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 * zone_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 * mapping->tree_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 * mapping->tree_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/pagemap.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/slab.h>
53 #include <linux/init.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/rcupdate.h>
57 #include <linux/export.h>
58 #include <linux/memcontrol.h>
59 #include <linux/mmu_notifier.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/backing-dev.h>
63 #include <linux/page_idle.h>
65 #include <asm/tlbflush.h>
67 #include <trace/events/tlb.h>
71 static struct kmem_cache *anon_vma_cachep;
72 static struct kmem_cache *anon_vma_chain_cachep;
74 static inline struct anon_vma *anon_vma_alloc(void)
76 struct anon_vma *anon_vma;
78 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
80 atomic_set(&anon_vma->refcount, 1);
81 anon_vma->num_children = 0;
82 anon_vma->num_active_vmas = 0;
83 anon_vma->parent = anon_vma;
85 * Initialise the anon_vma root to point to itself. If called
86 * from fork, the root will be reset to the parents anon_vma.
88 anon_vma->root = anon_vma;
94 static inline void anon_vma_free(struct anon_vma *anon_vma)
96 VM_BUG_ON(atomic_read(&anon_vma->refcount));
99 * Synchronize against page_lock_anon_vma_read() such that
100 * we can safely hold the lock without the anon_vma getting
103 * Relies on the full mb implied by the atomic_dec_and_test() from
104 * put_anon_vma() against the acquire barrier implied by
105 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
107 * page_lock_anon_vma_read() VS put_anon_vma()
108 * down_read_trylock() atomic_dec_and_test()
110 * atomic_read() rwsem_is_locked()
112 * LOCK should suffice since the actual taking of the lock must
113 * happen _before_ what follows.
116 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
117 anon_vma_lock_write(anon_vma);
118 anon_vma_unlock_write(anon_vma);
121 kmem_cache_free(anon_vma_cachep, anon_vma);
124 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
126 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
129 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
131 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
134 static void anon_vma_chain_link(struct vm_area_struct *vma,
135 struct anon_vma_chain *avc,
136 struct anon_vma *anon_vma)
139 avc->anon_vma = anon_vma;
140 list_add(&avc->same_vma, &vma->anon_vma_chain);
141 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
145 * anon_vma_prepare - attach an anon_vma to a memory region
146 * @vma: the memory region in question
148 * This makes sure the memory mapping described by 'vma' has
149 * an 'anon_vma' attached to it, so that we can associate the
150 * anonymous pages mapped into it with that anon_vma.
152 * The common case will be that we already have one, but if
153 * not we either need to find an adjacent mapping that we
154 * can re-use the anon_vma from (very common when the only
155 * reason for splitting a vma has been mprotect()), or we
156 * allocate a new one.
158 * Anon-vma allocations are very subtle, because we may have
159 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
160 * and that may actually touch the spinlock even in the newly
161 * allocated vma (it depends on RCU to make sure that the
162 * anon_vma isn't actually destroyed).
164 * As a result, we need to do proper anon_vma locking even
165 * for the new allocation. At the same time, we do not want
166 * to do any locking for the common case of already having
169 * This must be called with the mmap_sem held for reading.
171 int anon_vma_prepare(struct vm_area_struct *vma)
173 struct anon_vma *anon_vma = vma->anon_vma;
174 struct anon_vma_chain *avc;
177 if (unlikely(!anon_vma)) {
178 struct mm_struct *mm = vma->vm_mm;
179 struct anon_vma *allocated;
181 avc = anon_vma_chain_alloc(GFP_KERNEL);
185 anon_vma = find_mergeable_anon_vma(vma);
188 anon_vma = anon_vma_alloc();
189 if (unlikely(!anon_vma))
190 goto out_enomem_free_avc;
191 anon_vma->num_children++; /* self-parent link for new root */
192 allocated = anon_vma;
195 anon_vma_lock_write(anon_vma);
196 /* page_table_lock to protect against threads */
197 spin_lock(&mm->page_table_lock);
198 if (likely(!vma->anon_vma)) {
199 vma->anon_vma = anon_vma;
200 anon_vma_chain_link(vma, avc, anon_vma);
201 anon_vma->num_active_vmas++;
205 spin_unlock(&mm->page_table_lock);
206 anon_vma_unlock_write(anon_vma);
208 if (unlikely(allocated))
209 put_anon_vma(allocated);
211 anon_vma_chain_free(avc);
216 anon_vma_chain_free(avc);
222 * This is a useful helper function for locking the anon_vma root as
223 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
226 * Such anon_vma's should have the same root, so you'd expect to see
227 * just a single mutex_lock for the whole traversal.
229 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
231 struct anon_vma *new_root = anon_vma->root;
232 if (new_root != root) {
233 if (WARN_ON_ONCE(root))
234 up_write(&root->rwsem);
236 down_write(&root->rwsem);
241 static inline void unlock_anon_vma_root(struct anon_vma *root)
244 up_write(&root->rwsem);
248 * Attach the anon_vmas from src to dst.
249 * Returns 0 on success, -ENOMEM on failure.
251 * If dst->anon_vma is NULL this function tries to find and reuse existing
252 * anon_vma which has no vmas and only one child anon_vma. This prevents
253 * degradation of anon_vma hierarchy to endless linear chain in case of
254 * constantly forking task. On the other hand, an anon_vma with more than one
255 * child isn't reused even if there was no alive vma, thus rmap walker has a
256 * good chance of avoiding scanning the whole hierarchy when it searches where
259 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
261 struct anon_vma_chain *avc, *pavc;
262 struct anon_vma *root = NULL;
264 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
265 struct anon_vma *anon_vma;
267 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
268 if (unlikely(!avc)) {
269 unlock_anon_vma_root(root);
271 avc = anon_vma_chain_alloc(GFP_KERNEL);
275 anon_vma = pavc->anon_vma;
276 root = lock_anon_vma_root(root, anon_vma);
277 anon_vma_chain_link(dst, avc, anon_vma);
280 * Reuse existing anon_vma if it has no vma and only one
283 * Root anon_vma is never reused:
284 * it has self-parent reference and at least one child.
286 if (!dst->anon_vma &&
287 anon_vma->num_children < 2 &&
288 anon_vma->num_active_vmas == 0)
289 dst->anon_vma = anon_vma;
292 dst->anon_vma->num_active_vmas++;
293 unlock_anon_vma_root(root);
298 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
299 * decremented in unlink_anon_vmas().
300 * We can safely do this because callers of anon_vma_clone() don't care
301 * about dst->anon_vma if anon_vma_clone() failed.
303 dst->anon_vma = NULL;
304 unlink_anon_vmas(dst);
309 * Attach vma to its own anon_vma, as well as to the anon_vmas that
310 * the corresponding VMA in the parent process is attached to.
311 * Returns 0 on success, non-zero on failure.
313 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
315 struct anon_vma_chain *avc;
316 struct anon_vma *anon_vma;
319 /* Don't bother if the parent process has no anon_vma here. */
323 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
324 vma->anon_vma = NULL;
327 * First, attach the new VMA to the parent VMA's anon_vmas,
328 * so rmap can find non-COWed pages in child processes.
330 error = anon_vma_clone(vma, pvma);
334 /* An existing anon_vma has been reused, all done then. */
338 /* Then add our own anon_vma. */
339 anon_vma = anon_vma_alloc();
342 anon_vma->num_active_vmas++;
343 avc = anon_vma_chain_alloc(GFP_KERNEL);
345 goto out_error_free_anon_vma;
348 * The root anon_vma's spinlock is the lock actually used when we
349 * lock any of the anon_vmas in this anon_vma tree.
351 anon_vma->root = pvma->anon_vma->root;
352 anon_vma->parent = pvma->anon_vma;
354 * With refcounts, an anon_vma can stay around longer than the
355 * process it belongs to. The root anon_vma needs to be pinned until
356 * this anon_vma is freed, because the lock lives in the root.
358 get_anon_vma(anon_vma->root);
359 /* Mark this anon_vma as the one where our new (COWed) pages go. */
360 vma->anon_vma = anon_vma;
361 anon_vma_lock_write(anon_vma);
362 anon_vma_chain_link(vma, avc, anon_vma);
363 anon_vma->parent->num_children++;
364 anon_vma_unlock_write(anon_vma);
368 out_error_free_anon_vma:
369 put_anon_vma(anon_vma);
371 unlink_anon_vmas(vma);
375 void unlink_anon_vmas(struct vm_area_struct *vma)
377 struct anon_vma_chain *avc, *next;
378 struct anon_vma *root = NULL;
381 * Unlink each anon_vma chained to the VMA. This list is ordered
382 * from newest to oldest, ensuring the root anon_vma gets freed last.
384 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
385 struct anon_vma *anon_vma = avc->anon_vma;
387 root = lock_anon_vma_root(root, anon_vma);
388 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
391 * Leave empty anon_vmas on the list - we'll need
392 * to free them outside the lock.
394 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
395 anon_vma->parent->num_children--;
399 list_del(&avc->same_vma);
400 anon_vma_chain_free(avc);
403 vma->anon_vma->num_active_vmas--;
404 unlock_anon_vma_root(root);
407 * Iterate the list once more, it now only contains empty and unlinked
408 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
409 * needing to write-acquire the anon_vma->root->rwsem.
411 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
412 struct anon_vma *anon_vma = avc->anon_vma;
414 VM_WARN_ON(anon_vma->num_children);
415 VM_WARN_ON(anon_vma->num_active_vmas);
416 put_anon_vma(anon_vma);
418 list_del(&avc->same_vma);
419 anon_vma_chain_free(avc);
423 static void anon_vma_ctor(void *data)
425 struct anon_vma *anon_vma = data;
427 init_rwsem(&anon_vma->rwsem);
428 atomic_set(&anon_vma->refcount, 0);
429 anon_vma->rb_root = RB_ROOT;
432 void __init anon_vma_init(void)
434 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
435 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
437 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
438 SLAB_PANIC|SLAB_ACCOUNT);
442 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
444 * Since there is no serialization what so ever against page_remove_rmap()
445 * the best this function can do is return a locked anon_vma that might
446 * have been relevant to this page.
448 * The page might have been remapped to a different anon_vma or the anon_vma
449 * returned may already be freed (and even reused).
451 * In case it was remapped to a different anon_vma, the new anon_vma will be a
452 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
453 * ensure that any anon_vma obtained from the page will still be valid for as
454 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
456 * All users of this function must be very careful when walking the anon_vma
457 * chain and verify that the page in question is indeed mapped in it
458 * [ something equivalent to page_mapped_in_vma() ].
460 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
461 * that the anon_vma pointer from page->mapping is valid if there is a
462 * mapcount, we can dereference the anon_vma after observing those.
464 struct anon_vma *page_get_anon_vma(struct page *page)
466 struct anon_vma *anon_vma = NULL;
467 unsigned long anon_mapping;
470 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
471 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
473 if (!page_mapped(page))
476 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
477 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
483 * If this page is still mapped, then its anon_vma cannot have been
484 * freed. But if it has been unmapped, we have no security against the
485 * anon_vma structure being freed and reused (for another anon_vma:
486 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
487 * above cannot corrupt).
489 if (!page_mapped(page)) {
491 put_anon_vma(anon_vma);
501 * Similar to page_get_anon_vma() except it locks the anon_vma.
503 * Its a little more complex as it tries to keep the fast path to a single
504 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
505 * reference like with page_get_anon_vma() and then block on the mutex.
507 struct anon_vma *page_lock_anon_vma_read(struct page *page)
509 struct anon_vma *anon_vma = NULL;
510 struct anon_vma *root_anon_vma;
511 unsigned long anon_mapping;
514 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
515 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
517 if (!page_mapped(page))
520 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
521 root_anon_vma = READ_ONCE(anon_vma->root);
522 if (down_read_trylock(&root_anon_vma->rwsem)) {
524 * If the page is still mapped, then this anon_vma is still
525 * its anon_vma, and holding the mutex ensures that it will
526 * not go away, see anon_vma_free().
528 if (!page_mapped(page)) {
529 up_read(&root_anon_vma->rwsem);
535 /* trylock failed, we got to sleep */
536 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
541 if (!page_mapped(page)) {
543 put_anon_vma(anon_vma);
547 /* we pinned the anon_vma, its safe to sleep */
549 anon_vma_lock_read(anon_vma);
551 if (atomic_dec_and_test(&anon_vma->refcount)) {
553 * Oops, we held the last refcount, release the lock
554 * and bail -- can't simply use put_anon_vma() because
555 * we'll deadlock on the anon_vma_lock_write() recursion.
557 anon_vma_unlock_read(anon_vma);
558 __put_anon_vma(anon_vma);
569 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
571 anon_vma_unlock_read(anon_vma);
574 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
576 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
577 * important if a PTE was dirty when it was unmapped that it's flushed
578 * before any IO is initiated on the page to prevent lost writes. Similarly,
579 * it must be flushed before freeing to prevent data leakage.
581 void try_to_unmap_flush(void)
583 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
586 if (!tlb_ubc->flush_required)
591 if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
592 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
594 trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
597 if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
598 flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
599 cpumask_clear(&tlb_ubc->cpumask);
600 tlb_ubc->flush_required = false;
601 tlb_ubc->writable = false;
605 /* Flush iff there are potentially writable TLB entries that can race with IO */
606 void try_to_unmap_flush_dirty(void)
608 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
610 if (tlb_ubc->writable)
611 try_to_unmap_flush();
614 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
615 struct page *page, bool writable)
617 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
619 cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
620 tlb_ubc->flush_required = true;
623 * Ensure compiler does not re-order the setting of tlb_flush_batched
624 * before the PTE is cleared.
627 mm->tlb_flush_batched = true;
630 * If the PTE was dirty then it's best to assume it's writable. The
631 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
632 * before the page is queued for IO.
635 tlb_ubc->writable = true;
639 * Returns true if the TLB flush should be deferred to the end of a batch of
640 * unmap operations to reduce IPIs.
642 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
644 bool should_defer = false;
646 if (!(flags & TTU_BATCH_FLUSH))
649 /* If remote CPUs need to be flushed then defer batch the flush */
650 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
658 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
659 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
660 * operation such as mprotect or munmap to race between reclaim unmapping
661 * the page and flushing the page. If this race occurs, it potentially allows
662 * access to data via a stale TLB entry. Tracking all mm's that have TLB
663 * batching in flight would be expensive during reclaim so instead track
664 * whether TLB batching occurred in the past and if so then do a flush here
665 * if required. This will cost one additional flush per reclaim cycle paid
666 * by the first operation at risk such as mprotect and mumap.
668 * This must be called under the PTL so that an access to tlb_flush_batched
669 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
672 void flush_tlb_batched_pending(struct mm_struct *mm)
674 if (mm->tlb_flush_batched) {
678 * Do not allow the compiler to re-order the clearing of
679 * tlb_flush_batched before the tlb is flushed.
682 mm->tlb_flush_batched = false;
686 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
687 struct page *page, 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 unsigned long address;
704 if (PageAnon(page)) {
705 struct anon_vma *page__anon_vma = page_anon_vma(page);
707 * Note: swapoff's unuse_vma() is more efficient with this
708 * check, and needs it to match anon_vma when KSM is active.
710 if (!vma->anon_vma || !page__anon_vma ||
711 vma->anon_vma->root != page__anon_vma->root)
713 } else if (page->mapping) {
714 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
718 address = __vma_address(page, vma);
719 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
724 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
731 pgd = pgd_offset(mm, address);
732 if (!pgd_present(*pgd))
735 pud = pud_offset(pgd, address);
736 if (!pud_present(*pud))
739 pmd = pmd_offset(pud, address);
741 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
742 * without holding anon_vma lock for write. So when looking for a
743 * genuine pmde (in which to find pte), test present and !THP together.
747 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
754 * Check that @page is mapped at @address into @mm.
756 * If @sync is false, page_check_address may perform a racy check to avoid
757 * the page table lock when the pte is not present (helpful when reclaiming
758 * highly shared pages).
760 * On success returns with pte mapped and locked.
762 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
763 unsigned long address, spinlock_t **ptlp, int sync)
769 if (unlikely(PageHuge(page))) {
770 /* when pud is not present, pte will be NULL */
771 pte = huge_pte_offset(mm, address);
775 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
779 pmd = mm_find_pmd(mm, address);
783 pte = pte_offset_map(pmd, address);
784 /* Make a quick check before getting the lock */
785 if (!sync && !pte_present(*pte)) {
790 ptl = pte_lockptr(mm, pmd);
793 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
797 pte_unmap_unlock(pte, ptl);
802 * page_mapped_in_vma - check whether a page is really mapped in a VMA
803 * @page: the page to test
804 * @vma: the VMA to test
806 * Returns 1 if the page is mapped into the page tables of the VMA, 0
807 * if the page is not mapped into the page tables of this VMA. Only
808 * valid for normal file or anonymous VMAs.
810 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
812 unsigned long address;
816 address = __vma_address(page, vma);
817 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
819 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
820 if (!pte) /* the page is not in this mm */
822 pte_unmap_unlock(pte, ptl);
827 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
829 * Check that @page is mapped at @address into @mm. In contrast to
830 * page_check_address(), this function can handle transparent huge pages.
832 * On success returns true with pte mapped and locked. For PMD-mapped
833 * transparent huge pages *@ptep is set to NULL.
835 bool page_check_address_transhuge(struct page *page, struct mm_struct *mm,
836 unsigned long address, pmd_t **pmdp,
837 pte_t **ptep, spinlock_t **ptlp)
845 if (unlikely(PageHuge(page))) {
846 /* when pud is not present, pte will be NULL */
847 pte = huge_pte_offset(mm, address);
851 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
856 pgd = pgd_offset(mm, address);
857 if (!pgd_present(*pgd))
859 pud = pud_offset(pgd, address);
860 if (!pud_present(*pud))
862 pmd = pmd_offset(pud, address);
864 if (pmd_trans_huge(*pmd)) {
865 ptl = pmd_lock(mm, pmd);
866 if (!pmd_present(*pmd))
868 if (unlikely(!pmd_trans_huge(*pmd))) {
873 if (pmd_page(*pmd) != page)
885 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
889 pte = pte_offset_map(pmd, address);
890 if (!pte_present(*pte)) {
895 ptl = pte_lockptr(mm, pmd);
899 if (!pte_present(*pte)) {
900 pte_unmap_unlock(pte, ptl);
904 /* THP can be referenced by any subpage */
905 if (pte_pfn(*pte) - page_to_pfn(page) >= hpage_nr_pages(page)) {
906 pte_unmap_unlock(pte, ptl);
915 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
917 struct page_referenced_arg {
920 unsigned long vm_flags;
921 struct mem_cgroup *memcg;
924 * arg: page_referenced_arg will be passed
926 static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
927 unsigned long address, void *arg)
929 struct mm_struct *mm = vma->vm_mm;
930 struct page_referenced_arg *pra = arg;
936 if (!page_check_address_transhuge(page, mm, address, &pmd, &pte, &ptl))
939 if (vma->vm_flags & VM_LOCKED) {
943 pra->vm_flags |= VM_LOCKED;
944 return SWAP_FAIL; /* To break the loop */
948 if (ptep_clear_flush_young_notify(vma, address, pte)) {
950 * Don't treat a reference through a sequentially read
951 * mapping as such. If the page has been used in
952 * another mapping, we will catch it; if this other
953 * mapping is already gone, the unmap path will have
954 * set PG_referenced or activated the page.
956 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
960 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
961 if (pmdp_clear_flush_young_notify(vma, address, pmd))
964 /* unexpected pmd-mapped page? */
970 clear_page_idle(page);
971 if (test_and_clear_page_young(page))
976 pra->vm_flags |= vma->vm_flags;
981 return SWAP_SUCCESS; /* To break the loop */
986 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
988 struct page_referenced_arg *pra = arg;
989 struct mem_cgroup *memcg = pra->memcg;
991 if (!mm_match_cgroup(vma->vm_mm, memcg))
998 * page_referenced - test if the page was referenced
999 * @page: the page to test
1000 * @is_locked: caller holds lock on the page
1001 * @memcg: target memory cgroup
1002 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
1004 * Quick test_and_clear_referenced for all mappings to a page,
1005 * returns the number of ptes which referenced the page.
1007 int page_referenced(struct page *page,
1009 struct mem_cgroup *memcg,
1010 unsigned long *vm_flags)
1014 struct page_referenced_arg pra = {
1015 .mapcount = total_mapcount(page),
1018 struct rmap_walk_control rwc = {
1019 .rmap_one = page_referenced_one,
1020 .arg = (void *)&pra,
1021 .anon_lock = page_lock_anon_vma_read,
1025 if (!page_mapped(page))
1028 if (!page_rmapping(page))
1031 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
1032 we_locked = trylock_page(page);
1038 * If we are reclaiming on behalf of a cgroup, skip
1039 * counting on behalf of references from different
1043 rwc.invalid_vma = invalid_page_referenced_vma;
1046 ret = rmap_walk(page, &rwc);
1047 *vm_flags = pra.vm_flags;
1052 return pra.referenced;
1055 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
1056 unsigned long address, void *arg)
1058 struct mm_struct *mm = vma->vm_mm;
1064 pte = page_check_address(page, mm, address, &ptl, 1);
1068 if (pte_dirty(*pte) || pte_write(*pte)) {
1071 flush_cache_page(vma, address, pte_pfn(*pte));
1072 entry = ptep_clear_flush(vma, address, pte);
1073 entry = pte_wrprotect(entry);
1074 entry = pte_mkclean(entry);
1075 set_pte_at(mm, address, pte, entry);
1079 pte_unmap_unlock(pte, ptl);
1082 mmu_notifier_invalidate_page(mm, address);
1089 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1091 if (vma->vm_flags & VM_SHARED)
1097 int page_mkclean(struct page *page)
1100 struct address_space *mapping;
1101 struct rmap_walk_control rwc = {
1102 .arg = (void *)&cleaned,
1103 .rmap_one = page_mkclean_one,
1104 .invalid_vma = invalid_mkclean_vma,
1107 BUG_ON(!PageLocked(page));
1109 if (!page_mapped(page))
1112 mapping = page_mapping(page);
1116 rmap_walk(page, &rwc);
1120 EXPORT_SYMBOL_GPL(page_mkclean);
1123 * page_move_anon_rmap - move a page to our anon_vma
1124 * @page: the page to move to our anon_vma
1125 * @vma: the vma the page belongs to
1127 * When a page belongs exclusively to one process after a COW event,
1128 * that page can be moved into the anon_vma that belongs to just that
1129 * process, so the rmap code will not search the parent or sibling
1132 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1134 struct anon_vma *anon_vma = vma->anon_vma;
1136 page = compound_head(page);
1138 VM_BUG_ON_PAGE(!PageLocked(page), page);
1139 VM_BUG_ON_VMA(!anon_vma, vma);
1141 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1143 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1144 * simultaneously, so a concurrent reader (eg page_referenced()'s
1145 * PageAnon()) will not see one without the other.
1147 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1151 * __page_set_anon_rmap - set up new anonymous rmap
1152 * @page: Page to add to rmap
1153 * @vma: VM area to add page to.
1154 * @address: User virtual address of the mapping
1155 * @exclusive: the page is exclusively owned by the current process
1157 static void __page_set_anon_rmap(struct page *page,
1158 struct vm_area_struct *vma, unsigned long address, int exclusive)
1160 struct anon_vma *anon_vma = vma->anon_vma;
1168 * If the page isn't exclusively mapped into this vma,
1169 * we must use the _oldest_ possible anon_vma for the
1173 anon_vma = anon_vma->root;
1175 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1176 page->mapping = (struct address_space *) anon_vma;
1177 page->index = linear_page_index(vma, address);
1181 * __page_check_anon_rmap - sanity check anonymous rmap addition
1182 * @page: the page to add the mapping to
1183 * @vma: the vm area in which the mapping is added
1184 * @address: the user virtual address mapped
1186 static void __page_check_anon_rmap(struct page *page,
1187 struct vm_area_struct *vma, unsigned long address)
1189 #ifdef CONFIG_DEBUG_VM
1191 * The page's anon-rmap details (mapping and index) are guaranteed to
1192 * be set up correctly at this point.
1194 * We have exclusion against page_add_anon_rmap because the caller
1195 * always holds the page locked, except if called from page_dup_rmap,
1196 * in which case the page is already known to be setup.
1198 * We have exclusion against page_add_new_anon_rmap because those pages
1199 * are initially only visible via the pagetables, and the pte is locked
1200 * over the call to page_add_new_anon_rmap.
1202 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1203 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1208 * page_add_anon_rmap - add pte mapping to an anonymous page
1209 * @page: the page to add the mapping to
1210 * @vma: the vm area in which the mapping is added
1211 * @address: the user virtual address mapped
1212 * @compound: charge the page as compound or small page
1214 * The caller needs to hold the pte lock, and the page must be locked in
1215 * the anon_vma case: to serialize mapping,index checking after setting,
1216 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1217 * (but PageKsm is never downgraded to PageAnon).
1219 void page_add_anon_rmap(struct page *page,
1220 struct vm_area_struct *vma, unsigned long address, bool compound)
1222 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1226 * Special version of the above for do_swap_page, which often runs
1227 * into pages that are exclusively owned by the current process.
1228 * Everybody else should continue to use page_add_anon_rmap above.
1230 void do_page_add_anon_rmap(struct page *page,
1231 struct vm_area_struct *vma, unsigned long address, int flags)
1233 bool compound = flags & RMAP_COMPOUND;
1238 VM_BUG_ON_PAGE(!PageLocked(page), page);
1239 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1240 mapcount = compound_mapcount_ptr(page);
1241 first = atomic_inc_and_test(mapcount);
1243 first = atomic_inc_and_test(&page->_mapcount);
1247 int nr = compound ? hpage_nr_pages(page) : 1;
1249 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1250 * these counters are not modified in interrupt context, and
1251 * pte lock(a spinlock) is held, which implies preemption
1255 __inc_node_page_state(page, NR_ANON_THPS);
1256 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1258 if (unlikely(PageKsm(page)))
1261 VM_BUG_ON_PAGE(!PageLocked(page), page);
1263 /* address might be in next vma when migration races vma_adjust */
1265 __page_set_anon_rmap(page, vma, address,
1266 flags & RMAP_EXCLUSIVE);
1268 __page_check_anon_rmap(page, vma, address);
1272 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1273 * @page: the page to add the mapping to
1274 * @vma: the vm area in which the mapping is added
1275 * @address: the user virtual address mapped
1276 * @compound: charge the page as compound or small page
1278 * Same as page_add_anon_rmap but must only be called on *new* pages.
1279 * This means the inc-and-test can be bypassed.
1280 * Page does not have to be locked.
1282 void page_add_new_anon_rmap(struct page *page,
1283 struct vm_area_struct *vma, unsigned long address, bool compound)
1285 int nr = compound ? hpage_nr_pages(page) : 1;
1287 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1288 __SetPageSwapBacked(page);
1290 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1291 /* increment count (starts at -1) */
1292 atomic_set(compound_mapcount_ptr(page), 0);
1293 __inc_node_page_state(page, NR_ANON_THPS);
1295 /* Anon THP always mapped first with PMD */
1296 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1297 /* increment count (starts at -1) */
1298 atomic_set(&page->_mapcount, 0);
1300 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1301 __page_set_anon_rmap(page, vma, address, 1);
1305 * page_add_file_rmap - add pte mapping to a file page
1306 * @page: the page to add the mapping to
1308 * The caller needs to hold the pte lock.
1310 void page_add_file_rmap(struct page *page, bool compound)
1314 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1315 lock_page_memcg(page);
1316 if (compound && PageTransHuge(page)) {
1317 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1318 if (atomic_inc_and_test(&page[i]._mapcount))
1321 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1323 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1324 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1326 if (PageTransCompound(page) && page_mapping(page)) {
1327 VM_WARN_ON_ONCE(!PageLocked(page));
1329 SetPageDoubleMap(compound_head(page));
1330 if (PageMlocked(page))
1331 clear_page_mlock(compound_head(page));
1333 if (!atomic_inc_and_test(&page->_mapcount))
1336 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, nr);
1337 mem_cgroup_update_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, nr);
1339 unlock_page_memcg(page);
1342 static void page_remove_file_rmap(struct page *page, bool compound)
1346 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1347 lock_page_memcg(page);
1349 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1350 if (unlikely(PageHuge(page))) {
1351 /* hugetlb pages are always mapped with pmds */
1352 atomic_dec(compound_mapcount_ptr(page));
1356 /* page still mapped by someone else? */
1357 if (compound && PageTransHuge(page)) {
1358 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1359 if (atomic_add_negative(-1, &page[i]._mapcount))
1362 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1364 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1365 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1367 if (!atomic_add_negative(-1, &page->_mapcount))
1372 * We use the irq-unsafe __{inc|mod}_zone_page_state because
1373 * these counters are not modified in interrupt context, and
1374 * pte lock(a spinlock) is held, which implies preemption disabled.
1376 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, -nr);
1377 mem_cgroup_update_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, -nr);
1379 if (unlikely(PageMlocked(page)))
1380 clear_page_mlock(page);
1382 unlock_page_memcg(page);
1385 static void page_remove_anon_compound_rmap(struct page *page)
1389 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1392 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1393 if (unlikely(PageHuge(page)))
1396 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1399 __dec_node_page_state(page, NR_ANON_THPS);
1401 if (TestClearPageDoubleMap(page)) {
1403 * Subpages can be mapped with PTEs too. Check how many of
1404 * themi are still mapped.
1406 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1407 if (atomic_add_negative(-1, &page[i]._mapcount))
1414 if (unlikely(PageMlocked(page)))
1415 clear_page_mlock(page);
1418 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1419 deferred_split_huge_page(page);
1424 * page_remove_rmap - take down pte mapping from a page
1425 * @page: page to remove mapping from
1426 * @compound: uncharge the page as compound or small page
1428 * The caller needs to hold the pte lock.
1430 void page_remove_rmap(struct page *page, bool compound)
1432 if (!PageAnon(page))
1433 return page_remove_file_rmap(page, compound);
1436 return page_remove_anon_compound_rmap(page);
1438 /* page still mapped by someone else? */
1439 if (!atomic_add_negative(-1, &page->_mapcount))
1443 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1444 * these counters are not modified in interrupt context, and
1445 * pte lock(a spinlock) is held, which implies preemption disabled.
1447 __dec_node_page_state(page, NR_ANON_MAPPED);
1449 if (unlikely(PageMlocked(page)))
1450 clear_page_mlock(page);
1452 if (PageTransCompound(page))
1453 deferred_split_huge_page(compound_head(page));
1456 * It would be tidy to reset the PageAnon mapping here,
1457 * but that might overwrite a racing page_add_anon_rmap
1458 * which increments mapcount after us but sets mapping
1459 * before us: so leave the reset to free_hot_cold_page,
1460 * and remember that it's only reliable while mapped.
1461 * Leaving it set also helps swapoff to reinstate ptes
1462 * faster for those pages still in swapcache.
1466 struct rmap_private {
1467 enum ttu_flags flags;
1472 * @arg: enum ttu_flags will be passed to this argument
1474 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1475 unsigned long address, void *arg)
1477 struct mm_struct *mm = vma->vm_mm;
1481 int ret = SWAP_AGAIN;
1482 unsigned long sh_address;
1483 bool pmd_sharing_possible = false;
1484 unsigned long spmd_start, spmd_end;
1485 struct rmap_private *rp = arg;
1486 enum ttu_flags flags = rp->flags;
1488 /* munlock has nothing to gain from examining un-locked vmas */
1489 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1492 if (flags & TTU_SPLIT_HUGE_PMD) {
1493 split_huge_pmd_address(vma, address,
1494 flags & TTU_MIGRATION, page);
1495 /* check if we have anything to do after split */
1496 if (page_mapcount(page) == 0)
1501 * Only use the range_start/end mmu notifiers if huge pmd sharing
1502 * is possible. In the normal case, mmu_notifier_invalidate_page
1503 * is sufficient as we only unmap a page. However, if we unshare
1504 * a pmd, we will unmap a PUD_SIZE range.
1506 if (PageHuge(page)) {
1507 spmd_start = address;
1508 spmd_end = spmd_start + vma_mmu_pagesize(vma);
1511 * Check if pmd sharing is possible. If possible, we could
1512 * unmap a PUD_SIZE range. spmd_start/spmd_end will be
1513 * modified if sharing is possible.
1515 adjust_range_if_pmd_sharing_possible(vma, &spmd_start,
1517 if (spmd_end - spmd_start != vma_mmu_pagesize(vma)) {
1518 sh_address = address;
1520 pmd_sharing_possible = true;
1521 mmu_notifier_invalidate_range_start(vma->vm_mm,
1522 spmd_start, spmd_end);
1526 pte = page_check_address(page, mm, address, &ptl,
1527 PageTransCompound(page));
1532 * If the page is mlock()d, we cannot swap it out.
1533 * If it's recently referenced (perhaps page_referenced
1534 * skipped over this mm) then we should reactivate it.
1536 if (!(flags & TTU_IGNORE_MLOCK)) {
1537 if (vma->vm_flags & VM_LOCKED) {
1538 /* PTE-mapped THP are never mlocked */
1539 if (!PageTransCompound(page)) {
1541 * Holding pte lock, we do *not* need
1544 mlock_vma_page(page);
1549 if (flags & TTU_MUNLOCK)
1552 if (!(flags & TTU_IGNORE_ACCESS)) {
1553 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1560 * Call huge_pmd_unshare to potentially unshare a huge pmd. Pass
1561 * sh_address as it will be modified if unsharing is successful.
1563 if (PageHuge(page) && huge_pmd_unshare(mm, &sh_address, pte)) {
1565 * huge_pmd_unshare unmapped an entire PMD page. There is
1566 * no way of knowing exactly which PMDs may be cached for
1567 * this mm, so flush them all. spmd_start/spmd_end cover
1568 * this PUD_SIZE range.
1570 flush_cache_range(vma, spmd_start, spmd_end);
1571 flush_tlb_range(vma, spmd_start, spmd_end);
1574 * The ref count of the PMD page was dropped which is part
1575 * of the way map counting is done for shared PMDs. When
1576 * there is no other sharing, huge_pmd_unshare returns false
1577 * and we will unmap the actual page and drop map count
1583 /* Nuke the page table entry. */
1584 flush_cache_page(vma, address, page_to_pfn(page));
1585 if (should_defer_flush(mm, flags)) {
1587 * We clear the PTE but do not flush so potentially a remote
1588 * CPU could still be writing to the page. If the entry was
1589 * previously clean then the architecture must guarantee that
1590 * a clear->dirty transition on a cached TLB entry is written
1591 * through and traps if the PTE is unmapped.
1593 pteval = ptep_get_and_clear(mm, address, pte);
1595 set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
1597 pteval = ptep_clear_flush(vma, address, pte);
1600 /* Move the dirty bit to the physical page now the pte is gone. */
1601 if (pte_dirty(pteval))
1602 set_page_dirty(page);
1604 /* Update high watermark before we lower rss */
1605 update_hiwater_rss(mm);
1607 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1608 if (PageHuge(page)) {
1609 hugetlb_count_sub(1 << compound_order(page), mm);
1611 dec_mm_counter(mm, mm_counter(page));
1613 set_pte_at(mm, address, pte,
1614 swp_entry_to_pte(make_hwpoison_entry(page)));
1615 } else if (pte_unused(pteval)) {
1617 * The guest indicated that the page content is of no
1618 * interest anymore. Simply discard the pte, vmscan
1619 * will take care of the rest.
1621 dec_mm_counter(mm, mm_counter(page));
1622 } else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) {
1626 * Store the pfn of the page in a special migration
1627 * pte. do_swap_page() will wait until the migration
1628 * pte is removed and then restart fault handling.
1630 entry = make_migration_entry(page, pte_write(pteval));
1631 swp_pte = swp_entry_to_pte(entry);
1632 if (pte_soft_dirty(pteval))
1633 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1634 set_pte_at(mm, address, pte, swp_pte);
1635 } else if (PageAnon(page)) {
1636 swp_entry_t entry = { .val = page_private(page) };
1639 * Store the swap location in the pte.
1640 * See handle_pte_fault() ...
1642 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
1644 if (flags & TTU_LZFREE) {
1645 int ref_count, map_count;
1648 * Synchronize with gup_pte_range():
1649 * - clear PTE; barrier; read refcount
1650 * - inc refcount; barrier; read PTE
1654 ref_count = page_ref_count(page);
1655 map_count = page_mapcount(page);
1658 * Order reads for page refcount and dirty flag
1659 * (see comments in __remove_mapping()).
1664 * The only page refs must be one from isolation
1665 * plus the rmap(s) (dropped by discard:).
1667 if (ref_count == 1 + map_count &&
1669 /* It's a freeable page by MADV_FREE */
1670 dec_mm_counter(mm, MM_ANONPAGES);
1676 if (swap_duplicate(entry) < 0) {
1677 set_pte_at(mm, address, pte, pteval);
1681 if (list_empty(&mm->mmlist)) {
1682 spin_lock(&mmlist_lock);
1683 if (list_empty(&mm->mmlist))
1684 list_add(&mm->mmlist, &init_mm.mmlist);
1685 spin_unlock(&mmlist_lock);
1687 dec_mm_counter(mm, MM_ANONPAGES);
1688 inc_mm_counter(mm, MM_SWAPENTS);
1689 swp_pte = swp_entry_to_pte(entry);
1690 if (pte_soft_dirty(pteval))
1691 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1692 set_pte_at(mm, address, pte, swp_pte);
1694 dec_mm_counter(mm, mm_counter_file(page));
1697 page_remove_rmap(page, PageHuge(page));
1701 pte_unmap_unlock(pte, ptl);
1702 if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
1703 mmu_notifier_invalidate_page(mm, address);
1705 if (pmd_sharing_possible)
1706 mmu_notifier_invalidate_range_end(vma->vm_mm,
1707 spmd_start, spmd_end);
1711 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1713 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1718 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1719 VM_STACK_INCOMPLETE_SETUP)
1725 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1727 return is_vma_temporary_stack(vma);
1730 static int page_mapcount_is_zero(struct page *page)
1732 return !page_mapcount(page);
1736 * try_to_unmap - try to remove all page table mappings to a page
1737 * @page: the page to get unmapped
1738 * @flags: action and flags
1740 * Tries to remove all the page table entries which are mapping this
1741 * page, used in the pageout path. Caller must hold the page lock.
1742 * Return values are:
1744 * SWAP_SUCCESS - we succeeded in removing all mappings
1745 * SWAP_AGAIN - we missed a mapping, try again later
1746 * SWAP_FAIL - the page is unswappable
1747 * SWAP_MLOCK - page is mlocked.
1749 int try_to_unmap(struct page *page, enum ttu_flags flags)
1752 struct rmap_private rp = {
1757 struct rmap_walk_control rwc = {
1758 .rmap_one = try_to_unmap_one,
1760 .done = page_mapcount_is_zero,
1761 .anon_lock = page_lock_anon_vma_read,
1765 * During exec, a temporary VMA is setup and later moved.
1766 * The VMA is moved under the anon_vma lock but not the
1767 * page tables leading to a race where migration cannot
1768 * find the migration ptes. Rather than increasing the
1769 * locking requirements of exec(), migration skips
1770 * temporary VMAs until after exec() completes.
1772 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1773 rwc.invalid_vma = invalid_migration_vma;
1775 if (flags & TTU_RMAP_LOCKED)
1776 ret = rmap_walk_locked(page, &rwc);
1778 ret = rmap_walk(page, &rwc);
1780 if (ret != SWAP_MLOCK && !page_mapcount(page)) {
1782 if (rp.lazyfreed && !PageDirty(page))
1788 static int page_not_mapped(struct page *page)
1790 return !page_mapped(page);
1794 * try_to_munlock - try to munlock a page
1795 * @page: the page to be munlocked
1797 * Called from munlock code. Checks all of the VMAs mapping the page
1798 * to make sure nobody else has this page mlocked. The page will be
1799 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1801 * Return values are:
1803 * SWAP_AGAIN - no vma is holding page mlocked, or,
1804 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1805 * SWAP_FAIL - page cannot be located at present
1806 * SWAP_MLOCK - page is now mlocked.
1808 int try_to_munlock(struct page *page)
1811 struct rmap_private rp = {
1812 .flags = TTU_MUNLOCK,
1816 struct rmap_walk_control rwc = {
1817 .rmap_one = try_to_unmap_one,
1819 .done = page_not_mapped,
1820 .anon_lock = page_lock_anon_vma_read,
1824 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1826 ret = rmap_walk(page, &rwc);
1830 void __put_anon_vma(struct anon_vma *anon_vma)
1832 struct anon_vma *root = anon_vma->root;
1834 anon_vma_free(anon_vma);
1835 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1836 anon_vma_free(root);
1839 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1840 struct rmap_walk_control *rwc)
1842 struct anon_vma *anon_vma;
1845 return rwc->anon_lock(page);
1848 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1849 * because that depends on page_mapped(); but not all its usages
1850 * are holding mmap_sem. Users without mmap_sem are required to
1851 * take a reference count to prevent the anon_vma disappearing
1853 anon_vma = page_anon_vma(page);
1857 anon_vma_lock_read(anon_vma);
1862 * rmap_walk_anon - do something to anonymous page using the object-based
1864 * @page: the page to be handled
1865 * @rwc: control variable according to each walk type
1867 * Find all the mappings of a page using the mapping pointer and the vma chains
1868 * contained in the anon_vma struct it points to.
1870 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1871 * where the page was found will be held for write. So, we won't recheck
1872 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1875 static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1878 struct anon_vma *anon_vma;
1880 struct anon_vma_chain *avc;
1881 int ret = SWAP_AGAIN;
1884 anon_vma = page_anon_vma(page);
1885 /* anon_vma disappear under us? */
1886 VM_BUG_ON_PAGE(!anon_vma, page);
1888 anon_vma = rmap_walk_anon_lock(page, rwc);
1893 pgoff = page_to_pgoff(page);
1894 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1895 struct vm_area_struct *vma = avc->vma;
1896 unsigned long address = vma_address(page, vma);
1900 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1903 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1904 if (ret != SWAP_AGAIN)
1906 if (rwc->done && rwc->done(page))
1911 anon_vma_unlock_read(anon_vma);
1916 * rmap_walk_file - do something to file page using the object-based rmap method
1917 * @page: the page to be handled
1918 * @rwc: control variable according to each walk type
1920 * Find all the mappings of a page using the mapping pointer and the vma chains
1921 * contained in the address_space struct it points to.
1923 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1924 * where the page was found will be held for write. So, we won't recheck
1925 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1928 static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1931 struct address_space *mapping = page_mapping(page);
1933 struct vm_area_struct *vma;
1934 int ret = SWAP_AGAIN;
1937 * The page lock not only makes sure that page->mapping cannot
1938 * suddenly be NULLified by truncation, it makes sure that the
1939 * structure at mapping cannot be freed and reused yet,
1940 * so we can safely take mapping->i_mmap_rwsem.
1942 VM_BUG_ON_PAGE(!PageLocked(page), page);
1947 pgoff = page_to_pgoff(page);
1949 i_mmap_lock_read(mapping);
1950 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1951 unsigned long address = vma_address(page, vma);
1955 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1958 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1959 if (ret != SWAP_AGAIN)
1961 if (rwc->done && rwc->done(page))
1967 i_mmap_unlock_read(mapping);
1971 int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1973 if (unlikely(PageKsm(page)))
1974 return rmap_walk_ksm(page, rwc);
1975 else if (PageAnon(page))
1976 return rmap_walk_anon(page, rwc, false);
1978 return rmap_walk_file(page, rwc, false);
1981 /* Like rmap_walk, but caller holds relevant rmap lock */
1982 int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1984 /* no ksm support for now */
1985 VM_BUG_ON_PAGE(PageKsm(page), page);
1987 return rmap_walk_anon(page, rwc, true);
1989 return rmap_walk_file(page, rwc, true);
1992 #ifdef CONFIG_HUGETLB_PAGE
1994 * The following three functions are for anonymous (private mapped) hugepages.
1995 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1996 * and no lru code, because we handle hugepages differently from common pages.
1998 static void __hugepage_set_anon_rmap(struct page *page,
1999 struct vm_area_struct *vma, unsigned long address, int exclusive)
2001 struct anon_vma *anon_vma = vma->anon_vma;
2008 anon_vma = anon_vma->root;
2010 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
2011 page->mapping = (struct address_space *) anon_vma;
2012 page->index = linear_page_index(vma, address);
2015 void hugepage_add_anon_rmap(struct page *page,
2016 struct vm_area_struct *vma, unsigned long address)
2018 struct anon_vma *anon_vma = vma->anon_vma;
2021 BUG_ON(!PageLocked(page));
2023 /* address might be in next vma when migration races vma_adjust */
2024 first = atomic_inc_and_test(compound_mapcount_ptr(page));
2026 __hugepage_set_anon_rmap(page, vma, address, 0);
2029 void hugepage_add_new_anon_rmap(struct page *page,
2030 struct vm_area_struct *vma, unsigned long address)
2032 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2033 atomic_set(compound_mapcount_ptr(page), 0);
2034 __hugepage_set_anon_rmap(page, vma, address, 1);
2036 #endif /* CONFIG_HUGETLB_PAGE */