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->degree = 1; /* Reference for first vma */
93 anon_vma->parent = anon_vma;
95 * Initialise the anon_vma root to point to itself. If called
96 * from fork, the root will be reset to the parents anon_vma.
98 anon_vma->root = anon_vma;
104 static inline void anon_vma_free(struct anon_vma *anon_vma)
106 VM_BUG_ON(atomic_read(&anon_vma->refcount));
109 * Synchronize against page_lock_anon_vma_read() such that
110 * we can safely hold the lock without the anon_vma getting
113 * Relies on the full mb implied by the atomic_dec_and_test() from
114 * put_anon_vma() against the acquire barrier implied by
115 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
117 * page_lock_anon_vma_read() VS put_anon_vma()
118 * down_read_trylock() atomic_dec_and_test()
120 * atomic_read() rwsem_is_locked()
122 * LOCK should suffice since the actual taking of the lock must
123 * happen _before_ what follows.
126 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
127 anon_vma_lock_write(anon_vma);
128 anon_vma_unlock_write(anon_vma);
131 kmem_cache_free(anon_vma_cachep, anon_vma);
134 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
136 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
139 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
141 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
144 static void anon_vma_chain_link(struct vm_area_struct *vma,
145 struct anon_vma_chain *avc,
146 struct anon_vma *anon_vma)
149 avc->anon_vma = anon_vma;
150 list_add(&avc->same_vma, &vma->anon_vma_chain);
151 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
155 * __anon_vma_prepare - attach an anon_vma to a memory region
156 * @vma: the memory region in question
158 * This makes sure the memory mapping described by 'vma' has
159 * an 'anon_vma' attached to it, so that we can associate the
160 * anonymous pages mapped into it with that anon_vma.
162 * The common case will be that we already have one, which
163 * is handled inline by anon_vma_prepare(). But if
164 * not we either need to find an adjacent mapping that we
165 * can re-use the anon_vma from (very common when the only
166 * reason for splitting a vma has been mprotect()), or we
167 * allocate a new one.
169 * Anon-vma allocations are very subtle, because we may have
170 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
171 * and that may actually touch the spinlock even in the newly
172 * allocated vma (it depends on RCU to make sure that the
173 * anon_vma isn't actually destroyed).
175 * As a result, we need to do proper anon_vma locking even
176 * for the new allocation. At the same time, we do not want
177 * to do any locking for the common case of already having
180 * This must be called with the mmap_lock held for reading.
182 int __anon_vma_prepare(struct vm_area_struct *vma)
184 struct mm_struct *mm = vma->vm_mm;
185 struct anon_vma *anon_vma, *allocated;
186 struct anon_vma_chain *avc;
190 avc = anon_vma_chain_alloc(GFP_KERNEL);
194 anon_vma = find_mergeable_anon_vma(vma);
197 anon_vma = anon_vma_alloc();
198 if (unlikely(!anon_vma))
199 goto out_enomem_free_avc;
200 allocated = anon_vma;
203 anon_vma_lock_write(anon_vma);
204 /* page_table_lock to protect against threads */
205 spin_lock(&mm->page_table_lock);
206 if (likely(!vma->anon_vma)) {
207 vma->anon_vma = anon_vma;
208 anon_vma_chain_link(vma, avc, anon_vma);
209 /* vma reference or self-parent link for new root */
214 spin_unlock(&mm->page_table_lock);
215 anon_vma_unlock_write(anon_vma);
217 if (unlikely(allocated))
218 put_anon_vma(allocated);
220 anon_vma_chain_free(avc);
225 anon_vma_chain_free(avc);
231 * This is a useful helper function for locking the anon_vma root as
232 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
235 * Such anon_vma's should have the same root, so you'd expect to see
236 * just a single mutex_lock for the whole traversal.
238 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
240 struct anon_vma *new_root = anon_vma->root;
241 if (new_root != root) {
242 if (WARN_ON_ONCE(root))
243 up_write(&root->rwsem);
245 down_write(&root->rwsem);
250 static inline void unlock_anon_vma_root(struct anon_vma *root)
253 up_write(&root->rwsem);
257 * Attach the anon_vmas from src to dst.
258 * Returns 0 on success, -ENOMEM on failure.
260 * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and
261 * anon_vma_fork(). The first three want an exact copy of src, while the last
262 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
263 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
264 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
266 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
267 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
268 * This prevents degradation of anon_vma hierarchy to endless linear chain in
269 * case of constantly forking task. On the other hand, an anon_vma with more
270 * than one child isn't reused even if there was no alive vma, thus rmap
271 * walker has a good chance of avoiding scanning the whole hierarchy when it
272 * searches where page is mapped.
274 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
276 struct anon_vma_chain *avc, *pavc;
277 struct anon_vma *root = NULL;
279 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
280 struct anon_vma *anon_vma;
282 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
283 if (unlikely(!avc)) {
284 unlock_anon_vma_root(root);
286 avc = anon_vma_chain_alloc(GFP_KERNEL);
290 anon_vma = pavc->anon_vma;
291 root = lock_anon_vma_root(root, anon_vma);
292 anon_vma_chain_link(dst, avc, anon_vma);
295 * Reuse existing anon_vma if its degree lower than two,
296 * that means it has no vma and only one anon_vma child.
298 * Do not chose parent anon_vma, otherwise first child
299 * will always reuse it. 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 != src->anon_vma && anon_vma->degree < 2)
304 dst->anon_vma = anon_vma;
307 dst->anon_vma->degree++;
308 unlock_anon_vma_root(root);
313 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
314 * decremented in unlink_anon_vmas().
315 * We can safely do this because callers of anon_vma_clone() don't care
316 * about dst->anon_vma if anon_vma_clone() failed.
318 dst->anon_vma = NULL;
319 unlink_anon_vmas(dst);
324 * Attach vma to its own anon_vma, as well as to the anon_vmas that
325 * the corresponding VMA in the parent process is attached to.
326 * Returns 0 on success, non-zero on failure.
328 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
330 struct anon_vma_chain *avc;
331 struct anon_vma *anon_vma;
334 /* Don't bother if the parent process has no anon_vma here. */
338 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
339 vma->anon_vma = NULL;
342 * First, attach the new VMA to the parent VMA's anon_vmas,
343 * so rmap can find non-COWed pages in child processes.
345 error = anon_vma_clone(vma, pvma);
349 /* An existing anon_vma has been reused, all done then. */
353 /* Then add our own anon_vma. */
354 anon_vma = anon_vma_alloc();
357 avc = anon_vma_chain_alloc(GFP_KERNEL);
359 goto out_error_free_anon_vma;
362 * The root anon_vma's spinlock is the lock actually used when we
363 * lock any of the anon_vmas in this anon_vma tree.
365 anon_vma->root = pvma->anon_vma->root;
366 anon_vma->parent = pvma->anon_vma;
368 * With refcounts, an anon_vma can stay around longer than the
369 * process it belongs to. The root anon_vma needs to be pinned until
370 * this anon_vma is freed, because the lock lives in the root.
372 get_anon_vma(anon_vma->root);
373 /* Mark this anon_vma as the one where our new (COWed) pages go. */
374 vma->anon_vma = anon_vma;
375 anon_vma_lock_write(anon_vma);
376 anon_vma_chain_link(vma, avc, anon_vma);
377 anon_vma->parent->degree++;
378 anon_vma_unlock_write(anon_vma);
382 out_error_free_anon_vma:
383 put_anon_vma(anon_vma);
385 unlink_anon_vmas(vma);
389 void unlink_anon_vmas(struct vm_area_struct *vma)
391 struct anon_vma_chain *avc, *next;
392 struct anon_vma *root = NULL;
395 * Unlink each anon_vma chained to the VMA. This list is ordered
396 * from newest to oldest, ensuring the root anon_vma gets freed last.
398 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
399 struct anon_vma *anon_vma = avc->anon_vma;
401 root = lock_anon_vma_root(root, anon_vma);
402 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
405 * Leave empty anon_vmas on the list - we'll need
406 * to free them outside the lock.
408 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
409 anon_vma->parent->degree--;
413 list_del(&avc->same_vma);
414 anon_vma_chain_free(avc);
417 vma->anon_vma->degree--;
418 unlock_anon_vma_root(root);
421 * Iterate the list once more, it now only contains empty and unlinked
422 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
423 * needing to write-acquire the anon_vma->root->rwsem.
425 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
426 struct anon_vma *anon_vma = avc->anon_vma;
428 VM_WARN_ON(anon_vma->degree);
429 put_anon_vma(anon_vma);
431 list_del(&avc->same_vma);
432 anon_vma_chain_free(avc);
436 static void anon_vma_ctor(void *data)
438 struct anon_vma *anon_vma = data;
440 init_rwsem(&anon_vma->rwsem);
441 atomic_set(&anon_vma->refcount, 0);
442 anon_vma->rb_root = RB_ROOT_CACHED;
445 void __init anon_vma_init(void)
447 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
448 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
450 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
451 SLAB_PANIC|SLAB_ACCOUNT);
455 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
457 * Since there is no serialization what so ever against page_remove_rmap()
458 * the best this function can do is return a locked anon_vma that might
459 * have been relevant to this page.
461 * The page might have been remapped to a different anon_vma or the anon_vma
462 * returned may already be freed (and even reused).
464 * In case it was remapped to a different anon_vma, the new anon_vma will be a
465 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
466 * ensure that any anon_vma obtained from the page will still be valid for as
467 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
469 * All users of this function must be very careful when walking the anon_vma
470 * chain and verify that the page in question is indeed mapped in it
471 * [ something equivalent to page_mapped_in_vma() ].
473 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
474 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
475 * if there is a mapcount, we can dereference the anon_vma after observing
478 struct anon_vma *page_get_anon_vma(struct page *page)
480 struct anon_vma *anon_vma = NULL;
481 unsigned long anon_mapping;
484 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
485 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
487 if (!page_mapped(page))
490 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
491 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
497 * If this page is still mapped, then its anon_vma cannot have been
498 * freed. But if it has been unmapped, we have no security against the
499 * anon_vma structure being freed and reused (for another anon_vma:
500 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
501 * above cannot corrupt).
503 if (!page_mapped(page)) {
505 put_anon_vma(anon_vma);
515 * Similar to page_get_anon_vma() except it locks the anon_vma.
517 * Its a little more complex as it tries to keep the fast path to a single
518 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
519 * reference like with page_get_anon_vma() and then block on the mutex.
521 struct anon_vma *page_lock_anon_vma_read(struct page *page)
523 struct anon_vma *anon_vma = NULL;
524 struct anon_vma *root_anon_vma;
525 unsigned long anon_mapping;
528 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
529 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
531 if (!page_mapped(page))
534 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
535 root_anon_vma = READ_ONCE(anon_vma->root);
536 if (down_read_trylock(&root_anon_vma->rwsem)) {
538 * If the page is still mapped, then this anon_vma is still
539 * its anon_vma, and holding the mutex ensures that it will
540 * not go away, see anon_vma_free().
542 if (!page_mapped(page)) {
543 up_read(&root_anon_vma->rwsem);
549 /* trylock failed, we got to sleep */
550 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
555 if (!page_mapped(page)) {
557 put_anon_vma(anon_vma);
561 /* we pinned the anon_vma, its safe to sleep */
563 anon_vma_lock_read(anon_vma);
565 if (atomic_dec_and_test(&anon_vma->refcount)) {
567 * Oops, we held the last refcount, release the lock
568 * and bail -- can't simply use put_anon_vma() because
569 * we'll deadlock on the anon_vma_lock_write() recursion.
571 anon_vma_unlock_read(anon_vma);
572 __put_anon_vma(anon_vma);
583 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
585 anon_vma_unlock_read(anon_vma);
588 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
590 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
591 * important if a PTE was dirty when it was unmapped that it's flushed
592 * before any IO is initiated on the page to prevent lost writes. Similarly,
593 * it must be flushed before freeing to prevent data leakage.
595 void try_to_unmap_flush(void)
597 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
599 if (!tlb_ubc->flush_required)
602 arch_tlbbatch_flush(&tlb_ubc->arch);
603 tlb_ubc->flush_required = false;
604 tlb_ubc->writable = false;
607 /* Flush iff there are potentially writable TLB entries that can race with IO */
608 void try_to_unmap_flush_dirty(void)
610 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
612 if (tlb_ubc->writable)
613 try_to_unmap_flush();
616 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
618 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
620 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
621 tlb_ubc->flush_required = true;
624 * Ensure compiler does not re-order the setting of tlb_flush_batched
625 * before the PTE is cleared.
628 mm->tlb_flush_batched = true;
631 * If the PTE was dirty then it's best to assume it's writable. The
632 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
633 * before the page is queued for IO.
636 tlb_ubc->writable = true;
640 * Returns true if the TLB flush should be deferred to the end of a batch of
641 * unmap operations to reduce IPIs.
643 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
645 bool should_defer = false;
647 if (!(flags & TTU_BATCH_FLUSH))
650 /* If remote CPUs need to be flushed then defer batch the flush */
651 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
659 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
660 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
661 * operation such as mprotect or munmap to race between reclaim unmapping
662 * the page and flushing the page. If this race occurs, it potentially allows
663 * access to data via a stale TLB entry. Tracking all mm's that have TLB
664 * batching in flight would be expensive during reclaim so instead track
665 * whether TLB batching occurred in the past and if so then do a flush here
666 * if required. This will cost one additional flush per reclaim cycle paid
667 * by the first operation at risk such as mprotect and mumap.
669 * This must be called under the PTL so that an access to tlb_flush_batched
670 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
673 void flush_tlb_batched_pending(struct mm_struct *mm)
675 if (data_race(mm->tlb_flush_batched)) {
679 * Do not allow the compiler to re-order the clearing of
680 * tlb_flush_batched before the tlb is flushed.
683 mm->tlb_flush_batched = false;
687 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
691 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
695 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
698 * At what user virtual address is page expected in vma?
699 * Caller should check the page is actually part of the vma.
701 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
703 if (PageAnon(page)) {
704 struct anon_vma *page__anon_vma = page_anon_vma(page);
706 * Note: swapoff's unuse_vma() is more efficient with this
707 * check, and needs it to match anon_vma when KSM is active.
709 if (!vma->anon_vma || !page__anon_vma ||
710 vma->anon_vma->root != page__anon_vma->root)
712 } else if (!vma->vm_file) {
714 } else if (vma->vm_file->f_mapping != compound_head(page)->mapping) {
718 return vma_address(page, vma);
721 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
729 pgd = pgd_offset(mm, address);
730 if (!pgd_present(*pgd))
733 p4d = p4d_offset(pgd, address);
734 if (!p4d_present(*p4d))
737 pud = pud_offset(p4d, address);
738 if (!pud_present(*pud))
741 pmd = pmd_offset(pud, address);
743 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
744 * without holding anon_vma lock for write. So when looking for a
745 * genuine pmde (in which to find pte), test present and !THP together.
749 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
755 struct page_referenced_arg {
758 unsigned long vm_flags;
759 struct mem_cgroup *memcg;
762 * arg: page_referenced_arg will be passed
764 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
765 unsigned long address, void *arg)
767 struct page_referenced_arg *pra = arg;
768 struct page_vma_mapped_walk pvmw = {
775 while (page_vma_mapped_walk(&pvmw)) {
776 address = pvmw.address;
778 if (vma->vm_flags & VM_LOCKED) {
779 page_vma_mapped_walk_done(&pvmw);
780 pra->vm_flags |= VM_LOCKED;
781 return false; /* To break the loop */
785 if (ptep_clear_flush_young_notify(vma, address,
788 * Don't treat a reference through
789 * a sequentially read mapping as such.
790 * If the page has been used in another mapping,
791 * we will catch it; if this other mapping is
792 * already gone, the unmap path will have set
793 * PG_referenced or activated the page.
795 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
798 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
799 if (pmdp_clear_flush_young_notify(vma, address,
803 /* unexpected pmd-mapped page? */
811 clear_page_idle(page);
812 if (test_and_clear_page_young(page))
817 pra->vm_flags |= vma->vm_flags;
821 return false; /* To break the loop */
826 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
828 struct page_referenced_arg *pra = arg;
829 struct mem_cgroup *memcg = pra->memcg;
831 if (!mm_match_cgroup(vma->vm_mm, memcg))
838 * page_referenced - test if the page was referenced
839 * @page: the page to test
840 * @is_locked: caller holds lock on the page
841 * @memcg: target memory cgroup
842 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
844 * Quick test_and_clear_referenced for all mappings to a page,
845 * returns the number of ptes which referenced the page.
847 int page_referenced(struct page *page,
849 struct mem_cgroup *memcg,
850 unsigned long *vm_flags)
853 struct page_referenced_arg pra = {
854 .mapcount = total_mapcount(page),
857 struct rmap_walk_control rwc = {
858 .rmap_one = page_referenced_one,
860 .anon_lock = page_lock_anon_vma_read,
867 if (!page_rmapping(page))
870 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
871 we_locked = trylock_page(page);
877 * If we are reclaiming on behalf of a cgroup, skip
878 * counting on behalf of references from different
882 rwc.invalid_vma = invalid_page_referenced_vma;
885 rmap_walk(page, &rwc);
886 *vm_flags = pra.vm_flags;
891 return pra.referenced;
894 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
895 unsigned long address, void *arg)
897 struct page_vma_mapped_walk pvmw = {
903 struct mmu_notifier_range range;
907 * We have to assume the worse case ie pmd for invalidation. Note that
908 * the page can not be free from this function.
910 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
911 0, vma, vma->vm_mm, address,
912 vma_address_end(page, vma));
913 mmu_notifier_invalidate_range_start(&range);
915 while (page_vma_mapped_walk(&pvmw)) {
918 address = pvmw.address;
921 pte_t *pte = pvmw.pte;
923 if (!pte_dirty(*pte) && !pte_write(*pte))
926 flush_cache_page(vma, address, pte_pfn(*pte));
927 entry = ptep_clear_flush(vma, address, pte);
928 entry = pte_wrprotect(entry);
929 entry = pte_mkclean(entry);
930 set_pte_at(vma->vm_mm, address, pte, entry);
933 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
934 pmd_t *pmd = pvmw.pmd;
937 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
940 flush_cache_page(vma, address, page_to_pfn(page));
941 entry = pmdp_invalidate(vma, address, pmd);
942 entry = pmd_wrprotect(entry);
943 entry = pmd_mkclean(entry);
944 set_pmd_at(vma->vm_mm, address, pmd, entry);
947 /* unexpected pmd-mapped page? */
953 * No need to call mmu_notifier_invalidate_range() as we are
954 * downgrading page table protection not changing it to point
957 * See Documentation/vm/mmu_notifier.rst
963 mmu_notifier_invalidate_range_end(&range);
968 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
970 if (vma->vm_flags & VM_SHARED)
976 int page_mkclean(struct page *page)
979 struct address_space *mapping;
980 struct rmap_walk_control rwc = {
981 .arg = (void *)&cleaned,
982 .rmap_one = page_mkclean_one,
983 .invalid_vma = invalid_mkclean_vma,
986 BUG_ON(!PageLocked(page));
988 if (!page_mapped(page))
991 mapping = page_mapping(page);
995 rmap_walk(page, &rwc);
999 EXPORT_SYMBOL_GPL(page_mkclean);
1002 * page_move_anon_rmap - move a page to our anon_vma
1003 * @page: the page to move to our anon_vma
1004 * @vma: the vma the page belongs to
1006 * When a page belongs exclusively to one process after a COW event,
1007 * that page can be moved into the anon_vma that belongs to just that
1008 * process, so the rmap code will not search the parent or sibling
1011 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1013 struct anon_vma *anon_vma = vma->anon_vma;
1015 page = compound_head(page);
1017 VM_BUG_ON_PAGE(!PageLocked(page), page);
1018 VM_BUG_ON_VMA(!anon_vma, vma);
1020 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1022 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1023 * simultaneously, so a concurrent reader (eg page_referenced()'s
1024 * PageAnon()) will not see one without the other.
1026 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1030 * __page_set_anon_rmap - set up new anonymous rmap
1031 * @page: Page or Hugepage to add to rmap
1032 * @vma: VM area to add page to.
1033 * @address: User virtual address of the mapping
1034 * @exclusive: the page is exclusively owned by the current process
1036 static void __page_set_anon_rmap(struct page *page,
1037 struct vm_area_struct *vma, unsigned long address, int exclusive)
1039 struct anon_vma *anon_vma = vma->anon_vma;
1047 * If the page isn't exclusively mapped into this vma,
1048 * we must use the _oldest_ possible anon_vma for the
1052 anon_vma = anon_vma->root;
1054 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1055 page->mapping = (struct address_space *) anon_vma;
1056 page->index = linear_page_index(vma, address);
1060 * __page_check_anon_rmap - sanity check anonymous rmap addition
1061 * @page: the page to add the mapping to
1062 * @vma: the vm area in which the mapping is added
1063 * @address: the user virtual address mapped
1065 static void __page_check_anon_rmap(struct page *page,
1066 struct vm_area_struct *vma, unsigned long address)
1069 * The page's anon-rmap details (mapping and index) are guaranteed to
1070 * be set up correctly at this point.
1072 * We have exclusion against page_add_anon_rmap because the caller
1073 * always holds the page locked, except if called from page_dup_rmap,
1074 * in which case the page is already known to be setup.
1076 * We have exclusion against page_add_new_anon_rmap because those pages
1077 * are initially only visible via the pagetables, and the pte is locked
1078 * over the call to page_add_new_anon_rmap.
1080 VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page);
1081 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1086 * page_add_anon_rmap - add pte mapping to an anonymous page
1087 * @page: the page to add the mapping to
1088 * @vma: the vm area in which the mapping is added
1089 * @address: the user virtual address mapped
1090 * @compound: charge the page as compound or small page
1092 * The caller needs to hold the pte lock, and the page must be locked in
1093 * the anon_vma case: to serialize mapping,index checking after setting,
1094 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1095 * (but PageKsm is never downgraded to PageAnon).
1097 void page_add_anon_rmap(struct page *page,
1098 struct vm_area_struct *vma, unsigned long address, bool compound)
1100 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1104 * Special version of the above for do_swap_page, which often runs
1105 * into pages that are exclusively owned by the current process.
1106 * Everybody else should continue to use page_add_anon_rmap above.
1108 void do_page_add_anon_rmap(struct page *page,
1109 struct vm_area_struct *vma, unsigned long address, int flags)
1111 bool compound = flags & RMAP_COMPOUND;
1114 if (unlikely(PageKsm(page)))
1115 lock_page_memcg(page);
1117 VM_BUG_ON_PAGE(!PageLocked(page), page);
1121 VM_BUG_ON_PAGE(!PageLocked(page), page);
1122 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1123 mapcount = compound_mapcount_ptr(page);
1124 first = atomic_inc_and_test(mapcount);
1126 first = atomic_inc_and_test(&page->_mapcount);
1130 int nr = compound ? thp_nr_pages(page) : 1;
1132 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1133 * these counters are not modified in interrupt context, and
1134 * pte lock(a spinlock) is held, which implies preemption
1138 #ifdef CONFIG_FINEGRAINED_THP
1140 if (nr == HPAGE_PMD_NR)
1141 __inc_lruvec_page_state(page, NR_ANON_THPS);
1143 __inc_lruvec_page_state(page, NR_ANON_64KB_THPS);
1145 #else /* CONFIG_FINEGRAINED_THP */
1146 __inc_lruvec_page_state(page, NR_ANON_THPS);
1147 #endif /* CONFIG_FINEGRAINED_THP */
1148 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1151 if (unlikely(PageKsm(page))) {
1152 unlock_page_memcg(page);
1156 /* address might be in next vma when migration races vma_adjust */
1158 __page_set_anon_rmap(page, vma, address,
1159 flags & RMAP_EXCLUSIVE);
1161 __page_check_anon_rmap(page, vma, address);
1165 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1166 * @page: the page to add the mapping to
1167 * @vma: the vm area in which the mapping is added
1168 * @address: the user virtual address mapped
1169 * @compound: charge the page as compound or small page
1171 * Same as page_add_anon_rmap but must only be called on *new* pages.
1172 * This means the inc-and-test can be bypassed.
1173 * Page does not have to be locked.
1175 void page_add_new_anon_rmap(struct page *page,
1176 struct vm_area_struct *vma, unsigned long address, bool compound)
1178 int nr = compound ? thp_nr_pages(page) : 1;
1180 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1181 __SetPageSwapBacked(page);
1183 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1184 /* increment count (starts at -1) */
1185 atomic_set(compound_mapcount_ptr(page), 0);
1186 if (hpage_pincount_available(page))
1187 atomic_set(compound_pincount_ptr(page), 0);
1188 #ifdef CONFIG_FINEGRAINED_THP
1189 if (nr == HPAGE_PMD_NR)
1190 __inc_lruvec_page_state(page, NR_ANON_THPS);
1192 __inc_lruvec_page_state(page, NR_ANON_64KB_THPS);
1193 #else /* CONFIG_FINEGRAINED_THP */
1194 __inc_lruvec_page_state(page, NR_ANON_THPS);
1195 #endif /* CONFIG_FINEGRAINED_THP */
1197 /* Anon THP always mapped first with PMD */
1198 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1199 /* increment count (starts at -1) */
1200 atomic_set(&page->_mapcount, 0);
1202 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1203 __page_set_anon_rmap(page, vma, address, 1);
1207 * page_add_file_rmap - add pte mapping to a file page
1208 * @page: the page to add the mapping to
1209 * @compound: charge the page as compound or small page
1211 * The caller needs to hold the pte lock.
1213 void page_add_file_rmap(struct page *page, bool compound)
1217 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1218 lock_page_memcg(page);
1219 if (compound && PageTransHuge(page)) {
1220 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1221 if (atomic_inc_and_test(&page[i]._mapcount))
1224 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1226 if (PageSwapBacked(page))
1227 #ifdef CONFIG_FINEGRAINED_THP
1228 __inc_node_page_state(page, nr == HPAGE_PMD_NR ?
1229 NR_SHMEM_PMDMAPPED : NR_SHMEM_PTEMAPPED);
1231 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1234 #ifdef CONFIG_FINEGRAINED_THP
1235 __inc_node_page_state(page, nr == HPAGE_PMD_NR ?
1236 NR_FILE_PMDMAPPED : NR_FILE_PTEMAPPED);
1238 __inc_node_page_state(page, NR_FILE_PMDMAPPED);
1241 if (PageTransCompound(page) && page_mapping(page)) {
1242 VM_WARN_ON_ONCE(!PageLocked(page));
1244 SetPageDoubleMap(compound_head(page));
1245 if (PageMlocked(page))
1246 clear_page_mlock(compound_head(page));
1248 if (!atomic_inc_and_test(&page->_mapcount))
1251 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1253 unlock_page_memcg(page);
1256 static void page_remove_file_rmap(struct page *page, bool compound)
1260 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1262 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1263 if (unlikely(PageHuge(page))) {
1264 /* hugetlb pages are always mapped with pmds */
1265 atomic_dec(compound_mapcount_ptr(page));
1269 /* page still mapped by someone else? */
1270 if (compound && PageTransHuge(page)) {
1271 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1272 if (atomic_add_negative(-1, &page[i]._mapcount))
1275 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1277 if (PageSwapBacked(page))
1278 #ifdef CONFIG_FINEGRAINED_THP
1279 __dec_node_page_state(page, nr == HPAGE_PMD_NR ?
1280 NR_SHMEM_PMDMAPPED : NR_SHMEM_PTEMAPPED);
1282 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1285 #ifdef CONFIG_FINEGRAINED_THP
1286 __dec_node_page_state(page, nr == HPAGE_PMD_NR ?
1287 NR_FILE_PMDMAPPED : NR_FILE_PTEMAPPED);
1289 __dec_node_page_state(page, NR_FILE_PMDMAPPED);
1292 if (!atomic_add_negative(-1, &page->_mapcount))
1297 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1298 * these counters are not modified in interrupt context, and
1299 * pte lock(a spinlock) is held, which implies preemption disabled.
1301 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1303 if (unlikely(PageMlocked(page)))
1304 clear_page_mlock(page);
1307 static void page_remove_anon_compound_rmap(struct page *page)
1311 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1314 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1315 if (unlikely(PageHuge(page)))
1318 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1321 #ifdef CONFIG_FINEGRAINED_THP
1322 if (thp_nr_pages(page) == HPAGE_PMD_NR)
1323 __dec_lruvec_page_state(page, NR_ANON_THPS);
1325 __dec_lruvec_page_state(page, NR_ANON_64KB_THPS);
1326 #else /* CONFIG_FINEGRAINED_THP */
1327 __dec_lruvec_page_state(page, NR_ANON_THPS);
1328 #endif /* CONFIG_FINEGRAINED_THP */
1330 if (TestClearPageDoubleMap(page)) {
1332 * Subpages can be mapped with PTEs too. Check how many of
1333 * them are still mapped.
1335 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1336 if (atomic_add_negative(-1, &page[i]._mapcount))
1341 * Queue the page for deferred split if at least one small
1342 * page of the compound page is unmapped, but at least one
1343 * small page is still mapped.
1345 if (nr && nr < thp_nr_pages(page))
1346 deferred_split_huge_page(page);
1348 nr = thp_nr_pages(page);
1351 if (unlikely(PageMlocked(page)))
1352 clear_page_mlock(page);
1355 __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
1359 * page_remove_rmap - take down pte mapping from a page
1360 * @page: page to remove mapping from
1361 * @compound: uncharge the page as compound or small page
1363 * The caller needs to hold the pte lock.
1365 void page_remove_rmap(struct page *page, bool compound)
1367 lock_page_memcg(page);
1369 if (!PageAnon(page)) {
1370 page_remove_file_rmap(page, compound);
1375 page_remove_anon_compound_rmap(page);
1379 /* page still mapped by someone else? */
1380 if (!atomic_add_negative(-1, &page->_mapcount))
1384 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1385 * these counters are not modified in interrupt context, and
1386 * pte lock(a spinlock) is held, which implies preemption disabled.
1388 __dec_lruvec_page_state(page, NR_ANON_MAPPED);
1390 if (unlikely(PageMlocked(page))) {
1391 if (unlikely(PageTransCompound(page)))
1392 clear_page_mlock(compound_head(page));
1394 clear_page_mlock(page);
1397 if (PageTransCompound(page))
1398 deferred_split_huge_page(compound_head(page));
1401 * It would be tidy to reset the PageAnon mapping here,
1402 * but that might overwrite a racing page_add_anon_rmap
1403 * which increments mapcount after us but sets mapping
1404 * before us: so leave the reset to free_unref_page,
1405 * and remember that it's only reliable while mapped.
1406 * Leaving it set also helps swapoff to reinstate ptes
1407 * faster for those pages still in swapcache.
1410 unlock_page_memcg(page);
1414 * @arg: enum ttu_flags will be passed to this argument
1416 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1417 unsigned long address, void *arg)
1419 struct mm_struct *mm = vma->vm_mm;
1420 struct page_vma_mapped_walk pvmw = {
1426 struct page *subpage;
1428 struct mmu_notifier_range range;
1429 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1432 * When racing against e.g. zap_pte_range() on another cpu,
1433 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1434 * try_to_unmap() may return false when it is about to become true,
1435 * if page table locking is skipped: use TTU_SYNC to wait for that.
1437 if (flags & TTU_SYNC)
1438 pvmw.flags = PVMW_SYNC;
1440 /* munlock has nothing to gain from examining un-locked vmas */
1441 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1444 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1445 is_zone_device_page(page) && !is_device_private_page(page))
1448 if (flags & TTU_SPLIT_HUGE_PMD) {
1449 split_huge_pmd_address(vma, address,
1450 flags & TTU_SPLIT_FREEZE, page);
1453 #ifdef CONFIG_FINEGRAINED_THP
1454 if (flags & TTU_SPLIT_HUGE_PTE)
1455 split_huge_pte_address(vma, address,
1456 flags & TTU_SPLIT_FREEZE, page);
1460 * For THP, we have to assume the worse case ie pmd for invalidation.
1461 * For hugetlb, it could be much worse if we need to do pud
1462 * invalidation in the case of pmd sharing.
1464 * Note that the page can not be free in this function as call of
1465 * try_to_unmap() must hold a reference on the page.
1467 range.end = PageKsm(page) ?
1468 address + PAGE_SIZE : vma_address_end(page, vma);
1469 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1470 address, range.end);
1471 if (PageHuge(page)) {
1473 * If sharing is possible, start and end will be adjusted
1476 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1479 mmu_notifier_invalidate_range_start(&range);
1481 while (page_vma_mapped_walk(&pvmw)) {
1482 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1483 /* PMD-mapped THP migration entry */
1484 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1485 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1487 set_pmd_migration_entry(&pvmw, page);
1490 #ifdef CONFIG_FINEGRAINED_THP
1491 if (pvmw.pte && pte_cont(*pvmw.pte) && (flags & TTU_MIGRATION)) {
1492 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1494 set_huge_pte_migration_entry(&pvmw, page);
1497 #endif /* CONFIG_FINEGRAINED_THP */
1501 * If the page is mlock()d, we cannot swap it out.
1502 * If it's recently referenced (perhaps page_referenced
1503 * skipped over this mm) then we should reactivate it.
1505 if (!(flags & TTU_IGNORE_MLOCK)) {
1506 if (vma->vm_flags & VM_LOCKED) {
1507 /* PTE-mapped THP are never mlocked */
1508 if (!PageTransCompound(page)) {
1510 * Holding pte lock, we do *not* need
1513 mlock_vma_page(page);
1516 page_vma_mapped_walk_done(&pvmw);
1519 if (flags & TTU_MUNLOCK)
1523 /* Unexpected PMD-mapped THP? */
1524 VM_BUG_ON_PAGE(!pvmw.pte, page);
1526 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1527 address = pvmw.address;
1529 if (PageHuge(page) && !PageAnon(page)) {
1531 * To call huge_pmd_unshare, i_mmap_rwsem must be
1532 * held in write mode. Caller needs to explicitly
1533 * do this outside rmap routines.
1535 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1536 #ifdef CONFIG_FINEGRAINED_THP
1537 if (thp_nr_pages(page) == HPAGE_PMD_NR &&
1538 huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1540 * huge_pmd_unshare unmapped an entire PMD
1541 * page. There is no way of knowing exactly
1542 * which PMDs may be cached for this mm, so
1543 * we must flush them all. start/end were
1544 * already adjusted above to cover this range.
1546 flush_cache_range(vma, range.start, range.end);
1547 flush_tlb_range(vma, range.start, range.end);
1548 mmu_notifier_invalidate_range(mm, range.start,
1551 * The ref count of the PMD page was dropped
1552 * which is part of the way map counting
1553 * is done for shared PMDs. Return 'true'
1554 * here. When there is no other sharing,
1555 * huge_pmd_unshare returns false and we will
1556 * unmap the actual page and drop map count
1559 page_vma_mapped_walk_done(&pvmw);
1562 #else /* CONFIG_FINEGRAINED_THP */
1563 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1565 * huge_pmd_unshare unmapped an entire PMD
1566 * page. There is no way of knowing exactly
1567 * which PMDs may be cached for this mm, so
1568 * we must flush them all. start/end were
1569 * already adjusted above to cover this range.
1571 flush_cache_range(vma, range.start, range.end);
1572 flush_tlb_range(vma, range.start, range.end);
1573 mmu_notifier_invalidate_range(mm, range.start,
1577 * The ref count of the PMD page was dropped
1578 * which is part of the way map counting
1579 * is done for shared PMDs. Return 'true'
1580 * here. When there is no other sharing,
1581 * huge_pmd_unshare returns false and we will
1582 * unmap the actual page and drop map count
1585 page_vma_mapped_walk_done(&pvmw);
1588 #endif /* CONFIG_FINEGRAINED_THP */
1591 if (IS_ENABLED(CONFIG_MIGRATION) &&
1592 (flags & TTU_MIGRATION) &&
1593 is_zone_device_page(page)) {
1597 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1600 * Store the pfn of the page in a special migration
1601 * pte. do_swap_page() will wait until the migration
1602 * pte is removed and then restart fault handling.
1604 entry = make_migration_entry(page, 0);
1605 swp_pte = swp_entry_to_pte(entry);
1608 * pteval maps a zone device page and is therefore
1611 if (pte_swp_soft_dirty(pteval))
1612 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1613 if (pte_swp_uffd_wp(pteval))
1614 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1615 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1617 * No need to invalidate here it will synchronize on
1618 * against the special swap migration pte.
1620 * The assignment to subpage above was computed from a
1621 * swap PTE which results in an invalid pointer.
1622 * Since only PAGE_SIZE pages can currently be
1623 * migrated, just set it to page. This will need to be
1624 * changed when hugepage migrations to device private
1625 * memory are supported.
1631 /* Nuke the page table entry. */
1632 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1633 if (should_defer_flush(mm, flags)) {
1635 * We clear the PTE but do not flush so potentially
1636 * a remote CPU could still be writing to the page.
1637 * If the entry was previously clean then the
1638 * architecture must guarantee that a clear->dirty
1639 * transition on a cached TLB entry is written through
1640 * and traps if the PTE is unmapped.
1642 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1644 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1646 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1649 /* Move the dirty bit to the page. Now the pte is gone. */
1650 if (pte_dirty(pteval))
1651 set_page_dirty(page);
1653 /* Update high watermark before we lower rss */
1654 update_hiwater_rss(mm);
1656 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1657 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1658 if (PageHuge(page)) {
1659 hugetlb_count_sub(compound_nr(page), mm);
1660 set_huge_swap_pte_at(mm, address,
1662 vma_mmu_pagesize(vma));
1664 dec_mm_counter(mm, mm_counter(page));
1665 set_pte_at(mm, address, pvmw.pte, pteval);
1668 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1670 * The guest indicated that the page content is of no
1671 * interest anymore. Simply discard the pte, vmscan
1672 * will take care of the rest.
1673 * A future reference will then fault in a new zero
1674 * page. When userfaultfd is active, we must not drop
1675 * this page though, as its main user (postcopy
1676 * migration) will not expect userfaults on already
1679 dec_mm_counter(mm, mm_counter(page));
1680 /* We have to invalidate as we cleared the pte */
1681 mmu_notifier_invalidate_range(mm, address,
1682 address + PAGE_SIZE);
1683 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1684 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1688 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1689 set_pte_at(mm, address, pvmw.pte, pteval);
1691 page_vma_mapped_walk_done(&pvmw);
1696 * Store the pfn of the page in a special migration
1697 * pte. do_swap_page() will wait until the migration
1698 * pte is removed and then restart fault handling.
1700 entry = make_migration_entry(subpage,
1702 swp_pte = swp_entry_to_pte(entry);
1703 if (pte_soft_dirty(pteval))
1704 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1705 if (pte_uffd_wp(pteval))
1706 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1707 set_pte_at(mm, address, pvmw.pte, swp_pte);
1709 * No need to invalidate here it will synchronize on
1710 * against the special swap migration pte.
1712 } else if (PageAnon(page)) {
1713 swp_entry_t entry = { .val = page_private(subpage) };
1716 * Store the swap location in the pte.
1717 * See handle_pte_fault() ...
1719 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1722 /* We have to invalidate as we cleared the pte */
1723 mmu_notifier_invalidate_range(mm, address,
1724 address + PAGE_SIZE);
1725 page_vma_mapped_walk_done(&pvmw);
1729 /* MADV_FREE page check */
1730 if (!PageSwapBacked(page)) {
1731 if (!PageDirty(page)) {
1732 /* Invalidate as we cleared the pte */
1733 mmu_notifier_invalidate_range(mm,
1734 address, address + PAGE_SIZE);
1735 dec_mm_counter(mm, MM_ANONPAGES);
1740 * If the page was redirtied, it cannot be
1741 * discarded. Remap the page to page table.
1743 set_pte_at(mm, address, pvmw.pte, pteval);
1744 SetPageSwapBacked(page);
1746 page_vma_mapped_walk_done(&pvmw);
1750 if (swap_duplicate(entry) < 0) {
1751 set_pte_at(mm, address, pvmw.pte, pteval);
1753 page_vma_mapped_walk_done(&pvmw);
1756 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1757 set_pte_at(mm, address, pvmw.pte, pteval);
1759 page_vma_mapped_walk_done(&pvmw);
1762 if (list_empty(&mm->mmlist)) {
1763 spin_lock(&mmlist_lock);
1764 if (list_empty(&mm->mmlist))
1765 list_add(&mm->mmlist, &init_mm.mmlist);
1766 spin_unlock(&mmlist_lock);
1768 dec_mm_counter(mm, MM_ANONPAGES);
1769 inc_mm_counter(mm, MM_SWAPENTS);
1770 swp_pte = swp_entry_to_pte(entry);
1771 if (pte_soft_dirty(pteval))
1772 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1773 if (pte_uffd_wp(pteval))
1774 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1775 set_pte_at(mm, address, pvmw.pte, swp_pte);
1776 /* Invalidate as we cleared the pte */
1777 mmu_notifier_invalidate_range(mm, address,
1778 address + PAGE_SIZE);
1781 * This is a locked file-backed page, thus it cannot
1782 * be removed from the page cache and replaced by a new
1783 * page before mmu_notifier_invalidate_range_end, so no
1784 * concurrent thread might update its page table to
1785 * point at new page while a device still is using this
1788 * See Documentation/vm/mmu_notifier.rst
1790 dec_mm_counter(mm, mm_counter_file(page));
1794 * No need to call mmu_notifier_invalidate_range() it has be
1795 * done above for all cases requiring it to happen under page
1796 * table lock before mmu_notifier_invalidate_range_end()
1798 * See Documentation/vm/mmu_notifier.rst
1800 page_remove_rmap(subpage, PageHuge(page));
1804 mmu_notifier_invalidate_range_end(&range);
1809 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1811 return vma_is_temporary_stack(vma);
1814 static int page_not_mapped(struct page *page)
1816 return !page_mapped(page);
1820 * try_to_unmap - try to remove all page table mappings to a page
1821 * @page: the page to get unmapped
1822 * @flags: action and flags
1824 * Tries to remove all the page table entries which are mapping this
1825 * page, used in the pageout path. Caller must hold the page lock.
1827 * If unmap is successful, return true. Otherwise, false.
1829 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1831 struct rmap_walk_control rwc = {
1832 .rmap_one = try_to_unmap_one,
1833 .arg = (void *)flags,
1834 .done = page_not_mapped,
1835 .anon_lock = page_lock_anon_vma_read,
1839 * During exec, a temporary VMA is setup and later moved.
1840 * The VMA is moved under the anon_vma lock but not the
1841 * page tables leading to a race where migration cannot
1842 * find the migration ptes. Rather than increasing the
1843 * locking requirements of exec(), migration skips
1844 * temporary VMAs until after exec() completes.
1846 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1847 && !PageKsm(page) && PageAnon(page))
1848 rwc.invalid_vma = invalid_migration_vma;
1850 if (flags & TTU_RMAP_LOCKED)
1851 rmap_walk_locked(page, &rwc);
1853 rmap_walk(page, &rwc);
1856 * When racing against e.g. zap_pte_range() on another cpu,
1857 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1858 * try_to_unmap() may return false when it is about to become true,
1859 * if page table locking is skipped: use TTU_SYNC to wait for that.
1861 return !page_mapcount(page);
1865 * try_to_munlock - try to munlock a page
1866 * @page: the page to be munlocked
1868 * Called from munlock code. Checks all of the VMAs mapping the page
1869 * to make sure nobody else has this page mlocked. The page will be
1870 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1873 void try_to_munlock(struct page *page)
1875 struct rmap_walk_control rwc = {
1876 .rmap_one = try_to_unmap_one,
1877 .arg = (void *)TTU_MUNLOCK,
1878 .done = page_not_mapped,
1879 .anon_lock = page_lock_anon_vma_read,
1883 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1884 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1886 rmap_walk(page, &rwc);
1889 void __put_anon_vma(struct anon_vma *anon_vma)
1891 struct anon_vma *root = anon_vma->root;
1893 anon_vma_free(anon_vma);
1894 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1895 anon_vma_free(root);
1898 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1899 struct rmap_walk_control *rwc)
1901 struct anon_vma *anon_vma;
1904 return rwc->anon_lock(page);
1907 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1908 * because that depends on page_mapped(); but not all its usages
1909 * are holding mmap_lock. Users without mmap_lock are required to
1910 * take a reference count to prevent the anon_vma disappearing
1912 anon_vma = page_anon_vma(page);
1916 anon_vma_lock_read(anon_vma);
1921 * rmap_walk_anon - do something to anonymous page using the object-based
1923 * @page: the page to be handled
1924 * @rwc: control variable according to each walk type
1926 * Find all the mappings of a page using the mapping pointer and the vma chains
1927 * contained in the anon_vma struct it points to.
1929 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1930 * where the page was found will be held for write. So, we won't recheck
1931 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1934 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1937 struct anon_vma *anon_vma;
1938 pgoff_t pgoff_start, pgoff_end;
1939 struct anon_vma_chain *avc;
1942 anon_vma = page_anon_vma(page);
1943 /* anon_vma disappear under us? */
1944 VM_BUG_ON_PAGE(!anon_vma, page);
1946 anon_vma = rmap_walk_anon_lock(page, rwc);
1951 pgoff_start = page_to_pgoff(page);
1952 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
1953 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1954 pgoff_start, pgoff_end) {
1955 struct vm_area_struct *vma = avc->vma;
1956 unsigned long address = vma_address(page, vma);
1958 VM_BUG_ON_VMA(address == -EFAULT, vma);
1961 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1964 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1966 if (rwc->done && rwc->done(page))
1971 anon_vma_unlock_read(anon_vma);
1975 * rmap_walk_file - do something to file page using the object-based rmap method
1976 * @page: the page to be handled
1977 * @rwc: control variable according to each walk type
1979 * Find all the mappings of a page using the mapping pointer and the vma chains
1980 * contained in the address_space struct it points to.
1982 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1983 * where the page was found will be held for write. So, we won't recheck
1984 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1987 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1990 struct address_space *mapping = page_mapping(page);
1991 pgoff_t pgoff_start, pgoff_end;
1992 struct vm_area_struct *vma;
1995 * The page lock not only makes sure that page->mapping cannot
1996 * suddenly be NULLified by truncation, it makes sure that the
1997 * structure at mapping cannot be freed and reused yet,
1998 * so we can safely take mapping->i_mmap_rwsem.
2000 VM_BUG_ON_PAGE(!PageLocked(page), page);
2005 pgoff_start = page_to_pgoff(page);
2006 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
2008 i_mmap_lock_read(mapping);
2009 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2010 pgoff_start, pgoff_end) {
2011 unsigned long address = vma_address(page, vma);
2013 VM_BUG_ON_VMA(address == -EFAULT, vma);
2016 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2019 if (!rwc->rmap_one(page, vma, address, rwc->arg))
2021 if (rwc->done && rwc->done(page))
2027 i_mmap_unlock_read(mapping);
2030 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
2032 if (unlikely(PageKsm(page)))
2033 rmap_walk_ksm(page, rwc);
2034 else if (PageAnon(page))
2035 rmap_walk_anon(page, rwc, false);
2037 rmap_walk_file(page, rwc, false);
2040 /* Like rmap_walk, but caller holds relevant rmap lock */
2041 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
2043 /* no ksm support for now */
2044 VM_BUG_ON_PAGE(PageKsm(page), page);
2046 rmap_walk_anon(page, rwc, true);
2048 rmap_walk_file(page, rwc, true);
2051 #ifdef CONFIG_HUGETLB_PAGE
2053 * The following two functions are for anonymous (private mapped) hugepages.
2054 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2055 * and no lru code, because we handle hugepages differently from common pages.
2057 void hugepage_add_anon_rmap(struct page *page,
2058 struct vm_area_struct *vma, unsigned long address)
2060 struct anon_vma *anon_vma = vma->anon_vma;
2063 BUG_ON(!PageLocked(page));
2065 /* address might be in next vma when migration races vma_adjust */
2066 first = atomic_inc_and_test(compound_mapcount_ptr(page));
2068 __page_set_anon_rmap(page, vma, address, 0);
2071 void hugepage_add_new_anon_rmap(struct page *page,
2072 struct vm_area_struct *vma, unsigned long address)
2074 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2075 atomic_set(compound_mapcount_ptr(page), 0);
2076 if (hpage_pincount_available(page))
2077 atomic_set(compound_pincount_ptr(page), 0);
2079 __page_set_anon_rmap(page, vma, address, 1);
2081 #endif /* CONFIG_HUGETLB_PAGE */