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->degree = 1; /* Reference for first vma */
87 anon_vma->parent = anon_vma;
89 * Initialise the anon_vma root to point to itself. If called
90 * from fork, the root will be reset to the parents anon_vma.
92 anon_vma->root = anon_vma;
98 static inline void anon_vma_free(struct anon_vma *anon_vma)
100 VM_BUG_ON(atomic_read(&anon_vma->refcount));
103 * Synchronize against page_lock_anon_vma_read() such that
104 * we can safely hold the lock without the anon_vma getting
107 * Relies on the full mb implied by the atomic_dec_and_test() from
108 * put_anon_vma() against the acquire barrier implied by
109 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
111 * page_lock_anon_vma_read() VS put_anon_vma()
112 * down_read_trylock() atomic_dec_and_test()
114 * atomic_read() rwsem_is_locked()
116 * LOCK should suffice since the actual taking of the lock must
117 * happen _before_ what follows.
120 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
121 anon_vma_lock_write(anon_vma);
122 anon_vma_unlock_write(anon_vma);
125 kmem_cache_free(anon_vma_cachep, anon_vma);
128 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
130 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
133 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
135 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
138 static void anon_vma_chain_link(struct vm_area_struct *vma,
139 struct anon_vma_chain *avc,
140 struct anon_vma *anon_vma)
143 avc->anon_vma = anon_vma;
144 list_add(&avc->same_vma, &vma->anon_vma_chain);
145 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
149 * __anon_vma_prepare - attach an anon_vma to a memory region
150 * @vma: the memory region in question
152 * This makes sure the memory mapping described by 'vma' has
153 * an 'anon_vma' attached to it, so that we can associate the
154 * anonymous pages mapped into it with that anon_vma.
156 * The common case will be that we already have one, which
157 * is handled inline by anon_vma_prepare(). But if
158 * not we either need to find an adjacent mapping that we
159 * can re-use the anon_vma from (very common when the only
160 * reason for splitting a vma has been mprotect()), or we
161 * allocate a new one.
163 * Anon-vma allocations are very subtle, because we may have
164 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
165 * and that may actually touch the spinlock even in the newly
166 * allocated vma (it depends on RCU to make sure that the
167 * anon_vma isn't actually destroyed).
169 * As a result, we need to do proper anon_vma locking even
170 * for the new allocation. At the same time, we do not want
171 * to do any locking for the common case of already having
174 * This must be called with the mmap_sem held for reading.
176 int __anon_vma_prepare(struct vm_area_struct *vma)
178 struct mm_struct *mm = vma->vm_mm;
179 struct anon_vma *anon_vma, *allocated;
180 struct anon_vma_chain *avc;
184 avc = anon_vma_chain_alloc(GFP_KERNEL);
188 anon_vma = find_mergeable_anon_vma(vma);
191 anon_vma = anon_vma_alloc();
192 if (unlikely(!anon_vma))
193 goto out_enomem_free_avc;
194 allocated = anon_vma;
197 anon_vma_lock_write(anon_vma);
198 /* page_table_lock to protect against threads */
199 spin_lock(&mm->page_table_lock);
200 if (likely(!vma->anon_vma)) {
201 vma->anon_vma = anon_vma;
202 anon_vma_chain_link(vma, avc, anon_vma);
203 /* vma reference or self-parent link for new root */
208 spin_unlock(&mm->page_table_lock);
209 anon_vma_unlock_write(anon_vma);
211 if (unlikely(allocated))
212 put_anon_vma(allocated);
214 anon_vma_chain_free(avc);
219 anon_vma_chain_free(avc);
225 * This is a useful helper function for locking the anon_vma root as
226 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
229 * Such anon_vma's should have the same root, so you'd expect to see
230 * just a single mutex_lock for the whole traversal.
232 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
234 struct anon_vma *new_root = anon_vma->root;
235 if (new_root != root) {
236 if (WARN_ON_ONCE(root))
237 up_write(&root->rwsem);
239 down_write(&root->rwsem);
244 static inline void unlock_anon_vma_root(struct anon_vma *root)
247 up_write(&root->rwsem);
251 * Attach the anon_vmas from src to dst.
252 * Returns 0 on success, -ENOMEM on failure.
254 * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and
255 * anon_vma_fork(). The first three want an exact copy of src, while the last
256 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
257 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
258 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
260 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
261 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
262 * This prevents degradation of anon_vma hierarchy to endless linear chain in
263 * case of constantly forking task. On the other hand, an anon_vma with more
264 * than one child isn't reused even if there was no alive vma, thus rmap
265 * walker has a good chance of avoiding scanning the whole hierarchy when it
266 * searches where page is mapped.
268 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
270 struct anon_vma_chain *avc, *pavc;
271 struct anon_vma *root = NULL;
272 struct vm_area_struct *prev = dst->vm_prev, *pprev = src->vm_prev;
275 * If parent share anon_vma with its vm_prev, keep this sharing in in
278 * 1. Parent has vm_prev, which implies we have vm_prev.
279 * 2. Parent and its vm_prev have the same anon_vma.
281 if (!dst->anon_vma && src->anon_vma &&
282 pprev && pprev->anon_vma == src->anon_vma)
283 dst->anon_vma = prev->anon_vma;
286 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
287 struct anon_vma *anon_vma;
289 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
290 if (unlikely(!avc)) {
291 unlock_anon_vma_root(root);
293 avc = anon_vma_chain_alloc(GFP_KERNEL);
297 anon_vma = pavc->anon_vma;
298 root = lock_anon_vma_root(root, anon_vma);
299 anon_vma_chain_link(dst, avc, anon_vma);
302 * Reuse existing anon_vma if its degree lower than two,
303 * that means it has no vma and only one anon_vma child.
305 * Do not chose parent anon_vma, otherwise first child
306 * will always reuse it. Root anon_vma is never reused:
307 * it has self-parent reference and at least one child.
309 if (!dst->anon_vma && src->anon_vma &&
310 anon_vma != src->anon_vma && anon_vma->degree < 2)
311 dst->anon_vma = anon_vma;
314 dst->anon_vma->degree++;
315 unlock_anon_vma_root(root);
320 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
321 * decremented in unlink_anon_vmas().
322 * We can safely do this because callers of anon_vma_clone() don't care
323 * about dst->anon_vma if anon_vma_clone() failed.
325 dst->anon_vma = NULL;
326 unlink_anon_vmas(dst);
331 * Attach vma to its own anon_vma, as well as to the anon_vmas that
332 * the corresponding VMA in the parent process is attached to.
333 * Returns 0 on success, non-zero on failure.
335 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
337 struct anon_vma_chain *avc;
338 struct anon_vma *anon_vma;
341 /* Don't bother if the parent process has no anon_vma here. */
345 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
346 vma->anon_vma = NULL;
349 * First, attach the new VMA to the parent VMA's anon_vmas,
350 * so rmap can find non-COWed pages in child processes.
352 error = anon_vma_clone(vma, pvma);
356 /* An existing anon_vma has been reused, all done then. */
360 /* Then add our own anon_vma. */
361 anon_vma = anon_vma_alloc();
364 avc = anon_vma_chain_alloc(GFP_KERNEL);
366 goto out_error_free_anon_vma;
369 * The root anon_vma's spinlock is the lock actually used when we
370 * lock any of the anon_vmas in this anon_vma tree.
372 anon_vma->root = pvma->anon_vma->root;
373 anon_vma->parent = pvma->anon_vma;
375 * With refcounts, an anon_vma can stay around longer than the
376 * process it belongs to. The root anon_vma needs to be pinned until
377 * this anon_vma is freed, because the lock lives in the root.
379 get_anon_vma(anon_vma->root);
380 /* Mark this anon_vma as the one where our new (COWed) pages go. */
381 vma->anon_vma = anon_vma;
382 anon_vma_lock_write(anon_vma);
383 anon_vma_chain_link(vma, avc, anon_vma);
384 anon_vma->parent->degree++;
385 anon_vma_unlock_write(anon_vma);
389 out_error_free_anon_vma:
390 put_anon_vma(anon_vma);
392 unlink_anon_vmas(vma);
396 void unlink_anon_vmas(struct vm_area_struct *vma)
398 struct anon_vma_chain *avc, *next;
399 struct anon_vma *root = NULL;
402 * Unlink each anon_vma chained to the VMA. This list is ordered
403 * from newest to oldest, ensuring the root anon_vma gets freed last.
405 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
406 struct anon_vma *anon_vma = avc->anon_vma;
408 root = lock_anon_vma_root(root, anon_vma);
409 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
412 * Leave empty anon_vmas on the list - we'll need
413 * to free them outside the lock.
415 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
416 anon_vma->parent->degree--;
420 list_del(&avc->same_vma);
421 anon_vma_chain_free(avc);
424 vma->anon_vma->degree--;
425 unlock_anon_vma_root(root);
428 * Iterate the list once more, it now only contains empty and unlinked
429 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
430 * needing to write-acquire the anon_vma->root->rwsem.
432 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
433 struct anon_vma *anon_vma = avc->anon_vma;
435 VM_WARN_ON(anon_vma->degree);
436 put_anon_vma(anon_vma);
438 list_del(&avc->same_vma);
439 anon_vma_chain_free(avc);
443 static void anon_vma_ctor(void *data)
445 struct anon_vma *anon_vma = data;
447 init_rwsem(&anon_vma->rwsem);
448 atomic_set(&anon_vma->refcount, 0);
449 anon_vma->rb_root = RB_ROOT_CACHED;
452 void __init anon_vma_init(void)
454 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
455 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
457 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
458 SLAB_PANIC|SLAB_ACCOUNT);
462 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
464 * Since there is no serialization what so ever against page_remove_rmap()
465 * the best this function can do is return a locked anon_vma that might
466 * have been relevant to this page.
468 * The page might have been remapped to a different anon_vma or the anon_vma
469 * returned may already be freed (and even reused).
471 * In case it was remapped to a different anon_vma, the new anon_vma will be a
472 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
473 * ensure that any anon_vma obtained from the page will still be valid for as
474 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
476 * All users of this function must be very careful when walking the anon_vma
477 * chain and verify that the page in question is indeed mapped in it
478 * [ something equivalent to page_mapped_in_vma() ].
480 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
481 * that the anon_vma pointer from page->mapping is valid if there is a
482 * mapcount, we can dereference the anon_vma after observing those.
484 struct anon_vma *page_get_anon_vma(struct page *page)
486 struct anon_vma *anon_vma = NULL;
487 unsigned long anon_mapping;
490 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
491 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
493 if (!page_mapped(page))
496 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
497 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
503 * If this page is still mapped, then its anon_vma cannot have been
504 * freed. But if it has been unmapped, we have no security against the
505 * anon_vma structure being freed and reused (for another anon_vma:
506 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
507 * above cannot corrupt).
509 if (!page_mapped(page)) {
511 put_anon_vma(anon_vma);
521 * Similar to page_get_anon_vma() except it locks the anon_vma.
523 * Its a little more complex as it tries to keep the fast path to a single
524 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
525 * reference like with page_get_anon_vma() and then block on the mutex.
527 struct anon_vma *page_lock_anon_vma_read(struct page *page)
529 struct anon_vma *anon_vma = NULL;
530 struct anon_vma *root_anon_vma;
531 unsigned long anon_mapping;
534 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
535 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
537 if (!page_mapped(page))
540 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
541 root_anon_vma = READ_ONCE(anon_vma->root);
542 if (down_read_trylock(&root_anon_vma->rwsem)) {
544 * If the page is still mapped, then this anon_vma is still
545 * its anon_vma, and holding the mutex ensures that it will
546 * not go away, see anon_vma_free().
548 if (!page_mapped(page)) {
549 up_read(&root_anon_vma->rwsem);
555 /* trylock failed, we got to sleep */
556 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
561 if (!page_mapped(page)) {
563 put_anon_vma(anon_vma);
567 /* we pinned the anon_vma, its safe to sleep */
569 anon_vma_lock_read(anon_vma);
571 if (atomic_dec_and_test(&anon_vma->refcount)) {
573 * Oops, we held the last refcount, release the lock
574 * and bail -- can't simply use put_anon_vma() because
575 * we'll deadlock on the anon_vma_lock_write() recursion.
577 anon_vma_unlock_read(anon_vma);
578 __put_anon_vma(anon_vma);
589 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
591 anon_vma_unlock_read(anon_vma);
594 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
596 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
597 * important if a PTE was dirty when it was unmapped that it's flushed
598 * before any IO is initiated on the page to prevent lost writes. Similarly,
599 * it must be flushed before freeing to prevent data leakage.
601 void try_to_unmap_flush(void)
603 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
605 if (!tlb_ubc->flush_required)
608 arch_tlbbatch_flush(&tlb_ubc->arch);
609 tlb_ubc->flush_required = false;
610 tlb_ubc->writable = false;
613 /* Flush iff there are potentially writable TLB entries that can race with IO */
614 void try_to_unmap_flush_dirty(void)
616 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
618 if (tlb_ubc->writable)
619 try_to_unmap_flush();
622 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
624 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
626 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
627 tlb_ubc->flush_required = true;
630 * Ensure compiler does not re-order the setting of tlb_flush_batched
631 * before the PTE is cleared.
634 mm->tlb_flush_batched = true;
637 * If the PTE was dirty then it's best to assume it's writable. The
638 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
639 * before the page is queued for IO.
642 tlb_ubc->writable = true;
646 * Returns true if the TLB flush should be deferred to the end of a batch of
647 * unmap operations to reduce IPIs.
649 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
651 bool should_defer = false;
653 if (!(flags & TTU_BATCH_FLUSH))
656 /* If remote CPUs need to be flushed then defer batch the flush */
657 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
665 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
666 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
667 * operation such as mprotect or munmap to race between reclaim unmapping
668 * the page and flushing the page. If this race occurs, it potentially allows
669 * access to data via a stale TLB entry. Tracking all mm's that have TLB
670 * batching in flight would be expensive during reclaim so instead track
671 * whether TLB batching occurred in the past and if so then do a flush here
672 * if required. This will cost one additional flush per reclaim cycle paid
673 * by the first operation at risk such as mprotect and mumap.
675 * This must be called under the PTL so that an access to tlb_flush_batched
676 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
679 void flush_tlb_batched_pending(struct mm_struct *mm)
681 if (mm->tlb_flush_batched) {
685 * Do not allow the compiler to re-order the clearing of
686 * tlb_flush_batched before the tlb is flushed.
689 mm->tlb_flush_batched = false;
693 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
697 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
701 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
704 * At what user virtual address is page expected in vma?
705 * Caller should check the page is actually part of the vma.
707 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
709 unsigned long address;
710 if (PageAnon(page)) {
711 struct anon_vma *page__anon_vma = page_anon_vma(page);
713 * Note: swapoff's unuse_vma() is more efficient with this
714 * check, and needs it to match anon_vma when KSM is active.
716 if (!vma->anon_vma || !page__anon_vma ||
717 vma->anon_vma->root != page__anon_vma->root)
719 } else if (page->mapping) {
720 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
724 address = __vma_address(page, vma);
725 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
730 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
738 pgd = pgd_offset(mm, address);
739 if (!pgd_present(*pgd))
742 p4d = p4d_offset(pgd, address);
743 if (!p4d_present(*p4d))
746 pud = pud_offset(p4d, address);
747 if (!pud_present(*pud))
750 pmd = pmd_offset(pud, address);
752 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
753 * without holding anon_vma lock for write. So when looking for a
754 * genuine pmde (in which to find pte), test present and !THP together.
758 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
764 struct page_referenced_arg {
767 unsigned long vm_flags;
768 struct mem_cgroup *memcg;
771 * arg: page_referenced_arg will be passed
773 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
774 unsigned long address, void *arg)
776 struct page_referenced_arg *pra = arg;
777 struct page_vma_mapped_walk pvmw = {
784 while (page_vma_mapped_walk(&pvmw)) {
785 address = pvmw.address;
787 if (vma->vm_flags & VM_LOCKED) {
788 page_vma_mapped_walk_done(&pvmw);
789 pra->vm_flags |= VM_LOCKED;
790 return false; /* To break the loop */
794 if (ptep_clear_flush_young_notify(vma, address,
797 * Don't treat a reference through
798 * a sequentially read mapping as such.
799 * If the page has been used in another mapping,
800 * we will catch it; if this other mapping is
801 * already gone, the unmap path will have set
802 * PG_referenced or activated the page.
804 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
807 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
808 if (pmdp_clear_flush_young_notify(vma, address,
812 /* unexpected pmd-mapped page? */
820 clear_page_idle(page);
821 if (test_and_clear_page_young(page))
826 pra->vm_flags |= vma->vm_flags;
830 return false; /* To break the loop */
835 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
837 struct page_referenced_arg *pra = arg;
838 struct mem_cgroup *memcg = pra->memcg;
840 if (!mm_match_cgroup(vma->vm_mm, memcg))
847 * page_referenced - test if the page was referenced
848 * @page: the page to test
849 * @is_locked: caller holds lock on the page
850 * @memcg: target memory cgroup
851 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
853 * Quick test_and_clear_referenced for all mappings to a page,
854 * returns the number of ptes which referenced the page.
856 int page_referenced(struct page *page,
858 struct mem_cgroup *memcg,
859 unsigned long *vm_flags)
862 struct page_referenced_arg pra = {
863 .mapcount = total_mapcount(page),
866 struct rmap_walk_control rwc = {
867 .rmap_one = page_referenced_one,
869 .anon_lock = page_lock_anon_vma_read,
876 if (!page_rmapping(page))
879 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
880 we_locked = trylock_page(page);
886 * If we are reclaiming on behalf of a cgroup, skip
887 * counting on behalf of references from different
891 rwc.invalid_vma = invalid_page_referenced_vma;
894 rmap_walk(page, &rwc);
895 *vm_flags = pra.vm_flags;
900 return pra.referenced;
903 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
904 unsigned long address, void *arg)
906 struct page_vma_mapped_walk pvmw = {
912 struct mmu_notifier_range range;
916 * We have to assume the worse case ie pmd for invalidation. Note that
917 * the page can not be free from this function.
919 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
920 0, vma, vma->vm_mm, address,
921 min(vma->vm_end, address + page_size(page)));
922 mmu_notifier_invalidate_range_start(&range);
924 while (page_vma_mapped_walk(&pvmw)) {
927 address = pvmw.address;
930 pte_t *pte = pvmw.pte;
932 if (!pte_dirty(*pte) && !pte_write(*pte))
935 flush_cache_page(vma, address, pte_pfn(*pte));
936 entry = ptep_clear_flush(vma, address, pte);
937 entry = pte_wrprotect(entry);
938 entry = pte_mkclean(entry);
939 set_pte_at(vma->vm_mm, address, pte, entry);
942 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
943 pmd_t *pmd = pvmw.pmd;
946 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
949 flush_cache_page(vma, address, page_to_pfn(page));
950 entry = pmdp_invalidate(vma, address, pmd);
951 entry = pmd_wrprotect(entry);
952 entry = pmd_mkclean(entry);
953 set_pmd_at(vma->vm_mm, address, pmd, entry);
956 /* unexpected pmd-mapped page? */
962 * No need to call mmu_notifier_invalidate_range() as we are
963 * downgrading page table protection not changing it to point
966 * See Documentation/vm/mmu_notifier.rst
972 mmu_notifier_invalidate_range_end(&range);
977 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
979 if (vma->vm_flags & VM_SHARED)
985 int page_mkclean(struct page *page)
988 struct address_space *mapping;
989 struct rmap_walk_control rwc = {
990 .arg = (void *)&cleaned,
991 .rmap_one = page_mkclean_one,
992 .invalid_vma = invalid_mkclean_vma,
995 BUG_ON(!PageLocked(page));
997 if (!page_mapped(page))
1000 mapping = page_mapping(page);
1004 rmap_walk(page, &rwc);
1008 EXPORT_SYMBOL_GPL(page_mkclean);
1011 * page_move_anon_rmap - move a page to our anon_vma
1012 * @page: the page to move to our anon_vma
1013 * @vma: the vma the page belongs to
1015 * When a page belongs exclusively to one process after a COW event,
1016 * that page can be moved into the anon_vma that belongs to just that
1017 * process, so the rmap code will not search the parent or sibling
1020 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1022 struct anon_vma *anon_vma = vma->anon_vma;
1024 page = compound_head(page);
1026 VM_BUG_ON_PAGE(!PageLocked(page), page);
1027 VM_BUG_ON_VMA(!anon_vma, vma);
1029 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1031 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1032 * simultaneously, so a concurrent reader (eg page_referenced()'s
1033 * PageAnon()) will not see one without the other.
1035 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1039 * __page_set_anon_rmap - set up new anonymous rmap
1040 * @page: Page or Hugepage to add to rmap
1041 * @vma: VM area to add page to.
1042 * @address: User virtual address of the mapping
1043 * @exclusive: the page is exclusively owned by the current process
1045 static void __page_set_anon_rmap(struct page *page,
1046 struct vm_area_struct *vma, unsigned long address, int exclusive)
1048 struct anon_vma *anon_vma = vma->anon_vma;
1056 * If the page isn't exclusively mapped into this vma,
1057 * we must use the _oldest_ possible anon_vma for the
1061 anon_vma = anon_vma->root;
1063 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1064 page->mapping = (struct address_space *) anon_vma;
1065 page->index = linear_page_index(vma, address);
1069 * __page_check_anon_rmap - sanity check anonymous rmap addition
1070 * @page: the page to add the mapping to
1071 * @vma: the vm area in which the mapping is added
1072 * @address: the user virtual address mapped
1074 static void __page_check_anon_rmap(struct page *page,
1075 struct vm_area_struct *vma, unsigned long address)
1077 #ifdef CONFIG_DEBUG_VM
1079 * The page's anon-rmap details (mapping and index) are guaranteed to
1080 * be set up correctly at this point.
1082 * We have exclusion against page_add_anon_rmap because the caller
1083 * always holds the page locked, except if called from page_dup_rmap,
1084 * in which case the page is already known to be setup.
1086 * We have exclusion against page_add_new_anon_rmap because those pages
1087 * are initially only visible via the pagetables, and the pte is locked
1088 * over the call to page_add_new_anon_rmap.
1090 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1091 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1096 * page_add_anon_rmap - add pte mapping to an anonymous page
1097 * @page: the page to add the mapping to
1098 * @vma: the vm area in which the mapping is added
1099 * @address: the user virtual address mapped
1100 * @compound: charge the page as compound or small page
1102 * The caller needs to hold the pte lock, and the page must be locked in
1103 * the anon_vma case: to serialize mapping,index checking after setting,
1104 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1105 * (but PageKsm is never downgraded to PageAnon).
1107 void page_add_anon_rmap(struct page *page,
1108 struct vm_area_struct *vma, unsigned long address, bool compound)
1110 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1114 * Special version of the above for do_swap_page, which often runs
1115 * into pages that are exclusively owned by the current process.
1116 * Everybody else should continue to use page_add_anon_rmap above.
1118 void do_page_add_anon_rmap(struct page *page,
1119 struct vm_area_struct *vma, unsigned long address, int flags)
1121 bool compound = flags & RMAP_COMPOUND;
1126 VM_BUG_ON_PAGE(!PageLocked(page), page);
1127 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1128 mapcount = compound_mapcount_ptr(page);
1129 first = atomic_inc_and_test(mapcount);
1131 first = atomic_inc_and_test(&page->_mapcount);
1135 int nr = compound ? hpage_nr_pages(page) : 1;
1137 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1138 * these counters are not modified in interrupt context, and
1139 * pte lock(a spinlock) is held, which implies preemption
1143 __inc_node_page_state(page, NR_ANON_THPS);
1144 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1146 if (unlikely(PageKsm(page)))
1149 VM_BUG_ON_PAGE(!PageLocked(page), page);
1151 /* address might be in next vma when migration races vma_adjust */
1153 __page_set_anon_rmap(page, vma, address,
1154 flags & RMAP_EXCLUSIVE);
1156 __page_check_anon_rmap(page, vma, address);
1160 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1161 * @page: the page to add the mapping to
1162 * @vma: the vm area in which the mapping is added
1163 * @address: the user virtual address mapped
1164 * @compound: charge the page as compound or small page
1166 * Same as page_add_anon_rmap but must only be called on *new* pages.
1167 * This means the inc-and-test can be bypassed.
1168 * Page does not have to be locked.
1170 void page_add_new_anon_rmap(struct page *page,
1171 struct vm_area_struct *vma, unsigned long address, bool compound)
1173 int nr = compound ? hpage_nr_pages(page) : 1;
1175 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1176 __SetPageSwapBacked(page);
1178 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1179 /* increment count (starts at -1) */
1180 atomic_set(compound_mapcount_ptr(page), 0);
1181 __inc_node_page_state(page, NR_ANON_THPS);
1183 /* Anon THP always mapped first with PMD */
1184 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1185 /* increment count (starts at -1) */
1186 atomic_set(&page->_mapcount, 0);
1188 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1189 __page_set_anon_rmap(page, vma, address, 1);
1193 * page_add_file_rmap - add pte mapping to a file page
1194 * @page: the page to add the mapping to
1195 * @compound: charge the page as compound or small page
1197 * The caller needs to hold the pte lock.
1199 void page_add_file_rmap(struct page *page, bool compound)
1203 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1204 lock_page_memcg(page);
1205 if (compound && PageTransHuge(page)) {
1206 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1207 if (atomic_inc_and_test(&page[i]._mapcount))
1210 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1212 if (PageSwapBacked(page))
1213 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1215 __inc_node_page_state(page, NR_FILE_PMDMAPPED);
1217 if (PageTransCompound(page) && page_mapping(page)) {
1218 VM_WARN_ON_ONCE(!PageLocked(page));
1220 SetPageDoubleMap(compound_head(page));
1221 if (PageMlocked(page))
1222 clear_page_mlock(compound_head(page));
1224 if (!atomic_inc_and_test(&page->_mapcount))
1227 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1229 unlock_page_memcg(page);
1232 static void page_remove_file_rmap(struct page *page, bool compound)
1236 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1237 lock_page_memcg(page);
1239 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1240 if (unlikely(PageHuge(page))) {
1241 /* hugetlb pages are always mapped with pmds */
1242 atomic_dec(compound_mapcount_ptr(page));
1246 /* page still mapped by someone else? */
1247 if (compound && PageTransHuge(page)) {
1248 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1249 if (atomic_add_negative(-1, &page[i]._mapcount))
1252 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1254 if (PageSwapBacked(page))
1255 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1257 __dec_node_page_state(page, NR_FILE_PMDMAPPED);
1259 if (!atomic_add_negative(-1, &page->_mapcount))
1264 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1265 * these counters are not modified in interrupt context, and
1266 * pte lock(a spinlock) is held, which implies preemption disabled.
1268 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1270 if (unlikely(PageMlocked(page)))
1271 clear_page_mlock(page);
1273 unlock_page_memcg(page);
1276 static void page_remove_anon_compound_rmap(struct page *page)
1280 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1283 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1284 if (unlikely(PageHuge(page)))
1287 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1290 __dec_node_page_state(page, NR_ANON_THPS);
1292 if (TestClearPageDoubleMap(page)) {
1294 * Subpages can be mapped with PTEs too. Check how many of
1295 * themi are still mapped.
1297 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1298 if (atomic_add_negative(-1, &page[i]._mapcount))
1305 if (unlikely(PageMlocked(page)))
1306 clear_page_mlock(page);
1309 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1310 deferred_split_huge_page(page);
1315 * page_remove_rmap - take down pte mapping from a page
1316 * @page: page to remove mapping from
1317 * @compound: uncharge the page as compound or small page
1319 * The caller needs to hold the pte lock.
1321 void page_remove_rmap(struct page *page, bool compound)
1323 if (!PageAnon(page))
1324 return page_remove_file_rmap(page, compound);
1327 return page_remove_anon_compound_rmap(page);
1329 /* page still mapped by someone else? */
1330 if (!atomic_add_negative(-1, &page->_mapcount))
1334 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1335 * these counters are not modified in interrupt context, and
1336 * pte lock(a spinlock) is held, which implies preemption disabled.
1338 __dec_node_page_state(page, NR_ANON_MAPPED);
1340 if (unlikely(PageMlocked(page)))
1341 clear_page_mlock(page);
1343 if (PageTransCompound(page))
1344 deferred_split_huge_page(compound_head(page));
1347 * It would be tidy to reset the PageAnon mapping here,
1348 * but that might overwrite a racing page_add_anon_rmap
1349 * which increments mapcount after us but sets mapping
1350 * before us: so leave the reset to free_unref_page,
1351 * and remember that it's only reliable while mapped.
1352 * Leaving it set also helps swapoff to reinstate ptes
1353 * faster for those pages still in swapcache.
1358 * @arg: enum ttu_flags will be passed to this argument
1360 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1361 unsigned long address, void *arg)
1363 struct mm_struct *mm = vma->vm_mm;
1364 struct page_vma_mapped_walk pvmw = {
1370 struct page *subpage;
1372 struct mmu_notifier_range range;
1373 enum ttu_flags flags = (enum ttu_flags)arg;
1375 /* munlock has nothing to gain from examining un-locked vmas */
1376 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1379 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1380 is_zone_device_page(page) && !is_device_private_page(page))
1383 if (flags & TTU_SPLIT_HUGE_PMD) {
1384 split_huge_pmd_address(vma, address,
1385 flags & TTU_SPLIT_FREEZE, page);
1389 * For THP, we have to assume the worse case ie pmd for invalidation.
1390 * For hugetlb, it could be much worse if we need to do pud
1391 * invalidation in the case of pmd sharing.
1393 * Note that the page can not be free in this function as call of
1394 * try_to_unmap() must hold a reference on the page.
1396 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1398 min(vma->vm_end, address + page_size(page)));
1399 if (PageHuge(page)) {
1401 * If sharing is possible, start and end will be adjusted
1404 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1407 mmu_notifier_invalidate_range_start(&range);
1409 while (page_vma_mapped_walk(&pvmw)) {
1410 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1411 /* PMD-mapped THP migration entry */
1412 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1413 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1415 set_pmd_migration_entry(&pvmw, page);
1421 * If the page is mlock()d, we cannot swap it out.
1422 * If it's recently referenced (perhaps page_referenced
1423 * skipped over this mm) then we should reactivate it.
1425 if (!(flags & TTU_IGNORE_MLOCK)) {
1426 if (vma->vm_flags & VM_LOCKED) {
1427 /* PTE-mapped THP are never mlocked */
1428 if (!PageTransCompound(page)) {
1430 * Holding pte lock, we do *not* need
1433 mlock_vma_page(page);
1436 page_vma_mapped_walk_done(&pvmw);
1439 if (flags & TTU_MUNLOCK)
1443 /* Unexpected PMD-mapped THP? */
1444 VM_BUG_ON_PAGE(!pvmw.pte, page);
1446 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1447 address = pvmw.address;
1449 if (PageHuge(page)) {
1450 if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1452 * huge_pmd_unshare unmapped an entire PMD
1453 * page. There is no way of knowing exactly
1454 * which PMDs may be cached for this mm, so
1455 * we must flush them all. start/end were
1456 * already adjusted above to cover this range.
1458 flush_cache_range(vma, range.start, range.end);
1459 flush_tlb_range(vma, range.start, range.end);
1460 mmu_notifier_invalidate_range(mm, range.start,
1464 * The ref count of the PMD page was dropped
1465 * which is part of the way map counting
1466 * is done for shared PMDs. Return 'true'
1467 * here. When there is no other sharing,
1468 * huge_pmd_unshare returns false and we will
1469 * unmap the actual page and drop map count
1472 page_vma_mapped_walk_done(&pvmw);
1477 if (IS_ENABLED(CONFIG_MIGRATION) &&
1478 (flags & TTU_MIGRATION) &&
1479 is_zone_device_page(page)) {
1483 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1486 * Store the pfn of the page in a special migration
1487 * pte. do_swap_page() will wait until the migration
1488 * pte is removed and then restart fault handling.
1490 entry = make_migration_entry(page, 0);
1491 swp_pte = swp_entry_to_pte(entry);
1492 if (pte_soft_dirty(pteval))
1493 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1494 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1496 * No need to invalidate here it will synchronize on
1497 * against the special swap migration pte.
1499 * The assignment to subpage above was computed from a
1500 * swap PTE which results in an invalid pointer.
1501 * Since only PAGE_SIZE pages can currently be
1502 * migrated, just set it to page. This will need to be
1503 * changed when hugepage migrations to device private
1504 * memory are supported.
1510 if (!(flags & TTU_IGNORE_ACCESS)) {
1511 if (ptep_clear_flush_young_notify(vma, address,
1514 page_vma_mapped_walk_done(&pvmw);
1519 /* Nuke the page table entry. */
1520 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1521 if (should_defer_flush(mm, flags)) {
1523 * We clear the PTE but do not flush so potentially
1524 * a remote CPU could still be writing to the page.
1525 * If the entry was previously clean then the
1526 * architecture must guarantee that a clear->dirty
1527 * transition on a cached TLB entry is written through
1528 * and traps if the PTE is unmapped.
1530 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1532 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1534 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1537 /* Move the dirty bit to the page. Now the pte is gone. */
1538 if (pte_dirty(pteval))
1539 set_page_dirty(page);
1541 /* Update high watermark before we lower rss */
1542 update_hiwater_rss(mm);
1544 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1545 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1546 if (PageHuge(page)) {
1547 hugetlb_count_sub(compound_nr(page), mm);
1548 set_huge_swap_pte_at(mm, address,
1550 vma_mmu_pagesize(vma));
1552 dec_mm_counter(mm, mm_counter(page));
1553 set_pte_at(mm, address, pvmw.pte, pteval);
1556 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1558 * The guest indicated that the page content is of no
1559 * interest anymore. Simply discard the pte, vmscan
1560 * will take care of the rest.
1561 * A future reference will then fault in a new zero
1562 * page. When userfaultfd is active, we must not drop
1563 * this page though, as its main user (postcopy
1564 * migration) will not expect userfaults on already
1567 dec_mm_counter(mm, mm_counter(page));
1568 /* We have to invalidate as we cleared the pte */
1569 mmu_notifier_invalidate_range(mm, address,
1570 address + PAGE_SIZE);
1571 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1572 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1576 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1577 set_pte_at(mm, address, pvmw.pte, pteval);
1579 page_vma_mapped_walk_done(&pvmw);
1584 * Store the pfn of the page in a special migration
1585 * pte. do_swap_page() will wait until the migration
1586 * pte is removed and then restart fault handling.
1588 entry = make_migration_entry(subpage,
1590 swp_pte = swp_entry_to_pte(entry);
1591 if (pte_soft_dirty(pteval))
1592 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1593 set_pte_at(mm, address, pvmw.pte, swp_pte);
1595 * No need to invalidate here it will synchronize on
1596 * against the special swap migration pte.
1598 } else if (PageAnon(page)) {
1599 swp_entry_t entry = { .val = page_private(subpage) };
1602 * Store the swap location in the pte.
1603 * See handle_pte_fault() ...
1605 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1608 /* We have to invalidate as we cleared the pte */
1609 mmu_notifier_invalidate_range(mm, address,
1610 address + PAGE_SIZE);
1611 page_vma_mapped_walk_done(&pvmw);
1615 /* MADV_FREE page check */
1616 if (!PageSwapBacked(page)) {
1617 if (!PageDirty(page)) {
1618 /* Invalidate as we cleared the pte */
1619 mmu_notifier_invalidate_range(mm,
1620 address, address + PAGE_SIZE);
1621 dec_mm_counter(mm, MM_ANONPAGES);
1626 * If the page was redirtied, it cannot be
1627 * discarded. Remap the page to page table.
1629 set_pte_at(mm, address, pvmw.pte, pteval);
1630 SetPageSwapBacked(page);
1632 page_vma_mapped_walk_done(&pvmw);
1636 if (swap_duplicate(entry) < 0) {
1637 set_pte_at(mm, address, pvmw.pte, pteval);
1639 page_vma_mapped_walk_done(&pvmw);
1642 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1643 set_pte_at(mm, address, pvmw.pte, pteval);
1645 page_vma_mapped_walk_done(&pvmw);
1648 if (list_empty(&mm->mmlist)) {
1649 spin_lock(&mmlist_lock);
1650 if (list_empty(&mm->mmlist))
1651 list_add(&mm->mmlist, &init_mm.mmlist);
1652 spin_unlock(&mmlist_lock);
1654 dec_mm_counter(mm, MM_ANONPAGES);
1655 inc_mm_counter(mm, MM_SWAPENTS);
1656 swp_pte = swp_entry_to_pte(entry);
1657 if (pte_soft_dirty(pteval))
1658 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1659 set_pte_at(mm, address, pvmw.pte, swp_pte);
1660 /* Invalidate as we cleared the pte */
1661 mmu_notifier_invalidate_range(mm, address,
1662 address + PAGE_SIZE);
1665 * This is a locked file-backed page, thus it cannot
1666 * be removed from the page cache and replaced by a new
1667 * page before mmu_notifier_invalidate_range_end, so no
1668 * concurrent thread might update its page table to
1669 * point at new page while a device still is using this
1672 * See Documentation/vm/mmu_notifier.rst
1674 dec_mm_counter(mm, mm_counter_file(page));
1678 * No need to call mmu_notifier_invalidate_range() it has be
1679 * done above for all cases requiring it to happen under page
1680 * table lock before mmu_notifier_invalidate_range_end()
1682 * See Documentation/vm/mmu_notifier.rst
1684 page_remove_rmap(subpage, PageHuge(page));
1688 mmu_notifier_invalidate_range_end(&range);
1693 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1695 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1700 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1701 VM_STACK_INCOMPLETE_SETUP)
1707 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1709 return is_vma_temporary_stack(vma);
1712 static int page_mapcount_is_zero(struct page *page)
1714 return !total_mapcount(page);
1718 * try_to_unmap - try to remove all page table mappings to a page
1719 * @page: the page to get unmapped
1720 * @flags: action and flags
1722 * Tries to remove all the page table entries which are mapping this
1723 * page, used in the pageout path. Caller must hold the page lock.
1725 * If unmap is successful, return true. Otherwise, false.
1727 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1729 struct rmap_walk_control rwc = {
1730 .rmap_one = try_to_unmap_one,
1731 .arg = (void *)flags,
1732 .done = page_mapcount_is_zero,
1733 .anon_lock = page_lock_anon_vma_read,
1737 * During exec, a temporary VMA is setup and later moved.
1738 * The VMA is moved under the anon_vma lock but not the
1739 * page tables leading to a race where migration cannot
1740 * find the migration ptes. Rather than increasing the
1741 * locking requirements of exec(), migration skips
1742 * temporary VMAs until after exec() completes.
1744 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1745 && !PageKsm(page) && PageAnon(page))
1746 rwc.invalid_vma = invalid_migration_vma;
1748 if (flags & TTU_RMAP_LOCKED)
1749 rmap_walk_locked(page, &rwc);
1751 rmap_walk(page, &rwc);
1753 return !page_mapcount(page) ? true : false;
1756 static int page_not_mapped(struct page *page)
1758 return !page_mapped(page);
1762 * try_to_munlock - try to munlock a page
1763 * @page: the page to be munlocked
1765 * Called from munlock code. Checks all of the VMAs mapping the page
1766 * to make sure nobody else has this page mlocked. The page will be
1767 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1770 void try_to_munlock(struct page *page)
1772 struct rmap_walk_control rwc = {
1773 .rmap_one = try_to_unmap_one,
1774 .arg = (void *)TTU_MUNLOCK,
1775 .done = page_not_mapped,
1776 .anon_lock = page_lock_anon_vma_read,
1780 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1781 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1783 rmap_walk(page, &rwc);
1786 void __put_anon_vma(struct anon_vma *anon_vma)
1788 struct anon_vma *root = anon_vma->root;
1790 anon_vma_free(anon_vma);
1791 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1792 anon_vma_free(root);
1795 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1796 struct rmap_walk_control *rwc)
1798 struct anon_vma *anon_vma;
1801 return rwc->anon_lock(page);
1804 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1805 * because that depends on page_mapped(); but not all its usages
1806 * are holding mmap_sem. Users without mmap_sem are required to
1807 * take a reference count to prevent the anon_vma disappearing
1809 anon_vma = page_anon_vma(page);
1813 anon_vma_lock_read(anon_vma);
1818 * rmap_walk_anon - do something to anonymous page using the object-based
1820 * @page: the page to be handled
1821 * @rwc: control variable according to each walk type
1823 * Find all the mappings of a page using the mapping pointer and the vma chains
1824 * contained in the anon_vma struct it points to.
1826 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1827 * where the page was found will be held for write. So, we won't recheck
1828 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1831 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1834 struct anon_vma *anon_vma;
1835 pgoff_t pgoff_start, pgoff_end;
1836 struct anon_vma_chain *avc;
1839 anon_vma = page_anon_vma(page);
1840 /* anon_vma disappear under us? */
1841 VM_BUG_ON_PAGE(!anon_vma, page);
1843 anon_vma = rmap_walk_anon_lock(page, rwc);
1848 pgoff_start = page_to_pgoff(page);
1849 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1850 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1851 pgoff_start, pgoff_end) {
1852 struct vm_area_struct *vma = avc->vma;
1853 unsigned long address = vma_address(page, vma);
1857 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1860 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1862 if (rwc->done && rwc->done(page))
1867 anon_vma_unlock_read(anon_vma);
1871 * rmap_walk_file - do something to file page using the object-based rmap method
1872 * @page: the page to be handled
1873 * @rwc: control variable according to each walk type
1875 * Find all the mappings of a page using the mapping pointer and the vma chains
1876 * contained in the address_space struct it points to.
1878 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1879 * where the page was found will be held for write. So, we won't recheck
1880 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1883 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1886 struct address_space *mapping = page_mapping(page);
1887 pgoff_t pgoff_start, pgoff_end;
1888 struct vm_area_struct *vma;
1891 * The page lock not only makes sure that page->mapping cannot
1892 * suddenly be NULLified by truncation, it makes sure that the
1893 * structure at mapping cannot be freed and reused yet,
1894 * so we can safely take mapping->i_mmap_rwsem.
1896 VM_BUG_ON_PAGE(!PageLocked(page), page);
1901 pgoff_start = page_to_pgoff(page);
1902 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1904 i_mmap_lock_read(mapping);
1905 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1906 pgoff_start, pgoff_end) {
1907 unsigned long address = vma_address(page, vma);
1911 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1914 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1916 if (rwc->done && rwc->done(page))
1922 i_mmap_unlock_read(mapping);
1925 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1927 if (unlikely(PageKsm(page)))
1928 rmap_walk_ksm(page, rwc);
1929 else if (PageAnon(page))
1930 rmap_walk_anon(page, rwc, false);
1932 rmap_walk_file(page, rwc, false);
1935 /* Like rmap_walk, but caller holds relevant rmap lock */
1936 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1938 /* no ksm support for now */
1939 VM_BUG_ON_PAGE(PageKsm(page), page);
1941 rmap_walk_anon(page, rwc, true);
1943 rmap_walk_file(page, rwc, true);
1946 #ifdef CONFIG_HUGETLB_PAGE
1948 * The following two functions are for anonymous (private mapped) hugepages.
1949 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1950 * and no lru code, because we handle hugepages differently from common pages.
1952 void hugepage_add_anon_rmap(struct page *page,
1953 struct vm_area_struct *vma, unsigned long address)
1955 struct anon_vma *anon_vma = vma->anon_vma;
1958 BUG_ON(!PageLocked(page));
1960 /* address might be in next vma when migration races vma_adjust */
1961 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1963 __page_set_anon_rmap(page, vma, address, 0);
1966 void hugepage_add_new_anon_rmap(struct page *page,
1967 struct vm_area_struct *vma, unsigned long address)
1969 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1970 atomic_set(compound_mapcount_ptr(page), 0);
1971 __page_set_anon_rmap(page, vma, address, 1);
1973 #endif /* CONFIG_HUGETLB_PAGE */