process
1026 static void __page_set_anon_rmap(struct page *page,
1027 struct vm_area_struct *vma, unsigned long address, int exclusive)
1029 struct anon_vma *anon_vma = vma->anon_vma;
1037 * If the page isn't exclusively mapped into this vma,
1038 * we must use the _oldest_ possible anon_vma for the
1042 anon_vma = anon_vma->root;
1044 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1045 page->mapping = (struct address_space *) anon_vma;
1046 page->index = linear_page_index(vma, address);
1050 * __page_check_anon_rmap - sanity check anonymous rmap addition
1051 * @page: the page to add the mapping to
1052 * @vma: the vm area in which the mapping is added
1053 * @address: the user virtual address mapped
1055 static void __page_check_anon_rmap(struct page *page,
1056 struct vm_area_struct *vma, unsigned long address)
1058 #ifdef CONFIG_DEBUG_VM
1060 * The page's anon-rmap details (mapping and index) are guaranteed to
1061 * be set up correctly at this point.
1063 * We have exclusion against page_add_anon_rmap because the caller
1064 * always holds the page locked, except if called from page_dup_rmap,
1065 * in which case the page is already known to be setup.
1067 * We have exclusion against page_add_new_anon_rmap because those pages
1068 * are initially only visible via the pagetables, and the pte is locked
1069 * over the call to page_add_new_anon_rmap.
1071 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1072 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1077 * page_add_anon_rmap - add pte mapping to an anonymous page
1078 * @page: the page to add the mapping to
1079 * @vma: the vm area in which the mapping is added
1080 * @address: the user virtual address mapped
1081 * @compound: charge the page as compound or small page
1083 * The caller needs to hold the pte lock, and the page must be locked in
1084 * the anon_vma case: to serialize mapping,index checking after setting,
1085 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1086 * (but PageKsm is never downgraded to PageAnon).
1088 void page_add_anon_rmap(struct page *page,
1089 struct vm_area_struct *vma, unsigned long address, bool compound)
1091 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1095 * Special version of the above for do_swap_page, which often runs
1096 * into pages that are exclusively owned by the current process.
1097 * Everybody else should continue to use page_add_anon_rmap above.
1099 void do_page_add_anon_rmap(struct page *page,
1100 struct vm_area_struct *vma, unsigned long address, int flags)
1102 bool compound = flags & RMAP_COMPOUND;
1107 VM_BUG_ON_PAGE(!PageLocked(page), page);
1108 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1109 mapcount = compound_mapcount_ptr(page);
1110 first = atomic_inc_and_test(mapcount);
1112 first = atomic_inc_and_test(&page->_mapcount);
1116 int nr = compound ? hpage_nr_pages(page) : 1;
1118 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1119 * these counters are not modified in interrupt context, and
1120 * pte lock(a spinlock) is held, which implies preemption
1124 __inc_node_page_state(page, NR_ANON_THPS);
1125 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1127 if (unlikely(PageKsm(page)))
1130 VM_BUG_ON_PAGE(!PageLocked(page), page);
1132 /* address might be in next vma when migration races vma_adjust */
1134 __page_set_anon_rmap(page, vma, address,
1135 flags & RMAP_EXCLUSIVE);
1137 __page_check_anon_rmap(page, vma, address);
1141 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1142 * @page: the page to add the mapping to
1143 * @vma: the vm area in which the mapping is added
1144 * @address: the user virtual address mapped
1145 * @compound: charge the page as compound or small page
1147 * Same as page_add_anon_rmap but must only be called on *new* pages.
1148 * This means the inc-and-test can be bypassed.
1149 * Page does not have to be locked.
1151 void page_add_new_anon_rmap(struct page *page,
1152 struct vm_area_struct *vma, unsigned long address, bool compound)
1154 int nr = compound ? hpage_nr_pages(page) : 1;
1156 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1157 __SetPageSwapBacked(page);
1159 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1160 /* increment count (starts at -1) */
1161 atomic_set(compound_mapcount_ptr(page), 0);
1162 __inc_node_page_state(page, NR_ANON_THPS);
1164 /* Anon THP always mapped first with PMD */
1165 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1166 /* increment count (starts at -1) */
1167 atomic_set(&page->_mapcount, 0);
1169 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1170 __page_set_anon_rmap(page, vma, address, 1);
1174 * page_add_file_rmap - add pte mapping to a file page
1175 * @page: the page to add the mapping to
1176 * @compound: charge the page as compound or small page
1178 * The caller needs to hold the pte lock.
1180 void page_add_file_rmap(struct page *page, bool compound)
1184 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1185 lock_page_memcg(page);
1186 if (compound && PageTransHuge(page)) {
1187 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1188 if (atomic_inc_and_test(&page[i]._mapcount))
1191 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1193 if (PageSwapBacked(page))
1194 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1196 __inc_node_page_state(page, NR_FILE_PMDMAPPED);
1198 if (PageTransCompound(page) && page_mapping(page)) {
1199 VM_WARN_ON_ONCE(!PageLocked(page));
1201 SetPageDoubleMap(compound_head(page));
1202 if (PageMlocked(page))
1203 clear_page_mlock(compound_head(page));
1205 if (!atomic_inc_and_test(&page->_mapcount))
1208 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1210 unlock_page_memcg(page);
1213 static void page_remove_file_rmap(struct page *page, bool compound)
1217 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1218 lock_page_memcg(page);
1220 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1221 if (unlikely(PageHuge(page))) {
1222 /* hugetlb pages are always mapped with pmds */
1223 atomic_dec(compound_mapcount_ptr(page));
1227 /* page still mapped by someone else? */
1228 if (compound && PageTransHuge(page)) {
1229 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1230 if (atomic_add_negative(-1, &page[i]._mapcount))
1233 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1235 if (PageSwapBacked(page))
1236 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1238 __dec_node_page_state(page, NR_FILE_PMDMAPPED);
1240 if (!atomic_add_negative(-1, &page->_mapcount))
1245 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1246 * these counters are not modified in interrupt context, and
1247 * pte lock(a spinlock) is held, which implies preemption disabled.
1249 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1251 if (unlikely(PageMlocked(page)))
1252 clear_page_mlock(page);
1254 unlock_page_memcg(page);
1257 static void page_remove_anon_compound_rmap(struct page *page)
1261 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1264 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1265 if (unlikely(PageHuge(page)))
1268 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1271 __dec_node_page_state(page, NR_ANON_THPS);
1273 if (TestClearPageDoubleMap(page)) {
1275 * Subpages can be mapped with PTEs too. Check how many of
1276 * themi are still mapped.
1278 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1279 if (atomic_add_negative(-1, &page[i]._mapcount))
1286 if (unlikely(PageMlocked(page)))
1287 clear_page_mlock(page);
1290 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1291 deferred_split_huge_page(page);
1296 * page_remove_rmap - take down pte mapping from a page
1297 * @page: page to remove mapping from
1298 * @compound: uncharge the page as compound or small page
1300 * The caller needs to hold the pte lock.
1302 void page_remove_rmap(struct page *page, bool compound)
1304 if (!PageAnon(page))
1305 return page_remove_file_rmap(page, compound);
1308 return page_remove_anon_compound_rmap(page);
1310 /* page still mapped by someone else? */
1311 if (!atomic_add_negative(-1, &page->_mapcount))
1315 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1316 * these counters are not modified in interrupt context, and
1317 * pte lock(a spinlock) is held, which implies preemption disabled.
1319 __dec_node_page_state(page, NR_ANON_MAPPED);
1321 if (unlikely(PageMlocked(page)))
1322 clear_page_mlock(page);
1324 if (PageTransCompound(page))
1325 deferred_split_huge_page(compound_head(page));
1328 * It would be tidy to reset the PageAnon mapping here,
1329 * but that might overwrite a racing page_add_anon_rmap
1330 * which increments mapcount after us but sets mapping
1331 * before us: so leave the reset to free_unref_page,
1332 * and remember that it's only reliable while mapped.
1333 * Leaving it set also helps swapoff to reinstate ptes
1334 * faster for those pages still in swapcache.
1339 * @arg: enum ttu_flags will be passed to this argument
1341 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1342 unsigned long address, void *arg)
1344 struct mm_struct *mm = vma->vm_mm;
1345 struct page_vma_mapped_walk pvmw = {
1351 struct page *subpage;
1353 struct mmu_notifier_range range;
1354 enum ttu_flags flags = (enum ttu_flags)arg;
1357 * When racing against e.g. zap_pte_range() on another cpu,
1358 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1359 * try_to_unmap() may return false when it is about to become true,
1360 * if page table locking is skipped: use TTU_SYNC to wait for that.
1362 if (flags & TTU_SYNC)
1363 pvmw.flags = PVMW_SYNC;
1365 /* munlock has nothing to gain from examining un-locked vmas */
1366 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1369 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1370 is_zone_device_page(page) && !is_device_private_page(page))
1373 if (flags & TTU_SPLIT_HUGE_PMD) {
1374 split_huge_pmd_address(vma, address,
1375 flags & TTU_SPLIT_FREEZE, page);
1379 * For THP, we have to assume the worse case ie pmd for invalidation.
1380 * For hugetlb, it could be much worse if we need to do pud
1381 * invalidation in the case of pmd sharing.
1383 * Note that the page can not be free in this function as call of
1384 * try_to_unmap() must hold a reference on the page.
1386 range.end = PageKsm(page) ?
1387 address + PAGE_SIZE : vma_address_end(page, vma);
1388 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1389 address, range.end);
1390 if (PageHuge(page)) {
1392 * If sharing is possible, start and end will be adjusted
1395 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1398 mmu_notifier_invalidate_range_start(&range);
1400 while (page_vma_mapped_walk(&pvmw)) {
1401 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1402 /* PMD-mapped THP migration entry */
1403 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1404 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1406 set_pmd_migration_entry(&pvmw, page);
1412 * If the page is mlock()d, we cannot swap it out.
1413 * If it's recently referenced (perhaps page_referenced
1414 * skipped over this mm) then we should reactivate it.
1416 if (!(flags & TTU_IGNORE_MLOCK)) {
1417 if (vma->vm_flags & VM_LOCKED) {
1418 /* PTE-mapped THP are never mlocked */
1419 if (!PageTransCompound(page)) {
1421 * Holding pte lock, we do *not* need
1424 mlock_vma_page(page);
1427 page_vma_mapped_walk_done(&pvmw);
1430 if (flags & TTU_MUNLOCK)
1434 /* Unexpected PMD-mapped THP? */
1435 VM_BUG_ON_PAGE(!pvmw.pte, page);
1437 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1438 address = pvmw.address;
1440 if (PageHuge(page)) {
1441 if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1443 * huge_pmd_unshare unmapped an entire PMD
1444 * page. There is no way of knowing exactly
1445 * which PMDs may be cached for this mm, so
1446 * we must flush them all. start/end were
1447 * already adjusted above to cover this range.
1449 flush_cache_range(vma, range.start, range.end);
1450 flush_tlb_range(vma, range.start, range.end);
1451 mmu_notifier_invalidate_range(mm, range.start,
1455 * The ref count of the PMD page was dropped
1456 * which is part of the way map counting
1457 * is done for shared PMDs. Return 'true'
1458 * here. When there is no other sharing,
1459 * huge_pmd_unshare returns false and we will
1460 * unmap the actual page and drop map count
1463 page_vma_mapped_walk_done(&pvmw);
1468 if (IS_ENABLED(CONFIG_MIGRATION) &&
1469 (flags & TTU_MIGRATION) &&
1470 is_zone_device_page(page)) {
1474 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1477 * Store the pfn of the page in a special migration
1478 * pte. do_swap_page() will wait until the migration
1479 * pte is removed and then restart fault handling.
1481 entry = make_migration_entry(page, 0);
1482 swp_pte = swp_entry_to_pte(entry);
1483 if (pte_soft_dirty(pteval))
1484 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1485 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1487 * No need to invalidate here it will synchronize on
1488 * against the special swap migration pte.
1490 * The assignment to subpage above was computed from a
1491 * swap PTE which results in an invalid pointer.
1492 * Since only PAGE_SIZE pages can currently be
1493 * migrated, just set it to page. This will need to be
1494 * changed when hugepage migrations to device private
1495 * memory are supported.
1501 if (!(flags & TTU_IGNORE_ACCESS)) {
1502 if (ptep_clear_flush_young_notify(vma, address,
1505 page_vma_mapped_walk_done(&pvmw);
1510 /* Nuke the page table entry. */
1511 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1512 if (should_defer_flush(mm, flags)) {
1514 * We clear the PTE but do not flush so potentially
1515 * a remote CPU could still be writing to the page.
1516 * If the entry was previously clean then the
1517 * architecture must guarantee that a clear->dirty
1518 * transition on a cached TLB entry is written through
1519 * and traps if the PTE is unmapped.
1521 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1523 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1525 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1528 /* Move the dirty bit to the page. Now the pte is gone. */
1529 if (pte_dirty(pteval))
1530 set_page_dirty(page);
1532 /* Update high watermark before we lower rss */
1533 update_hiwater_rss(mm);
1535 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1536 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1537 if (PageHuge(page)) {
1538 hugetlb_count_sub(compound_nr(page), mm);
1539 set_huge_swap_pte_at(mm, address,
1541 vma_mmu_pagesize(vma));
1543 dec_mm_counter(mm, mm_counter(page));
1544 set_pte_at(mm, address, pvmw.pte, pteval);
1547 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1549 * The guest indicated that the page content is of no
1550 * interest anymore. Simply discard the pte, vmscan
1551 * will take care of the rest.
1552 * A future reference will then fault in a new zero
1553 * page. When userfaultfd is active, we must not drop
1554 * this page though, as its main user (postcopy
1555 * migration) will not expect userfaults on already
1558 dec_mm_counter(mm, mm_counter(page));
1559 /* We have to invalidate as we cleared the pte */
1560 mmu_notifier_invalidate_range(mm, address,
1561 address + PAGE_SIZE);
1562 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1563 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1567 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1568 set_pte_at(mm, address, pvmw.pte, pteval);
1570 page_vma_mapped_walk_done(&pvmw);
1575 * Store the pfn of the page in a special migration
1576 * pte. do_swap_page() will wait until the migration
1577 * pte is removed and then restart fault handling.
1579 entry = make_migration_entry(subpage,
1581 swp_pte = swp_entry_to_pte(entry);
1582 if (pte_soft_dirty(pteval))
1583 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1584 set_pte_at(mm, address, pvmw.pte, swp_pte);
1586 * No need to invalidate here it will synchronize on
1587 * against the special swap migration pte.
1589 } else if (PageAnon(page)) {
1590 swp_entry_t entry = { .val = page_private(subpage) };
1593 * Store the swap location in the pte.
1594 * See handle_pte_fault() ...
1596 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1599 /* We have to invalidate as we cleared the pte */
1600 mmu_notifier_invalidate_range(mm, address,
1601 address + PAGE_SIZE);
1602 page_vma_mapped_walk_done(&pvmw);
1606 /* MADV_FREE page check */
1607 if (!PageSwapBacked(page)) {
1608 int ref_count, map_count;
1611 * Synchronize with gup_pte_range():
1612 * - clear PTE; barrier; read refcount
1613 * - inc refcount; barrier; read PTE
1617 ref_count = page_ref_count(page);
1618 map_count = page_mapcount(page);
1621 * Order reads for page refcount and dirty flag
1622 * (see comments in __remove_mapping()).
1627 * The only page refs must be one from isolation
1628 * plus the rmap(s) (dropped by discard:).
1630 if (ref_count == 1 + map_count &&
1632 /* Invalidate as we cleared the pte */
1633 mmu_notifier_invalidate_range(mm,
1634 address, address + PAGE_SIZE);
1635 dec_mm_counter(mm, MM_ANONPAGES);
1640 * If the page was redirtied, it cannot be
1641 * discarded. Remap the page to page table.
1643 set_pte_at(mm, address, pvmw.pte, pteval);
1644 SetPageSwapBacked(page);
1646 page_vma_mapped_walk_done(&pvmw);
1650 if (swap_duplicate(entry) < 0) {
1651 set_pte_at(mm, address, pvmw.pte, pteval);
1653 page_vma_mapped_walk_done(&pvmw);
1656 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1657 set_pte_at(mm, address, pvmw.pte, pteval);
1659 page_vma_mapped_walk_done(&pvmw);
1662 if (list_empty(&mm->mmlist)) {
1663 spin_lock(&mmlist_lock);
1664 if (list_empty(&mm->mmlist))
1665 list_add(&mm->mmlist, &init_mm.mmlist);
1666 spin_unlock(&mmlist_lock);
1668 dec_mm_counter(mm, MM_ANONPAGES);
1669 inc_mm_counter(mm, MM_SWAPENTS);
1670 swp_pte = swp_entry_to_pte(entry);
1671 if (pte_soft_dirty(pteval))
1672 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1673 set_pte_at(mm, address, pvmw.pte, swp_pte);
1674 /* Invalidate as we cleared the pte */
1675 mmu_notifier_invalidate_range(mm, address,
1676 address + PAGE_SIZE);
1679 * This is a locked file-backed page, thus it cannot
1680 * be removed from the page cache and replaced by a new
1681 * page before mmu_notifier_invalidate_range_end, so no
1682 * concurrent thread might update its page table to
1683 * point at new page while a device still is using this
1686 * See Documentation/vm/mmu_notifier.rst
1688 dec_mm_counter(mm, mm_counter_file(page));
1692 * No need to call mmu_notifier_invalidate_range() it has be
1693 * done above for all cases requiring it to happen under page
1694 * table lock before mmu_notifier_invalidate_range_end()
1696 * See Documentation/vm/mmu_notifier.rst
1698 page_remove_rmap(subpage, PageHuge(page));
1702 mmu_notifier_invalidate_range_end(&range);
1707 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1709 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1714 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1715 VM_STACK_INCOMPLETE_SETUP)
1721 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1723 return is_vma_temporary_stack(vma);
1726 static int page_not_mapped(struct page *page)
1728 return !page_mapped(page);
1732 * try_to_unmap - try to remove all page table mappings to a page
1733 * @page: the page to get unmapped
1734 * @flags: action and flags
1736 * Tries to remove all the page table entries which are mapping this
1737 * page, used in the pageout path. Caller must hold the page lock.
1739 * If unmap is successful, return true. Otherwise, false.
1741 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1743 struct rmap_walk_control rwc = {
1744 .rmap_one = try_to_unmap_one,
1745 .arg = (void *)flags,
1746 .done = page_not_mapped,
1747 .anon_lock = page_lock_anon_vma_read,
1751 * During exec, a temporary VMA is setup and later moved.
1752 * The VMA is moved under the anon_vma lock but not the
1753 * page tables leading to a race where migration cannot
1754 * find the migration ptes. Rather than increasing the
1755 * locking requirements of exec(), migration skips
1756 * temporary VMAs until after exec() completes.
1758 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1759 && !PageKsm(page) && PageAnon(page))
1760 rwc.invalid_vma = invalid_migration_vma;
1762 if (flags & TTU_RMAP_LOCKED)
1763 rmap_walk_locked(page, &rwc);
1765 rmap_walk(page, &rwc);
1768 * When racing against e.g. zap_pte_range() on another cpu,
1769 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1770 * try_to_unmap() may return false when it is about to become true,
1771 * if page table locking is skipped: use TTU_SYNC to wait for that.
1773 return !page_mapcount(page);
1777 * try_to_munlock - try to munlock a page
1778 * @page: the page to be munlocked
1780 * Called from munlock code. Checks all of the VMAs mapping the page
1781 * to make sure nobody else has this page mlocked. The page will be
1782 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1785 void try_to_munlock(struct page *page)
1787 struct rmap_walk_control rwc = {
1788 .rmap_one = try_to_unmap_one,
1789 .arg = (void *)TTU_MUNLOCK,
1790 .done = page_not_mapped,
1791 .anon_lock = page_lock_anon_vma_read,
1795 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1796 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1798 rmap_walk(page, &rwc);
1801 void __put_anon_vma(struct anon_vma *anon_vma)
1803 struct anon_vma *root = anon_vma->root;
1805 anon_vma_free(anon_vma);
1806 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1807 anon_vma_free(root);
1810 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1811 struct rmap_walk_control *rwc)
1813 struct anon_vma *anon_vma;
1816 return rwc->anon_lock(page);
1819 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1820 * because that depends on page_mapped(); but not all its usages
1821 * are holding mmap_sem. Users without mmap_sem are required to
1822 * take a reference count to prevent the anon_vma disappearing
1824 anon_vma = page_anon_vma(page);
1828 anon_vma_lock_read(anon_vma);
1833 * rmap_walk_anon - do something to anonymous page using the object-based
1835 * @page: the page to be handled
1836 * @rwc: control variable according to each walk type
1838 * Find all the mappings of a page using the mapping pointer and the vma chains
1839 * contained in the anon_vma struct it points to.
1841 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1842 * where the page was found will be held for write. So, we won't recheck
1843 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1846 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1849 struct anon_vma *anon_vma;
1850 pgoff_t pgoff_start, pgoff_end;
1851 struct anon_vma_chain *avc;
1854 anon_vma = page_anon_vma(page);
1855 /* anon_vma disappear under us? */
1856 VM_BUG_ON_PAGE(!anon_vma, page);
1858 anon_vma = rmap_walk_anon_lock(page, rwc);
1863 pgoff_start = page_to_pgoff(page);
1864 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1865 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1866 pgoff_start, pgoff_end) {
1867 struct vm_area_struct *vma = avc->vma;
1868 unsigned long address = vma_address(page, vma);
1870 VM_BUG_ON_VMA(address == -EFAULT, vma);
1873 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1876 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1878 if (rwc->done && rwc->done(page))
1883 anon_vma_unlock_read(anon_vma);
1887 * rmap_walk_file - do something to file page using the object-based rmap method
1888 * @page: the page to be handled
1889 * @rwc: control variable according to each walk type
1891 * Find all the mappings of a page using the mapping pointer and the vma chains
1892 * contained in the address_space struct it points to.
1894 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1895 * where the page was found will be held for write. So, we won't recheck
1896 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1899 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1902 struct address_space *mapping = page_mapping(page);
1903 pgoff_t pgoff_start, pgoff_end;
1904 struct vm_area_struct *vma;
1907 * The page lock not only makes sure that page->mapping cannot
1908 * suddenly be NULLified by truncation, it makes sure that the
1909 * structure at mapping cannot be freed and reused yet,
1910 * so we can safely take mapping->i_mmap_rwsem.
1912 VM_BUG_ON_PAGE(!PageLocked(page), page);
1917 pgoff_start = page_to_pgoff(page);
1918 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1920 i_mmap_lock_read(mapping);
1921 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1922 pgoff_start, pgoff_end) {
1923 unsigned long address = vma_address(page, vma);
1925 VM_BUG_ON_VMA(address == -EFAULT, vma);
1928 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1931 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1933 if (rwc->done && rwc->done(page))
1939 i_mmap_unlock_read(mapping);
1942 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1944 if (unlikely(PageKsm(page)))
1945 rmap_walk_ksm(page, rwc);
1946 else if (PageAnon(page))
1947 rmap_walk_anon(page, rwc, false);
1949 rmap_walk_file(page, rwc, false);
1952 /* Like rmap_walk, but caller holds relevant rmap lock */
1953 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1955 /* no ksm support for now */
1956 VM_BUG_ON_PAGE(PageKsm(page), page);
1958 rmap_walk_anon(page, rwc, true);
1960 rmap_walk_file(page, rwc, true);
1963 #ifdef CONFIG_HUGETLB_PAGE
1965 * The following two functions are for anonymous (private mapped) hugepages.
1966 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1967 * and no lru code, because we handle hugepages differently from common pages.
1969 void hugepage_add_anon_rmap(struct page *page,
1970 struct vm_area_struct *vma, unsigned long address)
1972 struct anon_vma *anon_vma = vma->anon_vma;
1975 BUG_ON(!PageLocked(page));
1977 /* address might be in next vma when migration races vma_adjust */
1978 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1980 __page_set_anon_rmap(page, vma, address, 0);
1983 void hugepage_add_new_anon_rmap(struct page *page,
1984 struct vm_area_struct *vma, unsigned long address)
1986 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1987 atomic_set(compound_mapcount_ptr(page), 0);
1988 __page_set_anon_rmap(page, vma, address, 1);
1990 #endif /* CONFIG_HUGETLB_PAGE */