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) * (see huegtlbfs below)
26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
28 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
30 * mm->page_table_lock or pte_lock
31 * pgdat->lru_lock (in mark_page_accessed, isolate_lru_page)
32 * swap_lock (in swap_duplicate, swap_info_get)
33 * mmlist_lock (in mmput, drain_mmlist and others)
34 * mapping->private_lock (in __set_page_dirty_buffers)
35 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
36 * i_pages lock (widely used)
37 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39 * sb_lock (within inode_lock in fs/fs-writeback.c)
40 * i_pages lock (widely used, in set_page_dirty,
41 * in arch-dependent flush_dcache_mmap_lock,
42 * within bdi.wb->list_lock in __sync_single_inode)
44 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
48 * * hugetlbfs PageHuge() pages take locks in this order:
49 * mapping->i_mmap_rwsem
50 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * page->flags PG_locked (lock_page)
55 #include <linux/sched/mm.h>
56 #include <linux/sched/task.h>
57 #include <linux/pagemap.h>
58 #include <linux/swap.h>
59 #include <linux/swapops.h>
60 #include <linux/slab.h>
61 #include <linux/init.h>
62 #include <linux/ksm.h>
63 #include <linux/rmap.h>
64 #include <linux/rcupdate.h>
65 #include <linux/export.h>
66 #include <linux/memcontrol.h>
67 #include <linux/mmu_notifier.h>
68 #include <linux/migrate.h>
69 #include <linux/hugetlb.h>
70 #include <linux/huge_mm.h>
71 #include <linux/backing-dev.h>
72 #include <linux/page_idle.h>
73 #include <linux/memremap.h>
74 #include <linux/userfaultfd_k.h>
76 #include <asm/tlbflush.h>
78 #include <trace/events/tlb.h>
82 static struct kmem_cache *anon_vma_cachep;
83 static struct kmem_cache *anon_vma_chain_cachep;
85 static inline struct anon_vma *anon_vma_alloc(void)
87 struct anon_vma *anon_vma;
89 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
91 atomic_set(&anon_vma->refcount, 1);
92 anon_vma->num_children = 0;
93 anon_vma->num_active_vmas = 0;
94 anon_vma->parent = anon_vma;
96 * Initialise the anon_vma root to point to itself. If called
97 * from fork, the root will be reset to the parents anon_vma.
99 anon_vma->root = anon_vma;
105 static inline void anon_vma_free(struct anon_vma *anon_vma)
107 VM_BUG_ON(atomic_read(&anon_vma->refcount));
110 * Synchronize against page_lock_anon_vma_read() such that
111 * we can safely hold the lock without the anon_vma getting
114 * Relies on the full mb implied by the atomic_dec_and_test() from
115 * put_anon_vma() against the acquire barrier implied by
116 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
118 * page_lock_anon_vma_read() VS put_anon_vma()
119 * down_read_trylock() atomic_dec_and_test()
121 * atomic_read() rwsem_is_locked()
123 * LOCK should suffice since the actual taking of the lock must
124 * happen _before_ what follows.
127 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
128 anon_vma_lock_write(anon_vma);
129 anon_vma_unlock_write(anon_vma);
132 kmem_cache_free(anon_vma_cachep, anon_vma);
135 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
137 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
140 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
142 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
145 static void anon_vma_chain_link(struct vm_area_struct *vma,
146 struct anon_vma_chain *avc,
147 struct anon_vma *anon_vma)
150 avc->anon_vma = anon_vma;
151 list_add(&avc->same_vma, &vma->anon_vma_chain);
152 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
156 * __anon_vma_prepare - attach an anon_vma to a memory region
157 * @vma: the memory region in question
159 * This makes sure the memory mapping described by 'vma' has
160 * an 'anon_vma' attached to it, so that we can associate the
161 * anonymous pages mapped into it with that anon_vma.
163 * The common case will be that we already have one, which
164 * is handled inline by anon_vma_prepare(). But if
165 * not we either need to find an adjacent mapping that we
166 * can re-use the anon_vma from (very common when the only
167 * reason for splitting a vma has been mprotect()), or we
168 * allocate a new one.
170 * Anon-vma allocations are very subtle, because we may have
171 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
172 * and that may actually touch the spinlock even in the newly
173 * allocated vma (it depends on RCU to make sure that the
174 * anon_vma isn't actually destroyed).
176 * As a result, we need to do proper anon_vma locking even
177 * for the new allocation. At the same time, we do not want
178 * to do any locking for the common case of already having
181 * This must be called with the mmap_lock held for reading.
183 int __anon_vma_prepare(struct vm_area_struct *vma)
185 struct mm_struct *mm = vma->vm_mm;
186 struct anon_vma *anon_vma, *allocated;
187 struct anon_vma_chain *avc;
191 avc = anon_vma_chain_alloc(GFP_KERNEL);
195 anon_vma = find_mergeable_anon_vma(vma);
198 anon_vma = anon_vma_alloc();
199 if (unlikely(!anon_vma))
200 goto out_enomem_free_avc;
201 anon_vma->num_children++; /* self-parent link for new root */
202 allocated = anon_vma;
205 anon_vma_lock_write(anon_vma);
206 /* page_table_lock to protect against threads */
207 spin_lock(&mm->page_table_lock);
208 if (likely(!vma->anon_vma)) {
209 vma->anon_vma = anon_vma;
210 anon_vma_chain_link(vma, avc, anon_vma);
211 anon_vma->num_active_vmas++;
215 spin_unlock(&mm->page_table_lock);
216 anon_vma_unlock_write(anon_vma);
218 if (unlikely(allocated))
219 put_anon_vma(allocated);
221 anon_vma_chain_free(avc);
226 anon_vma_chain_free(avc);
232 * This is a useful helper function for locking the anon_vma root as
233 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
236 * Such anon_vma's should have the same root, so you'd expect to see
237 * just a single mutex_lock for the whole traversal.
239 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
241 struct anon_vma *new_root = anon_vma->root;
242 if (new_root != root) {
243 if (WARN_ON_ONCE(root))
244 up_write(&root->rwsem);
246 down_write(&root->rwsem);
251 static inline void unlock_anon_vma_root(struct anon_vma *root)
254 up_write(&root->rwsem);
258 * Attach the anon_vmas from src to dst.
259 * Returns 0 on success, -ENOMEM on failure.
261 * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and
262 * anon_vma_fork(). The first three want an exact copy of src, while the last
263 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
264 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
265 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
267 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
268 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
269 * This prevents degradation of anon_vma hierarchy to endless linear chain in
270 * case of constantly forking task. On the other hand, an anon_vma with more
271 * than one child isn't reused even if there was no alive vma, thus rmap
272 * walker has a good chance of avoiding scanning the whole hierarchy when it
273 * searches where page is mapped.
275 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
277 struct anon_vma_chain *avc, *pavc;
278 struct anon_vma *root = NULL;
280 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
281 struct anon_vma *anon_vma;
283 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
284 if (unlikely(!avc)) {
285 unlock_anon_vma_root(root);
287 avc = anon_vma_chain_alloc(GFP_KERNEL);
291 anon_vma = pavc->anon_vma;
292 root = lock_anon_vma_root(root, anon_vma);
293 anon_vma_chain_link(dst, avc, anon_vma);
296 * Reuse existing anon_vma if it has no vma and only one
299 * Root anon_vma is never reused:
300 * it has self-parent reference and at least one child.
302 if (!dst->anon_vma && src->anon_vma &&
303 anon_vma->num_children < 2 &&
304 anon_vma->num_active_vmas == 0)
305 dst->anon_vma = anon_vma;
308 dst->anon_vma->num_active_vmas++;
309 unlock_anon_vma_root(root);
314 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
315 * decremented in unlink_anon_vmas().
316 * We can safely do this because callers of anon_vma_clone() don't care
317 * about dst->anon_vma if anon_vma_clone() failed.
319 dst->anon_vma = NULL;
320 unlink_anon_vmas(dst);
325 * Attach vma to its own anon_vma, as well as to the anon_vmas that
326 * the corresponding VMA in the parent process is attached to.
327 * Returns 0 on success, non-zero on failure.
329 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
331 struct anon_vma_chain *avc;
332 struct anon_vma *anon_vma;
335 /* Don't bother if the parent process has no anon_vma here. */
339 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
340 vma->anon_vma = NULL;
343 * First, attach the new VMA to the parent VMA's anon_vmas,
344 * so rmap can find non-COWed pages in child processes.
346 error = anon_vma_clone(vma, pvma);
350 /* An existing anon_vma has been reused, all done then. */
354 /* Then add our own anon_vma. */
355 anon_vma = anon_vma_alloc();
358 anon_vma->num_active_vmas++;
359 avc = anon_vma_chain_alloc(GFP_KERNEL);
361 goto out_error_free_anon_vma;
364 * The root anon_vma's spinlock is the lock actually used when we
365 * lock any of the anon_vmas in this anon_vma tree.
367 anon_vma->root = pvma->anon_vma->root;
368 anon_vma->parent = pvma->anon_vma;
370 * With refcounts, an anon_vma can stay around longer than the
371 * process it belongs to. The root anon_vma needs to be pinned until
372 * this anon_vma is freed, because the lock lives in the root.
374 get_anon_vma(anon_vma->root);
375 /* Mark this anon_vma as the one where our new (COWed) pages go. */
376 vma->anon_vma = anon_vma;
377 anon_vma_lock_write(anon_vma);
378 anon_vma_chain_link(vma, avc, anon_vma);
379 anon_vma->parent->num_children++;
380 anon_vma_unlock_write(anon_vma);
384 out_error_free_anon_vma:
385 put_anon_vma(anon_vma);
387 unlink_anon_vmas(vma);
391 void unlink_anon_vmas(struct vm_area_struct *vma)
393 struct anon_vma_chain *avc, *next;
394 struct anon_vma *root = NULL;
397 * Unlink each anon_vma chained to the VMA. This list is ordered
398 * from newest to oldest, ensuring the root anon_vma gets freed last.
400 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
401 struct anon_vma *anon_vma = avc->anon_vma;
403 root = lock_anon_vma_root(root, anon_vma);
404 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
407 * Leave empty anon_vmas on the list - we'll need
408 * to free them outside the lock.
410 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
411 anon_vma->parent->num_children--;
415 list_del(&avc->same_vma);
416 anon_vma_chain_free(avc);
419 vma->anon_vma->num_active_vmas--;
420 unlock_anon_vma_root(root);
423 * Iterate the list once more, it now only contains empty and unlinked
424 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
425 * needing to write-acquire the anon_vma->root->rwsem.
427 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
428 struct anon_vma *anon_vma = avc->anon_vma;
430 VM_WARN_ON(anon_vma->num_children);
431 VM_WARN_ON(anon_vma->num_active_vmas);
432 put_anon_vma(anon_vma);
434 list_del(&avc->same_vma);
435 anon_vma_chain_free(avc);
439 static void anon_vma_ctor(void *data)
441 struct anon_vma *anon_vma = data;
443 init_rwsem(&anon_vma->rwsem);
444 atomic_set(&anon_vma->refcount, 0);
445 anon_vma->rb_root = RB_ROOT_CACHED;
448 void __init anon_vma_init(void)
450 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
451 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
453 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
454 SLAB_PANIC|SLAB_ACCOUNT);
458 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
460 * Since there is no serialization what so ever against page_remove_rmap()
461 * the best this function can do is return a locked anon_vma that might
462 * have been relevant to this page.
464 * The page might have been remapped to a different anon_vma or the anon_vma
465 * returned may already be freed (and even reused).
467 * In case it was remapped to a different anon_vma, the new anon_vma will be a
468 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
469 * ensure that any anon_vma obtained from the page will still be valid for as
470 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
472 * All users of this function must be very careful when walking the anon_vma
473 * chain and verify that the page in question is indeed mapped in it
474 * [ something equivalent to page_mapped_in_vma() ].
476 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
477 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
478 * if there is a mapcount, we can dereference the anon_vma after observing
481 struct anon_vma *page_get_anon_vma(struct page *page)
483 struct anon_vma *anon_vma = NULL;
484 unsigned long anon_mapping;
487 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
488 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
490 if (!page_mapped(page))
493 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
494 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
500 * If this page is still mapped, then its anon_vma cannot have been
501 * freed. But if it has been unmapped, we have no security against the
502 * anon_vma structure being freed and reused (for another anon_vma:
503 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
504 * above cannot corrupt).
506 if (!page_mapped(page)) {
508 put_anon_vma(anon_vma);
518 * Similar to page_get_anon_vma() except it locks the anon_vma.
520 * Its a little more complex as it tries to keep the fast path to a single
521 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
522 * reference like with page_get_anon_vma() and then block on the mutex.
524 struct anon_vma *page_lock_anon_vma_read(struct page *page)
526 struct anon_vma *anon_vma = NULL;
527 struct anon_vma *root_anon_vma;
528 unsigned long anon_mapping;
531 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
532 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
534 if (!page_mapped(page))
537 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
538 root_anon_vma = READ_ONCE(anon_vma->root);
539 if (down_read_trylock(&root_anon_vma->rwsem)) {
541 * If the page is still mapped, then this anon_vma is still
542 * its anon_vma, and holding the mutex ensures that it will
543 * not go away, see anon_vma_free().
545 if (!page_mapped(page)) {
546 up_read(&root_anon_vma->rwsem);
552 /* trylock failed, we got to sleep */
553 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
558 if (!page_mapped(page)) {
560 put_anon_vma(anon_vma);
564 /* we pinned the anon_vma, its safe to sleep */
566 anon_vma_lock_read(anon_vma);
568 if (atomic_dec_and_test(&anon_vma->refcount)) {
570 * Oops, we held the last refcount, release the lock
571 * and bail -- can't simply use put_anon_vma() because
572 * we'll deadlock on the anon_vma_lock_write() recursion.
574 anon_vma_unlock_read(anon_vma);
575 __put_anon_vma(anon_vma);
586 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
588 anon_vma_unlock_read(anon_vma);
591 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
593 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
594 * important if a PTE was dirty when it was unmapped that it's flushed
595 * before any IO is initiated on the page to prevent lost writes. Similarly,
596 * it must be flushed before freeing to prevent data leakage.
598 void try_to_unmap_flush(void)
600 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
602 if (!tlb_ubc->flush_required)
605 arch_tlbbatch_flush(&tlb_ubc->arch);
606 tlb_ubc->flush_required = false;
607 tlb_ubc->writable = false;
610 /* Flush iff there are potentially writable TLB entries that can race with IO */
611 void try_to_unmap_flush_dirty(void)
613 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
615 if (tlb_ubc->writable)
616 try_to_unmap_flush();
619 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
621 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
623 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
624 tlb_ubc->flush_required = true;
627 * Ensure compiler does not re-order the setting of tlb_flush_batched
628 * before the PTE is cleared.
631 mm->tlb_flush_batched = true;
634 * If the PTE was dirty then it's best to assume it's writable. The
635 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
636 * before the page is queued for IO.
639 tlb_ubc->writable = true;
643 * Returns true if the TLB flush should be deferred to the end of a batch of
644 * unmap operations to reduce IPIs.
646 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
648 bool should_defer = false;
650 if (!(flags & TTU_BATCH_FLUSH))
653 /* If remote CPUs need to be flushed then defer batch the flush */
654 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
662 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
663 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
664 * operation such as mprotect or munmap to race between reclaim unmapping
665 * the page and flushing the page. If this race occurs, it potentially allows
666 * access to data via a stale TLB entry. Tracking all mm's that have TLB
667 * batching in flight would be expensive during reclaim so instead track
668 * whether TLB batching occurred in the past and if so then do a flush here
669 * if required. This will cost one additional flush per reclaim cycle paid
670 * by the first operation at risk such as mprotect and mumap.
672 * This must be called under the PTL so that an access to tlb_flush_batched
673 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
676 void flush_tlb_batched_pending(struct mm_struct *mm)
678 if (data_race(mm->tlb_flush_batched)) {
682 * Do not allow the compiler to re-order the clearing of
683 * tlb_flush_batched before the tlb is flushed.
686 mm->tlb_flush_batched = false;
690 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
694 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
698 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
701 * At what user virtual address is page expected in vma?
702 * Caller should check the page is actually part of the vma.
704 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
706 if (PageAnon(page)) {
707 struct anon_vma *page__anon_vma = page_anon_vma(page);
709 * Note: swapoff's unuse_vma() is more efficient with this
710 * check, and needs it to match anon_vma when KSM is active.
712 if (!vma->anon_vma || !page__anon_vma ||
713 vma->anon_vma->root != page__anon_vma->root)
715 } else if (!vma->vm_file) {
717 } else if (vma->vm_file->f_mapping != compound_head(page)->mapping) {
721 return vma_address(page, vma);
724 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
732 pgd = pgd_offset(mm, address);
733 if (!pgd_present(*pgd))
736 p4d = p4d_offset(pgd, address);
737 if (!p4d_present(*p4d))
740 pud = pud_offset(p4d, address);
741 if (!pud_present(*pud))
744 pmd = pmd_offset(pud, address);
746 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
747 * without holding anon_vma lock for write. So when looking for a
748 * genuine pmde (in which to find pte), test present and !THP together.
752 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
758 struct page_referenced_arg {
761 unsigned long vm_flags;
762 struct mem_cgroup *memcg;
765 * arg: page_referenced_arg will be passed
767 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
768 unsigned long address, void *arg)
770 struct page_referenced_arg *pra = arg;
771 struct page_vma_mapped_walk pvmw = {
778 while (page_vma_mapped_walk(&pvmw)) {
779 address = pvmw.address;
781 if (vma->vm_flags & VM_LOCKED) {
782 page_vma_mapped_walk_done(&pvmw);
783 pra->vm_flags |= VM_LOCKED;
784 return false; /* To break the loop */
788 if (ptep_clear_flush_young_notify(vma, address,
791 * Don't treat a reference through
792 * a sequentially read mapping as such.
793 * If the page has been used in another mapping,
794 * we will catch it; if this other mapping is
795 * already gone, the unmap path will have set
796 * PG_referenced or activated the page.
798 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
801 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
802 if (pmdp_clear_flush_young_notify(vma, address,
806 /* unexpected pmd-mapped page? */
814 clear_page_idle(page);
815 if (test_and_clear_page_young(page))
820 pra->vm_flags |= vma->vm_flags;
824 return false; /* To break the loop */
829 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
831 struct page_referenced_arg *pra = arg;
832 struct mem_cgroup *memcg = pra->memcg;
834 if (!mm_match_cgroup(vma->vm_mm, memcg))
841 * page_referenced - test if the page was referenced
842 * @page: the page to test
843 * @is_locked: caller holds lock on the page
844 * @memcg: target memory cgroup
845 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
847 * Quick test_and_clear_referenced for all mappings to a page,
848 * returns the number of ptes which referenced the page.
850 int page_referenced(struct page *page,
852 struct mem_cgroup *memcg,
853 unsigned long *vm_flags)
856 struct page_referenced_arg pra = {
857 .mapcount = total_mapcount(page),
860 struct rmap_walk_control rwc = {
861 .rmap_one = page_referenced_one,
863 .anon_lock = page_lock_anon_vma_read,
870 if (!page_rmapping(page))
873 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
874 we_locked = trylock_page(page);
880 * If we are reclaiming on behalf of a cgroup, skip
881 * counting on behalf of references from different
885 rwc.invalid_vma = invalid_page_referenced_vma;
888 rmap_walk(page, &rwc);
889 *vm_flags = pra.vm_flags;
894 return pra.referenced;
897 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
898 unsigned long address, void *arg)
900 struct page_vma_mapped_walk pvmw = {
906 struct mmu_notifier_range range;
910 * We have to assume the worse case ie pmd for invalidation. Note that
911 * the page can not be free from this function.
913 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
914 0, vma, vma->vm_mm, address,
915 vma_address_end(page, vma));
916 mmu_notifier_invalidate_range_start(&range);
918 while (page_vma_mapped_walk(&pvmw)) {
921 address = pvmw.address;
924 pte_t *pte = pvmw.pte;
926 if (!pte_dirty(*pte) && !pte_write(*pte))
929 flush_cache_page(vma, address, pte_pfn(*pte));
930 entry = ptep_clear_flush(vma, address, pte);
931 entry = pte_wrprotect(entry);
932 entry = pte_mkclean(entry);
933 set_pte_at(vma->vm_mm, address, pte, entry);
936 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
937 pmd_t *pmd = pvmw.pmd;
940 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
943 flush_cache_page(vma, address, page_to_pfn(page));
944 entry = pmdp_invalidate(vma, address, pmd);
945 entry = pmd_wrprotect(entry);
946 entry = pmd_mkclean(entry);
947 set_pmd_at(vma->vm_mm, address, pmd, entry);
950 /* unexpected pmd-mapped page? */
956 * No need to call mmu_notifier_invalidate_range() as we are
957 * downgrading page table protection not changing it to point
960 * See Documentation/vm/mmu_notifier.rst
966 mmu_notifier_invalidate_range_end(&range);
971 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
973 if (vma->vm_flags & VM_SHARED)
979 int page_mkclean(struct page *page)
982 struct address_space *mapping;
983 struct rmap_walk_control rwc = {
984 .arg = (void *)&cleaned,
985 .rmap_one = page_mkclean_one,
986 .invalid_vma = invalid_mkclean_vma,
989 BUG_ON(!PageLocked(page));
991 if (!page_mapped(page))
994 mapping = page_mapping(page);
998 rmap_walk(page, &rwc);
1002 EXPORT_SYMBOL_GPL(page_mkclean);
1005 * page_move_anon_rmap - move a page to our anon_vma
1006 * @page: the page to move to our anon_vma
1007 * @vma: the vma the page belongs to
1009 * When a page belongs exclusively to one process after a COW event,
1010 * that page can be moved into the anon_vma that belongs to just that
1011 * process, so the rmap code will not search the parent or sibling
1014 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1016 struct anon_vma *anon_vma = vma->anon_vma;
1018 page = compound_head(page);
1020 VM_BUG_ON_PAGE(!PageLocked(page), page);
1021 VM_BUG_ON_VMA(!anon_vma, vma);
1023 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1025 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1026 * simultaneously, so a concurrent reader (eg page_referenced()'s
1027 * PageAnon()) will not see one without the other.
1029 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1033 * __page_set_anon_rmap - set up new anonymous rmap
1034 * @page: Page or Hugepage to add to rmap
1035 * @vma: VM area to add page to.
1036 * @address: User virtual address of the mapping
1037 * @exclusive: the page is exclusively owned by the current process
1039 static void __page_set_anon_rmap(struct page *page,
1040 struct vm_area_struct *vma, unsigned long address, int exclusive)
1042 struct anon_vma *anon_vma = vma->anon_vma;
1050 * If the page isn't exclusively mapped into this vma,
1051 * we must use the _oldest_ possible anon_vma for the
1055 anon_vma = anon_vma->root;
1057 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1058 page->mapping = (struct address_space *) anon_vma;
1059 page->index = linear_page_index(vma, address);
1063 * __page_check_anon_rmap - sanity check anonymous rmap addition
1064 * @page: the page to add the mapping to
1065 * @vma: the vm area in which the mapping is added
1066 * @address: the user virtual address mapped
1068 static void __page_check_anon_rmap(struct page *page,
1069 struct vm_area_struct *vma, unsigned long address)
1072 * The page's anon-rmap details (mapping and index) are guaranteed to
1073 * be set up correctly at this point.
1075 * We have exclusion against page_add_anon_rmap because the caller
1076 * always holds the page locked, except if called from page_dup_rmap,
1077 * in which case the page is already known to be setup.
1079 * We have exclusion against page_add_new_anon_rmap because those pages
1080 * are initially only visible via the pagetables, and the pte is locked
1081 * over the call to page_add_new_anon_rmap.
1083 VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page);
1084 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1089 * page_add_anon_rmap - add pte mapping to an anonymous page
1090 * @page: the page to add the mapping to
1091 * @vma: the vm area in which the mapping is added
1092 * @address: the user virtual address mapped
1093 * @compound: charge the page as compound or small page
1095 * The caller needs to hold the pte lock, and the page must be locked in
1096 * the anon_vma case: to serialize mapping,index checking after setting,
1097 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1098 * (but PageKsm is never downgraded to PageAnon).
1100 void page_add_anon_rmap(struct page *page,
1101 struct vm_area_struct *vma, unsigned long address, bool compound)
1103 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1107 * Special version of the above for do_swap_page, which often runs
1108 * into pages that are exclusively owned by the current process.
1109 * Everybody else should continue to use page_add_anon_rmap above.
1111 void do_page_add_anon_rmap(struct page *page,
1112 struct vm_area_struct *vma, unsigned long address, int flags)
1114 bool compound = flags & RMAP_COMPOUND;
1117 if (unlikely(PageKsm(page)))
1118 lock_page_memcg(page);
1120 VM_BUG_ON_PAGE(!PageLocked(page), page);
1124 VM_BUG_ON_PAGE(!PageLocked(page), page);
1125 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1126 mapcount = compound_mapcount_ptr(page);
1127 first = atomic_inc_and_test(mapcount);
1129 first = atomic_inc_and_test(&page->_mapcount);
1133 int nr = compound ? thp_nr_pages(page) : 1;
1135 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1136 * these counters are not modified in interrupt context, and
1137 * pte lock(a spinlock) is held, which implies preemption
1141 __inc_lruvec_page_state(page, NR_ANON_THPS);
1142 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1145 if (unlikely(PageKsm(page))) {
1146 unlock_page_memcg(page);
1150 /* address might be in next vma when migration races vma_adjust */
1152 __page_set_anon_rmap(page, vma, address,
1153 flags & RMAP_EXCLUSIVE);
1155 __page_check_anon_rmap(page, vma, address);
1159 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1160 * @page: the page to add the mapping to
1161 * @vma: the vm area in which the mapping is added
1162 * @address: the user virtual address mapped
1163 * @compound: charge the page as compound or small page
1165 * Same as page_add_anon_rmap but must only be called on *new* pages.
1166 * This means the inc-and-test can be bypassed.
1167 * Page does not have to be locked.
1169 void page_add_new_anon_rmap(struct page *page,
1170 struct vm_area_struct *vma, unsigned long address, bool compound)
1172 int nr = compound ? thp_nr_pages(page) : 1;
1174 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1175 __SetPageSwapBacked(page);
1177 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1178 /* increment count (starts at -1) */
1179 atomic_set(compound_mapcount_ptr(page), 0);
1180 if (hpage_pincount_available(page))
1181 atomic_set(compound_pincount_ptr(page), 0);
1183 __inc_lruvec_page_state(page, NR_ANON_THPS);
1185 /* Anon THP always mapped first with PMD */
1186 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1187 /* increment count (starts at -1) */
1188 atomic_set(&page->_mapcount, 0);
1190 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1191 __page_set_anon_rmap(page, vma, address, 1);
1195 * page_add_file_rmap - add pte mapping to a file page
1196 * @page: the page to add the mapping to
1197 * @compound: charge the page as compound or small page
1199 * The caller needs to hold the pte lock.
1201 void page_add_file_rmap(struct page *page, bool compound)
1205 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1206 lock_page_memcg(page);
1207 if (compound && PageTransHuge(page)) {
1208 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1209 if (atomic_inc_and_test(&page[i]._mapcount))
1212 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1214 if (PageSwapBacked(page))
1215 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1217 __inc_node_page_state(page, NR_FILE_PMDMAPPED);
1219 if (PageTransCompound(page) && page_mapping(page)) {
1220 VM_WARN_ON_ONCE(!PageLocked(page));
1222 SetPageDoubleMap(compound_head(page));
1223 if (PageMlocked(page))
1224 clear_page_mlock(compound_head(page));
1226 if (!atomic_inc_and_test(&page->_mapcount))
1229 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1231 unlock_page_memcg(page);
1234 static void page_remove_file_rmap(struct page *page, bool compound)
1238 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1240 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1241 if (unlikely(PageHuge(page))) {
1242 /* hugetlb pages are always mapped with pmds */
1243 atomic_dec(compound_mapcount_ptr(page));
1247 /* page still mapped by someone else? */
1248 if (compound && PageTransHuge(page)) {
1249 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1250 if (atomic_add_negative(-1, &page[i]._mapcount))
1253 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1255 if (PageSwapBacked(page))
1256 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1258 __dec_node_page_state(page, NR_FILE_PMDMAPPED);
1260 if (!atomic_add_negative(-1, &page->_mapcount))
1265 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1266 * these counters are not modified in interrupt context, and
1267 * pte lock(a spinlock) is held, which implies preemption disabled.
1269 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1271 if (unlikely(PageMlocked(page)))
1272 clear_page_mlock(page);
1275 static void page_remove_anon_compound_rmap(struct page *page)
1279 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1282 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1283 if (unlikely(PageHuge(page)))
1286 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1289 __dec_lruvec_page_state(page, NR_ANON_THPS);
1291 if (TestClearPageDoubleMap(page)) {
1293 * Subpages can be mapped with PTEs too. Check how many of
1294 * them are still mapped.
1296 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1297 if (atomic_add_negative(-1, &page[i]._mapcount))
1302 * Queue the page for deferred split if at least one small
1303 * page of the compound page is unmapped, but at least one
1304 * small page is still mapped.
1306 if (nr && nr < thp_nr_pages(page))
1307 deferred_split_huge_page(page);
1309 nr = thp_nr_pages(page);
1312 if (unlikely(PageMlocked(page)))
1313 clear_page_mlock(page);
1316 __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
1320 * page_remove_rmap - take down pte mapping from a page
1321 * @page: page to remove mapping from
1322 * @compound: uncharge the page as compound or small page
1324 * The caller needs to hold the pte lock.
1326 void page_remove_rmap(struct page *page, bool compound)
1328 lock_page_memcg(page);
1330 if (!PageAnon(page)) {
1331 page_remove_file_rmap(page, compound);
1336 page_remove_anon_compound_rmap(page);
1340 /* page still mapped by someone else? */
1341 if (!atomic_add_negative(-1, &page->_mapcount))
1345 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1346 * these counters are not modified in interrupt context, and
1347 * pte lock(a spinlock) is held, which implies preemption disabled.
1349 __dec_lruvec_page_state(page, NR_ANON_MAPPED);
1351 if (unlikely(PageMlocked(page)))
1352 clear_page_mlock(page);
1354 if (PageTransCompound(page))
1355 deferred_split_huge_page(compound_head(page));
1358 * It would be tidy to reset the PageAnon mapping here,
1359 * but that might overwrite a racing page_add_anon_rmap
1360 * which increments mapcount after us but sets mapping
1361 * before us: so leave the reset to free_unref_page,
1362 * and remember that it's only reliable while mapped.
1363 * Leaving it set also helps swapoff to reinstate ptes
1364 * faster for those pages still in swapcache.
1367 unlock_page_memcg(page);
1371 * @arg: enum ttu_flags will be passed to this argument
1373 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1374 unsigned long address, void *arg)
1376 struct mm_struct *mm = vma->vm_mm;
1377 struct page_vma_mapped_walk pvmw = {
1383 struct page *subpage;
1385 struct mmu_notifier_range range;
1386 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1389 * When racing against e.g. zap_pte_range() on another cpu,
1390 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1391 * try_to_unmap() may return false when it is about to become true,
1392 * if page table locking is skipped: use TTU_SYNC to wait for that.
1394 if (flags & TTU_SYNC)
1395 pvmw.flags = PVMW_SYNC;
1397 /* munlock has nothing to gain from examining un-locked vmas */
1398 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1401 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1402 is_zone_device_page(page) && !is_device_private_page(page))
1405 if (flags & TTU_SPLIT_HUGE_PMD) {
1406 split_huge_pmd_address(vma, address,
1407 flags & TTU_SPLIT_FREEZE, page);
1411 * For THP, we have to assume the worse case ie pmd for invalidation.
1412 * For hugetlb, it could be much worse if we need to do pud
1413 * invalidation in the case of pmd sharing.
1415 * Note that the page can not be free in this function as call of
1416 * try_to_unmap() must hold a reference on the page.
1418 range.end = PageKsm(page) ?
1419 address + PAGE_SIZE : vma_address_end(page, vma);
1420 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1421 address, range.end);
1422 if (PageHuge(page)) {
1424 * If sharing is possible, start and end will be adjusted
1427 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1430 mmu_notifier_invalidate_range_start(&range);
1432 while (page_vma_mapped_walk(&pvmw)) {
1433 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1434 /* PMD-mapped THP migration entry */
1435 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1436 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1438 set_pmd_migration_entry(&pvmw, page);
1444 * If the page is mlock()d, we cannot swap it out.
1445 * If it's recently referenced (perhaps page_referenced
1446 * skipped over this mm) then we should reactivate it.
1448 if (!(flags & TTU_IGNORE_MLOCK)) {
1449 if (vma->vm_flags & VM_LOCKED) {
1450 /* PTE-mapped THP are never mlocked */
1451 if (!PageTransCompound(page)) {
1453 * Holding pte lock, we do *not* need
1456 mlock_vma_page(page);
1459 page_vma_mapped_walk_done(&pvmw);
1462 if (flags & TTU_MUNLOCK)
1466 /* Unexpected PMD-mapped THP? */
1467 VM_BUG_ON_PAGE(!pvmw.pte, page);
1469 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1470 address = pvmw.address;
1472 if (PageHuge(page) && !PageAnon(page)) {
1474 * To call huge_pmd_unshare, i_mmap_rwsem must be
1475 * held in write mode. Caller needs to explicitly
1476 * do this outside rmap routines.
1478 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1479 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1481 * huge_pmd_unshare unmapped an entire PMD
1482 * page. There is no way of knowing exactly
1483 * which PMDs may be cached for this mm, so
1484 * we must flush them all. start/end were
1485 * already adjusted above to cover this range.
1487 flush_cache_range(vma, range.start, range.end);
1488 flush_tlb_range(vma, range.start, range.end);
1489 mmu_notifier_invalidate_range(mm, range.start,
1493 * The ref count of the PMD page was dropped
1494 * which is part of the way map counting
1495 * is done for shared PMDs. Return 'true'
1496 * here. When there is no other sharing,
1497 * huge_pmd_unshare returns false and we will
1498 * unmap the actual page and drop map count
1501 page_vma_mapped_walk_done(&pvmw);
1506 if (IS_ENABLED(CONFIG_MIGRATION) &&
1507 (flags & TTU_MIGRATION) &&
1508 is_zone_device_page(page)) {
1512 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1515 * Store the pfn of the page in a special migration
1516 * pte. do_swap_page() will wait until the migration
1517 * pte is removed and then restart fault handling.
1519 entry = make_migration_entry(page, 0);
1520 swp_pte = swp_entry_to_pte(entry);
1523 * pteval maps a zone device page and is therefore
1526 if (pte_swp_soft_dirty(pteval))
1527 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1528 if (pte_swp_uffd_wp(pteval))
1529 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1530 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1532 * No need to invalidate here it will synchronize on
1533 * against the special swap migration pte.
1535 * The assignment to subpage above was computed from a
1536 * swap PTE which results in an invalid pointer.
1537 * Since only PAGE_SIZE pages can currently be
1538 * migrated, just set it to page. This will need to be
1539 * changed when hugepage migrations to device private
1540 * memory are supported.
1546 /* Nuke the page table entry. */
1547 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1548 if (should_defer_flush(mm, flags)) {
1550 * We clear the PTE but do not flush so potentially
1551 * a remote CPU could still be writing to the page.
1552 * If the entry was previously clean then the
1553 * architecture must guarantee that a clear->dirty
1554 * transition on a cached TLB entry is written through
1555 * and traps if the PTE is unmapped.
1557 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1559 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1561 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1564 /* Move the dirty bit to the page. Now the pte is gone. */
1565 if (pte_dirty(pteval))
1566 set_page_dirty(page);
1568 /* Update high watermark before we lower rss */
1569 update_hiwater_rss(mm);
1571 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1572 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1573 if (PageHuge(page)) {
1574 hugetlb_count_sub(compound_nr(page), mm);
1575 set_huge_swap_pte_at(mm, address,
1577 vma_mmu_pagesize(vma));
1579 dec_mm_counter(mm, mm_counter(page));
1580 set_pte_at(mm, address, pvmw.pte, pteval);
1583 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1585 * The guest indicated that the page content is of no
1586 * interest anymore. Simply discard the pte, vmscan
1587 * will take care of the rest.
1588 * A future reference will then fault in a new zero
1589 * page. When userfaultfd is active, we must not drop
1590 * this page though, as its main user (postcopy
1591 * migration) will not expect userfaults on already
1594 dec_mm_counter(mm, mm_counter(page));
1595 /* We have to invalidate as we cleared the pte */
1596 mmu_notifier_invalidate_range(mm, address,
1597 address + PAGE_SIZE);
1598 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1599 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1603 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1604 set_pte_at(mm, address, pvmw.pte, pteval);
1606 page_vma_mapped_walk_done(&pvmw);
1611 * Store the pfn of the page in a special migration
1612 * pte. do_swap_page() will wait until the migration
1613 * pte is removed and then restart fault handling.
1615 entry = make_migration_entry(subpage,
1617 swp_pte = swp_entry_to_pte(entry);
1618 if (pte_soft_dirty(pteval))
1619 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1620 if (pte_uffd_wp(pteval))
1621 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1622 set_pte_at(mm, address, pvmw.pte, swp_pte);
1624 * No need to invalidate here it will synchronize on
1625 * against the special swap migration pte.
1627 } else if (PageAnon(page)) {
1628 swp_entry_t entry = { .val = page_private(subpage) };
1631 * Store the swap location in the pte.
1632 * See handle_pte_fault() ...
1634 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1637 /* We have to invalidate as we cleared the pte */
1638 mmu_notifier_invalidate_range(mm, address,
1639 address + PAGE_SIZE);
1640 page_vma_mapped_walk_done(&pvmw);
1644 /* MADV_FREE page check */
1645 if (!PageSwapBacked(page)) {
1646 int ref_count, map_count;
1649 * Synchronize with gup_pte_range():
1650 * - clear PTE; barrier; read refcount
1651 * - inc refcount; barrier; read PTE
1655 ref_count = page_ref_count(page);
1656 map_count = page_mapcount(page);
1659 * Order reads for page refcount and dirty flag
1660 * (see comments in __remove_mapping()).
1665 * The only page refs must be one from isolation
1666 * plus the rmap(s) (dropped by discard:).
1668 if (ref_count == 1 + map_count &&
1670 /* Invalidate as we cleared the pte */
1671 mmu_notifier_invalidate_range(mm,
1672 address, address + PAGE_SIZE);
1673 dec_mm_counter(mm, MM_ANONPAGES);
1678 * If the page was redirtied, it cannot be
1679 * discarded. Remap the page to page table.
1681 set_pte_at(mm, address, pvmw.pte, pteval);
1682 SetPageSwapBacked(page);
1684 page_vma_mapped_walk_done(&pvmw);
1688 if (swap_duplicate(entry) < 0) {
1689 set_pte_at(mm, address, pvmw.pte, pteval);
1691 page_vma_mapped_walk_done(&pvmw);
1694 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1695 set_pte_at(mm, address, pvmw.pte, pteval);
1697 page_vma_mapped_walk_done(&pvmw);
1700 if (list_empty(&mm->mmlist)) {
1701 spin_lock(&mmlist_lock);
1702 if (list_empty(&mm->mmlist))
1703 list_add(&mm->mmlist, &init_mm.mmlist);
1704 spin_unlock(&mmlist_lock);
1706 dec_mm_counter(mm, MM_ANONPAGES);
1707 inc_mm_counter(mm, MM_SWAPENTS);
1708 swp_pte = swp_entry_to_pte(entry);
1709 if (pte_soft_dirty(pteval))
1710 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1711 if (pte_uffd_wp(pteval))
1712 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1713 set_pte_at(mm, address, pvmw.pte, swp_pte);
1714 /* Invalidate as we cleared the pte */
1715 mmu_notifier_invalidate_range(mm, address,
1716 address + PAGE_SIZE);
1719 * This is a locked file-backed page, thus it cannot
1720 * be removed from the page cache and replaced by a new
1721 * page before mmu_notifier_invalidate_range_end, so no
1722 * concurrent thread might update its page table to
1723 * point at new page while a device still is using this
1726 * See Documentation/vm/mmu_notifier.rst
1728 dec_mm_counter(mm, mm_counter_file(page));
1732 * No need to call mmu_notifier_invalidate_range() it has be
1733 * done above for all cases requiring it to happen under page
1734 * table lock before mmu_notifier_invalidate_range_end()
1736 * See Documentation/vm/mmu_notifier.rst
1738 page_remove_rmap(subpage, PageHuge(page));
1742 mmu_notifier_invalidate_range_end(&range);
1747 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1749 return vma_is_temporary_stack(vma);
1752 static int page_not_mapped(struct page *page)
1754 return !page_mapped(page);
1758 * try_to_unmap - try to remove all page table mappings to a page
1759 * @page: the page to get unmapped
1760 * @flags: action and flags
1762 * Tries to remove all the page table entries which are mapping this
1763 * page, used in the pageout path. Caller must hold the page lock.
1765 * If unmap is successful, return true. Otherwise, false.
1767 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1769 struct rmap_walk_control rwc = {
1770 .rmap_one = try_to_unmap_one,
1771 .arg = (void *)flags,
1772 .done = page_not_mapped,
1773 .anon_lock = page_lock_anon_vma_read,
1777 * During exec, a temporary VMA is setup and later moved.
1778 * The VMA is moved under the anon_vma lock but not the
1779 * page tables leading to a race where migration cannot
1780 * find the migration ptes. Rather than increasing the
1781 * locking requirements of exec(), migration skips
1782 * temporary VMAs until after exec() completes.
1784 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1785 && !PageKsm(page) && PageAnon(page))
1786 rwc.invalid_vma = invalid_migration_vma;
1788 if (flags & TTU_RMAP_LOCKED)
1789 rmap_walk_locked(page, &rwc);
1791 rmap_walk(page, &rwc);
1794 * When racing against e.g. zap_pte_range() on another cpu,
1795 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1796 * try_to_unmap() may return false when it is about to become true,
1797 * if page table locking is skipped: use TTU_SYNC to wait for that.
1799 return !page_mapcount(page);
1803 * try_to_munlock - try to munlock a page
1804 * @page: the page to be munlocked
1806 * Called from munlock code. Checks all of the VMAs mapping the page
1807 * to make sure nobody else has this page mlocked. The page will be
1808 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1811 void try_to_munlock(struct page *page)
1813 struct rmap_walk_control rwc = {
1814 .rmap_one = try_to_unmap_one,
1815 .arg = (void *)TTU_MUNLOCK,
1816 .done = page_not_mapped,
1817 .anon_lock = page_lock_anon_vma_read,
1821 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1822 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1824 rmap_walk(page, &rwc);
1827 void __put_anon_vma(struct anon_vma *anon_vma)
1829 struct anon_vma *root = anon_vma->root;
1831 anon_vma_free(anon_vma);
1832 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1833 anon_vma_free(root);
1836 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1837 struct rmap_walk_control *rwc)
1839 struct anon_vma *anon_vma;
1842 return rwc->anon_lock(page);
1845 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1846 * because that depends on page_mapped(); but not all its usages
1847 * are holding mmap_lock. Users without mmap_lock are required to
1848 * take a reference count to prevent the anon_vma disappearing
1850 anon_vma = page_anon_vma(page);
1854 anon_vma_lock_read(anon_vma);
1859 * rmap_walk_anon - do something to anonymous page using the object-based
1861 * @page: the page to be handled
1862 * @rwc: control variable according to each walk type
1864 * Find all the mappings of a page using the mapping pointer and the vma chains
1865 * contained in the anon_vma struct it points to.
1867 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1868 * where the page was found will be held for write. So, we won't recheck
1869 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1872 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1875 struct anon_vma *anon_vma;
1876 pgoff_t pgoff_start, pgoff_end;
1877 struct anon_vma_chain *avc;
1880 anon_vma = page_anon_vma(page);
1881 /* anon_vma disappear under us? */
1882 VM_BUG_ON_PAGE(!anon_vma, page);
1884 anon_vma = rmap_walk_anon_lock(page, rwc);
1889 pgoff_start = page_to_pgoff(page);
1890 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
1891 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1892 pgoff_start, pgoff_end) {
1893 struct vm_area_struct *vma = avc->vma;
1894 unsigned long address = vma_address(page, vma);
1896 VM_BUG_ON_VMA(address == -EFAULT, vma);
1899 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1902 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1904 if (rwc->done && rwc->done(page))
1909 anon_vma_unlock_read(anon_vma);
1913 * rmap_walk_file - do something to file page using the object-based rmap method
1914 * @page: the page to be handled
1915 * @rwc: control variable according to each walk type
1917 * Find all the mappings of a page using the mapping pointer and the vma chains
1918 * contained in the address_space struct it points to.
1920 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1921 * where the page was found will be held for write. So, we won't recheck
1922 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1925 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1928 struct address_space *mapping = page_mapping(page);
1929 pgoff_t pgoff_start, pgoff_end;
1930 struct vm_area_struct *vma;
1933 * The page lock not only makes sure that page->mapping cannot
1934 * suddenly be NULLified by truncation, it makes sure that the
1935 * structure at mapping cannot be freed and reused yet,
1936 * so we can safely take mapping->i_mmap_rwsem.
1938 VM_BUG_ON_PAGE(!PageLocked(page), page);
1943 pgoff_start = page_to_pgoff(page);
1944 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
1946 i_mmap_lock_read(mapping);
1947 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1948 pgoff_start, pgoff_end) {
1949 unsigned long address = vma_address(page, vma);
1951 VM_BUG_ON_VMA(address == -EFAULT, vma);
1954 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1957 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1959 if (rwc->done && rwc->done(page))
1965 i_mmap_unlock_read(mapping);
1968 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1970 if (unlikely(PageKsm(page)))
1971 rmap_walk_ksm(page, rwc);
1972 else if (PageAnon(page))
1973 rmap_walk_anon(page, rwc, false);
1975 rmap_walk_file(page, rwc, false);
1978 /* Like rmap_walk, but caller holds relevant rmap lock */
1979 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1981 /* no ksm support for now */
1982 VM_BUG_ON_PAGE(PageKsm(page), page);
1984 rmap_walk_anon(page, rwc, true);
1986 rmap_walk_file(page, rwc, true);
1989 #ifdef CONFIG_HUGETLB_PAGE
1991 * The following two functions are for anonymous (private mapped) hugepages.
1992 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1993 * and no lru code, because we handle hugepages differently from common pages.
1995 void hugepage_add_anon_rmap(struct page *page,
1996 struct vm_area_struct *vma, unsigned long address)
1998 struct anon_vma *anon_vma = vma->anon_vma;
2001 BUG_ON(!PageLocked(page));
2003 /* address might be in next vma when migration races vma_adjust */
2004 first = atomic_inc_and_test(compound_mapcount_ptr(page));
2006 __page_set_anon_rmap(page, vma, address, 0);
2009 void hugepage_add_new_anon_rmap(struct page *page,
2010 struct vm_area_struct *vma, unsigned long address)
2012 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2013 atomic_set(compound_mapcount_ptr(page), 0);
2014 if (hpage_pincount_available(page))
2015 atomic_set(compound_pincount_ptr(page), 0);
2017 __page_set_anon_rmap(page, vma, address, 1);
2019 #endif /* CONFIG_HUGETLB_PAGE */